Hybrid System for fouling control in biomass boilers - CiteSeerX
ARTICLE IN PRESS
Engineering Applications of Artificial Intelligence 19 (2006) 915–925 www.elsevier.com/locate/engappai
Hybrid System for fouling control in biomass boilers Luis M. Romeo!, Raquel Gareta Centro de Investigacio´n de Recursos y Consumos Energe´ticos (CIRCE), Universidad de Zaragoza, Centro Polite´cnico Superior, Marı´a de Luna, 3, Zaragoza 50018, Spain Received 30 May 2005; received in revised form 26 January 2006; accepted 29 January 2006 Available online 23 March 2006
Abstract Renewable energy sources are essential paths towards sustainable development and CO2 emission reduction. For example, the European Union has set the target of achieving 22% of electricity generation from renewable sources by 2010. However, the extensive use of this energy source is being avoided by some technical problems as fouling and slagging in the surfaces of boiler heat exchangers. Although these phenomena were extensively studied in the last decades in order to optimize the behaviour of large coal power boilers, a simple, general and effective method for fouling control has not been developed. For biomass boilers, the feedstock variability and the presence of new components in ash chemistry increase the fouling influence in boiler performance. In particular, heat transfer is widely affected and the boiler capacity becomes dramatically reduced. Unfortunately, the classical approach of regular sootblowing cycles becomes clearly insufficient for them. Artificial Intelligence (AI) provides new means to undertake this problem. This paper illustrates a methodology based on Neural Networks (NNs) and Fuzzy-Logic Expert Systems to select the moment for activating sootblowing in an industrial biomass boiler. The main aim is to minimize the boiler energy and efficiency losses with a proper sootblowing activation. Although the NN type used in this work is well-known and the Hybrid Systems had been extensively used in the last decade, the excellent results obtained in the use of AI in industrial biomass boilers control with regard to previous approaches makes this work a novelty. r 2006 Elsevier Ltd. All rights reserved. Keywords: Biomass; Boiler fouling; Hybrid system
1. Introduction The European Union has the aim to increase the contribution of renewable energy sources up to 22% of electricity generation from renewable sources by 2010. Intensive research is being carried out in order to take advantage of biomass potential as a renewable energy source. The behaviour of the medium-sized biomass boilers started up recently shows a dramatically tendency to a rapid fouling of the superheater surfaces that reduce the boiler capacity (Zevenhoven-Onderwater et al., 2000). Therefore, the development of large biomass boilers and the wide use of the biomass to produce electricity is being questioned. Due to the inherent complexity of the biomass !Corresponding author. Tel.: +34 976 762570; fax: +34 976 732078.
E-mail address: [email protected] (L.M. Romeo). 0952-1976/$ - see front matter r 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.engappai.2006.01.019
fouling, this problem has not been completely solved. Biomass fouling not only influences the steam production, but also produces other important effects on boiler performance. A boiler efficiency reduction, an increase in carbon dioxide emissions, or a deteriorated combustion behaviour with higher combustion temperatures and an extra formation of nitrogen oxides and carbon monoxide have been reported. The mechanism of fouling formation is concerned with the high formation of incombustible volatiles belonging to several biomass components (Sami et al., 2000). The condensation of these vapours over the superheaters tubes creates a stick film for the solid soot particles. The increasing layer reduces heat transfer and evidently steam production, also boiler efficiency decreases dramatically (Heinzel et al., 1998; Peltola et al., 1999; Skrifvars et al., 1998; Dare, 2002). Although the principles and mechanisms that cause these phenomena have been clearly
ARTICLE IN PRESS L.M. Romeo, R. Gareta / Engineering Applications of Artificial Intelligence 19 (2006) 915–925
established, the inherent complexity of fouling and the uncertain evolution of mineral constituents have complicated to solve the problem. During the last decades, the fouling formation mechanisms in coal power plants has been widely analysed (Anthony et al., 2001; Fan et al., 2001; Lee et al., 2002; Wang and Harb, 1997; Bergeles et al., 1997). This includes different point of views: chemical fuel analysis, complex CFD simulations and experiments in pilot-sized boilers. Traditional methods for coal boilers to reduce boiler fouling (i.e. sootblowing) are not suitable without a proper boiler evaluation. In this case research has been focussed in boiler monitoring by means of special measuring instruments and on-line calculations. Monitoring could be accomplished by means of standard and special power plant instrumentation or a combination of both techniques. In any case external software is required for simulation and calculations. Market has some examples of boiler monitoring, especially for fouling phenomena (Electric Power Research Institute, 2000; Energy Research Center, 2005; BMS International Ltd., 2005), but there is a lack of information about the internal behaviour of these applications. Biomass fouling approaches use experimental correlations, based on biomass chemical analysis to estimate fouling probability (Jenkins et al., 1998; Seggiani, 1999). However, these correlations could establish the tendency to fouling of a selected biomass, but neither solve the problem, nor develop a methodology to operate the boiler minimizing the fouling effects. The introduction of the Neural Networks (NNs) to solve other problems concerned to power plants, power systems modelling, control and forecasting in energy systems has been progressively reported elsewhere (Olofsson et al., 1998; Lu and Hogg, 2000; Mihalakakou et al., 2000; Bechtler et al., 2001; Gareta et al., 2006). The application of the Artificial Intelligence (AI) techniques to renewable energies (Kalogirou, 2001, 2003) and to the evaluation of fouling deposits evolution inside the boilers (Afgan et al., 1996; Lalot and Lecoeuche, 2003) also have been reported. However, an integrated control methodology that evaluates actual and future boiler performance depending on sootblowing (and fouling) and decide to clean heat transfer surfaces has not been already developed. The present work is conceived as a contribution to this question. The objective of this paper is to present the application of a Hybrid System that combines NNs advantages with Fuzzy-Logic Expert System (FLES) to control boiler fouling and optimize boiler performance, minimizing the effect of fouling. This Hybrid System uses several sets of NN with different objectives: boiler monitoring, fouling forecasting, prediction of boiler behaviour and the cleaning effect if a sootblowing cycle was activated. The application is completed by the development of a FLES that takes the sootblowing decision. Validation shows important energy saving between Hybrid System outputs and real data obtained from a biomass boiler.
2. Problem description Fouling and slagging cause a severe reduction of the boiler heat transfer, these phenomena were firstly observed in coal boilers, specially in those using brown coals. But the medium-sized biomass boilers started up recently report a performance strongly affected by fouling/slagging problem, as pointed out in Fig. 1. A design load of 74.6 ton/h is reduced, at the end of the operating period (hour 4200), approximately 30%. The boiler uses biomass with a moving inclined grate. It produces a steam output of 74.6 ton/h with clean surfaces. The actual fuel is a mix of different types of biomass (bark and wood of indeterminate species) with 48.5% of moisture, 3% of ash and an average heating value of 8.5 MJ/kg. The boiler includes furnace, two superheaters, two convective evaporators and two economizers. The steam output temperature from the first superheater is 276 1C and the steam output temperature from the final superheater is 475 1C. Eight sootblowers on the superheater area and six sootblowers in the convection area are installed in the boiler, as shown in Fig. 2. Two automatic sootblowing cycles are possible. In short cycles, only the sootblowers in the superheater area are operated, but superheater and convective area sootblowers become activated in long cycles. One sootblowing cycle is usually performed per working shift; therefore, total steam consumption for sootblowing is 9.7 ton/day if three long cycles of sootblowing were performed. Despite this steam consumption, the fouling effect increases with operational time as is shown in Fig. 1. 3. Hybrid System design The introduction of the NN and AI techniques to analyse energy systems as previously commented has stimulated the development of an Hybrid System to minimize fouling influence. It includes a combination of NNs and a FLES to simulate, predict and control the biomass boiler fouling. 85 80 75 STEAM (ton/h)
916
70 65
30%
60 55 50 45 40
0
1000
2000
3000 4000 HOURS
5000
6000
7000
Fig. 1. Steam mass flow (ton/h) evolution over an operating period.
ARTICLE IN PRESS L.M. Romeo, R. Gareta / Engineering Applications of Artificial Intelligence 19 (2006) 915–925
917
LIVED STEAM
SUPERHEATER 2
DRUM
SUPERHEATER 1
SOOTBLOWER
NEW EVAPORATOR
CONVECTIVE EVAPORATOR
ECONOMIZER
FURNACE CIRCULATING ECONOMIZER
AIR PREHEATER
FEEDING WATER
Fig. 2. Sootblowers and heat transfer equipment in biomass boiler.
A fundamental step in the design of the fouling predicting and control system has been the development of an accurate thermal boiler simulation. The use of NN instead of other thermodynamic-based simulation has introduced important advantages as a better knowledge of the boiler performance based on the operation data and simple and accurate model to be used in on-line applications. Based on this simulation a design of different set of NN has been developed as shown in Fig. 3:
!
! !
Thermal monitoring system. An NN monitoring system based on thermal boiler simulation is needed in order to manage historic and on-line data, calculate boiler fouling and fouling indexes, select the objective variables and prepare the data to train following NN. Boiler fouling level evolution. Based on previous NN, historical data and boiler simulation, fouling evolution indexes with and without sootblowers operation are predicted. Evaluation of boiler thermal response with and without sootblowing. Once fouling evolution with and without
!
sootblowers operation is known, another NN set based on boiler simulations forecasts the energy responses in both scenarios to know the hourly energy improvements. Sootblowing decision maker. Previous information is used by Fuzzy-Logic System that decide the right time and type of sootblowing cycle.
This novel architecture is entirely conceived to optimize sootblower activation strategy taking into account the NN and Fuzzy-Logic advantages. This design has been developed to avoid the classical approach that estimate fouling evolution through a time exponential decay equation (Kern and Seaton, 1959; Allmon et al., 1991) with several coefficients calculated with complicated NN architectures. Moreover, the problem has been reinterpreted in order to use a wide operation plant data collection and to develop a strong and efficient prediction system that could be used in on-line applications. The developed NN architecture is a basic multilayer feedforward NN. It is a simple and well-known type of NN but it was chosen because it fits perfectly with the problem
ARTICLE IN PRESS 918
L.M. Romeo, R. Gareta / Engineering Applications of Artificial Intelligence 19 (2006) 915–925 POWER BOILER INSTRUMENTATION
NEURAL NETWORK-1
MONITORIZING SYSTEM
FOULING / SLAGGING SENSIBLE INDEX
HYBRID SYSTEM
NEURAL NETWORK -2 RED NEURONAL NEURAL NETWOR KN-2
NEURAL NETWORK -2 RED NEURONAL EURAL NET WORK -2
FOULING / SLAGGING SENSIBLE INDEX TEMPORAL SERIE SOOT BLOWING ACTIVATION
FORECASTING
FOULING / SLAGGING SENSIBLE INDEX TEMPORAL SERIE NOT SOOT BLOWING ACTIVATION
NEURAL EURAL NETNETWORK WORK -3 RED NEURONAL
NEURAL NETWOR KN RED NEURONAL -3 NEURAL NETWORK N-3
MODEL HEAT RECOVERY IN A FIXED TEMPORAL HORIZONT
HEAT RECOVERY IN A FIXED TEMPORAL HORIZONT
SOOT BLOWING ACTIVATION
NOT SOOT BLOWING ACTIVATION
FUZZYLOG LOGIC FUZZY IC EXPERT SYSTEM SYSTEM EXPERT
DECISION MAKER
DECISION
Fig. 3. Biomass boiler control system design.
to solve. Feedforward NN shows a great process capacity with a robust performance specially suitable to dealing with non-filtered and on-line real process data (Olofsson et al., 1998; Bechtler et al., 2001). Moreover, non-linear modelling of time series and forecasting utilization (Olofsson et al., 1998; Lu and Hogg, 2000; Mihalakakou et al., 2000) are another important NN characteristics that are appropriate for this problem. The general idea of the system is not to design a strong protocol to manage data for developing a ‘‘white box’’ (where the connection between inputs and outputs could be expressed using equations), but allow NN to deal with on-line real process data by means of a ‘‘grey box’’ model (Ikonen et al., 2000). In ‘‘grey box’’ models theoretical information is allowed to be introduced by the way of selecting an architecture of the global NN where the relationship between variables was implicit. Moreover, since irrelevant inputs have been reported of causing overfitting in the NN, special attention has been paid in inputs selection (Ikonen et al., 2000; Lertpalangsunti et al., 1999; Prieto et al., 2001). 4. Biomass boiler monitoring. Fouling calculation by NN Thermal monitoring is the basis of the knowledge of boiler performance and the design of the sootblowing
control system. An NN simulation module has been developed in order to simplify the control design and make stronger the boiler simulation, taking advantage of the great qualities of the NN. A set of simple NN was developed in three levels corresponding with thermal simulation strategy: combustion gases composition calculated from stack and fuel analysis, heat transferred in each exchanger and fouling index, as shown in Fig. 4. NN has the same structure and input variables as a thermal simulation and as much physical concepts as possible have been introduced in its design. Each NN has three layers of neurons following a feedforward architecture. A sigmoid activation function in the neuron hidden-layer and a linear one in the neuron output layer have been selected. Each NN has been trained independently following an iterative scheme. Non-relevant inputs have been identified and non-considered, 3000 data selected randomly have been used in the training process. To validate each NN, resubstitution test (with the 3000 training data) and resistance test (with a reserved group of 1455 data) have been accomplished in order to verify the accuracy without overfitting. Special attention has been paid in the input selection. Since a negative influence of an excessive researcher’s
ARTICLE IN PRESS 919
L.M. Romeo, R. Gareta / Engineering Applications of Artificial Intelligence 19 (2006) 915–925
%H2O fuel
% O2 stack % N2 % CO2 stack % O2 boiler
% H2O
% H2O stack
% CO2
Fuel mass flow Gases mass flow Air mass flow Furnace heat Final steam output SH 2 heat Final steam temperature SH 1 heat Steam temp. after atemp. Evapor. heat
Gases temp. Furnace exit
Econom. heat
Gases temp after econom
Furnace fouling index
Gases temp boiler exit Feed water temperature
SH 2 fouling index
Steam temp. before atemp.
SH 1 fouling index
Water temp before aire PH Primary aire Mass flow
Evapor fouling index
Air temp before aire PH.
Econom fouling index
Air temp after aire PH
Fig. 4. Relation between variables of the three monitoring NN sets.
subjectivity should be avoided, the selection of the inputs was made using only tools concerned with NN. In the first training of each NN, all the available inputs (23 variables)
were used and the mean square error (MSE) obtained was registered. After that, a variable was eliminated in every new training executed and the MSE obtained was also
ARTICLE IN PRESS 920
L.M. Romeo, R. Gareta / Engineering Applications of Artificial Intelligence 19 (2006) 915–925
Table 1 Training and validation error values of the fouling index monitoring NN Variable
Neurons number
Resubstitution average error (%)
Resistance average error (%)
Furnace fouling index SH 2 fouling index SH 1 fouling index Convective evaporator fouling index Economizer fouling index
3 4 3 4 6
1.03 1.60 1.08 2.35 0.86
1.01 1.54 1.09 2.36 0.88
registered and compared with the value obtained in the first training. The higher the influence of the absent input, the more the increase in the MSE value, and therefore the eliminated input variable was more relevant (Sung, 1998). This analysis of the input’s influence is an iterative task and time-consuming, but the highly logical input variables obtained compared with thermo-physical relations indicate a strong and coherent NN input variable selection. As an example, to calculate the exhaust gases composition, input variables related with exit gases composition in the stack and the moisture in the boiler have been selected, but those related with temperature and pressure signals have been rejected. Main results include new variables not collected in power plant data acquisition system as heat transferred from the combustion gases to the steam or water in each section, intermediate exhaust gas temperatures, steam mass flows in the heat exchangers and variables that include the fouling influence in heat transfer (fouling index). Every NN has been individually validated under two strategies in order to reject NN overfittings. Firstly, 3000 training output data were compared with the NN results (resubstitution test) concluding a good agreement with training data. Secondly, the comparison was made with 1495 data reserved from the data collection and nonused in the training stage. A resistance test was made in this way pointing out clearly any possible overfitting (Reich and Barai, 2000). The validation with the resubstitution–resistance methodology is achieved when the errors obtained with this analysis are similar, as shown in Table 1. Despite the low number of neurons, results clearly show a non-overfitting NN. Only a slightly more relevant deviation was pointed out in the convective evaporator fouling index, because of the lack of a representative data set. Since the output data obtained from NN were not included in the power plant data system variables, the validation cannot follow other traditional NN validation rules. In this case a final verification was made comparing the sensitivity analysis results of the calculated variables obtained from historical data. SH1 fouling index shows a deviation of 7.5% taking into account the uncertainty of the measure instruments and, indeed, the convective evaporator fouling index could reach till 9.8%. Results in Table 1 show deviation clearly below these limits.
5. Fouling and boiler thermal response forecasting The classical approach to estimate fouling evolution tried to adjust heat transfer with a time exponential decay equation (Kern and Seaton, 1959). In consequence, the great number of physical mechanisms involved in the deposit growing avoids a good agreement between real data and the calculation of heat transfer coefficient exponential decay constants (Allmon et al., 1991). It is proposed to analyse the fouling tendency by forecasting a fouling index (ratio between real heat transfer coefficient compared with clean conditions) hourly after the sootblowing. In opposition to the classical approach, a set of NN was developed in order to calculate two time series of fouling indexes (with and without sootblowing) for 12 h just after the moment of study, Fig. 3. The available data to develop the NN come from 2 years of boiler operation. Five hundred and seventy-three sootblowings were located during this period and 161 were selected as effective and with enough data to train an NN with a prediction of 12 h. Approximately 100 sootblowing series have been used to train, that means a 60% of the data, and the rest was reserved for the validation stage. Developed NN contains three layers of NN with three layers of neurons, every one following a feedforward architecture. A sigmoid activation function has been selected for the neuron hidden-layer and a linear one for the neuron output layer. The training stage has followed the same principles exposed in the monitoring module: an NN input selection as previously described, a training based in resubstitution test and a final validation following a resistance test. This methodology allows to point out curiosities as showed in Fig. 5, where the NN architecture design evidences the reduction of the influence of actual data, while the influence of the previous fouling forecasted values increases. The trend provided by the fouling forecasting NN is highly coherent with physical concepts and can be asserted as faithful in long-term analysis. Tables 2 and 3 display the number of neurons and deviations in reconstitution and resistance validations. Results are also excellent in superheaters (in Figs. 6 and 7 the results for superheaters 1 and 2 h after sootblowing are shown); also in convective evaporators results show a coherent evolution. The following step lies in the evaluation of the fouling influence in boiler performance. Two NNs have been
ARTICLE IN PRESS 921
L.M. Romeo, R. Gareta / Engineering Applications of Artificial Intelligence 19 (2006) 915–925 Time previous soot blowing Time Tiempo frfrom omdesde previous soplad sootb o an lowing teri or Hour 00FF.I./ Hour I. E. Hora .I./ 0/ Pre vious Hour Previous Hour Hora Previa
Sootbl. Number/Previous F.I. Or Sootbl. den de Numb soplado/F er/P revious .I. Previ F.I. o
+1
Fu flow stib le Fule flow Ca ud al de le Combu
Hour 1 F.I./ I. E. Hor a 1/ Hour Pr evious Hour Previous Hora Pr Hour evia
flow CaAir udAir al de flow aire
Hour 2 F.I./ Hour F.I./ I. E. Hor a 2/ Previous Hour Hour Hora Previa
Steam Caudal por Steadeflow mva flow
+3 3 3F.I./ Hour I.Hour E. Hor aF.I./ 3/ Previous Hour Previous Hour Hora Previa
PrHora evious F.I. Previous I. E. PF.I. revia
Fig. 5. Relation between variables of the fouling forecasting NN sets.
Table 2 Training and validation error values of the fouling index forecasting NN Variable
Neurons number
Resubstitution average error (%)
Resistance average error (%)
Superheater 1 Hour 0 fouling index Hour 1 fouling index Hour 2 fouling index Fouling index in the next hour
5 6 3 3
2.79 2.04 2.01 3.54
4.97 6.79 2.80 3.98
Convective evaporators Hour 0 fouling index Hour 1 fouling index Hour 2 fouling index Fouling index in the next hour
3 7 3 3
3.80 4.21 2.66 5.32
9.57 11.17 8.80 7.22
Table 3 Training and validation error values in the boiler performance evaluation after fouling forecasting Variable
Neurons number
Resubstitution average error (%)
Resistance average error (%)
Total heat Furnace heat Superheater 2 heat Superheater 1 heat Convective evaporator heat Economizer heat Steam flow
20 20 20 20 20 20 20
2.59 2.44 3.97 2.81 2.33 2.41 2.65
3.12 3.38 4.75 3.27 2.69 2.94 3.18
carried out in order to simulate boiler performance and compare the energy recovered with or without sootblowing, Fig. 3. Results are calculated as energy improvements for 4, 6 and 8 h after evaluation. This final NN uses as input data obtained from the fouling forecasting NN (heat exchanger fouling level values). Three thousand data have been used to train and 1550 data have been used to validate this model. This NN has the same structure as those explained in the monitoring and in the forecasting module. Table 3 displays the number of neurons used in this NN set and the average resubstitution and resistance errors. NN results for training and validation are shown in Fig. 8. A good agreement is observed between calculated and real boiler heat absorption, the key variable to analyse boiler
fouling influence. Although a slightly higher dispersion in the outputs than in previous results is noticed, it is explained by the reduced number of inputs required to this set of NN, compared for instance, to those needed for a thermodynamic model. 6. Sootblowing evaluation To complete the application and obtain a Hybrid System to for fouling control in on-line applications, a decision maker based on fuzzy-logic rules has been developed. The aim of the decision maker is to evaluate from the results provided by previous NN modules the activation of the sootblowing cycle. Therefore, the decision maker module
ARTICLE IN PRESS 922
L.M. Romeo, R. Gareta / Engineering Applications of Artificial Intelligence 19 (2006) 915–925
1.2
1.2 Training data (realvs. NN) SH2 Fouling level One hour after sootblowing
1.1 1
1.1 1
0.9
0.9
0.8
0.8
0.7
0.7
0.6
0.6
0.5
0.5
0.4 0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
1.2
0.4 0.4
Validation data (real vs. NN) SH2 Fouling level One hour after sootblowing
0.5
0.6
0.7
0.8
0.9
1
1.1
1.2
1
1.1
Fig. 6. Training and validation results of the SH2 fouling level in the sootblowing cycle initial hour.
1.1
1.1 Training data (real vs. NN) SH1 Fouling level Two hours after sootblowing
1 0.9
1 0.9
0.8
0.8
0.7
0.7
0.6
0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2 0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
0.2 0.2
Validation data (real vs. NN) SH1 Fouling level Two hours after sootblowing
0.3
0.4
0.5
0.6
0.7
0.8
0.9
70000
70000 65000 60000 55000 50000 45000 40000 35000 30000 30000 35000 40000 45000 50000 55000 60000 65000 70000 Boiler heat absorption (kW) - Real value
Boiler heat absorption (kW) - NN value
Boiler heat absorption (kW) - NN value
Fig. 7. Training and validation results of the SH1 fouling level 2 h after the sootblowing cycle.
65000 60000 55000 50000 45000 40000 35000 30000 30000 35000 40000 45000 50000 55000 60000 65000 70000 Boiler heat absorption (kW) - Real value
Fig. 8. Boiler heat absorption. Training (left) and validation (right).
should be developed attending to the experience implicit in the historical data available and the decision thresholds are extracted from a detailed study of previous results, Fig. 3. A FLES was selected because of the ability of arriving to conclusions dealing with non-precise data inputs (Frantti and Ma¨ho¨nen, 2001; Chen et al., 2003). The approach is
based on previous experience and do not require a physical law to calculate the system performance, becoming an adequate tool for the control of fouling in a biomass boiler (Fig. 9). The outputs of the Hybrid System decision maker were selected as: no activation of the sootblowers, activation of
ARTICLE IN PRESS L.M. Romeo, R. Gareta / Engineering Applications of Artificial Intelligence 19 (2006) 915–925
923
Energy recovery in the next hours when activati NOT sootblowing activation
Energy recovery in the next hours when sootblowing activation
EVALUATION
Energy improvement Energy improvement Energy improvement after 4 hours after 6 hours after 8 hours
¿SOOTBLOWING ACTIVATION?
YES
¿LONG CYCLE?
NO
NO
NEW EVALUATION THE NEXT HOUR
SHORT SOOTBLOWING CYCLE ACTIVATION
YES
LONG SOOTBLOWING CYCLE ACTIVATION
Fig. 9. Fuzzy-Logic Expert System design.
a short cycle or activation of a long cycle (depending on the fouling influence in the boiler). The inputs of the decision maker were exclusively the results obtained from the NN modules and the thresholds of activation were fixed in order to guarantee energy savings. With all these requirements the membership functions and pre- and postprocessing of the data were designed. The inputs selected from previous NN set are those related to the energy improvement caused by the sootblowing cycle, specially the prediction of the accumulated energy improvement due to the sootblowing after 4, 6 and 8 h. FLES module has been designed to emphasize the improvements in the first hours in order to recover the sootblowing energy as soon as possible, taking into account that the fouling evolution could not be as regular as predicted. The first calculation made in the FLES module assigns a numeric value to the prediction achieved by the NN modules. After that, this number is compared with the threshold values and a final decision is emitted. There are two threshold values: the first establishes the difference between not advising the activation of the sootblowers or advising a short cycle, the second has been calculated in order to advise a long cycle if the energy saved is enough. Forced by the availability of historical data, the adjustment and validation of the FLES module have been made with them. Results show that approximately 1020 sootblowing cycles would be activated in an operation period (6700 h considered), which represent approximately 200 more sootblowing cycles than traditional operating manoeuvres. From these 1020 sootblowing cycles, the FLES module recommends almost 800, but the rest will be automatically carried out by the system in a security limit established in 12 h. From the 800 sootblowing cycles
Table 4 Comparison of energy values in the application of NN–FL methodology to control boiler sootblowing cycles Energy values for the biomass boiler Average heat output produced in the boiler Energy recovered in historical sootblowing strategy Energy recovered in NN-FL sootblowing strategy Energy savings from historical to optimized strategy Percentage of energy savings Average heat output produced in the boiler
54.1 MW "7556 MWh +4953 MWh +12,509 MWh +3.45% 56.0 MW
recommended by the FLES module, 670 of them are long cycle type. Although it is known that the convective section is not a critical area, this also means a good level of cleanliness in it. Finally, it is really worth comparing the power production by the biomass boiler performing with historical sootblowing strategy and with the new strategy, Table 4. Previous strategy caused a decrease in energy saving of 7556.4 MWh due the sootblowing activation when the boiler was nearly clean or too fouled to be cleaned by steam, as Fig. 1 shows. The NN–FLES strategy shows an increase of energy saving up to 12509.0 MWh in an operational period. This means a power production augmentation of approximately 3.5%, from an average of heat produced as steam of 54.1 MW to a new average of 56.0 MW. 7. Conclusions There are some difficulties for extensive biomass utilization. In particular, biomass combustion produces fouling
ARTICLE IN PRESS 924
L.M. Romeo, R. Gareta / Engineering Applications of Artificial Intelligence 19 (2006) 915–925
in boiler heat transfer equipment, causing a reduction of steam output and boiler efficiency. The traditional approach uses experimental correlations but the problem has not been solved due to the inherent complexity of fouling and it is necessary to develop new techniques and strategies to minimize its consequences. The application of a Hybrid System combining NNs and a FLES has allowed to control boiler fouling and optimize its performance, although it has obligated to a new approach to the problem description. Different sets of NN have been used to monitor the boiler, to forecast fouling and the consequent future boiler performance. A set of fuzzy-logic rules based on real data that takes as inputs the results from previous NN has optimized boiler cleaning cycles and minimized fouling evolution. A procedure that compares the energy cost of a sootblowing cycle with the thermal improvements has been developed and it decides the sootblowing activation. NN and FLES system has been designed and trained obtaining encouraging results. Especially considering that any sootblowers control based on the state of the boiler has not been reported previously. A relative comparison between the old routine of sootblowing strategy and the new strategy programmed in the NN–FLES shows that approximately 3.5% of power production improvement can be obtained. The relevance of these results shows that techniques of AI as those proposed in this paper (NN–FLES) could help to reorient the study of complex phenomena, providing new point of views and tools to develop strong and easy-to-use control systems. Acknowledgements The work described in this paper was supported by the Fifth Framework Programme of the European Union, FP5–Energy, Environment and Sustainable Development, NNE5-2001-00128, Intelligent Process Control System for Biomass Fuelled industrial Power Plants. The authors are solely responsible for this article and it does not represent the opinion of the European Community. We gratefully acknowledge the inputs, collaboration and patience of the members of the INTCON project, TPS Terminska Processer AB, Tecnatom and CINAR. References Afgan, N., Carvalho, M.G., Coelho, P., 1996. Concept of expert system for boiler fouling assessment. Applied Thermal Engineering 16 (10), 835–844. Allmon, B.A., Watson, G.B., Carpenter, N.N., 1991. Fouling and enhancement interactions. ASME HTD-164, 61–70. Anthony, E.J., Iribarne, A.P., Iribarne, J.V., Talbot, R., Jia, L., Granatstein, D.L., 2001. Fouling in a 160 MWe FBC boiler firing coal and petroleum coke. Fuel 80 (7), 1009–1014. Bechtler, H., Browne, M.W., Bansal, P.K., Kecman, V., 2001. New approach to dynamic modelling of vapour-compression liquid chillers, artificial neural networks. Applied Thermal Engineering 21 (9), 941–953.
Bergeles, G., Bouris, D., Yianneskis, M., Balabani, S., Kravaritis, A., Itskos, S., 1997. Effects of fouling on the efficiency of heat exchangers in lignite utility boilers. Applied Thermal Engineering 17 (8–10), 739–749. BMS International Ltd., 2005. Intelligent Sootblower Control Systems for Efficient Power Generation. Currently available at http:// www.bmsint.com/power/index.html Chen, W.C., Chang, N.B., Chen, J.C., 2003. Rough set-based hybrid fuzzy-neural controller design for industrial wastewater treatment. Water Research 37 (1), 95–107. Dare, P., 2002. Combustion performance of biomass residue and purpose grown species. Fuel and Energy Abstracts 43 (4), 276. Electric Power Research Institute, 2000. Guidelines for intelligent sootblower control. Final Report, Report no. 1000410. Energy Research Center, Lehigh University, 2005. Optimize Sootblowing to Reduce NOx and Improve Heat Rate. Currently available at http:// www3.lehigh.edu/research/recenters.asp Fan, J.R., Zha, X.D., Sun, P., Cen, K.F., 2001. Simulation of ash deposit in a pulverized coal-fired boiler. Fuel 80 (5), 645–654. Frantti, T., Ma¨ho¨nen, P., 2001. Fuzzy logic-based forecasting model. Engineering Applications of Artificial Intelligence 14 (2), 189–201. Gareta, R., Romeo, L.M., Gil, A., 2006. Forecasting of electricity prices with neural networks. Energy Conversion and Management Journal. Currently available at http://www.sciencedirect.com Heinzel, T., Siegle, V., Spliethoff, H., Hein, K.R.G., 1998. Investigation of slagging in pulverized fuel co-combustion of biomass and coal at a pilot-scale test facility. Fuel Processing Technology 54 (1–3), 109–125. Ikonen, E., Najim, K., Kortela, U., 2000. Neuro-fuzzy modelling of power plant flue-gas emissions. Engineering Applications of Artificial Intelligence 13 (6), 705–717. Jenkins, B.M., Baxter, L.L., Miles Jr., T.R., Miles, T.R., 1998. Combustion properties of biomass. Fuel Processing Technology 54 (1–3), 17–46. Kalogirou, S.A., 2001. Artificial neural networks in renewable energy systems applications, a review. Renewable and Sustainable Energy Reviews 5, 373–401. Kalogirou, S.A., 2003. Artificial intelligence for the modelling and control of combustion processes, a review. Progress in Energy and Combustion Science 29, 515–566. Kern, D.Q., Seaton, R.E., 1959. A theoretical analysis of thermal surface fouling. British Chemical Engineering 4, 258–262. Lalot, S., Lecoeuche, S., 2003. Online fouling detection in electrical circulation heaters using neural networks. International Journal of Heat and Mass Transfer 46, 2445–2457. Lee, B.E., Fletcher, C.A.J., Shin, S.H., Kwon, S.B., 2002. Computational study of fouling deposit due to surface-coated particles in coal-fired power utility boilers. Fuel 81 (15), 2001–2008. Lertpalangsunti, N., Chan, C.W., Mason, R., Tontiwachwuthikul, P., 1999. A toolset for construction of hybrid intelligent forecasting systems, application for water demand prediction. Artificial Intelligence in Engineering 13 (1), 21–42. Lu, S., Hogg, B.W., 2000. Dynamic nonlinear modelling of power plant by physical principles and neural networks. International Journal of Electrical Power & Energy Systems 22 (1), 67–78. Mihalakakou, G., Santamouris, M., Tsangrassoulis, A., 2000. On the energy consumption in residential buildings. Energy and Buildings 34 (7), 727–736. Olofsson, T., Andersson, S., O¨stin, R., 1998. A method for predicting the annual building heating demand based on limited performance data. Energy and Buildings 28 (1), 101–108. Peltola, K., Hiltunen, M., Blomqvist, J.P., Skrifvars, B.J., Kurkela, J., Latva-Somppi, J., Kauppinen, E.I., 1999. Fouling of the surfaces in biofueled-fired fluidized bed boilers. In: Proceedings of the 15th International Conference on Fluidized Bed Combustion. ASME. Prieto, M., Vallina, J.M., Sua´rez, I., Martı´ n, I., 2001. Application of a design code for estimating fouling on-line in a power plant condenser
ARTICLE IN PRESS L.M. Romeo, R. Gareta / Engineering Applications of Artificial Intelligence 19 (2006) 915–925 cooled by seawater. Experimental Thermal and Fluid Science 25 (5), 329–336. Reich, Y., Barai, S.V., 2000. A methodology for building neural networks models from empirical engineering data. Engineering Applications of Artificial Intelligence 13 (6), 685–694. Sami, M., Annamalai, K., Wooldridge, M., 2000. Co-ring of coal and biomass fuel blends. Progress in Energy and Combustion Science 27, 171–214. Seggiani, M., 1999. Empirical correlations of the ash fusion temperatures and temperature of critical viscosity for coal and biomass ashes. Fuel 78 (9), 1121–1125. Skrifvars, B.J., Backman, R., Hupa, M., Sfiris, G., A˚byhammar, T., Lyngfelt, A., 1998. Ash behaviour in a CFB boiler during combustion of coal, peat or wood. Fuel 77 (1–2), 65–70. Sung, A.H., 1998. Ranking importance of input parameters of neural networks. Expert Systems with Applications 15 (3–4), 405–411. Wang, H., Harb, J.N., 1997. Modelling of ash deposition in large-scale combustion facilities burning pulverized coal. Progress in Energy and Combustion Science 23 (3), 267–282. Zevenhoven-Onderwater, M., Blomquist, J.P., Skrifvars, B.J., Backman, R., Hupa, M., 2000. The prediction of behaviour of ashes from five different solid fuels in fluidised bed combustion. Fuel 79 (11), 1353–1361.
925
Luis M. Romeo (Ph.D. Eng.) is Project Manager of the Centre of Research for Energy Resources and Consumption. He also teaches Heat Transfer and Thermodynamics in the Department of Mechanical Engineering, University of Zaragoza. He gained his professional experiences developing and managing several projects in different Spanish power plants such as the Pressurised Fluidised Bed Combustion Power Plant at Escatro´n, Pulverized Coal Teruel Power Plant, Combined Cycles and Biomass Boilers. He works on R&D projects and in activities related with efficiency improvements, power plant modelisation and CO2 capture. Ms. Raquel Gareta (Ph.D. Eng.) was a CIRCE researcher. She studied Mechanical Engineering at the University of Zaragoza with major speciality in Heat Transfer and Fluids Mechanics. She has worked on R&D projects since August 1999 in activities related with Gas Turbine performance optimisation, Neural Networks development, Fouling and Slagging influence minimization and Biomass Boilers improvement. Her Thesis ‘‘INTELLIGENT MODEL BASED SUPERVISION SYSTEM TO BIOMASS BOILER SOOTBLOWING CYCLES MANAGEMENT’’ has been one of the winners of the European Talent Award for Innovative Energy Systems 2005 (European Foundation for Power Engineering).
Engineering Applications of Artificial Intelligence 19 (2006) 915–925 www.elsevier.com/locate/engappai
Hybrid System for fouling control in biomass boilers Luis M. Romeo!, Raquel Gareta Centro de Investigacio´n de Recursos y Consumos Energe´ticos (CIRCE), Universidad de Zaragoza, Centro Polite´cnico Superior, Marı´a de Luna, 3, Zaragoza 50018, Spain Received 30 May 2005; received in revised form 26 January 2006; accepted 29 January 2006 Available online 23 March 2006
Abstract Renewable energy sources are essential paths towards sustainable development and CO2 emission reduction. For example, the European Union has set the target of achieving 22% of electricity generation from renewable sources by 2010. However, the extensive use of this energy source is being avoided by some technical problems as fouling and slagging in the surfaces of boiler heat exchangers. Although these phenomena were extensively studied in the last decades in order to optimize the behaviour of large coal power boilers, a simple, general and effective method for fouling control has not been developed. For biomass boilers, the feedstock variability and the presence of new components in ash chemistry increase the fouling influence in boiler performance. In particular, heat transfer is widely affected and the boiler capacity becomes dramatically reduced. Unfortunately, the classical approach of regular sootblowing cycles becomes clearly insufficient for them. Artificial Intelligence (AI) provides new means to undertake this problem. This paper illustrates a methodology based on Neural Networks (NNs) and Fuzzy-Logic Expert Systems to select the moment for activating sootblowing in an industrial biomass boiler. The main aim is to minimize the boiler energy and efficiency losses with a proper sootblowing activation. Although the NN type used in this work is well-known and the Hybrid Systems had been extensively used in the last decade, the excellent results obtained in the use of AI in industrial biomass boilers control with regard to previous approaches makes this work a novelty. r 2006 Elsevier Ltd. All rights reserved. Keywords: Biomass; Boiler fouling; Hybrid system
1. Introduction The European Union has the aim to increase the contribution of renewable energy sources up to 22% of electricity generation from renewable sources by 2010. Intensive research is being carried out in order to take advantage of biomass potential as a renewable energy source. The behaviour of the medium-sized biomass boilers started up recently shows a dramatically tendency to a rapid fouling of the superheater surfaces that reduce the boiler capacity (Zevenhoven-Onderwater et al., 2000). Therefore, the development of large biomass boilers and the wide use of the biomass to produce electricity is being questioned. Due to the inherent complexity of the biomass !Corresponding author. Tel.: +34 976 762570; fax: +34 976 732078.
E-mail address: [email protected] (L.M. Romeo). 0952-1976/$ - see front matter r 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.engappai.2006.01.019
fouling, this problem has not been completely solved. Biomass fouling not only influences the steam production, but also produces other important effects on boiler performance. A boiler efficiency reduction, an increase in carbon dioxide emissions, or a deteriorated combustion behaviour with higher combustion temperatures and an extra formation of nitrogen oxides and carbon monoxide have been reported. The mechanism of fouling formation is concerned with the high formation of incombustible volatiles belonging to several biomass components (Sami et al., 2000). The condensation of these vapours over the superheaters tubes creates a stick film for the solid soot particles. The increasing layer reduces heat transfer and evidently steam production, also boiler efficiency decreases dramatically (Heinzel et al., 1998; Peltola et al., 1999; Skrifvars et al., 1998; Dare, 2002). Although the principles and mechanisms that cause these phenomena have been clearly
ARTICLE IN PRESS L.M. Romeo, R. Gareta / Engineering Applications of Artificial Intelligence 19 (2006) 915–925
established, the inherent complexity of fouling and the uncertain evolution of mineral constituents have complicated to solve the problem. During the last decades, the fouling formation mechanisms in coal power plants has been widely analysed (Anthony et al., 2001; Fan et al., 2001; Lee et al., 2002; Wang and Harb, 1997; Bergeles et al., 1997). This includes different point of views: chemical fuel analysis, complex CFD simulations and experiments in pilot-sized boilers. Traditional methods for coal boilers to reduce boiler fouling (i.e. sootblowing) are not suitable without a proper boiler evaluation. In this case research has been focussed in boiler monitoring by means of special measuring instruments and on-line calculations. Monitoring could be accomplished by means of standard and special power plant instrumentation or a combination of both techniques. In any case external software is required for simulation and calculations. Market has some examples of boiler monitoring, especially for fouling phenomena (Electric Power Research Institute, 2000; Energy Research Center, 2005; BMS International Ltd., 2005), but there is a lack of information about the internal behaviour of these applications. Biomass fouling approaches use experimental correlations, based on biomass chemical analysis to estimate fouling probability (Jenkins et al., 1998; Seggiani, 1999). However, these correlations could establish the tendency to fouling of a selected biomass, but neither solve the problem, nor develop a methodology to operate the boiler minimizing the fouling effects. The introduction of the Neural Networks (NNs) to solve other problems concerned to power plants, power systems modelling, control and forecasting in energy systems has been progressively reported elsewhere (Olofsson et al., 1998; Lu and Hogg, 2000; Mihalakakou et al., 2000; Bechtler et al., 2001; Gareta et al., 2006). The application of the Artificial Intelligence (AI) techniques to renewable energies (Kalogirou, 2001, 2003) and to the evaluation of fouling deposits evolution inside the boilers (Afgan et al., 1996; Lalot and Lecoeuche, 2003) also have been reported. However, an integrated control methodology that evaluates actual and future boiler performance depending on sootblowing (and fouling) and decide to clean heat transfer surfaces has not been already developed. The present work is conceived as a contribution to this question. The objective of this paper is to present the application of a Hybrid System that combines NNs advantages with Fuzzy-Logic Expert System (FLES) to control boiler fouling and optimize boiler performance, minimizing the effect of fouling. This Hybrid System uses several sets of NN with different objectives: boiler monitoring, fouling forecasting, prediction of boiler behaviour and the cleaning effect if a sootblowing cycle was activated. The application is completed by the development of a FLES that takes the sootblowing decision. Validation shows important energy saving between Hybrid System outputs and real data obtained from a biomass boiler.
2. Problem description Fouling and slagging cause a severe reduction of the boiler heat transfer, these phenomena were firstly observed in coal boilers, specially in those using brown coals. But the medium-sized biomass boilers started up recently report a performance strongly affected by fouling/slagging problem, as pointed out in Fig. 1. A design load of 74.6 ton/h is reduced, at the end of the operating period (hour 4200), approximately 30%. The boiler uses biomass with a moving inclined grate. It produces a steam output of 74.6 ton/h with clean surfaces. The actual fuel is a mix of different types of biomass (bark and wood of indeterminate species) with 48.5% of moisture, 3% of ash and an average heating value of 8.5 MJ/kg. The boiler includes furnace, two superheaters, two convective evaporators and two economizers. The steam output temperature from the first superheater is 276 1C and the steam output temperature from the final superheater is 475 1C. Eight sootblowers on the superheater area and six sootblowers in the convection area are installed in the boiler, as shown in Fig. 2. Two automatic sootblowing cycles are possible. In short cycles, only the sootblowers in the superheater area are operated, but superheater and convective area sootblowers become activated in long cycles. One sootblowing cycle is usually performed per working shift; therefore, total steam consumption for sootblowing is 9.7 ton/day if three long cycles of sootblowing were performed. Despite this steam consumption, the fouling effect increases with operational time as is shown in Fig. 1. 3. Hybrid System design The introduction of the NN and AI techniques to analyse energy systems as previously commented has stimulated the development of an Hybrid System to minimize fouling influence. It includes a combination of NNs and a FLES to simulate, predict and control the biomass boiler fouling. 85 80 75 STEAM (ton/h)
916
70 65
30%
60 55 50 45 40
0
1000
2000
3000 4000 HOURS
5000
6000
7000
Fig. 1. Steam mass flow (ton/h) evolution over an operating period.
ARTICLE IN PRESS L.M. Romeo, R. Gareta / Engineering Applications of Artificial Intelligence 19 (2006) 915–925
917
LIVED STEAM
SUPERHEATER 2
DRUM
SUPERHEATER 1
SOOTBLOWER
NEW EVAPORATOR
CONVECTIVE EVAPORATOR
ECONOMIZER
FURNACE CIRCULATING ECONOMIZER
AIR PREHEATER
FEEDING WATER
Fig. 2. Sootblowers and heat transfer equipment in biomass boiler.
A fundamental step in the design of the fouling predicting and control system has been the development of an accurate thermal boiler simulation. The use of NN instead of other thermodynamic-based simulation has introduced important advantages as a better knowledge of the boiler performance based on the operation data and simple and accurate model to be used in on-line applications. Based on this simulation a design of different set of NN has been developed as shown in Fig. 3:
!
! !
Thermal monitoring system. An NN monitoring system based on thermal boiler simulation is needed in order to manage historic and on-line data, calculate boiler fouling and fouling indexes, select the objective variables and prepare the data to train following NN. Boiler fouling level evolution. Based on previous NN, historical data and boiler simulation, fouling evolution indexes with and without sootblowers operation are predicted. Evaluation of boiler thermal response with and without sootblowing. Once fouling evolution with and without
!
sootblowers operation is known, another NN set based on boiler simulations forecasts the energy responses in both scenarios to know the hourly energy improvements. Sootblowing decision maker. Previous information is used by Fuzzy-Logic System that decide the right time and type of sootblowing cycle.
This novel architecture is entirely conceived to optimize sootblower activation strategy taking into account the NN and Fuzzy-Logic advantages. This design has been developed to avoid the classical approach that estimate fouling evolution through a time exponential decay equation (Kern and Seaton, 1959; Allmon et al., 1991) with several coefficients calculated with complicated NN architectures. Moreover, the problem has been reinterpreted in order to use a wide operation plant data collection and to develop a strong and efficient prediction system that could be used in on-line applications. The developed NN architecture is a basic multilayer feedforward NN. It is a simple and well-known type of NN but it was chosen because it fits perfectly with the problem
ARTICLE IN PRESS 918
L.M. Romeo, R. Gareta / Engineering Applications of Artificial Intelligence 19 (2006) 915–925 POWER BOILER INSTRUMENTATION
NEURAL NETWORK-1
MONITORIZING SYSTEM
FOULING / SLAGGING SENSIBLE INDEX
HYBRID SYSTEM
NEURAL NETWORK -2 RED NEURONAL NEURAL NETWOR KN-2
NEURAL NETWORK -2 RED NEURONAL EURAL NET WORK -2
FOULING / SLAGGING SENSIBLE INDEX TEMPORAL SERIE SOOT BLOWING ACTIVATION
FORECASTING
FOULING / SLAGGING SENSIBLE INDEX TEMPORAL SERIE NOT SOOT BLOWING ACTIVATION
NEURAL EURAL NETNETWORK WORK -3 RED NEURONAL
NEURAL NETWOR KN RED NEURONAL -3 NEURAL NETWORK N-3
MODEL HEAT RECOVERY IN A FIXED TEMPORAL HORIZONT
HEAT RECOVERY IN A FIXED TEMPORAL HORIZONT
SOOT BLOWING ACTIVATION
NOT SOOT BLOWING ACTIVATION
FUZZYLOG LOGIC FUZZY IC EXPERT SYSTEM SYSTEM EXPERT
DECISION MAKER
DECISION
Fig. 3. Biomass boiler control system design.
to solve. Feedforward NN shows a great process capacity with a robust performance specially suitable to dealing with non-filtered and on-line real process data (Olofsson et al., 1998; Bechtler et al., 2001). Moreover, non-linear modelling of time series and forecasting utilization (Olofsson et al., 1998; Lu and Hogg, 2000; Mihalakakou et al., 2000) are another important NN characteristics that are appropriate for this problem. The general idea of the system is not to design a strong protocol to manage data for developing a ‘‘white box’’ (where the connection between inputs and outputs could be expressed using equations), but allow NN to deal with on-line real process data by means of a ‘‘grey box’’ model (Ikonen et al., 2000). In ‘‘grey box’’ models theoretical information is allowed to be introduced by the way of selecting an architecture of the global NN where the relationship between variables was implicit. Moreover, since irrelevant inputs have been reported of causing overfitting in the NN, special attention has been paid in inputs selection (Ikonen et al., 2000; Lertpalangsunti et al., 1999; Prieto et al., 2001). 4. Biomass boiler monitoring. Fouling calculation by NN Thermal monitoring is the basis of the knowledge of boiler performance and the design of the sootblowing
control system. An NN simulation module has been developed in order to simplify the control design and make stronger the boiler simulation, taking advantage of the great qualities of the NN. A set of simple NN was developed in three levels corresponding with thermal simulation strategy: combustion gases composition calculated from stack and fuel analysis, heat transferred in each exchanger and fouling index, as shown in Fig. 4. NN has the same structure and input variables as a thermal simulation and as much physical concepts as possible have been introduced in its design. Each NN has three layers of neurons following a feedforward architecture. A sigmoid activation function in the neuron hidden-layer and a linear one in the neuron output layer have been selected. Each NN has been trained independently following an iterative scheme. Non-relevant inputs have been identified and non-considered, 3000 data selected randomly have been used in the training process. To validate each NN, resubstitution test (with the 3000 training data) and resistance test (with a reserved group of 1455 data) have been accomplished in order to verify the accuracy without overfitting. Special attention has been paid in the input selection. Since a negative influence of an excessive researcher’s
ARTICLE IN PRESS 919
L.M. Romeo, R. Gareta / Engineering Applications of Artificial Intelligence 19 (2006) 915–925
%H2O fuel
% O2 stack % N2 % CO2 stack % O2 boiler
% H2O
% H2O stack
% CO2
Fuel mass flow Gases mass flow Air mass flow Furnace heat Final steam output SH 2 heat Final steam temperature SH 1 heat Steam temp. after atemp. Evapor. heat
Gases temp. Furnace exit
Econom. heat
Gases temp after econom
Furnace fouling index
Gases temp boiler exit Feed water temperature
SH 2 fouling index
Steam temp. before atemp.
SH 1 fouling index
Water temp before aire PH Primary aire Mass flow
Evapor fouling index
Air temp before aire PH.
Econom fouling index
Air temp after aire PH
Fig. 4. Relation between variables of the three monitoring NN sets.
subjectivity should be avoided, the selection of the inputs was made using only tools concerned with NN. In the first training of each NN, all the available inputs (23 variables)
were used and the mean square error (MSE) obtained was registered. After that, a variable was eliminated in every new training executed and the MSE obtained was also
ARTICLE IN PRESS 920
L.M. Romeo, R. Gareta / Engineering Applications of Artificial Intelligence 19 (2006) 915–925
Table 1 Training and validation error values of the fouling index monitoring NN Variable
Neurons number
Resubstitution average error (%)
Resistance average error (%)
Furnace fouling index SH 2 fouling index SH 1 fouling index Convective evaporator fouling index Economizer fouling index
3 4 3 4 6
1.03 1.60 1.08 2.35 0.86
1.01 1.54 1.09 2.36 0.88
registered and compared with the value obtained in the first training. The higher the influence of the absent input, the more the increase in the MSE value, and therefore the eliminated input variable was more relevant (Sung, 1998). This analysis of the input’s influence is an iterative task and time-consuming, but the highly logical input variables obtained compared with thermo-physical relations indicate a strong and coherent NN input variable selection. As an example, to calculate the exhaust gases composition, input variables related with exit gases composition in the stack and the moisture in the boiler have been selected, but those related with temperature and pressure signals have been rejected. Main results include new variables not collected in power plant data acquisition system as heat transferred from the combustion gases to the steam or water in each section, intermediate exhaust gas temperatures, steam mass flows in the heat exchangers and variables that include the fouling influence in heat transfer (fouling index). Every NN has been individually validated under two strategies in order to reject NN overfittings. Firstly, 3000 training output data were compared with the NN results (resubstitution test) concluding a good agreement with training data. Secondly, the comparison was made with 1495 data reserved from the data collection and nonused in the training stage. A resistance test was made in this way pointing out clearly any possible overfitting (Reich and Barai, 2000). The validation with the resubstitution–resistance methodology is achieved when the errors obtained with this analysis are similar, as shown in Table 1. Despite the low number of neurons, results clearly show a non-overfitting NN. Only a slightly more relevant deviation was pointed out in the convective evaporator fouling index, because of the lack of a representative data set. Since the output data obtained from NN were not included in the power plant data system variables, the validation cannot follow other traditional NN validation rules. In this case a final verification was made comparing the sensitivity analysis results of the calculated variables obtained from historical data. SH1 fouling index shows a deviation of 7.5% taking into account the uncertainty of the measure instruments and, indeed, the convective evaporator fouling index could reach till 9.8%. Results in Table 1 show deviation clearly below these limits.
5. Fouling and boiler thermal response forecasting The classical approach to estimate fouling evolution tried to adjust heat transfer with a time exponential decay equation (Kern and Seaton, 1959). In consequence, the great number of physical mechanisms involved in the deposit growing avoids a good agreement between real data and the calculation of heat transfer coefficient exponential decay constants (Allmon et al., 1991). It is proposed to analyse the fouling tendency by forecasting a fouling index (ratio between real heat transfer coefficient compared with clean conditions) hourly after the sootblowing. In opposition to the classical approach, a set of NN was developed in order to calculate two time series of fouling indexes (with and without sootblowing) for 12 h just after the moment of study, Fig. 3. The available data to develop the NN come from 2 years of boiler operation. Five hundred and seventy-three sootblowings were located during this period and 161 were selected as effective and with enough data to train an NN with a prediction of 12 h. Approximately 100 sootblowing series have been used to train, that means a 60% of the data, and the rest was reserved for the validation stage. Developed NN contains three layers of NN with three layers of neurons, every one following a feedforward architecture. A sigmoid activation function has been selected for the neuron hidden-layer and a linear one for the neuron output layer. The training stage has followed the same principles exposed in the monitoring module: an NN input selection as previously described, a training based in resubstitution test and a final validation following a resistance test. This methodology allows to point out curiosities as showed in Fig. 5, where the NN architecture design evidences the reduction of the influence of actual data, while the influence of the previous fouling forecasted values increases. The trend provided by the fouling forecasting NN is highly coherent with physical concepts and can be asserted as faithful in long-term analysis. Tables 2 and 3 display the number of neurons and deviations in reconstitution and resistance validations. Results are also excellent in superheaters (in Figs. 6 and 7 the results for superheaters 1 and 2 h after sootblowing are shown); also in convective evaporators results show a coherent evolution. The following step lies in the evaluation of the fouling influence in boiler performance. Two NNs have been
ARTICLE IN PRESS 921
L.M. Romeo, R. Gareta / Engineering Applications of Artificial Intelligence 19 (2006) 915–925 Time previous soot blowing Time Tiempo frfrom omdesde previous soplad sootb o an lowing teri or Hour 00FF.I./ Hour I. E. Hora .I./ 0/ Pre vious Hour Previous Hour Hora Previa
Sootbl. Number/Previous F.I. Or Sootbl. den de Numb soplado/F er/P revious .I. Previ F.I. o
+1
Fu flow stib le Fule flow Ca ud al de le Combu
Hour 1 F.I./ I. E. Hor a 1/ Hour Pr evious Hour Previous Hora Pr Hour evia
flow CaAir udAir al de flow aire
Hour 2 F.I./ Hour F.I./ I. E. Hor a 2/ Previous Hour Hour Hora Previa
Steam Caudal por Steadeflow mva flow
+3 3 3F.I./ Hour I.Hour E. Hor aF.I./ 3/ Previous Hour Previous Hour Hora Previa
PrHora evious F.I. Previous I. E. PF.I. revia
Fig. 5. Relation between variables of the fouling forecasting NN sets.
Table 2 Training and validation error values of the fouling index forecasting NN Variable
Neurons number
Resubstitution average error (%)
Resistance average error (%)
Superheater 1 Hour 0 fouling index Hour 1 fouling index Hour 2 fouling index Fouling index in the next hour
5 6 3 3
2.79 2.04 2.01 3.54
4.97 6.79 2.80 3.98
Convective evaporators Hour 0 fouling index Hour 1 fouling index Hour 2 fouling index Fouling index in the next hour
3 7 3 3
3.80 4.21 2.66 5.32
9.57 11.17 8.80 7.22
Table 3 Training and validation error values in the boiler performance evaluation after fouling forecasting Variable
Neurons number
Resubstitution average error (%)
Resistance average error (%)
Total heat Furnace heat Superheater 2 heat Superheater 1 heat Convective evaporator heat Economizer heat Steam flow
20 20 20 20 20 20 20
2.59 2.44 3.97 2.81 2.33 2.41 2.65
3.12 3.38 4.75 3.27 2.69 2.94 3.18
carried out in order to simulate boiler performance and compare the energy recovered with or without sootblowing, Fig. 3. Results are calculated as energy improvements for 4, 6 and 8 h after evaluation. This final NN uses as input data obtained from the fouling forecasting NN (heat exchanger fouling level values). Three thousand data have been used to train and 1550 data have been used to validate this model. This NN has the same structure as those explained in the monitoring and in the forecasting module. Table 3 displays the number of neurons used in this NN set and the average resubstitution and resistance errors. NN results for training and validation are shown in Fig. 8. A good agreement is observed between calculated and real boiler heat absorption, the key variable to analyse boiler
fouling influence. Although a slightly higher dispersion in the outputs than in previous results is noticed, it is explained by the reduced number of inputs required to this set of NN, compared for instance, to those needed for a thermodynamic model. 6. Sootblowing evaluation To complete the application and obtain a Hybrid System to for fouling control in on-line applications, a decision maker based on fuzzy-logic rules has been developed. The aim of the decision maker is to evaluate from the results provided by previous NN modules the activation of the sootblowing cycle. Therefore, the decision maker module
ARTICLE IN PRESS 922
L.M. Romeo, R. Gareta / Engineering Applications of Artificial Intelligence 19 (2006) 915–925
1.2
1.2 Training data (realvs. NN) SH2 Fouling level One hour after sootblowing
1.1 1
1.1 1
0.9
0.9
0.8
0.8
0.7
0.7
0.6
0.6
0.5
0.5
0.4 0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
1.2
0.4 0.4
Validation data (real vs. NN) SH2 Fouling level One hour after sootblowing
0.5
0.6
0.7
0.8
0.9
1
1.1
1.2
1
1.1
Fig. 6. Training and validation results of the SH2 fouling level in the sootblowing cycle initial hour.
1.1
1.1 Training data (real vs. NN) SH1 Fouling level Two hours after sootblowing
1 0.9
1 0.9
0.8
0.8
0.7
0.7
0.6
0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2 0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
0.2 0.2
Validation data (real vs. NN) SH1 Fouling level Two hours after sootblowing
0.3
0.4
0.5
0.6
0.7
0.8
0.9
70000
70000 65000 60000 55000 50000 45000 40000 35000 30000 30000 35000 40000 45000 50000 55000 60000 65000 70000 Boiler heat absorption (kW) - Real value
Boiler heat absorption (kW) - NN value
Boiler heat absorption (kW) - NN value
Fig. 7. Training and validation results of the SH1 fouling level 2 h after the sootblowing cycle.
65000 60000 55000 50000 45000 40000 35000 30000 30000 35000 40000 45000 50000 55000 60000 65000 70000 Boiler heat absorption (kW) - Real value
Fig. 8. Boiler heat absorption. Training (left) and validation (right).
should be developed attending to the experience implicit in the historical data available and the decision thresholds are extracted from a detailed study of previous results, Fig. 3. A FLES was selected because of the ability of arriving to conclusions dealing with non-precise data inputs (Frantti and Ma¨ho¨nen, 2001; Chen et al., 2003). The approach is
based on previous experience and do not require a physical law to calculate the system performance, becoming an adequate tool for the control of fouling in a biomass boiler (Fig. 9). The outputs of the Hybrid System decision maker were selected as: no activation of the sootblowers, activation of
ARTICLE IN PRESS L.M. Romeo, R. Gareta / Engineering Applications of Artificial Intelligence 19 (2006) 915–925
923
Energy recovery in the next hours when activati NOT sootblowing activation
Energy recovery in the next hours when sootblowing activation
EVALUATION
Energy improvement Energy improvement Energy improvement after 4 hours after 6 hours after 8 hours
¿SOOTBLOWING ACTIVATION?
YES
¿LONG CYCLE?
NO
NO
NEW EVALUATION THE NEXT HOUR
SHORT SOOTBLOWING CYCLE ACTIVATION
YES
LONG SOOTBLOWING CYCLE ACTIVATION
Fig. 9. Fuzzy-Logic Expert System design.
a short cycle or activation of a long cycle (depending on the fouling influence in the boiler). The inputs of the decision maker were exclusively the results obtained from the NN modules and the thresholds of activation were fixed in order to guarantee energy savings. With all these requirements the membership functions and pre- and postprocessing of the data were designed. The inputs selected from previous NN set are those related to the energy improvement caused by the sootblowing cycle, specially the prediction of the accumulated energy improvement due to the sootblowing after 4, 6 and 8 h. FLES module has been designed to emphasize the improvements in the first hours in order to recover the sootblowing energy as soon as possible, taking into account that the fouling evolution could not be as regular as predicted. The first calculation made in the FLES module assigns a numeric value to the prediction achieved by the NN modules. After that, this number is compared with the threshold values and a final decision is emitted. There are two threshold values: the first establishes the difference between not advising the activation of the sootblowers or advising a short cycle, the second has been calculated in order to advise a long cycle if the energy saved is enough. Forced by the availability of historical data, the adjustment and validation of the FLES module have been made with them. Results show that approximately 1020 sootblowing cycles would be activated in an operation period (6700 h considered), which represent approximately 200 more sootblowing cycles than traditional operating manoeuvres. From these 1020 sootblowing cycles, the FLES module recommends almost 800, but the rest will be automatically carried out by the system in a security limit established in 12 h. From the 800 sootblowing cycles
Table 4 Comparison of energy values in the application of NN–FL methodology to control boiler sootblowing cycles Energy values for the biomass boiler Average heat output produced in the boiler Energy recovered in historical sootblowing strategy Energy recovered in NN-FL sootblowing strategy Energy savings from historical to optimized strategy Percentage of energy savings Average heat output produced in the boiler
54.1 MW "7556 MWh +4953 MWh +12,509 MWh +3.45% 56.0 MW
recommended by the FLES module, 670 of them are long cycle type. Although it is known that the convective section is not a critical area, this also means a good level of cleanliness in it. Finally, it is really worth comparing the power production by the biomass boiler performing with historical sootblowing strategy and with the new strategy, Table 4. Previous strategy caused a decrease in energy saving of 7556.4 MWh due the sootblowing activation when the boiler was nearly clean or too fouled to be cleaned by steam, as Fig. 1 shows. The NN–FLES strategy shows an increase of energy saving up to 12509.0 MWh in an operational period. This means a power production augmentation of approximately 3.5%, from an average of heat produced as steam of 54.1 MW to a new average of 56.0 MW. 7. Conclusions There are some difficulties for extensive biomass utilization. In particular, biomass combustion produces fouling
ARTICLE IN PRESS 924
L.M. Romeo, R. Gareta / Engineering Applications of Artificial Intelligence 19 (2006) 915–925
in boiler heat transfer equipment, causing a reduction of steam output and boiler efficiency. The traditional approach uses experimental correlations but the problem has not been solved due to the inherent complexity of fouling and it is necessary to develop new techniques and strategies to minimize its consequences. The application of a Hybrid System combining NNs and a FLES has allowed to control boiler fouling and optimize its performance, although it has obligated to a new approach to the problem description. Different sets of NN have been used to monitor the boiler, to forecast fouling and the consequent future boiler performance. A set of fuzzy-logic rules based on real data that takes as inputs the results from previous NN has optimized boiler cleaning cycles and minimized fouling evolution. A procedure that compares the energy cost of a sootblowing cycle with the thermal improvements has been developed and it decides the sootblowing activation. NN and FLES system has been designed and trained obtaining encouraging results. Especially considering that any sootblowers control based on the state of the boiler has not been reported previously. A relative comparison between the old routine of sootblowing strategy and the new strategy programmed in the NN–FLES shows that approximately 3.5% of power production improvement can be obtained. The relevance of these results shows that techniques of AI as those proposed in this paper (NN–FLES) could help to reorient the study of complex phenomena, providing new point of views and tools to develop strong and easy-to-use control systems. Acknowledgements The work described in this paper was supported by the Fifth Framework Programme of the European Union, FP5–Energy, Environment and Sustainable Development, NNE5-2001-00128, Intelligent Process Control System for Biomass Fuelled industrial Power Plants. The authors are solely responsible for this article and it does not represent the opinion of the European Community. We gratefully acknowledge the inputs, collaboration and patience of the members of the INTCON project, TPS Terminska Processer AB, Tecnatom and CINAR. References Afgan, N., Carvalho, M.G., Coelho, P., 1996. Concept of expert system for boiler fouling assessment. Applied Thermal Engineering 16 (10), 835–844. Allmon, B.A., Watson, G.B., Carpenter, N.N., 1991. Fouling and enhancement interactions. ASME HTD-164, 61–70. Anthony, E.J., Iribarne, A.P., Iribarne, J.V., Talbot, R., Jia, L., Granatstein, D.L., 2001. Fouling in a 160 MWe FBC boiler firing coal and petroleum coke. Fuel 80 (7), 1009–1014. Bechtler, H., Browne, M.W., Bansal, P.K., Kecman, V., 2001. New approach to dynamic modelling of vapour-compression liquid chillers, artificial neural networks. Applied Thermal Engineering 21 (9), 941–953.
Bergeles, G., Bouris, D., Yianneskis, M., Balabani, S., Kravaritis, A., Itskos, S., 1997. Effects of fouling on the efficiency of heat exchangers in lignite utility boilers. Applied Thermal Engineering 17 (8–10), 739–749. BMS International Ltd., 2005. Intelligent Sootblower Control Systems for Efficient Power Generation. Currently available at http:// www.bmsint.com/power/index.html Chen, W.C., Chang, N.B., Chen, J.C., 2003. Rough set-based hybrid fuzzy-neural controller design for industrial wastewater treatment. Water Research 37 (1), 95–107. Dare, P., 2002. Combustion performance of biomass residue and purpose grown species. Fuel and Energy Abstracts 43 (4), 276. Electric Power Research Institute, 2000. Guidelines for intelligent sootblower control. Final Report, Report no. 1000410. Energy Research Center, Lehigh University, 2005. Optimize Sootblowing to Reduce NOx and Improve Heat Rate. Currently available at http:// www3.lehigh.edu/research/recenters.asp Fan, J.R., Zha, X.D., Sun, P., Cen, K.F., 2001. Simulation of ash deposit in a pulverized coal-fired boiler. Fuel 80 (5), 645–654. Frantti, T., Ma¨ho¨nen, P., 2001. Fuzzy logic-based forecasting model. Engineering Applications of Artificial Intelligence 14 (2), 189–201. Gareta, R., Romeo, L.M., Gil, A., 2006. Forecasting of electricity prices with neural networks. Energy Conversion and Management Journal. Currently available at http://www.sciencedirect.com Heinzel, T., Siegle, V., Spliethoff, H., Hein, K.R.G., 1998. Investigation of slagging in pulverized fuel co-combustion of biomass and coal at a pilot-scale test facility. Fuel Processing Technology 54 (1–3), 109–125. Ikonen, E., Najim, K., Kortela, U., 2000. Neuro-fuzzy modelling of power plant flue-gas emissions. Engineering Applications of Artificial Intelligence 13 (6), 705–717. Jenkins, B.M., Baxter, L.L., Miles Jr., T.R., Miles, T.R., 1998. Combustion properties of biomass. Fuel Processing Technology 54 (1–3), 17–46. Kalogirou, S.A., 2001. Artificial neural networks in renewable energy systems applications, a review. Renewable and Sustainable Energy Reviews 5, 373–401. Kalogirou, S.A., 2003. Artificial intelligence for the modelling and control of combustion processes, a review. Progress in Energy and Combustion Science 29, 515–566. Kern, D.Q., Seaton, R.E., 1959. A theoretical analysis of thermal surface fouling. British Chemical Engineering 4, 258–262. Lalot, S., Lecoeuche, S., 2003. Online fouling detection in electrical circulation heaters using neural networks. International Journal of Heat and Mass Transfer 46, 2445–2457. Lee, B.E., Fletcher, C.A.J., Shin, S.H., Kwon, S.B., 2002. Computational study of fouling deposit due to surface-coated particles in coal-fired power utility boilers. Fuel 81 (15), 2001–2008. Lertpalangsunti, N., Chan, C.W., Mason, R., Tontiwachwuthikul, P., 1999. A toolset for construction of hybrid intelligent forecasting systems, application for water demand prediction. Artificial Intelligence in Engineering 13 (1), 21–42. Lu, S., Hogg, B.W., 2000. Dynamic nonlinear modelling of power plant by physical principles and neural networks. International Journal of Electrical Power & Energy Systems 22 (1), 67–78. Mihalakakou, G., Santamouris, M., Tsangrassoulis, A., 2000. On the energy consumption in residential buildings. Energy and Buildings 34 (7), 727–736. Olofsson, T., Andersson, S., O¨stin, R., 1998. A method for predicting the annual building heating demand based on limited performance data. Energy and Buildings 28 (1), 101–108. Peltola, K., Hiltunen, M., Blomqvist, J.P., Skrifvars, B.J., Kurkela, J., Latva-Somppi, J., Kauppinen, E.I., 1999. Fouling of the surfaces in biofueled-fired fluidized bed boilers. In: Proceedings of the 15th International Conference on Fluidized Bed Combustion. ASME. Prieto, M., Vallina, J.M., Sua´rez, I., Martı´ n, I., 2001. Application of a design code for estimating fouling on-line in a power plant condenser
ARTICLE IN PRESS L.M. Romeo, R. Gareta / Engineering Applications of Artificial Intelligence 19 (2006) 915–925 cooled by seawater. Experimental Thermal and Fluid Science 25 (5), 329–336. Reich, Y., Barai, S.V., 2000. A methodology for building neural networks models from empirical engineering data. Engineering Applications of Artificial Intelligence 13 (6), 685–694. Sami, M., Annamalai, K., Wooldridge, M., 2000. Co-ring of coal and biomass fuel blends. Progress in Energy and Combustion Science 27, 171–214. Seggiani, M., 1999. Empirical correlations of the ash fusion temperatures and temperature of critical viscosity for coal and biomass ashes. Fuel 78 (9), 1121–1125. Skrifvars, B.J., Backman, R., Hupa, M., Sfiris, G., A˚byhammar, T., Lyngfelt, A., 1998. Ash behaviour in a CFB boiler during combustion of coal, peat or wood. Fuel 77 (1–2), 65–70. Sung, A.H., 1998. Ranking importance of input parameters of neural networks. Expert Systems with Applications 15 (3–4), 405–411. Wang, H., Harb, J.N., 1997. Modelling of ash deposition in large-scale combustion facilities burning pulverized coal. Progress in Energy and Combustion Science 23 (3), 267–282. Zevenhoven-Onderwater, M., Blomquist, J.P., Skrifvars, B.J., Backman, R., Hupa, M., 2000. The prediction of behaviour of ashes from five different solid fuels in fluidised bed combustion. Fuel 79 (11), 1353–1361.
925
Luis M. Romeo (Ph.D. Eng.) is Project Manager of the Centre of Research for Energy Resources and Consumption. He also teaches Heat Transfer and Thermodynamics in the Department of Mechanical Engineering, University of Zaragoza. He gained his professional experiences developing and managing several projects in different Spanish power plants such as the Pressurised Fluidised Bed Combustion Power Plant at Escatro´n, Pulverized Coal Teruel Power Plant, Combined Cycles and Biomass Boilers. He works on R&D projects and in activities related with efficiency improvements, power plant modelisation and CO2 capture. Ms. Raquel Gareta (Ph.D. Eng.) was a CIRCE researcher. She studied Mechanical Engineering at the University of Zaragoza with major speciality in Heat Transfer and Fluids Mechanics. She has worked on R&D projects since August 1999 in activities related with Gas Turbine performance optimisation, Neural Networks development, Fouling and Slagging influence minimization and Biomass Boilers improvement. Her Thesis ‘‘INTELLIGENT MODEL BASED SUPERVISION SYSTEM TO BIOMASS BOILER SOOTBLOWING CYCLES MANAGEMENT’’ has been one of the winners of the European Talent Award for Innovative Energy Systems 2005 (European Foundation for Power Engineering).
