Modeling and multi-objective optimization of a complex CHP process

Applied Energy 161 (2016) 309–319

Contents lists available at ScienceDirect

Applied Energy journal homepage: www.elsevier.com/locate/apenergy

Modeling and multi-objective optimization of a complex CHP process Sandra Seijo a,⇑, Inés del Campo a, Javier Echanobe a, Javier García-Sedano b a b

Department of Electricity and Electronics of the University of the Basque Country UPV/EHU Leioa, Vizcaya, Spain OPTIMITIVE S.L. Vitoria-Gasteiz, Álava, Spain

h i g h l i g h t s
Dynamic optimization of a CHP process using a multi-objective function is proposed.
ANNs and ANFISs are used to generate predictive models of the process.
Real-life data from a CHP process and a slurry drying treatment are available.
To optimize the plant 12 input parameters are selected as decision variables.

a r t i c l e

i n f o

Article history: Received 29 May 2015 Received in revised form 24 September 2015 Accepted 1 October 2015

Keywords: Artificial Neural Networks Adaptive Neuro-Fuzzy Inference System CHP Process modeling Multi-objective optimization

a b s t r a c t In this paper, the optimization of a real Combined Heat and Power (CHP) plant and a slurry drying process is proposed. Artificial Neural Networks (ANNs) and Adaptive Neuro-Fuzzy Inference Systems (ANFISs) are used to generate predictive models of the process. A dataset collected over a one-year period, with variables for the whole plant, is used to generate the predictive models. First, data mining techniques are used to obtain a representative dataset for the process as well as the input and target parameters for each model. Subsequently, models are used to optimize the plant performance in order to maximize the effective electrical efficiency of the process. For this purpose, 12 input parameters are selected as decision variables, i.e., variables which can change their values to optimize the plant. Plant performance optimization is a multi-objective problem with three goals: to maximize electrical production, minimize fuel consumption and maximize the amount of heat used in the slurry process. The optimization algorithm calculates the values of the decision variables for each time-step using Gradient Descent Methods (GDM). The simulation results show that optimization using a multi-objective function increases the CHP plant’s effective electrical efficiency by around 3% on average. Ó 2015 Elsevier Ltd. All rights reserved.

1. Introduction According to the European Association for the Promotion of Cogeneration [1], ‘‘Cogeneration (Combined Heat and Power or CHP), is the simultaneous production of electricity and heat, both of which are used”. On average, conventional power plants, operate at an efficiency of just 35%. This means that up to 65% of the potential energy is released as waste heat. More recently, CHP plants have been designed which can increase this efficiency up to 80% [2]. The objective of cogeneration is to use as much of the potential energy contained within a fuel as possible. In CHP, the flue gases produced by engines, gas turbines, or other machines are used to generate more electricity or for other process. This reduces costs, because the total amount of fuel needed to produce both electricity ⇑ Corresponding author. Tel.: +34 646615569. E-mail address: [email protected] (S. Seijo). http://dx.doi.org/10.1016/j.apenergy.2015.10.003 0306-2619/Ó 2015 Elsevier Ltd. All rights reserved.

and heat in a cogeneration plant is less than the total fuel needed to produce the same amount of electricity and heat in separate technologies (e.g., an electric utility generating plant and an industrial boiler). These fuel savings also result in less pollution. In light of these economic and environmental factors cogeneration plants currently play an important role in some EU countries [1]. Optimizing the behavior of CHP processes is a hard task for many reasons, including complexity, uncertainties, high dimensionality, non-linearity and time delays. Mathematical models with a large number of assumptions are necessary [3–5] to simulate these processes, they are often either practically impossible to model or takes too much computational time and resources. Since such physical models usually use iterative methods for final solutions they take a long time to yield results, thus they are impractical for ’real-time’ applications. Computational intelligence (CI) techniques are a set of methodologies, inspired by nature, that can deal with problems which are

310

S. Seijo et al. / Applied Energy 161 (2016) 309–319

Nomenclature ANN ANFIS CHP DivA DivB DivC DivD FCond FEv FFlueGas FGas_A FGas_B FGas_C FGas_D FSteam FIS GDM HAmb LHV PCond PEv POWA POWB POWC POWD POWST PSt_Gen QFuel QTH ST TAmb TB1_A

Artificial Neural Network Adaptive Neuro-Fuzzy Inference System Combined Heat and Power diverter engine A (%) diverter engine B (%) diverter engine C (%) diverter engine D (%) condensate effluent flow (kg/h) evaporator feed flow (kg/h) flue gases flow (kg/h) natural gas flow engine A (m3/h) natural gas flow engine B (m3/h) natural gas flow engine C (m3/h) natural gas flow engine D (m3/h) steam flow to steam turbine (kg/h) Fuzzy Inference System Gradient Descent Method ambient humidity (%) Low Heating Value (kWh/m3) condenser pressure (bar) evaporator pressure (bar) rated power engine A (%) rated power engine B (%) rated power engine C (%) rated power engine D (%) power of the ST (kW) steam generator pressure (bar) natural gas consumed by the engines (kW) heat used in the slurry process (kW) steam turbine ambient temperature (°C) air intake temperature engine A Bank1 (°C)

very hard to solve with standard methods. The most commonly used CI techniques are: Genetic Algorithms (GAs), Artificial Neural Networks (ANNs), Fuzzy Inference Systems (FISs) and hybrid approaches such as Neuro-Fuzzy systems (NFs). GAs are based on a heuristic search that mimics the process of natural selection. ANNs, are highly interconnected structures, inspired by the human brain, capable of learning from input–output samples. They are widely used in industry because they are data driven, adaptive, respond quickly and are highly accurate if trained with the correct data. FISs, based on Fuzzy Set theory, provide a framework to represent imprecise information and to reason using this kind of information. Neuro-Fuzzy Systems comprise a synergistic combination of ANNs and FISs. ANNs enhance the performance of FISs with their capacity for learning from input–output samples; this learning is used to adapt FIS parameters such as a membership function or rules. CI methods are widely used for modeling different industrial processes, for example in water treatment [6], steel production [7] and paper industries [8], or industrial boilers [9,10], amongst other uses. They are also very useful to detect and predict faults in industrial processes [11–13] and engines [14–18]. With respect to cogeneration plants, Neural/Fuzzy techniques are used in many applications, for example, the design of adaptive load-shedding models [19–21] or temperature controllers [22,23]. Fuzzy systems are used in [24] to identify failures in a boiler during the cogeneration process and GAs are used to alleviate these failures. ANNs are used to predict the baseline energy consumption in CHP plants because they are particularly robust to uncertainties that affect the measured values of input parameters [25]. Furthermore, ANNs have proven to be a useful tool when using simple

TB2_A air intake temperature engine A Bank2 (°C) TB1_B air intake temperature engine B Bank1 (°C) air intake temperature engine B Bank2 (°C) TB2_B TB1_C air intake temperature engine C Bank1 (°C) TB2_C air intake temperature engine C Bank2 (°C) TB1_D air intake temperature engine D Bank1 (°C) air intake temperature engine D Bank2 (°C) TB2_D TBank1_A gases temperature engine A bank 1 (°C) TBank2_A gases temperature engine A bank 2 (°C) TBank1_B gases temperature engine B bank 1 (°C) TBank2_B gases temperature engine B bank 2 (°C) TBank1_C gases temperature engine C bank 1 (°C) TBank2_C gases temperature engine C bank 2 (°C) TBank1_D gases temperature engine D bank 1 (°C) TBank2_D gases temperature engine D bank 2 (°C) TH2O_Ex exchange water temperature (°C) TH2O_SH superheated water temperature (°C) TH2O_TH water temperature tubular heater (°C) TH2O_Tow water temperature cooling tower (°C) TMixt_EngA water temperature to cooling the mixture engine (°C) TMixt_EngB water temperature to cooling the mixture engine (°C) TMixt_EngC water temperature to cooling the mixture engine (°C) TMixt_EngD water temperature to cooling the mixture engine (°C) TST_Cond steam temperature input condenser (°C) power generated by the engines and the ST (kW) WE eEE effective electrical efficiency (%)

A B C D

models with relatively few variables to predict the power generated in a cogeneration plant [26,27]. In [28] the power output was predicted based on ANN by studying the relationships between the electricity produced in a CHP plant and the properties of the fuel. Another study [29] predicted power by dividing the process into two submodules, each with its own ANN model. Although NF techniques are used to model cogeneration plant far less frequently than ANNs, some studies have employed them to obtain a predictive model for two pieces of equipment in a CHP plant with accurate results [30]. NF algorithms have been used by [31] to model some parameters of a heat recovery steam generator in a cogeneration and cooling plant. All the studies mentioned above model cogeneration systems with accurate prediction results, although only relatively few variables are employed. Some researchers not only address CHP modeling but also deal with the optimization of cogeneration processes. The process was modeled in [32] using ANNs, then the plant was statically optimized using dynamic neural networks (one optimum value per variable) in order to minimize the fuel mass flow. After ANN and thermodynamic modeling, [33] developed a static multi-objective optimization using GAs. In [34] a mathematical model of the process was created using thermodynamic principles, then multiobjective optimization is performed determining static optimum values for the decision variables using GAs. After thermodynamically modeling a CHP plant, [35] used GAs to produce a static multi-objective optimization in order to obtain the optimal operation planning for a CHP process under different scenarios while [36] used linear programming (i.e., a mathematical method of solving optimization problems by means of linear functions in which the

S. Seijo et al. / Applied Energy 161 (2016) 309–319

Fig. 1. Diagram of the system architecture for data collection from the plant.

variables involved are subject to linear constraints). In [37] a combined cooling, heating and power system is statically optimized using two objective functions under three different scenarios using GAs. Multi-objective statical optimization is performed in [38,39] using Particle Swarm Optimization (PSO). PSO is a populationbased optimization technique inspired by the social behavior of birds flocking or fish schooling. The main limitation of static optimization is that static values are not the best option due to possible changes in working and atmospheric conditions. Therefore, the best option for the optimization of a real cogeneration process is to obtain the optimum values for the decision variables in each time-step of the dataset (i.e., a dynamic optimization). In [40] the dynamic single-objective minimization through a network flow model by using linear programming was achieved. Several optimization methods, including GAs and PSO, are used in [41] to dynamically minimize the operating cost of the CHP system. A dynamic single-objective optimization through a mathematical model was proposed in [42] with the use of an exhaustive search method (i.e.,a mathematical method that consists in the of enumeration of all possible candidates for the solution and checking whether each candidate satisfies the problem’s statement). In [43] a PSO and fuzzy decision-making system was proposed to optimize a benchmark cogeneration system. The work carried out in [44] presented a thermodynamic model of the process and dynamically optimized the set points related to the economically efficient management of a combined heat, power and cooling plant using dynamic programming (a method for solving complex problems

311

by breaking them down into a collection of simpler subproblems). This allows to minimize the total daily cost based on the demand for the process. However, it did not improve the efficiency of the process or any of its components. In summary, the studies cited above provide a partial solution to the optimization/modeling problem: they either deal with only single-objective dynamic optimization, or just perform a static multi-objective optimization, or conduct a multi-objective dynamic optimization but only taking into account economically-objective functions (not the efficiency), or perform the modeling process for just some components but not for the complete plant, or obtain models from simulated data (not real-life data). On the other hand, only [32,33] used CI techniques, specifically ANNs, to model a CHP process. No previous multi-objective dynamic optimization work based on NF modeling using real-life data from an entire CHP plant has been reported. The present work performs the dynamic optimization of a CHP process using a global multi-objective function, i.e., the effective electrical efficiency (eEE ), by determining the optimal values for 12 decision variables for each time-step in a one-week dataset. Real-life data from a CHP process and a slurry drying treatment was available from two PLCs (Programmable Logic Controllers) connected to an OPC server (Open Platform Communication). This server was connected to an OPTIBATÒ tool with a client server obtaining data once per minute and storing them in a data-base (see Fig. 1). The dynamic optimization procedure was based on GDMs (Gradient Descent Methods), which searches for optimum value for a minimization problem with an objective function from a given point [45]. The experimental results revealed an average increase in efficiency of 3.05%. The CHP plant generates electricity with four internal combustion engines. The flue gases from the engines are used to generate steam that feeds a steam turbine and generates more electricity. The heat from the engines is also used in a slurry drying process. The multi-objective function considers the whole cogeneration and slurry process while taking into account three goals: to maximize the electrical energy generated by the engines and steam turbine; to minimize the use of primary energy i.e., the total fuel used to feed the engines; and to maximize the use of thermal energy from the engine’s flue gases in the slurry drying process. In Fig. 2, the different energy systems withy their

Fig. 2. Scheme for the energy systems and their relation with the objective functions.

312

S. Seijo et al. / Applied Energy 161 (2016) 309–319

key parameters and their relations with the three objective functions are shown. Two CI techniques were employed to model the cogeneration and the slurry treatment process, an ANN and an adaptive FIS, known as an Adaptive Neuro-Fuzzy Inference System (ANFIS). The two algorithms are compared and their results are discussed. The rest of this paper is organized as follows: Section 2 gives a brief description of the CHP process, and also introduces the slurry treatment and its components; Section 3 introduces the ANNs, FIS and ANFIS. Section 4 develops the data conditioning and modeling for both the CHP plant and the slurry process; Section 5 analyses the optimization of the process and its results; and finally, Section 6 presents the conclusions of the work. 2. The CHP plant Fig. 3 depicts a scheme of the CHP plant optimized in this paper. The plant is located in Monzón (Huesca), in the North of Spain [46]. The plant’s main systems are: four engines, four engine cooling circuits, an exhaust steam boiler, a steam turbine condenser, a steam turbine, and a slurry drying process. The plant generates electricity via the four internal combustion engines and the steam turbine. Heat from the engines is used to generate steam which is then used in the turbine and the slurry drying process. The slurry drying process uses slurry from nearby farms.

The four engines are identical, each with the same characteristics and the nominal power of 3700 kW. The engines have two banks of eight cylinders (Fig. 4) and operate with natural gas. The engines exchange heat with two circuits which take water from the cooling towers, as shown in Fig. 3. One cooling circuit maintains the temperature of the air–fuel mixture around 50 °C while the other circuit preheats the air intake around 35 °C. The engines generate electricity and flue gases. The electricity generated is sold. Each engine has a diverter which sends the flue gases to an exhaust steam boiler, whenever the given engine is operating at greater than 50% of its rated power, or to the chimney, if operating at less than 50% of rated power. Engines usually run at levels above this threshold and so the flue gases are generally sent to the exhaust steam boiler. Moreover, the steam generator uses heat from the exhaust steam boiler to create steam at around 400 °C and 22.5 bar. This steam is fed into a steam turbine where it generates more electricity, with 1000 kW of nominal power. The steam turbine condenser uses water from the cooling towers to force the steam from the steam turbine to condense before recirculating it to the system. As with the engines, the power generated from the steam turbine is sold. The slurry from surrounding farms consists of approximately 6% solids. It is first subjected to a mechanical treatment using rotatory equipment to remove the solid fraction from the liquid portion (Fig. 5). The liquid portion is then treated chemically to reduce

Fig. 3. Scheme for a combined heat and power plant with its related equipment. The red circuit represents steam, the purple circuit is the superheated water circuit, and the green and blue circuits are water circuits.

S. Seijo et al. / Applied Energy 161 (2016) 309–319

313

towers. Finally, the sterilizer uses superheated water to purify the condensed effluent and produces water suitable for irrigation.

3. Computational intelligence algorithms 3.1. Artificial Neural Networks

Fig. 4. Photograph of one of the four engines at the CHP plant.

the chemical load. The product of the chemical treatment is then treated thermally to separate condensable from non-condensable matter in an evaporator which uses superheated water generated in the exhaust steam boiler (water at a temperature of around 120 °C). A tubular heater recirculates the effluent to the evaporator and preheats it. The tubular heater takes water from the cooling circuit that preheats the engine’s air intake. The non-condensable portion is combined with the solid fraction produced by the mechanical treatment, this is then sold as fertilizer. The condensable effluent is condensed again using water from the cooling

An Artificial Neural Network can be defined as a data processing system consisting of simple processing elements (artificial neurons) highly interconnected in an architecture inspired by the structure of the brain’s cerebral cortex. These processing elements are organized into layers (i.e., the input layer, one or more hidden layers and an output layer) [47]. One of the most common ANNs is the Multiple Layer Perceptron (MLP). Previous works have used MLP to successfully model CHP plants [22,23,25,26,20]. In the MLP, a neuron in a particular layer is connected to all the neurons in the next layer, as shown in Fig. 6. A neuron is a single computational element in the network that receives one or more inputs (xi ) either from a neuron in a previous layer or (when it is situated in the input later) from the outside world (input parameters measured from the process). These are multiplied by weights (xij ) and then added together. They are subsequently passed through an activation function (in this study the sigmoid function, g, is used), making the activation of the neuron continuously valued. The neuron output oj is:

oj ¼ g

n X

xij xi

ð1Þ

i¼1

The scheme for an artificial neuron is shown in Fig. 7. When the neuron fires, it means that its output has a non-zero value which is transmitted to neurons in the next layer. Outputs from neurons situated in the output layer collectively form the output for the whole

Fig. 5. Scheme of the slurry drying process.

314

S. Seijo et al. / Applied Energy 161 (2016) 309–319

labels, each associated with a membership function lij ðxi Þ. In zeroorder Sugeno inference systems y ¼ p0j , i.e. p1j ; . . . ; pnj ¼ 0, and in first-order Sugeno inference systems, some pij have non-zero values. The inference procedure used to derive the conclusion for a specific input xi consists of two main steps. First, the firing strength or weight wj of each rule is calculated as

wj ¼

n Y

lij ðxi Þ

ð3Þ

i¼1

After that, the overall inference result, y, is obtained by means of the weighted average of the consequent,

Pm

j¼1 wij yj

y ¼ Pm

j¼1 wj

Fig. 6. Scheme of a MLP with n inputs, m outputs and L hidden nodes in a single hidden layer.

network. Analysis of the structure of a neural network, reveals that the information is stored in the weights and biases (a bias input always has the value 1 and is also weighted). Artificial Neural Networks can be trained using a collection of sample data based on different learning algorithms, such as the gradient descent methods (GDMs), used in this study. During the training process the weights are updated in order to minimize the square sum of the difference between real output (desired output) and the network output. Eq. (2) shows how a weight is updated in every iteration of the GDM algorithm, where P E ¼ 12 ki¼1 ðyi  y0i Þ2 represents the error, yi and y0i are respectively the actual and the desired i-th outputs and g denotes a learning rate.

ð4Þ

Eqs. (3) and (4) provide a compact representation of the inference model. Fig. 8 represents the computational scheme for Sugeno-type FIS. In the first layer, the inputs are fuzzified using membership functions, in this case the Gaussian function. In the second layer, the firing strength is calculated using Eq. (3). In layer 3, the firing strengths of the fuzzy rules are normalized. The next layer produces the output for each rule. Finally, in layer 5, the overall output is computed using Eq. (4). 3.3. Neuro-fuzzy inference systems

where Rj is the jth rule (1 6 j 6 m), xi ð1 6 i 6 nÞ are input variables, y is the output, pij are consequent constants, and Aij ðxi Þ are linguistic

The aim of NF systems is to combine the advantages of both the above mentioned approaches, ANNs and FISs. Knowledge of the system is expressed as a linguistic fuzzy relationship while the neural network learning schemes, which are capable of learning non-linear mappings of numerical data, are used to train the system. Furthermore, a NF system is capable of extracting fuzzy knowledge from numerical data. In this particular work, the well-known Adaptive Neuro-Fuzzy Inference System (ANFIS) algorithm has been used. Previous investigations by our research group have shown that NF systems are very useful when it comes to obtaining accurate models for some pieces of equipment at the CHP plant [49]. A fuzzy system has a network-type structure similar to that of a neural network, as shown in Fig. 8. In a fuzzy system, the parameters associated with the membership functions are fixed values. However, in the ANFIS algorithm, those parameters change throughout the learning process. ANFIS is trained with a hybrid learning algorithm comprising a GDM process to find the optimal values for the parameters of the antecedent membership functions. A least squares estimation (LSE) process is employed for the linear consequent parameters of the fuzzy rules. They are calculated so as to minimize the error between the calculated output and the measured output. An

Fig. 7. Performance of a neuron.

Fig. 8. Scheme of the network-type structure of a fuzzy inference system similar to that of a neural network with two inputs and one output.

wkij ¼ wk1 g ij

dE dwij

ð2Þ

3.2. Fuzzy inference systems FIS are based on the Fuzzy Set theory proposed by Zadeh in the sixties [48]. They are systems which can deal with imprecision, vagueness or incomplete information. They are composed of a rule base and an inference mechanism, where the rules are of the IF-THEN type:

Rj : IF x1 is A1j ðx1 Þ and x2 is A2j ðx2 Þ and xn is Anj ðxn Þ . . . THEN y ¼ p1j  x þ    þ pnj xn þ p0j

315

S. Seijo et al. / Applied Energy 161 (2016) 309–319 Table 1 CHP plant systems and their corresponding variables. CHP systems Cooling engine Cooling engine Cooling engine Cooling engine Engine A

A B C D

Engine B Engine C Engine D Exhaust steam boiler Steam turbine condenser Steam turbine Slurry process

Inputs of the model

Output

TH2O_Ex TH2O_TOW POWA TH2O_ExTH2O_TOW POWB TH2O_ExTH2O_TOW POWC TH2O_ExTH2O_TOW POWD TB1_A TB2_A TAmb HAmb LHV TBank1_A TBank2_A TMixt_EngA POWA DIVA TB1_B TB2_B TAmb HAmb LHV TBank1_B TBank2_B TMixt_EngB POWB DIVB TB1_C TB2_C TAmb HAmb LHV TBank1_C TBank2_C TMixt_EngC POWC DIVC TB1_D TB2_D TAmb HAmb LHV TBank1_D TBank2_D TMixt_EngD POWD DIVD PStGen FFlueGas TH2O_Tow TST_Cond

TMixt_EngA TMixt_EngB TMixt_EngC TMixt_EngD FGas_A

PStGen FSteamPCond PEv TH20_SH TH2O_TH TH2O_Ex FCond

FGas_B FGas_C FGas_D FSteam PCond POWST FEv

enhanced ANFIS algorithm with structure learning capability was used. The rule base is an incremental base, which means the number of rules increases gradually as long as the model improves its behavior. 4. Modeling the CHP plant In order to obtain a model for the whole CHP plant, a database with 213 variables from the entire cogeneration process collected over a one-year period was available to model the process. The values of the variables were read once per minute throughout the data collection period. 4.1. Data cleaning and variables selection Data mining techniques were applied using Rapid MinerÒ software to clean the database and to select variables. Data mining is the process of discovering and extracting useful and interesting information from databases. First, due to the large amount of samples and variables, the database was resampled every 30 min. Each variable was represented to analyze its behavior. Anomalous data for each variable were removed using filters. Non-informative variables (e.g. constant variables, zero variables or redundant variables) were removed from the database because they do not provide any useful information. The CHP plant was then separated into the different systems listed in Table 1. For each system, the target variable (output)

was selected according to knowledge of the process and the variables available in the database. The input variables for each systems were then obtained by taking the variables with the highest influence from each output system. To this end we employed a combination of mathematical techniques (correlation, mutual information, attribute selection methods, and covariance, amongst others) and in-depth knowledge of the process and systems. The different systems and variables involved in generating the models are shown in Table 1. The original dataset was separated into two datasets: a training set with data from February, April, June, August, October and December and a testing set for evaluating the predictability of the models with the remaining months. 4.2. Modeling the CHP systems An ANN model and an ANFIS model were prepared for each system. When selecting the structure of the ANN models, after some initial testing it was concluded that the accuracy did not show any significant improvement after increasing the number of hidden layers. Once the number of hidden nodes is sufficient the exact number is no longer relevant in terms of accuracy. Therefore, the structure chosen for the ANN models was always the same: the number of neurons in the input layer was the same as the number of inputs entering the model, the number of neurons in the hidden layer was twice the number of inputs (only one hidden layer), while a single output neuron was used. The stop criteria is the number of epochs. The ANFIS models were structured such that the number of rules increased until the model’s error was stabilized. P The Mean Absolute Error MAE ¼ N1 ki¼1 jyi  y0i j was calculated for each model. This testing error represents the model’s behavior better than the training error as it contains unseen data from the model and the model’s predictive behavior can be evaluated. For all models, the difference between the training error and the testing error was always less than 0.3%. This means that the models were capable of learning and could make accurate predictions when dealing with unseen data. The modeling and optimizations for all the experiments were carried out with the Optibat Trainer tool [50]. As can be seen in Table 2, both algorithms were, in general,very accurate. However, when there is a high number of inputs, as is the case for engines, the ANNs perform better than the ANFISs. Although ANFISs increase the number of rules generated, the engine systems are better modeled using ANNs. The testing errors for both models, ANNs and ANFISs, were compared and the model with the lowest error was selected. The

Table 2 Model features and results. System

Cooling engine A Cooling engine B Cooling engine C Cooling engine D Engine A Engine B Engine C Engine D Recovery boiler ST condenser Steam turbine Slurry process

ANN model

ANFIS model

Structurea

Epoch

Train error (%)

Test error (%)

N0 rules

Train error (%)

Test error (%)

3/6/1 3/6/1 3/6/1 3/6/1 10/20/1 10/20/1 10/20/1 10/20/1 2/4/1 2/4/1 3/6/1 5/10/1

6000 6000 6000 6000 20000 20000 20000 20000 4000 4000 6000 10000

0.21 0.28 0.10 0.49 0.39 0.41 0.38 0.38 0.61 1.01 0.67 2.35

0.23 0.26 0.13 0.33 0.42 0.41 0.42 0.37 0.63 0.96 0.70 2.52

16 16 16 16 22 22 22 22 12 5 16 21

0.21 0.23 0.11 0.48 0.51 0.5 0.49 0.51 0.61 2.27 0.75 2.88

0.22 0.23 0.12 0.30 0.54 0.51 0.49 0.52 0.62 1.97 0.90 2.84

The testing error is highlighted in italic and the testing error with the lowest error in each system is highlighted in bold. a ANN structure: input layer neurons/hidden layer neurons/output layer neurons.

316

S. Seijo et al. / Applied Energy 161 (2016) 309–319

testing error highlighted in bold in Table 2 corresponds to the selected models. 5. Optimization The aim of cogeneration optimization is to improve plant efficiency. The ANN and ANFIS models selected in the previous section were used to simulate the process, calculate plant efficiency, and then optimize it. In the present case the optimization problem is actually a multiobjective optimization problem in which three functions need to be optimized:

optimized values for the decision variables, a subset with data corresponding to one week in February was selected from year-long dataset and was then resampled at a sample time of ten minutes, i.e., the values of the variables were read every ten minutes. Values for the optimum decision variables were calculated for each sample in the week-long dataset, as shown Fig. 9. As all the air intake temperatures were very similar, only one of them is shown here. Also, the steam pressure in the steam generator is not included because the optimized and real values were in agreement for every dataset. Fig. 9a shows the temperature of the superheated water used in the slurry drying process and it can be observed that, in most cases, the optimized values were slightly higher than the real values


The net useful power output (needs to be maximized):

W E ¼ ðPOW A þ POW B þ POW C þ POW D Þ c1 þ POW ST

ð5Þ


The total natural gas consumed by the four engines (needs to be minimized):

Q FUEL ¼ ðF GAS A þ F GAS B þ F GAS C þ F GAS D Þ c2

ð6Þ


The useful thermal energy used in the slurry process (needs to be maximized):

Q TH ¼ F Ev c3

ð7Þ

where c1; c2; c3 are the scale factors after the dimensional analysis study. However, since the efficiency of CHP plants is measured in terms of the effective electrical efficiency (eEE ), the three objectives above can be formulated as a single optimization problem:

eEE ¼

WE Q FUEL  QaTH

100

ð8Þ

where a is the efficiency of conventional technology that otherwise would be used to produce the useful thermal energy output if the CHP system did not exist. The value of a, as established in Annex II of the European Commission’s Directive dated 21 December 2006 (2007/74/EC) and covered to in Article 1 of Spanish Royal Decree RD 661/2007, depends on the type of fuel and the manner in which the heat is used (direct/indirect use). For a gaseous fuel and indirect use of exhaust gases, a is 0.9. To achieve optimization following a careful study of the process, 12 variables were selected as decision variables (i.e., input parameters whose values can be adjusted to improve the value of eEE -highlighted by a gray circular background in Fig. 3). To provide a reasonable and realistic optimization, the decision variables must satisfy the following constraints:
TA/B/C/D_1and TA/B/C/D_2 (8 intake air temperatures, 2 for each engine): 30 6 T  38  C.
TH2O_EX(exchange water temperature): 61 6 T H2OEx  65  C.
PSt_Gen(pressure of the steam generator): 20 6 P StGen  22 bar.
PEv(evaporator pressure): 0:13 6 P Ev  0:17 bar.
TH2O_SH (water superheated temperature):110 6 T  125  C. TA/B/C/D_1and TA/B/C/D_2 constrains are system constraints. The remaining decision variables were selected from the recorded values and the knowledge of the systems. Besides the constraints related to the decision variables, the four engines have to operate at least 94% of rated power as an economic constraint dictated by the plant owner. The optimization algorithm calculates the values for the decision variables in order to obtain the maximum eEE for each sample of the dataset being optimized. This is realized through an exhaustive search based on the gradient descendent, which starts with an initial set of values for the variables and iteratively moves toward a set of values that optimizes the cost function. To obtain the

Fig. 9. Real values (dotted line) versus optimized values (continuous line) for the selected decision variables.

S. Seijo et al. / Applied Energy 161 (2016) 309–319

Fig. 10. Output for the slurry process model with and without optimization (FEv).

317

measured in the subset. Fig. 9b shows the real evaporator pressure and how optimization decreased the pressure thus helping to evaporate more slurry. Similarly, the optimization decreased the temperature of the exchange water in comparison with the real temperature (Fig. 9c). Moreover, with regard to the air intake temperatures for bank 1 in engine D, the optimized values were generally equal to the real values, except when the optimized values were 30 °C. As noted above, a system constraint was the air intake temperature had to be between 30 and 38 °C. Therefore, when the real value was lower than 30 °C, the optimized values were, at least, at the lower limit (Fig. 9d). Optimization causes some of the decision variables to change their values and, therefore some of the model’s outputs also change. For example, the system whose output experienced the biggest change was the slurry process. This is shown in Fig. 10 which predicts that a significantly greater amount of slurry would be treated after the optimization process. This is consistent with the optimized values for the superheated water temperature, the pressure in the evaporator and the exchange water temperature, because they all increase the volume of slurry treated. Finally, Fig. 11 represents each term in the eEE , with and without optimizing the decision variables. The optimization procedure achieves an average increase in eEE of 3.05%. This improvement is mainly derived from the increase in the useful thermal energy used in the slurry drying process which means the CHP plant can treat a greater quantity of slurry. This increase in eEE represents important savings when dealing with large, high-consumption industrial processes such as the CHP plant described in this work.

6. Conclusion

Fig. 11. Terms of the multi-objective function and multi-objective function with and without recommendations.

In this paper, ANN and ANFIS algorithms are proposed for the simultaneous modeling and optimization of a complex cogeneration process in conjunction with a slurry drying process. The plant was composed of several systems which were modeled using both ANNs or ANFISs. The results verified that both ANFIS and ANN methods are powerful tools when it comes to modeling a cogeneration plant. In particular, the results for all models revealed that the difference between the training error and the testing error was always less than 0.3%. This means that the models could learn and also make accurate predictions when dealing with unseen data. The best model for each system was selected by comparing the testing error. Optimization of a real CHP plant and slurry drying process was successfully achieved using the Neural/Fuzzy models developed for the process. A remarkable feature optimizing a cogeneration process was the selection of a multi-objective function to represent the effective electrical efficiency (eEE ). The efficiency was improved by changing some decision variables in each time-step, adapting to potentially changing conditions. The results of the optimization process using a multi-objective function (eEE ) show that the main improvement derives from the increased volume of slurry treated. This improvement is consistent with the expected results because, considering the engine load restrictions, the amount of electricity generated is basically the same, as is the natural gas used to feed the engine. The dynamic optimization has the advantage of adaptation to changes in atmospheric conditions and working conditions obtaining the maximum energy efficiency in each timestep. This potential increase in the volume of slurry treated results in an average increase of 3.05% for the multi-objective function. This percentage is a good result when dealing with very large, high-cost industrial processes, such as the CHP plant considered in this work. Further work will consider real-time plant optimization for maximum energy efficiency. This will be done by installing a pro-

318

S. Seijo et al. / Applied Energy 161 (2016) 309–319

totype in the Optimitive softwareÒ and connecting via OPC to close the plant’s control loop. The main advantage of the proposed approach is that no new investment or changes in the existing plant are required.

[21]

[22]

Acknowledgment This work was supported in part by the Basque Country Government under Grants IT733-13, and IG2012/221 (ICOGME), and the Zabalduz Program of the University of the Basque Country (Spain). References [1] http://www.cogeneurope.eu/. [2] Vatopoulos K, Andrews D, Carlsson J, Papaioannou I, Zubi G. Study on the state of play of energy efficiency of heat and electricity production technologies. JRC Sci Policy Rep. doi:http://dx.doi.org/10.2790/57624. [3] Zare V, Mahmoudi SMS, Yari M, Amidpour M. Thermoeconomic analysis and optimization of an ammonia-water power/cooling cogeneration cycle. Energy 2012;47(1):271–83. http://dx.doi.org/10.1016/j.energy.2012.09.002. [4] Feidt M, Costea M. Energy and exergy analysis and optimization of combined heat and power systems: comparison of various systems. Energies 2012;5 (9):3701–22. http://dx.doi.org/10.3390/en5093701. [5] Sen D, Panua R, Sen P, Das D. Thermodynamic analysis and cogeneration of a cement plant in india-a case study. In: 2013 International conference on energy efficient technologies for sustainability (ICEETS 2013); 2013. p. 641–6. doi:http://dx.doi.org/10.1109/ICEETS.2013.6533459. [6] Noshadi I, Salahi A, Hemmati M, Rekabdar F, Mohammadi T. Experimental and anfis modeling for fouling analysis of oily wastewater treatment using ultrafiltration. Asia–Pacific J Chem Eng 2013;8(4):527–38. http://dx.doi.org/ 10.1002/apj.1691. [7] Isazadeh G, Hooshmand R, Khodabakhshian A. Design of an adaptative dynamic load shedding algorithm using neural network in the steelmaking cogeneration facility. Iranian J Sci Technol – Trans Electr Eng 2012;36 (E1):67–82. http://dx.doi.org/10.1109/28.585846. [8] Zhang Y, Ai J. Zeta potential modeling of papermaking wastewater on neural network. In: Proceedings of the 2012 second international conference on instrumentation & measurement, computer, communication and control (IMCCC 2012); 2012. p. 63–6. doi:http://dx.doi.org/10.1109/IMCCC.2012.21. [9] Budnik M, Stanek W, Rusinowski H. Application of neural modelling in hybrid control model of fluidized bed boiler fired with coal and biomass. In: Proceedings of the 13th international carpathian control conference, ICCC 2012; 2012. p. 69–74. doi:http://dx.doi.org/10.1109/CarpathianCC.2012. 6228618. [10] Huang X, Wang H. Modelling of combustion process characteristics of a 600 mw boiler by t–s fuzzy neural networks. In: 2009 International conference on eenrgy and environment technology (ICEET 2009), vol. 1; 2009. p. 737–40. doi:http://dx.doi.org/10.1109/ICEET.2009.185. [11] Embrechts M, Benedek S. Hybrid identification of nuclear power plant transients with artificial neural networks. IEEE Trans Indust Electron 2004;51(3):686–93. http://dx.doi.org/10.1109/TIE.2004.824874. [12] Rakhshani E, Sariri I, Rouzbehi K. Application of data mining on fault detection and prediction in boiler of power plant using artificial neural network. In: 2nd International conference on power engineering, eenrgy and electrical drives proceedings (POWERENG 2009), Lisbon; 2009. p. 473–8. doi:http://dx.doi.org/ 10.1109/POWERENG.2009.4915186. [13] Lemma T, Hashim F. Ifdd: intelligent fault detection and diagnosis-application to a cogeneration and cooling plant. Asian J Sci Res 2013;6(3):478–87. http:// dx.doi.org/10.3923/ajsr.2013.478.487. [14] Shatnawi Y, Al-khassaweneh M. Fault diagnosis in internal combustion engines using extension neural network. IEEE Trans Indust Electron 2014;61 (3):1434–43. http://dx.doi.org/10.1109/TIE.2013.2261033. [15] Ghate VN, Dudul SV. Cascade neural-network-based fault classifier for threephase induction motor. IEEE Trans Indust Electron 2011;58(5):1555–63. http://dx.doi.org/10.1109/TIE.2010.2053337. [16] Wolkiewicz M, Kowalski CT. On-line neural network-based stator fault diagnosis system of the converter-fed induction motor drive. In: 39th Annual conference of the IEEE industrial electronics society (IECON 2013); 2013. p. 5561–6. doi:http://dx.doi.org/10.1109/IECON.2013.6700044. [17] Refaat SS, Abu-Rub H, Saad MS, Aboul-Zahab EM, Iqbal A. ANN-based for detection, diagnosis the bearing fault for three phase induction motors using current signal. In: 2013 IEEE international conference on industrial technology (ICIT); 2013. p. 253–8. doi:http://dx.doi.org/10.1109/ICIT.2013.6505681. [18] Ballal M, Khan Z, Suryawanshi H, Sonolikar R. Adaptive neural fuzzy inference system for the detection of inter-turn insulation and bearing wear faults in induction motor. IEEE Trans Indust Electron 2007;54(1):250–8. http://dx.doi. org/10.1109/TIE.2006.888789. [19] Kumar Kirar M, Agnihotri G. Design of high speed adaptive load shedding for industrial cogeneration system. Int Rev Electr Eng 2013;8(2):867–74. [20] Hsu C-T, Chuang H-J, Chen C-S. Artificial neural network based adaptive load shedding for an industrial cogeneration facility. In: Conference records – IAS

[23]

[24]

[25]

[26]

[27]

[28]

[29]

[30]

[31]

[32]

[33]

[34]

[35]

[36]

[37]

[38]

[39]

[40]

[41]

[42] [43]

annual meeting (IEEE industry aookications society); 2008. p. 1692–9. doi: http://dx.doi.org/10.1109/08IAS.2008.137. Isazadeh G, Hooshmand R-A, Khodabakhshian A. Modeling and optimization of an adaptive dynamic load shedding using the anfis-pso algorithm. Simulation 2012;88(2):181–96. http://dx.doi.org/10.1177/0037549711400452. Wang J-J, Zhang C-F, Jing Y-Y. Self-adaptive RBF neural network PID control in exhaust temperature of micro gas turbine. In: Proceedings of 2008 international conference on machine learning and cybernetics, vol. 1–7; 2008. p. 2131–6. doi:http://dx.doi.org/10.1109/ICMLC.2008.4620758. Fazlur Rahman M, Devanathan R, Kuanyi Z. Neural network approach for linearizing control of nonlinear process plants. IEEE Trans Indust Electron 2000;47(2):470–7. http://dx.doi.org/10.1109/41.836363. Mariajayaprakash A, Senthilvelan T, Gnanadass R. Optimization of process parameters through fuzzy logic and genetic algorithm – a case study in a process industry. Appl Soft Comput 2015;30:94–103. http://dx.doi.org/ 10.1016/j.asoc.2015.01.042. Rossi F, Velázquez D, Monedero I, Biscarri F. Artificial neural networks and physical modeling for determination of baseline consumption of chp plants. Expert Syst Appl 2014;41(10):4658–69. http://dx.doi.org/10.1016/j. eswa.2014.02.001. Nikpey H, Assadi M, Breuhaus P. Development of an optimized artificial neural network model for combined heat and power micro gas turbines. Appl Energy 2013;108:137–48. http://dx.doi.org/10.1016/j.apenergy.2013.03.016. Sisworahardjo N, El-Sharkh MY. Validation of artificial neural network based model of microturbine power plant. In: Proceedings of 2013 IEEE industry applications society annual meeting; 2013. doi:http://dx.doi.org/10.1109/IAS. 2013.6682526. Ozel Y, Guney I, Arca E. Neural network solution to the cogeneration system by using coal. Int J Energy 2007;1:105–12. http://dx.doi.org/10.15623/ ijret.2014.0322023. De S, Kaiadi M, Fast M, Assadi M. Development of an artificial neural network model for the steam process of a coal biomass cofired combined heat and power (CHP) plant in Sweden. Energy 2007;32(11):2099–109. http://dx.doi. org/10.1016/j.energy.2007.04.008. Mastacan L, Olah I, Dosoftei C, Ivana D. Neuro-fuzzy models of thermoelectric power station installations. In: Proceedings-international conference on computational intelligence for modelling, control and automation, CIMCA 2005, vol. 1; 2005. p. 899–904. doi:http://dx.doi.org/10.1109/CIMCA.2005. 1631378. Tamiru A, Rangkuti C, Hashim F. Neuro-fuzzy and pso based model for the steam and cooling sections of a cogeneration and cooling plant (ccp). In: Proceeding 2009 3rd international conference on energy and environment: advancement towards global sustainability; 2009. p. 27–33. doi:http://dx.doi. org/10.1109/ICEENVIRON.2009.5398677. Zomorodian R, Rezasoltani M, Ghofrani M. Static and dynamic neural networks for simulation and optimization of cogeneration systems. Int J Energy Environ Eng 2011;2(1):51–61. http://dx.doi.org/10.1115/GT2006-90236. Jamali A, Ahmadi P, Mohd Jaafar M. Optimization of a novel carbon dioxide cogeneration system using artificial neural network and multi-objective genetic algorithm. Appl Thermal Eng 2014;64(1–2):293–306. http://dx.doi. org/10.1016/j.applthermaleng.2013.11.071. Soltani R, Mohammadzadeh Keleshtery P, Vahdati M, Khoshgoftarmanesh M, Rosen M, Amidpour M. Multi-objective optimization of a solar-hybrid cogeneration cycle: application to cgam problem. Energy Convers Management 2014;81:60–71. http://dx.doi.org/10.1016/j. enconman.2014.02.013. Gopalakrishnan H, Kosanovic D. Operational planning of combined heat and power plants through genetic algorithms for mixed 0-1 nonlinear programming. Comput Oper Res 2015;56:51–67. http://dx.doi.org/10.1016/ j.cor.2014.11.001. Shaneb O, Taylor P, Coates G. Optimal online operation of residential microchp systems using linear programicro. Energy Build 2012;44:17–25. http://dx. doi.org/10.1016/j.enbuild.2011.10.003. Wang M, Wang J, Zhao P, Dai Y. Multi-objective optimization of a combined cooling, heating and power system driven by solar energy. Energy Convers Management 2015;89:289–97. http://dx.doi.org/10.1016/j. enconman.2014.10.009. Abdalisousan A, Fani M, Farhanieh B, Abbaspour M. Multi-objective thermoeconomic optimisation for combined-cycle power plant using particle swarm optimisation and compared with two approaches: an application. Int J Exergy 2015;16(4):430–63. http://dx.doi.org/10.1504/IJEX.2015.069112. Zhao P, Dai Y, Wang J. Performance assessment and optimization of a combined heat and power system based on compressed air energy storage system and humid air turbine cycle. Energy Convers Management 2015;103:562–72. http://dx.doi.org/10.1016/j.enconman.2015.07.004. Cho H, Luck R, Eksioglu SD, Chamra LM. Cost-optimized real-time operation of chp systems. Energy Build 2009;41(4):445–51. http://dx.doi.org/10.1016/j. enbuild.2008.11.011. Ikeda S, Ooka R. Metaheuristic optimization methods for a comprehensive operating schedule of battery, thermal energy storage, and heat source in a building energy system. Appl Energy 2015;151:192–205. http://dx.doi.org/ 10.1016/j.asoc.2015.01.042. Delgoshaei P, Treado S, Windham A. Real time optimization of building combined heat and power systems. ASHRAE Trans 2012;118(1):27–33. Sayyaadi H, Babaie M, Farmani MR. Implementing of the multi-objective particle swarm optimizer and fuzzy decision-maker in exergetic,

S. Seijo et al. / Applied Energy 161 (2016) 309–319 exergoeconomic and environmental optimization of a benchmark cogeneration system. Energy 2011;36(8):4777–89. http://dx.doi.org/10.1016/ j.energy.2011.05.012. {PRES} 2010. [44] Facci A, Andreassi L, Ubertini S. Optimization of chcp (combined heat power and cooling) systems operation strategy using dynamic programming. Energy 2014;66:387–400. http://dx.doi.org/10.1016/j.energy.2013.12.069. [45] Brown DG, Morgenstern B, editors. Algorithms in bioinformatics - 14th international workshop, WABI 2014, Wroclaw, Poland, September 8–10, 2014. Proceedings. Lecture notes in computer science, vol. 8701. Springer; 2014. [46] http://www.energyworks.com/.

319

[47] Tsoukalas LH, Uhrig RE. Fuzzy and neural approaches in engineering. Wiley Interscience; 1997. doi:http://dx.doi.org/10.1016/S0893-6080(97)00079-8. [48] Mendel JM. Fuzzy logic systems for engineering: a tutorial. Proc IEEE 1995;83 (3):345–77. http://dx.doi.org/10.1109/5.364485. [49] Seijo Fernández S, del Campo I, Echanobe J, García-Sedano J, Suso E, Arbizu E. Computational intelligence techniques for maximum energy efficiency of an internal combustion engine and a steam turbine of a cogeneration process. Int J Energy Environ Eng 2014;5(2–3):1–10. http://dx.doi.org/10.1007/s40095014-0121-5. [50] http://www.optimitive.com/.