1. Introduction
Combined cycle power plants (CCPPs) are one of the power plants types that could produce electricity with high efficiency and low air pollutant. However, the CCPPs production cost is relatively higher than the coal-fired power plant due to prime energy costs. Hence, operating CCPPs as efficiently as possible is required to decrease production costs. CCPP highest efficiency could be achieved with maximum load and operating in base load. Unfortunately, energy supply instability is a common problem of CCPPs in Indonesia. This issue leads the CCPP management staff struggling to predict the performance of their CCPP. Assessing historical performance data using machine learning will help to predict the CCPP performance accurately.
Some literature about gas turbine (GT) or CCPP modeling with machine learning has been studied. Wood [
1] used the transparent open box (TOB) to predict the combined cycle gas turbine’s power output. Liu et al. [
2] predicted gas turbine performance with high dimensional model representation (HDMR) and an artificial neural network (ANN). A distributed control system (DCS) was utilized as a tool to provide selected input data. It was presented that the holistic ANN model had a higher accuracy to predict gas turbine performance. The ANN was utilized to describe the micro gas turbine performance and effectively evaluate it in a real installation in any climate as presented by Bartolini et al. [
3]. Anvari et al. [
4] utilized an ANN on combined cooling heating and power (CCHP) to predict GT performance prediction. Some underlying parameters were chosen as inputs but did not include any fuel gas characteristics. Fast et al. [
5] developed an ANN model to predict the performance of an industrial GT. Relative humidity, ambient pressure and ambient temperature were used as input parameters. It was reported that the best possible modeling of the plant was optimized for 10,000 epochs. Rossi et al. [
6] reported ANN modeling for baseline consumption of a combined heat and power plant. Elfaki et al. [
7] predicted the electrical output power of CCPP using the regression ANN model. Moreover, ANN has been applied to other GT-based application [
8,
9,
10,
11]. ANN is also widely used in power plants with multiple objectives. Smrekar et al. [
12] predicted power generated by a coal-fired power plant. Thermal efficiency and power plant pollutant were also predicted with ANNs [
13,
14,
15,
16].
To date, there is no actual study predicting CCPP performance using fuel gas characteristics such as gas heat units in million British thermal units (MMBTU) or CO
2 percentage, even though those two parameters are usually used in the gas purchase agreement between the gas suppliers and power plant operator. Furthermore, the power plant operator usually has a plant efficiency written in their contract with the regulator. Omar et al. [
17] showed that decreasing power generation leads to higher specific fuel consumption.
According to the research gap, this paper proposes a prediction method of CCPP heat rate performance with gas heat unit in million British thermal units (MMBTUs), carbon dioxide (CO2) percentage contain in the fuel gas and generated power in megawatts (MW) as input parameters. Some studies above showed that ANN is effective in predicting gas turbine performance. Therefore, an ANN was used as a machine learning tool to predict the power plant heat rate.
2. Materials and Methods
Artificial neural networks are a technology-based on studies of the brain and nervous system. These networks emulate a biological neural network, but they use a reduced set of biological neural systems concepts. Specifically, ANN models simulate the electrical activity of the brain and nervous system. Processing elements (also known as either a neurodes or perceptrons) are connected to other processing elements. Typically, the neurodes are arranged in a layer or vector, with the output of one layer serving as the input to the next layer and possibly other layers. A neurode may be connected to all or a subset of the neurodes in the subsequent layer, with these connections simulating the brain’s synaptic connections. Weighted data signals entering a neurode simulate a nerve cell’s electrical excitation and, consequently, the transference of information within the network or brain. The input values to a processing element are multiplied by a connection weight that simulates the strengthening of neural pathways in the brain. Through the adjustment of the connection strengths or weights, learning is emulated in ANNs [
18].
An ANN is comprised of an input layer, hidden layers and an output layer. All input nodes have their weight connection of a neuron. Then the relationship between input parameters and output parameters is learned through the iterative training process. Linear and nonlinear activation functions are included in the ANN system, resulting in a good prediction of nonlinear behavior systems.
The components of ANNs are neurons, transfer functions and predefined layers. In a multilayer perceptron (MLP), the transfer function used are tan-sigmoid (tansig), log-sigmoid (logsig), and linear purelin functions [
19]. Tansig receives the signal in (−∞, +∞) and converts them to output signal within a (−1, +1) range. Logsig receives the input signals in the range of (−∞, +∞) and transfers them to the (0, 1) limit. Pureline is a linear function that maps the input data between (−∞, +∞) and results in the identical data in (−∞, +∞) interval [
4].
Our proposed method for predicting heat rate in a CCPP was developed and demonstrated for a 740 MW CCPP with configuration two GTs, two heat recovery Sseam generators (HRSGs), and one steam turbine. However, this paper only provides predictions with configuration 1-1-1 (one GT and one HRSG stop) due to a lack of gas supply. Real operation data were collected over one year (2019).
Figure 1 presents this paper’s flowchart. The input parameters in this paper were MMBTU, CO
2 percentage and power generated in MW. Total generated power of the CCPP, from a GT and steam turbine, recorded by a distributed control system (DCS) was used as an input parameter in this paper. Since the DCS did not provide a fuel gas heating unit in MMBTUs and CO
2 percentage, the gas supplier’s hourly data were used. The gas provided by the supplier went to the combustion chamber and was used as a fuel for the GT. The exhaust gas from the process was used to feed the HRSG. Since the total heat from these two processes is equal to the gas supplied, the gas data from supplier was used. Equation (1) was used to calculate the plant heat rate [
20].
where: HR = Heat Rate (kcal/kWh); Mf = Fuel gas flow (kNm
3/h); HHV = High Heating Value (BTU/SCF); P = power output (MW).
One input parameter was used with the ANN to see the result in predicting the heat rate. Two combined parameters were used with the ANN to see the result in predicting the heat rate. And finally, the three combined parameters were used with the ANN to see the result in predicting the heat rate. Fuel gas flow was recorded in a DCS and a High Heating Value (HHV) recorded by the gas suppliers and used in this paper. As shown in
Figure 1, the generated power recorded by the DCS was the key to get clean data, as per 15 min of data recorded by the DCS collected and evaluated. The data was normalized by rounding it to the nearest multiple of five, then evaluated by comparing it with data 15 min before and after. If the normalized data had an equal value with data 15 min before and after, the data would continue to the next step. If the value was not equal, then it was terminated. This step was necessary to determine whether the CCPP was in steady-state load condition or not. The measured data were selected carefully to remove the transient during start-up, shut down and load change. Finally, 4322 clean data were provided in steady-state power plant conditions, as presented in
Table 1. The input parameters were combined as shown in
Table 2 to find the best combination to predict heat rate.
A feed-forward backpropagation network type was implemented in all ANN models and trained using gradient descent with momentum and adaptive learning rate backpropagation (traingdx). Mean squared error (MSE) was used as the loss function. With a back-propagation algorithm, the learning function changed the weight between the neurons to reduce the difference. Since success implementation for power plant application has been shown in other studies [
2,
5,
8], the tansig and purelin activation function was implemented in this paper. One hidden layer was used with purelin as a transfer function. The total neuron used was one neuron. ANN structure for the model with one input parameter is shown in
Figure 2.
Two input parameters were combined as presented in the
Table 2. In these models, two hidden layers were used. The first hidden layer used four neurons as a rule of thumb of the hidden layer as the input layer square, and the second hidden layer used 1 neuron. The transfer function used in the first hidden layer was tansig, and the second hidden layer used purelin as a transfer function. The ANN structure for the model with two input parameters is shown in
Figure 3.
The last model was the combination of all three input parameters. Two hidden layers were used, with the first hidden layer using nine neurons as a rule of thumb of the hidden layer as the input layer square, and the second layer used 1 neuron. The first layer used tansig as a transfer function, and the second layer used purelin as a transfer function.
Figure 4 shows the ANN structure for the model with three input parameters:
The output parameter for all models was the plant heat rate, since the plant heat rate was usually used in a contractual agreement to judge whether the power plant was operated efficiently or not. Each model used 10,000 epoch iterations and 1000 validation checks. Data used for training was 50%, while 25% was used for validation and the other 25% used for the test.
4. Discussion
The experimental results showed that ANN could predict CCPP heat rate performance with correct input parameters. Those models with one input parameter were not able to execute 10,000 epoch iterations. The ANN always met the maximum validation check allowances. Combining two parameters could lead to a better heat rate prediction. P1 + P3 that used fuel gas heat input and power output was a more accurate model than the other models with two input parameters. P1 + P2 was the only two parameter input that did not meet the iteration expected. The training stopped at 5346 iterations because the validation check met the maximum allowance. The combination of P2 + P3 could meet the expected iteration. However, the R
2 regression was only 0.984, and the validation check was 17. Combining three input parameters, P1 + P2 + P3, could produce better accuracy.
Table 6 shows that P1 + P2 + P3 led to the highest regression R
2 value. It also had a zero validation check. A validation check refers to the number of errors from the dataset. Lower validation checks could lead to higher accuracy from the model to predict the desired data. The experiment data showed that adding CO
2 percentage as an input parameter could lead to higher accuracy in predicting CCPP heat rate value. All the variation had average error data lower than 2%. This indicates that the NN could predict heat rate accurately, and the P1 + P2 + P3 with the lowest error data could be the best variation to predict heat rate.
Input parameter, P1 + P2 + P3 as listed in
Table 5, were used in ANN method to predict the actual heat rate as presented in
Table 6 with 1028 data predictions. The average error for prediction was 2.52% with max error value 10.88%