Improved butterfly optimization algorithm applied to prediction of combined cycle power plant

Abstract

The electricity output is worth monitoring because of the rising electricity demand. Ambient temperature, air pressure, relative humidity, and exhaust pressure all impact the production output of a combined cycle power plant. This study proposes the BOAPPE algorithm, which combines the butterfly optimization algorithm (BOA) with the phasmatodea population evolution algorithm (PPE) to estimate power output better and reduce excessive cost waste. When used in conjunction with a support vector regression (SVR) model, such as BOAPPE-SVR, for estimating the power output of a power plant under basic load, it not only enhances the model’s prediction accuracy but also successfully avoids the problem of the model entering local optimization too soon. The parallel strategy of the BOAPPE algorithm in this research improves convergence speed, while the random walk strategy prevents the model from sliding into local optimization. The results reveal that the model paired with the BOAPPE algorithm is more accurate and better than the other models in this research when comparing the performance of the parameters such as mean square error, relative error, and correlation coefficient. As a result, the BOAPPE-SVR model is a viable model for power load forecasting.

Introduction

The utilization of combined cycle power plants (CCPP) is expanding all over the world as the demand for electricity rises [33]. CCPP is a large power plant that combines a gas turbine and a steam turbine. They use fossil fuels to achieve the highest technical efficiency of power generation, which has now exceeded 0.60 [23].

The essential load operation of a power plant is influenced by four key elements that determine the electric power output: ambient temperature, atmospheric pressure, relative humidity, and exhaust pressure [36]. Any change in these parameters may change the power output of the system. Effectively predicting the change of output caused by the change of parameters significantly impacts power production. Only by precisely anticipating the full-load power production of base-load power plants can electricity consumption be better managed and allocated.

In the literature [13], by using the artificial neural network (ANN) model of the multi-layer feedforward network model, the operation and performance parameters of a thermal gas turbine in different environments are accurately predicted. The application of intake heating technology in CCPP is proposed in Ref. [38]. The results show that when the inlet gas temperature is higher, the thermal efficiency of the thermal gas turbine will be reduced, and the energy conversion efficiency of CCPP will be improved. The gas turbine cycle is analyzed and studied in Ref. [27]. A better rational efficiency value of the turbine is obtained by using the response surface method (RSM) to optimize the input variables. In Ref. [7], a novel sequence model is proposed that can detect the degree of engine component degradation, minimize the computational burden of diagnosis procedures, and increase gas turbine reliability and energy efficiency. In [14], [22], [31], ANN is used to predict the operation characteristics and efficiency of burners and turbines.

Previous researchers mainly built models for evaluation through machine learning and neural networks. SVR in machine learning has achieved remarkable results in data regression prediction. In recent years, the public has widely known intelligent optimization algorithms because of their excellent optimization ability, and it is applied to the parameter optimization of the model. ANN is one of them, and it excels at coping with non-linear situations. At present, we are in an era of big data, and there must be a tremendous amount of information changes in the development of everything. Combining these data analyses with existing models for analysis and prediction, can effectively avoid some unnecessary losses. In [39], the model is applied to the field of ocean exploration and early warning to predict the height of waves, and hybrid thinking evolutionary algorithm (MEA) paired with BP neural network (BPNN) is utilized to improve the prediction and generalization capacity of BPNN [26]. By using chest X-rays to detect COVID-19 and other pneumonia, a new neural network structure is presented to speed up the doctor’s differential diagnosis of patients. SVR has also achieved good results in many application fields, such as predicting natural disasters [4], [11], [28]. In Ref. [24], the author uses a particle swarm optimization algorithm (PSO) combined with a support vector machine (SVM) to predict the charging state of lithium battery in fault state through cross-test. The authors of [21] found the ideal coefficients for the four kernel SVR model’s kernel functions using the Human Learning Optimization Algorithm (HLO), and then utilized the ideal SVR model to simulate the compressive strength of concrete.

The newly suggested BOAPPE algorithm is the primary focus of this research. The SVR model’s unknown parameters are retrieved using real data set by this algorithm, and the SVR model is continuously refined to precisely measure the power output. Additionally, in order to train the SVR model to estimate the maximum power of the base-load power plant equipment, comparative experiments were conducted using BOA, PPE, PSO, Moth Flame Optimization Algorithm (MFO), and Aquila Optimizer (AO). The following are the contributions of this paper:

1. To optimize the SVR model, this study uses a better hybrid heuristic method called BOAPPE [8], [29], which combines BOA and PPE. It also uses two search techniques to increase the search space and speed up convergence while maintaining a balance between development and exploration.

2. In this study, six hybrid mathematical models are constructed, studied, and compared for their ability to predict power generation: BOA-SVR, PPE-SVR, PSO-SVR, MFO-SVR, AO-SVR and BOAPPE-SVR. According to the experimental findings, the BOAPPE-SVR model is more accurate than other SVR models.

3. This paper extracts the unknown parameters of the four SVR kernel functions as precisely as possible and shows that the Gaussian kernel function is the most effective kernel function for the SVR model.

4. Use the proposed BOAPPE-SVR model to successfully analyze the CCPP data set in the UCI machine learning library.

The remainder of the paper’s structure is the following: Section 2 introduces similar works of earlier studies. Section 3 describes the methodology and materials employed in this paper. Section 4 introduces the revised algorithm; Section 5 contains the experimental data and performance analysis. The conclusion and future work are presented in Section 6.