Research On Short-term Electricity Price Forecasting Model Based On Feature Engineering And Deep Learning

Electricity price forecasting plays an indispensable role in the decision-making process of electricity market participants.On the one hand,accurate electricity price forecasts can help market participants make more reasonable and optimized decision plans,thus reducing costs and improving efficiency;on the other hand,inaccurate electricity price forecasts may lead to poor decisions by market participants,resulting in supply-demand imbalances and market fluctuations,thus affecting the stability and reliability of the entire electricity market.Therefore,the importance of electricity price forecasting in the electricity market is receiving more and more widespread attention.However,the liberalization of the power market and the massive input of renewable energy have changed the pattern and operation of the market,and the market supply-demand relationship has become more complex and unstable,which in turn leads to the increased volatility of electricity prices and brings greater challenges and difficulties to electricity price forecasting.In order to build a forecasting model that meets the current stage of electricity market electricity price and improve the accuracy of electricity price forecasting,this paper combines feature engineering and deep learning algorithms to make relevant attempts and innovations,which mainly include:(1)A benchmark model for electricity price forecasting based on Long Short-Term Memory network(LSTM)is constructed,and a reliable framework for short-term electricity price forecasting model is proposed by combining Bayesian Optimization and Hyperband optimization algorithm(BOHB)to optimize the hyperparameters.(2)A feature selection algorithm based on Copula entropy estimation mutual information is introduced and improved by using conditional mutual information for its non-adaptive drawbacks.An adaptive feature selection algorithm(ACBFS)based on the maximum correlation and minimum redundancy criterion is proposed for the selection of input features of the electricity price forecasting model.(3)Among the popular short-term electricity price forecasting methods,the hybrid forecasting method combining signal decomposition reconstruction and forecasting model based on "divide and conquer" strategy has the problems of multiple prediction of subseries of the original electricity price signal and difficulty in predicting subseries of high frequency signal.To solve these problems,this paper proposes a new signal decomposition reconstruction method based on the "decomposition and denoising" strategy.This method changes the subject of signal decomposition and reconstruction from the original electricity price sequence to the input features of the predicted electricity price,and decomposes and reconstructs the input features by the Variational Modal Decomposition(VMD)algorithm to improve the correlation and reduce the redundancy between the features themselves and the predicted electricity price,and finally constructs a hybrid model based on ACBFS-VMD-BOHB-LSTM to improve the accuracy of short-term electricity price forecasting.The model only needs to forecast the original electricity price once,avoiding the forecasting problem of high-frequency signal subsequences.In addition,the time to decompose and reconstruct the features is much smaller than the time to forecast each original electricity price subsequence,thus reducing the computational effort of the forecasting model.(4)The forecasting performance of a series of forecasting models is cross-compared using four electricity price datasets of PJM electricity market in spring,summer,autumn and winter in 2017,and the experimental results show that the hybrid model proposed in this paper has higher forecasting accuracy and better stability.