A data driven model based approach for medium-to-long-term electricity price forecasting in power markets

Accurate medium-to-long-term electricity price forecasting constitutes a critical prerequisite for electricity market participants to formulate optimal bidding strategies and mitigate energy procurement expenditures. However, medium-to-long-term electricity price forecasting encounters challenges including high-dimensional data acquisition difficulty, forecasting mode complexity with poor adaptability, and high volatility of the forecasted prices. In this manuscript, a data-driven-model-based approach for medium-to-long-term electricity price forecasting is proposed. Firstly, the key data sets are screened from the importance assessment based on the decision tree, and the historical electricity price, the coal price, and the natural gas price are chosen as the key price data to establish the medium-to-long-term electricity price forecasting model to reduce the forecasted data dimension effectively. Secondly, the key data sequences are denoised with the volatility reduction through combining Fast Fourier Transformation (FFT). Then, GWO-CNN-LSTM-Attention model, combining Gray Wolf Optimization (GWO) algorithm, Convolutional Neural Network (CNN), Long-Short-Term Memory (LSTM) network and Attention mechanism, is constructed for better forecasting performances and enhanced adaptability. Finally, the price time series are separated into the primary and residual frequencies. The primary frequency of the historical local electricity price is fed into GWO-CNN-LSTM-Attention algorithm to forecast the trend of the electricity price, while the primary frequencies of the historical coal price and the natural gas price are fed into GWO-CNN-LSTM-Attention algorithm to forecast the fluctuation of the electricity price, contributing to the high accuracy of the price forecasting in long term. The proposed FFT-GWO-CNN-LSTM-Attention (FGCLA) algorithm facilitates the dimension reduction of the forecasting dataset, the adaptability enhancement of the forecasting model, and the effective suppression of the high volatility, so as to improve the forecasting accuracy of the electricity price. The proposed algorithm is compared to the traditional LSTM and Transformer algorithms through simulation cases, with forecasting accuracy improvements of 57.21% and 49.69% respectively on average, concluding that the proposed algorithm could effectively reduce the forecasting errors of the medium-to-long-term electricity price.