Deep learning quantile regression multi-step electricity price prediction based on MIC-CEEMDAN decomposition

As the power market reform deepens, market player competition intensifies, and the challenge of formulating more rational trading strategies becomes significant. Electricity price forecasting, as the core basis for the formulation of trading strategies, has gradually become a focus of research. In this context, this paper proposes an electricity price forecasting model based on MIC-CEEMDAN-BiLSTM-Attention-QRNN. Firstly, the Maximal Information Coefficient (MIC) model is utilized to screen electricity price-related factors, proposing a reasonable factor matrix as input for the forecasting model. Secondly, Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) is employed to decompose the original data matrix, generating different decomposition sequence curves to reduce the impact of the volatility of the original data on the forecasting model and enhance its stability. Thirdly, Bidirectional Long Short-Term Memorymodel(BiLSTM)Attention-Quasi-Recurrent Neural Networks (QRNN) is used to predict the spot market electricity price based on the different decomposition curves, improving the feature learning and prediction capabilities of machine learning. Finally, a case analysis using PJM electricity market data demonstrates that the electricity price forecasting model proposed in this paper is highly practical and can meet the needs of spot market electricity price forecasting.