MW-EOT-XG: Volatility-decoupled ensemble learning for enhanced day-ahead electricity price forecasting

Accurate electricity price forecasting is vital for both power generation enterprises and electricity consumers. Existing data-driven methodologies mainly focus on the single volatility of electricity prices, making it difficult to achieve a high prediction accuracy. This paper rigorously analyzes three intrinsic volatility characteristics of electricity prices: mean-reverting volatility, uncertainty, and jumps. Volatility characteristics of electricity prices are decoupled, and a Multi-scale wavelet-Encoder only transformer-XGBoost(MW-EOT-XG) ensemble learning method is developed. Through targeted modeling of the decoupled features, the methodology effectively captures time-frequency characteristics and abrupt structural changes in price dynamics. The model is validated using day-ahead electricity price data from Shanxi and Gansu provinces in China. In the Shanxi electricity market, the proposed model achieves a mean absolute error (MAE) of 34.24 and a symmetric mean absolute percentage error (SMAPE) of 16.31%. In the Gansu market, the MAE is reduced to 27.15, with a SMAPE of 21.64% and an R 2 of 0.88. These results confirm the efficacy of volatility-decoupled ensemble algorithms in accomplishing day-ahead electricity price forecasting tasks.