Novel Multimodal Data for Enhanced Electricity Spot Price Forecasting Using A CNN-LSTM Ensemble Learning Model for the Japan Electric Power eXchange (JEPX) Spot Market

Electricity price forecasting (EPF) is critical in energy markets, particularly with the rise of renewable energy sources (RES) in Japan, which can cause day-ahead spot prices to drop to nearly zero JPY/kWh, impacting retailer profitability. This study demonstrates that the Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) model outperforms the LSTM model in both prediction accuracy and computational efficiency in Japan's electricity spot market. A novel ensemble learning strategy enhances both the robustness and accuracy of the EPF model is proposed. Novel multimodal explanatory variables, including electricity spot price, system price, discretionary cost, sell bid amount, buy bid amount, total contracted volume, actual power generation, actual solar power generation, meteorological forecasts and calendar forecasts, alongside the rolling features of spot prices, are utilized and verified. Furthermore, a “policy-versus-policy” approach addresses the zero-inflated regression issue of the zero price prediction is proposed. Our model, with a comprehensive feature integration, achieves an RMSE of 5.66 JPY/kWh and an R2 of 0.729 during the test range from 2022.01.01 to 2022.12.31. The paper also introduces a novel method for estimating confidence intervals using ensemble learning.