An Electricity Market Price Forecasting Model Based on Multi-Task Learning and Long Short-Term Memory Networks

Accurate electricity price forecasting is indispensable for efficient power market operation and risk management, yet it remains challenging due to the inherent volatility, non-linearity, and complex dependencies characterizing price dynamics. This paper proposes a novel forecasting model integrating Multi-Task Learning with Long Short-Term Memory networks (MTL-LSTM) to enhance prediction accuracy and robustness. The model leverages LSTM's proven ability to capture long-range temporal dependencies through its gating mechanisms, while employing MTL's hard parameter sharing paradigm to learn shared representations across related forecasting tasks (e.g., different day types), thereby improving generalization and integrating diverse influencing factors like historical prices, load demand, weather conditions, and calendar information. This paper details the theoretical underpinnings of LSTM and MTL, the specific architecture of the proposed MTL-LSTM model, and the comprehensive methodology encompassing data preprocessing, feature engineering, model training, and hyperparameter optimization. Empirical validation using extensive historical data (2022-2025) from the PJM electricity market demonstrates the model's effectiveness. Results show excellent convergence during offline training with negligible overfitting and superior performance in tracking actual price fluctuations for both weekdays and non-weekdays during online testing.