A novel reinforcement learning-based approach for short-term load and price forecasting in energy markets

Short-term forecasting of energy load and price is vital for the stability and cost-effectiveness of modern smart grids. This article introduces a novel Deep Reinforcement Learning (DRL) based method that models forecasting as an MDP with a customized reward function balancing load and price errors, outperforming traditional approaches. We treat the complex, step-by-step prediction process as a Markov Decision Process (MDP), defined by four main components: a multi-dimensional State that includes historical trends and extra features like weather and fuel; an Action space that represents the predicted load and price values for the next time step; a Reward function aimed at reducing prediction error; and an optimal Policy. The heart of the model is a Deep Q-Network (DQN), which uses deep neural networks (DNNs) to estimate the optimal Q-value function, Q(s, a). The agent learns the best prediction strategy by continuously interacting with both historical and real-time market data, applying the Q-learning algorithm to improve the value function over time. Using the PJM Interconnection dataset spanning 2021-2023, we have performed the proposed strategy for short-term load and price forecasting. Simulation results on real-world data show a 15-20% reduction in mean absolute percentage error (MAPE), outperforming baselines such as ARIMA, LSTM, and XGBoost. Also, the obtained results illustrate that this DRL-based method greatly reduces prediction error. This improvement leads to better operational efficiency, smarter resource use, and increased reliability in smart energy networks.