Power load forecasting using deep learning and reinforcement learning

Accurate power load forecasting plays a pivotal role in balancing power supply and demand within smart grid development. While hybrid forecasting technologies have gained popularity in power load forecast, existing forecasting modules often rely on traditional models such as statistical methods, machine learning, and long short-term memory (LSTM), which limits their diverse applications. To resolve this challenge, we develop a novel multi-factor and multi-scale power load forecasting method. The proposed method comprises three steps: firstly, power loads and meteorological factors are decomposed using the noise-assisted multivariate variational mode decomposition (NA-MVMD) method across multiple scales, capturing temporal characteristics. Next, individual forecast is generated for each decomposition using the adaptive Nesterov momentum algorithm (Adan) and inverted transformer (iTransformer). Finally, the weights corresponding to each prediction result are determined using the Dueling Deep Q-Network reinforcement learning method, and the results are aggregated by weighted summation. The empirical study using the New York City dataset demonstrates that the proposed model achieves robust stability and high forecast accuracy comparing with other competitors. These findings reveal its potential to inform smart grid optimization strategies effectively.