A Hybrid Deep Learning Method With Iterative Feature Selection for Electric Load Forecasting Considering Social Activities and User Behaviors

Electric load inherently reflects the collective patterns of social activities and user behaviors, making their accurate prediction a challenging task. Accurate electric load forecasting is crucial for the planning, operation, scheduling, and market management of modern power systems, especially under the increasing complexity of residential energy consumption behaviors. From a data-driven modeling perspective, traditional load forecasting based solely on time-series data often fails to capture the social and behavioral dimensions underlying demand fluctuations. To address these challenges, this work presents an innovative electric load forecasting approach by using multifactor and time-series forecasting concepts. A comprehensive feature pool is first constructed by combining social and environmental factors, feature decomposition, and basis function transformation. Then, a metaheuristic-enhanced feature selection and modeling framework is proposed, which leverages a simulated annealing (SA) algorithm in conjunction with a hybrid deep learning architecture. Specifically, it encodes selected features as a solution of SA and evaluates it by a hybrid deep learning model that incorporates an attention mechanism, convolutional neural networks, and long short-term memory networks. In this way, it can effectively capture both temporal dependencies and social-behavioral influences on load patterns. The proposed approach is validated on 26 real-world datasets of residential electric load, which reveals that forecasting performance directly reflects aggregated social behavior in energy usage. Their synergistic effect achieves a maximum$\boldsymbol{R^{2}}$of 0.97 with a prediction error margin of less than 5% and enables the proposed approach to outperform several state-of-the-art peers. These results highlight the value of integrating social system factors with computational intelligence, showcasing the potential of the proposed method for practical applications in electric load forecasting.