A short-term spatiotemporal load forecasting method for power systems based on an improved salp swarm algorithm

Traditional power system load forecasting methods frequently suffer from limited accuracy and stability when processing complex and volatile data, as they fail to simultaneously model spatial dependencies and effectively optimize model hyperparameters. To address these coupled challenges, this paper proposes a short-term spatiotemporal load forecasting framework that integrates an Improved Salp Swarm Algorithm (ISSA) with a Temporal Graph Convolutional Network (TGCN). The core contribution lies in a novel two-stage optimization strategy within ISSA, which incorporates an adaptive attenuation factor to balance global exploration and local exploitation and introduces a dynamic learning mechanism for followers to prevent premature convergence. This strategy synergistically optimizes the hyperparameters of the TGCN, which explicitly captures both temporal dynamics and grid-topology-based spatial correlations. Experimental results on a provincial power grid dataset demonstrate that the proposed ISSA-TGCN model achieves superior prediction accuracy, attaining a normalized fitness value of 0.14 within only 10 iterations, corresponding to a root mean square error (RMSE) of 17.12 MW on the original load scale. Its predictions closely track actual load trends, significantly outperforming state-of-the-art benchmarks across mean absolute percentage error (MAPE), RMSE, and mean absolute error (MAE) metrics, thereby validating the effectiveness of the proposed methodological innovations.