<p>Power load prediction plays a crucial role in the production and dispatching processes of power systems. Considering the current limitations of commonly used power load prediction methods in handling power load data with typical nonlinear, unstable, and sequential characteristics, this paper introduces a method called the dynamic hierarchical granularity support vector machine with temporal factors (DHGSVM-TF) algorithm. The algorithm granulates load data with different weight feature value ranges into multiple sub-granularities sets and allocates them across levels in the feature space that represent different weight value ranges; taking into consideration the temporal features of electric load data, it adjusts the feature weight values of sub-granularity sets and updates their levels. By dynamically updating the feature weight values of load data, the algorithm optimizes the weight values of load data in the predictive model, thereby enhancing the accuracy of power load prediction results. Using the GEFCom2012 electric load dataset for simulation experiments, the results reveal that, compared to traditional granularity support vector machines and other typical short-term power load prediction methods, DHGSVM-TF significantly outperforms in terms of prediction accuracy and robustness. This method also offers a new solution for electric power system load prediction that can be applied in actual power engineering projects.</p>