Abstract: To address the insufficient characterization of temperature-humidity coupling effects, the neglect of high-temperature lag accumulation, and the poor interpretability of traditional deep learning models in short-term load forecasting under complex meteorological conditions, this paper proposes a short-term load forecasting model based on temperature-humidity cumulative effects and the Temporal Kolmogorov-Arnold Network (Temporal KAN). First, the Temperature Humidity Index (THI) and heat accumulation features are introduced to construct a multivariate dynamic meteorological feature matrix, which captures the nonlinear impacts of temperature-humidity interactions and continuous high-temperature events on load variations. Furthermore, a Temporal KAN framework is developed by employing learnable B-spline functions on network edges instead of fixed node activation functions in conventional neural networks, enabling effective integration of temporal load patterns and meteorological driving factors. The proposed method provides an effective approach for accurate load forecasting and refined power system dispatch under complex weather conditions.