Abstract: To address the limitation of traditional models in accurately modeling complex nonlinear relationships in traditional short-term load forecasting models, a novel hybrid architecture integrating long short-term memory (LSTM) networks with a Chebyshev polynomial-based function network and termed the LSTM-ChebyKAN model—was proposed. In this model, the ChebyKAN module was constructed by replacing the spline functions used in traditional Kolmogorov-Arnold network (KAN) with numerically stable Chebyshev polynomials. This design retained the temporal modeling capability of LSTM while significantly enhancing both function approximation capacity and training stability. Experimental validation was conducted using actual load data from the UK National Grid spanning 2009—2024. The results indicate that the proposed model outperforms baseline models such as RNN, GRU, LSTM, and LSTM-KAN across all forecasting accuracy metrics. Compared with the strongest competitor, LSTM-KAN, the proposed model achieves a reduction of approximately 16.7% in mean squared error (MSE), a decrease of about 12.7% in mean absolute error (MAE), a reduction of around 5.9% in mean absolute percentage error (MAPE), and a significant improvement in the coefficient of determination (R²). Further cross-seasonal generalization experiments confirm its superior robustness. These findings demonstrate that the LSTM-ChebyKAN model captures complex nonlinear load characteristics more accurately and stably, offering an effective technical pathway to enhance the accuracy and practical reliability of short-term load forecasting.