Short-term load forecasting plays a critical role in the operation and planning of power systems. Accurate prediction of load variations not only optimizes grid dispatching and enhances grid reliability but also provides decision-makers with scientific evidence to reduce operational costs and improve system efficiency. The development of adaptable short-term load forecasting models has become pivotal to improving prediction accuracy. To address seasonal variations, external environmental factors, and dynamic user behavior, this study proposes the XLSTM model, which integrates SLSTM and MLSTM architectures. The model leverages their complementary strengths to capture spatial-temporal correlations within load data, effectively mitigating prediction accuracy degradation caused by data drift during extended forecasting periods. Simulations using cross-seasonally partitioned real-world power system load data were conducted for both single-step and multi-step forecasting, with comparative analyses against standalone LSTM and CNN-LSTM models. The results demonstrate that XLSTM outperforms other models in both forecasting scenarios, validating its robust generalization capability and effectiveness in alleviating data drift challenges in short-term load forecasting.