Park-level load exhibiting nonlinearity, multi-coupling and stochastic fluctuations, which pose challenges for accurate load forecasting. To address these issues, this study proposes a hybrid short-term load forecasting model based on CNN-LSTM-SAO-MHA. In this model, Convolutional Neural Networks (CNN) extract local temporal features, while Long Short-Term Memory (LSTM) captures long-term dependencies. The Multi-Head Attention (MHA) mechanism strengthens the model’s ability to assign adaptive weights to different time steps, thereby enhancing feature representation and improving the capture of temporal dependencies. Additionally, the Snowmelt Optimisation Algorithm (SAO) is employed for hyperparameter optimisation, enabling automatic adjustment of key parameters to enhance prediction accuracy and computational efficiency. To validate the effectiveness of the proposed model, experiments were conducted using real-world cooling load data from a typical park. The results demonstrate that the proposed CNN-LSTM-SAO-MHA hybrid model significantly outperforms benchmark models, achieving reductions of 28.74% in RMSE and CV-RMSE compared to CNN-LSTM, highlighting its superior performance in short-term park-level load forecasting.