In multi-energy complementary highway energy supply systems, the understanding of electrical load variation is considered essential for scientific energy planning and accurate dynamic allocation. To improve the accuracy of long-term electrical load forecasting, a prediction model based on long short-term memory networks (LSTM) and a multi-head attention (MHA) mechanism was proposed. Firstly, feature engineering was applied to expand the dimensionality of input features. Then, an LSTM network was built to model temporal dependencies in electrical load data. The MHA mechanism was added to highlight key time-step information in the sequence. These components were combined to form the MHA-LSTM forecasting model. Finally, real electrical load data from a specific region were used to validate the model. The proposed model achieves a mean absolute percentage error (MAPE) of 3.23%, an coefficient of determination (R²) of 0.882 3, and an relative percent deviation (RPD) of 2.914 2. Improved performance is shown compared to traditional LSTM and attention-based LSTM models. Strong fitting ability and high predictive accuracy are demonstrated for long-term forecasting with complex temporal patterns.