A hybrid model for power customer load forecasting has been introduced to tackle the challenges posed by the high dimensionality, complex features, and significant interference present in current power data. Utilizing the integrated empirical mode decomposition model, electricity consumption characteristics of power users are decomposed, separating the features into high-frequency and low-frequency components based on the zero crossing rate. Employing a multi-objective evolution-deep belief network, the low-frequency components are processed to accurately forecast the overall trends. Utilizing an enhanced long short-term memory network, the high-frequency components are processed, significantly improving the capability to handle complex nonlinear local behaviors and ensuring precise high-frequency load forecasting. Utilizing the superposition rule, the load forecasting is reconstructed to refine predictions of local fluctuations, markedly enhancing the model’s overall performance. Experimental results indicate that, compared to models such as KNN, BPNN, RNN, and LSTM, the proposed model achieves an average reduction in the mean absolute percentage error. This model demonstrates superior load forecasting accuracy and can offer insights for enhancing the safe operation and service quality of distribution networks.