Time-Frequency Domain Fusion Prediction Based on the Time-Varying Correlation of the Integrated Energy System Load

Introduction: Accurate integrated energy load forecasting plays an important role in optimizing energy scheduling, reducing carbon emissions, and promoting efficient energy utilization. To enhance forecasting accuracy by capturing complex temporal dynamics, a load forecasting method combining time-domain and frequency-domain characteristics is proposed, which considers the time-varying correlation of load. Methods: In this method, the sliding correlation analysis is conducted in the time domain section, and the feature influence weight is allocated accordingly. A load reconstruction and extraction module is designed, and the time domain load characteristics are learned in a multi-task manner. In the frequency domain section, periodic characteristics are captured by an attention mechanism, and the frequency domain information is extracted. Based on the energy proportion of the sequence in the high-frequency and low-frequency parts, the adaptive weight is designed to calculate and allocate the real-time weights of both the time-domain and frequency-domain sections, enabling accurate prediction of cooling, heating, and power loads within the integrated energy system. Results: The proposed model is compared with other models, and an ablation experiment is conducted on the proposed model. The effectiveness of the proposed method is evaluated in point prediction. Experimental results show that the proposed model outperforms traditional methods, achieving lower MAPE, RMSE, and MAE values, with a total MAPE as low as 4.091% in three- -step forecasting. Discussion: Experimental results indicate that the proposed method achieves smaller prediction errors than all comparison models. These findings indicate the potential value of our approach for integrated energy system load forecasting. Conclusion: The proposed model effectively captures periodicities and energy coupling relationships in integrated energy systems using a two-dimensional feature matrix approach. Future work should focus on higher-resolution data and improved feature engineering for better local pattern recognition.