Thermal Load Prediction in District Heating Systems Using GAN-Based Data Augmentation and a Dynamic Weighted LSTM-Prophet Hybrid Model

Efficient and precise regulation of heating networks constitutes a critical technological pathway for achieving energy optimization and carbon emission reduction objectives. To enhance the prediction accuracy of thermal load in district heating systems while addressing differentiated demands across heat substations, this study proposes a personalized prediction methodology integrating data augmentation and hybrid model optimization. Initially, temperature, humidity, and temporal features were identified as key influencing factors through Pearson correlation coefficient analysis. A deep generative adversarial network (GAN) was subsequently employed for multi-dimensional scenario data augmentation, constructing an extended dataset. Building upon this foundation, an innovative LSTM-Prophet dynamic weighted hybrid model was developed, tailored to the topological characteristics of a regional heating network in Tianjin, China, enabling personalized weight configuration for individual substations to fulfill their heterogeneous operational requirements. Taking heat substation No. 566 as the case study, the impact patterns of weight coefficient w (denoting the weight of Prophet) on prediction accuracy were systematically investigated. Experimental results demonstrate that when w was set to 0.5, 0.7, 0.9, and 0.95, the mean relative errors between predicted and actual thermal load values reached 4.9%, 2.1%, -0.65%, and -1.45%, respectively. Consequently, the model exhibited optimal comprehensive performance at w = 0.9 for this specific substation. Error analysis revealed that negative error values indicate moderately conservative predictions, which strategically enhance operational reliability of the heating system. This weight optimization strategy provides a technical reference for differentiated regulation in district heating systems.