An innovative load forecasting model for ground source heat pump systems integrating mechanistic constraints and CNN-ATT-LSTM framework

Ground-source heat pump systems (GSHPs) are extensively utilized due to their high efficiency and renewable energy characteristics. However, the complexity of factors affecting GSHPs loads poses significant challenges traditional data-driven methods in capturing the system's physical principles. Improving load predictive accuracy and generalizability remains a critical challenge. This study integrates GSHPs mechanism models into the CNN-ATT-LSTM framework, thereby developing the CNN-ATT-LSTM-Mechanistic model, which can enhance forecasting accuracy and generalization capability in complex conditions. The results demonstrate that the CNN-ATT-LSTM-Mechanistic model outperforms the CNN-ATT-LSTM, CNN-LSTM, and LSTM models in terms of MAE, RMSE, and R2. For instance, under the winter and summer conditions of Project A, as well as under the same conditions of Projects A and B, the model achieves a 20% improvement in these metrics, underscoring its robust predictive performance across the investigated case studies and its potential for transferability to similar GSHPs. The above indicates that effectively integrating physical mechanisms and data-driven methods is crucial for enhancing the accuracy and generalization of GSHPs load forecasting, significantly boosting its predictive performance and practical applicability under complex conditions.