Explainable AI for cooling load prediction and heat flow estimation using SHAP values

Effective load management for cooling energy can be realised through accurate cooling load predictions. This can reduce peak loads and ease the burden on the power grid. Accurate load prediction can be achieved using machine learning, which are black-box models. Explainable AI can be used to make machine learning models explainable. SHAP values are employed for this purpose. Currently, there is a research gap as no comprehensive and systematic investigation of the influence of building characteristic on the SHAP values of hourly cooling load predictions has been carried out. Moreover, it has not yet been investigated whether the heat flows in a room can be represented by the local SHAP values. This study investigates the influence of individual features on the cooling load predictions. The feature importance was examined both local, for a single point in time, and global, for the entire dataset, using SHAP values for a large number of various buildings with different characteristics. The results show a correlation between building characteristics and the importance of features for the cooling load prediction. Therefore, the building characteristics are implicit in the SHAP values. This enables an purely data-driven interpretation of the building characteristics. Additionally, the study demonstrated that the local SHAP values provide a satisfactory description of the heat flows in a room. This provides a purely data-driven insight into the heat flows of the building, which can reveal optimisation potential for the building.