Management of waste heat in residential buildings: Predictive modelling and sensitivity analysis of variables characterising shower heat exchanger conditions

The article presents developed predictive models capable of predicting the values of three output variables describing the functioning of a shower heat exchanger. The data set for training predictive models was obtained through comprehensive experimental tests, taking into account four input variables, two heat exchanger lengths and three configurations of its operation. Three machine learning methods were compared, i.e. artificial neural networks (ANN), eXtreme Gradient Boosting (XGB) and Random Forest (RF). The method that proved to guarantee the highest prediction efficiency was XGB. Using the developed XGB models, SHapley Additive exPlanations (SHAP) analysis was performed. This innovative approach enabled an understanding of the impact of individual input variables on the predictive model outputs, enhancing their interpretability, transparency, and reliability. In the case of balanced flow, the mixed water flow rate has the highest impact on the values of all three output variables. In the case of unbalanced flow, this input variable is the most important only for the heat transfer rate (P). For heat exchangers effectiveness (ε) and percentage of recovered heat (R), the most important input variable turned out to be the drain water temperature. This study presents accurate tools for predicting and analyzing performance of shower heat exchangers.