Accurate design and performance assessment of solar thermal domestic hot water systems coupled with a heat pump auxiliary typically requires transient simulation, as the system’s behavior depends on multiple interactions among collector characteristics, storage stratification, control logic, weather, and draw-off timing. Monthly methods such as the f-chart are useful for first-pass estimates, but they do not resolve stratification, thermostat operation, or demand timing, and they may become inaccurate for stratified thermostat-controlled systems. Direct comparisons of locally inspectable symbolic and black-box surrogate families for this system class remain limited. A 10,982-case development dataset was generated from minute-resolved annual MATLAB simulations, parameterized by collector area, optical efficiency, and first- and second-order loss coefficients. Three surrogate families were benchmarked under a unified protocol, random forest-assisted shape-constrained symbolic regression (SR), feed-forward artificial neural network (ANN) models, and Automatic Learning of Algebraic Models for Optimization (ALAMO), with the f-chart used as a monthly reference method. The targets were the 12 monthly solar fractions under the direct solar heat definition and the corresponding annual mean solar fraction, evaluated on the same independent 991-case test set. SR achieved the lowest average error (mean absolute percentage error, MAPE = 0.82%; root mean square error, RMSE = 0.006), followed by the ANN (MAPE = 2.07%, RMSE = 0.028) and ALAMO (MAPE = 3.67%, RMSE = 0.060), with Nash–Sutcliffe efficiency (NSE) values above 0.98 for all models. Evaluation times were 0.0026–0.124 s per target, compared with about 1000 s for one full-year simulation. These results define the study as a common protocol benchmark within the studied simulator-defined envelope. SR gives the strongest accuracy with local symbolic inspectability, the ANN remains the flexible retrainable option, and ALAMO provides compact algebraic evaluation with the shortest learned model runtime.
Sourgoutsidis et al. (Sat,) studied this question.