Reduced-order prediction of transient temperature-field evolution in solar photovoltaic water-heating systems using physically informed features

Accurate and fast prediction of thermal stratification in solar photovoltaic water-heating tanks is essential for real-time energy management, yet remains difficult because fluctuating PV power and water draw-off jointly induce strongly nonlinear, multi-timescale thermal dynamics. To address this challenge, this study proposes a physics-guided reduced-order learning framework (POD-PI-LSTM) for transient temperature-field prediction. A hybrid dataset is constructed from experiments and CFD simulations validated under both constant-power and time-varying PV heating conditions. Proper Orthogonal Decomposition (POD) is employed to extract dominant low-dimensional thermal modes, while physically informed features (PI) describing PV excitation dynamics are embedded into an LSTM network. In addition, a physics-guided regularization term is introduced into the training objective to enforce consistency between the predicted reduced coefficients and the global energy evolution, stratification development, and heater-dominated local thermal structures under the adopted reduced-order thermodynamic assumptions. Compared with POD-LSTM, the proposed model reduces RMSE and MAE by 45.71% and 39.80%, respectively, while achieving orders of magnitude lower computational cost than CFD. The framework maintains strong generalizable prediction under unseen operating conditions, offering a practical solution for real-time digital twins, predictive control, and intelligent optimization of PV-coupled thermal storage systems.