Machine learning based surrogate modeling for heat recovery and energy management

Enhancing heat-recovery efficiency is an important component of industrial energy management and emissions-reduction strategies. This study presents a machine-learning-based surrogate-modeling framework for predicting the logarithmic mean temperature difference (LMTD) in heat-exchanger applications. A synthetic thermal dataset containing 10,000 operating scenarios and 20 numerical variables was evaluated. Because the clean LMTD variable is deterministically related to other generated thermodynamic quantities, a noisy LMTD variable was selected as the prediction target to provide a more realistic assessment of predictive robustness. Four regression algorithms were compared: Linear Regression, Random Forest Regressor, Extra Trees Regressor, and Gradient Boosting Regressor. Dataset-quality screening identified four zero-variance predictors, which were removed, as well as a physically questionable cold-outlet-temperature variable that was explicitly treated as a limitation of the synthetic dataset. Three leakage-aware feature-set experiments were conducted to distinguish prediction based on direct thermodynamic relationships from prediction based on broader operating information. Using the reduced feature set after removing direct heat-load and LMTD-related variables, Linear Regression achieved the best performance, with R² = 0.7644, RMSE = 10.6492 K, and MAE = 8.4317 K. When the heat-load, LMTD, and temperature-equation inputs were removed, model performance decreased to approximately zero or became negative, demonstrating that predictive capability depended strongly on thermodynamically informative variables. Input-perturbation testing indicated limited performance degradation under moderate additional synthetic noise. A surrogate-based optimization case study was also presented as a numerical demonstration; however, its predicted operating point should not be interpreted as physically implementable because of the limitations of the synthetic dataset. The proposed framework is therefore positioned as a leakage-aware diagnostic and surrogate-modeling workflow rather than as a replacement for the analytical LMTD equation or as an experimentally validated heat-exchanger model.