Development of a heat-exchanger performance degradation model by integrating physics-based modeling with an LSTM approach

Abstract

Performance degradation in heat exchangers poses a significant risk to process stability and equipment safety. This study presents a predictive framework that combines the physical indicator of log mean temperature difference with a long short-term memory neural network to monitor degradation trends. In addition to LSTM, the study evaluates seven other time-series models, including AR, MA, ARMA, ARIMA, KNN, SVR, and Transformer. Model performance was assessed using five statistical metrics: mean absolute percentage error, mean squared error, root mean squared error, mean absolute error, and coefficient of determination. Among all models, LSTM consistently delivered the most reliable results across both training and test datasets. During testing, the LSTM model achieved a R² of 0.992, DTW similarity of 94.5 percent, and MAPE below 0.1 percent. These results confirm the model's strong fitting capability and generalizability. The proposed approach successfully addresses challenges such as the nonlinear behaviour of thermal signals and the lack of pronounced degradation features. It offers practical value for maintenance planning and process shutdown decision support in real industrial settings.