Improving heat transfer predictions in heterogeneous riparian zones using transfer learning techniques

Abstract

Data-driven deep learning models usually perform well in terms of improving computational efficiency for predicting heat transfer processes in heterogeneous riparian zones. However, traditional deep learning models often suffer from accuracy when data availability is limited. In this study, a novel physics-informed deep transfer learning (PDTL) approach is proposed to improve the accuracy of spatiotemporal temperature distribution predictions. The proposed PDTL model integrates the physical mechanisms described by an analytical model into the standard deep neural network (DNN) model using a transfer learning technique. To test the robustness of the proposed PDTL model, we analyze the influence of the number of observation points at different locations, streambed heterogeneity, and observation noise levels on the mean squared error MSE values between the observed and predicted temperature fields. Results indicate that the PDTL model significantly outperforms the DNN model in scenarios with scarce training data, and the MSE values decrease with increasing observation points for both PDTL and DNN models. Furthermore, increasing streambed heterogeneity and observation noise levels raises the MSE values of the PDTL and DNN models, with the PDTL model exhibiting greater robustness than the DNN model, highlighting its potential for practical applications in riparian zone management.