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 deep transfer learning (DTL) approach is proposed to improve the accuracy of spatiotemporal temperature distribution predictions. The proposed DTL model integrates the physical mechanisms described by an analytical model into the standard Deep Neural Networks (DNN) model using a transfer learning technique. To test the robustness of the proposed DTL model, the influence of the number of observation points at different locations, streambed heterogeneity (𝜎²_lnK =0, 0.2, 0.5, and 1.0), and observation noise levels (𝜎_{𝑁𝑜𝑖𝑠𝑒} =0.025, 0.05, 0.075) on the MSE values between the observed and predicted temperature fields. Results indicate that the DTL model significantly outperforms the DNN model in scenarios with scarce training data, and the mean MSE values decrease with increasing observation points for both DTL and DNN models. The mean MSE values for both the DTL and DNN models approach zero as the number of observation points increases to 200, indicating that both DTL and DNN models perform satisfactorily. Furthermore, increasing 𝜎²_lnK and 𝜎_{𝑁𝑜𝑖𝑠𝑒} raises the mean MSE values of the DTL and DNN models, with the DTL model exhibiting greater robustness than the DNN model, highlighting its potential for practical applications in riparian zone management.