Coupled SWAT and SWT-CNN-LSTM Model to Improve Watershed Streamflow Simulation
chengqing Ren,Jianxia Chang,Xuebin Wang,Chen Niu,Liyuan Wang,Junhao Zhang
DOI: https://doi.org/10.21203/rs.3.rs-5443509/v1
2024-01-01
Abstract:Accurately simulating hydrological conditions is a significant challenge for hydrologists globally, particularly in capturing patterns of high and low streamflow due to the ongoing impacts of climate change and human activities on water resource availability. To address this challenge, this study proposes an innovative integrated approach that combines the Soil and Water Assessment Tool (SWAT), Stationary Wavelet Transform (SWT), and interpretable machine learning models, specifically Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks. The hydrological and meteorological features generated by SWAT were preprocessed using SWT and then used as inputs for the CNN-LSTM model. During the testing period, the SWAT-SWT-CNN-LSTM model achieved R² and NSE values of 0.90 and 0.88, respectively, outperforming other machine learning models (e.g., Support Vector Machine, Random Forest, LSTM) as well as the calibrated SWAT model. Additionally, it effectively reduced the underestimation of high streamflow and the overestimation of low streamflow, with deviations in both categories maintained within 1.3%. From a machine learning perspective, solar radiation and percolation volume are identified as key factors influencing local streamflow, while precipitation shows the highest sensitivity to streamflow variation. Even with a reduced sliding window length, the model's R² and NSE during the training and testing periods remained above 0.88 and 0.80, respectively, demonstrating exceptional stability. Furthermore, the SWAT-CNN-LSTM comparative experiment showed that SWT effectively mitigates the overfitting issue in machine learning. This study highlights that the organic integration of SWAT, SWT, CNN, and LSTM, along with the application of interpretable methods such as Individual Conditional Expectation (ICE), Partial Dependence Plots (PDP), and Shapley Additive Explanations (SHAP), not only enhances model performance significantly but also increases the credibility of machine learning results, paving a valuable new pathway for long-term streamflow simulations in watersheds.