Application of a wavelet denoising-based LSTM-Transformer model for water quality prediction at river cross-sections

This study proposed a hybrid Long Short-Term Memory （LSTM）-Transformer model integrated with wavelet denoising for water quality prediction. Using hourly monitoring data （water temperature， turbidity， pH， conductivity， and dissolved oxygen） collected from two municipally controlled river cross-sections in South China from 2021 to 2024， the discrete wavelet transform was first applied for noise reduction. Subsequently， a predictive model combining LSTM and Transformer architectures was constructed. Experimental results demonstrated that the proposed model achieved outstanding performance in predicting dissolved oxygen （DO） concentrations for the next four hours at both sites （Site 1： coefficient of determination （R²）=0.8015， mean absolute error （MAE）=0.5169 mg/L， root mean square error （RMSE）=0.8494 mg/L； Site 2： R²=0.8873， MAE=0.4456 mg/L， RMSE=0.7143 mg/L）， significantly outperforming standalone LSTM and Transformer models （the R² of the proposed model increased by 5.7%， while MAE and RMSE decreased by 20.2% and 10.4%， respectively）.Furthermore， the SHAP interpretability method was employed for feature importance analysis and global impact interpretation， revealing that the key water quality factors influencing DO and their complex nonlinear relationships exhibited significant site-specific heterogeneity. This underscores the necessity of incorporating specific environmental contexts （e.g.， geographical features， hydrological conditions， and pollution source distribution） for mechanistic interpretation. The findings of this study provide an effective and interpretable technical reference for high-precision real-time prediction and intelligent management of regional river water quality.