Prediction of CODMn concentration in lakes based on spatiotemporal feature screening and interpretable learning methods - A study of Changdang Lake, China

•The paper proposes a methodology for a hybrid model that can capture complex non-linear relationships and help understand the interactions between input and output features.•The model overcomes the limitation of traditional neural network models in predicting across different spaces.•The proposed model is an intelligent water quality prediction model that can provide decision support for predicting CODMn concentration and managing watershed risks in the Changdang Lake basin, with good application prospects and practical value. The organic pollution of lake water can cause a tremendous threat to the water ecosystem and human health. The CODMn is one of the crucial indicators of lake water quality and is commonly utilized to gauge the extent of organic pollution in lake water. Therefore, this paper selected CODMn as the research object and used the water quality monitoring data of Changdang Lake in China and its upstream and downstream to predict the CODMn concentration in the lake. In order to study the spatial relationship between the lake and upstream and downstream water quality, reflect the joint action of multiple water quality factors in prediction and the interaction between different feature factors. This study combined the XGBoost feature filtering algorithm, maximum mutual information coefficient (MIC), and improved recurrent neural network (GRU) and proposes a hybrid model called XGB-MIC-GRU. The model first used XGBoost to screen and extract the relative importance of water quality characteristics and used the Shapley addition extension (SHAP) method to explain XGBoost feature extraction. Then, the correlation between the lake and the upstream and downstream water quality is calculated through MIC analysis. Finally, the selected water quality factor characteristics and spatial characteristics are input into the GRU model for prediction. The experimental results showed that water temperature, total phosphorus, and total nitrogen are the most important to CODMn, and the upstream US1 and downstream DS1 and DS2 stations are the most closely related to the concentration of CODMn in the lake. By comparing the prediction effect of the model in different time steps, the best 16-time steps related data were selected to predict the value of the next time. MAE, RMSE, and R2 of the model are 0.10, 0.13, and 0.96, respectively. The model has better prediction accuracy and correlation error than the traditional SVR and GPR. The proposed mixed model can accur