Medium and Long-term Runoff Forecasting Based on Two-stage Decomposition and Interpretable Machine Learning

To improve the prediction accuracy of monthly runoff, this paper proposes a prediction model based on a two-stage decomposition strategy. First, the original runoff sequence is decomposed into trend terms, seasonal terms and residual terms through the Seasonal-trend decomposition using Loess（STL）, and then the residual items with strong randomness are further decomposed through the Variational Mode Decomposition（VMD） to eliminate noise. Three machine learning models, namely, Long-Short Term Memory（LSTM） networks, Convolutional Neural Network（CNN） and Support Vector Regression（SVR）, are used to predict each component one by one, and the monthly runoff prediction result is a linear set of the predicted values of each component. Taking Shimen Station in Lishui Basin as the research object, indexes such as Mean Absolute Error（MAE）、Mean absolute percentage error（MAPE） and Nash Efficiency（NSE） were selected to comprehensively evaluate the model prediction accuracy, and combined with the SHAP（Shapley Additive explanations） interpretable machine learning method to explore the input characteristics in the optimal model contribution to runoff prediction results. The result shows that the overall prediction accuracy of LSTM and CNN is better than that of the SVR model, but the change in prediction accuracy caused by the difference in the model structure is smaller than that caused by the difference in input items; the contribution of each decomposition component in the optimal model STL-VMD-LSTM to the prediction results is better than that of other input characteristics.