An integrated framework for prediction and sensitivity analysis of water levels in front of pumping stations

Study region: The South-to-North Water Diversion Eastern Route Project section from the Nansihu-Dongpinghu pumping station cluster. Study focus: An integrated framework for prediction and sensitivity analysis of water levels in front of pumping stations is proposed to obtain more accurate predictive surrogate models and to simplify surrogate model inputs. The results show that among the three different water transport models, the Firefly-Support Vector Machine model has a smaller mean absolute error (<2.38 %), root mean square error (<4.76 %), and mean absolute percentage error (<0.07 %) with higher linear correlation (>0.85). The Firefly-Support Vector Machine model is more suitable for water level prediction than other models. The water level in front of the target pumping station and the t-ahead flow were the most sensitive parameters, and the longer the foresight period, the higher the importance. New hydrological insight for the region: Three water transportation modes are proposed according to the characteristics of regional hydrological connectivity in the long-distance water transportation system. This enables the water level prediction surrogate model to adapt to the complex connectivity of pumping stations and lakes in the region, improving the accuracy of water level prediction. Subsequently, the parameter sensitivity of the water level prediction surrogate model for each water transport mode was also tested. • First time to combine water level prediction and sensitivity analysis for LDWTS. • FA-SVM model has better robustness and higher prediction accuracy. • Sensitivity analysis can simplify the inputs to the prediction surrogate model. • The input simplified prediction model predicts similarly to the original model.