Application of BP neural network model based on runoff characteristics in medium and long term inflow forecasting — A case study of Yuqiao Reservoir in Tianjin City

Medium and long-term runoff forecasting is the key to the implementation of effective water resources scheduling and scientific management in the basin. The screening of forecasting factors is of great significance for improving the accuracy of forecasting. Yuqiao Reservoir in Tianjin City is selected as the forecast object, and the measured runoff process is analyzed. Based on the runoff characteristics, the dry and wet seasons are divided. The dry season is divided into two periods: from November to following February and from March to June. The wet season is from July to October. The influencing factors affecting the cross-section flow process of Yuqiao Reservoir were determined. The BP neural network model is used for segmented forecasting, and the overall annual runoff was forecasted and compared. The monthly data from 1999 to 2020 are used for training, and the overall results from 2021 to 2023 are used for verification. The results show that the R² of the segmented forecast by dividing the dry season and the wet season is 0.31 higher than that of the whole year, the mean absolute percentage error (MAPE) is optimized by 25.41%, and the relative error(RE) is reduced by 19.32%. Compared with the annual calculation, the annual relative errors of the segmented calculation and forecast from 2021 to 2023 are increased by 24.96%, 16.30% and 6.38%, respectively. The results of segmented forecast based on runoff characteristics in dry and wet seasons are better than those of annual forecasting, which can provide data basis for fine scheduling and scientific management of water resources in the basin.