An improved Back Propagation Neural Network framework and its application in the automatic calibration of Storm Water Management Model for an urban river watershed

Abstract

The use of the Storm Water Management Model (SWMM) to simulate flows in urban river watersheds necessitates the proper calibration of the various parameters involved in the process. Back Propagation Neural Network (BPNN) is often used to establish relationship between two sets of multivariate variables, such as parameters and simulation results of SWMM. The aim of this study is to establish an improved BPNN to calibrate SWMM. It was found that when using gauged flow data obtained from the urban river management system as calibration data, only using BPNN was not sufficient. An improved BPNN framework was proposed with integrating Principal Component Analysis (PCA) and Genetic Algorithm (GA) process, abbreviated as PCA-GA-BPNN. It was proved to be effective for calibration. The BPNN combined with GA process made 90 % of the predicted parameters within reasonable range, which was only 8 % using BPNN alone. The PCA process reduced the training time up to 64 %. Using a hydrograph of 196 h, compared with the nondominated sorting genetic algorithm (NSGA), PCA-GA-BPNN training time can be reduced from 18,142 s down to 4.5 s. Nash efficiency coefficients (NSE) of hydrographs fitting was 0.75. Including more rainfall events data in calibration achieved better fitting than including more gauging station data.