Application of Online Flow Measurement System Combined with Neural Network Model in Flow Prediction

To address the accuracy limitations of conventional multiple regression models for stage-discharge relationships at the Balijiang Hydrological Station in the lower Yangtze River, this study employs an online flow measurement system to collect 210 field datasets (stage range: 6.30-21.14 m, discharge range: 15, 300-73, 800 m³/s, float velocity range: 0.44-2.37 m/s). A systematic comparison was conducted to evaluate the performance of various neural network models including BP, CNN, ELM, LSTM, BiLSTM, RBF, and SVM in discharge simulation. Results demonstrate that all neural network models significantly outperform the multiple regression model in both training and validation phases. During training, the models achieved reductions of 314-414 m³/s in MAE, 1.20%-1.64% in MAPE, and 439-603 m³/s in RMSE, while improving NSE by 0.29%-0.36%. In validation, performance improvements included MAE reduction of 81-304 m³/s, MAPE decrease of 0.52%-2.40%, RMSE reduction of 80-221 m³/s, and NSE enhancement of 0.23%-0.57%. Among the baseline models, the BP neural network delivered the most comprehensive performance with training phase MAE, MAPE, RMSE, and NSE values of 468 m³/s, 2.02%, 636 m³/s, and 0.998 respectively, and validation phase values of 686 m³/s, 3.32%, 936 m³/s, and 0.9914 respectively. To further enhance performance, genetic algorithm (GA) and particle swarm optimization (PSO) were introduced to optimize the initial weights of the BP network. The PSO-BP model demonstrated superior performance, achieving training phase MAE, MAPE, RMSE, and NSE values of 437 m³/s, 1.88%, 581 m³/s, and 0.9988 respectively, and validation phase values of 612 m³/s, 2.92%, 825 m³/s, and 0.993 respectively. This research confirms that the PSO-BP hybrid model effectively improves the accuracy of discharge prediction at Balijiang Hydrological Station, providing a more reliable modeling approach for similar hydrological applications.