LU Li-feng, XU Chao-liang, MENG Zi-qiang, FU You, HUI Qun-zhi
Focuses on a water supply system at a pump station in Hunan as the research object. Through the two-equation characteristic line method, a mathematical model of the pump station's transient process is constructed. The complete characteristic curves of the pump, the closing time of the pipeline valve, the maximum pressure envelope, the minimum pressure envelope, and the pressure point after the valve are analyzed using Deep Forest, BP neural network, and polynomial regression models. The analysis yields the following conclusions: A reasonable combination of machine learning models and transient flow models can accurately predict and analyze the maximum and minimum pressure envelopes, and the pressure point after the valve in a pressurized water delivery system, providing reliable support for the digital twin optimization scheduling of the water delivery system. Regarding the regression prediction of the complete characteristic curves of the pump, the accuracy distribution trend is as follows: BP neural network>Deep Forest>polynomial regression method. Since the pump does not generally experience reverse rotation during the transition process, and the water flows forward (in the fourth quadrant, reverse dissipation area x∈(4.71, 6.28)), focusing on the regression results of the first, second, and third quadrants shows that the Deep Forest method performs better in regressing the complete characteristic curves of the pump. In terms of regression prediction for maximum/minimum pressure envelopes and valve pressure, the overall expected trend under the Deep Forest model is better than that of the BP neural network. When pressure oscillations caused by the water hammer effect are large and changes are dramatic, the regression results of the BP neural network may produce certain prediction discrepancies.