Pump Scheduling Optimization in Urban Water Supply Stations: A Physics‐Informed Multiagent Deep Reinforcement Learning Approach

Abstract

In the urban water supply system, a significant proportion of energy consumption is attributed to the water supply pumping station (WSPS). The conventional manual scheduling method employed by water supply enterprises imposes a considerable economic burden. In this paper, we intend to minimize the energy cost of WSPS by dynamically adjusting the combination of pumps and their operational states while considering the pressure difference of the main pipe and switching times of pump group. Achieving this goal is challenging due to the lack of accurate mechanistic models of pumps, uncertainty in environmental parameters, and temporal coupling constraints in the database. Consequently, a WSPS pump scheduling algorithm based on physics‐informed long short‐term memory (PI‐LSTM) surrogate model and multiagent deep deterministic policy gradient (MADDPG) is proposed. The proposed algorithm operates without prior knowledge of an accurate mechanistic model of the pump units. Combining data‐driven with the physical laws of fluid mechanics improves the prediction accuracy of the model compared to traditional data‐based deep learning models, especially when the amount of data is small. Simulation results based on real‐world trajectories show that the proposed algorithm can reduce energy consumption by 13.38% compared with the original scheduling scheme. This study highlights the potential of integrating physics‐informed deep learning and reinforcement learning to optimize energy consumption in urban water supply systems.