Energy-saving scheduling for multiple water intake pumping stations in water treatment plants based on personalized federated deep reinforcement learning

Abstract

Urban water treatment plants are among the largest energy consumers in municipal infrastructure, imposing significant economic burdens on their operators. This study employs a data-driven personalized federated learning-based multi-agent attention deep reinforcement learning (PFL-MAADRL) algorithm to address the intake scheduling problem of three water intake pumping stations in urban water treatment plants. Personalized federated learning (PFL) is combined with long short-term memory (LSTM) modeling to create environment models for water plants, focusing on energy consumption, reservoir levels, and mainline pressure. The average accuracies of PFL-based LSTM (PFL-LSTM) models are 0.012, 0.002, and 0.002 higher than those of the LSTM model in the three water plants. Evaluation metrics were established to quantify the effectiveness of each pumping station's energy-efficient scheduling, considering constraints such as reservoir water levels and mainline pressure. The results indicate that the proposed algorithm performs robustly under uncertainties, achieving a maximum energy consumption reduction of 10.6% compared to other benchmark methods.