Leveraging LSTM-based neuro-evolution for enhanced real-time control in urban drainage systems

Real-time control (RTC) of urban drainage systems helps reduce flooding and combined sewer overflow (CSO) spills, thereby alleviating the pressures of climate change and urbanization. In recent years, neural network-based methods for control policy formulation—such as deep reinforcement learning and neuro-evolution—have shown promise in RTC. Previous studies have focused on improving the performance of these methods by exploring aspects such as state variable selection, reward function formulation, and control frequency adjustment. However, the impact of neural network architecture remains underexplored. This study investigates the performance of Long Short-Term Memory (LSTM) as a policy network in neuro-evolution for RTC of urban drainage systems. The simulation results from the Astlingen benchmarking network show that LSTM-based policies reduce CSO volume by 9.8 %–10.1 % compared to passive control, outperforming Multi-Layer Perceptron-based policies, which achieve reductions of 7.0 %–9.0 % during 2000–2009. For LSTM-based policies, memory information contributes positively to decision-making, while zero-initialized memory leads to a 10.0 %–21.0 % increase in CSO volume, reflecting deteriorating control. Additionally, LSTM-based policies using a single tank level as input offer an efficient alternative in RTC, achieving greater improvements over BC than the global control strategy based on equal filling degree. These improvements also surpass those achieved by LSTM-based policies using multivariate input versus single-variable input. This study reveals the potential of LSTM-based neuro-evolution in RTC of urban drainage systems, supporting the intelligent management of water infrastructure.