Optimizing Chlorination in Water Distribution Systems via Surrogate-assisted Neuroevolution

TL;DR

This work tackles the challenge of maintaining microbiological safety in large water distribution systems by optimizing chlorine injections through a surrogate-assisted, neuroevolution framework. It combines ESP with NEAT-based prescriptors, trained against a learned surrogate of EPANET via knowledge distillation, and optimized with NSGA-II under a curricular progression of objectives. The approach yields diverse Pareto-optimal policies that outperform PPO baselines and demonstrates that surrogate fine-tuning accelerates exploration and regularization. The results suggest a practical pathway for safer, more efficient chlorination in urban water networks and point to broad applicability in other spatiotemporal control problems. The framework is release-ready and can be extended with longer simulations and additional constraints to further enhance robustness and deployment potential.

Abstract

Ensuring the microbiological safety of large, heterogeneous water distribution systems (WDS) typically requires managing appropriate levels of disinfectant residuals including chlorine. WDS include complex fluid interactions that are nonlinear and noisy, making such maintenance a challenging problem for traditional control algorithms. This paper proposes an evolutionary framework to this problem based on neuroevolution, multi-objective optimization, and surrogate modeling. Neural networks were evolved with NEAT to inject chlorine at strategic locations in the distribution network at select times. NSGA-II was employed to optimize four objectives: minimizing the total amount of chlorine injected, keeping chlorine concentrations homogeneous across the network, ensuring that maximum concentrations did not exceed safe bounds, and distributing the injections regularly over time. Each network was evaluated against a surrogate model, i.e. a neural network trained to emulate EPANET, an industry-level hydraulic WDS simulator that is accurate but infeasible in terms of computational cost to support machine learning. The evolved controllers produced a diverse range of Pareto-optimal policies that could be implemented in practice, outperforming standard reinforcement learning methods such as PPO. The results thus suggest a pathway toward improving urban water systems, and highlight the potential of using evolution with surrogate modeling to optimize complex real-world systems.

Figures (9)

• Figure 1: Topology of the water distribution network utilized in this paper. Red and yellow squares denote sensor and injection nodes, respectively; reservoirs indicate infinite water sources. The topology defines the optimization task and the challenges in it. In this case the large diameter introduces long-term effects and high density nonlinear effects, which the surrogate model needs to learn and the controller needs to optimize.
• Figure 2: The outer loop of the Evolutionary Surrogate-assisted Prescription (ESP) algorithm Francon_Gonzalez_Hodjat_Meyerson_Miikkulainen_Qiu_Shahrzad_2020. The predictor and prescriptor are co-optimized to guide learning of both models synergistically, encouraging innovation and regularizing the resulting policies.
• Figure 3: Comparison of $T_{\phi}$ vs. $S_{\theta}$ vs. true values for $o_{t+1}$/$\hat{o}_{t+1}$. Low error in $S_{\theta}$ predictions despite large variations in concentration suggest that the distillation from $T_{\phi}$ is reliable, and the resulting surrogate model can be used to optimize the controller.
• Figure 4: Pairwise population fronts with (a) two, (b) three, and (c) four objectives. The Pareto front evolves gradually as new objectives are added, and becomes fully formed when all objectives are included.
• Figure 5: NEAT innovation resulting from fine-tuning the surrogate model. The graphs illustrate the performance difference in bound violations between ESP iteration (a) zero to (b) one with fine tuning performed between them. Red lines indicate the level of best solutions after iteration zero. After fine-tuning, a new species (highlighted in yellow) emerges above this level. In this manner, fine-tuning accelerates exploration of the objective landscape.