[Paper] Optimizing Chlorination in Water Distribution Systems via Surrogate-assisted Neuroevolution

Source: arXiv - 2602.07299v1

Overview

Ensuring safe drinking water in sprawling urban networks is a classic “big‑data meets control” problem: chlorine must be dosed precisely enough to kill pathogens, but not so much that it harms users or the environment. The paper Optimizing Chlorination in Water Distribution Systems via Surrogate‑assisted Neuroevolution tackles this challenge by marrying three AI techniques—neuroevolution, multi‑objective optimization, and surrogate modeling—to automatically generate dosing policies that are both effective and computationally tractable.

Key Contributions

• Neuroevolution‑based controller design – Uses NEAT (NeuroEvolution of Augmenting Topologies) to evolve neural networks that decide where and when to inject chlorine.
• Multi‑objective formulation – Simultaneously optimizes four competing goals: (1) minimal total chlorine use, (2) spatial homogeneity of residual chlorine, (3) safety‑bound compliance (no over‑concentration), and (4) temporal regularity of injections.
• Surrogate model for fast evaluation – Trains a lightweight neural network to mimic EPANET, the industry‑standard hydraulic simulator, cutting evaluation time from minutes to milliseconds.
• Pareto‑optimal policy set – Generates a diverse portfolio of dosing strategies, giving operators a menu of trade‑offs rather than a single “one‑size‑fits‑all” solution.
• Empirical superiority over RL baselines – Demonstrates that the evolved controllers outperform Proximal Policy Optimization (PPO) on all objectives across several realistic water‑network scenarios.

Methodology

1. Problem Encoding

   • The water distribution system (WDS) is abstracted as a graph of nodes (junctions, tanks) and pipes.
   • At each decision step the controller outputs a vector of injection rates for a subset of “chlorination stations”.

2. Neuroevolution with NEAT

   • Starts from simple feed‑forward networks and lets evolution add hidden nodes and connections, automatically discovering the right network complexity.
   • Fitness is computed via the four objectives, combined using NSGA‑II’s Pareto ranking to preserve diversity.

3. Surrogate Modeling

   • A separate neural network (the surrogate) is trained on thousands of EPANET simulation runs, learning the mapping from injection actions → resulting chlorine concentrations throughout the network.
   • During evolution, each candidate controller is evaluated on the surrogate, enabling millions of cheap fitness evaluations.

4. Multi‑objective Optimization (NSGA‑II)

   • The population is sorted into non‑dominated fronts; crowding distance preserves a spread of solutions across the objective space.
   • After a fixed number of generations, the best Pareto front is extracted for further validation on the full EPANET simulator.

5. Baseline Comparison

   • A PPO agent is trained on the same surrogate environment, then tested on EPANET.
   • Performance metrics (chlorine usage, concentration variance, safety violations) are compared across methods.

Results & Findings

• The surrogate model achieved an R² ≈ 0.96 against EPANET, confirming its fidelity for evolutionary evaluation.
• The Pareto front revealed clear trade‑offs: policies that aggressively minimize chlorine use tended to increase spatial variance, while those emphasizing homogeneity used slightly more chlorine.
• Across three benchmark city‑scale networks (small, medium, large), the neuroevolution approach consistently outperformed PPO on all four objectives.

Practical Implications

• Plug‑and‑play dosing controllers – Utilities can deploy the evolved neural networks directly on SCADA systems; the inference cost is negligible (sub‑millisecond per decision).
• Decision‑support dashboard – The Pareto front gives operators a “policy picker” UI to choose a strategy that matches current operational priorities (e.g., cost‑saving vs. strict safety).
• Reduced chemical costs & environmental impact – By shaving ~10 % off chlorine usage, municipalities can lower procurement expenses and downstream chlorine by‑products.
• Scalable to other disinfectants – The same framework can be retrained for chloramine, ozone, or UV dosing, simply by swapping the surrogate’s training data.
• Accelerated simulation‑in‑the‑loop design – Surrogate‑assisted evolution sidesteps the need for high‑performance computing clusters, making the workflow feasible for mid‑size water utilities.

Limitations & Future Work

• Surrogate fidelity – While the surrogate performed well on the tested networks, rare hydraulic events (e.g., pipe bursts) may fall outside its training distribution, potentially degrading controller safety.
• Static network topology – The study assumes a fixed pipe layout; real‑world networks evolve (new connections, maintenance closures) and would require periodic retraining.
• Limited real‑world validation – Experiments were confined to simulated environments; field trials are needed to assess robustness against sensor noise and operational constraints.
• Scalability of multi‑objective space – Adding more objectives (e.g., energy consumption of pumps) could increase the complexity of the Pareto front, demanding more sophisticated selection mechanisms.

Future research directions

1. Online learning loops where the surrogate is continuously updated with live sensor data.
2. Hybrid approaches that combine neuroevolution with model‑based MPC for safety guarantees.
3. Extending the methodology to multi‑commodity water quality management (e.g., simultaneous chlorine and fluoride dosing).

Bottom line: By evolving neural controllers on a fast surrogate of a high‑fidelity hydraulic simulator, the authors demonstrate a practical, data‑driven pathway to smarter, greener chlorine dosing—an approach that could be adapted to many other complex infrastructure control problems.

Authors

• Rivaaj Monsia
• Daniel Young
• Olivier Francon
• Risto Miikkulainen

Paper Information

• arXiv ID: 2602.07299v1
• Categories: cs.NE, eess.SY
• Published: February 7, 2026
• PDF: Download PDF