A Proposed Methodological Framework for Data-Driven Deep Learning Surrogate Modelling of Non-Revenue Water Prediction and Operational Control in Small-Sized Municipal Water Distribution Systems

Non-Revenue Water (NRW) remains a major operational challenge for water utilities, particularly in small-sized distribution systems where limited monitoring and data availability constrain the development of efficient loss-management strategies. This study proposes a data-driven surrogate modelling framework based on deep learning (DL) algorithms to support NRW prediction within a telemetry-enabled operational control strategy. The proposed methodology framework integrates hydraulic modelling, pressure-dependent leakage representation, District Metered Areas (DMAs), Pressure-Reducing Valves (PRVs), and telemetry for the analysis of alternative operating conditions in the drinking water distribution system of Arboletes, Antioquia, Colombia. Real water losses were represented using the Germanopoulos pressure-dependent formulation with a pressure exponent of γ = 1, consistent with the predominantly high-density polyethylene network. Three operational scenarios were evaluated through 24 h extended-period simulations: current operation without sectorisation, DMA-based sectorisation, and DMA-based sectorisation combined with PRV-based pressure control. The resulting hydraulic simulations were used to develop and evaluate deep learning surrogate models—including Convolutional Neural Network (CNN), Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), and Multilayer Perceptron (MLP) architectures—for NRW prediction. The DL models successfully reproduced the temporal dynamics of NRW across all scenarios, achieving RMSE values of 0.551%, 0.482%, and 0.223% for the no-DMA, DMA, and DMA+PRV scenarios, respectively. Both the Gated Recurrent Unit (GRU) and Multilayer Perceptron (MLP) architectures demonstrated their applicability to NRW prediction under the simulated operating conditions, as part of an exploratory study. Hydraulic results showed that DMA implementation alone produced limited reductions in water losses, whereas the combined DMA+PRV strategy reduced the average daily NRW from 43.04% to 30.99% (with the daily volumetric NRW decreasing from 42.39% to 27.18%). The results demonstrate the potential of deep learning-based surrogate modelling to accurately reproduce NRW dynamics and provide a data-driven tool that, once trained, can reproduce NRW dynamics without repeatedly executing the hydraulic model, supporting the operational analysis of drinking water distribution systems. These results are based on a single 24 h simulation scenario for one case-study network and should be regarded as proof of concept rather than generalised to other seasons or system types without further validation.