Synthetic experience augmentation: A two-regime framework for extending LSTM extrapolation in rapid sand filter clogging prediction under water utility consolidation

Water utility consolidation and capacity expansion require estimates of rapid-filter head loss at unfamiliar filtration velocities. We developed synthetic experience augmentation (SEA) to generate training data for a long short-term memory (LSTM) model from operating records without identifying plant-specific bed parameters. Under laminar conditions, the Kozeny-Carman proportionality scales filtration velocity and head loss at the same target time while retaining measured water-quality conditions. Uniform SEA transforms the complete velocity history, whereas Instantaneous SEA transforms only the target-time velocity. We evaluated SEA using 24 months of records from Myoken Water Treatment Plant and three months from Furugō Water Treatment Plant, while withholding measured high-velocity targets from training. Operating histories ending at the target time were used to estimate head loss at that time. Across 30 matched seeds, Instantaneous SEA increased the baseline LSTM out-of-distribution R² from 0.818 to 0.837 at Myoken and from 0.424 to 0.564 at Furugō. The mean paired improvements were +0.019 (p=0.00134) and +0.140 (p=5.14×10^-6), respectively. Uniform SEA and a physics-informed neural network (PINN) were also evaluated over 15 seeds each. Differences between Instantaneous SEA and PINN were significant at Myoken and nonsignificant at Furugō, and the ranking of the SEA variants depended on the plant. Within the generation-factor ranges, deviations from the Ergun reference values remained within approximately 1.2%. SEA uses existing records to support head-loss estimation under specified velocity and water-quality conditions, providing a basis for assessing operating conditions before consolidation or capacity expansion.