Data Quality over Model Complexity: A Leakage-Aware, Horizon-Aware Forecasting Pipeline for Degraded Sensor Streams in a Small-Population Wastewater Treatment Plant

Reliable forecasting of wastewater treatment plant (WWTP) sensor streams is limited by corrupted measurements, nonstationary regimes, and temporally invalid evaluation. These constraints are especially severe in small-population plants with degraded sensors and irregular maintenance. This research presents a leakage-aware, temporally consistent framework that combines consensus anomaly detection, selective reconstruction, multi-resolution feature generation, and horizon-aware validation. It is evaluated on real operational data from a municipal WWTP serving 1,177 population equivalent, with 19 sensor channels and 7,020 hourly observations, across inlet pH, inlet oxidation-reduction potential, effluent ammonium, and effluent nitrate. Leakage-aware data preparation, rather than model complexity, is the dominant contributor to predictive reliability. Reported as a three-seed ensemble, the optimized Long Short-Term Memory architecture attains R² = 0.955 for inlet pH without meteorological inputs and 0.913 for effluent nitrate when ERA5 precipitation and air temperature are included as target-specific exogenous inputs (R² = 0.867 without them). In a separate benchmark without meteorological inputs, recurrent models retain skill across the full 72-hour horizon for all four targets, while tree-based baselines collapse to negative R² beyond 24 hours. These results support long-range stability, rather than single-step accuracy, as the operationally decisive property of a single-site pipeline whose multi-site transfer and closed-loop deployment remain future work.