An integrated data decomposition and feature selection–based hybrid machine learning framework for enhanced hourly wastewater inflow forecasting at a wastewater treatment plant in Oahu, Hawaii

Study region The Sand Island Wastewater Treatment Plant (WWTP), the primary wastewater management facility on Oahu, Hawaii. Study focus Wastewater inflow (WI) is a critical variable for efficient WWTP management. This study developed a hybrid framework for hourly WI forecasting at lead times of 1–48 h, integrating Variational Mode Decomposition (VMD) with two feature selection methods—K-Best (KBest) and Recursive Feature Elimination (RFE)—combined with a Cascade Forward Neural Network (CFNN), selected as the base learner after outperforming Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks. Historical WI and hour of day were the only inputs. The resulting VMD-KBest-CFNN and VMD-RFE-CFNN were compared with the standalone CFNN. New hydrogeological insights for the region The hybrid models outperformed the standalone CFNN, with VMD-RFE-CFNN most accurate in validation (most recent 25% of the 2018–2024 record): NSE of 0.9991–0.9422 and RMSE of 0.0129–0.1030 m³ /s across lead times. On the two-year test set (unseen data; May 2024–April 2026), the model retained NSE ≥ 0.9351 and RMSE ≤ 0.1549 m 3 /s up to the 24 h lead time and, unlike the standalone model, preserved accuracy for high-flow events. Under Oahu's separate sewer system and stable subtropical climate, WI is dominated by the anthropogenic diurnal cycle, enabling accurate forecasts from a parsimonious input set. These findings can support forecast-based plant operation, from real-time pump and aeration control to day-ahead planning.