A Graph-Augmented Spatio-Temporal Forecasting Framework with a Hybrid Coati–Osprey Synergistic Optimization Algorithm (COSOA) for Energy-Aware, Flood-Resilient Real-Time Control of Urban Drainage Networks under Rainfall Uncertainty

Abstract

Decentralised and adaptive operation of urban drainage networks (UDNs) is increasingly required to contain pluvial flooding and pumping-energy costs under intensifying and uncertain rainfall. Existing UDN studies have largely addressed layout design or post-event leakage diagnosis in isolation, while the closed loop that links short-horizon hydraulic forecasting to energy-aware pump/gate control under uncertainty remains under-explored. This paper proposes an integrated framework that couples a Graph-Augmented Spatio-Temporal Forecaster (GASTF) with a novel Hybrid Coati–Osprey Synergistic Optimization Algorithm (COSOA) for the real-time control (RTC) of UDNs. The UDN is encoded as a weighted graph; the GASTF predicts node water levels from lagged levels, graphaggregated neighbour states and rainfall, and supplies these forecasts to a robust multi-objective RTC formulation that jointly minimises flood volume and pumping energy while maximising hydraulic reliability across a rainfall ensemble. COSOA fuses the structured group-hunting exploration of the Coati Optimization Algorithm with the plunge-and-carry exploitation of the Osprey Optimization Algorithm, governed by an adaptive synergy factor and a memetic elite-refinement operator. On eight 30-dimensional benchmark functions COSOA attains the best mean rank (2.12) among nine optimizers. On a SWMM-calibrated RTC problem, COSOA reduces storm flood volume by 94.9% (from 33,163 m³ under passive operation to 1,702 m³) and raises hydraulic reliability from 47.7% to 97.8%, achieving the best mean robust objective (J = 0.0046) and the lowest single-run objective among all baselines, with statistically significant gains over the Osprey, genetic and search-and-rescue optimizers (Wilcoxon p < 0.01). Graph augmentation lowers forecasting RMSE by 5.1% over a non-graph neural baseline (R² = 0.961). The results indicate that the COSOA-driven framework is an effective and robust tool for energy-aware, flood-resilient UDN operation