Nonlinear pollution source switching under extreme rainfall revealed by budget-closed graph-based apportionment

Extreme precipitation is intensifying under climate change, yet watershed management frameworks built on annual or seasonal load budgets remain blind to the sub-weekly episodic pulses that can restructure pollutant transport. We present the Budget-Closed Graph-based Source Apportionment (BC-GSA) framework, which couples a graph attention network with an algebraic mass-balance constraint that forces exact closure (ε 0): nonpoint-source (NPS) loads are computed as the deterministic residual of observed flux minus upstream transport and point-source (PS) inventory, eliminating the 5-15% budget leakage typical of unconstrained data-driven models. Applied to the Yiluo River Basin (18,881 km², 2020-2021), BC-GSA revealed that extreme rainfall in July 2021 triggered a rapid, nonlinear reversal of the basin's PS-NPS contribution balance. Basin-wide TP NPS contributions surged from a flood-season mean of 41.5% to 89.2%, with the outlet station reaching 98.7%, and TN rose from 57.5% to 86.5% - the entire shift completing within 48 h. A hydrological forcing-response analysis across 15 independent storm events confirmed a reproducible relationship between hydrological amplification and NPS surge magnitude (r = 0.946, p < 0.001 for TP), with a physical saturation ceiling at 93-99% NPS. Five independent validation lines - spatial regression, C-Q signatures, turbidity proxies, SWAT benchmarking, and bootstrap analysis - confirmed the physical credibility of the residual-based NPS estimates. Ablation experiments further revealed an observed-load anchoring effect: because discharge dominates total flux during extreme events, NPS estimates remain stable (bias ≤ 5 pp) even when water quality predictions fail. These findings argue for replacing static Total Maximum Daily Loads with event-triggered adaptive control protocols tied to real-time hydrological thresholds.