Evaluating short-term irrigation demand forecasts: From accuracy metrics to forecast-state screening via mass-balance consistency

： Short-term operation of pressurized, on-demand irrigation networks requires flow forecasts assessed for both accuracy and branch–Total mass-balance consistency. Conventional error metrics do not test simultaneous branch and Total predictions as a complete network state. We developed a mass-balance consistency error (MCE)-first screening protocol for 10, 20 and 30 min forecasts in a valve-controlled lateral in northwestern China. Native 2025 telemetry supplied development data and horizon-specific historical-validation Ridge P95 thresholds. Persistence, Ridge, Random Forest and a long short-term memory hybrid generated candidate states. Temporal external testing yielded 10,026, 9994 and 9965 qualified network targets in 2026. Ridge MCE-only acceptance remained near 93.4%, whereas Random Forest acceptance fell from 86.97% to 75.51% despite a 30 min whole-cohort Total-flow Nash–Sutcliffe efficiency of 0.97570. Independent positive-supply evaluation showed lower mean realized errors in Accept, but Review/Reject errors were not consistently monotonic. Ridge large-error discrimination was limited (10 min area under the receiver operating characteristic curve, 0.612), and matched-acceptance error exceeded Persistence. A separate pressure modifier increased recall with substantial false positives and showed no overall baseline advantage. Weighted least-squares projection and physics-informed graph attention network reconstruction remained target-assisted; the graph model’s branch advantage was confined to positive supply. Frozen MCE-first screening exposed branch–Total disagreement overlooked by aggregate accuracy and organized it into evidence for operator review. Its realized-error discrimination depended on the forecast generator and supply regime. The findings support temporal transfer of this structural screen within the tested network, but do not establish pressure adequacy, hydraulic feasibility, spatial transfer or closed-loop scheduling benefit.