CORE: Household-Learned Adaptive Thresholds for Pump Scheduling in Intermittent Water Supply Systems

Abstract

Pump scheduling for household water tanks under intermittent water supply (IWS) almost universally relies on fixed trigger levels chosen by a designer — constants that cannot reflect what any specific household actually needs at any specific hour. We propose CORE (Causal Online Reserve Estimator): a per-household, per-hour-of-day adaptive threshold derived from each household’s own historical mean and standard deviation, re-estimated causally each night with no forecasting model and no anomaly classifier. Evaluated against two published baselines on 100 households over 90 days — a synthetic demand simulation (pySIMDEUM) and a real smart-meter dataset (DAIAD, Alicante, Spain; CC-BY-SA 4.0) — CORE outperforms Variable Trigger Levels (VTL; Alvisi & Franchini 2016) on both scarcity prevention (p < 0.006) and pump cycles (p < 0.000001) on both datasets. Against the Adaptive Seasonal Autoregressive model (ASAR; Chen & Boccelli 2018), CORE achieves practical equivalence in scarcity prevention (Two One-Sided Tests equivalence test, p < 0.001; absolute gap 0.40 events/household/year; Cohen’s d = 0.35) and negligible difference in pump cycles (Cohen’s d = 0.16), at 42× lower computation (1.01 ms vs. 42.5 ms per household). A 7-hour circular smoothing closes the remaining pump-cycle gap without autoregressive fitting. CORE is deployable on Arduino-class microcontrollers, requiring only three arithmetic operations per household per night.