Leakage-Controlled Multi-Horizon Forecasting and Rolling-Horizon MILP Dispatch for an Islanded PV–Battery Microgrid with Secondary Wind Generation

Forecast-driven dispatch studies for renewable-storage microgrids can be weakened by feature leakage, non-receding dispatch windows, and inconsistent planned/realized evaluation scopes. This article presents a reproducible workflow-oriented low-code audit for an islanded photovoltaic (PV)-battery benchmark with secondary wind generation using public NASA Prediction Of Worldwide Energy Resources (POWER) weather and New York Independent System Operator (NYISO) load proxy data. The main direct multi-horizon forecast uses only calendar, lag, and rolling features; contemporaneous meteorological variables are reserved for a weather-informed diagnostic case. A mixed-integer linear programming (MILP) dispatch implemented with PuLP and COIN-OR Branch-and-Cut (CBC) enforces binary charge/discharge exclusivity, terminal battery state-of-charge (SOC) control, and true rolling-horizon execution in which each 24 h solve implements only the first action. All dispatch comparisons are filtered to the same 1694 complete forecast origins. On this common horizon, actual load totals 817.204 MWh and renewable availability totals 730.291 MWh. Relative to a rule-based baseline, the MILP reduces realized load shedding from 9.201 to 6.474 MWh and the normalized weighted objective from 10.769 to 8.203 million scenario-cost units. Weight, battery-capacity, renewable-penetration, and throughput-penalty analyses, together with bootstrap uncertainty intervals, show where this conclusion is robust and where it is benchmark-dependent. The contribution is an auditable evidence chain connecting public data, leakage-controlled forecasts, rolling-horizon dispatch, realized recourse accounting, generated figures/tables, and supplementary quality records; it is not a field validation or proprietary microgrid deployment study.