Two-Stage Feeder Load Forecasting with Behind- the-Meter Rooftop Photovoltaic Using XGBoost

Short-term feeder load forecasting is increasingly complicated by behind-the-meter (BTM) rooftop photovoltaic (PV) systems, which offset local consumption without exporting power to the grid. Under a non-export scheme, the utility observes only the net feeder demand-the residual load after local solar offsets internal consumption-so the measured signal carries a hidden solar component that conventional load-only models capture poorly. This paper proposes a two-stage framework that makes this effect explicit. Stage 1 forecasts rooftop PV from meteorological and temporal features using XGBoost, SVR, LSTM, and persistence; the predicted PV is then an explanatory input to Stage 2, which forecasts the combined feeder load with the same model set, XGBoost as the primary model. An irradiance-based K-Means clustering label, with the cluster count set by consensus of internal criteria, encodes weather-regime conditions. The framework is evaluated on an industrial case study at PLN UP3 Majalaya: 5,880 hourly observations from a 2030 kWp non-export BTM PV system. Two chronological validation schemes-non-overlapping and modified overlap-based-are compared. Under the non-overlapping split, XGBoost attains the best feeder-load accuracy (RMSE 207.27 kW, nRMSE 5.66%), while LSTM leads Stage 1; the overlap-based split yields more optimistic validation results that do not transfer to the test set, confirming that a strictly chronological non-overlapping split is the more reliable basis for assessing generalization.