ZAPP: a zero-aware photovoltaic power prediction framework

Abstract

Predicting photovoltaic power generator outputs is not easy, especially when the values are near zero, because even very small errors can lead to large percentage errors. We introduce a new forecasting framework, ZAPP (Zero-Aware Photovoltaic Power Predictor), that handles near-zero errors while maintaining high prediction accuracy. The model does not rely on a single unified regression model but instead attaches two refinement stages to a base regression model. The first stage, called ZeroGate, predicts whether the output should be zero. The second refinement stage, the low-power MAPE reducer, fine-tunes the non-zero outputs near the zero level without significantly affecting scale-dependent regression metrics. The datasets are divided into training, validation, and test sets in chronological order, ensuring that test data is not used during model development, threshold tuning, ensemble weighting, or calibration. Experimental evaluation included base models such as Persistence, Ridge, XGBoost, LSTM, Transformer, and a hybrid Ridge–XGBoost model. Two separate datasets are employed to strengthen the quality of the results. For the hybrid model, sMAPE fell from 60.57% to 11.07% on Dataset 1. Its final MAE, RMSE, and R 2 were 22.37, 60.72, and 0.9637, respectively. On Dataset 2, sMAPE decreased from 63.30% to 11.92%, with MAE = 52.24, RMSE = 144.60, and R 2 = 0.9559. The zero-aware correction mainly reduced percentage errors without weakening the overall regression results.