Impact of High PV Penetration on Load Forecasting Accuracy in Smart Grid

Integration of photovoltaic (PV) systems in the smart grid presents important challenges for load forecasting due to the variability and uncertainty introduced by solar generation, especially the behind-the-meter system. This paper examines the effects of different levels of PV penetration (20%, 40%, and 70%) on the accuracy of three different forecasting models: Auto Regressive Integrated Moving Averages (ARIMA), Gradient Boosting (GB), and Long Short-Term Memory (LSTM). The study uses real-time load and solar generation data from the New York ISO zone K to assess how the inclusion of solar power shifts load peaks to later hours of the day and reduces the magnitude of peak demand. Results indicate a consistent decline in forecasting performance across all models as PV penetration increases. At 70% penetration, ARIMA shows increased MAE (60.34) and MAPE (2.99), while LSTM, despite having the lowest baseline errors, also degrades (MAE: 53.18, MAPE: 2.73). GB maintains stable performance at lower levels but deteriorates at higher penetration. The accuracy of all three forecasting models declines significantly. These findings reveal the limitations of existing forecasting methodologies under high PV penetration scenarios. Despite differing approaches, all models fail to maintain acceptable accuracy, underscoring the need for more advanced and adaptive forecasting techniques that can effectively capture the complexities introduced by widespread solar adoption.