Output Forecasting for Multiple Geographically Distributed PVs Without Meteorological Data

Photovoltaic (PV) output forecasting often uses meteorological and historical PV data, including cloud imagery and weather conditions. Access to such data can be limited for numerous dispersed PVs, particularly in remote areas, making accurate forecasting challenging. Recent advancements in distributed PVs and communication technologies, such as smart meters, have facilitated the collection of time-series data from numerous dispersed PV installations. This development has spurred research into new forecasting models that utilize these data to forecast the PV output across multiple locations. One notable technique is the optical flow algorithm, which estimates and forecasts PV power generation transitions by converting PV power generation data from various locations into images. This study introduces a hybrid model that combines optical flow with machine learning using historical PV generation, time, and location data from multiple installations. The proposed model has an 18.4% improvement in the mean absolute error (MAE) over traditional models that depend on weather data. It also exhibits a 5.8% improvement in MAE and a 10.8% improvement in the continuous ranking probability score compared to the optical flow alone.