A Comparison of Sampling Techniques for the Regional Forecast of Photovoltaic Power in Japan

The objective of this study is to investigate the use of sampling techniques and upscaling, to make regional forecasts of PV power. Sample sets of a group of 143 PV systems installed in the Kanto region in Japan were selected. Forecasts of PV power generation for the samples were done using their past power generation data, weather forecast data and support vector regression. After that, the result was upscaled to represent regional power generation, regarded here as the power generation of all 143 PV systems. The forecasts were done one-day-ahead in an hourly fashion for 1 year. Two theoretical sampling sizes, 18 and 39 PV systems, and two sampling techniques were used. The sampling sizes were based on different values for the expected standard deviation and error margin of the forecasts of the samples. Using 39 PV systems, similar median for the regional forecast error were achieved regardless the sampling technique. Namely, an rmse of 0.0724 kWh/kWrated with simple random sampling and of 0.0720 kWh/kWrated with stratified sampling were obtained. These values were less than 3% higher than the reference regional forecast error, indicating the suitability of sampling to forecast regional PV power generation.