INTRA-DAY FORECASTS OF PV POWER WITH NUMERICAL WEATHER PREDICTION DATA AND MACHINE LEARNING IN KYUSHU, JAPAN

摘要

In this study we verify the typical levels of error of intra-day forecasts of regional PV power generation for the Kyushu island in Japan. To do that, we applied a method to yield continuous intra-day forecasts with a variable lead time (2h to 24h ahead of time). The method requires the use of numerical weather prediction, NWP data, recently measured data, and machine learning to model the relation between input and output variables. Forecast of hourly PV power generation were done for 3 years considering a scenario of high penetration of PV power in Kyushu. The PV power forecasts were based on insolation measurements and NWP data yielded by the meso-scale model of the Japan Meteorological Agency, JMA. The results show that the annual root mean square error normalized by the PV installed capacity, NRMSE, varied from 0.0297 GWh/GW_(peak) to 0.0318 GWh/GW_(peak) for hour-ahead forecasts (calculated with values from 5h to 20h of each day of the target period). The same error varied from 0.0753 GWh/GW_(peak) to 0.0820 GWh/GW_(peak) for 24 hours ahead forecasts. Furthermore, the results also show for how long the use of recent clearness index values, or its persistence, is effective to reduce forecast errors. For all 3 years of results, past clearness index use was effective for forecasts of up to 6 hours-ahead. For a lead time longer than that its use did not have any meaningful effect on the forecasts.