[2K1-GS-10] AI application: Weather / Fluids

[2K1-GS-10-03] Investigation of a method for correcting systematic errors in solar radiation forecasting using deep learning

〇Shinichiro KANZAKI¹, Satoshi MIYAZAKI¹, Yuta KUNO¹, Koji YAMAGUCHI¹ (1. Japan Weather Association)

Keywords:Weather, Solar radiation forecast, Neural network, High resolution

Forecasting the amount of electricity generated by photovoltaics power has become important in recent years, as it has replaced a growing share of fossil fuel-based electricity generation, which is under pressure to be reduced due to global warming. The impact of occasional significant errors in a day or two days ahead solar radiation forecasting is substantial for operations of electrical grid. To address this issue, the Japan Weather Association is carrying out technological development under contract with the New Energy and Industrial Technology Development Organization (NEDO). One of the technological developments to mitigate this problem is to improve forecast accuracy by integrating forecast values from multiple meteorological organizations. However, the inadequate spatial resolution of global numerical weather prediction models is one of the reasons for the large errors in integrated forecasts. To solve this problem, we report on the results of an attempt to implement systematic error correction based on the spatial distribution of solar radiation and to increase the resolution by interposing the forecast values of overseas global numerical forecast models into a neural network using U-net and other methods.