Comparison of Solar Irradiance between WRF simulation and Deep Learning: Case study in Nishi Akisawa, Japan

Abstract

A This paper presents a study of evaluation for the reliability of seven-days direct normal irradiance (DNI) and diffuse irradiance (DIF) forecasts which provided by Weather Research and Forecasting (WRF) mesoscale model using six-hourly interval 0.5°x0.5° input dataset obtained from National Oceanic and Atmospheric Administration – National Operational Model Archive and Distribution System (NOMADS) website. 3 km spatial resolution was used to estimate seven-days simulation starting from 1st to 7th January 2017 for comparison purpose. Long short-term memory (LSTM) algorithm has been applied to make future prediction. The one-hourly values input dataset of LSTM obtained from Nishi Akisawa mega solar website is consist of one-year data of solar irradiance starting from 1st January to 31st December 2016. This dataset divided into training datasets (89%) and testing datasets(11%) where the testing datasets values was used to make future prediction of solar irradiance. The result shows the error of root mean square of LSTM algorithm is 129 W/m2 higher compare to 101 W/m2 from the WRF model for seven-days prediction but the result of this study proposed using LSTM algorithm for future prediction of solar irradiance or others parameter of weather.