Intra-day Forecast of Ground Horizontal Irradiance Using Long Short-term Memory Network (LSTM)

ournal of the Meteorological Society of Japan

article

Xiuhong CHEN¹  Xianglei HUANG¹  Yifan CAI²  Haoming SHEN³  Jiayue LU⁴ 

[1] Department of Climate and Space Sciences and Engineering, the University of Michigan;School of Software Engineering, the Shanghai Jiao Tong University;Department of Industrial and Operations Engineering, the University of Michigan;Department of Computer Science, University of Southern California

关键词: atmospheric radiation;    solar energy forecast;    recurrent neural network;   

DOI  :  10.2151/jmsj.2020-048

来源: Meteorological Society of Japan

【 摘 要 】

Accurate forecast of global horizontal irradiance (GHI) is one of the key issues for power grid managements with large penetration of solar energy. A challenge for solar forecasting is to forecast the solar irradiance with a lead time of 1–8 hours, here termed as intra-day forecast. This study investigated an algorithm using a long short-term memory (LSTM) model to predict the GHI in 1–8 hours. The LSTM model has been applied before for inter-day (> 24 hours) solar forecast but never for the intra-day forecast. Four years (2010–2013) of observations by the National Renewable Energy Laboratory (NREL) at Golden, Colorado were used to train the model. Observations in 2014 at the same site were used to test the model performance. According to the results, for a 1–4 hour lead time, the LSTM-based model can make predictions of GHIs with root-mean-square-errors (RMSE) ranging from 77 to 143 W m −2 , and normalized RMSEs around 18.4–33.0 %. With five-minute inputs, the forecast skill of LSTM with respect to smart persistence model is 0.34–0.42, better than random forest forecast (0.27) and the numerical weather forecast (−0.40) made by the Weather Research and Forecasting (WRF) model. The performance levels off beyond 4-hour lead time. The model performs better in fall and winter than in spring and summer, and better under clear-sky conditions than under cloudy conditions. Using adjacent information from the reanalysis as extra inputs can further improve the forecast performance.

【 授权许可】

Unknown