Intraday Ground Horizontal Irradiance Forecasting Using LSTM

Accurate forecasting of solar irradiance at the surface has been identified as one of the important issues for the operation of power grids with large penetration of solar energy. Current mainstream solar forecasting for GHI (ground horizontal irradiance) includes the intra-hour forecasting with a lead time of a few minutes to one hour, which heavily relies on extrapolation of cloud motion; and the inter-day forecasting with a lead time of >24 hours, relying on operational numerical weather prediction (NWP). A particular challenge for solar forecasting is the forecast with a lead time of 1-8 hours, so-called intra-day forecast. We here investigate a machine learning based on long short-term memory (LSTM) model to predict GHI with a 1-2 hour lead time. Four years (2010 - 2013) of 1-minute observed GHI by the National Renewable Energy Laboratory (NREL) at Golden (latitude 39.74N and longitude 105.18W) are used to train the model. We then used observed GHI in another year (2014) at the same site to test the model performance. The results show that the LSTM model has a good prediction performance with Root-Mean-Square-Error (RMSE) of 55.9 Wm^-2 and R-square of 0.96 for the 1-hour prediction. For 2-hour prediction, the RMSE increases to 101.1 Wm^-2 and R-square is reduced to 0.86. We also found the LSTM model predicts better results in fall and winter than in spring and summer. The sensitivities of model performance to different sets of the LSTM parameters are also discussed.