The Characteristics of Forecast Errors of the Global Horizontal Irradiance obtained from a Local Forecast Model

The installation of renewable energy is useful for reducing mass consumption of fossil fuels and for decreasing carbon dioxide. A tremendous amount of photovoltaic (PV) system has been installed in Japan after the summer in 2012. The power production obtained from the PV systems has temporal and/or spatial variability because solar irradiance depends on clouds, aerosols distributions and temperature etc. In order to stabilize the power supply, the forecast of the PV power production using the numerical weather prediction model (NWP) are desired.

Recently, NWP are becoming a useful tool to forecast the PV power production and the global horizontal irradiance (GHI). The Japan meteorological Agency developed a high resolution model, whose horizontal grid spacing is 2km (local forecasting model; LFM), and they started the operational use of the model from August 2012. Before using the model output in the field of the PV power production, forecast errors of the LFM should be investigated. In this study, seasonal and regional validation of the GHI obtained from the LFM are examined using the surface-observed GHI values.

The integration time of the LFM is nine hours and produced the output at each hour. In this study, the forecast data are analyzed during the period from 26 August to 31 December (approximately four months) in 2012. To verify the forecasted GHI data, we use the hourly surface-observed GHI values at the JMA stations (32 stations). We calculated the forecast errors based on the mean bias errors (MBE) and root mean square errors (RMSE) values as indices of efficiency evaluation.

Forecasted GHI values obtained from the LFM are overestimated (positive MBE) in the whole region of Japan from autumn to winter seasons. RMSE of GHI values over the Japan gradually decrease from August to December. Forecast errors normalized by the extraterrestrial solar irradiance ranges -0.2 to +0.4. Differences of forecast errors are not large in forecasts obtained from different initial conditions.

On case study, we also investigated forecast errors of hourly GHI values for various weather conditions. From the time series of the GHI values, the GHI forecasts are well reproduced the observations in cases of a clear sky or deep cloudy conditions (i.e., optically thick clouds). However, in cases of optical thin clouds, the GHI values have relatively large forecast errors.

Compared the cloud distribution of the LFM with that of a mesoscale model (MSM) with a horizontal grid spacing of 5km, we found some case that the LFM can reproduce cloud distribution along the coastline near the mountain range. We also compared with time series of the hourly GHI between the LFM and the MSM at a station, suggesting that hourly GHI values were close to the observation than that of the MSM. However, we also found some cases that the LFM has relatively large forecast errors compared with the MSM.