In the main islands of Japan including Hokkaido islands, there are nine power companies controlled power grid. Power supply has been performed by thermal power plants and hydropower plants. In the current situation, capacities of renewable energy is smaller than the thermal power plants and hydropower plants. Large penetration of PV power systems in Japan has been accelerated by an introduction of feed-in-tariff of 2012 after the Fukushima nuclear disaster. After 2016, liberalization of retail electricity sales will start, so more and more demand for GHI forecasts or PV power generation forecasts will be required in an electric power system and market.
In Japan, thermal power plants have been mainly used to adjust the power demand supply balance. To control the thermal power plants efficiently and economically, the PV power forecasts would be useful tools. However, forecast data of GHI values is sure to have forecast errors derived from uncertainty of the model schemes and meteorological observation errors as an input data. In particular, large forecast errors of GHI or PV power generation for a wide area would be caused severe risks (i.e., power outage). If the PV power production is overestimated, actual solar irradiance is weaker than expected. Thus, additional operations of thermal power plants or reserve capacity would be necessary. If the PV power production is underestimated, suppression of PV power generation exceeded power demand would be required.
Japan Meteorological Agency (JMA) developed and operated two NWPs to take measure against natural disasters. First one is a mesoscale model (called by MSM) with horizontal grid spacing of 5km. MSM performed 39hour forecast at eight times a day. Therefore, the MSM forecasts would be used as one day ahead or target day's forecast. Second one is a local forecast model (called by LFM) with horizontal grid spacing of 2km. This finer resolution model performed 9 hours forecasts at every hour. In this study, pyranometers observed at Japan Meteorological Agency and geostationary satellite data are used as model validation datasets. In order to validate GHI forecasts for a wide area (regional validation), spatially-averaged data are prepared for both observations and forecasts. For the definition of outlier event (large forecast errors), we used a parameter D, GHI forecast errors normalized by extraterrestrial solar irradiance for one day ahead forecast data (12 LST (03UTC), Japan local standard time initialization time). Here, if D=0.3 is given as a reference value, several outlier events (about 5%-10% for a year) are selected. These selected events were analyzed in this study.
For a target area, we selected the eastern Japan area (50Hz, including Tokyo, Tohoku and Hokkaido electric power companies). In this area, snow falls are often found in western coastal area of both Tohoku and Hokkaido areas in winter. Heavy rain falls has been caused by typhoons and intermittent rain in rainy seasons. Eastern part of Tohoku area are often found lower clouds (stratus clouds, cumulus clouds) are often found in summer season. Although large forecast errors are found in one day ahead forecast (12 LST initialization time forecasts of one day ahead), GHI forecast errors are reduced in intra-day forecasts (06 LST initialization time forecasts of target day) in several cases.
Finally, we want to try to forecast outlier events of GHI forecast using NWP. In order to improve GHI forecasts obtained from NWP, microphysics and/or radiation processes should be improved. However, it would take many times and efforts to develop physics processes in NWP. Therefore, we want to grasp a meteorological signal; whether GHI forecast would be outlier or not in advance (at least one day ahead). First of all, in order to understand outlier events of GHI forecasts, case studies of outlier events would be useful information.