16:15 〜 16:30
[MZZ40-04] Regional Global Horizontal Irradiance Forecast Using Global-Ensemble Prediction System
キーワード：全天日射量、全球アンサンブル予報システム (GEPS)、領域気象モデル (WRF)、太陽光発電
The use of photovoltaic power has increased. Because the electricity provided the solar energy fluctuates depending on the weather, weather forecasting of global horizontal irradiance (GHI) is an important technology for energy management. Especially, one- to several-day forecasts are important in planning for an appropriate reserve capacity. The predicted GHI could include forecast errors (uncertainty) and large forecast errors sometimes occur. The large forecast error causes blackouts, electric power surpluses, and an imbalance risk to a contracted power supply in the spot market (the day-ahead market). To avoid those problems, control power sources such as thermal power plants have been used. In 2022, the cost of control power in Japan was estimated to be 80 billion JPY (METI, 2022). Reducing the forecast error would help reduce the cost of this control power.
Numerical Weather Prediction (NWP) has been widely used to provide GHI forecasts. Since NWP is calculated from a deterministic approach, only one value is predicted. However, NWP includes an error in the boundary condition and initial values, which could grow forecast errors due to chaotic behaviour. To resolve this NWP problem, ensemble prediction systems (EPS) which add appropriate perturbations to the NWP have been used. Japan Meteorological Agency (JMA) provides two EPSs: the Meso-Ensemble Prediction System (MEPS) and the Global-Ensemble Prediction System (GEPS). MEPS generates 21 member forecasts, and it provided the up to 39 hour-ahead forecasts every six hours (initial time = 03, 09, 15, 21 JST (Japan standard time; 09 JST = 00 UTC)). And GEPS generates 51 member forecasts, and it provided the up to 11 day-ahead forecasts every 12 hours (initial time = 09, 21 JST). In the spot market, electricity power of the next day is contracted until 10 JST. The forecast hour of MEPS is a little short to use for the spot market. GEPS would help reduce forecast error and uncertainty. However, few studies used GEPS for GHI forecasting.
Here, we evaluate the reliability and predictability of GHI using GEPS. Because GEPS does not include GHI, downscaling method with the Weather Research and Forecasting (WRF) ver. 4.2.2 have conducted for large forecast error days. The study area is the area of power supplied by Tokyo electric power company holdings which is the largest PV generation area in Japan. We show the uncertainty of GHI and discuss the reasons for the large forecast error.
The mean value of the GHI predicted by a regional climate model developed by the Japan Weather Association (JWA) and the GHI provided by the European Centre for Medium-Range Weather Forecasts (ECMWF) were used as a benchmark for this study. And, observed GHI and estimated GHI using sunshine duration at JWA meteorological stations were used as the estimated GHI in the study area. 8 large forecast error days in 2021 were selected using the benchmark and estimated GHI. 51 ensemble simulations, which used each ensemble member of GEPS as the initial and boundary conditions were conducted for each day. The ensemble simulations are initialized at 21 JST and the model is run for 50 h. The results between 27 h and 50 h are used as the forecast values. For example, to forecast GHI on 3rd May 2021, the initial time of the model is 21 JST 1st May 2021.
As a result, all of each ensemble simulation for the 3 of 8 days overestimated GHI. Most of the ensemble members underestimated the low-level cloud fraction compared to the images from Himawari-8 for the visible and infrared bands. It was suggested that GEPS had a dry bias in the large forecast error days. The simulations for the small forecast error days under the weather conditions which were similar to the large forecast error days will be conducted as future work.
Numerical Weather Prediction (NWP) has been widely used to provide GHI forecasts. Since NWP is calculated from a deterministic approach, only one value is predicted. However, NWP includes an error in the boundary condition and initial values, which could grow forecast errors due to chaotic behaviour. To resolve this NWP problem, ensemble prediction systems (EPS) which add appropriate perturbations to the NWP have been used. Japan Meteorological Agency (JMA) provides two EPSs: the Meso-Ensemble Prediction System (MEPS) and the Global-Ensemble Prediction System (GEPS). MEPS generates 21 member forecasts, and it provided the up to 39 hour-ahead forecasts every six hours (initial time = 03, 09, 15, 21 JST (Japan standard time; 09 JST = 00 UTC)). And GEPS generates 51 member forecasts, and it provided the up to 11 day-ahead forecasts every 12 hours (initial time = 09, 21 JST). In the spot market, electricity power of the next day is contracted until 10 JST. The forecast hour of MEPS is a little short to use for the spot market. GEPS would help reduce forecast error and uncertainty. However, few studies used GEPS for GHI forecasting.
Here, we evaluate the reliability and predictability of GHI using GEPS. Because GEPS does not include GHI, downscaling method with the Weather Research and Forecasting (WRF) ver. 4.2.2 have conducted for large forecast error days. The study area is the area of power supplied by Tokyo electric power company holdings which is the largest PV generation area in Japan. We show the uncertainty of GHI and discuss the reasons for the large forecast error.
The mean value of the GHI predicted by a regional climate model developed by the Japan Weather Association (JWA) and the GHI provided by the European Centre for Medium-Range Weather Forecasts (ECMWF) were used as a benchmark for this study. And, observed GHI and estimated GHI using sunshine duration at JWA meteorological stations were used as the estimated GHI in the study area. 8 large forecast error days in 2021 were selected using the benchmark and estimated GHI. 51 ensemble simulations, which used each ensemble member of GEPS as the initial and boundary conditions were conducted for each day. The ensemble simulations are initialized at 21 JST and the model is run for 50 h. The results between 27 h and 50 h are used as the forecast values. For example, to forecast GHI on 3rd May 2021, the initial time of the model is 21 JST 1st May 2021.
As a result, all of each ensemble simulation for the 3 of 8 days overestimated GHI. Most of the ensemble members underestimated the low-level cloud fraction compared to the images from Himawari-8 for the visible and infrared bands. It was suggested that GEPS had a dry bias in the large forecast error days. The simulations for the small forecast error days under the weather conditions which were similar to the large forecast error days will be conducted as future work.