Large-ensemble generation based on mesoscale ensemble prediction system using machine learning

With increasing awareness regarding environmental issues, the adoption of renewable energy, particularly solar photovoltaics (PVs), is advancing rapidly worldwide. However, the output of PV systems fluctuates depending on weather conditions. Therefore, accurate PV power forecasting based on global horizontal irradiance (GHI) is necessary to ensure the stability of power systems. In general, a PV system outputs forecasts based on numerical weather predictions (NWPs), which are then used for day-ahead forecasting. In this study, the authors focus on ensemble predictions based on NWPs and develop a machine-learning-based large-ensemble generation method to improve the forecast accuracy of the GHI. A mesoscale ensemble-prediction system (MEPS) provided by the Japan Meteorological Agency is used, based on which large-ensemble predictions are generated. The proposed method comprises a Machine-Learning-based Large-Ensemble Generation (ML-LEG) model and a ensemble expansion method based on the ML-LEG model. This ensemble expansion method apply the ML-LEG model to the 21 results from the MEPS to generate 420 members large-ensemble GHI predictions. A control run of the MEPS is used as the input for the ML-LEG model, and the output is the generated ensemble members. The MEPS comprises 20 members, excluding the control-run; hence, the ML-LEG model provides outputs that are 20 times the input. Replacing the ML-LEG input by perturbed weather forecasts allows large-ensemble predictions to be generated. In this study, the MEPS is used as input for the ML-LEG model. Consequently, 420-member ensemble predictions are generated, and the ensemble means reduced the forecast error by 10.3% and 4.0% compared with the MSM and MEPS on dates with large forecast errors.