【Energy Company】Weather Prediction AI Model Building
Solar radiation forecasting model using deep learning with Japan Meteorological Agency API
Background
Photovoltaic power producers are obligated to submit forecasts of power generation to power transmission and distribution business operators. If there is a difference between their forecasts and actual results, they are required to pay a fine. In order to reduce it, solar radiation, which is directly related to the amount of electricity generated, must be forecasted precisely. For this reason, solar power generation companies generally utilize solar radiation forecasting APIs provided by major companies in order to make correct forecasts. We were asked to build a solar radiation forecasting model with the aim of minimizing costs associated with the amount of fines and API use.
Detail and Effect
We conducted the project from researching papers and built a deep learning solar radiation forecasting model using the JMA API. We conducted all processes, including requirement definition and software selection. LSTM was selected as the forecast model. As a result, we succeeded in developing a model with the same accuracy as the commercial service that had been used until then. The ability to obtain forecast values without relying on commercial services has resulted in annual cost savings of several tens of millions of yen.
Technology
- Survey of papers and selection of software
- Identification and Selection of Characteristics
- Identification and selection of prediction models
- LSTM to construct forecasting models
- Data infrastructure construction
Language, Software
- Python
- AWS(Amazon Redshift)