Forecasts
An integral part of running balancing operations on Tensor Cloud is creating accurate, reliable, and scalable asset generation forecasts. We currently provide downloadable next-day forecasts for solar PV assets only. Going forward, we will be expanding our forecasting capabilities to include
- Real-time intraday forecasts
- Creation of OCCTO-compliant XML files
- Automatic or semi-automatic submission of forecasts to OCCTO
- Probabilistic adjustment of forecasts before submission
- Wind asset forecasting
How to get started with forecasts
To get started with next-day generation forecasts for a given asset, you need to ensure the following:
- Your asset status is set to operational
- You have uploaded historical generation data for your asset
- You have entered the right location ID in the settings of your asset
- If your asset is part of a balancing group, you have added your asset to the right balancing group (optional)
Once you have completed step 1, Tensor Cloud automatically creates next-day generation forecasts every day at 06:00 JST. You can download forecasts for each balancing group in the main side bar menu under Balancing > Forecasts. Steps 2 and 3 are required for AI-based forecasts to increase forecast accuracy.
Although Tensor Cloud automatically creates forecasts for any asset set to operational, even without historical generation data, we strongly recommend that you upload historical generation data for each of your assets. This will allow Tensor Cloud to create a customized AI model for each asset and ultimately generate more accurate forecasts.
Forecast data format
You can download day-ahead forecasts in CSV format. The forecast CSV files contain the following columns in order:
| Column name | Description | Unit |
|---|---|---|
datetime | Start of forecasted 30-minute time slot | Timezone-aware ISO 8601 datetime format |
total | Forecasted total generation of all assets | kWh |
| Name of the asset | Forecasted generation of each asset | kWh |
Our forecast data format does not explicitly specify start and end of each forecast. Instead, as the Japanese electricity markets operate in 30-minute increments, we assume that each forecast covers the following 30 minutes. For example, a forecast with a timestamp of 2023-05-11 05:00:00+09:00 will cover 05:00 to 05:30 on that day.
Forecasting methodology
Generating accurate forecasts is a complex task that requires large amounts of data as well as a deep understanding of the underlying physical processes of electricity generation.
Tensor Cloud uses a combination of machine learning and physical simulations to generate forecasts. Physical simulations are what we also use for the long-term economic asset simulation on Tensor Cloud. They are based on the asset information you provide when adding new assets to Tensor Cloud. Physical simulations leverage a digital twin of your asset to estimate electricity generation and work in a similar fashion to simulations created with engineering software such as PVSyst. Machine learning prediction on the other hand uses historical generation and weather data to train an AI model that can then be used to generate forecasts.
For a given asset, if there is no or not enough historical generation data available, Tensor Cloud will default to a physical simulation-based forecast. If there is enough historical generation data available, Tensor Cloud will train a custom AI model for the asset and attempt to use it to generate forecasts.
The AI model is continuously evaluated against the physical simulation-based forecast and will only be used if it creates forecasts with a higher accuracy. From our experience, at least 50 days of historical generation data are needed to create an AI model that is more accurate than a physical simulation-based forecast.
On a high level, the forecasting process on Tensor Cloud for a given asset works as follows:
- Download the latest weather forecast for that specific asset location
- Ingest the weather forecast either into the physical asset simulation model, or into the AI model
- Use the chosen model to generate a forecast for the next day from 00:00 to 23:30
Tensor Cloud uses 11 different types of weather information including 3 solar irradiance indicators to generate forecasts. We get our weather forecasts through a third-party weather data provider, and the underlying raw data is provided by the Japan Meterological Agency (JMA) with an update frequency of 3 hours at a resolution of 5km.
Model re-training
Tensor Cloud regularly retrains the AI model of each asset. The frequency of re-training is based on the amount of historical generation data available: Assets with only very little historical data are re-trained more frequently than those assets with more data, as forecast accuracy would improve only marginally. Minimum re-training frequency is once per month, maximum re-training frequency is daily.