Solar generation forecasts
Balancing operations on Tensor Cloud depend on accurate generation forecasts that scale across many assets. Tensor Cloud forecasts the generation of solar PV systems up to 14 days ahead, and you can download forecasts up to 13 days ahead. Tensor Cloud also generates consolidated forecasts per balancing group (BG) for assets with solar PV systems, batteries, and loads.
How to get started with forecasts
To get next-day generation forecasts for an asset, make sure that:
- The balancing and trading feature is enabled on the asset (status is Operation).
- The asset's COD is in the past, or at most 14 days ahead.
- You have uploaded at least 50 days of historical generation data for the asset.
- For meter data, the right location ID is entered in the asset settings, so that Tensor Cloud can match the uploaded data to the asset.
Once steps 1 and 2 are met, Tensor Cloud automatically creates generation forecasts in 30-minute time slots, ten times a day. You can download them on the Forecasting page under Tasks in the main side navigation. These forecasts are based on physical simulation. Steps 3 and 4 are required for AI-based forecasts: Tensor Cloud trains an AI model for the asset once at least 50 days of historical generation data are available, and uses it when it is more accurate than the physical simulation.
Tensor Cloud creates forecasts for any asset that meets steps 1 and 2, even without historical generation data. We still strongly recommend uploading historical generation data for each asset: Tensor Cloud uses it to train an AI model for that asset, which produces more accurate forecasts.
If you schedule the balancing and trading feature with a future start date (status: Planning), Tensor Cloud starts creating forecasts 5 days before the scheduled start, and, if enough historical generation data is available, starts training the asset's AI model 14 days before it. Forecasts are therefore available from the first day of operation.
Downloading forecasts
The Forecasting page has two tabs:
- Daily lists the latest forecast for each generation balancing group, with the time it was created. Choose how many days the download covers (1, 3, 7, or 13 days) with the Forecast length selector at the top right, then click the download button of a balancing group.
- Day-ahead forecasts combines the day-ahead forecasts for a period into one file. Choose the Target (the entire workspace, selected balancing groups, or selected assets), the Start date, and the number of Days (up to 31), then click Download. For each day in the period, the file uses the 07:00 forecast created the previous morning, so forecasts are available up to the next day once that day's 07:00 forecast has been created.
Forecast data format
Forecasts are downloaded as CSV files with the following columns, in this order:
| Column name | Description | Unit |
|---|---|---|
Datetime | Start of the forecasted 30-minute time slot | Date and time in Japan Standard Time with UTC offset, for example 2023-05-11 05:00:00+09:00 |
Total (kWh) | Forecasted total generation of all assets | kWh |
Name of the asset (kWh) | Forecasted generation of each asset | kWh |
The forecast data format does not state the start and end of each forecast. Because Japanese electricity markets work in 30-minute increments, each forecast covers the 30 minutes that follow its timestamp. For example, a forecast with the timestamp 2023-05-11 05:00:00+09:00 covers 05:00 to 05:30 on that day.
The order of assets in the forecast CSV files is not guaranteed. Design any downstream systems to handle assets in any order.
Forecasting methodology
Accurate forecasts require large amounts of data and a good model of the physical processes behind electricity generation.
Tensor Cloud combines machine learning with physical simulations to generate forecasts. The physical simulations are the same ones used for long-term asset simulations on Tensor Cloud, and they are based on the asset information you provide when adding new assets. They use a digital twin of your asset to estimate generation, much like engineering software such as PVsyst. Machine learning, on the other hand, trains an AI model on historical generation and weather data and uses that model to generate forecasts.
If an asset has little or no historical generation data, Tensor Cloud uses a forecast based on physical simulation. If there is enough historical data, Tensor Cloud trains a custom AI model for the asset and tries to use it for forecasts.
Each time the AI model is trained, it is evaluated against the physical simulation forecast, and it is only used when it is more accurate. Tensor Cloud trains an AI model only for assets with at least 50 days of historical generation data.
For details on the forecasting methodology, see the technology section.
Model re-training
Tensor Cloud checks every night whether each asset's AI model needs retraining. It retrains the model once at least 7 days of new historical generation data are available, or when the asset's plant configuration changes. Otherwise retraining is skipped and the existing model stays in use.