Price forecasts
Background
All operational assets with batteries need price forecasts to optimize battery operation. For this, Tensor Cloud creates 13-day price forecasts for all Japanese grid areas and the system price once per day.
Components and architecture
The Tensor Cloud price forecast service comprises two main components:
- the training service, which trains the forecast models for each zone (and system).
- the prediction service, which loads the models trained and uses them to predict prices for each zone (and system).
Training service
The model training service runs once a week on Sunday at 23 JST. It trains the models that are used in the predictions, based on historical pricing data from JEPX.
Predictor data
Our models predict the electricity price for every 30-minute time slot for every zone, using a selection of available variables which may include:
| Category | Variable | Unit | Time resolution |
|---|---|---|---|
| JEPX market data | Ask volume | kWh | 30min |
| JEPX market data | Bid volume | kWh | 30min |
| JEPX market data | System price | JPY/kWh | 30min |
| JEPX market data | Zonal prices | JPY/kWh | 30min |
| JEPX market data | Daily price summaries | JPY/kWh | 1 day |
| JEPX market data | Intraday price info | JPY/kWh | 30min |
| Macroeconomic indicators | JPY/USD exchange rate | - | 1 day |
| Macroeconomic indicators | Coal price | USD/ton | 1 day |
| Macroeconomic indicators | Gas price | USD/m3 | 1 day |
| Macroeconomic indicators | Price of LNG delivered to Japan | USD/m3 | 1 day |
| Seasonal | Weekday | - | 1 |
| Seasonal | Holiday | - | 1 |
| Weather data | Air temperature at 2 m | °C | 1 hour |
| Weather data | Wind speed | m/s | 1 hour |
| Weather data | Global horizontal irradiance (GHI) | W/m2 | 1 hour |
| Weather data | Diffuse horizontal irradiance (DHI) | W/m2 | 1 hour |
| Weather data | Direct normal irradiance (DNI) | W/m2 | 1 hour |
| Weather data | Relative humidity at 2 m | % | 1 hour |
| Weather data | Dew point temperature at 2 m | °C | 1 hour |
| Weather data | Precipitation | mm | 1 hour |
| Weather data | Cloud cover | % | 1 hour |
| Weather data | Surface pressure | hPa | 1 hour |
| Weather data | Apparent temperature | °C | 1 hour |
We train 2 models for each zone: one model that predicts the prices, and another one that predicts the probability of market prices becoming zero. The probability of zero-price events is also used by the battery optimization to make more robust decisions: if there are several options to charge during forecasted zero-price events, the battery will charge when it is more likely.
Prediction service
The prediction service runs every day at 5am JST, and predicts the price on each zone as well as the system price for the next 13 days.
Price prediction FAQ
Q: For which markets are you creating price predictions?
A: We are currently supporting the JEPX day-ahead market with planned support for intraday and balancing markets. If you have urgent needs around price predictions for these markets, contact us.
Q: How can I download Tensor Cloud price prediction data?
A: Tensor Cloud mainly leverages price predictions internally, to optimize battery charge and discharge behavior, and while price forecasts can be downloaded through the Tensor Cloud UI as CSV files, we encourage using the Tensor API for integration with other systems.
Q: How accurate are your price predictions?
A: This depends on the grid area, time period, and prediction horizon. For the most common use-case, 1-day ahead prediction at 07:00 JST, our price prediction generally achieves RMSE between 2 and 3 JPW/kWh for most areas and the system price. We are constantly improving our price forecast accuracy, especially for predicting zero-price events, which are a major value driver for battery storage.