PREMIUM PV POWER FORECAST MODEL

Introduction

This page explains how the Premium PV Power Forecast Model works, what data it uses, and what the API returns. Premium PV is a site-trained, probabilistic power forecast for operating PV power plants, delivered through the Solcast API with global coverage from now to 14 days ahead.

Premium PV is managed by DNV. We quality-control your measurement data, configure and run the training, review the result before it goes live, and retrain the model periodically to capture ongoing improvements. The same Solcast API then serves the trained forecast.

1. Model overview

The Solcast Premium PV Power Forecast Model forecasts power for operational solar projects. It is designed for operators whose forecast accuracy affects penalties, imbalance settlement, day-ahead bidding or dispatch coordination.

The effectiveness of Premium PV is maximised when enough production history is available to train the model. For a new PV power plant, DNV can initially configure a physics-based forecast from the plant specifications, then train the full Premium PV model once enough production data has accumulated.

The Solcast Advanced PV model estimates plant behaviour from configured specifications, Premium PV also learns from measured generation history. This allows the model to reflect site-specific behaviour, including observed clipping and shading effects, that may not be fully represented by the plant specifications alone.  

Learn more about differences between Premium and Advanced PV models here.

Each Premium asset is identified by a unique resource ID. DNV creates your assets in the Solcast API as part of onboarding and provides the resource ID for each one. Forecasts are then retrieved by making a request to the Solcast API with that resource ID.

DNV makes the forecast available after training has been completed and reviewed.

2. Input data and model chain

Premium PV combines three classes of input: your plant's production history, Solcast's satellite-derived irradiance and weather data, and a site-specific weather ensemble built from multiple numerical weather prediction (NWP) sources.

• Solcast data inputs. The model uses Solcast's irradiance parameters (GHI, DNI, DHI and GTI) together with satellite cloud tracking and forecasting, ambient parameters including air temperature and wind speed, satellite-derived albedo, and a dynamic site-specific snow-soiling loss derived from irradiance, temperature and precipitation. PM2.5 and PM10 particulate concentration data is available through the API to support soiling analysis.
• Weather ensemble. Rather than depending on a single weather model, Premium ingests a wide range of global and regional NWP sources, downscales them to your site, and applies statistical bias correction against your site's own record. Using many sources reduces sensitivity to the biases of any one of them.
• Model chain. DNV fits multiple model types to your site’s production history and selects the best-performing approach based on measured performance. The model represents forecast uncertainty through P-levels, which correspond to specified probabilities of exceedance. A DNV forecasting specialist reviews the training results before the model goes live.

‍

3. Training on your production data

• What we need. One year of recent generation history is recommended to capture seasonal changes in plant performance. Shorter records can be used, but may not achieve the same forecast accuracy. For a new plant, or where insufficient production data is available, DNV can initially run a physics-based forecast from the plant specifications and train the full Premium PV model once more data has accumulated. Historical data provided for training should be at hourly resolution or better.
• What "trained" means. Premium PV is trained on your plant’s generation history, matched against the weather conditions at the site. That lets the forecast reflect site-specific behaviour such as clipping, shading, curtailment patterns and seasonal soiling, rather than relying only on plant specifications.
• Model maintenance. DNV reviews and updates trained models as modelling improvements become available. Tell DNV about material changes to the plant, including repowering, capacity changes or major equipment replacement, so that the model configuration can be reviewed.
• Curtailment in training data. If curtailment or outages are present in the historical data, they need to be marked separately from normal generation. Otherwise, the model may treat reduced output as normal site performance. The measurement template includes availability, curtailment and power fields so these periods can be identified during training. If availability and curtailment cannot be separated in your data, let us know during onboarding.

4. Probabilistic outputs and how to read them

Alongside the central forecast, the Max plan returns values at P90, P75, P25 and P10.

Premium PV uses the probability-of-exceedance convention:

• P90 is the level that generation has a 90% probability of exceeding. It is the lower, more conservative estimate.
• P10 is the level that generation has a 10% probability of exceeding. It is the higher, less conservative estimate.

Therefore, Premium PV’s P90 value is lower than its P10 value. We state the convention explicitly because P-level labels can be confused with percentiles which have the opposite order.

Note for Advanced PV : Advanced PV’s pv_power_advanced10 and pv_power_advanced90 are high and low cloudiness scenario values, respectively, rather than probability-of-exceedance values, and their numbering follows the opposite ordering.

5. Data handling

• What we use your data for. We use your measurement and plant data to produce forecasts for your sites, train your site models and analyse forecast performance. Access is limited to authorised DNV personnel who need the data to support and improve your service.
• Where your data is stored and processed. Unless different data-residency terms are agreed in your contract, customer data is stored and processed in Sydney, Australia. DNV uses Amazon Web Services in the ap-southeast-2 region to store and back up the data securely.
• How your data is protected. Customer data is isolated from other customers’ data using physical and logical controls, including access controls, permission boundaries and separation between customer environments. Access is restricted to the customer’s authorised users and the DNV personnel required to support the service.
• Data retention. Customer data is retained for the contract duration or for a maximum of five years, unless different terms are agreed in your contract or required under applicable DNV policies. For detailed retention requirements, DNV can refer specific questions to the appropriate QHSSE team.
• Regional arrangements. Where agreed in the contract, DNV can support analysis in a specified country or region. Commercial arrangements, availability and pricing for regional storage or processing are confirmed as part of the product and contract process.

6. Specifications

Plans

Product Specifications

7. Operating conditions: availability and curtailment

Premium PV can account for planned outages and curtailment when these limits are known ahead of time. Submit availability and curtailment schedules against the asset’s resource ID through the API or Web Toolkit. The API can then return both the unconstrained forecast and the constrained forecast, showing expected production before and after the submitted limits are applied.

Availability and curtailment scheduling are available on the Pro and Max plans.

8. Appendix A: asset specification parameters

The fields below describe the plant. Required fields must be supplied before the model can run. Optional fields can be left blank; where they are not supplied, the model uses estimated values. Supplying optional fields reduces the assumptions required by the physics-based forecast, particularly for a new plant while production data is accumulating.

Expand each section to view the parameter definitions and how each is used.

9. Appendix B: training measurement fields

Supply your generation history in the format below. One row per interval, per asset.

Expand each section to view the parameter definitions and how each is used.

10. Appendix C: output data parameters

Expand each section to view the parameter definitions and how each is used.

11. Appendix D: API request parameters