The new SolarAnywhere^® Probabilistic Forecast enhances risk management

PV forecasts are important for energy scheduling and grid stability. Traditional methods for near-term forecasts involve blending an ensemble of Numerical Weather Predictions (NWPs) into a single deterministic output, which is favored because it is simple to implement in operations. However, deterministic forecasts do not convey any information about marginal risks or the propensity for rapidly changing weather patterns. Additionally, with the advent of machine learning techniques, some deterministic forecasting approaches can be less accurate because they do not leverage extensive feature lists to adjust estimates.

To address these issues, Clean Power Research^® has developed a SolarAnywhere^® Probabilistic Forecast that can provide P1 – P99 values for day-ahead horizons. This capability allows end users to better manage financial risk in non-symmetric cost models, especially for PV assets in complex climate zones . The SolarAnywhere Probabilistic Forecast is available today as a trial for all SolarAnywhere Forecast licensed customers.

Figure 1 below displays the probabilistic range of irradiance (GHI) at any given hour, rather than a single value. For some days, the uncertainties have a larger spread that can impact decision making in day-ahead markets.

Probabilistic forecasts are particularly useful on cloudy days, where the range of irradiance is larger. As seen in the example in Figure 2, peak GHI has a P10 to P90 of about 210 W/m² to 380 W/m². This type of information helps day-ahead traders manage the downside risk of very low power output.

On the other hand, sunny days with near clear-sky conditions have a much narrower probabilistic distribution. In the example shown in Figure 3, the peak GHI P10 to P90 only varies from about 920 W/m² to 975 W/m², which gives operators a higher level of confidence in the forecast.

Training and validation

The SolarAnywhere Probabilistic Forecast was created using a probabilistic mixture-of-experts model architecture. To train the model and develop the probability curves, features from several satellite and model-based sources were ingested and the negative log-likelihood loss between predicted irradiance and ground-truth irradiance measurements were minimized.

The total training dataset covers three years of measurements from high-quality meteorological stations in various climate zones in North America. The SolarAnywhere Probabilistic Forecast was shown to outperform the Global Forecast System (GFS) benchmark on all accuracy metrics over a multi-year evaluation period using SolarAnywhere hindcasts. Figure 4 below summarizes the critical performance metrics, with a full validation paper available upon request.

Disclaimer: Summary accuracy metrics are not necessarily reflective of forecast performance at a specific location. For site-specific accuracy, we recommend using historical forecasts (hindcasts) for analysis.

The SolarAnywhere Probabilistic Forecast supports day-ahead operations with coverage up to 40 hours in the future. Full details are available in the support center. You can also refer to the SolarAnywhere Postman Collection for API details.

Try it out!

To access the Probabilistic Forecast, become a partner, or inquire about other SolarAnywhere data services, please contact support@solaranywhere.com.