New improvements & developments – Q4 2024

As we are slowly approaching the end of Q4 this year, it seems like a great time to share some of the latest work our R&D team has been focusing on. Being a global provider of weather and power forecasting, we understand how critical high accuracy is—and improving our forecasting capabilities is always on our daily agenda.

In the last months, we have expanded our NWP (Numerical Weather Prediction) model mix by adding two new European models and have introduced uncertainty flagging across multiple power forecasting setups. We have also developed two new models: a Physical icing model to predict power losses caused by icing and an Injection model that accounts for solar production alongside local consumption.

We have developed a new physical icing model to predict power loss due to icing, even with limited historical data. Currently, we operate four distinct icing models to meet various requirements:

• If you lack icing-related data, we can provide a physical estimation of icing events (yes/no) based on meteorological data and the Makkonen model. Additionally, we offer calibration-based estimates, utilizing data from a nearby wind farm where relevant data is available.
• If you have partial data, we can offer both a physical estimation of icing events (yes/no) and an estimation of physical power loss. These estimates will be calibrated using your available data, in combination with the Makkonen model and meteorological inputs.
• If you have a substantial amount of data, we can develop a machine-learning model specifically tailored to your site, leveraging both your data and meteorological information to optimize icing predictions.

The new power forecast add-on, Uncertainty Flagging, is fully operational and can be integrated into your current setup. We have already implemented uncertainty flags for some of our partners. Along with their standard wind or PV power forecasts, they receive signals indicating potential over- or under-estimation of power production.

Over the past few months, we have conducted multiple backtests to analyze the success rate of the uncertainty flag signals. Using a threshold of 25% potential error for wind power forecasts and 10% for solar, we evaluated both the risks of over- and under-predictions of the portfolio and compared the analysis with realised production data. Here are some of the results:

• Backtest for a small solar park located in Denmark: Our uncertainty flagging for overprediction was successful 67% of the time, and for underprediction, the success rate was 61%.
• Portfolio of 11 solar parks in the Balkan countries: Our uncertainty flagging for overprediction was successful 66% of the time, and for underprediction, the success rate was 71%.
• Portfolio of 40 on & offshore wind farms in the UK with a total capacity of over 1.4 GW: Our uncertainty flagging for overprediction was successful 73% of the time, and for underprediction, the success rate was 71%.

We have developed a new data-driven Injection model to provide a more accurate solar power forecast, considering the self-consumption from local businesses and households. Here are the main features of our Injection forecast:

• Upgraded Machine Learning Techniques: Our model uses advanced machine learning approaches that effectively handle variable weather conditions and seasonal changes, which are vital for accurately predicting solar energy production. 
• Time and Seasonal Considerations: The forecast uses time-based data – such as the time of day, day of the week, and month – to reflect the changing patterns of energy consumption and solar production. By recognizing how these patterns vary throughout the year, our model provides tailored predictions for different periods (e.g., increased consumption and production during the summer months). 
• Enhanced Meteorological Models: Continuously updated meteorological models, along with our in-house models, provide accurate forecasts for cloud cover, temperature, and solar radiation. This helps us make more precise predictions about solar energy production. 
• Aggregated Multiple Model Approach: Instead of relying on a single, centralized model, we segment our forecasting models based on specific consumption behaviours and geographic areas. This approach allows us to create tailored forecasts for each consumer group, ensuring we effectively meet the diverse needs of our clients. 

In recent months, we have conducted an analysis of a small solar portfolio in the Netherlands. The backtest results show an NMAE value for the aggregated forecast ranging from 1.57% to 2.01%.

Would you like to learn more about our latest developments? Contact us to discover how our power forecasting innovations can add value to your trading and operations.