Our machine learning (ML) icing forecasts have guided our clients through two winters, and now comes the third one. With close coordination with our clients and brilliant contributions from our team, we geared up our machine-learning icing models for this winter.
When talking about data-driven icing forecasts, most people would ask the following questions:
- What kind of data are needed?
- What types of icing forecasts can be provided?
- How do you evaluate icing forecasts?
We would like to address these questions point by point in the following.
What kind of data are needed?
There are several nuances regarding data resources and providers, which we can generalize into two following scenarios:
- Scenario 1: SCADA data at the turbine level for a park is required.
- Scenario 2: Data regarding icing events or power loss at a park level.
If either scenario fits your situation, we can always help you with ML icing forecasts. We will ensure that all the data will be analyzed and cleaned before modelling.
If neither scenario suits your needs, we have the capability to provide icing forecasts by using a physical model tailored to your specific requirements. You can read more about this approach in our previous article.
What kind of ML icing forecasts can be provided?
With the available data, machine learning models can be built for different applications. Currently, ConWX provides two main icing forecasts up to 48 hours ahead for each park:
- Icing binary events and their corresponding probability
- Icing power loss expected due to an icing event occurring at the plant.
How do we evaluate ML icing models?
We compare the ML icing forecasts with observational data from our clients based on different metrics depending on different icing forecasts.
ML icing models for icing events
Metrics to be used for icing events evaluation include confusion matrix, recall, precision, accuracy, and F1 score. We mostly use accuracy as a full evaluation for both icing and non-icing events and F1 score for icing-hitting events.
Accuracy score and F1 score improvement of upgraded model based on the original model with 5% and 50% (the higher the value, the better improvement).
ML icing models for icing power loss forecasts
Metrics for evaluating icing power loss forecasts include normalized RMSE (nRMSE) and normalized MAE (nMAE).
nRMSE and nMAE improvement of upgraded models based on the original models with -20% and -15.3% (the lower the value, the better the improvement).
We work closely with our clients for data gathering and model evaluations, thus upgrading our models based on performances and data availability year after year.
Do you own or operate wind parks located in cold regions? Or are you seeking information about power loss due to icing events? Reach out to us! We can schedule a call to talk further about your specific needs.