Machine learning predictive modeling for short term wind power forecast

It is often said that the best way to know the future is to look at the past. By exploring patterns, trends,  and relationships between various instances in historical data sets one aims to derive generalization rules and create a predictive model. Predictive model can be seen as a customized mapping function between a set of input data fields and target variables.
When it comes to large-scale offshore wind power production, a fast and accurate prediction of wind power is critical for optimal schedule and utilization of the energy. Wind power forecasting is closely related to wind speed forecasting, and if deterministic models are either too complex or require too long time for acquiring the results, mining the historical data can become the best way to obtain the required prediction.
Machine learning is used throughout the scientific world for handling the "information flow" and data mining. Developed from the artificial intelligence community, machine learning is a method of data analysis that automates predictive model building. Using algorithms that iteratively learn from data, machine learning allows computers to find hidden insights without explicit instructions on where to look. The iterative aspect of machine learning is important because as models are exposed to new data, they are able to adapt, i.e. to learn from previous computations to produce reliable results. After predictive model is built, learned from data, and validated, it is able to generalize the knowledge and predict the future.
Machine learning models are prone to both underfitting and overfitting problems, therefore data pre-processing is extremely important in order to create the successful predictive model.

For wind energy industry planning and optimization one of the most important prediction horizon is a short time ahead (beyond 12 hours and up to 48 hours ahead), and that is especially useful for wind energy penetration, electricity trading and grid integration. To increase the accuracy of prediction one might select to use local information at the wind turbines instead of numerical weather prediction. In such circumstances, machine learning pattern-based approach is the best way to get the accurate predictions. Even though there is a gap between the requirement of prediction accuracy and current achievements, the rapid progress in machine learning that arose from new big data and deep learning concepts is opening a new horizon for predictive modeling.


Alla Sapronova, Uni Research. Photo: Eivind Senneset


There are number of different models utilizing machine learning that can be suggested for discovering patterns and trends in wind data. For example, a double-module model can be suggested for accurate wind power prediction from 10 minutes to one hour ahead. In “Short time ahead wind power production forecast.” by Sapronova, Meissner, and Mana, a combination of artificial neural network module and kernel module were used to automatically discover the patterns hidden in the historical data of wind speed. Experimental results show that the proposed approach performs better than a single module approach and persistence forecasting. With double-module approach RMS percentage error has been lowered from 14.9 to 5.9 for 10 minutes ahead forecast for total power output. The following table shows the improvement in prediction accuracy achieved with double-module model and machine learning approach:


Prediction horizon:

5 minutes

10 minutes

20 minutes

30 minutes

Single module ANN model error





Double-module model error





Persistence approach





Another example of machine learning application to wind energy forecast has been demonstrated at the NORCOWE WP meeting in Grimstad this year. There, data from Sheringham Shoal wind park was used to train the model for wind speed prediction 10 to 180 minutes ahead.  In addition, NWP data from European Center for Medium-Range Weather Forecasts model, with a grid spacing of 0.7º in latitude and longitude has been used to improve the accuracy of the prediction.
The figure below illustrates the comparison of forecasting errors (RMSE) on validation data for three type of predictive models: machine learning based only (orange), with added  NWP data (yellow), and persistence model (blue).



Results on models validation for 10 to 180 minutes ahead wind prediction. Color codes of the models are the following:

In yellow: machine learning based model with NWP data added; In red: machine learning based model; In blue: persistence model.


For lead time up to about 45 minutes the machine learning based model with power output time series is shown to be more accurate (RMSE for one hour's lead time is 8% maximum rated farm power) and the model with added NWP data is shown to be more accurate for longer lead times (RMSE for 3 hours' lead time was 10% of maximum rated farm power).
As it was presented, the model with added NWP data can operate with fewer input parameters and can therefore be trained faster and with less training data. It was also concluded that NWP data added to energy production time series allows implicit categorization of energy production data.


Text and illustration: Alla Sapronova, Uni Research Computing




15. 02. 2017

NORCOWE presentations at DeepWind'17...


16. 12. 2016

New laser tracker and cameras in place at the Motion Lab...


Calendar/ Upcoming Events