By Dr Conaill Soraghan, Data & Digitalisation Team Lead
Machine Learning (ML) is a disruptive digital technology that, if deployed effectively in the wind industry, would enable predictive maintenance, automate blade defect detection, improve the accuracy of production forecasts, and many more valuable innovations. However, it’s a technology that’s massively underutilised in the wind sector compared to other data-heavy industries, such as health or finance.
This is a two-part blog examining why this is the case, and what we at ORE Catapult are doing to help the industry unlock the value of this promising technology.
ML is a branch of artificial intelligence, geared towards large datasets. It’s a technology that finds and exploits patterns in data for purposes such as making predictions or clustering related data.
There are three major recognised categories of ML, represented in the spider chart above .
Supervised learning is arguably the most popular paradigm for ML. It is easy to understand and relatively simple to implement. It involves feeding data with labels into an algorithm that can learn how to predict the label. Typical objectives are regression to predict a continuous variable or classification for discrete sets. For example, if you have a stack of wind turbine blade images with good labels specifying which blades have defects of which kind and with what severity, it would be possible to build a supervised model that could automatically classify the damage of previously unseen images of blades.
Unsupervised learning does not require any labels and focuses more on presenting the underlying structure or patterns in the data for a human to interpret. Typical objectives include clustering similar data together or reducing the dimension of big datasets with many sensors or variables. If you were to pour SCADA data into an unsupervised learning algorithm, it could reveal operative states of the system or recommend which combination of sensors is the best candidate for predicting a continuous variable, such as gearbox temperature.
Reinforcement learning uses reward mechanisms to drive learning. It requires generating a signal that rewards the algorithm for good behaviours and propagating this reward signal backwards from the end goal to the initial actions you take, which may be hundreds if not thousands of steps apart. Over time, the learning algorithm will make fewer mistakes and learn how best to achieve a predetermined goal. Robotics has a very promising future in the wind industry as a technology that will remove the need for dangerous and expensive human intervention. Reinforcement learning can be deployed to help a crawler familiarise itself with the task of walking along the outside of a wind turbine blade with complex aerofoil geometry.
Here are some questions to ask of data from the wind domain, mapped to appropriate ML techniques. Hopefully, this can act as a starting point for pilot demonstrations for those keen to explore how ML can unlock value. These are all areas we’ve been experimenting with at ORE Catapult.
|Paradigm||Objective||Technique||Wind Industry Applications|
|K-Means or DB Scan
|Cluster a single turbine SCADA to determine operative states to break down reasons for lost production.
Cluster the SCADA signals between the start and end of a turbine shutdown to help classify its severity.
|Principal Component Analysis (PCA)||Apply PCA on one turbine and explore the principal components to reveal how SCADA channels are related. Plot the principal components to identify operative states and seek outliers.
Then try PCA on multiple wind turbines in the same farm. Visualising the principal components will reveal if the turbines group in any interesting ways.
Compare PCA of real data to simulated twinned turbine data to validate the twin.
|K-Nearest Neighbours (KNN) or Decision Trees||Blade damage classification
Predicting the next alarm in a sequence
Predicting a stop type/classification.
|Linear, quadratic, logistic regression||Predict the temperature of a gearbox from other SCADA signals in that turbine. This can learn healthy operation then be used for anomaly detection.
Virtual sensors which predict the output of a foundation strain gauge from the others in case one fails.
|Neural Network||Predict the power output of a target turbine using anemometer readings from all wind turbines in the farm. This will act as an advanced power curve that can be used for power available readings or lost production analysis.|
|Genetic algorithms or the
|Teach a robot to walk along the outside of a wind turbine blade with complex aerofoil geometry.
The next and final part of this blog will build on this context and explore why ML approaches to address problems in the wind sector are not very common. Furthermore, it will explain what we at ORE Catapult are doing to help the industry unlock the value of this promising technology.