Korea Institute of Energy Research Develops AI-Based Predictive Maintenance Models for Offshore Wind Power
Prevent costly downtime for offshore wind turbines by identifying potential component failures before they occur
Use MATLAB to develop machine learning and deep learning algorithms that use existing sensor data to predict possible failures
- Development time cut in half
- 90%+ prediction accuracy achieved
- Aggressive deadline met
“Despite having little previous experience with AI, within a limited budget and a tight deadline, we completed a diagnostics model in MATLAB capable of detecting wind turbine component failure with over 90% accuracy.”Jung Chul Choi, Korea Institute of Energy Research
As part of its Renewable Energy 3020 Implementation Plan, the Korean government has set a goal of producing 20% of the country’s energy from renewable sources by 2030. As a result, the number of offshore power generation installations is expected to increase dramatically. Offshore facilities are more difficult to maintain than those onshore, increasing the need for effective predictive maintenance systems.
Korea Institute of Energy Research (KIER) engineers have developed a diagnostics model that uses AI to predict the structural load on individual turbine components so that preventive action can be taken before a failure occurs. Developed in MATLAB®, this model incorporates machine learning and deep learning algorithms and uses data collected from existing sensors in accordance with IEC 61400-13.
“For offshore power, the cost of sensor installation and operation is often an obstacle to predictive maintenance,” says Jung Chul Choi, senior researcher at KIER. “Our MATLAB based AI model saves cost by enabling us to diagnose the condition of those components with a small number of sensors.”
For predictive maintenance of wind turbines, KIER needed to estimate bending moments and stress on the turbine blades and other key components from sensor data. A typical turbine consists of approximately 8,000 components, and installing new sensors on each component would have been prohibitively expensive. KIER engineers needed to use data such as windspeed, turbine rotational speed, and power generated captured from existing sensors at various operating times. This would involve analyzing thousands of signals collected at a rate of 1 GB daily across multiple turbines.
Although the team lacked extensive experience with machine learning, they needed to rapidly evaluate a variety of machine learning and deep learning approaches to identify the best approach for the available data. In addition, they had to meet a tight deadline—they were tasked with delivering a dashboard for monitoring turbine operations within six months.
Working in MATLAB, the team preprocessed sensor data by removing outliers and performing smoothing and data reduction—for example, by eliminating sensor measurements taken when a turbine was stationary.
Using Statistics and Machine Learning Toolbox™ and Curve Fitting Toolbox™, the team implemented a number of machine learning algorithms, including algorithms based on regularized linear regression, polynomial curve fitting, and decision trees. They evaluated each algorithm’s ability to predict key turbine component load values, such as bending moment of the blade, shaft tilt moment, and shaft yaw moment. Next, the team used Deep Learning Toolbox™ to implement and train an artificial neural network (ANN) and evaluated it in the same way.
The engineers created a graphical interface for the algorithms using MATLAB App Designer. They packaged this interface together with the algorithms into an application using MATLAB Compiler™. They shared the application with KIER colleagues via a dashboard used to monitor turbine operations.
KIER uses a cumulative damage model to calculate remaining useful life (RUL) and determine when maintenance is necessary. They plan to install the MATLAB based algorithms in a health management system for turbines in KIER’s offshore wind facility on Jeju Island.
- Development time cut in half. “If we had used an open-source alternative such as Python, it would have taken more time to preprocess data, develop sound diagnostic algorithms, and create a dashboard,” says Choi. “We estimate that with MATLAB, we reduced development time by 50% or more compared with such an alternative.”
- 90%+ prediction accuracy achieved. “The model we developed in MATLAB has a predictive accuracy of over 90% across six major parts,” says Choi. “With this level of accuracy, we can develop wind turbine predictive maintenance systems that save millions of dollars annually per turbine by diagnosing failures in advance.”
- Aggressive deadline met. “MATLAB enabled us to process large amounts of data in a variety of file formats,” notes Choi. “With MATLAB we were able to analyze the correlation of multiple signals, reduce the data, and complete the algorithm development within a tight, six-month project deadline.”