Foundation AI partners with Ensemble Energy to identify component failures as much as 6 months in advance of traditional systems with 90% accuracy

Foundation AI is an Artificial Intelligence Solutions Provider. We help organizations process, manage, and leverage their unstructured data to automate labor-intensive tasks, make better data-driven decisions, and drive real business value.

Ensemble Energy is an advanced data analytics company dedicated to improving the way that energy is created. Their Energy.ML platform, developed in partnership with Foundation AI, combines edge machine learning and AI methods with deep energy industry expertise to help renewable energy producers reduce costs, improve efficiency, and increase production.


Goals

  • To configure Ensemble Energy’s Energy.ML platform to predict wind turbine failures earlier than standard condition monitoring systems (CMS).

Approach

  • Used historical SCADA data (measurements of temperature, current, voltage, power, and wind speed) taken in 10-minute increments from fleet of major energy utility company.
  • Applied support vector machines (Gaussian), artificial recurrent neural networks (LSTM), artificial neural networks (AutoEncoder), and gradient boosted decision tree (XGBoost).

Results

  • Predicted pitch bearing failure as far as 6 months in advance with 90% accuracy.
  • Predicted main bearing failure 2 months sooner than CMS with 90% accuracy.
  • Predicted transformer failure 1-2 months in advance.

Background

Clean energy plants lose $27 billion every year due to their reactive maintenance practices. This problem affects wind farm operators most acutely. Wind turbines have multiple inter-connected components which work collectively to generate power. Failure of any of these components can halt the entire machine. To restore the machine, engineers first need to identify the source of the failure (often by drilling holes in the turbine) and then repair or replace the failed component. The size and placement of wind turbines (which are sometimes offshore) make these repairs extremely expensive and time consuming—cranes and other heavy equipment are often necessary to conduct the inspection, maintenance, and repair. The process can take between a few weeks and a couple months, all the while the turbine may be offline.

Challenge

If, using sensor data from these turbines, wind operators can accurately predict when components will fail, they can schedule maintenance and repair before the turbine is forced offline, and procure the necessary equipment, labor, and parts in advance. They can also optimize their spend by performing prescriptive maintenance and repairs on several turbines at the same time.

Ensemble Energy approached Foundation AI to utilize the Extract Learning Platform to help develop a solution that could accurately predict wind turbine component failure in advance.

Solution

Ensemble Energy and Foundation AI partnered with a major utility to leverage its historical SCADA data to build predictive failure models for a particular model of wind turbine and configure their Energy.ML platform to prescribe maintenance actions to mitigate the costs.

Our AI and machine learning experts worked closely with Ensemble’s deep expertise in wind turbine controls, loads prediction, and mechanical failure modes to identify the key features to focus on for training, and experimented with various models including Multivariate Gaussian Anomaly Detection, LSTM, AutoEncoders, and XGBoost. Since the performance of each model varies based on how the data was collected, our data scientists built ensemble models for each prediction type.

Ensemble Energy’s Energy.ML platform provides descriptive dashboards of the inner functioning of component parts, prescribes actions to improve efficiency and reduce downtime, and forecast future maintenance and replacement costs of a windfarm.

Results

We were able to predict pitch bearing failures as far as 6 months in advance, and main bearing failures 2 months before traditional CMS systems, with 90% accuracy. The models can also identify which component is most likely to fail, the type of failure, and the remaining useful life of constituent components. In addition, the team built a model to identify turbines with misaligned yaw (which significantly decreases energy production) with 96% accuracy (using just 2 months of historical data).

If you are interested in deploying a solution built on our Extract Learning Platform, contact us to see what Foundation AI can do for you.
Artificial Intelligence for the Real World
© 2021 Foundation AI