Machine Learning in the Wind Industry

From Machine Learning
Jump to: navigation, search

The wind industry is developing interest in machine learning, particularly in regards to improving preventative and predictive maintenance. Most wind farms and most wind turbines within them have many sensors and alarms that provide colossal amounts of information about the immediate outside environment and the internal status of the turbines. Nearby met masts often also provide a lot of information about the surroundings. There is huge potential for all this data to be turned into local weather forecasts and status updates and predictions.

An example application of machine learning in the wind industry is illustrated with the following question: what is the predicted power a turbine can generate at a given wind speed?

ML in wind.jpg

The model for this could look like the following (where the outline is the "black box" of a neural network for example):

ML in wind detail.png

The following projects are ones that the ORE Catapult have been involved in. Some are or were exploratory projects, to investigate what kind of insights the data could provide. Others attempted to provide models that will get real, actionable insights.

University of Strathclyde using Microsoft Azure

This project trained a Neural Network to classify data as healthy or unhealthy.

The neural network was trained on the same data in two ways:

  • 80/20 training/validation dataset split
  • 80/20 5-fold cross validation
StrathNN.png

Wolfram Mathematica

Dimensionality Reduction

This project plotted 20 SCADA signals on a 2D plane. Each colour refers to a different turbine, and the size of the bubble corresponds to wind speeds. Points that are close to each other are similar.

Wolfram1.png Wolfram2.png

From the above two images it is clear that turbine 2 (the dark green in the first image and light green in the second image) is different from the other turbines. The shape of the clusters may be showing different modes of operation.

Classification distance graph

This project plotted turbines on a plane where turbines that are closer are more similar to each other.

Wolfram Alpha

Wolfram alpha is a language interpreter and maths engine, good at understanding human language and solving complicated mathematical functions.

It could potentially be used to develop a data based interface combining

  • SCADA signals
  • Work orders
  • Project finance

The interface would have the goal of answering questions related to lost energy production and any downtime a wind farm experiences for example.

Hackathon!

Problem statement for the hackathon held in October 2019:

National Grid needs reliable information on how much power a windfarm can produce so that they can balance the power supply.

The 5 finalists of the hackathon all used machine learning methods.

Hackathon.png

Local Sea State: Stop and Go Offshore (LoSSGO)

The aim of the project is to build an application that predicts local sea-state using

  • Buoy and turbine mounted devices
  • Windspeed measurements
  • Weather forecasts
  • Open ocean buoys
  • Satellite data
  • Tide and current information

The insights are to be used to better plan turbine access at specific times.

BLEEP Classification

The goal of this project is to standardise damage categorisation and imaging, to then lead to automatic classification of blade damage. The source photos would be taken by drones.

Data Pilots

What can we do with SCADA data? The problem statement for this project was:

How can machine learning approaches be used to predict power and identify leading edge erosion from 10-minute SCADA data?

Input data was the active power from neighbouring turbines, and the output was a production estimation for the target turbine.

It was found that neural networks are a valid method to identify leading edge erosion and potentially as a means to quantify its impact on turbine performance.

Other Machine Learning example projects in industry

  • ALICE Project (completed 2014) used reinforcement methods as neural network models to optimise wind turbine performance.
  • There is research ongoing at the Advanced Forming Research Centre (AFRC) on deep reinforcement learning in the context of robotic control for manipulating objects and machines.
  • ROMEO Project is ongoing using Horizon 2020 funding to develop machine learning methods to improve wind turbine and wind farm production.
  • A paper[1] has been written on the development of a Random Forest classifier for fault diagnosis of a wind turbine's gearbox

References