Offshore wind farms can make an enormous contribution to energy supply from renewable sources with their above-average wind turbines. The challenge: planned maintenance or spontaneous service calls are extremely time-consuming and costly – offshore wind farms are highly complex and difficult to access. A large energy supplier therefore wanted to be able to determine with high accuracy which sensor values measured at the turbines are normal for the current condition of a wind turbine and thus be able to detect irregular behaviour more quickly in the same time. The aim was to implement an intelligent monitoring system for wind turbines in order to quickly identify problems inside the turbines and ultimately reduce downtimes and breakdowns. Data Reply provided the architecture and implementation for the project. The end product is a special and efficient tool for site operators that has now been deployed in 30 onshore and offshore wind farms with more to follow.
There are already many tools on the market for monitoring sensor data, but they are often not mature enough to meet the special requirements and specific needs of the wind energy industry. The wind power sector needs intelligent tools which can be made possible through machine learning making them suitable for complex plants. The challenge was that due to the different configurations and after-running effects, each turbine is unique in its behaviour. Therefore, the models had to be trained on a turbine-by-turbine basis. The sensor data required for this had already been recorded and have been processed in the course of the project.
To predict the expected turbine condition, Random Forest Regressor Models are used, an advanced decision tree concept that is less prone to overfitting.
The models allow a deviation of the real data from the expected value to be determined with high accuracy. This makes it very easy to identify even small anomalies in the custom dashboard.
Data Reply's experts used open source tools for the development, among other things to avoid license fees. These include on-demand Spark Clusters in the Azure Cloud.
As soon as enough data is available, an automatic trigger creates a cluster and starts data processing automatically. After processing, the cluster is shut down again, resulting in high cost savings. Even if the solution with Spark would be possible as a 24/7 streaming job, the "On-Demand Batch Job" variant is preferred, since it is a cluster of about 2TB RAM, which is thus only switched on 2-3 hours a day.
Along with the customer, Data Reply's Data Scientists and developers built a highly scalable Spark Pipeline that uses trained machine learning models to identify deviating sensor readings in wind turbines. The results are presented in the wind farm monitoring tool and provide valuable input for site operators. They can now identify problems before they lead to significant downtime. The project is currently under development as further functions are to be introduced. Current features of the development include vibration analysis, categorisation of downtime and tracking of work orders. In addition to the expansion of the functions, the system will also be implemented step by step for further wind farms.