Big data enabled farm wide drivetrain load distribution analysis for O&M and lifetime optimization

Introduction

There is increasing attention for condition monitoring techniques to minimize the influence of downtime on turbine revenue. The goal is to get early warning about failing subcomponents, such as gearboxes, main bearings, and generators in order to cluster maintenance actions during low wind days. In addition there is a clear trend towards lifetime optimization for different turbines in a wind farm. Since not all turbines are loaded identically each turbine has a different load history and hence a different consumed lifetime and remaining production potential.
Detailed knowledge of the load history for each turbine brings essential information to the table for gaining insights in loading conditions leading to failure initiation and the real loads driving the consumed lifetime. Equipping turbines with strain gauges can provide the necessary load information. However, these devices have a cost and limited lifetime. To gain additional insights in a more cost effective way this paper suggests the use of traditional design calculation approaches running on an integrated big data platform that stores 1Hz sampled operating signals from turbine Supervisory Control and Data acquisition systems (SCADA) for all turbines in the farm.


Approach

The paper discusses an integrated approach to assess the difference in loading conditions for the different turbine drivetrains in a wind farm by assessing load information from 1Hz sampled SCADA data. Traditional design load calculation techniques, such as Load Duration Distribution (LDD) and Load Revolution distribution (LRD), are performed to compare load history between turbines. In addition it is shown that fatigue consuming and potentially failure-initiating events, such as emergency stop, can be derived from the 1Hz data. These events give complementary information to existing condition monitoring approaches and open the path towards prognosis.

Main body of abstract

For rotating components, such as gearboxes and generators, bearings are critical particularly with regard to failure. Therefore, we focus on those loads influencing the rotating components. Since SCADA signals are used our focus is on torque.

Drivetrain fatigue is accurately captured in the design standards. We use two metrics: LDD and LRD. First it is shown that there is a significant difference between the summary metrics LDD and LRD, generated from 10-minute averages and the ones from the 1Hz data. Second it is shown that the load metrics change strongly throughout the farm, showing the need for individual turbine load assessment.

Events are an important aspect for lifetime and failure initiation. This paper shows the advantage of 1Hz data over traditional 10-minute averages with regard to events. It is shown, based on a real-life storm event, how a wind speed dip triggers a start-up of the turbine and corresponding shut-down afterwards. Both initiation trigger and lifetime consuming events are identified automatically from the time series data allowing event counting.

Six months of real life data from all turbines in an offshore wind farm is used. This data is part of a bigger consistent dataset containing all SCADA sensor data, SCADA status codes, custom high frequency accelerometer measurements, and maintenance data from the wind farm. Using a scalable big-data platform allows to perform the load analysis in one calculation iteration rather than having to load the different datasets sequentially.


Conclusion

This paper showed a big data enabled approach to use available high frequency SCADA data for gaining insights in drivetrain load distribution throughout a wind farm. It was shown that each turbine should be investigated individually as fatigue loading changes within the farm. Moreover, it illustrated an approach to detect fatigue consuming and potential failure initiating events from the available dataset.


Learning objectives
Understand the potential of SCADA based analysis for lifetime
Show differences in loading within a farm
Show how failure-initiating events can be identified from the data