This August on our blog we talk about the Predictive Maintenance and RCM to prevent failures in renewable energy plants, Timo Lichtenstein, Research Associate at Fraunhofer Institute for Wind Energy Systems. We have been talking about this items.
How can early fault detection help to improve the efficiency and lower the costs of renewable energy plants?
Nowadays, although wind turbines already have an availability of at least around 97%, this 3% downtime can be responsible for up to 11% lost energy: Regarding the installed capacity of 220 gigawatts (GW) in Europe and assuming a mean capacity factor of 33%, this equals to a lost 70 terawatt-hours (TWh) of energy per year. With an economical per head consumption of 1500 kilowatt-hour per year (kWh/a) this is as much as the annual energy consumption of over 45 million people.
Furthermore, the costs for operation and maintenance can still be up to 25% of the whole life cycle costs for an onshore turbine – a big share here are the costs for service and repair. Small parts might not carry that much weight, but larger parts like gearbox, generator, or converter – although their failure rates might be quite low – need to be bought, transported, and installed, which involves both high personnel capacity and costs. Moreover, inevitably associated with this comes downtime that leads to a loss of revenue – additional virtual costs. All these costs further increase regarding offshore wind farms where logistical challenges are higher due to weather, heavy seas, or simply the unavailability of a service vessel.
This leads to two key improvement possibilities: First, increasing the reliability of wind turbines will significantly increase the energy yield from the wind. Second, knowing about the condition of a component beforehand or about the time when it will most probably fail helps to better plan service assignments. In combination, this means a reduction of the downtimes as well as the operational costs, both directly entering the levelized costs of electricity (LCOE). Certainly, the latter is a direct measure for the acceptance and competitiveness of wind energy compared to other (renewable) energy sources.
In our team “Technical Reliability” at Fraunhofer Institute for Wind Energy Systems IWES in Hannover we address the issues of finding the root causes of wind turbine failures and understanding the fault mechanisms behind them, hence, investigating the challenges mentioned above. Before we develop solution-oriented methods that inform about the health of components of a wind turbine, we start by identifying those components that are often defective and cause the turbine to fail and, consequently, to stop generating power. A second important aspect is to answer the question of why a particular component fails. These inputs can then be used as a basis to understand and eventually model parts of a wind turbine that can further be used for fault detection.
Nowadays, one of the most common approach for early fault detection is by monitoring components such as the drivetrain by rich high-frequency data. They play a key role in the technical availability of wind turbines. Though typically, such sensors are not part of the standard equipment of turbines and need to be installed explicitly. Furthermore, users complain about a sometimes-insufficient reliability and fault-detection performance. At Fraunhofer IWES, we support operators to select the systems which make sense from economic and technical points of view. Using field data, systematic assessments of the detection performance or probabilistic cost-benefit analyses can be given.
The Fraunhofer Institute for Wind Energy Systems IWES is involved in a project called WiSA big data, could you explain to us what this project is about?
Modern wind turbines already have a vast number of sensors with high sampling rates. But usually only a selection of these data is stored and, furthermore, unfortunately, most of them are only recorded as 10-minute average values. This temporal aggregation of the signal saves a lot of space on a computer. However, as such data are often available in a higher temporal resolution, they can contain relevant information, although considerable research is still required to further exploit this.
To use this information is the focus of the “Wind farm virtual Site Assistant big data” (WiSA big data) research project. Here, we utilize the high-resolution operating data from the Supervisory Control and Data Acquisition system (SCADA), the standard equipment of a turbine. We then combine these data sets with status codes, service reports, and other maintenance information to distinguish between a healthy and non-healthy state of the turbines. On this so-called “labelled data”, we apply partly already existing, partly new methods not only from the wind energy sector but as well from other disciplines and investigate them for the purpose of early fault detection. All these methods will be improved to benefit from the rich datasets and, if possible, concatenated with each other to optimize detection performance. Within the project at Fraunhofer IWES, we are responsible for the development, application, and evaluation of methods for early fault detection and for deriving decision support to optimize operation and maintenance. We will also investigate the possibilities to expand the methods to estimate residual useful lifetimes that would allow for an even more detailed scheduling of services.
The final product of the project will be a soft- and hardware platform as the core system for our virtual wind farm site assistant. We call it “demonstrator” because this platform will contain a graphical dashboard hosting the successful methods using live data for analytics and to test the industrial use case.
How does big data help to detect early faults?
Our approach takes all available signals of a turbine in the highest sampling rates available into account, ranging from 1 hertz (Hz), i.e., on the seconds scale, up to several kilohertz (kHz). Using this maximum of information in the data will increase the detection of anomalies and atypical behaviors. Such detections are nowadays often made by signal values such as temperatures outside a certain allowed range. Temperatures change their values rather slowly. In fact, we found out that their loss of information is negligible for certain components comparing data with a resolution of 1 s to 10 min average values. We already gave a talk at the Wind Energy Science Conference 2021 discussing the loss of information when temporally aggregating data. Unfortunately, an increasing temperature in a component might only be a symptom of second order, when, e.g., a bearing is starting to deteriorate. The first order anomaly should be visible in a more intrinsic sensor of a component: For a bearing, vibrational data will give an insight to developing faults and for electrical components deviations in the oscillation of the voltages might hint to a defect. Nevertheless, also other sensor values that are not directly associated with a component, but rather have a direct or maybe even hidden link can help in detecting upcoming failures. These deviations might occur even before other signs of an imminent failure appear. Consequently, operation and maintenance can improve its service assignments and order spare parts early enough or even fix little issues that could otherwise become bigger and costly problems. All this will also reduce possible downtimes to a minimum reducing a loss of revenue.
Furthermore, such a big data approach does not only need operating, but also service and event data are necessary. These data are crucial for the learning step in most of our method’s algorithms, but typically exist in all kinds of formats, some of them occasionally handwritten paperwork. But even with the texts in digital form, the information must be processed into a standardized structure for further use. In WiSAbigdata, we chose to work with the reference designation system RDS-PP for wind turbines in combination with the state-event-cause code ZEUS. I highly encourage to use these standards as they greatly support the first step in data analysis. Still, much of the time required for the knowledge gain is spent pre-processing such unstructured event data. Be it man or machine, standardization leaps the quality of diagnosis that in return will be much cheaper due to the independence from closed platforms.
In summary, with the availability of more high-resolution data of a turbine and especially for a component come more precise and detailed answers to the current state they are in. Furthermore, also earlier predictions of a preventable event are possible due to decreased uncertainties. In the future, the combination of a foundation of big data sets with multiple highly precise predictive models might, in the end, also lead to the development of highly dependable digital twins.
Regarding wind turbines, what magnitudes (vibrations, noise, thermography, oil analysis…) do you consider to be the most important for early fault detection?
For many components afflicted with high repair costs, systems for a reliable condition monitoring are not available on the market yet. In addition, our project is still in an early phase. Due to the diversity of possible detection models, a list with the most important signal names for early fault detection in general is therefore not available. Right now, we focus on the combination of all available data and we investigate which information we want to obtain. In any case, the finalized methods for particular components will then only rely on their validated signals..
To give you an insight into our considerations: For a deviation of the normal condition of a whole turbine wind speed and the active power output in a resolution of 10 minutes might be enough. But then you do not know the reason for the deviation yet. If you want to pinpoint the affected component, temperatures measured at the component’s location can give you a clue. However, this will not give any information to the root of the deviation and if a failure will occur at all. Now, for a deeper look important information about the health of an electrical component might be obtainable from high-resolution voltages and currents. Then again, for insight into a rotating component, vibrational data could be the key data source. Dedicated condition monitoring systems for bearings, e.g., make use of such vibrational sensors but often do not link their dataset to data from the rest of the turbine.
Which data quality will be necessary for a certain analysis is still under investigation. This will also be one of the results of WiSAbigdata in the form of an assessment of the forecast accuracy of the methods with varying temporal resolution as an input variable. The studies about the additional information in high-resolution data will serve as a basis for this.
For possible life extensions of wind turbines or after the end of a contract, could high-resolution operating data provide us with the lifespan left of the main components of a wind turbine?
Information about the lifespan left is certainly hidden somewhere in operating data of wind turbines. To access such a valuable resource a variety of well-known models must be trained with a profound data basis: Ideally a nearly complete archive of all operating data and a documentation about the maintenance carried out since commissioning is available. With this historic turbine data, economic decisions on the lifetime extension can be more reliable.
Although it is one goal of WiSAbigdata to also expand the methods of the project to also give an estimate of the residual useful lifetime of a component, first, we would focus on forecasts of several months. However, combining this principal approach with load calculations for damage equivalent might give the essential insights to good estimations with low uncertainties.