By Sabari Ram Subbaraman, Data and Operational Analytics Engineer, DNV GL – Energy
The rate of decarbonisation worldwide is expected to rise rapidly in the next few years due to climate change action and global policy shifts. The role of wind power in the energy transition is undeniable, and DNV GL’s Energy Transition Outlook 2019 forecasts wind-powered generation to increase over fifteen times from 1.1 PWh in 2018 to 17 PWh in 2050. However, migitating climate change is not as easy as plugging wind generation sources into the grid. Wind turbines comprise rotating machinery and equipment that are susceptible to failures and require routine maintenance during their lifetime. Research from Wood Mackenzie shows that global onshore wind operations and maintenance (O&M) costs will reach nearly $15 billion in 2019, of which $8.5 billion will be spent on unplanned repairs and corrective maintenance. Failure statistics from major reliability studies suggest that drivetrain components such as gearboxs and generators have a long downtime per failure and large cost implications. From DNV GL’s experience, bearing, gear and lubricant failures are common gearbox failures while stator, rotor and bearing failures are common generator failures. The deterioration of components occurs due to loading during high wind speeds, start-up, grid connection, emergency stops, shutdowns and other transient load events.
Adopting a proactive maintenance strategy is essential to cut down O&M costs and downtime. An effective approach is condition-based monitoring of components to predict failures. Due to the increasing importance of detecting incipient failures, many monitoring systems and approaches are being developed. These approaches are broadly divided into diagnostic and prognostic. The implementation of prognostics is relatively new in the wind industry. Typically, diagnostics deals with the detection of anomalies, identification of the affected component and the extent of the fault while prognostics refers to the estimation of time to failure, the remaining useful life and future failure modes. Some examples of commonly used condition monitoring systems (CMS) are supervisory control and data acquisition (SCADA) and specialised CMS. SCADA receives data feed on temperature parameters, turbine operational parameters, tower/drivetrain acceleration and status codes while specialised CMS incorporates additional hardware components to detect vibrations, acoustic emissions and particle measurements of drivetrain components. Market estimates price specialised vibration CMS at $7,000-$9,000 per turbine excluding the annual costs associated with maintenance, monitoring and software. This could be an expensive investment for a large windfarm stakeholder.
Due to low costs and existing installations, SCADA-based condition monitoring is widely used in the industry. Raw SCADA data is not very successful in detecting incipient faults due to the varying operational and environmental conditions stressed on the turbine. To overcome this, multiple SCADA-based data models are being researched and deployed using commercial software. Furthermore, machine mearning based models such as polynomial regression and artificial neural networks (ANNs) are used for detecting any abnormal behaviour. ANNs can identify complex relationships between input and output variables, and are now being widely used in other areas such as forecasting and control. Other machine learning approaches explored are random forest, and ANN-based adaptive neuro fuzzy inference systems (ANFIS) and self-organising maps. Although these models are now becoming easier to implement, more failure cases from a range of wind farms and manufacturers are needed for demonstrating the success of a model. Data models have limited physical understanding. In contrast, physics-based damage models represent the failure with physical significance when applied to condition monitoring. Hybrid models that incorporate data models with underlying physical understanding can be effective in predicting failures. In addition to detecting anomalies, tools and processes must be developed keeping in mind the failure modes. In doing so, the industry is faced with various challenges.
In conclusion, integrating multiple CMS, historical failure patterns, parameter relationships, alarm logs and operations records can help arrive at a quick prediction and provide actionable insights for a proactive maintenance strategy. Also, the integration of planning and forecasting tools can help refine the O&M process and feed valuable information back into the chain. From DNV GL’s experience, incorporating a proactive maintenance strategy has helped predict generator bearing failures well in advance and prevented a potential downtime of 500 hours and estimated revenue loss of $28,000. Therefore, at a time when the world is adopting various renewable energy generation sources, including wind, continuous monitoring and analysis is required for preventing unplanned failures and optimising the value of assets.