The global renewable energy industry has seen unprecedented growth over the last several years. According to the Global Wind Energy Council (GWEC), the cumulative installed capacity of wind power projects has increased from 24 GW in 2001 to 432 GW in 2015 and is expected to grow to 703 GW by 2020 (Global Data). Much of this growth has resulted from great state and national level policies, tax incentives and high electricity prices. However, many of these incentives have since been reduced and now wind power companies face pressure to improve profitability, while scaling operational excellence.
Mechanical engineering and physics based improvements have long been used to increase plant operations. However, these improvements plateau after a period of time thus providing diminishing returns. In the increasingly connected world of today, sensors are being used to achieve operation excellence based on real time, plant specific data. Companies now are realizing the value of using data to maximize returns as compared to making any physic improvements to the components. In this day and age of the Internet of Things (IoT), sensors are being used to record data from every device, and the renewable energy industry is not one to be left behind.
Governments around the world are slowly scaling down renewable energy incentives such as tax benefits and generation based incentives while renewable energy electricity prices are falling. Companies are thus focusing on minimizing operations and maintenance (O&M) costs and maximizing power production. Data analytics enables O&M teams to achieve both goals by taking a closer look at the data from every plant device. According to a recent report by McKinsey, depending upon the plant’s existing level of performance, better O&M could account for as much as a 20% increase in IRR. The type of data analytics for better wind farm O&M falls into three categories: a) forecasting, b) alarms based on threshold values, and c) condition monitoring.
Power and weather forecasts are useful for plant owners to understand what the future generation of a turbine will be. These forecasts are used regularly in energy trading markets which are an additional source of income generated from the wind farms.
The second type of data analytics is identification of threshold values for alarm systems. Data from plants is collected and mapped to identify trends during normal operation, that is operation during which time a failure does not occur. These figures are used to create threshold values and raise an alarm any data point from the plant falls outside the range contained by the thresholds. These alarms based on threshold values signal any issues that need attention, including but not limited to failures and system downtime.
The third type of data analytics is condition monitoring to predict a failure well before it occurs, thereby reducing downtime. By identifying changes in significant parameters that lead to equipment malfunctions, it is easy to prevent major failures, especially during high wind season when maintenance charges skyrocket. Furthermore, it provides an in-depth description every turbine’s behavior. For example, the average lifetime of a turbine is said to be is 20-25 years, but these figures are difficult to justify. With condition monitoring, the data is analyzed to understand individual turbine behavior, the pressure it has been exposed to, etc. in order to better approximate how long it will keep performing at its peak. This analysis can also contribute to customized repair and maintenance schedules based on the component’s health instead of a fixed schedule. Out of the three data analytics categories, condition monitoring has the power to provide maximum monetary benefits to wind farm owners by enabling maximum uptime.
Furthermore, condition monitoring can be categorized into three types: a) physics model, b) statistical model c) artificial intelligence model.
The physics model is mainly used for engineering purposes. For example, to find out the impact that force has on a component’s load and fatigue. If all the data is available, a physics model is highly accurate. The drawbacks are that a physics model is not scalable because each component of a wind power plant behaves differently. For example, a gearbox’s specification differs depending on its manufacturer. It is thus difficult to use a physics model created for one gearbox on another gearbox built by a different manufacturer. Furthermore, physics models are computation heavy which makes them less than ideal when deployed at scale. A physics model also requires in-depth knowledge of a component’s behavior as well as a complete list of component parameters that can be difficult to acquire.
The second type of model, statistical models are currently the most popular choice because they have been tried and tested. These models can use existing SCADA data to predict a component’s lifetime behavior. Unlike the physics model, statistical models can be highly accurate without component specific information. These models are easy to scale and can be built for any equipment that has sensors attached. Typically, in the wind industry, condition monitoring is performed exclusively for the gearbox. Statistical models expand that scope to all the components within a wind turbine. The drawbacks of statistical models are that they are computationally heavy and can be difficult for the end user to understand. However, the biggest drawback of statistical models is that they are only as good as the data used to build them. The model will be unable to predict future failures unless it has learnt them from the past data. For a statistical model to give accurate results over a period of time, the model will have to be “refreshed” by re-building it at regular intervals using new data that will capture all the known failures. This is not only time consuming but is also a significant crutch when trying to scale the models.
The third, and most effective method to predict failures is Artificial intelligence (AI). Like the statistical models, AI models use existing SCADA data and don’t require detailed information about the device which makes them easy to scale across different plant equipment. What sets AI models apart from statistical models is that they don’t need any manual “refreshment”. AI models can improve with every additional data point. These models are built to learn from new data and improve any past errors so that they are not repeated in the future. This makes AI models very accurate but at the cost of heavy computation and in most cases, it is challenging to explain the complex mathematics behind the results. For example, one of the most popular AI models is artificial neural networks (ANN). This model processes data similar to how the human brain processes information. For every new experience, the human brain begins by first recording information, acting on that information and then replicating actions based on new information. Similarly, as more data is fed to an ANN, the more it learns and the more accurate its results get. Since ANN is a “black box”, which is to say that it is difficult to say what is going on in each neuron, it is difficult to explain the results. However, there are numerous methods used to ensure that the results can be trusted, even if, their accuracy can’t be explained.
Ultimately, data analytics can boost a wind farm’s uptime thereby adding value for any future transactions. Thus, O&M powered by data analytics is the key differentiator between well performing and under-performing wind farms.