Predictive models (based on machine learning algorithms) aim to assess future outcomes based on analysis of past & present data. Increased processing power, availability of cloud infrastructure and commoditization of data mining has helped these models to move from the “lab” to the realm of “real time decision making”. Some common examples of use of predictive models include-
In this article, we will look at predictive maintenance as applied to wind turbines & solar plants. Maintaining and operating these renewable assets according to data driven approach requires tons of data: real time data, historical data, data from similar assets and past maintenance records. The standard maintenance methods adopted currently are one of the following –
The problem with these approaches is their cost. While reactive requires us to waiting until a component failure, this approach results in lost generation time and revenue. Periodic maintenance through regular inspections are also expensive and can lead to unnecessarily replacements.
The approach to developing a predictive model for these assets involves parameter selection based on correlation with historic failures. We construct a model that correlates past failures using multivariate analysis. The models can mine all the variables and conditions that contributed to past failures in order to predict future failures. New or current data is then fed to the trained model to generate a health scores on a real time basis.
Algo Engines leverages R for predictive analytics. With numerous machine learning libraries, strong community support and easy code integration, R stands as a strong candidate for among others for building machine intelligence. Some predictive modelling approaches that we have evaluated & developed include:
Asset Health Score
The Asset Health Score model is based on linear regression. It computes the likelihood that the asset will fail. We have often heard of the Body Mass Index (BMI) as an indicator of fitness level of the person. The Asset Health Score of a turbine or inverter is like the BMI. To construct the Health Score model we use historical failure data, operational information and environmental data to determine the current operational health. These would typical include temperatures, pressure parameters, vibration levels at different points of the turbine drive train or solar plant structure. A more representative health score will also quantify cracks and structural weaknesses. The health score value is presented as a number between 0 and 1. Values closer to 1 indicate a healthier asset. Once health scores across a farm are compared, it is an easy exercise to pick out the ill performers and investigate individual contributions of components. This helps plan and prioritize specific maintenance activities with maximized benefit.
Time to failure Analysis
The Time to failure Analysis model estimates an assets remaining lifespan under standard operating conditions. The unit of time can be in days, hours, minutes (or stress cycles, or any other metric). Inputs to this model include component failure data over a long period of time which helps the model understand failures of different components over different seasons and condition. It is also essential to feed in component –wise maintenance records which adds offsets to failure and improves failure predictions. For example, a gearbox with a recent oil change is likely to run for a much longer period than it would have without an oil change. In a solar plant, a replaced panel or string of panel is less likely to have drift or dip in current generation owing to aging.
Time series models:
Time series models are used for predicting or forecasting the future behaviour of variables. These models leverage the fact that data points analyzed over time exhibit patterns such as trends or seasonal variations. Models capable of learning these are able to capture anomalies and aberrations in real time data which can be investigated as abnormal behaviour. Standard regression techniques cannot be applied to time series data as they do not provide a methodology to identify the pattern. We can use the time series models like ARMA to predict say the trends in vibration of a gearbox which can be compared against control limits to identify when the vibration will cross the control limit. These can be applied to different temperatures, generator parameters like voltages, power or even to metrics like performance ratio of a solar plant.
In summary, predictive maintenance has the following advantages:
In the current power economy where performance of renewable energy assets is key to its project feasibility and future investment, predictive maintenance is the way to go…
For more on data mining for forecasting and predictive analytics, stay tuned to the next post from Algo Engines..