Greenbyte Documentation

Frequently asked questions

Energy Cloud officially launched Predict in April 2019. We have compiled some frequently asked questions to give you a better understanding of what Predict is and how it works.

Please contact your Customer Success Manager if you would like more information. We are always happy to help you optimize your Energy Cloud experience.

How does Predict work?

Predict is a temperature-based condition monitoring system designed to anticipate maintenance needs based on the data collected from a wind turbine SCADA system. Predict utilizes a mathematical modeling method called artificial neural networks (ANN) to create a model of various temperatures being measured in the turbine. The ANN models are designed to emulate the thermal equilibrium condition within the nacelle. During normal operation, the thermal equilibrium should not be disturbed, and the modeled and actual temperature values will not be significantly different. However, when a component abnormality leads to a change in the thermal equilibrium, the modeled and the actual temperature values will deviate. Predict detects these deviations and notifies the user of a potential issue in the turbine component.

Predict utilizes the ambient temperature, nacelle temperature, generated power, and rotor/generator RPM as inputs to create a model for component temperatures. Each of the inputs represents a source of heat. The component temperatures increase or decrease based on the ambient and nacelle temperatures. Generated power is highly correlated to electrical losses in the system, and rotational speed is highly correlated to mechanical losses inside the nacelle. This has a direct effect on the component temperatures and can therefore be used to accurately predict maintenance issues.

How does Predict differ from a vibration-based monitoring system?

A vibration-based CMS utilizes vibration data provided by specially installed sensors on the turbine. However, Predict works on measurements that are already available in the turbine SCADA system. This simplifies the onboarding process because it does not require installation of any new sensors. Vibration-based CMS is good at detecting faults in the drivetrain system, whereas Predict is effective in detecting faults in other sub-systems of the turbine, such as the converter system, the hydraulic system, etc. Predict is especially effective in detecting faults that manifest as a change in the operating temperature of the component.

How much historical data is required to run Predict?

Predict requires data available from at least one year of operation to train the ANN models.

Which components are monitored by Predict?

  • Transmission

  • Drivetrain system

  • Rotor system

  • Power generation system

  • Hydraulic system

  • Nacelle

  • Cooling system

How are the components defined in Predict?

The RDS-PP taxonomy is used to map various temperature measurements to components defined in the taxonomy.

How is the alert generated in Predict?

The ANN models generate a simulated component temperature for each 10-minute timestamp. This simulated value signifies the expected temperature of the component’s operation. The simulated temperature is compared to the actual value and an alert is generated when the defined conditions are met.  

Event: An event is created when the difference between the simulated signal and the actual signal is greater than the threshold value for at least half an hour:

(Simulated Signal – Actual Signal) > Threshold Value.

The threshold value is calculated based on the training performance of the model. The threshold value is defined as Mean + 3*Standard deviation of the training error. This method of calculating the threshold value avoids alerts being generated due to high training errors as a result of relatively bad data during the training period.

A severity measure is calculated for each event, which depends on the amplitude of the deviation and its duration.  

Severity: If the cumulative severity measure crosses the Low Severity threshold of 600, then a low severity alert is generated. Similarly, if the severity measure crosses the High Severity threshold of 1200, a high severity alert is generated. These thresholds are based on extensive testing done during the development phase of Predict. The thresholds will be updated if necessary.  

Decay: After the severity measure for a model crosses the Low Severity threshold, an alert with Low Severity is created. If no new events occur for the model, then the severity measure starts decaying at a defined rate. Once the severity measure reaches below the Low Severity threshold, the alert is automatically acknowledged. However, if the severity measure breaches the High Severity threshold, even if no new events occur, the alert is not automatically acknowledged. A user intervention is needed to remove the alert from open issues.

Where can I find which year was used to model the data?

In general, the data from the recent past one year is used to train the models. For example, if the Predict application starts on January 2019, then data from January 2018 to December 2018 is used to train the models. This date information is not visible in the UI. If a user needs to view this information, we recommend that they please contact Support.

Predict has shown an alert for a component in the turbine, what should I do next?

Predict triggers alerts on small deviations from the normal behavior. This sensitivity means that in most cases, there is ample time to fix any potential issues. We recommend that when users receive an alert from Predict, they perform a first analysis by quickly browsing through the Predict / Details page. The information on the Predict / Details page can then be provided to the maintenance team who can schedule an inspection of the specific component. Once the maintenance team has resolved the component issue, the user is highly recommended to use the Resolve action in the Predict / Details page. Here, the user can provide information about what action and which component were repaired and/or replaced during the maintenance. This feedback information will be used to train the recommendation system, which will provide details about the most probable reason for a certain alert.

How do I interpret the plots on the view page?

Currently, there are four plots on the Predict / Details page.  

First plot: This plot displays the simulated temperature vs the actual temperature for the component. The time frame to be displayed is chosen automatically between First anomaly date – 1 week and Last anomaly date + 1 week. The purpose of this plot is to show when and by how much the temperature has deviated from the expected temperature for the component.  

Second plot: This plot displays which operating region shows the highest deviations in the temperature. Here, the power from the turbine is divided into 50 bins. The operating temperature of the component is averaged for each bin. The data from the period between First anomaly date and Last anomaly date is compared to data from the same period one year prior.  

Third plot: The third plot is a scatter plot of component temperature vs power generated. As with the second plot, the data from the current period is compared to the data from one year prior. This plot provides information about how often the temperature has deviated from the normal operation over the period in which the alert has existed.  

Fourth plot: The final plot is a scatter plot of the component temperature vs the ambient temperature. The data from the current period is compared to data from the previous year. Since the ambient temperatures are highly correlated to some component temperatures, it might happen that an unusually high ambient temperature causes higher temperatures in the component. This plot is designed to indicate any issue due to high ambient temperatures. If the component temperatures are high due to high ambient temperatures, all the data points from the current year will be concentrated in the top right corner of the plot. If the data points from the current year are not concentrated in the top right corner, then ambient temperature is not the cause of high component temperatures.