Temperature Prediction Neural Network Toolbox - Typical Wind Turbine Component Temperature Across a 50 Turbine Wind Farm

We want to detect outlier component temperatures across a wind farm and focus maintenance efforts. One plan is to make a neural network time series model for the temperature of a major component on the "typical turbine" on a wind farm. Turbines with a poor model fit (where the component is significantly hotter than predicted) are examined further.

Example: predict the temperature of the rotor bearing based on previous and current measurements of other quantities like ambient temperature, active power, wind speed. Previous measurements are included because the component has a large thermal mass. Previous values of the rotor bearing (modelled component) temperature are not used as predictors because model error is used to determine whether a particular turbine's rotor bearing is running hotter than it should. Consider there are 50 wind turbines on a wind farm, we have 50 sets of time series data available to train the network (say one month of 10-minute measurements). What are some methods to generalise the network keeping in mind that because the temperature of the component at t=0 depends on values at t=-10min, t=-20min, the data must stay in order (or at least be split into chunks that remain in order). Given that we are looking for outliers and most "typical" turbines have a similar range of predictor values over the entire examined period, possibly we could take the median of each predictor signal (across the 50 turbines) and use these "median predictors" to target the median rotor bearing temperature when training. But in some situations this might limit the extents of the data and force the model to extrapolate beyond its training set. Is there a better method? Should we instead make 50 neural networks (one for each turbine) and then feed a constant input set into each model and look for the network that gives the highest temperature prediction? Thanks, Tom

Answer 1

[Greg Heath]

I would first assume that the combined time-series of all variables for one turbine is enough to predict it's bearing temperature.

To that end, calculate the auto and crosscorrelation functions to determine the time lags needed to make the prediction.

Failure prediction would just involve a threshold.

Next, consider the advantage of using measurements from the other 49.

Hope this helps.

Thank you for formally accepting my answer

Greg

[SCADA Miner]

Hi Greg, Thanks for your suggestion. There is of course some variation between each turbine. Picking a representative one is an option (e.g. the one with the wind farm median value of the measurement we are examining). But I am interested in whether there is an accepted way of creating a general model by somehow using all available data rather than just limiting the training data to that from one turbine.

[Greg Heath]

I really don't see how using input data for 50 turbines is going to clarify anything.

The important data is the predicted and true outputs from the 50 turbines.

Outliers should be easy to spot.