support

Input Data:

All input signals are 10 min values

• wind vane direction
• wind speed
• wind direction
• wind speed standard deviation
• pitch angle of the three blades
• turbine power
• generator rotation speed
• turbine power of 5 neighboring turbines

6 months of data are required. Better to have 1 year of data.

Physical background

The industry distinguishes between dynamic and static yaw mis-alignment. static yaw misalignment is characterized by a static offset between wind direction and and nacelle orientation. The most common cause is a wind vane that is not correctly calibrated. Dynamic yaw mis-alignment is characterized by an error of the turbine control during wind shifts. The most common fault is that the turbine reacts to slowly to wind shifts, increasing the yaw misalignment temporarily while yawing than a faster yawing turbine.

Commonly, the expected annual power production loss associated with yaw misalignment can be approximated by E_AEP(1 − cos²ϕ_yaw) where ϕ_yaw denotes the angle between nacelle and wind direction and E_AEP the annual power production of a healthy turbine.

Algorithm logic

Filter data

Standard filter: 

• Filter stuck values  
• Filter values by physical normal values

Furthermore, algorithm specific filter are used 

• Filter out low and high power data points 
• relative wind direction exists 
• group performance index exists 
• Active Power is a realistic value 
• Region 3 of the power curve is filtered out by filtering all data points that show a power production above 90% of the nominal power production

Train Regression models

To estimate the performance index a regression model is trained that predicts the power of the current turbine. The following input features are used to train the regression model:

• yaw error (the angle between nacelle direction and absolute wind direction. also known as relative wind direction)
• wind speed
• pitch angle
• wind direction
• power production of 5 closest neighboring turbines

The output feature is the the power of the turbine. A random forest regressor is used. This ensures that non significant input features are neglected. The regression model neglects temperature, air pressure or air density measurements.

The model is trained on 1 year of data if available. The minimum amount of data necessary is 6 months.

to compute the group performance index the actual power is divided by the regressed power

I_perf = P_act/P_reg

where I_perf denotes the performance index, P_act the actual power and P_reg the regressed power. This index, expresses, thus, if the turbines performance is as expected based on the current operating point and the neighboring turbines.  The performance index is computed for the same time frame and the same data as the regression model was trained on.

This gives a performance index for each data point in the last 6 months or year which were not filtered out by the pre-algorithm data grooming. It is now possible to plot a distribution of the performance index over the yaw error.

The distribution of the group performance index over the yaw error as shown in the UI of Wind Advanced Analytic

Alarm Trigger Logic

The alarm trigger logic identifies if the group performance index is higher for a certain yaw error that is not 0.

An alarm is triggered if the maximum group performance index is at a yaw error that is larger than 5 or smaller than -5 (the exact value can be set in the threshold management) and the AEP loss is larger than 0 % (can be set in threshold management).

Before an alarm is triggered it is checked if the highest performance index value is statistically significant. For this, a t-test is executed that checks if the bin contains enough data points to make the difference to the expected value of cos²(θ_yaw) statistically relevant with a confidence level of 90%. Only if the test is passed the alarm is triggered.

AEP Loss Calculation

The annual energy production (AEP) loss is computed by fitting a weibull wind distribution to the recorded data. The expected annual power production can then be computed by apply the contractual local power curve to that weibull distribution. The expected power output for each yaw error can then be determined by apply the cos²(θ_yaw) correction factor mentioned above. The expected energy product loss is thus θ_yaw.

The AEP loss is computed for a given yaw angle. If the algorithm determines that a yaw misalignment of θ_yaw is occurring the above formula estimates that yaw misalignment for a static yaw error.