Robust data filtering in wind power systems By: Andrés Llombart-Estopiñán CIRCE Foundation – Zaragoza University
Index Objective Introduction: the need of filtering The LMS fitting technique The LMedS methodology Experimental results Conclusions
Index Objective Introduction: the need of filtering The LMS fitting technique The LMedS methodology Experimental results Conclusions
Objective To assess the performance of the Least Median of Squares method when it is used to filter wind power data
Index Objective Introduction: the need of filtering The LMS fitting technique The LMedS methodology Experimental results Conclusions
Introduction Why it is needed? Operation Maintenance Production Control Characterization of the P – v curves High quality P – v data
Introduction Circumstances that affect the data quality Sensor accuracy EMI Information processing errors Storage faults Faults in the communication systems Alarms in the wind turbine etc
Introduction An example of P – v data
Introduction P – v data after considering the SCADA alarms
Index Objective Introduction: the need of filtering The LMS fitting technique The LMedS methodology Experimental results Conclusions
The LMS fitting technique Gets the curve that minimizes the Mean Square Error All measurements can be interpreted with the same model Very sensitive to outliers Breakdown of 0% of spurious data
The LMS fitting technique
Index Objective Introduction: the need of filtering The LMS fitting technique The LMedS methodology Experimental results Conclusions
The LMedS fitting technique It is based in the existence of redundancy LMedS method uses the Median whereas the LMS method uses the mean Unfortunately the LMedS method don’t have analytical solution
The LMS fitting technique 1 2 3 4 5 6 8 7
The LMedS fitting technique Example Fitting with a polynomial with 4 coefficients n measurements m possible solutions, where
The LMedS fitting technique Steps to get the fitting: Calculate the m subsets of the minimum number of measurements required to fit your curve For each subset S, we compute a power curve in closed form PS For each solution PS, the median MS of the squares of the residue with respect to all the measurements is computed We store the solution PS which gives the least median MS
The LMedS fitting technique Rejection of wrong data: Estimate de standard deviation Probability of accepting a measure being good: 99 % Threshold = 2.57 s
Index Objective Introduction: the need of filtering The LMS fitting technique The LMedS methodology Experimental results Conclusions
Experimental results Methodology A year of historical data 5 different tests Alarm Records (AR) AR + classical statistic method AR + robust statistic Classical statistic Robust statistic
Experimental results Rough data Considered Alarms
Experimental results AR + Class. Stat AR + Robust Stat.
Experimental results Classic Stat. Robust Stat.
Index Objective Introduction: the need of filtering The LMS fitting technique The LMedS methodology Experimental results Conclusions
Conclusions A robust filtering method has been proposed It has been proved successfully The method have shown a good robustness Some research is needed Considering the wind direction
Robust data filtering in wind power systems Thanks for your attention
The LMedS fitting technique Example: fitting a polynomial of 4 coefficients for a 3 months period of data, that implies ~ 12.750 data The computational cost is huge
The LMedS fitting technique Solution: selecting randomly subsets Compromise: Minimizing the number of subsets Warranting a reasonable probability of not failing So, the first method step is substituted by a Monte Carlo technique to randomly select k subsets of 4 elements
The LMedS fitting technique How many subsets? A selection of k subsets is good if at least in one subset all the measurements are good Pns is the probability that a measurement is not spurious Pm is the probability of not reaching a good solution
The LMedS fitting technique In our example considering: Pns = 75 % Pm = 0,001