Data Cleaning for Wind Turbine Systems Based on Iterative learning and Neural Network

The power curve is an important indicator for evaluating the power generation performance of wind turbine systems. Equipment failure and human intervention are the main factors leading abnormal data of power curve, resulting in serious deterioration of the data availability. In order to solve the negative impact of abnormal data, a novel data cleaning framework is proposed for wind turbine systems. First, the nonlinear regression model between the wind speed and the power is established by the back propagation (BP) neural network. Then, the distances between the real data and the nonlinear regression model are approximated by employing the designed auxiliary points. The candidate abnormal data and the current normal data are separated by a threshold judgment strategy after sorting all the distances in descending order. At the same time, based on the set of candidate abnormal data, the candidate normal data is selected through another threshold judgment strategy after sorting all the distances in ascending order. On the basic of both the current normal data and the candidate normal data, the neural network based on iterative learning is applied to update the nonlinear regression model. Finally, the effectiveness of the proposed method is verified upon two sets of data collected by real wind turbine systems.