Learning-based adaptive imputation method with kNN algorithm for missing power data

Abstract

We address the imputation of missing power consumption data in AMI. As the power consumption data are collected from more various power consumers, we propose a method to improve imputation accuracy by improving limitations of existing methods. In detail, we propose a method selection that takes into account the variability of the missing situation. Based on past similar situations, the kNN classification algorithm is used to select a more appropriate imputation method between the linear interpolation method and the historical average method. Next, we propose a method to select historical data useful for imputation and improve the existing historical average method based on kNN regression algorithm. Finally, it is shown through actual measured power data that the imputation accuracy is improved by applying the proposed method.