Spatio-Temporal Missing Data Imputation for Smart Power Grids

Sanmukh R. Kuppannagari, Yao Fu, Chung Ming Cheung, Viktor K. Prasanna

Abstract: Availability of high fidelity timeseries data is imperative for critical power grid operational tasks such as state estimation, DER scheduling, etc. However, the data obtained from the metering infrastructure is prone to disruptions due to communication outages leading to missing values. State-of-the-art smart power grid Missing Data Imputation (MDI) algorithms either operate on individual timeseries and are unable to capture spatial dependencies due to the power grid topology or they operate on the entire dataset, requiring complex models which lead to overfitting.In this work, we develop a novel technique to perform spatiotemporal missing data imputation. Using the power grid topology and timeseries data obtained from the metering infrastructure in the grid as input, we develop a Spatial-Temporal Graph Neural Network based Denoising Autoencoder (STGNN-DAE) that performs MDI by accounting for both temporal and spatial correlations. Using a real dataset obtained from a distribution test grid in Midwest, Iowa, we compare the proposed model with existing solutions for MDI. We show that our GNN based autoencoder obtains an improvement of 4.3% to 25.2% in error metrics such as Mean Absolute Error compared to the state-of-the-art missing data imputation methods.