Primary Voltage Forecasting in Distribution Systems using Principal Component Analysis

This paper presents an accurate forecasting method for primary voltage measurements of the distribution feeder. The characteristics of the measurements associated with the distribution system, which is high dimensional, with noisy conditions and anomalous data, require a low-complexity algorithm for the predictive data analytics. To this end, this paper proposes a two-level framework using (1) AdaOja-based streaming principal component analysis for dimensionality reduction and anomaly detection and (2) Vector Auto Regression (AdaOja-PCA-VAR) model to forecast the low-dimensional projection of the measurement data. The dimensionality-reduction approach relies on a strong correlation observed in voltage measurements of the primary voltage data resulting from the physics of the power flow. The low dimensional PCA projection is used to detect and eliminate measurement anomalies and to obtain forecasts. The forecasts of the low-dimensional data are projected back to the original subspace using eigenvectors of the PCA to obtain the actual voltage magnitude forecasts. The proposed methods are validated using the IEEE test feeder with simulated data.