Predictive Health Management of Smart Meters: Daily Measurement Error Forecasting Under Complex Environmental Conditions

Abstract

Daily measurement error (ME) forecasting is critical for the health management of smart meters (SMs) under complex environmental conditions. This paper proposes a tailored short-term ME prediction framework employing Gaussian Process Regression (GPR) enhanced by a Weighted Automatic Relevance Determination (WARD) kernel and time-frequency feature augmentation. A dual constraint screening mechanism using Pearson Correlation Analysis (PPMCC) and Pareto Smoothed Importance Sampling (PSIS) is introduced to optimize input features. To further improve predictive capabilities, an Adaptive S-transform (AST) decomposes ME, capturing time-frequency information for GPR input enhancement. Experimental validation with real-world SM data under extreme conditions demonstrates that the proposed AST-MKGPR(WARD) model achieves superior interpretability and predictive accuracy compared to state-of-the-art approaches, offering a robust solution for daily SM health assessments.