Deep learning framework for predicting measurement error drift in smart meter sensors under harsh coastal environments

Abstract

Smart meter sensors in coastal environments suffer severe accuracy degradation due to salt fog corrosion, high humidity, and temperature cycling affecting current transformers, voltage dividers, and measurement circuits. This study proposes a reliability prediction framework addressing multi-stress degradation in harsh marine atmospheres. The framework integrates an enhanced Transformer with multi-scale attention, a Bidirectional Long Short-Term Memory (BiLSTM) network for local degradation patterns, and a standard Transformer, unified through particle swarm optimization for dynamic weight adjustment. Validation using six years of field data comprising 10.5 million records from 200 sensors on Meizhou Island, China, demonstrates superior performance, achieving a coefficient of determination of 0.944 and a root mean square error of 0.0121%, representing a 6.9% improvement over the best individual model. To capture coastal-specific effects, two novel indices are introduced: the Salt Fog Corrosion Index, quantifying cumulative chloride deposition, and the Electrochemical Activity Factor, modeling electrochemical corrosion potential. Feature analysis identifies load current-temperature interaction with correlation 0.76 as the dominant drift mechanism, consistent with Joule heating theory. Field deployment verifies the framework’s ability to predict failures about 5 days in advance with 0.01 percentage point accuracy, enabling condition-based maintenance and reducing out-of-specification operation time by 85%. This research provides practical tools for sensor reliability in harsh environments beyond smart metering.