Smart meter health state prediction based on residual-connected self-attention ConvLSTM network.

Accurate health state prediction of smart meters is essential for ensuring metering reliability and enabling condition-based maintenance in advanced metering infrastructure. Existing approaches based on conventional machine learning or standard deep learning architectures struggle to simultaneously capture complex spatiotemporal degradation patterns, model long-range dependencies, and maintain stable training in deeper network configurations. This paper proposes a health state prediction framework built upon a Residual-Connected Self-Attention Convolutional Long Short-Term Memory (ConvLSTM) network. The framework introduces residual shortcut connections across stacked ConvLSTM layers to mitigate gradient vanishing and network degradation, and integrates a self-attention mechanism to dynamically recalibrate spatiotemporal feature representations according to their diagnostic relevance. An end-to-end pipeline encompassing multi-dimensional feature construction, data preprocessing, and health state classification is developed. Experiments conducted on a real-world dataset of 12,860 smart meters over 36 months show that, under the tested provincial grid conditions, the proposed model attains 94.38% overall accuracy and 89.94% macro-F1, outperforming LSTM, ConvLSTM, CNN-LSTM, and Transformer baselines. Ablation studies confirm the synergistic contribution of both self-attention and residual connections, while attention weight visualisations cross-verified by SHAP-based attribution reveal that the model autonomously concentrates on critical degradation signatures. We acknowledge that broader cross-regional validation remains an open task and outline this together with computational scalability and interpretability as the principal directions for future work.