Adaptive spatiotemporal feature fusion attention-enhanced ConvLSTM for smart meter lifespan assessment

Accurate lifespan assessment of smart meters is essential for condition-based maintenance and efficient asset management in modern power grids. Existing approaches struggle to jointly exploit the heterogeneous, multi-source nature of meter operational data and the intertwined spatial and temporal degradation patterns that govern meter aging. This paper proposes an end-to-end framework that integrates a multi-scale Convolutional Long Short-Term Memory (ConvLSTM) network with an adaptive spatiotemporal attention fusion mechanism for remaining useful life prediction. Multi-source meter data are first organized into spatiotemporal tensors, which are then processed by dual-branch ConvLSTM streams with different receptive fields to capture both localized and regional degradation correlations. A channel attention sub-module and a spatial attention sub-module subsequently recalibrate feature importance, and a learned gating unit dynamically balances their contributions according to each sample’s degradation characteristics. Experiments on a real-world dataset comprising 12,580 smart meters over eight years of field operation show that the proposed model achieves an RMSE of 4.62 months and an \:R² of 0.972, outperforming recent Transformer and graph-neural-network baselines as well as vanilla ConvLSTM (by 29.8%) and standard LSTM (by 45.4%) in RMSE. Beyond fixed-weight fusion schemes adopted in earlier attention-augmented ConvLSTM designs, the gating unit acts at the sample level, redirecting the model toward whichever attention pathway is more diagnostic for a given degradation phase. Ablation studies confirm that every proposed component contributes measurably; statistical-significance testing across multiple training seeds further indicates that the gains over the strongest baseline are non-trivial. Robustness tests demonstrate resilience to both injected sensor noise and held-out operating regions, and the trained network runs in under a millisecond per sample on a GPU—and under 30 ms on a low-cost Raspberry Pi—so fleet-scale screening stays practical even at the grid edge.