Metering Automation System 3.0 Base Version Based on Machine Learning

Abstract

The accurate identification of equipment base versions in Metering Automation System 3.0 (MAS 3.0) is critical for ensuring interoperability and maintenance efficiency in modern smart grids. However, traditional machine learning methods and standalone deep learning architectures struggle to balance spatiotemporal feature extraction, computational efficiency, and deployment constraints for high-frequency multivariate metering data. This study proposes a hybrid DSCNN-CBAM-BiLSTM framework that synergistically integrates depthwise separable convolutions, dual attention mechanisms, and bidirectional temporal modeling to address these challenges. The depthwise separable convolutional neural network (DSCNN) minimizes parameter overhead while capturing spatial correlations across distributed grid nodes, followed by convolutional block attention modules (CBAM) that dynamically recalibrate channel and spatial features to amplify discriminative patterns. Bidirectional LSTM (BiLSTM) layers then model long-range temporal dependencies in both forward and backward directions, enabling robust contextual analysis of energy consumption sequences. Validated on 14 TB of operational data from China Southern Power Grid, the framework achieves 96.7% classification accuracy with an inference latency of 8.9 ms—outperforming CNNs (89.2%), Transformers (90.5%), and GRUs (92.1%) while reducing GPU memory usage by 35.7–72.7%. Edge deployment tests on NVIDIA Jetson AGX Xavier demonstrate real-time compatibility with IEC 61850-7-420 protocols, maintaining <15 ms latency at 200-node resolution. These advancements establish a highly effective and resource-efficient framework. For resource-efficient, edge-deployable analytics in smart grid infrastructure, effectively bridging the gap between high-accuracy version identification and industrial computational constraints.