Online electric vehicle charging detection based on memory-based transformer using smart meter data

The increasing adoption of Electric Vehicles (EVs) presents significant challenges for electricity grid management, particularly in accurately identifying EV charging patterns to optimize distribution network planning. Most recent works focus on EV load forecasting for future grid planning and do not take into account real-time EV charging identification. This paper introduces a novel unsupervised memory-based transformer (M-TR) model for real-time detection of EV charging events from streaming smart meter data. The proposed method is designed for deployment by Distribution Network Operators (DNOs) and can be integrated into smart grid management systems to enhance grid stability and demand response strategies. The M-TR model leverages an encoder-decoder architecture: the encoder captures coarse-scale historical patterns over an extended temporal window, while the decoder focuses on fine-scale details within a local window. A key innovation of this approach is the dual memory comparison mechanism, which enables faster real-time inference by dynamically integrating historical and recent data. Unlike existing methods, the M-TR requires no prior knowledge of EV charging profiles and only uses real power consumption data from non-EV users. The goal of this approach is to automatically identify EV charging events from behind the meter aggregated smart meter data without requiring labeled EV data or additional sensors, thereby enabling real-time monitoring, reducing grid uncertainties, and supporting load balancing strategies. Benchmarking against state-of-the-art unsupervised models, the M-TR achieves superior performance with an average F-score of 0.844 and an AUC of 0.910, demonstrating its accuracy and robustness. Additionally, the model’s efficient execution time of 1.2 s per 1-min data sample highlights its suitability for real-time applications. This work underscores the importance of incorporating global temporal context and dual memory mechanisms in EV charging detection, offering a scalable solution for integrating EVs into the power grid. • An online unsupervised learning framework for real-time EV charging identification using aggregated smart meter data. • The proposed approach is based on anomaly detection method, using only smart meter data without EV which are abundant. • A memory-based transformer that effectively captures both long and short temporal information to enhance EVs charging detection. • Dual memory compression is proposed to ensure the model operates with linear time complexity, enabling real-time performance. • Streaming peak-over-threshold (SPoT) is proposed to dynamically define threshold values.