Anomaly detection for real-world electric vehicle charging data using a convolutional autoencoder with multiscale convolution, attention mechanism, and BiLSTM

With the rapid expansion of electric vehicle (EV) charging, charging safety has become increasingly critical. Real-world charging data exhibit complex characteristics, including high dimensionality, nonlinearity, and strong volatility, while existing anomaly detection methods remain immature. To improve the detection of anomalies in complex EV charging data, this study proposes a method based on an enhanced convolutional autoencoder (CoAE), MA-BiLSTM-MCoAE. The model builds upon the traditional CoAE by integrating a multi-scale convolutional autoencoder (MCoAE), bidirectional long short-term memory (BiLSTM) networks, and multi-head attention (MA) mechanisms, enabling it to effectively capture the intricate temporal and spatial patterns in charging data. An exponential weighted moving average (EWMA) is then introduced to optimize error processing. Experiments are conducted on real charging data and alarm cases from multiple vehicle types. The results show that, compared with the baseline CoAE model, the mean squared error (MSE) and mean absolute error (MAE) are reduced by 95.85 % and 83.47 %, respectively, the inference time for a single charging session reaches 17.69 ms, and potential over-temperature and voltage anomalies can be detected 8.3 min and 11.55 min in advance. The proposed method provides a novel solution for high-dimensional time-series feature extraction, multi-target collaborative prediction, and noise-suppressed anomaly warning, offering broad prospects for engineering applications.