With the rapid growth of the new energy vehicle fleet, precise early warning for power battery charging safety has become a critical challenge requiring urgent solutions. This paper proposes an electric vehicle charging anomaly detection method based on voltage prediction residual analysis. The method leverages a CNN-LSTM-Attention hybrid model to predict the measured voltage values during the charging process, subsequently calculating the residuals between the predicted voltage and the actual measured voltage. Building on this, a two-tier threshold system including the anomaly threshold and the fault threshold, based on historical residual percentiles is established to enable early identification and risk stratification of battery voltage anomalies: when the residual exceeds the anomaly threshold, it indicates a potential thermal runaway risk; when it exceeds the fault threshold, it warns of an imminent failure. A vehicle-specific modeling strategy is introduced, constructing a dedicated prediction model based on the same anomaly detection method for each individual electric vehicle. Cross-vehicle testing results demonstrate that non-dedicated models lead to significant missed detections and false alarms, strongly validating the necessity and importance of vehicle-specific modeling for improving anomaly detection accuracy.