EVCA-Net: An MDN-Based Explainable Framework for Robust Secure Profiling in Smart EV Charging Infrastructure

The rapid proliferation of Electric Vehicle (EV) charging infrastructure within smart grid ecosystems has exposed these cyber-physical assets to complex threats that compromise data privacy, billing integrity, and hardware safety. Conventional intrusion detection mechanisms struggle with non-linear attack vectors, lack structural interpretability, and impose high computational overheads. To overcome these challenges, this paper presents EVCA-Net, a novel, high-performance hybrid expert system framework for automated EV charger threat profiling and explainable security diagnostics. Validated on physical Raspberry Pi-managed control unit telemetry using the CIC-EVSE-2024 dataset, EVCA-Net deploys an advanced preprocessing pipeline and a Transformer-driven Mixture Density Network (T-MDN) module to extract deep multimodal spatiotemporal features. To optimize computational efficiency, a hybrid feature-selection strategy combines Shapley Additive exPlanations (SHAP) with Marine Predator Optimization (MPO) to isolate critical threat indicators. Classification is executed using a Deep Additive Neural Network (DANet) integrated with Local Interpretable Model-agnostic Explanations (LIME), delivering transparent, operator-verifiable security maps. EVCA-Net achieves 99.95% accuracy along with 99.639% precision, 99.266% recall, and 99.452% F1-score, establishing a robust, scalable, and fully explainable decision-support paradigm for modern EV charging infrastructure.