Abstract
<title>Abstract</title> <p>Modern power systems face escalating cyber threats that simultaneously target digital communication protocols and physical infrastructure, making real-time detection essential for secure grid operation. Current data-driven cyber attack detections rely on deterministic temporal data encoding and struggle with high sample complexity, overfitting. This paper proposes the Deep Twin-Stream Sparse Spatiotemporal Quantum Diffusion (STQD) framework for cyber attack detection in smart grid systems. The proposed STQD model proposes a sparse twin-stream spatiotemporal transformer autoencoder compresses high-dimensional sensor inputs into a lower dimensional latent representation using sparse spatiotemporal attention, adaptive gating fusion, and supervised contrastive learning. The extracted spatiotemporal representation are mapped to quantum Hilbert space via quantum encoding circuits. A quantum diffusion model learns the latent spatiotemporal embeddings probability distribution function through a quantum-native denoising process on density matrices, learning the class-conditional probability distribution of the encoded quantum states. Finally, a Quantum Support Vector Machine (QSVM) classifies the quantum spatiotemporal representations using a fidelity-based quantum kernel across three encoding strategies, including angle, amplitude, and instantaneous quantum polynomial. The proposed STQD framework is evaluated on two distinct publicly available smart grid benchmarks including Mississippi State University and Oak Ridge National Laboratory (MSU-ORNL) and the Smart Digital Substation (SDS) dataset, achieving weighted F1-scores of 0.95 and 0.90, respectively. Consistent performance across two architecturally diverse grid environments demonstrates the superior generalization capacity of STQD compared to state-of-the-art approaches.</p>