A Spiking Neural Network–Driven Neuromorphic Digital Twin Architecture for Real-Time Resilient Resource Allocation in Smart Sports Clubs

Document Type : Original Article

Author

PhD in Sports Management, Department of Sports Management, Faculty of Sports Sciences, University of Mazandaran, Babolsar, Iran

Abstract
Smart sports clubs combine large conditioned spaces, volatile member demand, and growing IoT exposure, yet their energy, space, and storage resources are often managed using static rules. This paper presents NeuroTwin, a spiking neural network (SNN)-driven neuromorphic digital twin for real-time resource allocation that remains resilient to grid outages, equipment failures, demand surges, and cyberattacks while preserving member privacy. The proposed seven-layer cyber-physical architecture integrates IoT sensing with local differential privacy, event-driven spiking perception for occupancy detection, demand forecasting, and intrusion detection, a Spike-Gated Trust layer that isolates compromised sensors and substitutes twin-based estimates, and seven decentralized spiking agents for cooling control, space allocation, and battery management. Agents are trained through twin-in-the-loop privileged distillation using a receding-horizon MPC teacher and the DAgger framework. The digital twin is developed from publicly available datasets, including Building Data Genome 2, UC Berkeley gym occupancy records, the UCI Occupancy Dataset, and TON_IoT telemetry data. Experimental results show occupancy detection accuracies of 97.79% and 98.33% and an intrusion-detection macro-F1 of 0.95. Across 15 held-out weeks, NeuroTwin improved the weekly operational objective by 22.5% during grid outages and 17.8% under compound disruptions while maintaining comparable performance under nominal conditions. It also outperformed the strongest model-free MARL baseline during outages by 10.1%, generalized to a second club without retraining, and tolerated local differential privacy with ε = 2 without measurable performance loss. The seven spiking agents required only 386.1 nJ per hourly decision, significantly outperforming equivalent ANN and online MPC approaches in energy efficiency.

Keywords

Subjects

  • Receive Date 04 October 2026
  • Revise Date 09 November 2026
  • Accept Date 16 November 2026
  • First Publish Date 16 November 2026
  • Publish Date 01 January 2026