UrbanInsight: A Distributed Edge Computing Framework with LLM-Powered Data Filtering for Smart City Digital Twins

Kishor Datta Gupta, Md. Manjurul Ahsan, Mohd Ariful Haque, Roy George, Azmine Toushik WasiarXiv 2025 · DOI · arXiv

Introduction

The rapid expansion of smart city infrastructure is pushing centralized cloud-based data processing to its limits. Projections indicate that urban environments will generate over 190 zettabytes of data annually by 2025. Current "cloud-centric" architectures—where every raw sensor feed is transmitted to a central hub—face critical bottlenecks in latency, bandwidth costs, and data fragmentation.

The research paper "UrbanInsight: A Distributed Edge Computing Framework with LLM-Powered Data Filtering for Smart City Digital Twins" introduces a novel approach to address these challenges. By deploying Large Language Models (LLMs) at the network edge, UrbanInsight intelligently filters and prioritizes data at the source, ensuring that only contextually relevant information is transmitted to update the city's Digital Twin.

System Architecture

UrbanInsight shifts the intelligence from the cloud to the edge through a multi-layered distributed framework. This architecture is designed to handle diverse data streams while maintaining high fidelity in the central simulation model.

The UrbanInsight distributed architecture, showing the flow from data acquisition to the Digital Twin Core.

The framework is built upon four primary components:

  1. LLM-Powered Edge Filtering: Optimized LLMs (quantized for edge hardware) generate adaptive filtering rules based on real-time sensor data, historical trends, and specific urban policies.
  2. Physics-Informed Reasoning: To ensure reliability, the filtering engine is grounded in physical laws, preventing the AI from making decisions that contradict real-world dynamics (e.g., impossible fluid pressure spikes).
  3. Knowledge Graph (KG) Backbone: Data reaching the central hub is integrated into a semantic Knowledge Graph that maps spatial, temporal, and causal relationships between urban entities.
  4. Priority-Based Transmission: Data packets are assigned priority scores based on anomaly levels and semantic richness, ensuring critical alerts are processed with minimal delay.

Experimental Evaluation

Due to the privacy concerns and fragmentation of real-world urban data, the authors evaluated UrbanInsight using a comprehensive synthetic dataset. The evaluation environment simulated 5,000 multi-sensor observations across 100 virtual locations over a 30-day period, totaling 2.3 TB of raw data. This dataset included traffic flow, environmental metrics, CCTV metadata, and social media feeds, with various anomalies injected at a 2% rate to test detection capabilities.

Key Results

The experimental results demonstrate that UrbanInsight significantly outperforms traditional cloud-centric baselines in efficiency and accuracy.

Data reduction rates and accuracy retention across different sensor types.
  • Data Reduction: The system achieved an overall 88.28% reduction in transmitted data. Notably, CCTV metadata transmission was reduced by 92% without losing critical event information.
  • Accuracy Retention: Despite the massive data reduction, the system maintained high accuracy for downstream machine learning tasks. For instance, video metadata tasks retained 97% accuracy while using only 8% of the original data volume.
  • Anomaly Detection: The F1-score for anomaly detection improved from a baseline of 0.253 to 0.598, with a True Positive Rate of 94%.
  • Operational Efficiency: Monthly operating costs and total energy consumption were reduced by approximately 51%, largely due to the decreased load on network infrastructure.

Discussion and Future Work

UrbanInsight demonstrates that "context is king" in urban data management. By understanding that high traffic near a stadium after a concert is normal, the LLM-driven engine avoids the false positives common in traditional threshold-based systems.

However, several challenges remain for real-world deployment. The reliance on synthetic data means the system must still be tested against the "messiness" of actual urban environments, including sensor failures and unpredictable human behavior. Additionally, moving LLMs to the edge introduces new security considerations regarding the integrity of filtering rules. Future research will focus on hardening edge-to-cloud protocols and improving the model's robustness against missing or conflicting sensor inputs.

Summary updated

Abstract

Cities today generate enormous streams of data from sensors, cameras, and connected infrastructure. While this information offers unprecedented opportunities to improve urban life, most existing systems struggle with scale, latency, and fragmented insights. This work introduces a framework that blends physics-informed machine learning, multimodal data fusion, and knowledge graph representation with adaptive, rule-based intelligence powered by large language models (LLMs). Physics-informed methods ground learning in real-world constraints, ensuring predictions remain meaningful and consistent with physical dynamics. Knowledge graphs act as the semantic backbone, integrating heterogeneous sensor data into a connected, queryable structure. At the edge, LLMs generate context-aware rules that adapt filtering and decision-making in real time, enabling efficient operation even under constrained resources. Together, these elements form a foundation for digital twin systems that go beyond passive monitoring to provide actionable insights. By uniting physics-based reasoning, semantic data fusion, and adaptive rule generation, this approach opens new possibilities for creating responsive, trustworthy, and sustainable smart infrastructures.
BibTeX
@misc{gupta2025urbaninsight,
  author = {Kishor Datta Gupta and Md. Manjurul Ahsan and Mohd Ariful Haque and Roy George and Azmine Toushik Wasi},
  title = {{UrbanInsight: A Distributed Edge Computing Framework with LLM-Powered Data Filtering for Smart City Digital Twins}},
  year = {2025},
  eprint = {2509.00936},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.00936},
}
Figure