Quantifying the "Planning-Practice Gap" in Urban Resilience: Validating Taipei’s Disaster Governance via HSEEP and Unsupervised Neural Networks

Drafting a governance framework for urban resilience is merely the first step; validating its operational feasibility under dynamic disruption is where the real challenge lies. This study addresses the critical gap between policy planning and operational practice within the Taipei City Government (TCG).

First, we constructed a high-granularity Business Continuity Plan (BCP) predicated on a worst-case scenario: a Magnitude 6.6 earthquake along the Shanchiao Fault. Aligned with ISO 22320 and the Sendai Framework, the plan categorizes mission-critical functions into nine operational chapters with prioritized recovery timelines. It incorporates Naismith’s Rule for realistic personnel mobilization and establishes a multi-tier resource reserve system.

To rigorously measure the resilience of these protocols, we implemented the U.S. HSEEP guidelines, aggregating outcomes from 17 diverse tabletop exercises (TTX) conducted across Taipei (2024-2025). This comprehensive dataset includes six critical sessions focused on compounding disruptors: acute manpower shortages and communication blackouts.

Addressing the limitation of subjective feedback in traditional governance assessments, we propose a novel quantitative framework. We integrated semantic text mining (SentenceTransformer) with Self-Organizing Maps (SOM) to process 636 unstructured feedback entries. This data was projected onto a topological map, distilling complex responses into quantifiable spatial clusters.

The AI-driven analysis revealed a critical divergence: while the BCP policy emphasized physical redundancies (multi-site backups), the neural network identified "Information Systems and Tools" as the dominant bottleneck in the cognitive map of participants. This finding highlights a hidden vulnerability in inter-agency data integration that traditional reporting missed. By coupling rigorous BCP formulation with unsupervised machine learning, this research offers a reproducible methodology for transforming subjective observations into objective, actionable data for urban risk governance.