A collaborative research team from The Hong Kong Polytechnic University and Guangzhou Institute of Industrial Technology has published a study in Engineering, outlining a self-driven intelligent agent framework that unites spatiotemporal fire prediction deep learning and domain-tuned large language models (LLMs) to address persistent limitations of conventional smart firefighting systems. Existing data-driven fire forecast platforms face two core practical barriers: sensor network vulnerability under harsh fire conditions and rigid output formats that fail to adapt to varied, real-time demands of trapped occupants, firefighters and incident commanders. The work constructs a closed-loop hybrid AI workflow to deliver scenario-specific, executable emergency guidance without manual model operation or data translation.

The study first introduces ConvLSTM-Fire, a fire-resilient spatiotemporal deep learning architecture built by integrating convolutional neural networks (CNNs) for spatial feature extraction and long short-term memory (LSTM) layers for temporal sequence modeling. The model targets two core sensing challenges in building fire environments: data remediation for damaged temperature sensors and inversion of discrete single-point sensor readings into continuous two-dimensional temperature contour slices. When temperature monitoring nodes lose signal due to heat, smoke or structural collapse, ConvLSTM-Fire reconstructs missing temperature values by mapping historical measurement correlations between intact sensors and failed units; normalized incomplete sensor time series matrices serve as input, with convolutional layers extracting spatial patterns and LSTM modules capturing time-dependent thermal evolution before fully connected layers output repaired temperature readings. Deconvolution layers further upscale one-dimensional sensor data into full spatial temperature distribution maps, forming the real-time situational awareness foundation for subsequent LLM processing. Model training relies on benchmark datasets generated via Fire Dynamics Simulator (FDS) 6.7 computational fluid dynamics simulations, covering 60 distinct fire scenarios with variable ignition positions, heat release rates and ventilation layouts, and the model’s predictive reliability is further validated against full-scale experimental fire tests conducted at the Guangzhou Institute of Industrial Technology.

To resolve the limited actionability of generic LLMs for fire incidents, the research designs a self-driven agent core centered on an LLM benchmarked against ChatGPT-4o, fine-tuned with fire engineering domain knowledge including ConvLSTM-Fire invocation parameters and building fire safety standards. Unlike standalone LLMs that generate generalized guidance without live on-site thermal data, the hybrid agent establishes bidirectional links between the language model and the ConvLSTM-Fire sensing module through an IoT environmental database. When users submit natural language queries via wearable, handheld or graphical user interface devices, the LLM classifies requests to separate general knowledge inquiries from tasks requiring real-time fire field data. Scenario-dependent demands trigger automatic sequential task decomposition and executable Python code generation; the self-driven framework runs these scripts within restricted subprocess environments to call ConvLSTM-Fire, retrieve reconstructed temperature maps, compute aggregated thermal metrics and segment hazardous zones by predefined temperature thresholds, with runtime error correction embedded to reprocess faulty code outputs.

Three distinct response modes are quantitatively evaluated to measure the impact of dynamic fire situational inputs on LLM output quality: LLM-Generic relying solely on static domain knowledge, LLM+6-Sensors supplied with raw IoT temperature readings, and LLM+2DTemp paired with complete ConvLSTM-Fire-generated temperature contours. Evaluations adopt BLEU, ROUGE and BERT-based cosine semantic similarity metrics to compare model outputs against manually compiled standard emergency response plans aligned with Hong Kong’s fire safety building code. The test results indicate that integrating full two-dimensional temperature field data improves lexical matching and semantic alignment between LLM recommendations and reference operational protocols, as thermal spatial information enables granular room-level risk classification, and targeted firefighter entry guidance unavailable from generic or single-point sensor-only inputs. Voice interaction functions within the graphical interface lower operational cognitive burden for responders under high-stress emergency conditions, translating visualized temperature field analysis into spoken safety instructions tailored to each stakeholder’s role.

The paper acknowledges existing boundaries of the proposed framework, noting that multi-sensor simultaneous failure scenarios require further ConvLSTM-Fire structural optimization, while ongoing work will target reduced model inference latency and calibrated sensor survival thresholds to activate data remediation mechanisms more accurately. The hybrid agent architecture establishes a reusable engineering paradigm for AI-powered smart firefighting, balancing robust real-time fire environmental perception with flexible natural language human–AI interaction to advance deployable intelligent emergency response systems for smart buildings and urban disaster management.

The paper “Integrating Smart Fire Forecast with LLM-Powered Emergency Response,” is authored by Weikang Xie, Yuxin Zhang, Tong Lu, Xianjia Huang, Jihao Shi, Xinyan Huang, Fu Xiao, Asif Usmani. Full text of the open access paper: https://doi.org/10.1016/j.eng.2026.02.023. For more information about Engineering, visit the website at https://www.sciencedirect.com/journal/engineering.