Dynamic scenario deduction analysis for hazardous chemical accident based on CNN-LSTM model with attention mechanism

The evolution of hazardous chemical accidents (HCAs) is characterized by uncertainty and complexity. It is challenging for decision-makers to expeditiously adapt emergency response plans in response to dynamically changing scenario states. This study proposes a data-driven methodology for constructing accident scenarios and develops a novel hybrid deep learning model for scenario deduction analysis. This model aids in accurately predicting the evolution of HCAs, enabling emergency responders to prepare and implement targeted interventions proactively. First, a framework for constructing an accident scenario database is presented, based on the time-sequential characteristics of accident progression. This framework employs a data-driven approach to describe the evolution process of accident scenarios. Second, a deep learning model (CNN-LSTM-Attention) that integrates convolutional neural network (CNN), long short-term memory (LSTM), and attention mechanism (AM) is developed for accident scenario deduction analysis. Finally, to illustrate practical application, a scenario database for HCAs is established. A major HCA case study is conducted to demonstrate the ability of this model to analyze various scenarios, thereby improving emergency decision-making efficiency. Compared with algorithms such as CNN, LSTM, and CNN-LSTM, the prediction accuracy of this method ranges from 86% to 93%, signifying an improvement of over 7%. This work provides a reliable framework for supporting decision-making in emergency management.