Prediction Of Fire Smoke Flow And Temperature Distribution Based On CFAST

The prediction of temperature and smoke layer height during a fire hazard is crucial in guiding an emergency rescue and fire analysis.Existing prediction models are not able to offer an accurate and reliable result due to the prediction errors and uncertainties of input parameters including heat release rate,venting rate etc.CFAST is a fire prediction model that is highly recognized among the field of construction and fire hydrants.Although various attempts to improve the model have been made,several flaws are yet to be eliminated,including the oversimplified experiment set-up,shortened prediction steps and the limitations of simulation systems.This thesis proposes a model of the prediction of fire smoke flow and temperature distribution based on CFAST to solve the problem of low prediction accuracy and cumbersome procedures of existing models.The model uses deep learning methods to train and predict the time series data of fire smoke flow and temperature distribution.Such model also extracts the fire trend feature of the time sequence to improve the accuracy of prediction.Furthermore,we have combined the deep learning methods with a data assimilation method to dynamically adjust the model.The major contributions are summarized as follows:(1)Different from the data assimilation method,which relies on real-time synchronization and correction of data,we use LSTM to learn,save and express latent knowledge and laws from previous records of fires,which improves our model’s overall accuracy and efficiency.In order to satisfy the amount of data required by the NNs,we combine multiple datasets from varied situations,and design the input and output of the network to establish the prediction model.(2)In order to further improve the performance,we have designed the LSTM-TFV algorithm.The trend characteristics of fire time series are extracted and analyzed by our algorithm.These characteristics are taken as prior knowledge to optimize the training process of deep neural network.(3)EnLSTM-TFV algorithm is proposed to improve the flexibility and accuracy of fire prediction model by combining deep neural network with improved data assimilation method.Improved data assimilation method is used to fuse observed values and predicted values.Compared with traditional step-by-step correction method of data assimilation,in our model,the time series only selects the time point of step length for data assimilation,while in other period,the trend features are used to update the Kalman gain to correct the prediction sequence.This research has made significant contributions to fire research by unveiling its origins and development patterns.It also provides additional information for firefighters to help people evacuate.Experimental result shows that the prediction algorithm proposed in this thesis improves the accuracy of the prediction of fire smoke movement and temperature distribution,and achieves efficient and convenient fire time series data prediction.