(62a) A Novel Prediction Method of the Chemical Leakage Range with Artificial Intelligence: Machine Learning for Chemical Species

This study focuses on the prediction of the chemical leakage range by using artificial intelligence (AI). A multi-variable regression model is developed by deep neural network (DNN) method. The DNN model is formulated by Python with Tensorflow library. The model contains 12 input variables, 70 hidden layers, and 3 output variables. Each layer has Rectified Linear Unit (ReLu) function as an activation function and layers are fully connected. All weights are initialized by Xavier initializer. The chemical leakage range data set is generated by the commercial software, ALOHA. The simulation results show the hazardous zones (AEGL-1, AEGL-2, AEGL-3) under various conditions. To train the AI, the data set is configured with variables: chemical species, wind speed (1~7 m/s), temperature (10~35 ℃), ground roughness (open country, open water), stability classes (A~F), and tank storages (200~4,000 kg). Total date sets of each chemical are 30,240, which are divided into 25,920 training data and 4,320 test data. Training data set is the given information to artificial intelligence and the test data set is unknown information. As the results, the developed model performs high accuracy over 95 % on average for all tested chemicals. The result shows that artificial intelligence figure out the regularity of the data on its own and make predictions with high accuracy. The developed model can be extended to predict chemical leakage ranges for unknown chemicals as well as chemical mixtures by using chemical property information as the input data. The same method could also be used for quick estimation of fire and explosion accidents. This kind of research is expected to contribute to real-time response of corresponding organization in the chemical process safety field.

This research was supported by Disaster-Safety Platform Technology Development Program of the National Research Foundation of Korea(NRF) funded by the Ministry of Science and ICT. (NRF-2017M3D7A1085361)