Consequence prediction using variable-length concentration time series for gas turbine enclosure

Flammable gas leakage in a semi-enclosed scenario can lead to catastrophic consequences, such as vapor cloud explosions. To reduce casualties and environmental damage, predicting the consequences based on the initial concentration time series monitored by sensors is of paramount importance. This paper proposes a consequence prediction model based on deep learning using variable-length concentration time series. Incomplete concentration values are padded and then passed through a masking layer, enabling the network to focus exclusively on valid data. The temporal correlations are extracted using a long short-term memory (LSTM) network, and the final prediction results are obtained by passing these features into a feedforward neural network (FNN). Computational fluid dynamics (CFD) software was used to simulate the leakage of hydrogen-mixed natural gas. Experiments were carried out for nine distinct prediction targets, derived from combinations of the mass and centroid coordinates of vapor clouds formed by various gases. These prediction targets were modeled using both fixed-length and variable-length input sequences. The high accuracy of the experimental results validates the effectiveness of the proposed method.