Dynamic prediction of multisensor gas concentration in semi-closed spaces: A unified spatiotemporal inter-dependencies approach

Abstract

Flammable gas leakage in industrial environments poses significant risks to human health and environmental safety. Developing accurate and efficient spatiotemporal models for gas dispersion is essential for mitigating these dangers. The gas diffusion is inherently a spatiotemporal process, yet most research has focused on modeling spatial or temporal correlations separately, failing to capture the dynamic relationships between various variables. To overcome this limitation, we propose a novel approach based on FourierGNN, a multivariate time series forecasting (MTS) method, which treats concentration values from multiple sensors as multivariates and predicts their future trends. By utilizing a fully-connected hypervariate graph structure, the model adaptively learns high-resolution representations across different timestamps and variates simultaneously. Experimental data are generated by simulating a methane leak scenario in a semi-closed gas turbine enclosure using computational fluid dynamics (CFD) software. The method is evaluated on the dataset with three distinct prediction horizons and compared with FC-LSTM and StemGNN. Results indicate that the approach outperforms others in terms of MAPE, MAE and RMSE across different prediction horizons while reducing parameter counts by 61.26% and 82.83%, respectively. Furthermore, the method demonstrates robustness under varying noise levels, confirming its reliability.