Scientific machine learning based framework to forecast time evolving pressure signals during hydrogen leaks

Accurate forecasting of pressure signals over time, such as those recorded during hydrogen tank leaks, is crucial for safety assessment and risk mitigation. This work presents a novel methodology for predicting pressure signal evolution over time in scenarios relevant to hydrogen leaks by integrating Computational Fluid Dynamics (CFD) for synthetic data generation with scientific machine learning for forecasting. Our approach focuses on predicting both the maximum pressure signal of the hydrogen jet and pressure signals at different spatial locations. To enhance forecasting accuracy, we apply a Fast Fourier Transform (FFT)-based smoothing technique to reduce noise in the pressure signal. We then evaluated multiple machine learning architectures, including linear models, dense neural networks, LSTMs, and CNNs, achieving high accuracy in forecasting the smoothed maximum pressure signal. We extend our approach to two multivariate datasets containing pressure measurements from 4 and 72 spatially distributed probes at varying distances from the hydrogen leak source. Results demonstrate high predictive accuracy of our framework even at a multivariate scenario, providing a robust tool for predicting pressure signals at hydrogen leak scenarios and establishing the foundation for real-time safety applications in hydrogen systems. • Machine Learning based hydrogen leak forecasting using CFD-generated pressure signal data. • FFT-based denoising to remove low-energy frequencies and improve forecast accuracy. • Compared ML models (linear, DNN, LSTM, CNN) for pressure evolution prediction accuracy. • Extended to multivariate data with 4 and 72 probes at varying distances.