Evaluating LSTM-Based Surrogate Models for Predicting Productivity at UBGH2-6 Gas-Hydrate Site

다국어 초록 (Multilingual Abstract)

Data-driven, site-specific surrogate models are developed to enable rapid and accurate prediction of gas and water production at the UBGH2-6 gas-hydrate site. Synthetic datasets are generated under various depressurization scenarios. Considering the high computational cost and time requirements of full-physics simulations, long short-term memory (LSTM) networks are employed as surrogate models to efficiently approximate production behavior. LSTM surrogate models per phase (water and gas) are trained on datasets of differing sizes—basic and expanded—with the expanded dataset encompassing a broader range of operational conditions. Water-production predictions are highly accurate even under limited input features, whereas gas-production predictions improve significantly when trained on the expanded dataset. Sensitivity analyses highlight critical architectural components, including batch normalization, learning rate, and batch size, as key drivers of model performance.
These findings indicate that LSTM-based surrogates are a computationally efficient and reliable alternative to conventional full-physics simulations for forecasting production in complex gas-hydrate reservoirs.