Multi-horizon probabilistic shale gas production forecasting using VAE-GAN

Applying Deep Learning (DL) methods that utilize historical data from wells drilled in similar formations has proven to be a powerful approach to production forecasting. Nevertheless, most of these DL models are deterministic and do not represent the uncertainty encapsulated by the range of possible outcomes, leading to potentially misleading predictions. Existing probabilistic models, such as quantile regression, provide an incomplete representation of the variability, while Bayesian models focus on parameter uncertainty rather than directly modeling the data distribution, and their predictions can be overly influenced by prior assumptions. This limitation highlights the need for a robust model to accurately learn the underlying data distributions from the training dataset, thereby providing predictions that align more closely with the inherent variability in the data. To overcome these limitations, we develop a tailored Variational Autoencoder with Generative Adversarial Network (VAE-GAN) tailored for probabilistic production forecasting. The model employs a hybrid loss function that combines adversarial, reconstruction, and latent regularization terms. The adversarial component encourages learning the data distribution by implicitly minimizing the Jensen–Shannon (JS) divergence. The model incorporates historical production through a history embedding component, which employs a Gated Recurrent Unit (GRU) to capture temporal dependencies and uses hidden states to propagate conditioning information across components. It applies a non-generative approach for multi-horizon forecasting, eliminating the accumulation of errors typically encountered in rolling predictions. The model generates confidence intervals by applying Monte Carlo sampling to the latent vector, enabling a comprehensive range of possible outcomes. The proposed model has been validated to generate data distribution closely resembling real data, providing a reliable representation of potential scenarios. Our study utilizes data from 6,605 wells in the Montney shale gas play, incorporating well features and a 5-month production history to predict the future production curve over 55 months. The proposed model generates highly realistic production curves and representative prediction distributions, demonstrating superior performance on multiple evaluation metrics.