Multi-horizon well performance forecasting with temporal fusion transformers

Abstract

Forecasting fluid flow in subsurface resources such as groundwater, geothermal, and oil and gas is essential to maximize project economics and maximize resource recovery. We propose a novel workflow for a data-driven surrogate flow model for subsurface forecasting using an attention-based deep neural network architecture. The proposed workflow allows the inclusion of static and dynamic predictor features and forecasting across multiple wells with low error predictions. The study reveals that the proposed workflow can be applied over a wide range of subsurface resources and can improve well productivity, maximize resource recovery, and project economics. The proposed workflow also provides a more interpretable model with diagnostics to evaluate the quality of the resulting model. This study contributes to the development of new methods and workflows to integrate widely available historical production data into multi-step time series forecasting, which can be applied to subsurface resources beyond the ones studied here.

Keywords

• Deep learning
• Temporal fusion transformer
• Machine learning
• Interpretability
• Explainable AI
• Multi-well forecasting