DeepONet-embedded physics-informed neural network for production prediction of multiscale shale matrix–fracture system

Physics of Fluids, Journal Year: 2025, Volume and Issue: 37(1)

Published: Jan. 1, 2025

As a rising method for reservoir-scale production analysis, machine learning (ML) models possess high computational efficiency with robust capability of nonlinear mapping. However, their accuracy and interpretability are commonly limited owing to the absence intrinsic physical mechanisms, solely by data fitting. This work proposes novel DeepONet-embedded physics-informed neural network (DE-PINN), which comprises forward connect matrix/fracture characteristics performance, sampling acquire location points within shale reservoirs. DeepONets constructed selected layers these networks output field variables in governing equations that include mass/momentum conservation coupled multiscale transport mechanisms. Through automatic differentiation method, solved obtained variables, residuals generated during solution integrated into loss function as constraints. Compared traditional data-driven models, DE-PINN exhibits better performance forecasting rate cumulative production, achieving mean absolute percentage error (MAPE) approximately 3% adjusted R2 values test set exceeding 0.98. model demonstrates advantage realizing superior predictive precision fewer samples under complex geological conditions

Language: Английский

A Study on a Novel Production Forecasting Method of Unconventional Oil and Gas Wells Based on Adaptive Fusion

Processes, Journal Year: 2024, Volume and Issue: 12(11), P. 2515 - 2515

Published: Nov. 12, 2024

Reliable forecasting of unconventional oil and gas well production has consistently been a hot challenging issue. Most existing data-driven models rely solely on single methodology, with the application effects other mainstream algorithms remaining unclear, which to some extent hinders generalization utilization these models. To address this, this study commences data preparation systematically develops novel model based adaptive fusion multiple such as random forest support vector machine. The validity is verified using actual wells in Marcellus. A comprehensive evaluation feature engineering extraction techniques concludes that main controlling factors affecting Marcellus are horizontal segment length, fracturing fluid volume, vertical depth, section, reservoir thickness. Evaluation primary reveal significant differences prediction performance among methods when applied dataset. newly developed optimized by genetic outperforms individual across various metrics, can effectively improve accuracy forecasting, demonstrating its potential for promoting production. Furthermore, will assist enterprises allocating resources more effectively, optimizing strategies, reducing costs stemming from inaccurate predictions.

Language: Английский

How to Optimize the Investment Decision of Shale Gas Multi-Objective Green Development for Energy Sustainable Development?

Published: Jan. 1, 2024

Download This Paper Open PDF in Browser Add to My Library Share: Permalink Using these links will ensure access this page indefinitely Copy URL DOI

Language: Английский