Reservoir parameters prediction based on spatially transferred long short-term memory network

Area of Science:

• Petroleum Geoscience
• Artificial Intelligence in Exploration
• Machine Learning for Reservoir Engineering

Background:

• Reservoir reconstruction is critical for oil and gas exploration, heavily relying on accurate parameter prediction.
• Traditional petrophysical models are being superseded by deep learning for improved accuracy and efficiency in parameter prediction.
• Deep learning faces challenges due to the high cost and technical difficulties of acquiring sufficient data for complex geological parameters.

Purpose of the Study:

• To address the data acquisition limitations in deep learning for reservoir parameter prediction.
• To propose and validate a novel transfer learning prediction model for enhancing reservoir exploration.
• To improve the accuracy and efficiency of reservoir parameter prediction in data-scarce environments.

Main Methods:

• Development of a transfer learning prediction model utilizing long short-term memory (LSTM) neural networks.
• Determination of the optimal model structure through parameter search and optimization techniques.
• Knowledge transfer from historical datasets to new well prediction by sharing neural network parameters.

Main Results:

• The proposed transfer learning model demonstrated significant improvements in reservoir parameter prediction accuracy.
• Effectiveness was validated through comparative analysis on two distinct block datasets.
• The method successfully mitigated the challenges posed by data shortage in deep learning applications.

Conclusions:

• Transfer learning with LSTM networks offers a viable solution for data-scarce reservoir exploration.
• The approach effectively enhances prediction accuracy by transferring knowledge from existing data.
• This method provides a practical and efficient tool for improving reservoir reconstruction in the oil and gas industry.