Using Transfer Learning and LSTM Neural Networks for Reservoir Parameter Prediction

By Muhammad OsamaReviewed by Susha Cheriyedath, M.Sc.Feb 7 2024

In an article published in the journal Plos One, researchers from China proposed an innovative method for predicting reservoir parameters, such as porosity and permeability, using transfer learning and long short-term memory (LSTM) neural networks.

The authors discussed how their approach could address the challenge of data scarcity and distribution discrepancy in reservoir parameter prediction by leveraging historical data from different blocks and applying a pre-train-finetune strategy. Moreover, they claimed that their technique can effectively extract common features from source and target domains and achieve higher prediction accuracy and efficiency than conventional methods.

Background

Reservoir reconstruction is an important step in oil and gas exploration and production. It involves estimating the physical properties and geological structures of the reservoirs. Logging curves are majorly used to provide continuous records of measurements along the wellbore. However, incomplete and noisy logging data pose challenges for accurate parameter prediction. Moreover, acquiring logging data is an expensive task.

Previously, machine learning techniques, especially deep learning models, have shown promising results in reservoir parameter prediction in terms of accuracy and efficiency. However, these models require large training datasets, which may not be available for new or unconventional reservoirs. Therefore, there is a need for methods that can leverage existing historical data and transfer knowledge to new wells with different geological conditions.

About the Research

In the present paper, the authors introduced a transfer learning approach based on LSTM neural networks for forecasting reservoir parameters. It can extract common features from datasets that obey different data distributions and enhance the prediction performance on the target domain.

Transfer learning is a technique that leverages the knowledge learned from one domain to another domain that has insufficient data. LSTM is a special kind of recurrent neural network that can capture sequential and temporal information in logging data and avoid the gradient vanishing or exploding problem.

The designed method consists of two stages: pre-training and fine-tuning. In the pre-training step, the LSTM model was trained on a large amount of historical data from a source block, which had similar geological conditions to the target block. In the fine-tuning phase, the LSTM model was adapted to the target block by freezing the first few layers and updating the last few layers with a small amount of data from the target block. This allowed the model to transfer common features from the source block and adjust to the specific features of the target block. The model structure and parameters were determined through grid search and optimization methods.

The model was tested on two datasets from two major oil and gas well blocks in Sichuan Province, China, namely Gaomo and Chuanxi. The Gaomo block had more complex geological conditions and less logging data, while the Chuanxi block had more data and simpler structures. The study used the Gaomo block as the target domain and the Chuanxi block as the source domain. Moreover, nine input variables, such as acoustic, caliper, and water saturation, were used to estimate one target variable (porosity of the formations).

The authors compared the newly presented method with four baseline models: a model trained on the source domain only, a model trained on the target domain only, a model trained on mixed data of source and target domains, and a model trained on the target domain with data augmentation. Mean absolute error (MAE), mean squared error (MSE), root mean square error (RMSE), and mean absolute percentage error (MAPE) were used as the performance factors.

Research Findings

The outcomes showed that the new technique outperformed the baseline models on all metrics and demonstrated its effectiveness in utilizing historical data and transferring knowledge to the new domain. It achieved an MSE of 0.0015, an MAE of 0.029, a MAPE of 0.062, and an RMSE of 0.039, while the best baseline model (trained on the target domain only) achieved an MSE of 0.0023, an MAE of 0.036, a MAPE of 0.077, and an RMSE of 0.048. Moreover, the developed model highlighted robustness and stability when varying the number of frozen layers and the size of the target dataset.

The proposed method has potential applications in reservoir reconstruction and management. It can provide reliable and efficient predictions of reservoir parameters from logging data. Additionally, it can be extended to other domains that face data scarcity problems, such as traffic flow prediction, location-based services, and natural language processing. Moreover, it can reduce the cost and time of data acquisition and improve the quality and accuracy of data analysis.

Conclusion

In summary, the novel methodology is efficient for predicting reservoir parameters. It effectively leverages the potential of the transfer learning technique LSTM neural networks and the historical data from different sources to overcome the data scarcity problem in reservoir reconstruction. The authors demonstrated that their technique highlighted superior performance and robustness compared to baseline models on two datasets from Sichuan Province, China.

The researchers acknowledged the limitations and challenges of the study and suggested some directions for future research, such as exploring more transfer learning techniques, incorporating more geological information, and applying the method to other reservoir parameters or regions. The research also calls for more collaboration and data sharing among scientists and practitioners in reservoir reconstruction.