Oilfield Production Prediction Method Based on Multi-Input CNN-LSTM With Attention Mechanism
Oil production prediction is crucial for the formulation of adjustment strategies, enhancement of recovery rates, and guidance of production in oilfields. Traditional production prediction methods based on reservoir numerical simulation are costly, challenging, and heavily influenced by human experi...
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Wiley
2025-01-01
|
| Series: | Geofluids |
| Online Access: | http://dx.doi.org/10.1155/gfl/6195991 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Oil production prediction is crucial for the formulation of adjustment strategies, enhancement of recovery rates, and guidance of production in oilfields. Traditional production prediction methods based on reservoir numerical simulation are costly, challenging, and heavily influenced by human experience, while the application of production prediction models such as decline curves yields poor results. To achieve rapid, low-cost, and intelligent oil production prediction, we propose a multi-input deep neural network model combining convolutional neural networks (CNNs) and long short-term memory (LSTM) networks with an attention mechanism. This model achieves prediction through two primary input paths: firstly, utilizing CNN to extract spatial dynamic features between wells to capture interwell production relationships and secondly, employing LSTM to extract temporal dynamic features of the oilfield. The model combines the attention mechanism to strengthen the key information. Additionally, to quantify the impact of different input features on production, we adopt a random forest algorithm to assess feature importance and optimize data input through assigned weights. Finally, the trained model is used to forecast oilfield production. Three sets of comparative experiments are conducted in this paper. Experiment 1 confirms that the new method outperforms previous methods in prediction performance. Experiment 2 demonstrates that the multi-input model exhibits superior prediction performance compared to single-input models. Experiment 3 verifies that the combination of importance weight initialization and the attention mechanism significantly enhances the accuracy of the model’s predictions. |
|---|---|
| ISSN: | 1468-8123 |