An uncertainty approach for Electric Submersible Pump modeling through Deep Neural Network

Area of Science:

• Petroleum Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Artificial oil lift systems using submersible electric pumps often require accurate predictive models.
• Traditional deep learning models necessitate large datasets, posing a challenge for complex systems.
• Evaluating the uncertainty of AI models is crucial for reliable system operation.

Purpose of the Study:

• To propose a novel methodology for identifying and validating deep learning models for artificial oil lift systems.
• To enable joint and systematic evaluation of model performance and prediction uncertainty.
• To develop computationally lighter AI dynamic models suitable for digital twin construction, control, and optimization.

Main Methods:

• Utilized a nonlinear model to generate synthetic training and validation data, mitigating the need for extensive real-world datasets.
• Employed the Markov Chain Monte Carlo algorithm to systematically assess the epistemic uncertainty of neural networks.
• Implemented the methodology in Python using Tensorflow and Keras for neural network construction and hyperparameter tuning.
• Validated the developed deep learning models against experimental data, including uncertainty assessment.

Main Results:

• The proposed methodology successfully generated deep learning models that accurately represent both the nonlinear model's dynamics and experimental data.
• The models provided a most probable value close to experimental data, with prediction uncertainty comparable to the nonlinear model.
• Uncertainty assessment confirmed the adequate validation of the developed AI dynamic models.
• The resulting models are computationally efficient and reliable.

Conclusions:

• The developed methodology offers a robust approach for creating and validating deep learning models for artificial oil lift systems.
• The AI dynamic models are suitable for applications requiring high reliability and computational efficiency, such as digital twins.
• This work advances the application of AI in optimizing oil and gas production systems.