Simulated Oil Well Sensor Data for Virtual Flow Metering Benchmarking

5,762 simulated time-series samples per well for two gas-lifted subsea oil wells, used to assess 10 machine learning algorithms for predicting oil and gas flow rates. It was created by Neville Aloysius D’Souza to evaluate model performance, noise filtering, and uncertainty quantification in a virtual flow metering context.

Use Cases

• Compare the prediction accuracy of 10 ML algorithms, including LSTM and XGBoost, for forecasting oil and gas flow rates from sensor inputs.
• Assess the impact of measurement noise and the effect of a median noise filter on model performance for bottom-hole pressure and temperature data.
• Implement uncertainty quantification by generating 95% confidence intervals for flow rate predictions using XGBoost models.
• Train regression models using key input variables like choke opening and wellhead pressures to predict multiphase flow rates.

Strengths

• Provides 5,762 time-series samples per well, enabling robust training and testing splits for model evaluation.
• Benchmarks 10 distinct machine learning methods, offering a direct comparison of algorithm performance on identical data.
• Covers a full machine learning workflow including data preprocessing, hyperparameter tuning, and uncertainty quantification.

Limitations

• Data is synthetic, generated from a simulation model, and may not fully capture the complexities and noise of real-world field measurements.
• The dataset is limited to two specific simulated well scenarios, which may not represent the full diversity of operational conditions in oil production.

Provenance

Source

Neville Aloysius D’Souza via DataverseNO Harvested Dataverse.

Collection Method

Data was generated from validated open-loop simulation models of gas-lifted oil wells, as described by Janatian et al. (2022).

Freshness

Last updated on 2026-03-01.