ML-based virtual flow meter

The solution had to process a wide variety of sensor types (temperature, pressure, vibration, acoustic), each with unique formats and noise characteristics.

Solution: We implemented a flexible data ingestion and preprocessing pipeline capable of normalizing and validating diverse input formats in real time. The system supports secure, scalable streaming from edge devices and sensors, ensuring consistent data quality across all environments.

The lack of labeled flow rate data made it difficult to train supervised models from scratch.

Solution: We applied a hybrid modeling approach, combining physics-based models with ML to reduce reliance on labeled data and maintain consistency with physical constraints.

Multiphase pipelines often exhibit non-steady behavior, such as slug flow or abrupt regime changes, which are difficult for static models to capture.

Solution: Our pipeline incorporated LSTM-based time-series models alongside first-principles equations to handle transitions and produce stable predictions under transient conditions.

Ensuring sub-second inference latency and compatibility with existing SCADA systems required architectural optimization.

Solution: We deployed models within a cloud-native inference framework, using event-triggered execution and lightweight serverless functions to ensure low-latency predictions. The system exposed secure API endpoints for seamless integration with external platforms, enabling real-time responsiveness across diverse operational environments.

Over time, physical sensors (especially in harsh oilfield environments) degrade, introduce noise, or drift from calibration, resulting in reduced model accuracy.

Solution: We implemented automated data validation, anomaly detection, and reconciliation rules based on conservation laws (mass/energy balance). This enabled ongoing correction of flawed sensor inputs without human intervention.

ML models degrade as operating conditions, fluid properties, or equipment configurations change, making predictions stale.

Solution: We built an MLOps pipeline to monitor model performance, detect drift, and automatically trigger retraining cycles using fresh data. This ensured sustained model accuracy across changing operational regimes.

Statistical checks detect outliers or missing values.

Tuning model parameters (e.g., friction factors, choke discharge
coefficients) to match established operating envelopes.

Cross-check predictions against multiphase flow meters or test
separators for periodic calibration where available.

• Neural Networks (CNN/MLP) for direct sensor-based flow prediction.
• Recurrent architectures (LSTM/RNN) for dynamic or transient phenomena.
• Ensemble or meta-learning methods to consolidate multiple model outputs.

• Automated retraining triggered by new data or drift detection.
• Version control and rollback for consistent reliability in production.