Time series causal relationships discovery through feature importance and ensemble models

Area of Science:

• Machine Learning
• Causal Inference
• Time Series Analysis

Background:

• Interpreting complex Machine Learning models from observational data is challenging.
• Increasing data complexity hinders understanding of model decision-making processes.
• Causal inference is crucial for reliable forecasting and model interpretability.

Purpose of the Study:

• To propose a novel methodology for inferring causal relationships from observational time series data.
• To leverage ensemble models for causal discovery in complex datasets.
• To establish the effectiveness of the proposed method in identifying interwell connections in oil field production.

Main Methods:

• Utilized ensemble models, specifically Random Forest, for feature importance analysis.
• Developed an iterative forecasting approach to identify causal drivers.
• Validated the methodology using synthetic datasets and real-world oil field production data with tracer information.

Main Results:

• The proposed method successfully identified causal relationships in both synthetic and real oil field datasets.
• Causal analysis results align with confirmed interwell connections from tracer data in oil fields.
• Demonstrated the ability to build causal networks by assessing feature importance in forecasting models.

Conclusions:

• The methodology provides a reliable approach for causal discovery using observational time series data.
• Ensemble models offer a powerful tool for uncovering hidden causal links in complex systems.
• This research pioneers the use of production data for causal analysis of interwell connections in oil fields.