Causal Model Segmentation for Well Spacing Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine-learning models used in hydrocarbon well drilling are ineffective in accurately predicting well performance for scenarios outside their training data, particularly in terms of well spacing, leading to inaccurate predictions and reliability issues.
Innovation Solution
The development of a causal model that separates the impact of well spacing from other features, using multiple machine-learning models to predict production errors and apply counterfactual scaling factors to estimate production outputs for unconventional well spacings, thereby controlling for confounding variables and improving prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional machine-learning models are trained on historical well data, then they can predict production for similar scenarios, but they fail to accurately predict production for scenarios with different well spacings not represented in the training data
Solution Approach 1:
The patent segments the prediction task into multiple components: (1) a base model trained on historical data to capture general production patterns, and (2) a separate causal model specifically trained to predict the effect of well spacing on production. This segmentation allows each model to specialize in its respective function, with the causal model handling the adaptability to unseen spacing scenarios while the base model provides reliable baseline predictions.
Solution Approach 2:
The patent introduces an intermediary causal model that acts as a mediator between the base model and the final prediction. This intermediary model specifically captures the spacing effect and adjusts the base model's predictions accordingly. The causal model serves as a bridge that enables the system to adapt to new spacing scenarios without retraining the entire system, thus improving both reliability and adaptability.
2Measurement precision
If models are trained to be highly accurate on training data, then they achieve good fit on historical scenarios, but they overfit and cannot generalize to new spacing configurations
Solution Approach 1:
By segmenting the prediction task into a base model and a causal spacing model, the patent prevents overfitting. The base model can be trained to high accuracy on historical data without capturing spacing-specific patterns, while the causal model learns the generalizable spacing effects. This division allows each component to be optimally trained for its specific purpose without overfitting to the entire dataset.
Solution Approach 2:
The patent extracts the spacing effect as a separate causal factor from the overall production prediction task. By taking out the spacing variable and modeling its effect independently through a dedicated causal model, the system avoids overfitting spacing patterns to the base model. This extraction allows the base model to focus on general production patterns while the causal model handles spacing-specific generalization.
3Reliability
If additional features including well spacing are included in the training data, then the model can capture spacing effects, but the model complexity increases and computational requirements rise
Solution Approach 1:
The patent segments the feature processing by creating a dedicated causal model that focuses exclusively on spacing features. This segmentation allows the base model to remain relatively simple while the causal model handles the complexity of spacing relationships. The modular structure reduces overall system complexity compared to a single monolithic model that would need to process all features simultaneously.
Solution Approach 2:
The causal model acts as an intermediary that simplifies the integration of spacing features into the prediction system. Instead of directly incorporating complex spacing features into the base model, the causal model processes these features and provides adjusted predictions. This intermediary approach reduces the computational burden on the base model while maintaining prediction reliability.
Data Source
AI summary
Example embodiments involve obtaining training data comprising a first set of features relating to drilling wells, a second set of features relating to geology at the well locations, a third set of features relating to spacings between wells, and a fourth set of features relating to production output of the wells; training a first model to predict the third set of features given the first and second set of features; determining a spacing error based on differences between the third set of features in the training data and as predicted; training a second model to predict the fourth set of features given the first and second set of features; determining production error based on differences between the fourth set of features in the training data and as predicted; and training a third model to predict the production error given the spacing error and the first and second set of features.