Support Vector Regression Based on the Particle Swarm Optimization Algorithm for Tight Oil Recovery Prediction

Area of Science:

• Petroleum Engineering
• Data Science
• Machine Learning Applications

Background:

• Tight oil recovery rates are difficult to predict accurately due to complex geological, technological, and developmental factors.
• Traditional prediction methods (formula calculations, curve fitting) lack applicability and deviate from actual reservoir conditions.
• Accurate recovery rate prediction is vital for evaluating reservoir development and performing dynamic analysis.

Purpose of the Study:

• To develop accurate machine learning models for predicting tight oil reservoir recovery rates.
• To compare the performance of Support Vector Regression (SVR) and optimized Particle Swarm Optimization-SVR (PSO-SVR) models.
• To assess the impact of data scarcity on model accuracy for tight oil recovery prediction.

Main Methods:

• Utilized Pearson correlation coefficient and Random Forest (RF) to analyze nonlinear influences of factors on oil well recovery.
• Developed and trained SVR and PSO-SVR models using 75% of production data from 200 horizontal wells in M tight oil reservoirs.
• Validated model performance using the remaining 25% of the data.

Main Results:

• The PSO-SVR model demonstrated a 10.85% improvement in accuracy over the unoptimized SVR model when data is scarce.
• Both RF and SVR-based models provided more accurate predictions compared to traditional methods.
• Identified over 14 factors influencing recovery rates, with a wide range of recovery observed (8.8% to 27.6%).

Conclusions:

• Optimized machine learning models, such as PSO-SVR, offer superior accuracy for tight oil recovery prediction, particularly in data-limited scenarios.
• Machine learning provides a more accurate and applicable approach to predicting oil recovery rates.
• Future optimization of construction and production systems can be informed by these accurate predictions to enhance overall oil recovery.