An intelligent prediction method for ROP in drilling based on optimized PSO-BP neural network

Area of Science:

• Petroleum Engineering
• Machine Learning
• Drilling Optimization

Background:

• Accurate rate of penetration (ROP) prediction is crucial for optimizing drilling operations, resource allocation, and cost management.
• Existing ROP prediction methods often lack accuracy and efficiency in complex geological formations.

Purpose of the Study:

• To develop a novel ROP prediction model integrating particle swarm optimization (PSO) with a momentum-adaptive backpropagation (BP) neural network.
• To enhance the accuracy, efficiency, and generalizability of ROP prediction models.

Main Methods:

• Systematic selection of critical input parameters including engineering indices, hydraulic characteristics, and lithological properties.
• Development of a hybrid PSO-BP algorithm incorporating momentum acceleration and adaptive learning rate adjustment.
• Validation using 1200 data points (80% training, 20% testing) and comparative experiments with field data.

Main Results:

• Significant correlations (coefficient > 0.5) identified between ROP and selected drilling parameters.
• The optimized PSO-BP model achieved superior performance compared to standard BP and GA-BP models.
• Achieved Mean Absolute Error (MAE) of 0.30 m/h, Mean Absolute Percentage Error (MAPE) of 11.35%, and R-squared (R²) of 0.93.

Conclusions:

• The proposed data-driven PSO-BP model provides a reliable framework for accurate ROP prediction in well-characterized formations.
• This approach enhances operational efficiency and supports cost-effective drilling strategies.
• The model demonstrates improved predictive accuracy and generalizability through heuristic optimization techniques.