CO2 Phase Change Fracturing Effect Prediction and Sensitivity Analysis Based on SAPSO-BP

Literature Overview and Research Context

This paper presents a prediction model for CO2 phase change fracturing effects using a Simulated Annealing Particle Swarm Optimization (SAPSO) optimized Back Propagation (BP) neural network approach. CO2 phase change fracturing is an emerging technology in reservoir stimulation where liquid CO2 is injected into formations, undergoes phase change to gas, and generates fracture networks through pressure buildup and thermal effects. The study develops a data-driven model to predict fracturing effectiveness based on multiple input parameters and performs sensitivity analysis to identify the most influential factors.

Core Technical Content

The SAPSO-BP model integrates simulated annealing (SA) with particle swarm optimization (PSO) to optimize the initial weights and thresholds of a BP neural network, avoiding local optima and improving prediction accuracy. The model takes reservoir parameters, injection parameters, and formation properties as inputs and predicts fracturing effectiveness metrics including fracture length, fracture width, and stimulated reservoir volume.

The model achieves prediction accuracy with mean absolute percentage error (MAPE) below 5% for fracture length and 8% for fracture width. The sensitivity analysis reveals that injection pressure, formation permeability, and overburden stress are the three most influential parameters, collectively accounting for over 65% of the prediction variance. The CO2 injection volume and temperature gradient are secondary factors, while formation thickness and CO2 viscosity have relatively minor influence on fracturing effectiveness.

Model Architecture and Training Process

The BP neural network architecture consists of an input layer with 10 neurons, two hidden layers with 20 and 15 neurons respectively, and an output layer with 3 neurons. The SAPSO algorithm optimizes the network weights and biases by combining the global exploration capability of PSO with the local exploitation capability of simulated annealing. The cooling schedule of the simulated annealing component follows a geometric decay function with initial temperature T0 = 100 and cooling rate α = 0.95. The training process uses a dataset of 200 samples, with 70% for training, 15% for validation, and 15% for testing.

Integration with Engineering Practice

For petroleum and natural gas reservoir stimulation engineers, the SAPSO-BP model provides a rapid and accurate tool for predicting CO2 phase change fracturing outcomes without the need for computationally expensive numerical simulations. The sensitivity analysis results guide parameter optimization during fracturing design, indicating that injection pressure and formation permeability should be prioritized for optimization. In practice, this means that for low-permeability formations, higher injection pressures are needed to achieve effective fracturing, while for high-permeability formations, injection rate control becomes more critical to prevent fluid loss.

The model can be integrated into real-time fracturing operation monitoring systems to provide predictive guidance for adjusting injection parameters during operations. The sensitivity analysis also informs risk assessment, identifying which parameters require the most precise control to ensure fracturing effectiveness. For example, a 10% error in injection pressure measurement has a significantly larger impact on prediction accuracy than the same percentage error in formation thickness.

Key Questions and Reflections

A significant limitation of data-driven prediction models is their dependence on the quality and representativeness of the training dataset. The model's performance for formations outside the range of the training data may degrade significantly. Engineers should exercise caution when applying the model to formations with properties substantially different from those in the training dataset. Additionally, the model does not capture the complex multiphase flow, heat transfer, and geomechanical interactions that occur during CO2 phase change fracturing, and should be used as a screening tool rather than a replacement for detailed numerical simulation in critical applications.

The sensitivity analysis results also raise questions about the interaction effects between parameters. While the analysis identifies individual parameter influences, the combined effect of multiple parameters may be non-linear and synergistic. For instance, the combined effect of high injection pressure and low permeability may produce fracturing outcomes that are not simply the sum of their individual effects. Future work should incorporate interaction terms and nonlinear sensitivity analysis to provide a more complete understanding of parameter influences.

Study Insights and Implications

The SAPSO-BP approach demonstrates that intelligent optimization algorithms can significantly improve the prediction accuracy of neural network models for complex engineering problems. For engineers involved in reservoir stimulation and hydraulic fracturing, the combination of optimization algorithms with neural networks provides a practical tool for rapid prediction and parameter sensitivity analysis. The key insight is that multi-objective optimization approaches like SAPSO are superior to single-objective methods for training neural networks, as they better explore the solution space and avoid local optima. This principle can be extended to other engineering prediction problems where multiple input parameters influence complex output behaviors, providing a general framework for developing accurate and efficient prediction models in petroleum engineering and related fields.