Stability analysis and prediction of hazardous rock mass in cold regions based on hybrid algorithm model

In the complex geological environments of cold regions, traditional methods struggle to address the multifactorial coupling and nonlinear dynamic evolution of hazardous rock mass driven by freeze‒thaw cycles. To overcome these challenges, this study investigates the applicability and optimization of intelligent prediction models tailored to cold regions. A long-term stability prediction framework is constructed by integrating the freeze–thaw–gravity coupling mechanism mechanism. Unlike generic hybrid models, this research systematically compares and optimizes various metaheuristic algorithms (SSA, PSO, GA) coupled with neural networks to identify an effective strategy for the high-dimensional, nonlinear characteristics of rock mass in these regions. Focusing on hazardous rock mass in western China, six primary influencing factors—cohesion, freezing depth, lowest temperature, freezing load, sunshine duration, and foot of slope displacement—were selected on the basis of the typical freeze–thaw–gravity coupling mechanism damage mechanism. Key control parameters were identified via gray relational analysis (GRA), and data normalization was applied to enhance model generalizability. The evaluation results demonstrate that hybrid algorithm models outperform traditional single-algorithm models for the investigated cases, with improved prediction accuracy and adaptability under freeze-thaw-dominated conditions. Specifically, the SSA-BP model reduced the root mean square error (RMSE) by approximately 30% compared with the standalone BP model, whereas the mean absolute error (MAE) and mean squared error (MSE) decreased by 28% and 35%, respectively, and achieved a goodness-of-fit with measured data exceeding 90%. Moreover, the PSO-BP model improved computational efficiency by approximately 40% while maintaining prediction accuracy, rendering it suitable for real-time monitoring and rapid warning scenarios. These findings indicate that hybrid algorithm models partially alleviate the limitations of single models—such as poor generalizability and susceptibility to local optima—by incorporating global optimization mechanisms and adaptive parameter adjustment, thereby demonstrating improved robustness and potential engineering-oriented applicability.