Abstract: The interior of large power transformers features densely packed windings, narrow gaps, and interlaced iron cores, rendering such complex confined environments extremely challenging for inspection tasks. Hyper-redundant snake-like robots with slender configurations are ideal carriers for inspection in such environments. However, these robots suffer from insufficient environmental adaptability and dynamic obstacle avoidance capabilities in narrow confined spaces. This paper proposes an obstacle avoidance control strategy that integrates deep learning-based speed adaptation with model predictive control (MPC) for path optimization. First, a simulation environment of the transformer interior incorporating multiple types of geometric constraints is constructed. Second, a sliding-window kinematics method is introduced to enable the robot body to follow the head trajectory while reducing the high-dimensional computational burden. Finally, a deep learning-based speed prediction model is designed to output adaptive speed coefficients based on environmental features, and is combined with an MPC controller to optimize the moving direction and step size of the robot head. When the robot falls into local extrema, a yaw-priority heuristic strategy is employed to enhance robustness. Simulation results demonstrate that the proposed method reduces the collision rate from 19.8% to 6.2% in high-density obstacle scenarios, validating its effectiveness in improving dynamic obstacle avoidance safety in confined spaces.