The inspection of hydropower stations using autonomous robots is vital for ensuring operational safety and efficiency. Obstacle avoidance plays a crucial role in enabling these robots to navigate complex environments that are filled with both static and dynamic hazards. However, existing obstacle avoidance methods often struggle to handle real-time changes in spatial and temporal contexts, resulting in suboptimal path planning and an increased risk of collision. To address these limitations, this research proposes a Spatio-Temporal Graph Convolutional Network (ST-GCN) based framework that effectively models both the spatial layout and dynamic movements of obstacles over time. The proposed ST-GCN framework processes real-time sensor data (LiDAR, cameras, IMU) and historical movement patterns to predict obstacle trajectories and adapt the robot’s navigation path accordingly. This approach allows the inspection robot to dynamically adjust its route in environments such as turbine halls, where moving machinery and personnel are common. Experimental evaluations in a simulated hydropower station environment demonstrated that the ST-GCN-based method significantly outperformed traditional reactive models, achieving higher accuracy in obstacle prediction and safer, more efficient navigation. These findings validate the effectiveness of spatio-temporal modeling for intelligent obstacle avoidance in industrial robotic inspection tasks.