Variable guidance lanes are designed to alleviate traffic congestion by adjusting the functional configurations of intersection entrance lanes. Existing variable guidance lanes predominantly operate on fixed-time schedules，which fail to fully utilize their adaptive potential. An intelligent control method integrating Kolmogorov-Arnold networks（KAN），convolutional neural networks（CNN），and long short-term memory（LSTM） networks was proposed to predict short-term traffic flow and temporal occupancy rates for straight-through and left-turn movements. Traffic data predictions were processed using a firefly algorithm-optimized fuzzy C-means（FCM） clustering algorithm，categorizing traffic states into three modes： “straight-through”“combined straight-left” and “left-turn”. A constrained switching framework was developed to dynamically adjust lane functions based on real-time traffic conditions. A microscopic traffic simulation was implemented in VISSIM to evaluate operational performance under different lane configurations. Results demonstrate that the proposed system reduces average travel time by 5.33%，queue length by 3.52%，and traffic delay by 11.79% in variable guidance lane approaches. The intelligent control method effectively enhances intersection operational efficiency while maintaining compliance with traffic regulations. This approach provides a data-driven solution for adaptive traffic management in urban road networks.