Optimization of pattern recognition algorithm for traffic flow prediction and road condition recognition in intelligent transportation system

Aiming at the key problems of traffic flow prediction and road condition recognition in Intelligent Transportation System (ITS), this paper puts forward an optimized framework of pattern recognition algorithm which integrates multi-source data and dynamic weight adjustment. Firstly, a spatio-temporal semantic model integrating sensor, traffic flow and event information is constructed, and graph neural network (GNN) is used to capture the link association. Then, an adaptive weighting mechanism based on attention and reinforcement learning (RL) is designed to improve the prediction robustness in complex scenes. Combined with Transformer-LSTM and Particle Swarm Optimization (PSO), the computational overhead is reduced and edge deployment is supported. Experiments are carried out based on PeMSD4 data and real-time data of an industrial park. The results show that the proposed method is superior to traditional and deep learning models in traffic flow forecasting performance. The root mean square error (RMSE) of 15-minute forecasting task is 23.4, the mean absolute percentage error (MAPE) is 8.7%, and the sudden congestion error is 28.9%. In the aspect of road condition recognition, the accuracy rate is 94.7% in normal weather, F1-score is 0.93, 88.4% in heavy rain, F1-score 0.85, 85.2% in sensor failure, and F1-score 0.82, which shows high robustness and reliability. At the same time, the parameters of the model are reduced from 43.2MB to 9.8MB, the reasoning delay is reduced from 121ms to 36ms, and the energy consumption is reduced from 293mJ to 87mJ, which has obvious lightweight characteristics and provides an efficient, accurate and adaptable solution for traffic flow prediction and road condition identification in ITS.