Highway networks are expanding and traffic congestion problems are gradually emerging. Accurate prediction of highway traffic speed is of great significance to traffic management and scheduling, and it is a key means to alleviate congestion and improve operational efficiency. In this study, a short-term traffic speed prediction system for highways is constructed based on the multilayer perceptual machine model and historical flow data. By analyzing the toll data of Chongqing Yuxiang Expressway in 2018 and the traffic speed data from 2023 to 2024, the spatial and temporal distribution characteristics and regular fluctuation patterns of traffic flow are revealed. The study adopts the multilayer perceptron model to establish the mapping relationship between the traffic flows of relevant road sections, introduces the BP algorithm for model training, and evaluates the prediction effect by the average relative error and the mean coefficient of parity. The experimental results show that the R² value of the constructed multilayer perceptron model reaches 0.837, which is 0.158 higher than that of the traditional RNN model and 0.089 higher than that of the GRU model, and the prediction accuracies are improved by 0.012 and 0.034 compared with those of the RNN and the GRU, respectively, which effectively captures the cyclic change characteristics of the traffic speed, and it is of great value to be applied in the support of decision-making of traffic management. The study confirms that the spatial and temporal features embedded in the historical flow data are of great value for short-term traffic speed prediction, and the short-term traffic speed prediction method based on multilayer perceptron can provide a scientific basis for highway traffic management.
