In order to alleviate urban congestion

it is necessary to establish an intelligent traffic management platform that can predict traffic flow. In this paper

the outlier detection algorithm based on the Spark is used to clean the real-time massive traffic flow data. Load balancing rules are designed for the parallel data registration and storage. Semantic parsing and logical optimization are used to realize distributed semantic queries. And the STARIMA model based on multi-order spatial weight matrix is designed to realize the traffic flow forecasting. By the comparison experiments

it is proved that: ① The traffic flow data cleaning

statistics

registration and storage methods effectively utilize the advantages of memory computing and iterative computing of the Spark framework. In the big data environment

this method reduces the time consumption by about 60% compared with the MPI method or MapReduce method. And it can complete the data preprocessing in one prediction cycle. ② The semantic query method provides data to the traffic flow prediction model. The multi-order spatial weight matrix of the model can reflect the multi-order traffic flow assignment law more accurately. Compared with the dynamic STARIMA model

the accuracy of prediction analysis can be increased by about 25%. And the method can provide reference for traffic guidance information publication.