Research And Implementation On Flow Prediction Algorithm For Logistics Data

With the rapid development of the Internet, Big data and Internet?Internet of things?Internet of Vehicles interacts with each other, it also affects the development of logistics industry. Logistics enterprises produce a wide range of logistics data, and which is important. Effectively dealing with traffic big data will provide a reliable theoretical basis and technical support. Real-time and accurate traffic prediction is the premise and foundation of the effective data processing, it also brings new problems to the prediction of traffic data, and it is worth further research.For the time series model, only the historical time series can be used to predict the traffic flow. This paper firstly analyses the characteristics of the traffic flow and then decomposes with the LMD(Local Mean Decomposition), which is also proved that the decomposition of the flow sequence with short correlation feature. Therefore, we put forward a traffic flow time series prediction algorithm based on LMD and GARCH(Generalized AutoRegressive Conditional Heteroskedasticity), the prediction accuracy is significantly higher than the traditional time series models. However, time series model can only achieve offline prediction, in order to forecast for data stream, this paper proposes a RBFNN online prediction algorithm based on SKmeans and SGD. This algorithm is an improvement on RBFNN, the training data was online clustering firstly, and then through the improved SGD algorithm to train the parameters. Experimental results show that the prediction accuracy and training efficiency of the algorithm were significantly higher than the nearest neighbor clustering online training algorithm, and achieve online prediction of the traffic flow effectively.With the increasing of training samples, the improved algorithm can not meet the actual application requirements. Therefore, this paper will achieve the RBFNN online prediction algorithm with the real-time stream processing platform Storm. In order to realize the prediction effectively, the parallel implementation of the algorithm is designed, and then the overall implementation of the algorithm is presented in combination with the vertical parallelism and the level. Finally, we build the Storm cluster environment. Experiments show that, compared to the local mode, the cluster model has a faster training speed on the prediction of traffic data, and the acceleration effect is obvious.