Traffic Flow Forecast Based On Combined Models

In recent years,the number of road vehicles in China has continuously created new records.The conflict between people's growing travel needs and limited road infrastructure capabilities is growing.This contradiction eventually leads to a growing problem of road congestion.In order to alleviate the traffic congestion problem and increase the road utilization rate of urban traffic,people's demand for predicting road traffic conditions has gradually increased.Accurate traffic flow prediction results can improve the operation of the entire transportation system.Therefore,as one of the core technologies in intelligent transportation systems,the study of short-term traffic flow prediction has important research value and application value.Based on the above viewpoints and the previous research,and combined with the idea of deep learning,this paper proposes several combined model prediction methods based on autoregressive moving average model(ARIMA)and long short-term memory(LSTM).The method can effectively combine the advantages of the ARIMA model and the LSTM network,thereby improving the prediction effect of the combined model.On this basis,in order to increase the fitting ability of the ARIMA model to the neighbor samples,this paper establishes the rolling-based ARIMA model,which can re-fitting the observed new samples,thus increasing the short-term prediction ability of the model.Moreover,for the problem that the scale of information of the training samples is too small,this paper proposes a data processing method based on information backtracking,which can increase the information of a single sample by increasing the prediction interval and increasing the influence weight.The method can improve the training efficiency and prediction effect of the LSTM network.At the same time,for the problems of long training time and high energy consumption in deep learning,this paper adopts a series of optimization measures when training LSTM network,which further improves the training efficiency and prediction effect of LSTM network.Finally,based on the traffic flow characteristics,considering the huge impact of the neighbor historical samples on future predictions,combined with the idea of online learning,a model combination method based on variance and sliding window is proposed.The method can recalculate the variance of the model according to the newly observed data samples and the prediction results of each model in the prediction,and dynamically and rationally allocate the weight of the model.On this basis,this paper combines the idea of data-driven and further proposes an improved model combination method,which can fit the linear relationship model between models by observing data and historical prediction results,and dynamically assign the weight value of the model.In order to prove the efficiency and superiority of the model proposed in this paper,a detailed comparative experiment was carried out in this paper.The experimental results show that the proposed method has better prediction effect,and each index is better than a single model and has good versatility and stability.