Railway Freight Volume Forecasting Method Based On High-dimensional Panel Data

The freight volume predicting is related to the construction of railway infrastructure,the formulation of freight plans and the management of freight yards.Therefore,the railway department has been trying to predict the volume of railway freight accurately.The traditional predicting method is limited by data collection,so it only predict based on the time series development of the railway freight volume itself,which the information relies on is not sufficient.With the development of data collection technology,the using of more data for more adequate data modeling has become the trend of current predicting development.High-dimensional panel data is a form of data whose cross-section dimension is much larger than the length of the time series.The information contained in is richer than the single time series and can provide sufficient basis for prediction.However,if the prediction is based on multidimensional data,the contradiction between the indexes makes the fusion of data a new problem.The thesis proposes a predicting method of freight volume based on high-dimensional panel data.First,data filtering and fuzzy clustering are used to achieve data dimensionality reduction,and the clustering results are tested by using chaos.Next,it pointing out that there is multiple collinearity between the prediction results of each class,and the data is fused by principal component regression.Then,according to the advantage of generalized regression neural network,for example,the calculation speed is fast,there is no local convergence,the function fitting ability is good,and the number of parameters is small,it is chose as the prediction tool.At the same time,pointing out the shortcomings of generalized regression neural network and proposing improvement: First,it does not distinguish the importance of different dimensions when calculating the distance between samples.Second,the prediction result cannot exceed the maximum and minimum values of the training sample set.For the first shortcoming,introducing weighted distance to change the role of different dimensions.For the second shortcoming,adding residuals in the molecular terms of the network summation layer to change the network performance;and using the growth rate of unequal spacing to expand the sample set.Finally,based on the macro annual report of Guangzhou City from 1999 to 2015,the prediction method proposed in the thesis is cross-tested.The experimental results show that the prediction error is between 1.0% and 9.0%.The method of predicting railway freight volume based on high-dimensional panel data proposed in the thesis has higher accuracy and better stability.However,there are still some shortcomings in the research of improving generalized regression neural network.Although the improvement of adding residuals in generalized neural network can make the network prediction break through the original limit,the scope of breakthrough is not controlled,so further research is needed in the future.