The Research Of Temporal Flow Prediction And Spatial-Temporal Flow Prediction Based On Neural Networks

The prediction of time seires and spatial-temporal series is an important way for researchers to analyze and understand the object world and social phenomena.With the rapid development of the information technology in modern society,massive time series data and spatial-temporal series have been produced in various fields.People can avoid risks,save resources,and improve efficiency by analyzing the development trend and formulating effective management measures.The traditional methods of temporal prediction and spatial-temporal prediction are statistical models which are base on linear relationship.However,due to the high non-linear features in real problems,the performance of these models is inaccurate.In recent years,the deep learning model based on artificial neural network provides more accurate prediction results when dealing with non-linear problems in real world.Compared with the traditional methods,neural network does not rely on prior assumptions.It can flexibly build the network and capture the dependence of the data.The researchs of neural network model in temporal prediction and spatial-temporal prediction are developing rapidly,but there are still some problems,such as the difficulty to capture long-distance spatial dependence and the poor prediction performance when data is scarce.Therefore,making accurate flow prediction in real world problem by deep learning models based on neural networks to capture the temporal and spatial-temporal characteristics are the key issues of this study.In order to improve the prediction accuracy,we focus on the the following four aspects in this work:(1)In order to improve the stability and accuracy of multi-step temporal prediction,a deep learning model named MEMORY-LSTM is proposed to capture dynamic temporal features and static pattern features.MEMORY-LSTM uses two parallel LSTM network to capture the sequential and periodic characteristics of time series.However,it is difficult for the model to learn these features if the training set is lack of adjacent time data.Therefore,a static pattern module is designed in MEMORY-LSTM to learn the representation of different patterns and extract the pattern information,which can improve the accuracy and stability of the multi-step temporal prediction.The experimental results show that MEMORY-LSTM can improve the prediction accuracy by combining temporal features and pattern features for multi-step prediction of time series.In addition,the influence of pattern feature information on prediction performance is further analyzed.(2)The accuracy of spatial-temporal prediction is not only affected by time factors,but also related to spatial dependence.In this work,a novel deep learning model which called LDRSN is proposed to make spatial-temporal prediction.LDRSN combines local convolutions with dilated convolutions to learn the nearby and distant spatial dependency.Furthermore,a region-level tempoal-shifting attention mechanism is proposed to model the temporal shifting which varies by region.In the experiments,we compare the proposed method with other stateof-the-art methods in two real-world crowd flows datasets.The experiment results show the effectiveness of the proposed model.(3)Neural network is based on data-driven,and data scarcity will affect the prediction performance.In this work,a novel transfer learning method ARG-STNet is proposed to make spatial-temporal prediction with only a small collection of data.Specifically,our model is designed as a spatial-temporal network based on a first-order meta-learning algorithm Reptile with an attention mechanism of the spatial-temporal distribution similarity.In addition,a generation mechanism is designed to learn and transfer long-term temporal features from source data which have abundant data to the target data.In the experiments,we compare our model with other state-of-the-art methods in real-world traffic prediction task.The experiments demonstrate that ARG-STNet can improve the prediction performance.(4)In this work,a transfer learning method GD-STNet based on a generative adversarial network is proposed to make spatial-temporal prediction with only a small collection of the target data and different types of source data which are under the same temporal-spatial conditions with the target data.GD-STNet uses discriminator network to classify the type of source domains,and then uses basic spatial-temporal network to learn spatial-temporal features and generator network to extract common spatial-temporal features.Finally,GD-STNet combines spatial-temporal features and common spatial-temporal features to make the prediction.The prediction results based on wireless traffic data show that GD-STNet can effectively improve the accuracy of the prediction.In summary,in order to solve the shortcomings of the existing temporal prediction and spatial-temporal prediction methods and improve the prediction performance,this paper makes further focuses on multi-step temporal prediction,multi-scale distance of spatial-temporal prediction and the transfer learning method of spatial-temporal prediction.