Multivariate Time Prediction Based On Neural Network

With the continuous development of cloud computing and big data,people can calculate and store more and more data.People hope to predict the future and assist decisionmaking from a large amount of similar historical time series data.Therefore,the shortterm time series forecasting problem is a very meaningful research topic,and its results can be used in the fields of commodity inventory turnover,traffic flow warning and financial investment.With the deep learning technologies and theories,this thesis conducts research on two different forecasting methods:single-step forecasting and multi-step forecasting.The main work of this thesis includes:(1)For the multivariate single-step prediction problem,this thesis proposes the GCNNDeepAR model,which consists of The multi-kernel gated convolutional neural network and the LSTM long-short-term memory neural network,and by mixing the two neural networks,the model can enhance the capture of short-period and long-period temporal patterns.The model first makes probability distribution assumptions on the time series data,then trains the parameter set in the time series probability distribution through the probability distribution loss function,and finally generates a single-step predicted value by sampling the distribution,and gives the upper and lower prediction boundary.The comparison results on the two datasets show that the GCNN-DeepAR model achieves better results.(2)The traditional method generates multi-step forecast by iterative single-step forecast,which will lead to accumulative error affecting the accuracy.one-step iteratation method is only applicable to univariate time series prediction,not multivariate time series prediction.For the multivariable multi-step prediction problem,this thesis proposes a Transformer-AR model that can directly predict multi-step.This model consists of a multi-kernel gated convolutional nerual network and a Transformer encoder,and by mixing the two neural networks,the model can enhance the capture of short-period and long-period temporal patterns.For the attention mechanism of the Transformer nerual network is global and bidirectional,which does not satisfy the time dependence of time series,this thesis designs a mask to apply the Transformer encoder to the multi-step series prediction problem.This model is trained using the quantile loss function,which gives?10=0.1,?50=0.5,?90=0.9 quantile predicted values,quantifying the uncertainty of the prediction.The comparison results on two real datasets show that TransformerAR achieves better results.