Research On Time Series Data Modeling Method For Internet Of Things

With the continuous development of the Internet of Things(Io Ts)technology,the time series data that people can obtain is increasing day by day,which not only provides basic data for the development of the time series field,but also poses new challenges to the technical development of this field.Long-term and accurate forecasting of future data helps people make more reasonable planning and reduce the waste of resources.How to accurately complete the long-term prediction of data in various application scenarios has become a hot issue in current research.Aiming at the time series data modeling problem of the Io Ts,this thesis focuses on the multivariate long-sequence time series forecasting problem,aiming to propose a time series forecasting framework suitable for multiple Io Ts scenarios,achieve high forecasting performance and apply in scenarios of time series forecasting.Specific research contents are as follows:(1)Aiming at the time series forecasting problem of multivariate long series,this thesis proposes a Transformer-based prediction model.On the basis of this model,temporal convolutional network is introduced to make up for the local insensitivity of the Transformer,so that the model can capture the time dependence of data(short-term dependence and long-term dependence); Graph convolutional network is also introduced to dynamically learn the relationship between variables relationship,thereby extracting the inter-variable dependencies of the data.(2)Aiming at the forecasting problems with complex data dependencies,in order to further improve the predictability of data and improve the prediction performance of the model,this thesis introduces the data decomposition module,data reconstruction module and data interaction module,so as to develop the model into a decomposition-based multivariate long-sequence time series prediction hybrid framework.In this framework,the data decomposition module is used to process data with complex patterns and decompose it into more predictable component data; the data reconstruction module is to reconstruct the predicted component data into target data; the data interaction module is responsible for interacting different components of information,so as to learn the relationship between them,and improve the prediction performance of the model under the combination of the three modules.Among them,the method of data decomposition supports dynamic increase,and the appropriate decomposition method can be selected according to different application scenarios,which further improves the flexibility of the framework.(3)Aiming at the application of Transformer in time series prediction,analyzed its shortcomings and its specific role in the training process,and proposed an optimization method using convolutional neural network to replace the multi-head attention mechanism,which reduces the FLOPs and the parameters of the model to improve the calculation speed of the model.At the same time,the introduction of convolutional neural network also introduces effective bias information to the Transformer,which enhances the adaptability of the model on small datasets.