Time sequence prediction method based on differential fusion Transform

Abstract

The invention discloses a time sequence prediction method based on differential fusion Transform, and the method comprises the steps: designing a differential layer, a neighbor attention mechanism, a sliding fusion mechanism and a residual layer on the basis of a classical Transform architecture through employing an encoder and decoder structure; specifically, feature differences of adjacent time points of time series data are extracted and focused through a difference layering and neighbor attention mechanism, various nonlinear features of each time point are fused through a sliding fusion mechanism, multi-granularity key features in a multivariable time series can be effectively extracted, and a core component further comprises a one-dimensional convolution and LSTM fusion residual layer. Therefore, the mutual dependency relationship between time points of the time series data is further learned, and the deep feature learning capability of the model on the complex multivariable time series data is improved. Compared with an existing method, the method has the advantages of being good in stability, high in prediction precision, high in generalization ability and the like.

Description

technical field [0001] The invention relates to a time sequence prediction method based on differential fusion Transformer, and belongs to the technical field of time sequence prediction. Background technique [0002] In recent years, with the rapid application of the Internet of Things and the rapid development of big data and artificial intelligence, the operation and maintenance of smart city traffic is moving towards an autonomous, large-scale and intelligent direction. With the explosive growth of urban spatiotemporal data, traditional management methods can no longer meet the business needs of the digital age. Therefore, it is urgent to develop more autonomous and efficient technical solutions to optimize traffic resources and improve the availability and stability of resources. It is necessary to continuously measure and monitor various time series data. Through the prediction of various time series variables, time series trends can be found in time, early warning and...

AI Technical Summary

Problems solved by technology

[0007] To sum up, the existing time series data prediction methods still have shortcomings such as weak ability to control the overall situation, insufficient learning of time dependence, and poor generalization.

Images