Research On Multivariate Time Series Forecasting Based On Time Series Decomposition

Multivariate time series forecasting is widely used in social production and daily life and provides predictive data support for various decision-making processes.In recent years,deep learning technology has effectively promoted the development of multivariate time series forecasting.However,it still has some problems,such as the difficulty of mining complex data with long-term dependency,the bottleneck of time context perception,the high consumption of computing resources and the challenges of model universality.To solve the above problems,this paper studies the time series decomposition and information utilization efficiency.The main research contents are:(1)Based on the multi-head attention mechanism and decomposition method,the decomposition collaboration model Segformer is proposed.Firstly,to solve the problem of multi-time pattern mixing in complex data,Segformer decompose the original time series into three types of pattern information integrating global attributes and local attributes by Fast Fourier Transform(FFT)technology.Through the information collaboration in the feature dimension,the modeling process of pattern information is coordinated,which improves the generalization and robustness of the model.Secondly,aiming at the problems of context awareness bottleneck,large consumption of computing resources and difficulty in capturing long-term dependency relationship,Segformer combines the recessive periodic representation with multi-head attention mechanism,extracts segment sequences and highdimensional features into high-level features by using two-dimensional convolution operation,which reduces computing load and established an efficient segment-level information aggregation mechanism.Finally,experiments on seven data sets show that the mean square error of Segformer is about 19%,10% and 13% lower than the three benchmarks,respectively,and Segformer improves the prediction accuracy.(2)Based on the decomposition method and the modeling method of frequencydomain information,a frequency-domain information processing model FRQformer is proposed.Firstly,a new adaptive decomposition algorithm based on FFT is proposed to solve the problem of the diversity of datasets in various fields and the complexity of data dependency,which also avoids the requirement of prior knowledge.Secondly,in view of the large consumption of computing resources in Segformer and the error introduced by using covariable to pre-fill the interval to be predicted,FRQformer designs a lightweight information mapping(extrapolation)module and a "preprocess-encode-extrapolate-decode" model structure by combining FFT,element multiplication and frequency domain shrinking operation.The pre-filling method is abandoned,and the time dependency of the sequence is modeled efficiently,which improves the generalization of the model.Finally,experiments show that,while maintaining good prediction effect,the cost of time and space of FRQformer is reduced by more than 35% compared with Segformer,and the performance of FRQformer is better than other benchmark models in many aspects,and thus the practicability is greatly improved.(3)Based on the proposed multivariate time series forecasting methods and scene requirements,a meteorological index forecasting system in the field of meteorological forecasting is designed and implemented.The system supports the forecasting of specific meteorological indicators and the training of specific forecasting models,and provides support for data management and model management.The system has certain practical significance.