Research And Implementation Of Multivariable Time Series Forecasting Algorithm Integrated By Data Augmentation

In the era of big data,time series data is widely present in people’s daily life.The evolution laws of things often exist in time series data.Using historical data to generate more accurate predictions has become a research hotspot in various application fields,attracting attention from academia and industry.Multiple sets of time series often coexist in practical applications.How to handle complex data patterns and the interdependence between series effectively is an important issue in the field of time series prediction.In addition,the insufficiency of time series data and information often leads to the failure of prediction.The problem of data quality raises a major challenge to the prediction task.Therefore,this thesis focuses on the field of time series forecasting,aiming to improve the performance of prediction algorithms and data quality.I try to explore the integration scheme of multivariate time series prediction algorithms and data augmentation methods.1)The idea of the hybrid model is introduced to address the complex internal and external relationships in multivariate time series.I propose a new type of parallelseries hybrid structure,which is "local series,global parallel".Then I propose the NP-ALSTM,a time series forecasting hybrid model based on NeuralProphet,LSTM,and attention and constructed by the parallel-series hybrid structure to integrate the strength of individual models.NP-ALSTM can effectively extract long and short-term periodic patterns within the series while also considering the interdependence between series.2)To address the problem of poor time series data quality,I propose a hybrid method of data augmentation in time series forecasting.This method adapts the standardization process to handle the original time series data.It generates new data using interpolation,jittering,and feature engineering methods,expanding the data scale and feature dimensionality simultaneously.This method is flexible,and easy-operating,and can be applied to time series data of different scales and fields,achieving the improvement of data quality.3)To build an effective bridge between algorithm research and practical application,I design and implement a time series forecasting visual system.This system is based on the Django framework and provides a web service for simulation scenarios of time series forecasting.It implements functions such as prediction,result visualization,model performance comparison,and data display.This system can present the research results of this thesis intuitively.I conduct comparative experiments on several public datasets to verify the effectiveness of our research.The experimental results show that the proposed NP-ALSTM has excellent forecasting capabilities and performs better than baseline models in time series prediction.The data augmentation hybrid method in time series forecasting can effectively improve data quality and support forecasting algorithms in generating more accurate predictions.In addition,I have implemented the time series forecasting visual system,which provides basic functions for time series prediction tasks,contributing to the development of time series forecasting research and applications.