A Study Of Deep Learning Based Multivariate Time Series Forecasting Methods

Time series are data recorded in chronological order at equally interval time points.With the advent of the era of big data,massive amounts of time series data are generated in the fields of energy management,economic activities,weather forecasting and disease trans-mission prediction.It is of great practical importance to use large-scale time series data to provide support for advance planning or decision making.Time series are usually com-posed of the superposition of multiple types of temporal patterns(seasonal,trend,cyclical and irregular components),increasing the difficulty of accurately capturing multi-scale time dependencies and correlations among variables.The main challenge of multivariate time series(also known as multivariate time series)forecasting is to accurately extract the multi-scale dependencies of the series as well as the correlations among variables.Current deep learning-based forecasting methods mainly capture dependencies in entangled temporal pat-terns using generic sequence modeling methods,which are not conducive to improving the accuracy of long-term forecasting.In addition,current approaches tend to design more com-plex models,which makes the prediction accuracy vulnerable to redundant information in the series and increases the model overhead.This study aims to improve the accuracy and efficiency of multivariate time series long-term prediction,and establishes a method system to improve the accuracy and efficiency of long-term time series prediction based on deep learning methods,autocorrelation mechanisms and time series decomposition methods,and validates the effectiveness of the proposed method on time series datasets with periodic and aperiodic and different sampling frequencies.The research content of this dissertation is summarized as follows:(1)A hybrid time series forecasting method based on convolutional neural networks and locally grouped autocorrelation.To address the problem of weak multi-scale time dependence extraction ability of the combined method of autocorrelation mechanism and Transformer architecture,a hybrid prediction method combining convolutional neural net-works and locally grouped autocorrelation mechanism is proposed to enhance the multi-scale temporal dependence extraction ability of the model.This scheme solves the problem of the autocorrelation mechanism being unable to capture the local dynamics and trends of non-periodic sequences through multi-layer convolutional neural networks.The locally grouped autocorrelation mechanism captures periodic time patterns in time series fragments,improv-ing the model’s ability to capture multi-scale time repeating patterns.In addition,the time series decomposition method decomposes the sequence into seasonal and trend components,avoiding interference caused by multiple time patterns in accurately extracting dependen-cies and improving the interpretability of the model.The experimental results show that the proposed method achieves a relative prediction accuracy improvement of 11.75%.On the dataset with no significant periodicity,the proposed method achieves a prediction accuracy improvement of 18.89%.(2)A time series forecasting method based on convolutional attention mechanism and time series decomposition.Aiming at the lack of self-learning ability of the autocorrelation mechanism and the current prediction method mainly focusing on the dependency capture in the time dimension and ignoring the correlation between variables,a forecasting model named CNformer based on Transformer architecture is proposed to capture time dependen-cies on dimensions and variable dimensions.Different from existing Transformer-based forecasting methods,CNformer uses dilated convolutional networks to extract the seasonal patterns of time series and the dependencies between variables,and uses the historical sea-sonal patterns extracted by the encoder to optimize the seasonal and trend components in the forecast sequence.The time series decomposition component separates the seasonal compo-nents with obvious periodicity from the series,enabling the convolutional neural network to take full advantage of its ability to capture temporally repetitive patterns.Experimental re-sults show that the proposed method achieves a 20.29% improvement in prediction accuracy.CNformer achieves 28.34% prediction accuracy improvement on datasets without obvious periodicity.Experimental results demonstrate that CNformer can significantly improve the long-term forecasting performance of multivariate time series.(3)A lightweight forecasting method for time series based on the modelling of temporal patterns and correlations between variables.Aiming at the problem that the prediction model based on Transformer architecture has a large space overhead and the redundant in-formation in the sequence affects the prediction accuracy,a concise and efficient time series prediction model SDCNet based on convolutional neural networks is designed.Unlike the point-wise dependence discovery of the self-attention mechanisms and the sequence-level dependence discovery of the autocorrelation mechanism,the multilayer convolutional neu-ral networks in SDCNet achieve multi-scale temporal pattern extraction at both time series elements and subsequences granularity,which improves the prediction accuracy and reduces the redundancy of temporal features.Experimental results show that the proposed method significantly reduces the space and time overhead of the model while achieving performance improvements.SDCNet achieves 16.73% prediction accuracy improvement and 23.87%prediction accuracy improvement on the dataset without obvious periodicity.