A Study Of Long Time-series Forecasting Based On Deep Learning

In recent years time-series forecasting has become a key research task in time-series data mining and analysis.Time series data,which can be found everywhere in daily life,is a series of ordered sets of points that describe trends in events in chronological order.These data are affected and disturbed by many factors.Time series forecasting is to pre-process and analyze the data according to their implied characteristics,and build model algorithms to find the intrinsic laws of the data,and finally use the model to forecast the trend of the series.In the face of massive data,it is of great research significance to establish accurate and effective time series prediction models,which will have a great impact on production life.Time series forecasting,a popular research topic in analytics today,is critical in areas such as climate change,financial markets,energy consumption,and traffic flow.In the above applications,accurate prediction with extended prediction time has become a pressing problem when making data-driven decisions.Therefore,this paper conducts a study of the long-time series forecasting problem by deep learning.The work in this paper is mainly as follows:(1)A forecasting model based on Transformer with dynamic convolution is designed.It first extracts valid information by downsampling the sequence.Secondly,a feature mapping module(F-Map)is constructed to adjust the sequence feature distribution and generate more feature information revealing the intrinsic association through low-cost linear operations on the time series.It not only compensates for the feature loss caused by the sparse mechanism,but also reduces the redundancy of feature sequences to reduce the encoder computation.Finally,the conditional convolutional network(Con Net)is built to replace the static fully connected paradigm by dynamically learning convolutional kernels.It learns specific convolutional kernels for each input sequence,increasing network capacity while maintaining efficient inference,and improving model performance while maintaining coupling relationships between sequences.Experimental results on five real datasets show that the model outperforms other Transformer variants in all aspects by producing different degrees of reduction in the evaluation metrics(MSE)for both multivariate and univariate time-series predictions,and by effectively reducing the runtime and number of parameters compared to the advanced model.(2)A decomposition-based DFNet prediction model was constructed.The general timeseries has some similarity between the subsequences of each cycle,which can effectively improve the prediction performance through the timing characteristics.Therefore,complex sequence patterns within the sequence are separated by decomposition architectures to mine long-term trend cycle similarities.The traditional decomposition is improved by using least squares adjustment in obtaining the seasonal pattern characteristics,which is closer to the true seasonal characteristics than the traditional average adjustment.Secondly,different sequence patterns are processed individually to obtain the internal regular,periodic and irregular perturbations precisely,and specific filters,linear layers and differential processing are used to extract the key information effectively.At the same time,a segmented polynomial activation function is proposed to solve the difficult problem of negative data loss and slow operation speed.And the computational effort is small.Finally,through the information fusion transfer mechanism,the extracted long-term dependencies are passed to each other for aggregation to enhance the coupling between sequences,which achieves the purpose of improving the prediction accuracy of the model.By conducting experiments on nine datasets,the prediction performance of the new model was significantly improved compared with the existing prediction model.The results demonstrate the superior performance of the model for long time series prediction.