Research And Application Of Multistep Prediction Method For Time Series Data Based On RNN

Time series data is defined as a set of observations that are continuously measured at equal time intervals.It appears in many real-life applications,such as economics and finance,industrial production,resources and environment,and scientific innovation.We can achieve the description and explanation of time series data by analyzing the potential laws and mechanisms of time series data,to better understand the phenomena in various fields and explore the scientific value of them.Furthermore,mathematical modeling of the time series data can be used to achieve accurate prediction,thereby providing decision-makers with reference and forward-looking guidance in different application fields.With the advent of the era of big data,time series data is constantly updated iteratively,reflecting more and more new features.The explosive growth of data scale,the continuous expansion of dimensions,and the complexity of the structure have brought great challenges to the forecasting task.On the one hand,the traditional time series data prediction model has the problem that the accuracy of multi-step prediction decreases with increasing time step.In order to ensure the accuracy of predictions,the more conventional prediction methods usually carry out single-step prediction modeling.However,in reality,we often need to obtain the predicted values for a longer period of time in the future to provide support for practical applications.In this case,You need to carry out multi-step prediction work.In the current research on multi-step prediction algorithms,most algorithms use fusion single-step and multi-step direct prediction,which will cause the accumulation of prediction errors and greatly reduce the accuracy of prediction;on the other hand,traditional time series data prediction models have training The problem of long time and large time overhead.In the case of massive data,it cannot adapt to the sequence data whose distribution changes with time,and cannot realize real-time processing and short-term response.In order to solve the above two problems,this paper proposes a time series data multi-step prediction model based on VMD-Hybrid-RNN,to improve the accuracy of multi-step prediction,get a longer period of sequence prediction value,It provides a better and feasible solution for capturing the key information of time series data in advance.The main research content of this article includes the following three parts:1.By analyzing the effects and advantages and disadvantages of wavelet,STL,empirical mode,and variational mode decomposition methods,select the optimal variational mode decomposition method as the time series data processing method in this paper.Using this method to extract the frequency characteristics of the original data,to solve the shortcomings of empirical modal decomposition method can not deal with modal aliasing phenomenon,and according to the extracted effective variational modal components were substituted into the recurrent neural network model for training.2.On the basis of the previous step,use the integrated learning method to model the different frequency sequences obtained by decomposition using RNN related models,and at the same time readjust the underlying optimization parameters for each sequence model to make the optimization environment smoother.Allow any gradient-based training algorithm to take a larger Step without encountering sudden changes in Loss,and iteratively train to achieve continuous updating of network structure weights and parameters.Finally,the sequence of each prediction result is reconstructed to further reduce the model training error.3.Build ARIMA,SVR,BPNN,GRU and other models to compare the prediction accuracy with the VMD-Hybrid-RNN model to verify the superiority of the model proposed in this article.In addition,the climate data daily value data set(V3.0)was used to verify the model's adaptability,and to explore the prediction effect and performance evaluation of the model in the field of multivariate time series data.Relevant experiments show that the above method can solve the problem that the traditional multi-step prediction scheme cannot effectively decompose the input data and extract important information,and the problem that the prediction error will accumulate during the prediction process,resulting in the decline of the multi-step prediction accuracy.By horizontally comparing the model proposed in this paper with other prediction models,we can find that the model in this paper has higher time efficiency and accuracy,and can also avoid possible gradient disappearance and internal covariate shifts,which has a great advantage in multi-step time series data prediction.At the same time,using the idea of integrated learning,the prediction model of each component can be calculated and optimized in parallel to ensure that the accuracy of the model is further improved under the condition of a certain time efficiency.