Research On Multi-step Forecasting Algorithm Of Time Series Based On Memory Network

In recent years,with the rapid development of information technology,problems related to the prediction of time series data continue to emerge in the fields of finance,transportation,and environment.Different application scenarios have different processing requirements for time series data,and these problems cannot be solved by traditional time series methods.For example,in the study of multivariate time series forecasting problems,traditional time series forecasting methods are difficult to model time series data with high complexity and nonlinearity,and it is difficult to capture the time dependence of data.For non-stationary and catastrophic multivariate time series,how to consider the interaction between various variables and capture the dependent patterns in the data is still a big challenge for multivariate time series prediction.Aiming at the problem of poor performance of non-stationary sequence prediction,this paper proposes a time series single-step prediction(MNTS)model based on memory network.By improving the loss function,a multi-step prediction model(MNTS-MS)is further proposed.The content and results are as follows:(1)Considering the different importance of data in different historical periods,this article divides the input data into two parts,one part is short-term historical data and the other part is long-term historical data,which are used to capture the short-term dependence and long-term dependence patterns in the data.On this basis,the time series single-step prediction model MNTS is proposed.MNTS consists of two encoders with the same structure and a memory component.The encoder can effectively extract local and global features of the data and model the correlation between variables.The memory component is used to store the encoder code.In addition,a temporal attention model is designed to extract the information stored in the memory component by embedding short-term historical data,which makes the MNTS model have certain interpretability,and it is easy to know which part of the historical data is more important to the prediction results..(2)For scenarios with short data intervals and the need to predict multiple time values in the future,this paper further studies the problem of time series multi-step prediction based on the above single-step prediction.Non-stationary sequence multi-step prediction is easy to cause error amplification and the accuracy is not high.Therefore,the loss function is improved based on the single-step prediction MNTSmodel,and a multi-step prediction MNTS-MS model is proposed.Consider combining the shape error,time error and mean square error between the predicted sequence and the real sequence as its loss function,so that the improved loss function is adapted to the overall difference between the evaluation sequences,and is no longer limited to the evaluation points Individual differences with points.(3)This paper selects a standard data set of a multivariate time series and a real data set of a train braking system,and experiments are carried out on the single-step prediction model MNTS and the multi-step prediction model MNTS-MS.The single-step and multi-step prediction effects of the proposed MNTS and MNTS-MS models on multivariate time series are better than the existing prediction methods,indicating that the model proposed in this paper is effective in modeling non-stationary sequences and can be better Mining the dependent patterns in time series data.In this paper,the proposed multi-step prediction model MNTS-MS is used as a fault warning engine,which has been practically applied in the train brake system fault warning system.