Search

CN-121996960-A - Multivariable time sequence prediction method, system, device and storage medium

CN121996960ACN 121996960 ACN121996960 ACN 121996960ACN-121996960-A

Abstract

The invention provides a multivariable time sequence prediction method, a system, equipment and a storage medium, wherein the method comprises the steps of performing polynomial interpolation and minimum-maximum standardized preprocessing on original multivariable time sequence data; the method comprises the steps of constructing a multidimensional feature set comprising time features, holiday features and index history values based on preprocessing data, wherein a core model is a time network connected in series, each time block extracts a main period through fast Fourier transform, data of each period is remodelled into a two-dimensional tensor, the two-dimensional tensor extracts features through two-dimensional convolution and then reduces dimensions, the features are weighted and fused according to frequency intensity, residual errors among blocks are connected, a Prophet model is introduced to integrate the Prophet model with the Prophet model, trends and holiday effects are captured together, and finally the trained integrated model is utilized to output predicted values of all key indexes in a specified time period in the future. The method can process multivariable input simultaneously, avoid error accumulation, effectively model long-term dependence and capture mutation points, and remarkably improve the accuracy and stability of multi-index time sequence prediction.

Inventors

  • SHAO FENGFENG
  • ZHANG XIAOQIANG
  • GAO HAORAN

Assignees

  • 携程旅游网络技术（上海）有限公司

Dates

Publication Date
20260508
Application Date
20260120

Claims (10)

  1. 1. A method of multivariate time series prediction comprising the steps of: s110, carrying out missing value complementation and dimension unified processing on original time sequence data containing a plurality of key indexes, wherein the missing value adopts a polynomial interpolation method complementation algorithm, and the dimension unified processing adopts a minimum-maximum standardization method; S120, constructing a multi-dimensional feature set based on the preprocessed data, wherein the multi-dimensional feature set comprises time features, holiday features and historical value features of all key indexes, and the time features at least comprise whether a prediction day is a holiday, whether the prediction day is a working day, what day the prediction day is in the holiday, the day of the prediction day is far from the next holiday, the day of the prediction day is a day of the week, the week number of the year of the prediction day and the season of the prediction day; S130, adopting a time network as a core prediction model, wherein the time network consists of a plurality of time blocks which are sequentially connected in series, each time block performs fast Fourier transform on input one-dimensional time sequence characteristics to extract the most remarkable k periods, remodels one-dimensional time sequence data segments corresponding to each period into two-dimensional tensors, applies two-dimensional convolution operation on each two-dimensional tensor to extract the characteristic representation, reduces the dimension of the extracted two-dimensional characteristics back to one dimension, performs weighted summation according to the frequency intensity corresponding to each period, and fuses to obtain the output of the time block, and transmits information through residual connection among the time blocks; s140, integrating the Prophet model with the time network to jointly capture trend change points and holiday effect characteristics in the time sequence, and S150, outputting predicted values of all key indexes in a specified time period in the future by using the trained integrated model.
  2. 2. The method of predicting a multivariate time series according to claim 1, wherein in step S110, the formula of the polynomial interpolation completion algorithm is: Where P (x) is the interpolation polynomial, x is the abscissa of the point to be complemented, Is the abscissa of a known data point, Is a polynomial coefficient.
  3. 3. The multi-variable time series prediction method according to claim 2, wherein in the step S110, the formula used in the min-max normalization method is: where x is the original eigenvalue, x' is the normalized eigenvalue, and min (x) and max (x) are the minimum and maximum values of the feature in all samples, respectively.
  4. 4. A multi-variable time series prediction method according to claim 3, wherein in step S130, the internal processing of each time block specifically includes: s131, inputting one-dimensional time sequence characteristics through fast Fourier transformation Extracting period, selecting k frequencies with maximum intensity And the corresponding period And converts the one-dimensional data into a two-dimensional tensor: Wherein, the Representing the intensity of each frequency component, period representing the fast fourier transform and the process of selecting top-k frequency and Period, reshape representing the operation of reshaping one-dimensional data into a two-dimensional tensor, and Padding representing the zero Padding operation before convolution; s132, for each two-dimensional tensor Extracting features by using a two-dimensional convolutional neural network to obtain a characterized two-dimensional tensor: Wherein Inception denotes the adoption of Inception network structure; s133, an intermediate result dimension reduction sub-step, namely converting the two-dimensional characteristics back into a one-dimensional space: wherein trunk represents an operation of removing zeros supplemented by Padding; s134, carrying out weighted summation on the one-dimensional characterization after the dimension reduction according to the intensity of each frequency component to obtain the output of the time block: Wherein A is the frequency intensity and  is the normalized weight.
  5. 5. The multi-variable time series prediction method according to claim 4, wherein in the step S140, the time network is integrated with the prediction result of the Prophet model by using a stacking method.
  6. 6. The method according to claim 4, wherein in the step S140, the Inception network structure is inception _v3 network structure.
  7. 7. The method of multivariate time series prediction according to claim 1, wherein in step S150, the specified period of time is 30 days in the future.
  8. 8. A multivariate time series prediction method system for implementing the multivariate time series prediction method of claim 1, comprising: the data preprocessing module is used for carrying out missing value complementation and dimension unified processing on original time sequence data containing a plurality of key indexes, wherein the missing value adopts a polynomial interpolation method complementation algorithm, and the dimension unified processing adopts a minimum-maximum standardization method; The feature extraction module is used for constructing a multi-dimensional feature set based on the preprocessed data, wherein the multi-dimensional feature set comprises time features, holiday features and historical value features of each key index, and the time features at least comprise whether a prediction day is a holiday, whether the prediction day is a working day, what day the prediction day is in the holiday, the day of the prediction day is far from the next holiday, what day is the day of the week, the number of weeks of the year in which the prediction day is located and the season in which the prediction day is located; The model training module adopts a time network as a core prediction model, and comprises a plurality of time blocks which are sequentially connected in series, wherein each time block carries out fast Fourier transform on input one-dimensional time sequence characteristics to extract the most remarkable k periods, remodels one-dimensional time sequence data segments corresponding to each period into two-dimensional tensors; The model integration module is used for introducing a Prophet model to integrate with the time network so as to jointly capture trend change points and holiday effect characteristics in a time sequence, and And the prediction output module is used for outputting predicted values of all the key indexes in a specified time period in the future by using the trained integrated model.
  9. 9. A multivariable time-series prediction method apparatus, comprising: A processor; A memory having stored therein executable instructions of the processor; Wherein the processor is configured to perform the steps of the multivariate time series prediction method of any one of claims 1 to 7 via execution of the executable instructions.
  10. 10. A computer readable storage medium storing a program, wherein the program when executed by a processor implements the steps of the multivariate time series prediction method of any one of claims 1 to 7.

Description

Multivariable time sequence prediction method, system, device and storage medium Technical Field The invention relates to the field of travel itinerary customization, in particular to a multivariable time sequence prediction method, a multivariable time sequence prediction system, multivariable time sequence prediction equipment and a multivariable time sequence storage medium. Background In the internet industry, key metrics such as daily active user volume (DAU), order volume, sales, etc., directly affect corporate strategic decisions and operational planning. The future trend of the indexes is accurately predicted, and the method can help companies to layout marketing activities in advance and optimize resource allocation, so that the benefit maximization is realized. This is essentially a time series prediction problem, i.e. predicting future values based on historical data of the index. At present, methods for solving the problem of time series prediction are mainly divided into two main categories, namely a traditional time series model and a machine learning model. Traditional models such as moving average, autoregressive integral moving average (ARIMA) model, exponential smoothing method and the like have the advantages of mature theory, strong interpretability and the like. However, these models are typically limited to univariate predictions, and it is difficult to handle multivariate inputs and interactions between variables. In the prediction scene of a plurality of key indexes, if a traditional model is used, a model is required to be independently built and optimized for each index, and the workload and the cost are high. In addition, when the conventional model performs multi-step prediction, a rolling prediction strategy is often adopted, that is, a predicted value of a previous period is input as an actual value to predict a next period, and the method can cause continuous accumulation of prediction errors and decline of long-term prediction accuracy. Deep learning-based methods, such as cyclic neural networks (RNNs), time-series convolutional networks (TCNs), and transformers, have been widely used for time-series prediction tasks. The methods support multiple variable input and output, can adaptively extract characteristics, and can directly perform multi-step prediction. However, RNN and TCN approaches have limited ability to capture long-term time-dependent relationships. While Transformer is adept at modeling for long periods of time, it relies on the mechanism of attention between discrete points in time, and it may be difficult to mine out robust timing dependencies for complex and diverse timing patterns in the real world. In recent years, with the rise of large model technology, time series prediction large models (such as Time-LLM, chronos, etc.) are continuously emerging. These models are capable of integrating multimodal information and providing predictive interpretation. However, the ubiquitous "illusion" problem of large models makes their predicted results extremely sensitive to input disturbances, facing stability and reliability challenges in practical industrial scenarios. Accordingly, the present invention provides a method, system, apparatus, and storage medium for multivariate time series prediction. Disclosure of Invention Aiming at the problems in the prior art, the invention aims to provide a multivariable time sequence prediction method, a system, equipment and a storage medium, which overcome the difficulties in the prior art, can simultaneously process multivariable input, avoid error accumulation, effectively model long-term dependence and capture mutation points, and remarkably improve the accuracy and stability of multi-index time sequence prediction. The embodiment of the invention provides a multivariable time sequence prediction method, which comprises the following steps of: s110, carrying out missing value complementation and dimension unified processing on original time sequence data containing a plurality of key indexes, wherein the missing value adopts a polynomial interpolation method complementation algorithm, and the dimension unified processing adopts a minimum-maximum standardization method; S120, constructing a multi-dimensional feature set based on the preprocessed data, wherein the multi-dimensional feature set comprises time features, holiday features and historical value features of all key indexes, and the time features at least comprise whether a prediction day is a holiday, whether the prediction day is a working day, what day the prediction day is in the holiday, the day of the prediction day is far from the next holiday, the day of the prediction day is a day of the week, the week number of the year of the prediction day and the season of the prediction day; S130, adopting a time network as a core prediction model, wherein the time network consists of a plurality of time blocks which are sequentially connected in series, each time block