Prediction system, prediction method, and prediction program

Abstract

To provide a prediction system, a prediction method, and a prediction program capable of improving prediction accuracy of time series information.SOLUTION: The prediction system includes a parameter determination unit 114 that determines a parameter defining an end of a range of target time series information on which variational mode decomposition is performed, a variational mode decomposition unit 112 that performs variational mode decomposition on the target time series information in the range defined by the determined parameter, a time series information prediction unit 120 that predicts time series information for each mode subsequent to each of the time series information of a plurality of modes using the time series information of the plurality of modes obtained by performing the variational mode decomposition, a time series information addition unit 131 that adds the predicted time series information for each mode, and a time series information output unit 132 that outputs the time series information obtained by adding the time series information for each mode. The parameters may be determined using a trained model.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to a prediction system, a prediction method, and a prediction program for predicting time-series information. [Background technology]

[0002] Various techniques for predicting time series information have been proposed. Non-Patent Document 1, as one example of such a technique, discloses a new integrated model, the Variational Mode Decomposition-Autoregressive Integrated Moving Average (VMD-ARIMA) model, which reduces the required data input and improves the accuracy of predictions based on complex mechanisms related to data dependency and prediction. This model uses variational mode decomposition (VMD) to extract trend features and detailed features contained in time series information and remove noise. Then, a corresponding autoregressive integrated moving average (ARIMA) model is created to reflect the different characteristics of each component.

[0003] Furthermore, Patent Document 1 discloses a simulation device for selecting an appropriate model even when actual measurement values ​​are unknown in a design support device. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2022-185713 [Non-patent literature]

[0005] [Non-Patent Document 1] Huan Wang, 4 others, “An Integrated Variational Mode Decomposition and ARIMA Model to Forecast Air Temperature”, [online], [November 27, 2020], Internet (URL: https: / / www.mdpi.com / 2071-1050 / 11 / 15 / 4018) Summary of the Invention [Problem to be solved by the invention]

[0006] However, the results of feature extraction of time series information vary depending on the setting of the endpoint (EP), which defines the end of the range of time series information for which variational mode decomposition is performed. For example, as shown in Figure 5, if two endpoints, EP1 and EP2, are set for the same time series information, the envelope of the time series information values ​​at EP1 will be different from the envelope of the time series information values ​​at EP2. Therefore, even if the time series information is the same, the mode value, which is the average value of the envelope of the time series information at EP1, will be different from the mode value of the time series information at EP2. Therefore, if the optimal endpoint cannot be set, the features of the time series information cannot be accurately extracted, resulting in a problem of low forecast accuracy of the time series information.

[0007] The present disclosure provides a prediction system, a prediction method, and a prediction program that can improve the prediction accuracy of time-series information. [Means for solving the problem]

[0008] The prediction system for predicting time-series information according to the present disclosure includes: a parameter determination unit that determines a parameter that defines the end of a range of time series information on which variational mode decomposition is performed; a variational mode decomposition unit that performs variational mode decomposition on target time-series information within a range defined by the parameters determined by the parameter determination unit; a time series information prediction unit that predicts time series information for each mode subsequent to each of the time series information for the plurality of modes, using the time series information for the plurality of modes obtained by executing the variational mode decomposition; a time series information adding unit that adds the time series information for each mode predicted by the time series information predicting unit; a time-series information output unit that outputs time-series information obtained by adding the time-series information for each mode by the time-series information adding unit, The parameter determination unit determines parameters that minimize the average entropy of the time-series information of each mode obtained by performing variational mode decomposition on the target time-series information.

[0009] The prediction system may further include a mode number determination unit that determines the number of modes that minimizes the average entropy; The variational mode decomposition unit decomposes the target time series information into time series information of the number of modes determined by the mode number determination unit.

[0010] The method for predicting time-series information according to the present disclosure includes: determining parameters that minimize an average entropy of time series information of each mode obtained by performing variational mode decomposition on the time series information of the target, the parameters specifying the end of a range of the time series information of the target on which variational mode decomposition is performed; Performing variational mode decomposition on the time series information of the target within a range defined by the determined parameters; predicting time series information for each mode subsequent to each of the time series information for the plurality of modes using time series information for the plurality of modes obtained by performing variational mode decomposition; Add the time series information for each predicted mode, The time series information obtained by adding the time series information for each mode is output.

[0011] In addition, the computer Determine the number of modes that minimizes the mean entropy, The time series information of the object may be decomposed into time series information of a determined number of modes.

[0012] A prediction program for predicting time-series information executed by a computer according to the present disclosure includes: determining parameters that minimize the average entropy of time series information of each mode obtained by performing variational mode decomposition on the time series information of the target, the parameters defining the end of a range of the time series information of the target on which variational mode decomposition is performed; A step of performing variational mode decomposition on time series information of the target within a range defined by the determined parameters; a step of predicting time series information for each mode subsequent to each of the time series information for the plurality of modes, using the time series information for the plurality of modes obtained by performing variational mode decomposition; adding up time series information for each predicted mode; and outputting the time-series information obtained by adding the time-series information for each mode. [Effects of the Invention]

[0013] The present disclosure makes it possible to provide a prediction system, a prediction method, and a prediction program that can improve the prediction accuracy of time-series information. [Brief explanation of the drawings]

[0014] [Figure 1] FIG. 1 is a block diagram illustrating a configuration of a prediction device according to the present disclosure. [Figure 2] FIG. 1 is a conceptual diagram illustrating variational mode decomposition according to the present disclosure. [Figure 3] FIG. 10 is a diagram illustrating an example of processing executed by a prediction device according to the present disclosure. [Figure 4] FIG. 10 is a diagram illustrating another example of processing executed by the prediction device according to the present disclosure. [Figure 5] FIG. 10 is a diagram for explaining the influence of the setting of the end point of the time-series information on feature extraction. DETAILED DESCRIPTION OF THE INVENTION

[0015] An embodiment will be described below with reference to the drawings. FIG. 1 is a block diagram showing the configuration of a prediction device 10 according to the present disclosure. The prediction device 10 is a device that uses various types of time-series information to predict time-series information that follows the time-series information. The prediction device 10 corresponds to a prediction system.

[0016] The prediction device 10 includes a processor 100, a communication interface (I / F) 140, and a storage device 150. The communication interface 140 is an interface for data communication between the prediction device 10 and other devices. The storage device 150 is a storage device that stores various data such as programs executed by the prediction device 10 and time-series information. The time-series information processed by the prediction device 10 can be stored in the storage device 150 via the communication interface 140.

[0017] The processor 100 is a processor such as a CPU (Central Processing Unit), GPU (Graphics Processing Unit), or MPU (Micro Processing Unit) that controls electronic circuits and devices included in the prediction device 10. The processor 100 executes a method defined by a program by reading the program from the storage device 150 and executing the program. The program executed by the processor 100 may be executed by an integrated circuit such as an FPGA (Field-Programmable Gate Array) or an ASIC (Application Specific Integrated Circuit). The processor 100 and these integrated circuits correspond to a computer.

[0018] The processor 100 includes a variational mode decomposition module 110, a sub-prediction model creation module 120, and a sub-prediction model fusion module 130. These modules can be realized by a program.

[0019] The variational mode decomposition module 110 includes an acquisition unit 111 , a variational mode decomposition unit 112 , an entropy calculation unit 113 , a parameter determination unit 114 , and an extraction unit 115 .

[0020] The acquisition unit 111 acquires time series information used to predict time series information (hereinafter referred to as "target time series information"), time information indicating time T, and the decomposition number K of the variational mode decomposition. The acquisition unit 111 can acquire this information from the storage device 150 or an external device. Specific examples of the target time series information include time series payment information in online payments and time series traffic volume in a network system. Note that the target time series information is not limited to these pieces of information, and various other time series information can be adopted. Furthermore, the target time series information is not limited to continuous information, and may include discrete information.

[0021] The time T is a time that defines the extraction range of time-series information by the extraction unit 115, which will be described later. The decomposition number K of the variational mode decomposition corresponds to the number of modes generated by the variational mode decomposition performed by the variational mode decomposition unit 112. Any number can be specified as the decomposition number K.

[0022] The variational mode decomposition unit 112 executes variational mode decomposition to decompose time-series information within a range defined by the parameters determined by the parameter determination unit 114 into time-series information of a plurality of modes. As shown in Fig. 2, each of the plurality of modes obtained by executing the variational mode decomposition includes time-series information. Note that the variational mode decomposition of the time-series information can be performed using vmdpy (Variational mode decomposition) in Python (registered trademark) or the like.

[0023] The entropy calculation unit 113 calculates the entropy of the time-series information of each mode using the time-series information of each mode generated by the variational mode decomposition. In this embodiment, for example, envelope entropy or sample entropy can be used as the entropy.

[0024] The parameter determination unit 114 determines parameters used in the variational mode decomposition performed by the variational mode decomposition unit 112. The parameters indicate endpoints (EPs) that define the end of the range of time-series information for which the variational mode decomposition is performed, as shown in Fig. 2. In other words, the parameters are parameters that control behavior.

[0025] More specifically, the parameter determination unit 114 determines parameters that minimize the average entropy of time-series information of each mode obtained by performing variational mode decomposition on the target time-series information. For example, the parameter determination unit 114 can determine parameters that minimize the average entropy of each mode when the target time-series information is decomposed, as shown in Equation 1. As a method for deriving optimal parameters that minimize the average entropy, metaheuristic algorithms such as particle swarm optimization (PSO) and Beluga Whale optimization (BWO) can be adopted. In other words, the equation shown in Equation 1 is used as the objective function, and parameters indicating the endpoints are optimized as control variables.

number

[0026] The extraction unit 115 extracts time series information of each mode from a predetermined time that defines the start point of the range of time series information for which variational mode decomposition is performed to time T. Any time before time T can be used as the predetermined time.

[0027] The sub-prediction model creation module 120 uses time series information of multiple modes obtained by executing variational mode decomposition to predict time series information for each mode that follows each of the time series information of the multiple modes. In other words, the sub-prediction model creation module 120 predicts time series information for each mode from time T onwards. The processing performed by the sub-prediction model creation module 120 will be described later with reference to FIG. 4.

[0028] The sub-prediction model fusion module 130 includes a time-series information addition unit 131 and a time-series information output unit 132 .

[0029] The time series information addition unit 131 adds the time series information for each mode predicted by the sub-prediction model creation module 120. Specifically, the time series information addition unit 131 adds the time series information for each mode that has the same date and time or time.

[0030] The time-series information output unit 132 outputs the time-series information obtained by the time-series information adder 131 adding the time-series information for each mode.

[0031] 3 is a diagram illustrating an example of processing executed by the prediction device 10. In step S1, the acquisition unit 111 of the variational mode decomposition module 110 acquires target time series information, time information indicating time T, and the decomposition number K.

[0032] In step S2, the variational mode decomposition unit 112 of the variational mode decomposition module 110 performs variational mode decomposition on the time series information within a range defined by parameters indicating an arbitrary endpoint, to generate time series information for each mode. The arbitrary endpoint is a point corresponding to a time before time T.

[0033] In step S3, the entropy calculation unit 113 of the variational mode decomposition module 110 calculates the entropy of the time-series information of each mode using the time-series information of each mode generated in step S2.

[0034] In step S4, the parameter determination unit 114 of the variational mode decomposition module 110 determines parameters that minimize the average entropy of the time-series information of each mode calculated in step S3.

[0035] In step S5, the variational mode decomposition unit 112 of the variational mode decomposition module 110 performs variational mode decomposition on the time series information within the range defined by the parameters determined by the parameter determination unit 114. In step S6, the extraction unit 115 of the variational mode decomposition module 110 extracts the time series information of each mode from a predetermined time to a time T indicated by the time information.

[0036] In step S7, the sub-prediction model creation module 120 executes the process shown in Fig. 4 to construct an ARIMA model for each mode using the time series information for each mode obtained by executing variational mode decomposition. In step S8, the sub-prediction model creation module 120 uses the ARIMA model for each mode to predict subsequent time series information for each mode that continues to time T.

[0037] In step S9, the time series information addition unit 131 of the sub-prediction model fusion module 130 adds the subsequent time series information for each mode predicted by the sub-prediction model creation module 120. In step S10, the time series information output unit 132 of the sub-prediction model fusion module 130 outputs the time series information obtained by the time series information addition unit 131 adding the subsequent time series information for each mode, and the processing in FIG. 3 ends.

[0038] 4 is a diagram illustrating an example of processing executed by the sub-prediction model creation module 120. In step S11, the sub-prediction model creation module 120 performs an Augmented Dickey Fuller Test (ADF) on the time-series information of mode k identified by k, where k is an integer greater than or equal to 1 and less than or equal to the number of decompositions K.

[0039] In step S12, the sub-prediction model creation module 120 determines whether the time series information of mode k is stationary based on the result of the ADF test on the time series information of mode k. If it is determined that the time series information of mode k is stationary (YES), the process branches to step S17. On the other hand, if it is determined that the time series information of mode k is not stationary (NO), the process branches to step S13.

[0040] In step S13, i+1 is set to a variable i representing the differential order. The initial value of the variable i is 0. In step S14, the sub-prediction model creation module 120 calculates the i-th order differential information of the time series information of mode k. The i-th order differential information of the time series information of mode k is obtained by repeatedly calculating the difference between two pieces of time series information adjacent in time i times. More specifically, when the time series information of mode k is x _1 ,x _2 ,…x _t When this is the case, the first difference is calculated as follows: _1 =x _2 -x _1 , x' _2 =x _3 -x _2 ,…,x' _t-1 =x _t -x _t-1 The method for calculating the second difference is x'' _1 =x' _2 -x' _1 , x'' _2 =x' _3 -x' _2 ,…,x'' _t-2 =x' _t -x' _t-1 Similarly, the i-th difference can be derived.

[0041] In step S15, the sub-prediction model creation module 120 performs an ADF test on the i-th order difference information of the time-series information of mode k calculated in step S14.

[0042] In step S16, the sub-prediction model creation module 120 determines whether the i-th order difference information of the time-series information of mode k is stationary based on the result of the ADF test on the i-th order difference information of the time-series information of mode k. If it is determined that the i-th order difference information of the time-series information of mode k is stationary (YES), the process branches to step S17. On the other hand, if it is determined that the i-th order difference information of the time-series information of mode k is not stationary (NO), the process returns to step S13.

[0043] In step S17, the sub-prediction model creation module 120 constructs an ARIMA model for each mode. More specifically, the sub-prediction model creation module 120 constructs the ARIMA model shown in Equation 2. Here, the sub-prediction model creation module 120 sets the autoregressive order from 1 to p for the i-th differential information. max , the moving average rank is 1 to q max Set to p max denotes the maximum autoregressive rank. q max represents the maximum moving average rank.

number

[0044] In step S18, the sub-prediction model creation module 120 calculates the AIC (Akaike Information Criterion) of each ARIMA model constructed in step S17. Specifically, the sub-prediction model creation module 120 can calculate the AIC of the ARIMA model based on Equation 3.

number

[0045] In step S19, the sub-prediction model creation module 120 calculates the optimal value (p * ,q * ) and minimize AIC (p * ,q * ) is output as the ARIMA regression result for mode k. * ,q * are the optimal values ​​of the autoregressive rank p and the moving average rank q, respectively.

[0046] In this way, the sub-forecasting model creation module 120 outputs an ARIMA model for each mode. Using each output ARIMA model, the sub-forecasting model creation module 120 can predict time series information subsequent to time T for the mode corresponding to each ARIMA model.

[0047] As described above, the parameter determination unit 114 determines parameters that minimize the average entropy of the time series information for each mode obtained by executing the variational mode decomposition. Next, the variational mode decomposition unit 112 performs variational mode decomposition on the time series information within the range defined by the parameters determined by the parameter determination unit 114. Next, the sub-prediction model creation module 120, which corresponds to the time series information prediction unit, predicts the time series information for each mode that follows each of the time series information for the multiple modes, using the time series information for the multiple modes obtained by executing the variational mode decomposition. Next, the time series information addition unit 131 adds the predicted time series information for each mode. Then, the time series information output unit 132 outputs the time series information obtained by the addition of the time series information for each mode by the time series information addition unit 131.

[0048] Non-Patent Document 1 shows that prediction accuracy can be improved by setting parameters that minimize the average entropy of time-series information. Therefore, by adopting the above configuration, it is possible to minimize the average entropy of time-series information for each mode generated by variational mode decomposition, thereby improving the prediction accuracy of time-series information.

[0049] In the above-described embodiment, variational mode decomposition is performed using a specified decomposition number K. However, in other embodiments, the mode number determination unit may determine the decomposition number K that minimizes the average entropy of the time-series information of each mode when the target time-series information is decomposed by the variational mode decomposition unit 112. As a method for deriving the optimal decomposition number K that minimizes the average entropy, metaheuristic algorithms such as particle swarm optimization (PSO) and BWO can be adopted. In other words, optimization can be performed using the formula shown in Equation 4 as the objective function, and the parameter indicating the endpoint EP and the decomposition number K as control variables.

number

[0050] In the above examples, the program includes instructions (or software code) that, when loaded into a computer, cause the computer to perform one or more functions described in the embodiments. The program may be stored on a non-transitory computer-readable medium or a tangible storage medium. By way of example and not limitation, computer-readable medium or tangible storage medium includes random-access memory (RAM), read-only memory (ROM), flash memory, solid-state drive (SSD) or other memory technology, CD-ROM, digital versatile disk (DVD), Blu-ray® disk or other optical disk storage, magnetic cassette, magnetic tape, magnetic disk storage or other magnetic storage device. The program may also be transmitted on a transitory computer-readable medium or communication medium. By way of example and not limitation, transitory computer-readable medium or communication medium includes electrical, optical, acoustic, or other forms of propagated signals.

[0051] The present disclosure is not limited to the above-described embodiment, and can be modified as appropriate without departing from the spirit of the present disclosure. For example, in the above-described embodiment, a single prediction device 10 executes the above-described program, but in other embodiments, multiple individual prediction devices may execute the above-described program in a distributed manner.

[0052] In another embodiment, the parameter determination unit 114 may use a model trained by machine learning to determine parameters that minimize the average entropy of time-series information for each mode obtained by variational mode decomposition. This model can be trained using training data that takes the target time-series information as input information and outputs parameters that minimize the average entropy of time-series information for each mode obtained by variational mode decomposition of the target time-series information. When the target time-series information is input, this model outputs parameters that indicate the endpoint that minimizes the average entropy of the time-series information for each mode. [Explanation of symbols]

[0053] 10: Prediction device, 100: Processor, 110: Variational mode decomposition module, 111: Acquisition unit, 112: Variational mode decomposition unit, 113: Entropy calculation unit, 114: Parameter determination unit, 115: Extraction unit, 120: Sub-prediction model creation module / time series information prediction unit, 130: Sub-prediction model fusion module, 131: Time series information addition unit, 132: Time series information output unit, 140: Communication interface, 150: Storage device

Claims

1. A prediction system for predicting time series information, a parameter determination unit that determines a parameter that defines the end of a range of time series information on which variational mode decomposition is performed; a variational mode decomposition unit that performs variational mode decomposition on the time-series information of the target within a range defined by the parameters determined by the parameter determination unit; a time series information prediction unit that predicts time series information for each mode subsequent to each of the time series information for the plurality of modes, using the time series information for the plurality of modes obtained by executing the variational mode decomposition; a time-series information adding unit that adds the time-series information for each mode predicted by the time-series information predicting unit; a time-series information output unit that outputs time-series information obtained by adding the time-series information for each mode by the time-series information adding unit, the parameter determination unit determines parameters that minimize an average entropy of time-series information of each mode obtained by performing variational mode decomposition on the time-series information of the target; Prediction system.

2. a mode number determination unit that determines the number of modes that minimizes the average entropy; The prediction system according to claim 1 , wherein the variational mode decomposition unit decomposes the time series information of the target into time series information of the number of modes determined by the number-of-modes determination unit.

3. A prediction method in which a computer predicts time-series information, the computer determining parameters that minimize an average entropy of time series information of each mode obtained by performing variational mode decomposition on the time series information of the target, the parameters defining the end of a range of the time series information of the target on which variational mode decomposition is performed; performing variational mode decomposition on the time series information of the target within a range defined by the determined parameters; predicting time series information for each mode subsequent to each of the time series information for the plurality of modes using time series information for the plurality of modes obtained by executing the variational mode decomposition; Adding the predicted time series information for each mode; outputting the time series information obtained by adding the time series information for each mode; Forecasting methods.

4. The computer determining the number of modes that minimizes the mean entropy; The prediction method according to claim 3 , wherein the time series information of the target is decomposed into time series information of a determined number of modes.

5. A prediction program for predicting time-series information executed by a computer, the program comprising: determining parameters that minimize the average entropy of time series information of each mode obtained by performing variational mode decomposition on the time series information of the target, the parameters defining the end of a range of the time series information of the target on which variational mode decomposition is performed; performing variational mode decomposition on the time series information of the target within a range defined by the determined parameters; predicting time series information for each mode subsequent to each of the time series information for the plurality of modes using time series information for the plurality of modes obtained by executing the variational mode decomposition; adding the predicted time series information for each mode; and outputting the time-series information obtained by adding the time-series information for each mode. Prediction program.

Images