Time sequence prediction method and system based on visual graph algorithm
By converting the time series data into a view and selecting the node sequence using the graph kernel and the maximum correlation minimum redundancy method, the problem of insufficient time series prediction accuracy and generalization capabilities in the prior art is solved, and more efficient and accurate prediction effects are achieved.
Patent Information
- Application Number
- CN202510083959.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-20
- Publication Date
- 2025-05-23
AI Technical Summary
The existing time series prediction methods have shortcomings in improving prediction accuracy and generalization capabilities, especially when processing large-scale time series data, information redundancy and computational complexity are prone to problems.
The timing prediction method based on the view algorithm is adopted to convert the timing data into a view, calculate the similarity between nodes through the graph core idea, and select the best node sequence for prediction using the maximum correlation and minimum redundancy method.
It improves the accuracy and efficiency of time series prediction, reduces information redundancy, and performs better than traditional ARIMA models under various error indicators.
Smart Images
Figure CN120030359A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of time series prediction, and in particular to a time series prediction method and system based on a visibility graph algorithm. Background Art
[0002] Time series (TS), as its name suggests, is an ordered set of observation data arranged in chronological order. Time series data is closely related to our lives. Finance, medicine, agriculture, climate and other fields will generate a large amount of time series data, and these data also contain various historical information when they were observed. Using these data to mine the hidden information and predict the future development and change trend of the time series is widely used in finance, transportation, environment and other fields.
[0003] From the time when the idea of time prediction was proposed to the present, the commonly used methods can be roughly divided into two categories. One is the traditional regression model based on statistics. This method originated from the auto-regressive model (AR) and the moving average model (MA) proposed by Yule, and many improved models such as ARMA and ARIMA were developed on this basis. These models have low computational cost and complexity, but with the increasing demand for prediction accuracy and the increasing scale of time series data, traditional models often cannot meet the requirements for accuracy, and the models themselves do not have good generalization capabilities. With the development of computer science, machine learning methods such as LSTM and CNN have also been widely used in time series prediction tasks. Unlike traditional regression models, models with machine learning as the core can use data sets to train models, explore the characteristics and change patterns within the data, and thus improve prediction accuracy. However, these models often require a large amount of accurate data for training, and require higher hardware conditions and more computing time.
[0004] The visualization graph algorithm proposed by Lacasa et al. can quickly transform time series data into complex networks. Its low complexity and the ability to retain the characteristics of the original data while extracting features make it one of the effective methods for analyzing time series data. On this basis, researchers have proposed a variety of methods for calculating node similarity and future value prediction, which can be mainly divided into two directions. One direction is to use one or more nodes with the maximum similarity to the last observed node under a certain similarity index for prediction. Although this can obtain more accurate prediction results to a certain extent, it often ignores a lot of useful information in the time series, making it difficult to significantly improve the prediction accuracy. In another direction, all observed nodes are used to predict future moments by assigning different weights to the nodes. This method can make full use of all the information in the time series, but due to the characteristics of the time series that "the observation value at each moment contains the data information of the previous moment", this will cause a lot of information redundancy, which will affect the prediction accuracy when the time series scale increases. Summary of the invention
[0005] In view of this, the present application provides a time series prediction method and system based on a visibility graph algorithm to achieve better prediction accuracy.
[0006] To achieve the above objectives, the technical solutions adopted in this application are as follows: A time series prediction method based on a visibility graph algorithm, comprising: Step 1: Convert the time series data into a visual graph through a visual graph algorithm; Step 2: Define the visible range of each node on the visible graph; Step 3: Extract the characteristic subgraph of each node in the graph according to the defined visual range; Step 4: Calculate the similarity between different feature subgraphs through the graph kernel idea and use it as the similarity distribution between nodes; Step 5: According to the similarity distribution between nodes, the maximum relevance minimum redundancy method is used to select the best node sequence for prediction; Step 6: Predict the future values of the time series data based on the selected best node sequence for prediction.
[0007] Furthermore, the method further comprises: Step 7: Evaluate the prediction results.
[0008] Furthermore, in step 1, the time series data is converted into a visible graph by using a visible graph algorithm as follows: Step 1.1: Convert the time series data into a time series diagram, where each observation value is represented graphically on the ordinate axis according to the size of the observation value, and arranged in the order of observation time on the abscissa; Step 1.2: Determine whether the tops of different graphics are connected according to visual criteria, wherein the visual criteria are: For two nodes and and any nodes between them , if the condition of formula (1) is satisfied, then and will be visible and connected in the network; , in (1) Step 1.3: Connect the tops of the images that can be connected, so that the timing diagram is converted into a visual diagram of the network structure ,in, Represents a collection of nodes, Represents a collection of edges.
[0009] Furthermore, the specific method for defining the visible range of each node on the visible graph in step 2 is: like Is a node The farthest node that can be connected to, the visible range for ,in .
[0010] Furthermore, the specific method for calculating the similarity between different feature subgraphs by using the graph core idea in step 4 is: Step 4.1: Using Node The characteristic subgraph of and nodes The characteristic subgraph of , construct the direct product graph ; is defined as: (4) (5) Therefore, the two figures The calculation formula of the random walk kernel is: (6) in, express The adjacency matrix of The matrix is Elements represent The Nodes and The length between nodes is The number of common walks, Indicates the length is The weight of the random walk of Step 4.2: According to equation (6), calculate the node and nodes The similarity between Step 4.3: Repeat steps 4.1 and 4.2 until the similarities between all nodes are calculated.
[0011] Furthermore, in step 5, the specific method of selecting the best node sequence for prediction using the maximum relevance minimum redundancy method according to the similarity distribution between nodes is as follows: Step 5.1: For the Node , remember it and the last observation node The similarity between them is S; the similarity between them and other nodes is R; Step 5.2: Node SD score ; Step 5.3: Repeat step 5.2 to get the SD scores of the first n-1 nodes ; Step 5.4: Set a lower limit for the score. Based on the SD score, nodes with scores higher than the lower limit are used in the prediction phase, while nodes with scores lower than the lower limit are not used in the prediction phase.
[0012] Furthermore, in step 6, the future value of the time series data is predicted according to the selected optimal node sequence for prediction as follows: Step 6.1: For each node selected for prediction, the prediction value is calculated by the following formula: (12) Step 6.2: Contribution weight of each node in the prediction phase By normalization get; (13) Step 6.3: Finally, the predicted value of the n+1th node is: (14).
[0013] Furthermore, the prediction results in step 7 are evaluated specifically according to the mean absolute error (MAE), mean absolute percentage error (MAPE), symmetric mean absolute percentage error (SMAPE) and root mean square error (RMSE), and the calculation formulas for each error are: (15) (16) (17) (18).
[0014] The present application also provides a time series prediction system based on a visibility graph algorithm, the system can implement a time series prediction method based on a visibility graph algorithm in the present application, the system comprising: A visual graph conversion module is used to convert time series data into a visual graph through a visual graph algorithm; Definition module, used to define the visible range of each node on the visible graph; The extraction module is used to extract the characteristic subgraph of each node in the graph according to the defined visual range; The calculation module is used to calculate the similarity between different feature subgraphs through the graph kernel idea and use it as the similarity distribution between nodes; A selection module is used to select the best node sequence for prediction based on the similarity distribution between nodes using the maximum relevance minimum redundancy method; The prediction module is used to predict the future value of the time series data based on the selected best node sequence for prediction.
[0015] Furthermore, the system also includes: Evaluation module, used to evaluate the prediction results.
[0016] Compared with the prior art, the beneficial effects of this application are: 0. Use the Graph Kernel algorithm to calculate the similarity between different feature subgraphs. It is directly oriented to graph structure data, which not only retains the advantages of efficient kernel function calculation, but also contains the structural information of graph data in Hilbert high-dimensional space. At the same time, different kernel functions can be defined for different graph structures for calculation; 1. The idea of maximum relevance and minimum redundancy (mRMR) is introduced. According to the similarity distribution between nodes, the best node sequence for prediction is selected to solve the problem of information redundancy that may occur in the prediction process; 2. The prediction error of the algorithm in this application is lower than that of the ARIMA model under the four error indicators, and has good prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.
[0018] Figure 1 A flow chart of a time series prediction method based on a visibility graph algorithm for this application; Figure 2 This is a structural block diagram of a time series prediction system based on a visibility graph algorithm for this application; Figure 3 It is a schematic diagram of the function of the visual diagram; Figure 4 Schematic diagram of mutual information derivation; Figure 5 (a) is a visual graph in a specific implementation of the present application, and (b) is a characteristic subgraph of a node in a specific implementation of the present application; Figure 6 This is a schematic diagram of multi-step prediction in a specific implementation manner of this application; Figure 7 The SD score distribution of each node in the CCI time series in the specific implementation mode of this application; Figure 8 This is a diagram of the prediction results in the specific implementation manner of this application. DETAILED DESCRIPTION
[0019] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments.
[0020] like Figure 1 As shown, a time series prediction method based on a visibility graph algorithm includes: Step 1: Convert the time series data into a visual graph through a visual graph algorithm; As a new method to study time series, visualization graphs can link time series with complex networks, such as Figure 3 shown.
[0021] As a further implementation, in step 1, converting the time series data into a visible graph by a visible graph algorithm is specifically as follows: Step 1.1: Convert the time series data into a time series diagram, where each observation value is represented graphically on the ordinate axis according to the size of the observation value, and arranged in the order of observation time on the abscissa; Specifically, the graph may be a bar graph or a column graph, or may be designed as a graph of other shapes according to actual conditions, and this application does not limit this.
[0022] Step 1.2: Determine whether the tops of different graphics are connected according to visual criteria, wherein the visual criteria are: For two nodes and and any nodes between them , if the condition of formula (1) is satisfied, then and will be visible and connected in the network; , in (1) Given a time series ,in represents the observation time, represents the observed value at that moment, is the length of the sequence, and each data point in the time series is represented as a node in the network. It should be noted that two adjacent nodes in the visibility graph must be visible because there are no other nodes between them. Through the visibility graph algorithm, the time series S will be transformed into a complex network .
[0023] Step 1.3: Connect the tops of the images that can be connected, so that the timing diagram is converted into a visual diagram of the network structure ,in, Represents a collection of nodes, Represents a collection of edges.
[0024] By analyzing the geometric characteristics of the network, the information contained in the time series can be mined.
[0025] For example, given a time series The converted visual representation is as follows Figure 5 (a) shown.
[0026] Step 2: Define the visible range of each node on the visible graph; Specifically, if Is a node The farthest node that can be connected to, the visible range for ,in .
[0027] Step 3: Extract the characteristic subgraph of each node in the graph according to the defined visual range; Specifically, if The visible range is , then its characteristic subgraph The midpoint range is , and includes a visual representation All edges in the range of the above nodes.
[0028] The characteristic subgraph of node 8 is as follows Figure 5 (b) as shown.
[0029] Step 4: Calculate the similarity between different feature subgraphs through the graph kernel idea and use it as the similarity distribution between nodes; The graph kernel method extends the kernel method in machine learning to graph structure data. It is a method for calculating the similarity between graphs. There are two main types of kernel methods in graph structure data: one is the graph embedding algorithm. The core idea of this algorithm is to embed the graph structure into the vector space to obtain the vector representation of the graph, and then use the kernel function based on the vector space for further processing. However, this type of method loses a lot of structural information in the process of reducing the dimension of structured data to the vector space. The other is the graph kernel algorithm, which is directly oriented to graph structure data. It not only retains the advantages of efficient kernel function calculation, but also contains the structural information of graph data in the Hilbert high-dimensional space. At the same time, for different graph structures, different kernel functions can be defined for calculation. At present, most graph kernels are developed and evolved on the basis of R-convolution kernels. The core idea of this method is to map the graph into a certain Hilbert space. According to the framework of R-convolution, the kernel value of the graph can be calculated by the following formula:
[0030] (2) in, and is a graph structure of two inputs, represents an arbitrary positive definite kernel defined on a node. The random walk kernel is one of the most studied graph kernels, which counts the number of walks in common between two graphs. The random walk kernel of as follows:
[0031] (3) in is the coefficient, Representation Node The neighbor nodes of is the step size of the random walk in the two graphs being compared.
[0032] In the figure and Performing a random walk on both graphs is equivalent to performing a random walk on their direct product graph. is defined as:
[0033] (4) (5) Therefore, the two figures The calculation formula of the random walk kernel is: (6) in, express The adjacency matrix of The matrix is The elements represent The Nodes and The length between nodes is The number of common walks, Indicates the length The weight of the random walk.
[0034] As a further implementation, the specific method for calculating the similarity between different feature subgraphs by using the graph core idea in step 4 is: Step 4.1: Using Node The characteristic subgraph of and nodes The characteristic subgraph of , construct the direct product graph ; Constructing a direct product graph to facilitate random walks on the direct product graph is equivalent to executing random walks in two graphs at the same time, which also facilitates the calculation of subsequent graph kernels.
[0035] Step 4.2: According to equation (6), calculate the node and nodes The similarity between K(G1, G2) calculated by formula (6) is the similarity score between graphs G1 and G2. The larger K(G1, G2) is, the more similar the two graphs are.
[0036] Step 4.3: Repeat steps 4.1 and 4.2 until the similarities between all nodes are calculated.
[0037] Step 5: According to the similarity distribution between nodes, the maximum relevance minimum redundancy method is used to select the best node sequence for prediction; Max-relevance and min-redundancy (mRMR) is a feature selection technique based on mutual information. This method calculates the similarity between features and target variables and between features to select a set of features with the highest correlation with the target variable and the lowest correlation with the remaining features from the initial feature set.
[0038] The similarity between features is obtained by calculating the mutual information between them. The expression of mutual information is:
[0039] (7) in, Representation characteristics a and Features b The joint probability density of p(a) and p(b) Respectively represent characteristics a and Features b The probability density of . Figure 4 It was further explained.
[0040] according to Figure 4 It can be seen that formula (7) can be derived through information entropy.
[0041] (8) set up is a feature vector consisting of m samples; the target variable ,in represents the number of eigenvectors, Indicates the number of samples. S Represents the selected feature subset.
[0042] Definition 1 Maximum correlation: (9) Definition 2 Minimum redundancy: (10) Definition 3 mRMR selection criteria: The goal of the mRMR algorithm is to select a feature subset S that maximizes the maximum correlation between S and the target category c minus the minimum redundancy of S.
[0043] (11) As a further implementation, in step 5, the specific method of selecting the best node sequence for prediction using the maximum relevance minimum redundancy method according to the similarity distribution between nodes is as follows: Step 5.1: For the Node , remember it and the last observation node The similarity between them is S; the similarity between them and other nodes is R; Step 5.2: Node SR points ; Step 5.3: Repeat step 5.2 to get the SR scores of the first n-1 nodes ; Step 5.4: Set a lower limit for the score. According to the SR score, nodes with scores higher than the lower limit are used in the prediction phase, while nodes with scores lower than the lower limit are not used in the prediction phase.
[0044] Specifically, the score lower limit is generally 0, and may also have different values according to actual conditions, which is not limited in this application. Nodes with scores higher than the lower limit will be used in the prediction phase; nodes with scores lower than the lower limit are considered not to bring additional useful information to the prediction, but may reduce the model performance.
[0045] Step 6: Predict the future values of the time series data based on the selected best node sequence for prediction.
[0046] As a further implementation, in step 6, predicting the future value of the time series data according to the selected optimal node sequence for prediction is specifically as follows: Step 6.1: For each node selected for prediction, the prediction value is calculated by the following formula: (12) Step 6.2: Contribution weight of each node in the prediction phase By normalization get; (13) Step 6.3: Finally, the predicted value of the n+1th node is: (14).
[0047] As a further embodiment, the method further comprises: Step 7: Evaluate the prediction results.
[0048] As a further implementation, the prediction results are evaluated in step 7 specifically according to the mean absolute error (MAE), mean absolute percentage error (MAPE), symmetric mean absolute percentage error (SMAPE) and root mean square error (RMSE), and the calculation formulas for each error are: (15) (16) (17) (18).
[0049] verify: In this section, the construction cost index CCI time series data is used to verify the effectiveness of the proposed model. The CCI data is divided into a training set (containing the first 240 data points) and a test set (containing the remaining 55 data points). The training set is used to construct the visualization graph, while the test set is used as the data to be predicted.
[0050] Multi-step prediction is another important problem in time series prediction tasks, such as predicting the future price changes of commodities, predicting the precipitation within a period of time, etc. Its implementation process is as follows Figure 6 shown.
[0051] Figure 7 Shows the nodes in the training set S-R From the scoring situation, we can see that a small number of nodes have scores below the threshold of 0. These nodes will be ignored in the prediction phase because their information redundancy is higher, which affects the performance of the model.
[0052] In order to further evaluate the effectiveness of the model in the prediction task, several error indicators are used to evaluate the prediction results. They are: mean absolute error (MAE), mean absolute percentage error (MAPE), symmetric mean absolute percentage error (SMAPE) and root mean square error (RMSE). The error calculation methods are detailed in formulas (13)-(16).
[0053] According to the prediction step, the best node is used for prediction. The predicted values of the 55 observation points in the test set are as follows Figure 8 In Table 1, four errors between the prediction results and the true values are calculated; at the same time, they are compared with the prediction results of the traditional classic prediction model ARIMA. It can be seen that the system model and method of the present application are lower than the prediction error of the ARIMA model under the four error indicators, and have good prediction accuracy. The bold in Table 1 represents emphasis.
[0054] Table 1 Comparison of CCI time series prediction errors
[0055] like Figure 2 As shown, the present application also provides a time series prediction system based on a visible graph algorithm, which can implement a time series prediction method based on a visible graph algorithm in the present application, and the system includes: A visual graph conversion module 210, used to convert the time series data into a visual graph through a visual graph algorithm; A definition module 220, used to define the visible range of each node on the visible graph; An extraction module 230 is used to extract a characteristic subgraph of each node in the graph according to a defined visual range; A calculation module 240 is used to calculate the similarity between different feature subgraphs through a graph kernel concept and use it as the similarity distribution between nodes; A selection module 250, for selecting an optimal node sequence for prediction according to the similarity distribution between nodes; The prediction module 260 is used to predict the future value of the time series data according to the selected best node sequence for prediction.
[0056] As a further embodiment, the system further includes: Evaluation module, used to evaluate the prediction results.
[0057] As for the device embodiment, since it basically corresponds to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The device embodiment described above is only illustrative, and the units described as separate components may or may not be physically separated.
[0058] The components may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units, such as distributed on a server and a client. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art may understand and implement it without creative work.
[0059] The present application uses a graph kernel algorithm to calculate the similarity between different feature subgraphs. It is directly oriented to graph structure data, retaining the advantage of efficient kernel function calculation and including the structured information of graph data in Hilbert high-dimensional space. At the same time, different kernel functions can be defined for calculation for different graph structures. Moreover, the idea of maximum relevance minimum redundancy (mRMR) is introduced to select the best node sequence for prediction according to the similarity distribution between nodes, thus solving the problem of information redundancy that may occur in the prediction process. In addition, the algorithm of the present application is lower than the prediction error of the ARIMA model under the four error indicators and has good prediction accuracy.
[0060] The above are only specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.
Claims
1. A time series prediction method based on a visibility graph algorithm, characterized in that: include: Step 1: Convert the time series data into a visual graph through a visual graph algorithm; Step 2: Define the visible range of each node on the visible graph; Step 3: Extract the characteristic subgraph of each node in the graph according to the defined visual range; Step 4: Calculate the similarity between different feature subgraphs through the graph core idea and use it as the similarity distribution between nodes; Step 5: According to the similarity distribution between nodes, the maximum relevance minimum redundancy method is used to select the best node sequence for prediction; Step 6: Predict the future values of the time series data based on the selected best node sequence for prediction.
2. A time series prediction method based on a visibility graph algorithm as claimed in claim 1, characterized in that: The method further comprises: Step 7: Evaluate the prediction results.
3. A time series prediction method based on a visibility graph algorithm as claimed in claim 2, characterized in that: In step 1, the time series data is converted into a visible graph by using a visible graph algorithm as follows: Step 1.1: Convert the time series data into a time series diagram, where each observation value is represented graphically on the ordinate axis according to the size of the observation value, and arranged in the order of observation time on the abscissa; Step 1.2: Determine whether the tops of different graphics are connected according to visual criteria, wherein the visual criteria are: For two nodes and and any nodes between them , if the condition of formula (1) is satisfied, then and will be visible and connected in the network; ,in (1) Step 1.3: Connect the tops of the images that can be connected, so that the timing diagram is converted into a visual diagram of the network structure ,in, Represents a collection of nodes, Represents a collection of edges.
4. A time series prediction method based on a visibility graph algorithm as claimed in claim 3, characterized in that: The specific method of defining the visible range of each node on the visible graph in step 2 is: like Is a node The farthest node that can be connected to, the visible range for ,in .
5. A time series prediction method based on a visibility graph algorithm as claimed in claim 4, characterized in that: The specific method for calculating the similarity between different feature subgraphs by using the graph core idea in step 4 is: Step 4.1: Using Node The characteristic subgraph of and nodes The characteristic subgraph of , construct the direct product graph ; is defined as: (4) (5) Therefore, the two figures The calculation formula of the random walk kernel is: (6) in, express The adjacency matrix of The matrix is The elements represent The Nodes and The length between nodes is The number of common walks, Indicates the length The weight of the random walk of Step 4.2: According to equation (6), calculate the node and nodes The similarity between Step 4.3: Repeat steps 4.1 and 4.2 until the similarities between all nodes are calculated.
6. A time series prediction method based on a visibility graph algorithm as claimed in claim 5, characterized in that: In step 5, the method of selecting the best node sequence for prediction using the maximum relevance minimum redundancy method according to the similarity distribution between nodes is as follows: Step 5.1: For the Node , remember it and the last observation node The similarity between them is S; the similarity between them and other nodes is R; Step 5.2: Node SD score ; Step 5.3: Repeat step 5.2 to get the SD scores of the first n-1 nodes ; Step 5.4: Set a lower limit for the score. Based on the SD score, nodes with scores higher than the lower limit are used in the prediction phase, while nodes with scores lower than the lower limit are not used in the prediction phase.
7. A time series prediction method based on visibility graph algorithm as claimed in claim 6, characterized in that: In step 6, the future value of the time series data is predicted according to the selected optimal node sequence for prediction as follows: Step 6.1: For each node selected for prediction, the prediction value is calculated by the following formula: (12) Step 6.2: Contribution weight of each node in the prediction phase By normalization get; (13) Step 6.3: Finally, the predicted value of the n+1th node is: (14)。 8. A time series prediction method based on a visibility graph algorithm as claimed in claim 7, characterized in that: The prediction results in step 7 are evaluated specifically according to MAE, MAPE, SMAPE and RMSE, and the error calculation formulas are: (15) (16) (17) (18)。 9. A time series prediction system based on a visibility graph algorithm, characterized in that: The system can implement the method according to any one of claims 1 to 8, and the system includes: A visual graph conversion module is used to convert time series data into a visual graph through a visual graph algorithm; Definition module, used to define the visible range of each node on the visible graph; The extraction module is used to extract the characteristic subgraph of each node in the graph according to the defined visual range; The calculation module is used to calculate the similarity between different feature subgraphs through the graph core idea and use it as the similarity distribution between nodes; A selection module is used to select the best node sequence for prediction based on the similarity distribution between nodes using the maximum relevance minimum redundancy method; The prediction module is used to predict the future value of the time series data based on the selected best node sequence for prediction.
10. A time series prediction system based on a visibility graph algorithm as claimed in claim 9, characterized in that: The system further comprises: Evaluation module, used to evaluate the prediction results.