Time sequence causal-driven meteorological disaster prediction method

Abstract

The invention discloses a sequential causal-driven meteorological disaster prediction method, and belongs to the field of causal relationship discovery and meteorological disaster prediction. According to the specific scheme, the method comprises the following steps: preprocessing time series data of meteorological disasters to obtain an optimal lag window and an optimal data length; dividing the meteorological data into a plurality of meteorological data segments with the same length according to the optimal lagging window and the optimal data length; respectively applying a time sequence causal relationship discovery method to the divided meteorological data segments to obtain window causal structure diagrams; and fusing the causal structure diagrams in the plurality of data fragments in an ensemble learning mode to obtain a globally consistent time sequence causal structure diagram. In meteorological disaster prediction, the time sequence causal relationship between disaster risk levels and key meteorological factors in meteorological data can be mined, the accuracy of meteorological disaster prediction by a meteorological institution is improved, and effective prevention of meteorological disasters is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of time series causal discovery and meteorological disaster prediction, and specifically relates to a meteorological disaster prediction method driven by time series causal relationship. Background Art

[0002] Meteorological disaster prediction refers to the use of multidisciplinary theories and technical means such as meteorology and computer science to make advance estimates and judgments on abnormal weather events that are about to occur or may occur. Its purpose is to provide timely and accurate meteorological information to multiple fields such as government, agriculture, transportation, and water conservancy, so as to take corresponding prevention and response measures, thereby minimizing the threat of meteorological disasters to the economy, society, and the safety of life and property of the people. However, the current mainstream method of meteorological disaster prediction focuses on the mining and analysis of correlations between variables. Although this method can reveal the correlation between weather phenomena, it fails to fully capture the deep and potential causal relationships between data. This lack of causal insight may significantly weaken the accuracy and stability of the prediction model when encountering extreme or atypical weather conditions, thereby increasing the difficulty and uncertainty of responding to meteorological disasters.

[0003] Time series causal discovery is an analytical method for finding and inferring causal relationships in time series data. It aims to reveal the dynamic causal relationships between variables in time series data, thereby helping people better understand the laws behind the data, predict future trends, and make more effective decisions. Existing time series causal discovery methods may need to rely on some initial parameters or assumptions when determining the optimal lag window, which may cause the algorithm to fail to accurately identify causal relationships or produce misleading results. In addition, although existing methods are able to determine the optimal lag window, this process may involve complex calculations, which usually require multiple experiments or optimization algorithms to achieve, which may lead to a significant increase in the amount of calculation. Therefore, traditional time series causal discovery methods face the problems of sensitivity to initial parameters and high computational complexity when predicting meteorological disasters. Summary of the invention

[0004] In view of the problem that the current meteorological agencies have poor accuracy in predicting meteorological disasters, the present invention provides a time series causal driven meteorological disaster prediction method.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A time series causal driven meteorological disaster prediction method comprises the following steps:

[0007] Step 1: preprocess the time series data of meteorological disasters and calculate the optimal lag window and slice length;

[0008] Step 2: Divide the time series data into multiple time segments of equal length according to the slice length;

[0009] Step 3: Perform time series causal discovery independently for each time segment to obtain a window causal structure diagram;

[0010] Step 4: Integrate the window causal structure graphs obtained from multiple time segments through an ensemble learning method to obtain a globally consistent temporal causal structure graph.

[0011] Furthermore, the step 1, preprocessing the time series data of meteorological disasters, and calculating the optimal lag window and slice length is specifically: preprocessing the time series data of meteorological disasters, initializing the lag window, increasing the size of the lag window and analyzing the changes in the number of causal relationships caused thereby, and determining the lag window when the causal relationship no longer increases with the increase of the lag window as the optimal lag window; initializing the slice length, reducing the slice length and analyzing the changes in the number of causal relationships caused thereby, and determining the slice length when the causal relationship no longer decreases with the decrease of the slice length as the optimal slice length.

[0012] Further, the calculation of the optimal hysteresis window and slice length in step 1 specifically includes the following steps:

[0013] Step 1.1.1, input the time series data X of meteorological disasters T R , including the number of variables R and the number of time T; the initial lag window τ; the initial time series window causal graph G; the causal relationship accumulation threshold α; the causal relationship increment β;

[0014] Step 1.1.2: Use the PCMCI algorithm to process the meteorological disaster time series data to obtain the current time series window causal graph G τ ;

[0015] Step 1.1.3, the current time series window causal graph G τ Compare with the initial time series window causal graph G to check whether there is a new causal relationship added. If a new causal relationship is found, reset the causal relationship increment β to 0; otherwise, increase the value of the causal relationship increment β;

[0016] Step 1.1.4, determine whether the causal relationship increment β is less than the set causal relationship accumulation threshold α. If β<α, the current hysteresis window τ is the optimal hysteresis window; otherwise, the current G τ Update to G, τ = τ + 1, and return to step 1.1.2 for the next round of iteration;

[0017] Step 1.1.5, by iterating steps 1.1.2 to 1.1.4, finally determine the optimal lag window τ x .

[0018] Step 1.2.1, Input: Time series data of meteorological disasters Including the number of variables R and the time number T; the starting time point t; the time length Optimal hysteresis window τ x ; Initial time window causal graph G; ideal causal relationship reduction ratio γ, actual causal relationship reduction ratio σ;

[0019] Step 1.2.2, use the PCMCI algorithm to process the meteorological disaster time series data and the lag window τ. For detailed implementation, see step 4 to obtain the current time series window causal diagram

[0020] Step 1.2.3, the current time series window cause and effect diagram Compare with the initial time series window causal graph G to check whether there is a reduction in causal relationships; if a reduction in causal relationships is found, the actual causal relationship is reduced by a certain ratio. Where φ(DAG) is used to represent the number of edges in the statistical graph DAG; if there is none, the actual causal relationship reduction ratio σ=0;

[0021] Step 1.2.4, determine whether the actual causal relationship reduction ratio σ is greater than the set ideal causal relationship reduction ratio γ. If σ>γ, the slice length at this time is is the optimal slice length; otherwise, keep G unchanged. And return to step 1.2.2 for the next round of iteration;

[0022] Step 1.2.5, by iterating steps 2.2 to 2.4, finally determine the optimal slice length

[0023] Furthermore, in step 2, the time series data is divided into a plurality of time segments of equal length according to the slice length;

[0024] According to step 2, the optimal slice length is obtained Time series data Divide into multiple segments of equal length; the length of each segment is the optimal slice length, until all data are segmented or the remaining length is not enough to form a complete time series segment; obtain time segments, the optimal lag window for each time segment is τ x ;

[0025] Furthermore, in step 3, time series causal discovery is performed independently for each time segment to obtain a window causal structure diagram;

[0026] Step 3.1, Input: Time series data fragment Optimal hysteresis window τx .

[0027] Step 3.2, PC1 stage: Initialize various variables The parent node of Then the first iteration (p = 0) is performed to test unconditional independence. and The hypothesis of independence cannot be established at the significance level α PC If rejected, from The parent node of In the next iteration (p→p+1), first sort the parent nodes according to the test statistic, and then test if S is known. and Whether it is independent, where S is Zhongyu The set of p nodes with the strongest relationship; after each iteration, the parent nodes independent of the test node are removed from Remove until no further tests are possible and the algorithm converges;

[0028] Step 3.3, MCI stage: Based on the edges retained in step 3.2 Consider at the same time and Proceed to step 3.2, which accounts for autocorrelation while identifying indirect and common cause links, thereby controlling the false positive rate to within the expected level;

[0029] Step 3.4, after the above two stages, what is finally obtained is the time series window causal graph.

[0030] Furthermore, in step 4, the window causal structure graphs obtained from multiple time segments are integrated through an integrated learning method to obtain a globally consistent temporal causal structure graph.

[0031] Step 4.1, align the k time series window causal graphs obtained in step 3 in time and variables to ensure that they maintain a corresponding relationship between time points and time series variables;

[0032] Step 4.2, counting the number of causal relationships at all positions in the causal graph of each time series window to form a causal relationship quantity matrix;

[0033] Step 4.3, process the causal relationship quantity matrix;

[0034] Step 4.3.1, first calculate the mean value of the causal relationship quantity matrix and store the result in the variable mean;

[0035] Step 4.3.2, then calculate the absolute difference between each element and the mean, and store these differences in an array diff;

[0036] Step 4.3.3, finally find the indexes of the elements in the difference array whose absolute difference is greater than the preset threshold, set the values ​​of these indexes in the corresponding positions of the causal relationship matrix to 1, and set the values ​​of the remaining positions to 0;

[0037] Step 4.4, performing mean processing on the k test statistic value matrices obtained in step 4 to obtain a mean test statistic value matrix;

[0038] Step 4.5, compress the causal relationship matrix in the time dimension. If there is a value of 1 between two time series variables in the entire time dimension, the corresponding position in the final causal relationship matrix is ​​set to 1; otherwise, it is set to 0;

[0039] Step 4.6, based on the mean test statistic matrix, determine the basic causal diagram rules for the causal relationship matrix; when a bidirectional edge is found, give priority to the side with a larger test statistic value, discard the side with a smaller test statistic value, and ensure that there is no loop in the graph;

[0040] Step 4.7, through the above operations, a final globally consistent temporal causal structure diagram is formed.

[0041] Compared with the prior art, the present invention has the following advantages:

[0042] The method of the present invention not only significantly improves the ability of meteorological forecasting agencies to efficiently predict the occurrence of meteorological disasters, but also greatly promotes the accuracy and operability of the prediction model. It alleviates the problems of sensitivity to initial parameters and high computational complexity faced by traditional time series causal discovery methods when processing meteorological disaster data. By accurately capturing the variables and their interrelationships in meteorological disaster data, the method can automatically predict meteorological disasters and provide timely and accurate meteorological information so that corresponding prevention and response measures can be taken, thereby minimizing the threat of meteorological disasters to the economy, society, and the safety of life and property of the people. At the same time, it helps to improve the accuracy of meteorological agencies' predictions of meteorological disasters and achieve effective prevention of meteorological disasters. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is a flow chart of the method of the present invention;

[0044] Figure 2 It is a schematic diagram of the specific technical route of the present invention. DETAILED DESCRIPTION

[0045] In order to gain a deeper understanding of the present invention, we will provide a comprehensive and detailed description of the present invention. However, the present invention has multiple implementations and is not limited to the specific examples listed herein. The presentation of these examples is intended to deepen the comprehensive understanding of the disclosure of the present invention.

[0046] Example 1

[0047] A time series causal driven meteorological disaster prediction method comprises the following steps:

[0048] Step 1, preprocess the time series data of meteorological disasters and calculate its optimal lag window and slice length; preprocess the time series data of meteorological disasters, initialize the lag window, increase the size of the lag window and analyze the changes in the number of causal relationships caused by it, and determine the lag window when the causal relationship does not increase with the increase of the lag window as the optimal lag window; initialize the slice length, reduce the slice length and analyze the changes in the number of causal relationships caused by it, and determine the slice length when the causal relationship does not decrease with the decrease of the slice length as the optimal slice length.

[0049] Step 1.1.1, input the time series data of meteorological disasters Including the number of variables R and the number of time T; the initial lag window τ; the initial time series window causal graph G; the causal relationship accumulation threshold α; the causal relationship increment β;

[0050] Step 1.1.2: Use the PCMCI algorithm to process the meteorological disaster time series data to obtain the current time series window causal graph G τ ;

[0051] Step 1.1.3, the current time series window causal graph G τ Compare with the initial time series window causal graph G to check whether there is a new causal relationship added. If a new causal relationship is found, reset the causal relationship increment β to 0; otherwise, increase the value of the causal relationship increment β;

[0052] Step 1.1.4, determine whether the causal relationship increment β is less than the set causal relationship accumulation threshold α. If β<α, the current hysteresis window τ is the optimal hysteresis window; otherwise, the current G τ Update to G, τ = τ + 1, and return to step 1.1.2 for the next round of iteration;

[0053] Step 1.1.5, by iterating steps 1.1.2 to 1.1.4, finally determine the optimal lag window τ x .

[0054] Step 1.2.1, Input: Time series data of meteorological disasters Including the number of variables R and the time number T; the starting time point t; the time length Optimal hysteresis window τ x ; Initial time window causal graph G; ideal causal relationship reduction ratio γ, actual causal relationship reduction ratio σ;

[0055] Step 1.2.2, use the PCMCI algorithm to process the meteorological disaster time series data and the lag window τ. For detailed implementation, see step 4 to obtain the current time series window causal diagram

[0056] Step 1.2.3, the current time series window cause and effect diagram Compare with the initial time series window causal graph G to check whether there is a reduction in causal relationships; if a reduction in causal relationships is found, the actual causal relationship is reduced by a certain ratio. Where φ(DAG) is used to represent the number of edges in the statistical graph DAG; if there is none, the actual causal relationship reduction ratio σ=0;

[0057] Step 1.2.4, determine whether the actual causal relationship reduction ratio σ is greater than the set ideal causal relationship reduction ratio γ. If σ>γ, the slice length at this time is is the optimal slice length; otherwise, keep G unchanged. And return to step 1.2.2 for the next round of iteration;

[0058] Step 1.2.5, by iterating steps 2.2 to 2.4, finally determine the optimal slice length

[0059] Step 2: Divide the time series data into multiple time segments of equal length according to the slice length;

[0060] According to step 2, the optimal slice length is obtained Time series data Divide into multiple segments of equal length; the length of each segment is the optimal slice length, until all data are segmented or the remaining length is not enough to form a complete time series segment; obtain time segments, the optimal lag window for each time segment is τ x ;

[0061] Step 3: Perform time series causal discovery independently for each time segment to obtain a window causal structure diagram;

[0062] Step 3.1, Input: Time series data fragment Optimal hysteresis window τ x .

[0063] Step 3.2, PC1 stage: Initialize various variables The parent node of Then the first iteration (p = 0) is performed to test unconditional independence. and The hypothesis of independence cannot be established at the significance level α PC If rejected, from The parent node of In the next iteration (p→p+1), first sort the parent nodes according to the test statistic, and then test if S is known. and Whether it is independent, where S is Zhongyu The set of p nodes with the strongest relationship; after each iteration, the parent nodes independent of the test node are removed from Remove until no further tests are possible and the algorithm converges;

[0064] Step 3.3, MCI stage: Based on the edges retained in step 3.2 Consider at the same time and Proceed to step 3.2, which accounts for autocorrelation while identifying indirect and common cause links, thereby controlling the false positive rate to within the expected level;

[0065] Step 3.4, after the above two stages, what is finally obtained is the time series window causal graph.

[0066] Step 4: Integrate the window causal structure graphs obtained from multiple time segments through ensemble learning to obtain a globally consistent temporal causal structure graph:

[0067] Step 4.1, align the k time series window causal graphs obtained in step 3 in time and variables to ensure that they maintain a corresponding relationship between time points and time series variables;

[0068] Step 4.2, counting the number of causal relationships at all positions in the causal graph of each time series window to form a causal relationship quantity matrix;

[0069] Step 4.3, process the causal relationship quantity matrix;

[0070] Step 4.3.1, first calculate the mean value of the causal relationship quantity matrix and store the result in the variable mean;

[0071] Step 4.3.2, then calculate the absolute difference between each element and the mean, and store these differences in an array diff;

[0072] Step 4.3.3, finally find the indexes of the elements in the difference array whose absolute difference is greater than the preset threshold, set the values ​​of these indexes in the corresponding positions of the causal relationship matrix to 1, and set the values ​​of the remaining positions to 0;

[0073] Step 4.4, performing mean processing on the k test statistic value matrices obtained in step 4 to obtain a mean test statistic value matrix;

[0074] Step 4.5, compress the causal relationship matrix in the time dimension. If there is a value of 1 between two time series variables in the entire time dimension, the corresponding position in the final causal relationship matrix is ​​set to 1; otherwise, it is set to 0;

[0075] Step 4.6, based on the mean test statistic matrix, determine the basic causal diagram rules for the causal relationship matrix; when a bidirectional edge is found, give priority to the side with a larger test statistic value, discard the side with a smaller test statistic value, and ensure that there is no loop in the graph;

[0076] Step 4.7, through the above operations, a final globally consistent temporal causal structure diagram is formed.

[0077] In order to better demonstrate the effect of this time series causal discovery method, the results of this method and existing methods on the meteorological disaster dataset are compared. The following is a detailed description of this dataset:

[0078] Meteorological disaster dataset: This dataset is generated by a time series data generator based on the known mechanism of flood occurrence and the key meteorological factors related to floods. It contains 10 time series with 1,000 sets of data with different noise.

[0079] For the evaluation of temporal causal relationship discovery, this paper uses precision (PRE), recall (Recall, Rec), and F1 value (F1Score) as evaluation indicators of the causal network structure learned by the algorithm to verify the correctness and effectiveness of the algorithm in this paper. Precision refers to the ratio of the number of edges of correct causal relationships identified from temporal causal data to the total number of edges identified; recall refers to the ratio of the number of edges correctly identified by the algorithm to the total number of edges in the original true causal temporal graph; F1 value combines precision and recall, and is obtained by the harmonic mean of precision and recall, which can balance between false positive control and high detection ability to restore true causal relationships.

[0080] This paper selects four state-of-the-art temporal causal discovery methods as baselines, which are described in detail as follows:

[0081] μ-PC: A constraint-based temporal causal discovery method that extends the PC algorithm to the temporal domain based on the μ-separation theory.

[0082] PCMCI: A constraint-based approach to temporal causal discovery, especially when dealing with high-dimensional data. Its goal is to identify causal relationships and dependency structures from data.

[0083] ITGH: This method proposes a general information-theoretic framework for causal inference on heterogeneous datasets using generalized linear models.

[0084] CUTE: This method effectively handles uncertain data through conditional independence detection and constructs a robust causal graph structure to reveal the causal relationship between variables.

[0085] Table 1: Comparison results on meteorological disaster datasets

[0086]

[0087] Table 2: Comparison results of meteorological disaster datasets at different lag window sizes

[0088]

[0089] From the above two tables, it can be found that the results of the temporal causal discovery method proposed in the present invention on the data set are basically better than the existing methods.

[0090] The contents not described in detail in the specification of the present invention belong to the prior art known to the professional and technical personnel in the field. Although the illustrative specific embodiments of the present invention are described above to facilitate the understanding of the present invention by the technical personnel in the field, it should be clear that the present invention is not limited to the scope of the specific embodiments. For the ordinary technical personnel in the field, as long as various changes are within the spirit and scope of the present invention defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.

Claims

1. A time series causal driven meteorological disaster prediction method, characterized in that: The following steps are involved: Step 1: preprocess the time series data of meteorological disasters and calculate the optimal lag window and slice length; Step 2: Divide the time series data into multiple time segments of equal length according to the slice length; Step 3: Perform time series causal discovery independently for each time segment to obtain a window causal structure diagram; Step 4: Integrate the window causal structure graphs obtained from multiple time segments through an ensemble learning method to obtain a globally consistent temporal causal structure graph.

2. The method for predicting meteorological disasters driven by temporal causality according to claim 1, characterized in that: The step 1 is to pre-process the time series data of meteorological disasters and calculate the optimal lag window and slice length. Specifically, the time series data of meteorological disasters are pre-processed, the lag window is initialized, the size of the lag window is increased and the quantitative changes of causal relationships caused thereby are analyzed, and the lag window when the causal relationship no longer increases with the increase of the lag window is determined as the optimal lag window; the slice length is initialized, the slice length is reduced and the quantitative changes of causal relationships caused thereby are analyzed, and the slice length when the causal relationship no longer decreases with the decrease of the slice length is determined as the optimal slice length.

3. A time series causal driven meteorological disaster prediction method according to any one of claims 1 or 2, characterized in that: Calculating the optimal hysteresis window and slice length in step 1 specifically includes the following steps: Step 1.1.1, input the time series data of meteorological disasters Including the number of variables R and the number of time T; the initial lag window τ; the initial time series window causal graph G; the causal relationship accumulation threshold α; the causal relationship increment β; Step 1.1.2: Use the PCMCI algorithm to process the meteorological disaster time series data to obtain the current time series window causal graph G τ ; Step 1.1.3, the current time series window causal graph G τ Compare with the initial time series window causal graph G to check whether there is a new causal relationship added. If a new causal relationship is found, reset the causal relationship increment β to 0; otherwise, increase the value of the causal relationship increment β; Step 1.1.4, determine whether the causal relationship increment β is less than the set causal relationship accumulation threshold α. If β<α, the current hysteresis window τ is the optimal hysteresis window; otherwise, the current G τ Update to G, τ = τ + 1, and return to step 1.1.2 for the next round of iteration; Step 1.1.5, by iterating steps 1.1.2 to 1.1.4, finally determine the optimal lag window τ x . Step 1.2.1, Input: Time series data of meteorological disasters Including the number of variables R and the time number T; the starting time point t; the time length Optimal hysteresis window τ x ; Initial time window causal graph G; ideal causal relationship reduction ratio γ, actual causal relationship reduction ratio σ; Step 1.2.2, use the PCMCI algorithm to process the meteorological disaster time series data and the lag window τ. For detailed implementation, see step 4 to obtain the current time series window causal diagram Step 1.2.3, the current time series window cause and effect diagram Compare with the initial time series window causal graph G to check whether there is a reduction in causal relationships; if a reduction in causal relationships is found, the actual causal relationship is reduced by a certain ratio. Where φ(DAG) is used to represent the number of edges in the statistical graph DAG; if there is none, the actual causal relationship reduction ratio σ=0; Step 1.2.4, determine whether the actual causal relationship reduction ratio σ is greater than the set ideal causal relationship reduction ratio γ. If σ>γ, the slice length at this time is is the optimal slice length; otherwise, keep G unchanged. And return to step 1.2.2 for the next round of iteration; Step 1.2.5, by iterating steps 2.2 to 2.4, finally determine the optimal slice length 4. The method for predicting meteorological disasters driven by temporal causality according to claim 1, characterized in that: The step 2 is to divide the time series data into a plurality of time segments of equal length according to the slice length; According to step 2, the optimal slice length is obtained Time series data Divide into multiple segments of equal length; the length of each segment is the optimal slice length, until all data are segmented or the remaining length is not enough to form a complete time series segment; obtain time segments, the optimal lag window for each time segment is τ x。 5. The method for predicting meteorological disasters driven by temporal causality according to claim 1, characterized in that: In step 3, time series causal discovery is performed independently for each time segment to obtain a window causal structure diagram; Step 3.1, Input: Time series data fragment Optimal hysteresis window τ x . Step 3.2, PC1 stage: Initialize various variables The parent node of Then the first iteration (p = 0) is performed to test unconditional independence. and The hypothesis of independence cannot be established at the significance level α PC If rejected, from The parent node of In the next iteration (p→p+1), first sort the parent nodes according to the test statistic, and then test if S is known. and Whether it is independent, where S is Zhongyu The set of p nodes with the strongest relationship; After each iteration, the parent nodes that are independent of the test node are removed from Remove until no further tests are possible and the algorithm converges; Step 3.3, MCI stage: Based on the edges retained in step 3.2 Consider at the same time and Proceed to step 3.2, which accounts for autocorrelation while identifying indirect and common cause links, thereby controlling the false positive rate to within the expected level; Step 3.4, after the above two stages, what is finally obtained is the time series window causal graph.

6. The method for predicting meteorological disasters driven by temporal causality according to claim 1, characterized in that: The step 4 integrates the window causal structure graphs obtained from multiple time segments through an integrated learning method to obtain a globally consistent temporal causal structure graph. Step 4.1, align the k time series window causal graphs obtained in step 3 in time and variables to ensure that they maintain a corresponding relationship between time points and time series variables; Step 4.2, counting the number of causal relationships at all positions in the causal graph of each time series window to form a causal relationship quantity matrix; Step 4.3, process the causal relationship quantity matrix; Step 4.3.1, first calculate the mean value of the causal relationship quantity matrix and store the result in the variable mean; Step 4.3.2, then calculate the absolute difference between each element and the mean, and store these differences in an array diff; Step 4.3.3, finally find the indexes of the elements in the difference array whose absolute difference is greater than the preset threshold, set the values ​​of these indexes in the corresponding positions of the causal relationship matrix to 1, and set the values ​​of the remaining positions to 0; Step 4.4, performing mean processing on the k test statistic value matrices obtained in step 4 to obtain a mean test statistic value matrix; Step 4.5, compress the causal relationship matrix in the time dimension. If there is a value of 1 between two time series variables in the entire time dimension, the corresponding position in the final causal relationship matrix is ​​set to 1; Otherwise set to 0; Step 4.6, based on the mean test statistic matrix, determine the basic causal diagram rules for the causal relationship matrix; when a bidirectional edge is found, give priority to the side with a larger test statistic value, discard the side with a smaller test statistic value, and ensure that there is no loop in the graph; Step 4.7, through the above operations, a final globally consistent temporal causal structure diagram is formed.

Images