Traffic meteorological disaster early warning method and system based on multi-source data fusion

Abstract

This invention discloses a traffic meteorological disaster early warning method and system based on multi-source data fusion, belonging to the technical field of power system energy storage equipment. The method includes: responding to the acquisition of meteorological observation data, extracting and obtaining a first feature parameter, and determining the baseline history of disasters for each road segment based on the first feature parameter; constructing a structural causal model; applying an intervention operation to the structural causal model, and inferring counterfactual history feature parameters of disasters for each road segment through the structural causal model after the intervention operation; comparing the baseline history with the counterfactual history feature parameters, and generating graded early warning information based on the comparison result. This invention, by constructing a structural causal model, explicitly encodes the causal relationship between meteorological observation data and road segment attribute data into a directed acyclic graph, solving the deficiency of existing methods that can only capture correlations but cannot identify causal driving mechanisms, and reducing false alarms caused by confounding variables.

Description

Technical Field

[0001] This invention relates to the field of meteorological early warning technology, specifically to a method and system for early warning of traffic meteorological disasters based on multi-source data fusion. Background Technology

[0002] Traffic meteorological disasters pose a significant threat to highway driving safety. Relevant departments typically rely on multi-source equipment such as meteorological stations and road surface sensors along the route to continuously collect meteorological observation data, and based on this, assess and issue warnings about the disaster risks of each road section.

[0003] Most existing early warning methods are based on threshold models established by statistical regularities of observational data. These models can only reflect the correlation between meteorological elements and disaster risk, but cannot identify the causal mechanism between the two. In practical applications, when multiple meteorological elements occur abnormally at the same time, existing methods struggle to distinguish which elements are the true driving factors of the disaster and which are merely accompanying phenomena. This leads to problems such as high false alarm rates and inaccurate risk level assessments in early warning results. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention provides a method and system for early warning of traffic and meteorological disasters based on multi-source data fusion.

[0005] Therefore, the technical problem solved by this invention is that most existing early warning methods are based on the statistical regularity of observation data to establish threshold models, which can only reflect the correlation between meteorological elements and disaster risk, but cannot identify the causal mechanism between the two.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a traffic meteorological disaster early warning method based on multi-source data fusion, comprising: in response to the collection of meteorological observation data, extracting and obtaining a first feature parameter, and determining the baseline history of disasters for each road segment based on the first feature parameter; constructing a structural causal model using the meteorological observation data and attribute data of each road segment as input; applying an intervention operation to the structural causal model, and inferring the counterfactual history feature parameters of disasters for each road segment through the structural causal model after the intervention operation; comparing the baseline history with the counterfactual history feature parameters, and generating graded early warning information based on the comparison result.

[0007] As a preferred embodiment of the traffic meteorological disaster early warning method based on multi-source data fusion described in this invention, the step of extracting the first feature parameter includes: synchronizing the collected meteorological observation data with timestamps, and preprocessing the synchronized meteorological observation data to obtain a meteorological element matrix; extracting the original measurement values ​​of meteorological elements for each road segment from the meteorological element matrix; performing statistical feature calculation on the original measurement values ​​of meteorological elements using a preset sliding time window, and using the calculation result as the first feature parameter.

[0008] As a preferred embodiment of the traffic meteorological disaster early warning method based on multi-source data fusion described in this invention, the step of preprocessing the synchronized meteorological observation data includes: filling in the missing values ​​in the meteorological element matrix column by column along the time axis to obtain a meteorological element matrix without missing values; calculating the mean and standard deviation of the meteorological observation data in each column of the meteorological element matrix without missing values, and replacing values ​​that deviate from the mean by more than M times the standard deviation with the mean of the two adjacent valid times of that column to obtain a primary element matrix; calculating the minimum and maximum values ​​of each column of the primary element matrix, and linearly mapping each column of data to the [0,1] interval to obtain the meteorological element matrix.

[0009] As a preferred embodiment of the traffic meteorological disaster early warning method based on multi-source data fusion described in this invention, the step of constructing a structural causal model includes: using meteorological observation data and road segment attribute data as nodes, performing a conditional independence test on the partial correlation coefficient between each pair of node variables, and obtaining an undirected skeleton graph based on the results of the conditional independence test; performing V-shaped structure recognition on each edge of the undirected skeleton graph, detecting whether the node triples satisfy the pattern that X and Y are independent when not conditional on Z, and not independent when conditional on Z, and assigning triples that satisfy this pattern to the nodes. Two edges in a tuple are determined as directed edges from X to Z and from Y to Z, respectively, where (X, Y, Z) is a node triplet. The remaining edges with undetermined directions are sequentially inferred to obtain a partially directed acyclic graph. For edges in the partially directed acyclic graph whose directions cannot be determined, the final direction is determined based on the minimum length criterion to obtain a complete directed acyclic graph. Based on the complete directed acyclic graph, for each node, the conditional probability distribution parameters of the node are fitted using the maximum likelihood estimation method, given its set of parent nodes, to obtain a structural causal model.

[0010] As a preferred embodiment of the traffic meteorological disaster early warning method based on multi-source data fusion described in this invention, the step of applying intervention to the structural causal model includes: performing a noise locking mechanism on the structural causal model based on the actual observation values ​​of each node at the current time to obtain a noise locking state; retrieving the current intervention scenario from the intervention scenario set, performing counterfactual inference under the noise locking state to obtain the counterfactual values ​​of each road segment disaster node under the intervention scenario at the current time; rolling forward with a fixed time step within a preset prediction time window, splicing the counterfactual values ​​in chronological order to form the counterfactual history of each road segment disaster under the intervention scenario; extracting counterfactual feature parameters from the counterfactual history, and taking the maximum value of the counterfactual feature parameters of each road segment under all intervention scenarios as the final output counterfactual history feature parameters of the road segment.

[0011] As a preferred embodiment of the traffic and meteorological disaster early warning method based on multi-source data fusion described in this invention, the noise locking mechanism includes the following steps: substituting the actual observation values ​​of each node at the current moment into the structural causal model; for each non-intervention node, according to the reverse order of the topological sorting of the directed acyclic graph, substituting the actual observation value of the node's parent node into the conditional probability distribution function of the node, and solving the current value of the node's exogenous noise variable in reverse; fixing the values ​​of the exogenous noise variables of all nodes to obtain the noise locking state.

[0012] As a preferred embodiment of the traffic meteorological disaster early warning method based on multi-source data fusion described in this invention, the step of setting the intervention scenario set includes: for each meteorological element node in the structural causal model, fixing the meteorological element node to the 95th percentile and 50th percentile of its historical observation sequence, deleting all incoming edges of the node, performing forward propagation on each road segment disaster node in the truncated model, calculating the difference between the output probabilities of each road segment disaster node under the two fixed value conditions, and using this difference as the average causal effect of the meteorological element node on each road segment disaster node; sorting all meteorological element nodes in descending order according to the mean of their average causal effect on all road segment disaster nodes, selecting the meteorological element nodes ranked in the top K positions as intervention variables, where K is a preset positive integer; for each selected intervention variable, sampling M values ​​between the upper and lower bounds according to a preset step size, using the 95th percentile of its historical observation sequence as the upper bound and the 5th percentile as the lower bound, constructing the intervention value sequence of the intervention variable; combining the intervention value sequences of each intervention variable to obtain the intervention scenario set.

[0013] As a preferred embodiment of the traffic meteorological disaster early warning method based on multi-source data fusion described in this invention, the step of performing counterfactual inference includes: replacing the values ​​of intervention variable nodes with corresponding intervention values ​​and deleting all their incoming edges; in the noise-locked state, the values ​​of exogenous noise variables of all non-intervention nodes remain unchanged, and forward propagation is re-executed for all non-intervention nodes according to topological sorting to obtain the counterfactual values ​​of disaster nodes of each road segment under the intervention scenario at the current time.

[0014] As a preferred embodiment of the traffic meteorological disaster early warning method based on multi-source data fusion described in this invention, the step of comparing the baseline history with the counterfactual history feature parameters includes: extracting baseline history feature parameters of the same type as the counterfactual history feature parameters from the baseline history; calculating the deviation rate between the counterfactual history feature parameters and the baseline history feature parameters; comparing the deviation rate with a preset set of hierarchical thresholds, the set of hierarchical thresholds including a first threshold, a second threshold, and a third threshold that increase sequentially; when the deviation rate is less than the first threshold, determining that the current disaster risk of the road segment is in a normal state and not triggering an early warning; when the deviation rate is greater than or equal to the first threshold and less than the second threshold, determining that the disaster risk of the road segment is low risk; when the deviation rate is greater than or equal to the second threshold and less than the third threshold, determining that the disaster risk of the road segment is medium risk; when the deviation rate is greater than or equal to the third threshold, determining that the disaster risk of the road segment is high risk; and taking the highest level of the early warning level for each moment within a preset prediction time window for the same road segment as the final output early warning level for that road segment.

[0015] This invention provides a traffic meteorological disaster early warning system based on multi-source data fusion.

[0016] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a traffic meteorological disaster early warning system based on multi-source data fusion, comprising: a data acquisition module, used to acquire meteorological observation data, extract and obtain a first feature parameter, and determine the baseline history of disasters in each road section based on the first feature parameter;

[0017] The model building module is used to construct a structural causal model using the meteorological observation data and the attribute data of each road section as input.

[0018] The judgment module is used to apply an intervention operation to the structural causal model and infer the counterfactual history characteristic parameters of the disaster in each road segment through the structural causal model after the intervention operation is applied.

[0019] The early warning module is used to compare the baseline history with the counterfactual history feature parameters and generate graded early warning information based on the comparison results.

[0020] The beneficial effects of this invention are as follows: By constructing a structural causal model, the causal relationship between meteorological observation data and road segment attribute data is explicitly encoded into a directed acyclic graph, overcoming the shortcomings of existing methods that can only capture correlations but cannot identify causal driving mechanisms, thus reducing false alarms caused by confounding variables. Furthermore, by introducing a counterfactual inference mechanism and applying intervention operations to key meteorological element nodes, the evolution of disaster risks for each road segment under various extreme weather scenarios is simulated under noise-locked conditions, achieving forward-looking predictions of future adverse scenarios and compensating for the lack of scenario extrapolation capabilities in existing methods. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 The above is a flowchart of a traffic meteorological disaster early warning method based on multi-source data fusion, which is an embodiment of the present invention. Detailed Implementation

[0023] To make the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0024] Example 1, referring to Figure 1 This is one embodiment of the present invention, which provides a traffic and meteorological disaster early warning method based on multi-source data fusion, including the following steps:

[0025] S1. In response to the collection of meteorological observation data, extract the first feature parameter, and determine the baseline history of disasters in each road section based on the first feature parameter.

[0026] S2. Using the meteorological observation data and the attribute data of each road section as input, construct a structural causal model.

[0027] S3. Apply an intervention operation to the structural causal model, and infer the counterfactual history characteristic parameters of each road segment disaster through the structural causal model after the intervention operation.

[0028] S4. Compare the baseline history with the counterfactual history feature parameters, and generate graded early warning information based on the comparison results.

[0029] Example 2, an embodiment of the present invention, provides a traffic meteorological disaster early warning method based on multi-source data fusion, based on the previous embodiment, including the following steps:

[0030] S1. In response to the collection of meteorological observation data, extract the first feature parameter, and determine the baseline history of disasters in each road section based on the first feature parameter.

[0031] In this embodiment, meteorological observation data is collected by multiple sources of equipment such as automatic weather stations, visibility meters, road surface condition sensors and radar echoes deployed along the route, including but not limited to air temperature, relative humidity, wind speed, wind direction, precipitation, visibility and road surface friction coefficient.

[0032] The steps for extracting the first feature parameter include S1.1 to S1.3:

[0033] S1.1. The collected meteorological observation data is timestamped and then preprocessed to obtain a meteorological element matrix.

[0034] It is important to know that, due to the sampling time discrepancies between different meteorological observation equipment, the timestamps of the data reported by each equipment are first aligned using Coordinated Universal Time as the reference, and all observation values ​​are interpolated or rounded to a unified time grid.

[0035] The preprocessing steps for synchronized meteorological observation data include A1 to A3:

[0036] A1. Fill in the missing values ​​in the meteorological element matrix column by column along the time axis to obtain a meteorological element matrix without missing values.

[0037] Missing values ​​in the meteorological element matrix are marked as null values. For each column of the matrix, linear interpolation is performed to fill in the missing values ​​along the time axis. If an observation value at a certain time is missing, linear interpolation is performed using the value of the nearest previous valid time and the value of the next valid time in that column. The calculation formula is as follows:

[0038] ;

[0039] in, The interpolation fill-in result at time t, The previous valid time closest to time t The observed values, The next valid time closest to time t The observed values.

[0040] For missing values ​​located at the beginning or end of the sequence, i.e., when there are no valid neighboring values ​​in either the forward or backward direction, nearest neighbor imputation is used, i.e., the nearest valid time observation is used to directly fill the missing value.

[0041] A2. Calculate the mean and standard deviation of meteorological observation data in each column of the missing meteorological element matrix. Replace the values ​​that deviate from the mean by more than M times the standard deviation with the mean of the two adjacent valid times in that column to obtain the primary element matrix.

[0042] The formula for calculating the mean is expressed as:

[0043] ;

[0044] The formula for calculating standard deviation is as follows:

[0045] ;

[0046] in, The mean, Standard deviation, is the observation value of the f-th meteorological element channel at time t; T is the total number of time steps.

[0047] For the column that satisfy The observed value is identified as an outlier and replaced with the mean of the observations at the two adjacent valid times on the time axis. The calculation formula is as follows:

[0048] ;

[0049] in, The values ​​for the real-time meteorological elements are after outlier correction. The time t is the non-abnormal valid time that is closest to the time in the forward direction on the time axis. Observations at that location The time t is the non-abnormal valid time that is closest to the time axis in the backward direction. The observed value at point M is a preset positive integer used to control the sensitivity of outlier detection; in this embodiment, it is set to 3. After outlier correction is completed for all columns, a primary feature matrix is ​​obtained.

[0050] In a specific example, the wind speed channel of a certain road segment has a total of T=100 observations, which are calculated as follows: , The outlier threshold range is [2.0, 14.0] m / s, meaning that observations with values ​​outside this range are considered outliers.

[0051] If the observed wind speed at t=35 is 22.0 m / s, exceeding the upper bound, and the wind speed at the preceding non-abnormal valid time point is 9.0 m / s and the wind speed at the following non-abnormal valid time point is 8.5 m / s, then the corrected wind speed for that time is: This replaced the anomalous observations that deviated from reality.

[0052] A3. Calculate the minimum and maximum values ​​for each column of the primary element matrix, and linearly map each column of data to the [0,1] interval to obtain the meteorological element matrix.

[0053] In this embodiment, after outlier correction, the minimum value of the channel within 100 time points is 2.1 m / s and the maximum value is 13.8 m / s. After normalization using the Min-Max linear normalization formula, the value range of each column of data is uniformly mapped to the interval [0,1].

[0054] When the maximum and minimum values ​​in a column are equal, indicating that the element remained unchanged during the observation period, all values ​​in that column are uniformly assigned to 0 to avoid division by zero anomalies. After normalizing all columns, the final meteorological element matrix is ​​obtained.

[0055] S1.2 Extract the original measurements of meteorological elements for each road segment from the meteorological element matrix.

[0056] In one embodiment, a pre-established mapping relationship exists between the column indexes of the meteorological element matrix and the meteorological observation stations and element channels of each road segment. Based on this mapping relationship, for each road segment, all element channel columns belonging to that road segment are retrieved to form a meteorological element submatrix for that road segment.

[0057] If a road segment covers multiple observation stations, the normalized values ​​of the same type of element at each station are taken as the arithmetic mean and used as the representative measurement value of that element for that road segment.

[0058] S1.3 Statistical feature calculations are performed on the original measurements of the meteorological elements using a preset sliding time window, and the calculation results are used as the first feature parameter.

[0059] In one embodiment, the sliding time window width is set to W (for example, W=12, corresponding to 60 minutes of historical observation), and the sliding step size is 1 time step. For each road segment r and each meteorological element channel f, at time t, W consecutive normalized observations of that element within the time window [t-W+1,t] are taken.

[0060] For calculating the mean When, it is represented as: ;

[0061] With respect to standard deviation When, it is represented as: ;

[0062] For calculating the rate of change When, it is represented as: .

[0063] in, Let be the normalized value, and i be the index of the time within the window. This is an index for meteorological element channels.

[0064] The calculated statistics are sequentially concatenated across all feature channels of road segment r to form the first feature parameter vector of road segment r at time t.

[0065] S2. Using the meteorological observation data and the attribute data of each road section as input, construct a structural causal model.

[0066] In this embodiment, road segment attribute data includes, but is not limited to, road segment slope, pavement material type, curve curvature, elevation, and historical disaster frequency. The steps for constructing the structural causal model include S2.1 to S2.5:

[0067] S2.1. Using meteorological observation data and road segment attribute data as nodes, perform a conditional independence test on the partial correlation coefficient between each pair of node variables, and obtain an undirected skeleton graph based on the results of the conditional independence test.

[0068] In this embodiment, there exists an undirected edge between any two nodes in the node set. For each pair of nodes, from the empty condition set... Initially, the size of the condition set is increased sequentially. From the nodes that are currently adjacent to the two nodes, a condition set of size |S| is enumerated, and a partial correlation coefficient test is performed on each candidate condition set S.

[0069] Calculate the partial correlation coefficient between two nodes under a given condition set S. If the partial correlation coefficient meets the set standardized value, the two nodes are considered conditionally independent under the given condition set S. Otherwise, increase the size of the condition set and continue the test until the size of the condition set exceeds the current minimum number of neighboring nodes. Traverse all node pairs to complete the conditional independence test, retaining the edges that have not been deleted, and obtain the undirected skeleton graph.

[0070] S2.2 Perform V-shaped structure recognition on each edge of the undirected skeleton graph, and detect whether the node triple (X,Y,Z) satisfies the pattern that X and Y are independent when not defined by Z and not independent when defined by Z. Determine the two edges in the triple that satisfies the pattern as directed edges pointing from X to Z and from Y to Z, respectively.

[0071] In this embodiment, the ternary group represents precipitation. Road surface water depth ,visibility If precipitation and visibility are not adjacent, and The data does not include the road surface water depth, so it is identified as a V-shaped structure. The directional edge precipitation is determined to point to the road surface water depth and the visibility is determined to point to the road surface water depth.

[0072] S2.3. Perform direction deduction on the remaining edges with undetermined directions in turn to obtain a partially directed acyclic graph.

[0073] In this embodiment, for the undirected edges whose direction is still undetermined after the V-shaped structure identification in step S2.2, the direction inference rules are applied sequentially until no new direction can be determined.

[0074] For example, if a directed path exists point to ,and and If there exists an undirected edge with an undetermined direction, then the direction of that edge shall be determined as follows: point to If it is directed to point to If a new V-shaped structure is introduced, it will also be oriented as follows: point to .

[0075] For example, if and The direction is not determined, and there are directed paths in the graph. Then and Determined as point to To avoid introducing directed cycles.

[0076] S2.4 For edges in the partially directed acyclic graph whose direction cannot be determined, the final direction is determined based on the minimum length criterion to obtain the complete directed acyclic graph.

[0077] In this embodiment, for some undirected edges in a directed acyclic graph whose direction is not determined, the minimum description length score is calculated. For candidate directions point to The calculation formula is expressed as:

[0078] ;

[0079] in, For nodes Observational data, For the current graph The set of parent nodes that have been determined The parameters are the maximum likelihood estimates under this directional assumption. Let n be the degree of freedom of the conditional probability distribution, and n be the sample size. The negative log-likelihood. This is a penalty term for model complexity.

[0080] In a specific example, relative humidity coefficient of friction with road surface The direction is still not determined after inferring it according to the rules, so calculation is performed. The MDL score is 126.4. The MDL score is 143.7, which is smaller than the former. Therefore, the direction is determined to be relative humidity towards the road surface friction coefficient, which is consistent with the physical causal logic that increased humidity leads to a decrease in the road surface friction coefficient.

[0081] S2.5 Based on the complete directed acyclic graph, for each node, using its set of parent nodes as a condition, the maximum likelihood estimation method is used to fit the conditional probability distribution parameters of the node to obtain the structural causal model.

[0082] In this embodiment, the complete directed acyclic graph (DAG) defines the causal topology between nodes, i.e., which nodes directly influence each node. Based on this, quantitative parameter estimation of the conditional probability distribution of each node is required, enabling the model to calculate the specific output of that node given the values ​​of its parent nodes.

[0083] For each node in the graph, a linear conditional probability distribution is established, with the node's historical observation data as the dependent variable and the historical observation data of all its parent nodes in the directed acyclic graph as the independent variables. This distribution is specifically represented as follows:

[0084] ;

[0085] in, This represents the value of the current node. This is the intercept term for the node, reflecting the baseline level of the node when all values ​​of its parent nodes are zero. Let be the set of parent nodes of the node in the directed acyclic graph; parent node The causality coefficient of the nodes, This is an exogenous noise variable.

[0086] The maximum likelihood estimation method is used to estimate the intercept term and causality coefficient in the structural equation by using samples from all times in the historical observation data as the training set.

[0087] Maximum likelihood estimation finds a set of parameter values ​​that maximize the probability of the observed data occurring at those parameters. For linear Gaussian models, maximum likelihood estimation is equivalent to least squares estimation, which solves for the parameters by minimizing the sum of squared residuals between the predicted and actual observations at each node.

[0088] For root nodes without parent nodes (such as static attribute nodes of certain road segments), their structural equations degenerate into marginal distributions containing only intercept and noise terms. The intercept term is taken as the historical observation mean, and the standard deviation of the noise term is taken as the historical observation standard deviation.

[0089] After performing parameter estimation on all nodes in the directed acyclic graph, each node corresponds to a known set of intercept terms, causality coefficients, and noise distribution parameters. The structural equations of all nodes are then combined with the topological structure of the directed acyclic graph to form a complete structural causal model.

[0090] S3. Apply an intervention operation to the structural causal model, and infer the counterfactual history characteristic parameters of each road segment disaster through the structural causal model after the intervention operation.

[0091] In this embodiment, the intervention operation involves artificially fixing the value of a certain meteorological element node to a specified value, while simultaneously severing all causal connections between that node and its cause node. This ensures that the value of that node is no longer affected by its upstream factors and is determined solely by the externally specified value. The changes in the output of disaster nodes in each road segment are observed under this forced intervention.

[0092] The steps for applying the intervention operation to the structural causal model include S3.1 to S3.4:

[0093] S3.1. Based on the actual observation values ​​of each node at the current time, a noise locking mechanism is executed on the structural causal model to obtain the noise locking state.

[0094] The noise lock-in mechanism includes steps B1 to B2:

[0095] B1. Substitute the actual observation values ​​of each node at the current moment into the structural causal model. For each non-intervention node, sort it in reverse order of the directed acyclic graph topology. Substitute the actual observation value of the node's parent node into the conditional probability distribution function of the node, and solve the current value of the exogenous noise variable of the node in reverse.

[0096] In this embodiment, node The structural equation is:

[0097] ;

[0098] What needs to be known is that nodes The value of is equal to the weighted sum of the values ​​of all its parent nodes plus the intercept. In addition to exogenous noise .

[0099] Therefore, the current time node Actual observed values The actual observation values ​​of its parent node By substituting the values, we can reverse the process and find the exogenous noise value of the node at the current time.

[0100] ;

[0101] This can be expressed as the exogenous noise of a node at the current time being equal to the actual observation value minus the part that the model can explain based on the parent node, i.e., the residual of that node at the current time.

[0102] Perform the above calculations on all non-interventional nodes in reverse order of the directed acyclic graph topological sort, ensuring that the actual observations of the parent node are ready for each node when calculating the residuals. For the root node (without a parent node), its exogenous noise value is the difference between its own actual observation and the mean of the marginal distribution. .

[0103] Specifically, road surface friction coefficient nodes The parent node is The observed value at the current time is The parent node observations are respectively , , Substitute the estimated parameters , , , The calculation yields:

[0104] .

[0105] B2. Fix the values ​​of all exogenous noise variables at all nodes to obtain a noise-locked state.

[0106] In this embodiment, the calculated exogenous noise values ​​for all nodes are uniformly recorded and frozen, and are not recalculated with changes in intervention operations. At this point, the structural causal model enters a noise-locked state.

[0107] S3.2. Extract the current intervention scenario from the set of intervention scenarios, perform counterfactual inference under the noise-locked state, and obtain the counterfactual values ​​of each road segment disaster node under the intervention scenario at the current time.

[0108] The steps for setting the intervention scenario set include C1 to C4:

[0109] C1. For each meteorological element node in the structural causal model, fix the meteorological element node to the 95th percentile and 50th percentile of its historical observation sequence, delete all incoming edges of the node, perform forward propagation on each road segment disaster node in the truncated model, calculate the difference in output probability of each road segment disaster node under the two fixed value conditions, and take the difference as the average causal effect of the meteorological element node on each road segment disaster node.

[0110] In this embodiment, the 95th percentile represents the extreme high value state of the meteorological element, and the 50th percentile represents its typical median state. The meteorological element nodes are... Fix them to these two values ​​respectively, and delete them. All incoming edges are forward propagated along the topological sorting direction to each road segment disaster node in the truncated model to obtain the output probability of each road segment disaster node under the two states.

[0111] Specifically, when the precipitation node is fixed at the 95th percentile, the output probability of the disaster node for a certain road segment is 0.78; when fixed at the 50th percentile, the output probability is 0.31. Therefore, the average causal effect of precipitation on this road segment is 0.78−0.31=0.47, indicating that extreme increases in precipitation have a positive driving effect on the disaster risk of this road segment.

[0112] C2. All meteorological element nodes are sorted in descending order according to the mean of their average causal effect on all road disaster nodes. The meteorological element nodes in the top K positions are selected as intervention variables, where K is a preset positive integer.

[0113] In this embodiment, for each meteorological element node, the arithmetic mean of its average causal effect on all road disaster nodes is taken to obtain the global causal importance score of that meteorological element node. All meteorological element nodes are sorted in descending order of this score, and the top K nodes with the highest scores are selected as the set of intervention variables. In this embodiment, K=3.

[0114] Interventions are only implemented on meteorological factors that actually drive disaster risk, while ignoring meteorological factors with weak causal effects, in order to avoid a combined explosion of intervention scenarios, while ensuring that counterfactual inferences focus on the most critical risk drivers.

[0115] For example, the global causal importance scores, from highest to lowest, are: precipitation (0.43), visibility (0.38), wind speed (0.29), temperature (0.11), and relative humidity (0.09). If K=3, then the set of intervention variables is {precipitation, visibility, and wind speed}.

[0116] C3. For each selected intervention variable, take the 95th percentile of its historical observation sequence as the upper bound and the 5th percentile as the lower bound, and sample M values ​​between the upper and lower bounds at a preset step size to construct the intervention value sequence of the intervention variable.

[0117] In this embodiment, the upper and lower bounds correspond to the extreme high and low values ​​that the intervention variable has historically experienced, respectively. M values ​​are uniformly sampled between the upper and lower bounds at equal intervals to form the intervention value sequence for the intervention variable. This design ensures that the distribution of intervention values ​​covers the complete risk range from extreme low to extreme high values, guaranteeing that counterfactual inference can explore the impact of various extreme states of the meteorological element in history on disaster risk.

[0118] In this embodiment, M is selected as 5. For example, the historical 95th percentile of the precipitation node is 0.91 and the 5th percentile is 0.08. Five values ​​are sampled at equal intervals within the interval [0.08, 0.91] to obtain the intervention value sequence {0.08, 0.29, 0.50, 0.71, 0.91}.

[0119] C4. Combine the intervention value sequences of each intervention variable to obtain the intervention scenario set.

[0120] The steps for performing counterfactual inference include D1~D2:

[0121] D1. Replace the values ​​of the intervention variable nodes with the corresponding intervention values ​​and delete all their incoming edges.

[0122] In this embodiment, for the current intervention scenario i, all incoming edges of each node in the intervention variable set in the directed acyclic graph are deleted, and the node values ​​are forcibly replaced with the corresponding intervention values ​​in the scenario.

[0123] D2. In the noise-locked state, the values ​​of exogenous noise variables of all non-intervention nodes remain unchanged. Forward propagation is re-executed for all non-intervention nodes according to topological sorting to obtain the counterfactual values ​​of disaster nodes of each road segment under the intervention scenario at the current time.

[0124] In this embodiment, keeping the exogenous noise values ​​of all non-intervention nodes frozen in step B2 unchanged, the forward computation is re-executed for each non-intervention node in the direction from the root node to the leaf node according to the topological sorting order of the directed acyclic graph:

[0125] ;

[0126] It is important to know that, in an intervention scenario, non-intervention nodes The counterfactual value is obtained by weighting the counterfactual values ​​of its parent node (if the parent node is an intervention variable, then the intervention value is taken; otherwise, the counterfactual value obtained from forward propagation) according to the causality coefficient, and then adding the individualized exogenous noise that has been locked in that node. This ensures that counterfactual inferences spread the intervention effect while the individual's current background remains unchanged.

[0127] After traversing all non-intervention nodes and completing the forward propagation, the output value of each road segment's disaster node is read, which is the counterfactual value under the intervention scenario at the current moment. .

[0128] For example, in intervention scenarios Under conditions of extreme heavy rainfall, extremely low visibility, and strong winds occurring simultaneously, the counterfactual value of a disaster node on a certain road section is 0.87, while the actual observed value of that node at the current moment is only 0.34. This indicates that if meteorological conditions evolve in this extreme direction, the disaster risk on that road section will increase.

[0129] S3.3. Within the preset prediction time window, the counterfactual values ​​are spliced ​​together in chronological order to form the counterfactual history of each road segment's disaster under the intervention scenario.

[0130] In this embodiment, the preset prediction time window length is 6, and the fixed time step is 1 time step. Starting from the current moment, at each time step, the rolling prediction value of each node at the current moment is used as the replacement for the actual observation value, and the noise locking in step S3.1 and the counterfactual forward propagation in steps D1~D2 are repeated to obtain the counterfactual values ​​of the disaster nodes of each road segment at that time step.

[0131] S3.4 Extract counterfactual feature parameters from the counterfactual process, and take the maximum value of the counterfactual feature parameters of each road segment under all intervention scenarios as the final output counterfactual process feature parameters of that road segment.

[0132] In this embodiment, the mean, maximum value and rate of change of the counterfactual history sequence of road segment r under the intervention scenario are calculated to constitute the counterfactual characteristic parameters under the intervention scenario.

[0133] For the counterfactual characteristic parameters of road segment r under all intervention scenarios in the intervention scenario set, the maximum value is taken in each dimension. In order to select the characteristic parameter corresponding to the scenario with the most severe disaster risk estimate for the road segment among all possible extreme weather intervention combinations, the counterfactual history characteristic parameter of the road segment is used as the final output of the road segment.

[0134] In a specific example, a certain road segment was considered in 125 intervention scenarios. The maximum counterfactual history value under this scenario is 0.91, which is higher than the corresponding values ​​of the other 124 scenarios. The counterfactual feature parameter vector under this scenario was finally selected as the counterfactual history feature parameter of this road segment.

[0135] S4. Compare the baseline history with the counterfactual history feature parameters, and generate graded early warning information based on the comparison results.

[0136] The step of comparing the baseline history with the counterfactual history characteristic parameters includes S4.1 to S4.5:

[0137] S4.1 Extract baseline process feature parameters from the baseline process that are of the same type as the counterfactual process feature parameters.

[0138] In this embodiment, in step S1, the first characteristic parameter of each road segment is obtained based on the current time and historical observation data. This characteristic parameter describes the baseline level of disaster risk of each road segment under the current actual meteorological conditions, and its corresponding history is called the baseline history.

[0139] The counterfactual history characteristic parameters include the history mean, the history maximum, and the history rate of change. The benchmark history characteristic parameters also extract these three types of statistics, and the statistical window length used in both is consistent, which is the preset prediction time window length.

[0140] In a specific example, the counterfactual history characteristic parameters of road segment r include a maximum history value of 0.87, a mean history value of 0.74, and a history change rate of 0.31; correspondingly, the maximum history value of 0.36, the mean history value of 0.29, and the history change rate of 0.08 are extracted from the baseline history of this road segment as the baseline history characteristic parameters.

[0141] S4.2 Calculate the deviation rate between the counterfactual process characteristic parameters and the benchmark process characteristic parameters.

[0142] In this embodiment, the relative increase of the deviation rate counterfactual history characteristic parameter relative to the baseline history characteristic parameter is used to measure the potential deterioration of the disaster risk of the road section under extreme weather intervention relative to the current normal state.

[0143] For each dimension in the feature parameter vector, when calculating the deviation rate, the counterfactual feature parameter is subtracted from the benchmark feature parameter, and then divided by the benchmark feature parameter to obtain the relative deviation ratio of that dimension. When the benchmark feature parameter is close to zero, a preset minimum benchmark value (0.01 in this embodiment) is used to replace the denominator in the calculation.

[0144] The arithmetic mean of the deviation rates of all dimensions in the feature parameter vector is taken to obtain the final comprehensive deviation rate of the road segment.

[0145] In the example, the deviation rate of the maximum historical value is (0.87−0.36) / 0.36=1.42, the deviation rate of the mean historical value is (0.74−0.29) / 0.29=1.55, and the deviation rate of the historical change rate is (0.31−0.08) / 0.08=2.88. The average of the three is (1.42+1.55+2.88) / 3=1.95, which means that the comprehensive deviation rate of this road segment is 1.95, indicating that under the worst-case intervention scenario, the disaster risk index of this road segment is 195% higher than the current normal state.

[0146] S4.3. The deviation rate is compared with a preset set of graded thresholds step by step, wherein the set of graded thresholds includes a first threshold, a second threshold, and a third threshold that increase sequentially.

[0147] In this embodiment, the first threshold, the second threshold, and the third threshold are set to 0.5, 1.0, and 2.0, respectively.

[0148] S4.4 When the deviation rate is less than the first threshold, the current disaster risk of the road segment is determined to be in a normal state, and no warning is triggered; when the deviation rate is greater than or equal to the first threshold and less than the second threshold, the disaster risk of the road segment is determined to be low risk; when the deviation rate is greater than or equal to the second threshold and less than the third threshold, the disaster risk of the road segment is determined to be medium risk; when the deviation rate is greater than or equal to the third threshold, the disaster risk of the road segment is determined to be high risk.

[0149] When the deviation rate is less than 0.5, it indicates that even if the meteorological conditions change drastically, the potential increase in the disaster risk of this road section relative to the current state is still within an acceptable range. There is no need to trigger an early warning at present, and the routine monitoring status should be maintained.

[0150] When the deviation rate is in the range of [0.5, 1.0), it indicates that the trend of extreme weather has potentially increased the disaster risk of this road section to a certain extent. It is recommended to send a reminder to the relevant management departments to pay attention to weather changes and deploy patrol resources in advance if necessary.

[0151] When the overall deviation rate is in the range of [1.0, 2.0), it indicates that the disaster risk of this road section is significantly increased under potential extreme weather scenarios. It is recommended to activate the yellow warning response, issue driving safety tips to road management departments and driving vehicles, and take speed limit or increased patrol measures on high-risk road sections as appropriate.

[0152] A comprehensive deviation rate greater than or equal to 2.0 indicates that the road section faces extremely severe disaster risks under extreme weather conditions. It is recommended to activate an orange or red alert response, take measures such as road closure and control, mandatory detours, or emergency rescue preparations, and notify the emergency management department simultaneously.

[0153] For example, the overall deviation rate of road segment r is 1.95, which falls within the range of [1.0, 2.0) and is judged as a medium risk.

[0154] S4.5. For the same road segment, the highest warning level at each moment within the preset prediction time window shall be taken as the final warning level output for that road segment.

[0155] In this embodiment, a counterfactual history is formed for each time step within the prediction time window. The deviation rate calculation and level determination are also performed independently at each time step. Therefore, the same road segment corresponds to a warning level at each of the H time steps within the prediction time window.

[0156] To avoid missing early warnings due to the average of high-risk conditions at a certain time step, the highest early warning level for the same road segment across all time steps within the prediction time window is taken as the final early warning level to be issued to the public at the current moment.

[0157] For example, the warning levels for road segment r at the six time steps within the prediction time window are: normal, low, low, medium, medium, and low. The highest level is taken as medium risk and is released as the final warning level for that road segment.

[0158] Example 3 is an embodiment of the present invention. This embodiment provides a traffic meteorological disaster early warning system based on multi-source data fusion, including a data acquisition module for acquiring meteorological observation data, extracting and obtaining a first feature parameter, and determining the baseline history of disasters on each road section based on the first feature parameter.

[0159] The model building module is used to construct a structural causal model using the meteorological observation data and the attribute data of each road section as input.

[0160] The judgment module is used to apply an intervention operation to the structural causal model and infer the counterfactual history characteristic parameters of the disaster in each road segment through the structural causal model after the intervention operation is applied.

[0161] The early warning module is used to compare the baseline history with the counterfactual history feature parameters and generate graded early warning information based on the comparison results.

[0162] This embodiment also provides an electronic device applicable to a traffic and meteorological disaster early warning method based on multi-source data fusion, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the traffic and meteorological disaster early warning method based on multi-source data fusion as proposed in the above embodiment.

[0163] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements a traffic and meteorological disaster early warning method based on multi-source data fusion as proposed in the above embodiment.

[0164] The storage medium proposed in this embodiment belongs to the same inventive concept as the method for implementing traffic and meteorological disaster early warning based on multi-source data fusion proposed in the above embodiments. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0165] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0166] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A traffic meteorological disaster early warning method based on multi-source data fusion, characterized in that: Includes the following steps, In response to the collection of meteorological observation data, the first feature parameter is extracted and obtained, and the baseline history of disasters in each road section is determined based on the first feature parameter; Using the meteorological observation data and the attribute data of each road section as input, a structural causal model is constructed; An intervention operation is applied to the structural causal model, and the counterfactual history characteristic parameters of the disaster in each road segment are inferred from the structural causal model after the intervention operation is applied; The baseline history is compared with the counterfactual history feature parameters, and graded early warning information is generated based on the comparison results.

2. The traffic meteorological disaster early warning method based on multi-source data fusion as described in claim 1, characterized in that: The steps for extracting and obtaining the first feature parameter include: The collected meteorological observation data is timestamped and synchronized, and the synchronized meteorological observation data is preprocessed to obtain a meteorological element matrix; Extract the original meteorological element measurements of each road segment from the meteorological element matrix; The original measurements of the meteorological elements are statistically analyzed using a preset sliding time window, and the results are used as the first feature parameter.

3. The traffic meteorological disaster early warning method based on multi-source data fusion as described in claim 2, characterized in that: The preprocessing steps for synchronized meteorological observation data include: The missing values ​​in the meteorological element matrix are filled column by column along the time axis to obtain a meteorological element matrix without missing values. Calculate the mean and standard deviation of meteorological observation data in each column of the missing meteorological element matrix, and replace the values ​​that deviate from the mean by more than M times the standard deviation with the mean of the two adjacent valid times in that column to obtain the primary element matrix. The minimum and maximum values ​​of the primary element matrix are calculated for each column, and the data in each column are linearly mapped to the interval [0,1] to obtain the meteorological element matrix.

4. The traffic meteorological disaster early warning method based on multi-source data fusion as described in claim 3, characterized in that: The steps to construct a structural causal model include: Meteorological observation data and road segment attribute data are used as nodes. Conditional independence tests of partial correlation coefficients are performed between each pair of node variables. An undirected skeleton graph is obtained based on the results of the conditional independence tests. V-shaped structure recognition is performed on each edge of the undirected skeleton graph. The node triples are checked to see if they satisfy the pattern that X and Y are independent when not defined by Z and not independent when defined by Z. The two edges in the triples that satisfy the pattern are determined as directed edges from X to Z and from Y to Z, respectively. X, Y, Z are node triples. By sequentially inferring the direction of the remaining undetermined edges, a partially directed acyclic graph is obtained. For edges in the aforementioned directed acyclic graph whose direction cannot be determined, the final direction is determined based on the minimum length criterion, thus obtaining a complete directed acyclic graph; Based on the complete directed acyclic graph, for each node, the conditional probability distribution parameters of the node are fitted using the maximum likelihood estimation method, given its set of parent nodes, to obtain the structural causal model.

5. The traffic meteorological disaster early warning method based on multi-source data fusion as described in claim 4, characterized in that: The steps for applying an intervention to the structural causal model include: Based on the actual observations of each node at the current moment, a noise locking mechanism is executed on the structural causal model to obtain the noise locking state; Take the current intervention scenario from the set of intervention scenarios, perform counterfactual inference under the noise-locked state, and obtain the counterfactual values ​​of the disaster nodes of each road segment under the intervention scenario at the current time; Within a preset prediction time window, the prediction is rolled forward at a fixed time step, and the counterfactual values ​​are spliced ​​together in chronological order to form the counterfactual history of each road section's disaster under the intervention scenario. Counterfactual feature parameters are extracted from the counterfactual process, and the maximum value of the counterfactual feature parameters of each road segment under all intervention scenarios is taken as the final output counterfactual process feature parameter of that road segment.

6. The traffic meteorological disaster early warning method based on multi-source data fusion as described in claim 5, characterized in that: The steps of the noise lock-in mechanism include: Substitute the actual observation values ​​of each node at the current moment into the structural causal model. For each non-intervention node, sort it in reverse order of the topological order of the directed acyclic graph. Substitute the actual observation value of the node's parent node into the conditional probability distribution function of the node to solve the current value of the exogenous noise variable of the node in reverse. By fixing the values ​​of all exogenous noise variables at all nodes, a noise-locked state is obtained.

7. The traffic meteorological disaster early warning method based on multi-source data fusion as described in claim 6, characterized in that: The steps for setting up the intervention scenario set include: For each meteorological element node in the structural causal model, the meteorological element node is fixed at the 95th percentile and 50th percentile of its historical observation sequence, and all incoming edges of the node are deleted. Forward propagation is performed on each road segment disaster node in the truncated model, and the difference between the output probabilities of each road segment disaster node under the two fixed value conditions is calculated. This difference is taken as the average causal effect of the meteorological element node on each road segment disaster node. All meteorological element nodes are sorted in descending order according to the mean of their average causal effect on all road disaster nodes. The meteorological element nodes in the top K positions are selected as intervention variables, where K is a preset positive integer. For each selected intervention variable, the 95th percentile of its historical observation sequence is used as the upper bound and the 5th percentile as the lower bound. M values ​​are sampled between the upper and lower bounds at a preset step size to construct the intervention value sequence of the intervention variable. The intervention value sequences of each intervention variable are combined to obtain the intervention scenario set.

8. The traffic meteorological disaster early warning method based on multi-source data fusion as described in claim 7, characterized in that: The steps for performing counterfactual inferences include: Replace the values ​​of the intervention variable nodes with the corresponding intervention values ​​and delete all their incoming edges; In the noise-locked state, the values ​​of exogenous noise variables of all non-intervention nodes remain unchanged. Forward propagation is re-executed for all non-intervention nodes according to topological sorting to obtain the counterfactual values ​​of disaster nodes of each road segment under the intervention scenario at the current time.

9. The traffic meteorological disaster early warning method based on multi-source data fusion as described in claim 8, characterized in that: The steps of comparing the baseline history with the counterfactual history characteristic parameters include: Extract baseline history feature parameters of the same type as the counterfactual history feature parameters from the baseline history; Calculate the deviation rate between the counterfactual history characteristic parameters and the baseline history characteristic parameters; The deviation rate is compared step by step with a preset set of graded thresholds, which includes a first threshold, a second threshold, and a third threshold that increase sequentially. When the deviation rate is less than the first threshold, the current disaster risk of the road segment is determined to be normal, and no warning is triggered; when the deviation rate is greater than or equal to the first threshold and less than the second threshold, the disaster risk of the road segment is determined to be low; when the deviation rate is greater than or equal to the second threshold and less than the third threshold, the disaster risk of the road segment is determined to be medium; when the deviation rate is greater than or equal to the third threshold, the disaster risk of the road segment is determined to be high. For the same road segment, the highest warning level at any given time within the preset prediction time window is taken as the final warning level output for that road segment.

10. A traffic meteorological disaster early warning system based on multi-source data fusion, employing the traffic meteorological disaster early warning method based on multi-source data fusion as described in any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to collect meteorological observation data, extract the first feature parameter, and determine the baseline history of disasters in each road section based on the first feature parameter. The model building module is used to construct a structural causal model using the meteorological observation data and the attribute data of each road section as input. The judgment module is used to apply an intervention operation to the structural causal model and infer the counterfactual history characteristic parameters of the disaster in each road segment through the structural causal model after the intervention operation is applied. The early warning module is used to compare the baseline history with the counterfactual history feature parameters and generate graded early warning information based on the comparison results.

Images