Intelligent traffic flow prediction method based on multi-source data fusion

Abstract

The invention provides a traffic flow intelligent prediction method based on multi-source data fusion, and the method comprises the steps: calculating the historical proportional relation between floating traffic flow and actual traffic flow at different time periods and different nodes according to historical time sequence data; identifying a change mode of the permeability and calculating to obtain a reference dynamic permeability value; establishing a permeability time relation matrix based on the reference dynamic permeability value; inputting the floating traffic flow data into the permeability time relation matrix to obtain a query dynamic permeability value; dividing the floating traffic flow data by the corresponding query dynamic permeability value to obtain a continuous estimation sequence of the traffic flow of the whole road network; fusing the continuous estimation sequence with multi-source historical traffic data to generate space-time fusion features; based on the space-time fusion features, the traffic flow prediction result of the future time period is output, the traffic flow prediction precision is significantly improved, the model generalization ability is enhanced, the prediction real-time performance and continuity are improved, and the problem of large prediction deviation is effectively solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent transportation systems, and in particular to a traffic flow intelligent prediction method based on multi-source data fusion. BACKGROUND

[0002] Traffic flow prediction is an important research direction of intelligent transportation systems, and its goal is to predict the trend of future traffic state changes by analyzing historical data. Currently, this field mainly uses data collected by fixed detectors combined with various prediction models to achieve flow estimation. This method can provide certain prediction reference under normal conditions when the traffic flow is relatively stable.

[0003] However, existing prediction methods face significant challenges in dealing with the complex dynamic characteristics of actual road networks, especially in terms of data utilization. The existing technology has limited integration capabilities for multi-source information, making it difficult to effectively capture the dynamic change rules of traffic flow under different conditions. As a result, when traffic conditions fluctuate dramatically, the prediction results often deviate significantly, which cannot meet the actual needs of intelligent transportation systems for accurate prediction. SUMMARY

[0004] The present application provides a traffic flow intelligent prediction method based on multi-source data fusion to solve the technical problem of large traffic flow prediction deviation in the prior art.

[0005] In one aspect, the present application provides a traffic flow intelligent prediction method based on multi-source data fusion, comprising: Obtaining historical time series data of multiple nodes in the traffic network; According to the historical time series data, the historical proportional relationship between the probe car flow and the actual traffic flow in different time periods and different nodes is calculated; Based on the historical proportional relationship, the change pattern of the penetration rate is identified and the baseline dynamic penetration rate value is calculated; Based on the baseline dynamic penetration rate value, a penetration rate time relationship matrix is established; wherein the penetration rate time relationship matrix reflects the penetration rate distribution characteristics of different time periods and spatial positions; Input the probe car flow data into the penetration rate time relationship matrix to obtain the query dynamic penetration rate value; Divide the probe car flow data by the corresponding query dynamic penetration rate value to obtain the continuous estimation sequence of the traffic flow of the entire road network; Fuse the continuous estimation sequence with multi-source historical traffic data to generate spatio-temporal fusion features; Based on the spatio-temporal fusion features, output the traffic flow prediction results of the future time period.

[0006] According to the application, a traffic flow intelligent prediction method based on multi-source data fusion is provided, which comprises the following steps: The historical data is dynamically divided in time and space dimensions to generate a time-space data cube; In the time-space data cube, the conditional probability distribution of the probe car flow and the actual traffic flow is calculated to obtain a penetration rate probability model under different time-space situations; Based on the penetration rate probability model, the expected penetration rate value of each period and each node is extracted as the historical proportion relationship of the period and the node.

[0007] According to the application, a traffic flow intelligent prediction method based on multi-source data fusion is provided, which comprises the following steps: Based on the time series analysis method, the historical penetration rate sequence is decomposed in multiple scales to obtain a first period template representing the intra-day variation rule and a second period template representing the intra-week variation rule; The current time is mapped to the first period template and the second period template, and is weighted and fused to generate a reference penetration rate prediction value of the current time; The reference penetration rate prediction value is corrected by coupling the real-time traffic state characteristics to capture the sudden fluctuations other than the periodic rules, and the final reference dynamic penetration rate value is output.

[0008] According to the application, a traffic flow intelligent prediction method based on multi-source data fusion is provided, which comprises the following steps: A two-dimensional matrix is constructed with time periods as rows and node positions as columns; wherein each matrix element is the reference dynamic penetration rate value of the period and the node; The penetration rate of the missing position is estimated to form a complete penetration rate time relationship matrix.

[0009] According to the application, a traffic flow intelligent prediction method based on multi-source data fusion is provided, which comprises the following steps: For the probe car flow data of each node in the whole road network at different time stamps, the corresponding query dynamic penetration rate value is queried from the penetration rate time relationship matrix according to the time stamp and the node position of each node; Each probe car flow data is divided by the corresponding query dynamic penetration rate value to obtain the estimated traffic flow of each node at the moment; By integrating the estimated traffic flow of all nodes at different timestamps, a time-continuous traffic flow estimation sequence covering the whole road network is formed.

[0010] According to the traffic flow intelligent prediction method based on multi-source data fusion provided by the application, the continuous estimation sequence is fused with multi-source historical traffic data to generate spatiotemporal fusion features, including: The continuous estimation sequence is spliced with multi-source auxiliary information from navigation software to form a multi-source feature tensor; wherein the multi-source auxiliary information includes at least one of the floating car average speed, the truck proportion, the positioning flow data and the environmental feature data; The multi-source feature tensor is input into a feature attention dynamic module, and the feature importance is screened through an XGBoost algorithm to obtain a screened feature subset; The self-attention mechanism is applied to the feature subset to calculate the attention weight of each feature in the spatiotemporal dimension; The original feature and the weighted attention feature are connected in residual, and the spatiotemporal fusion feature after dynamic fusion is output through a ReLU activation function.

[0011] According to the traffic flow intelligent prediction method based on multi-source data fusion provided by the application, the self-attention weight in the feature attention dynamic module is calculated by the following formula: ; Wherein, is the attention weight; is a learnable parameter for scaling the attention weight; is an activation function; , , is a learnable parameter representing a weight matrix; is a learnable parameter representing a bias term; is the screened feature; is the multi-source feature tensor.

[0012] According to the traffic flow intelligent prediction method based on multi-source data fusion provided by the application, based on the spatiotemporal fusion feature, the traffic flow prediction result of the future time period is output, including: The spatiotemporal fusion feature is input into a spatiotemporal graph convolution network for spatiotemporal feature extraction; wherein the spatiotemporal graph convolution network includes a graph convolution layer and a time convolution layer, which are used to capture the spatial dependence and temporal dynamics in the traffic network, respectively; The spatiotemporal feature is dynamically weighted through an attention mechanism to enhance the representation of key nodes and key time steps; The weighted spatiotemporal feature is input into a fully connected layer to output the traffic flow prediction value of the future preset time period.

[0013] According to the traffic flow intelligent prediction method based on multi-source data fusion provided by the application, the feature importance screening is performed through the XGBoost algorithm, and after obtaining the screened feature subset, the following steps are further included: Real-time monitoring of sudden traffic event information in the traffic network; the sudden traffic event information includes traffic accidents, road construction or temporary control; When the sudden traffic event information is detected, the initial weight of each feature in the XGBoost algorithm or the bias term of the split gain calculation is dynamically adjusted to improve the priority of the real-time feature related to traffic flow mutation in the feature importance screening process; Among them, the real-time features related to traffic flow mutation at least include: the difference between the current time and the previous time of the probe car flow, the instantaneous drop rate of the probe car average speed, and the user route change request frequency from the navigation software.

[0014] According to the traffic flow intelligent prediction method based on multi-source data fusion provided by the application, the initial weight of each feature in the XGBoost algorithm or the bias term of the split gain calculation is dynamically adjusted, including: According to the type and location of the sudden traffic event, the real-time traffic mutation index of the nodes in the event influence area is calculated; When the real-time traffic mutation index exceeds the preset threshold, the weight adjustment mechanism is triggered; In the weight adjustment mechanism, higher initial weight is given to the real-time features related to traffic mutation, and positive bias is applied to the split gain of these features when the decision tree node is split.

[0015] The traffic flow intelligent prediction method based on multi-source data fusion provided by the application, by obtaining the historical time series data of multiple nodes in the traffic network and calculating the historical proportional relationship between the probe car flow and the actual traffic flow, identifying the penetration rate change mode and establishing the penetration rate time relationship matrix, then combining the probe car flow data with the matrix to obtain the continuous estimation sequence of the whole network traffic flow, and then fusing with the multi-source historical traffic data to generate spatio-temporal fusion features, finally outputting the future traffic flow prediction result based on the spatio-temporal fusion features, realizes the technical effects of significantly improving the traffic flow prediction accuracy, enhancing the model generalization ability and improving the prediction real-time and continuity, effectively solves the technical problems of large prediction deviation and difficulty in adapting to complex scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0017] Figure 1 is a flowchart of a traffic flow intelligent prediction method based on multi-source data fusion provided by an embodiment of the present application; Figure 2 is a structural schematic diagram of a traffic flow intelligent prediction device based on multi-source data fusion provided by an embodiment of the present application; Figure 3 is a structural schematic diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0018] In order to make the objectives, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below with reference to the drawings in the present application. Obviously, the described embodiments are some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0019] Figure 1 is a flowchart of a traffic flow intelligent prediction method based on multi-source data fusion provided by an embodiment of the present application.

[0020] Referring to Figure 1 , the traffic flow intelligent prediction method based on multi-source data fusion includes the following steps 101 to 108.

[0021] Step 101, historical time series data of multiple nodes in a traffic network is acquired.

[0022] In this step, the nodes generally include road intersections, road sections, etc. The historical time series data refers to traffic flow, speed, density data and truck proportion collected from sensors, cameras or floating car devices, and the historical time series data can also be preprocessed by using a time stamp alignment and missing value filling method.

[0023] Step 102, the historical proportional relationship between floating car flow and actual traffic flow in different time periods and different nodes is calculated according to the historical time series data.

[0024] In this step, the floating car flow refers to the flow data collected by floating cars (such as vehicles equipped with navigation devices), and the actual traffic flow is measured by fixed detectors (such as radars, cameras). The historical proportional relationship refers to the ratio between the floating car flow and the actual traffic flow, which is generally different at different times and places.

[0025] Step 103, based on the historical proportional relationship, the change mode of the penetration rate is identified and the reference dynamic penetration rate value is calculated.

[0026] In this step, the permeability changes over time and space, for example, there are big differences in the morning and evening peak hours, or on weekdays and weekends. By analyzing historical data, identify the change pattern of permeability (such as daily cycle, weekly cycle), and calculate the reference dynamic permeability value at the current time. For example, through analysis, it is found that the permeability is low (such as 0.15) in the morning peak (7:00-9:00) of weekdays, and the permeability is high (such as 0.25) in the evening peak (17:00-19:00). The reference dynamic permeability value at the current time is 8:30 on weekdays, which may be 0.15. That is, the reference dynamic permeability value at the current time or in the future can be obtained according to the historical proportion relationship. It can also be calculated by kernel density estimation and neural network residual correction method to eliminate the systematic deviation caused by fixed proportion coefficient.

[0027] Step 104, based on the reference dynamic permeability value, a permeability time relationship matrix is established; wherein the permeability time relationship matrix reflects the permeability distribution characteristics of different time periods and spatial positions.

[0028] In this step, the permeability time relationship matrix refers to a two-dimensional data structure with time periods as rows and node positions as columns, which can be completed by cubic spline interpolation or random forest algorithm to store the permeability distribution characteristics in time and space dimensions.

[0029] Step 105, input the probe car flow data into the permeability time relationship matrix to obtain the query dynamic permeability value.

[0030] In this step, for the probe car flow data at the current time, the corresponding query dynamic permeability value is queried from the permeability time relationship matrix according to its time and node position.

[0031] Step 106, divide the probe car flow data by the corresponding query dynamic permeability value to obtain the continuous estimation sequence of the traffic flow of the whole road network.

[0032] In this step, the continuous estimation sequence refers to the traffic flow data generated by the element-by-element division operation of the probe car flow and the query dynamic permeability value, which can be smoothed by the sliding window method to construct a spatiotemporal continuous dataset covering the whole road network.

[0033] Step 107, fuse the continuous estimation sequence with multi-source historical traffic data to generate spatiotemporal fusion features.

[0034] In this step, the spatiotemporal fusion features refer to the multi-dimensional feature vector formed by splicing the continuous estimation sequence with multi-source auxiliary information, which can be dynamically weighted by attention mechanism and residual connection to enhance the model's ability to capture traffic sudden events.

[0035] Step 108, output the traffic flow prediction result of the future time period based on the spatio-temporal fusion feature.

[0036] In the embodiment, by obtaining the historical time series data of multiple nodes in the traffic network and calculating the historical proportional relationship between the probe car flow and the actual traffic flow, the penetration rate change pattern is identified and the penetration rate time relationship matrix is established, and then the probe car flow data is combined with the matrix to obtain the continuous estimation sequence of the traffic flow of the whole network, and the spatio-temporal fusion feature is generated by fusing the multi-source historical traffic data, and finally the future traffic flow prediction result is output based on the spatio-temporal fusion feature, which realizes the technical effects of significantly improving the traffic flow prediction accuracy, enhancing the model generalization ability and improving the prediction real-time and continuity, and effectively solves the technical problems of large prediction deviation and difficulty in adapting to complex scenes.

[0037] In an embodiment of the present specification, the historical proportional relationship between the probe car flow and the actual traffic flow at different time periods and different nodes is calculated according to the historical time series data, including: Step one, dynamically divide the historical data in the space-time dimension to generate a space-time data cube; In this step, the space-time data cube refers to a three-dimensional data structure formed by dynamically dividing the historical data according to the time window and the geographical location, which can be realized by combining the sliding time window and the geographical grid division, and is used to solve the data fragmentation problem caused by static division. For example, in the three-dimensional data structure, one coordinate axis can represent the spatial position, one coordinate axis can represent the time window, and one coordinate axis can represent the traffic parameter.

[0038] Step two, in the space-time data cube, the conditional probability distribution of the probe car flow and the actual traffic flow is calculated by using the kernel density estimation method to obtain the penetration rate probability model under different spatio-temporal situations; In this step, the kernel density estimation method refers to a probability density estimation technique based on non-parametric statistics, which can specifically use a Gaussian kernel function to smooth estimate the joint distribution of the probe car flow and the actual traffic flow, and is used to adapt to the sparsity and non-uniformity of the traffic data in the space-time dimension. The conditional probability distribution refers to the joint probability relationship of the probe car flow and the actual traffic flow under a given spatio-temporal position, which can be calculated by integrating the probability density function output by the kernel density estimation, and is used to quantify the correlation strength of the probe car flow and the actual traffic flow in different spatio-temporal situations.

[0039] Step three, based on the penetration rate probability model, the expected penetration rate value of each time period and each node is extracted as the historical proportional relationship of the time period and the node; In this step, the expected permeability value refers to the mathematical expectation of the permeability parameter in the conditional probability distribution, which can be calculated by integrating the permeability probability model or Monte Carlo sampling, and is used to represent the stable proportional relationship between the probe vehicle flow and the actual traffic flow in a specific space-time unit.

[0040] Specifically, first, the original historical traffic data is dynamically divided according to the preset time granularity and spatial grid to form a data cube structure containing time, space and flow dimensions. In each space-time unit of the cube, the joint distribution of probe vehicle flow and actual traffic flow is modeled using a kernel density estimation method to obtain a probability density surface reflecting the dynamic relationship between the two. By marginalizing the probability model, the conditional probability distribution of the permeability in each space-time unit is calculated. Finally, by calculating the expected value, the stable statistics of the permeability in each space-time unit are extracted to form a historical proportion relationship matrix covering the entire road network.

[0041] In this embodiment, by combining dynamic space-time division and kernel density estimation, the space-time heterogeneity of data distribution can be adaptively processed, especially in low-permeability areas or traffic mutation periods, the nonlinear relationship between probe vehicle flow and actual flow can be accurately reflected through the probability model. This embodiment effectively solves the permeability estimation deviation problem caused by rigid space-time division, and significantly improves the accuracy and robustness of historical proportion relationship modeling.

[0042] In an embodiment of the present application, based on the historical proportion relationship, the change pattern of the permeability is identified and the reference dynamic permeability value is calculated, including: Step one, based on time series analysis method, the historical permeability sequence is decomposed into first period template (such as daily period template) representing the daily variation law and second period template (such as weekly period template) representing the weekly variation law; In this step, multi-scale periodic decomposition refers to decomposing the original permeability sequence into periodic components of different time scales using time series decomposition algorithm, which can be realized by using seasonal trend decomposition or Fourier transform method, and is used to separate the variation law of different time dimensions such as daily period and weekly period. The period template refers to the standardized period pattern formed by historical data analysis, which can be realized by pattern extraction of historical period components through clustering algorithm, and is used to represent the typical permeability fluctuation characteristics at different time scales.

[0043] Step two, map the current time to the first period template and the second period template, and perform weighted fusion to generate the reference permeability prediction value at the current time; wherein the weight is dynamically configured according to whether the current date is a weekday or a holiday; In this step, the weighted fusion refers to dynamically combining the prediction results of multiple periodic templates according to the date type, which can be realized by using linear weighting or adaptive weight allocation algorithm, for adapting to different traffic patterns on weekdays and holidays.

[0044] Step three, coupling real-time traffic state features, a light neural network model is used to correct the residual of the benchmark permeability prediction value to capture the sudden fluctuations (also known as short-term fluctuations) other than the periodic regularity, and output the final benchmark dynamic permeability value; In this step, the residual correction refers to compensating the deviation of the periodic prediction result by using real-time data, which can be realized by constructing a neural network model containing input features such as real-time speed and sudden traffic events, for capturing abnormal fluctuations of permeability caused by sudden factors such as traffic accidents. Short-term fluctuations refer to sudden or temporary changes in traffic flow in addition to general regularity (such as daily cycle, weekly cycle). Short-term fluctuations are usually caused by sudden traffic events (such as traffic accidents, road construction, temporary restrictions, etc.) or random factors (such as weather changes, special activities, etc.). Such changes can have a significant impact on traffic flow in a short period of time, and existing prediction methods based on fixed patterns often fail to capture these sudden changes. The light neural network model can include multi-layer perceptron, one-dimensional convolutional neural network, gated recurrent unit, etc.

[0045] Specifically, first, the historical permeability data is decomposed into time series, the components with daily and weekly periodicity are extracted, and a standardized template library is constructed. When real-time prediction is performed, the benchmark value is matched in the daily and weekly cycle templates according to the time position of the current time, and the fusion weights of the two templates are dynamically allocated according to the date type. For example, in the weekday scenario, the weight of the weekly cycle template can be set to 0.3, and the weight of the daily cycle template can be set to 0.7. Then, the real-time collected traffic state parameters are input into the pre-trained neural network model, which dynamically adjusts the preliminary fused benchmark value by learning the residual pattern outside the periodic regularity in the historical data. The final output dynamic permeability value not only retains the periodicity of the regularity, but also incorporates the influence of real-time traffic state.

[0046] In this embodiment, through multi-scale decomposition and dynamic weighting mechanism, the periodic regularity of different time dimensions can be captured simultaneously, and adaptive adjustment can be made based on the date type. In addition, by introducing a light neural network for residual correction, the problem of prediction lag in sudden traffic event scenarios is effectively solved. This embodiment can improve the time resolution and environmental adaptability of the permeability estimation value. By fusing the periodic features of multiple time scales, the expression ability of the periodic fluctuation of traffic flow is enhanced.

[0047] In an embodiment of the present specification, based on the benchmark dynamic permeability value, a permeability time relationship matrix is established, including: Step one, construct a two-dimensional matrix with time period as row and node position as column; wherein each matrix element is the benchmark dynamic permeability value of the node in the period; Step two, estimate the permeability of the missing position using interpolation or machine learning method to form a complete permeability time relationship matrix; In this step, the interpolation or machine learning method refers to the data completion technology, which can be realized by using cubic spline interpolation method or random forest regression model, and the missing values are reasonably inferred by using the permeability data characteristics of adjacent space-time units to ensure the space-time continuity of the matrix.

[0048] Specifically, when constructing the permeability time relationship matrix, first divide the time axis into multiple continuous time period units, for example, divide the daily cycle time window with 15 minutes interval. At the same time, the nodes in the traffic network are grid encoded according to the geographical position to form the spatial dimension coordinates. For each space-time cell, if there is historical benchmark permeability data, it is directly filled, and for the data missing cell, the cubic spline interpolation can be used based on the permeability change trend of adjacent time period to complete the data, or the random forest model is trained to predict the missing value by using the permeability characteristics of the surrounding nodes in the same period. For example, when the node data of a road intersection is missing in the morning peak period, the permeability values of the previous and next periods of the node can be used for linear interpolation, or the permeability average of the adjacent three intersections in the period is extracted for filling.

[0049] In this embodiment, by constructing the dynamic permeability matrix and using intelligent data completion technology, the permeability fluctuation characteristics of different regions in different periods can be accurately described, especially when there is a blind area in data collection, the space-time correlation law is effectively mined through machine learning model, which significantly improves the integrity and reliability of permeability estimation.

[0050] In an embodiment of the present specification, the probe vehicle flow data is divided by the corresponding query dynamic permeability value to obtain a continuous estimation sequence of the whole road network traffic flow, including: Step one, for the probe vehicle flow data of each node in the whole road network at different time stamps, according to the time stamp and node position of each node, the corresponding query dynamic permeability value is queried from the permeability time relationship matrix; In this step, the query dynamic permeability value refers to the real-time permeability parameter extracted from the matrix according to the current time stamp and node position, which can be realized by matrix index matching or adjacent period data interpolation, and is used to eliminate the deviation caused by incomplete coverage of probe vehicle data.

[0051] Step two, divide each probe car flow data by its corresponding query dynamic penetration rate value to obtain the estimated traffic flow of each node at the moment; Step three, integrate the estimated traffic flow of all nodes at different timestamps to form a time-continuous traffic flow estimation sequence covering the whole road network; In this step, the continuous estimation sequence refers to the flow data stream covering the whole road network generated by calculating and integrating node by node and timestamp by timestamp. It can be realized by time sliding window or spatial grid aggregation method, and is used to build uninterrupted traffic state monitoring basic data.

[0052] Specifically, by dynamically matching the penetration rate parameter in time and space dimensions, the locally collected probe car data is converted into global traffic flow estimation value. For example, the probe car flow of a certain intersection during the morning peak period is 200 vehicles per minute, and if the penetration rate query value of the node during this period is 0.25, the actual traffic flow estimation value is 800 vehicles per minute. The calculation results of all nodes at different timestamps are time-aligned and spatially connected, and finally a flow sequence containing continuous time changes of each position in the whole road network is formed.

[0053] In this embodiment, by constructing the spatiotemporal associated penetration rate matrix, the dynamic characteristics of different nodes at different time periods can be accurately matched, and the local data calculation error can be reduced. The application effectively solves the traffic flow estimation fault caused by incomplete coverage of probe car data, and realizes continuous monitoring of the whole road network traffic state. For example, in the scenario where the probe car data collection interval is 5 minutes, this method can generate a continuous flow sequence with a time resolution of 1 minute through dynamic query and real-time calculation of the penetration rate matrix, providing high-precision input data for subsequent prediction models. At the same time, through node-level calculation in the spatial dimension, this method avoids the spatial detail loss caused by traditional regional aggregation methods, so that the traffic fluctuations of key nodes such as interchanges and ramps can be accurately captured.

[0054] In an embodiment of the present application, the continuous estimation sequence is fused with multi-source historical traffic data to generate spatiotemporal fusion features, including: Step one, splice the continuous estimation sequence with multi-source auxiliary information from navigation software to form a multi-source feature tensor; wherein the multi-source auxiliary information includes at least one of probe car average speed, truck proportion, positioning flow data, and environmental feature data (such as temperature, wind power, rainfall, etc.); In this step, the multi-source auxiliary information refers to the real-time dynamic parameters related to traffic flow extracted from navigation software, which can be realized by using floating car trajectory data, vehicle type identification data, satellite positioning data and meteorological sensor data to supplement the real-time traffic state details not contained in the continuous estimation sequence. Splicing is a data fusion operation, which refers to directly connecting multiple feature vectors in the dimension direction to form a new feature vector that is longer and contains all the information.

[0055] Step two, input the multi-source feature tensor into the feature attention dynamic module, and select the feature subset after screening through the XGBoost algorithm; In this step, the feature attention dynamic module refers to a feature screening and enhancement module based on machine learning and attention mechanism, which can be realized by using the XGBoost algorithm combined with the self-attention mechanism, and is used to screen key information from high-dimensional features and strengthen the spatio-temporal correlation. Specifically, as shown in the following formula (1): (1); Among them, is a multi-source feature tensor; is a feature subset after screening; is the corresponding XGBoost algorithm, is the dimension of the feature.

[0056] Step three, apply the self-attention mechanism to the feature subset, and calculate the attention weight of each feature in the space-time dimension; In this step, the self-attention mechanism refers to a method of dynamically allocating weights by calculating the correlation between features, which can be realized by calculating the attention score after linear transformation of the feature by using a learnable parameter matrix, and is used to capture the importance difference of the feature in different space-time dimensions.

[0057] Step four, residual connection is performed on the original feature and the weighted attention feature, and the spatio-temporal fusion feature after dynamic fusion is output through the ReLU activation function. Specifically, as shown in the following formula (2): (2); Among them, is an attention weight matrix, represents a feature weighting operation, represents residual connection, is an activation function, is a spatio-temporal fusion feature, represents the meaning of dynamic.

[0058] Specifically, the multi-source feature tensor is constructed by concatenating the continuous estimation sequence and the real-time auxiliary data provided by the navigation software, such as adding the average speed of the floating car and the proportion of the truck as the supplementary dimension to the tensor. Then, the XGBoost algorithm sorts the multi-source features according to the feature split gain, and selects a feature subset that is strongly related to the traffic flow prediction. In the self-attention calculation stage, the feature subset is transformed into a query vector and a key vector through a weight matrix, and after calculating the attention weight of each feature, the residual connection is performed with the original feature, and finally the fused spatio-temporal feature is output through a nonlinear activation function. Through dynamic feature screening and attention weighting, this process effectively eliminates the interference of redundant information while retaining key signals in different spatio-temporal dimensions.

[0059] In this embodiment, through the dual screening mechanism of XGBoost and self-attention mechanism, the key influencing factors can be dynamically identified and strengthened in the feature dimension, such as automatically increasing the weight of the floating car speed mutation feature in the traffic congestion scenario, thereby enhancing the model's ability to represent complex traffic states. The present application solves the problem of feature redundancy and key information loss in the multi-source data fusion process, and through the dynamic feature screening and attention weighting mechanism, effectively improves the discriminability of the spatio-temporal fusion feature, so that the subsequent prediction model can more accurately capture the spatio-temporal pattern of traffic flow changes, especially significantly reduces the prediction bias in the traffic state mutation scenario.

[0060] In an embodiment of the present application, the self-attention weight in the feature attention dynamic module is calculated by the following formula (3): (3); wherein, is the attention weight; is a learnable parameter for scaling the attention weight; is an activation function; , , is a learnable parameter representing a weight matrix; is a learnable parameter representing a bias term; is the selected feature; is the multi-source feature tensor.

[0061] In this embodiment, the attention weight refers to a parameter for dynamically assigning importance between features by calculating the correlation between features, which can be specifically implemented by linear transformation of a learnable parameter matrix and a feature tensor combined with an activation function, for capturing the contribution difference of different features to traffic flow prediction in the space-time dimension. The learnable parameter refers to a matrix variable that is automatically optimized through model training, which can be specifically implemented by updating the parameter value through the back propagation algorithm after random initialization, for dynamically adjusting the strength of feature interaction according to the data distribution. The screened feature refers to a feature subset screened by feature importance, which can be specifically implemented by retaining high importance features after calculating feature split gain by the XGBoost algorithm, for reducing the interference of redundant information on attention calculation.

[0062] In this embodiment, the dynamic attention mechanism is constructed by a learnable parameter matrix and an activation function, which can automatically adjust the interaction mode between features according to real-time data, effectively capturing the nonlinear correlation of multi-source features in the space-time dimension. The application can improve the refinement degree of multi-source traffic data fusion, so that the model can still accurately identify key influencing factors under complex traffic conditions. Through the dynamic attention weight distribution mechanism, the capture ability of features related to sudden traffic events is enhanced.

[0063] In an embodiment of the present application, based on the spatio-temporal fusion feature, a traffic flow prediction result of a future time period is output, including: Step one, input the spatio-temporal fusion feature into the spatio-temporal graph convolution network for spatio-temporal feature extraction; wherein the spatio-temporal graph convolution network includes a graph convolution layer and a time convolution layer, for capturing the spatial dependence and temporal dynamics in the traffic network, respectively; In this step, the spatio-temporal graph convolution network refers to a deep learning architecture combining graph convolution operation and time convolution operation, which can be specifically implemented by extracting the topological connection relationship of the road network by the graph convolution layer and extracting the periodic trend of the traffic flow sequence by the time convolution layer, and its role is to model the correlation of traffic flow in the space dimension and the evolution law in the time dimension. The graph convolution layer is shown in the following formula (4): (4); wherein, is the result of the graph convolution layer, is a parameterized filter function, is the parameter of the filter; denotes the graph convolution operation applied to the Laplacian matrix of the graph ; K is the order of the Chebyshev polynomial, used for the order of approximation; is the coefficient in the Chebyshev polynomial approximation, for each k, the range is from 0 to K-1; is the input feature signal, is a scaled graph Laplacian matrix, is a scaled graph Laplacian matrix, is the k-th order Chebyshev polynomial acting on the matrix .

[0064] The graph convolution layer adopts a first-order approximation based on Chebyshev polynomials to implement an approximate graph convolution operation to efficiently aggregate spatial information, the principle of which is to approximate the frequency domain filter, as shown in the following formula (5): (5); wherein, is an eigenvalue matrix of the graph Laplacian matrix; is a scaled eigenvalue matrix; is the k-th order Chebyshev polynomial.

[0065] Step two, dynamically weight the spatio-temporal features through the attention mechanism to enhance the representation of key nodes and key time steps; In this step, the action is to strengthen the feature representation of the traffic mutation area or peak period, and suppress the interference of non-key information. Key nodes can include traffic hubs, bridges, tunnels, schools, and the center of sudden traffic events.

[0066] Step three, input the weighted spatio-temporal features into the fully connected layer to output the traffic flow prediction value of the future preset time period; In this step, the fully connected layer refers to the linear transformation structure for feature mapping in the neural network, which can be realized by mapping high-dimensional spatio-temporal features to prediction value sequences through a multi-layer perceptron. Its role is to convert abstract features into interpretable traffic flow numerical output.

[0067] Step four, use mean squared error or mean absolute error as the loss function to optimize the network parameters through back propagation.

[0068] Specifically, the spatio-temporal graph convolution network models the adjacency relationship between road nodes through the graph convolution layer, for example, constructs an adjacency matrix using the actual connection distance between nodes to capture the spatial correlation of congestion propagation. The time convolution layer adopts a dilated causal convolution structure, for example, sets the convolution kernel size to 3 and the dilation coefficient to 2 to capture multi-scale time patterns such as hourly and daily. The dynamic weighting module generates attention weights by calculating the cosine similarity between node feature vectors, for example, assigns higher weight values to nodes in sudden congestion areas than to nodes in regular areas. The final prediction layer maps the weighted feature vectors to the traffic flow prediction values of each node in the next hour, and iteratively optimizes the model parameters through the mean squared error loss function.

[0069] In this embodiment, the joint modeling of spatial dependence and temporal dynamics is achieved through the cascaded structure of the graph convolution layer and the temporal convolution layer. The adaptive attention mechanism introduced in this embodiment can dynamically adjust the weight distribution according to real-time features, significantly improving the response sensitivity to traffic anomaly events. This embodiment solves the problem of limited prediction accuracy caused by insufficient modeling of the spatio-temporal correlation of multi-source traffic data, and effectively improves the accuracy of long-term traffic flow prediction in complex road network environments through the composite network structure of graph convolution and temporal convolution.

[0070] In an embodiment of the present specification, after the feature importance screening is performed by the XGBoost algorithm, the screened feature subset is obtained, and further includes: Step one, real-time monitoring of sudden traffic event information in the traffic network, the sudden traffic event information including traffic accidents, road construction or temporary regulation; Step two, when detecting the sudden traffic event information, dynamically adjusting the initial weight of each feature in the XGBoost algorithm or the bias term of the split gain calculation to improve the priority of real-time features strongly related to traffic flow mutation in the feature importance screening process; Among them, the real-time features strongly related to traffic flow mutation at least include: the difference between the current time and the previous time floating car flow, the instantaneous drop rate of floating car average speed, and the user route change request frequency from the navigation software.

[0071] In this embodiment, the sudden traffic event information refers to the abnormal traffic state data obtained in real time through traffic monitoring equipment or third-party data interface, which can be realized by video recognition technology or API interface transmission method, and is used to trigger the model parameter dynamic adjustment mechanism. The real-time traffic mutation index refers to the quantitative index calculated by the deviation of the floating car flow rate from the historical baseline data, which can be realized by the sliding window statistical method, and is used to evaluate the propagation range and influence intensity of the sudden traffic event. The weight adjustment mechanism refers to the operation logic of dynamically modifying the feature weight in the XGBoost algorithm according to the real-time traffic mutation index, which can be realized by the pre-set event response rule or online learning module, and is used to enhance the sensitivity of the model to sudden traffic events. The positive bias refers to the artificial incremental adjustment of the gain value of a specific feature in the decision tree splitting calculation process, which can be realized by modifying the offset term in the split gain calculation formula, and is used to guide the model to preferentially select features related to traffic mutation.

[0072] Specifically, when the traffic monitoring system detects a traffic accident on a certain section of road, the floating car flow difference in the event impact area will increase sharply, at which time the system automatically calculates the real-time flow mutation index of the area. If the index exceeds the preset threshold, for example, reaches 3 times the standard deviation of the historical average during peak hours, the weight adjustment mechanism is triggered. Under this mechanism, the initial weight of the floating car flow difference in the XGBoost algorithm is increased, and at the same time, a positive bias value is added to the split gain calculation of this feature when building the decision tree. This adjustment allows the model to prioritize retaining feature combinations that contain flow mutation signals during the feature selection stage, such as including the combination of floating car flow difference and user route change request frequency in the final feature subset.

[0073] In some embodiments, the impact range of road construction events can be defined by geographic fence technology, and the system only adjusts the weights of nodes within 500 meters upstream of the construction area. For temporary regulation events, the degree of overlap between the regulation start time and the evening peak traffic flow can be combined to dynamically set different bias adjustment magnitudes. For example, when the regulation occurs during the evening peak period, the positive bias value can be set to 1.5 times that of the non-peak period.

[0074] In this embodiment, by establishing an event-driven dynamic adjustment mechanism, the feature selection process can adapt to traffic state mutations, effectively solving the problem of mutation signals being overwhelmed by regular features. This embodiment can quickly capture key mutation features when a sudden traffic event occurs, avoiding important signals from being filtered out during the feature selection stage, ensuring the accuracy and timeliness of traffic flow prediction results in sudden traffic event scenarios.

[0075] In an embodiment of the present specification, the bias term of the initial weight or split gain calculation of each feature in the XGBoost algorithm is dynamically adjusted, including: Step one, according to the type and location of the sudden traffic event, calculate the real-time flow mutation index of the nodes in the event impact area; In this step, the real-time flow mutation index refers to a quantitative indicator constructed by parameters such as floating car flow difference, instantaneous drop rate of floating car average speed, and navigation software user route change request frequency. Specifically, the feature change rate between the current time and the previous time period can be calculated using a sliding window statistical method, and this index is used to objectively evaluate the impact of sudden traffic events on the local road network.

[0076] Step two, when the real-time flow mutation index exceeds the preset threshold, trigger the weight adjustment mechanism; In this step, the weight adjustment mechanism refers to the control logic that dynamically changes the feature selection strategy according to the real-time traffic state. Specifically, it is realized by modifying the feature gain calculation function of the XGBoost algorithm, which allows the model to quickly respond to abnormal state changes in the road network.

[0077] Step 3: In the weight adjustment mechanism, higher initial weights are assigned to real-time features related to traffic mutations, and a positive bias is applied to the split gain of these features when the decision tree node splits, so that the XGBoost model is more inclined to retain and amplify these mutation signal features during the feature selection stage.

[0078] In this step, positive bias refers to the operation of numerically enhancing the split gain of a specific feature during the decision tree splitting process. Specifically, it can be achieved by superimposing a preset gain correction amount when calculating the information gain. This operation can improve the priority of key mutation features in model decision-making.

[0079] Specifically, when a traffic accident or road construction event is detected, the system first locates the affected road network nodes and calculates the traffic fluctuation amplitude of each node in the area in real time based on floating car data streams. If the traffic mutation index of a node exceeds a preset warning value, the feature weight adjustment module is activated. During this process, features directly related to traffic mutations are assigned higher initial weights; for example, the initial weight coefficient of the floating car traffic difference feature can be increased from the default value of 0.5 to 0.8. When constructing the decision tree, the algorithm automatically adds a gain correction of 0.2 when calculating the information gain of these features, making it easier to prioritize split nodes containing these features. Through this dynamic adjustment mechanism, the model can effectively capture short-term traffic fluctuation patterns caused by traffic mutation events, avoiding the weakening or omission of important features during conventional screening.

[0080] In this embodiment, by establishing an event-driven dynamic adjustment mechanism, key abrupt change features can be accurately identified based on real-time traffic conditions. The influence of these features is strengthened during model training, thereby significantly improving the prediction model's response speed to sudden traffic conditions. This application effectively solves the problem of insufficient sensitivity in feature selection in existing traffic prediction systems during sudden traffic events. By dynamically adjusting feature weights and gain calculation strategies, it ensures that the model can promptly capture traffic flow abrupt change signals, thereby improving the accuracy and reliability of short-term traffic flow prediction.

[0081] Table 1 below lists the average prediction results over the next 15, 30, and 60 minutes compared to various baseline models. In short-term predictions (15 minutes), STIE-GCN showed improvements in both MAE and MAPE compared to state-of-the-art baseline models. In long-term predictions (60 minutes), improvements were observed in mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE).

[0082] Table 1. Performance index prediction results of different prediction models at different time scales.

[0083] Based on the same overall inventive concept, the present application also protects an intelligent traffic flow prediction device based on multi-source data fusion, as shown in Figure 2 Figure 2 is a structural schematic diagram of the intelligent traffic flow prediction device based on multi-source data fusion provided by the embodiments of the present application. The intelligent traffic flow prediction device based on multi-source data fusion provided by the present application is described below, and the intelligent traffic flow prediction device based on multi-source data fusion described below can be mutually corresponding to the intelligent traffic flow prediction method based on multi-source data fusion described above.

[0084] The intelligent traffic flow prediction device based on multi-source data fusion comprises: The acquisition module 201 acquires historical time series data of a plurality of nodes in a traffic network; The historical proportion module 202 calculates historical proportional relationships of probe car flow and actual traffic flow in different time periods and different nodes according to the historical time series data; The reference dynamic module 203 identifies a change mode of the penetration rate and calculates a reference dynamic penetration rate value based on the historical proportional relationships; The matrix module 204 establishes a penetration rate time relationship matrix based on the reference dynamic penetration rate value; wherein the penetration rate time relationship matrix reflects the distribution characteristics of the penetration rate in different time periods and spatial positions; The query dynamic module 205 inputs the probe car flow data into the penetration rate time relationship matrix to obtain a query dynamic penetration rate value; The sequence module 206 divides the probe car flow data by the corresponding query dynamic penetration rate value to obtain a continuous estimation sequence of the traffic flow of the whole network; The fusion module 207 fuses the continuous estimation sequence with multi-source historical traffic data to generate spatio-temporal fusion features; The prediction module 208 outputs a traffic flow prediction result of a future time period based on the spatio-temporal fusion features.

[0085] Figure 3 is a structural schematic diagram of an electronic device provided by the embodiments of the present application.

[0086] As shown in Figure 3 , the electronic device can comprise a processor 310, a communications interface 320, a memory 330 and a communications bus 340, wherein the processor 310, the communications interface 320 and the memory 330 complete mutual communication through the communications bus 340. The processor 310 can invoke logical instructions in the memory 330 to execute the intelligent traffic flow prediction method based on multi-source data fusion. ​

[0087] In addition, the logical instructions in the memory 330 described above can be implemented in the form of a software function unit and sold or used as an independent product, and can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application or the part of the prior art that contributes essentially or the part of the technical solutions can be embodied in the form of a software product, which is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in various embodiments of the present application. The aforementioned storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0088] In another aspect, the present application also provides a computer program product, which comprises a computer program, the computer program can be stored on a non-transitory computer readable storage medium, and the computer program can be executed by a processor to enable a computer to execute the intelligent traffic flow prediction method based on multi-source data fusion provided by the above-mentioned methods.

[0089] In yet another aspect, the present application also provides a non-transitory computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the intelligent traffic flow prediction method based on multi-source data fusion provided by the above-mentioned methods.

[0090] The device embodiments described above are only schematic, wherein the units illustrated as separate components can or can not be physically separated, and the components illustrated as units can or can not be physical units, i.e., they can be located in one place, or distributed on multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the present embodiment. Those skilled in the art can understand and implement it without creative labor.

[0091] Those skilled in the art can clearly understand the technical solutions of the various embodiments from the above description of the embodiments, and the various embodiments can be implemented by means of software with the necessary general hardware platforms, and of course, can also be implemented by hardware. Based on such understanding, the above technical solutions, essentially or in other words, the part of the prior art that makes a contribution, can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, and the like, and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0092] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for some technical features therein; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A traffic flow intelligent prediction method based on multi-source data fusion, characterized in that, The method comprises the following steps: obtaining historical time series data of multiple nodes in a traffic network; calculating historical proportional relationships between probe car flow and actual traffic flow in different time periods and at different nodes according to the historical time series data; identifying a change pattern of the penetration rate and calculating a reference dynamic penetration rate value based on the historical proportional relationships; establishing a penetration rate time relationship matrix based on the reference dynamic penetration rate value; wherein the penetration rate time relationship matrix reflects the distribution characteristics of the penetration rate in different time periods and spatial locations; inputting the probe car flow data into the penetration rate time relationship matrix to obtain a query dynamic penetration rate value; dividing the probe car flow data by the corresponding query dynamic penetration rate value to obtain a continuous estimation sequence of the traffic flow of the whole network; fusing the continuous estimation sequence with multi-source historical traffic data to generate spatio-temporal fusion features; outputting a traffic flow prediction result of a future time period based on the spatio-temporal fusion features. 2.The intelligent traffic flow prediction method based on multi-source data fusion according to claim 1, characterized in that, The method of calculating the historical proportional relationships between the probe car flow and the actual traffic flow in different time periods and at different nodes according to the historical time series data comprises the following steps: dynamically dividing the historical data in the time and space dimensions to generate a spatio-temporal data cube; calculating the conditional probability distribution of the probe car flow and the actual traffic flow in the spatio-temporal data cube to obtain a penetration rate probability model under different spatio-temporal situations; extracting the expected penetration rate value of each time period and each node based on the penetration rate probability model, which is used as the historical proportional relationship of the time period and the node. 3.The intelligent traffic flow prediction method based on multi-source data fusion of claim 2, wherein, The method of identifying the change pattern of the penetration rate and calculating the reference dynamic penetration rate value based on the historical proportional relationships comprises the following steps: performing multi-scale periodic decomposition on the historical penetration rate sequence based on a time series analysis method to obtain a first periodic template representing the intra-day variation rule and a second periodic template representing the intra-week variation rule; mapping the current time to the first periodic template and the second periodic template, and performing weighted fusion to generate a reference penetration rate prediction value of the current time; coupling real-time traffic state features to perform residual correction on the reference penetration rate prediction value to capture sudden fluctuations other than periodic rules, and output the final reference dynamic penetration rate value. 4.The intelligent traffic flow prediction method based on multi-source data fusion of claim 1, wherein, The method of establishing the penetration rate time relationship matrix based on the reference dynamic penetration rate value comprises the following steps: constructing a two-dimensional matrix with time periods as rows and node locations as columns; wherein each matrix element is the reference dynamic penetration rate value of the time period and the node; estimating the penetration rate of the missing positions to form a complete penetration rate time relationship matrix. 5.The intelligent traffic flow prediction method based on multi-source data fusion according to claim 1, wherein, The method of dividing the probe car flow data by the corresponding query dynamic penetration rate value to obtain a continuous estimation sequence of the traffic flow of the whole network comprises the following steps: for the probe car flow data of each node in the whole network at different time stamps, querying the corresponding query dynamic penetration rate value from the penetration rate time relationship matrix according to the time stamps and node locations thereof; dividing each probe car flow data by the corresponding query dynamic penetration rate value to obtain the estimated traffic flow of each node at the time combination; integrating the estimated traffic flow of all nodes at different time stamps to form a time-continuous traffic flow estimation sequence covering the whole network. 6.The intelligent traffic flow prediction method based on multi-source data fusion according to claim 1, wherein, The continuous estimation sequence is fused with multi-source historical traffic data to generate spatiotemporal fusion features, including: The continuous estimation sequence is spliced with multi-source auxiliary information from navigation software to form a multi-source feature tensor, wherein the multi-source auxiliary information includes at least one of floating car average speed, truck proportion, positioning flow data, and environmental feature data; The multi-source feature tensor is input into a feature attention dynamic module to perform feature importance screening through an XGBoost algorithm to obtain a screened feature subset; A self-attention mechanism is applied to the feature subset to calculate the attention weight of each feature in the spatiotemporal dimension; The original features are connected in residual with the weighted attention features, and the spatiotemporal fusion features after dynamic fusion are output through a ReLU activation function.

7. The intelligent traffic flow prediction method based on multi-source data fusion according to claim 6, characterized in that, The self-attention weight in the feature attention dynamic module is calculated by the following formula: ； wherein, is an attention weight; is a learnable parameter for scaling the attention weight; is an activation function; , , is a learnable parameter representing a weight matrix; is a learnable parameter representing a bias term; is a filtered feature; is a multi-source feature tensor. 8.The intelligent traffic flow prediction method based on multi-source data fusion of claim 1, wherein, Based on the spatiotemporal fusion features, the traffic flow prediction results in the future time period are output, including: The spatiotemporal fusion features are input into a spatiotemporal graph convolution network for spatiotemporal feature extraction; wherein the spatiotemporal graph convolution network includes a graph convolution layer and a time convolution layer for capturing spatial dependence and temporal dynamics in the traffic network, respectively; The spatiotemporal features are dynamically weighted through an attention mechanism to enhance the representation of key nodes and key time steps; The weighted spatiotemporal features are input into a fully connected layer to output traffic flow prediction values in the future preset time period. 9.The intelligent traffic flow prediction method based on multi-source data fusion according to claim 6, characterized in that, After the feature importance screening through the XGBoost algorithm to obtain the screened feature subset, it further includes: Real-time monitoring of sudden traffic event information in the traffic network; the sudden traffic event information includes traffic accidents, road construction, or temporary regulation; When detecting the sudden traffic event information, the initial weight of each feature in the XGBoost algorithm or the bias term of the split gain calculation is dynamically adjusted to improve the priority of real-time features related to traffic flow mutation in the feature importance screening process; Wherein, the real-time features related to traffic flow mutation at least include: the floating car flow difference between the current time and the previous time, the instantaneous drop rate of floating car average speed, and the user route change request frequency from navigation software. 10.The intelligent traffic flow prediction method based on multi-source data fusion according to claim 9, characterized in that, The dynamic adjustment of the initial weight of each feature in the XGBoost algorithm or the bias term of the split gain calculation includes: According to the type and location of the sudden traffic event, the real-time flow mutation index of the nodes in the event influence area is calculated; When the real-time flow mutation index exceeds a preset threshold, a weight adjustment mechanism is triggered; In the weight adjustment mechanism, the real-time features related to flow mutation are given higher initial weights, and when the decision tree node is split, the split gain of these features is positively biased.

Images