Road network status prediction method and system combined with ETC gantry data

Abstract

The present application provides a method and system for predicting a road network state in combination with ETC gantry data, which obtains prediction indication information input by a user during intelligent interaction with a traffic management system, including traffic network data to be predicted, and intervention prediction instructions for the traffic network data to be predicted; extracts instruction representation vectors for the intervention prediction instructions to obtain intervention prediction instruction representation vectors, extracts road network data representation vectors for the traffic network data to be predicted to obtain road network data representation vectors; performs intervention prediction on the road network data representation vector based on the intervention prediction instruction representation vector to obtain target road network data representation vectors, and performs road network information restoration mapping on the target road network data representation vector to obtain target traffic network data corresponding to the traffic network data to be predicted, and each traffic element of the target traffic network data is generated under the intervention mode. The present application can increase the accuracy of traffic network data prediction.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a road network status prediction method and system combining ETC gantry data. Background Art

[0002] With the acceleration of urbanization and the continuous growth of traffic demand, the intelligent and precise traffic management system has become the key to improving urban traffic efficiency and alleviating traffic congestion. Traditional traffic network data prediction methods often rely on statistical analysis of historical data or simple model prediction, which are incapable of dealing with complex and changing traffic conditions.

[0003] In the existing technology, although some traffic network data prediction methods based on machine learning have emerged, most of these methods focus on the mining and analysis of historical data, while ignoring the interactivity between users' actual needs and traffic management systems. Users can only passively accept the prediction results provided by the system, and cannot intervene or customize the prediction process according to their actual needs. This one-way information flow not only limits the flexibility and accuracy of the prediction, but also reduces user experience and satisfaction.

[0004] In addition, the existing traffic network data prediction methods also face many challenges when processing large-scale, high-dimensional traffic data. For example, how to effectively extract key features from traffic network data, how to build accurate prediction models, and how to handle the complex relationships between different traffic elements are all difficult problems that are difficult to effectively solve with current technology. These problems not only affect the accuracy and reliability of the prediction results, but also limit the application effect of traffic management systems in real-time traffic monitoring, dynamic traffic scheduling, and intelligent traffic signal control. Summary of the invention

[0005] In view of this, the present application provides a road network status prediction method and system combined with ETC gantry data. The technical solution of the present application is implemented as follows:

[0006] On the one hand, the present application provides a method for predicting a road network state in combination with ETC gantry data, the method comprising: obtaining prediction indication information input by a user during intelligent interaction with a traffic management system, the prediction indication information comprising traffic network data to be predicted, and intervention prediction instructions for the traffic network data to be predicted; wherein the intervention prediction instructions are used to indicate an intervention method for the traffic network data to be predicted on one or more traffic elements; the traffic network data to be predicted comprises a plurality of ETC gantry data in a monitored traffic area; performing instruction representation vector extraction on the intervention prediction instructions to obtain an intervention prediction instruction representation vector, and performing network data representation vector extraction on the traffic network data to be predicted to obtain a network data representation vector; performing intervention prediction on the network data representation vector based on the intervention prediction instruction representation vector to obtain a target network data representation vector, and performing network information restoration mapping on the target network data representation vector to obtain target traffic network data corresponding to the traffic network data to be predicted; wherein the traffic elements presented by the target traffic network data are generated under the intervention method.

[0007] In one implementation, the intervention prediction instruction includes one or more sub-intervention prediction instructions, and the instruction types of the sub-intervention prediction instructions include label instruction types and text instruction types. The extracting the instruction representation vector of the intervention prediction instruction to obtain the intervention prediction instruction representation vector includes:

[0008] For each of the sub-intervention prediction instructions, if the instruction type of the sub-intervention prediction instruction is the label instruction type, the sub-intervention prediction instruction is mapped to a label vector to obtain a sub-intervention representation vector corresponding to the sub-intervention prediction instruction; if the instruction type of the sub-intervention prediction instruction is the text instruction type, the sub-intervention prediction instruction is encoded with text semantics to obtain a sub-intervention representation vector corresponding to the sub-intervention prediction instruction;

[0009] The sub-intervention representation vectors corresponding to each of the sub-intervention prediction instructions are integrated to obtain the intervention prediction instruction representation vector; the sub-intervention prediction instruction corresponds to each of the traffic elements, and the sub-intervention prediction instruction is used to indicate the intervention method of the traffic network data to be predicted on the corresponding traffic element.

[0010] In one implementation, the traffic network data to be predicted includes a plurality of traffic confluences and roads connecting the traffic confluences, and extracting a network data representation vector from the traffic network data to be predicted to obtain the network data representation vector includes:

[0011] Performing a traffic directed graph transformation on the traffic network data to be predicted, obtaining a target traffic directed graph corresponding to the traffic network data to be predicted, wherein the graph vertices in the target traffic directed graph correspond to the traffic confluences in the traffic network data to be predicted, and the connecting lines in the target traffic directed graph correspond to the roads in the traffic network data to be predicted; wherein the graph vertex representation information of the graph vertices in the target traffic directed graph is used to represent the traffic information of the corresponding traffic confluence;

[0012] A road network data representation vector is extracted from the target traffic directed graph to obtain the road network data representation vector.

[0013] In one implementation, the road network data representation vector extraction is completed based on a graph recursive network model, the number of components of the message transmission component in the graph recursive network model is X, X>1; the road network data representation vector extraction is performed on the target traffic directed graph to obtain the road network data representation vector, including:

[0014] Acquire the vertex state array, relationship array and vertex connectivity array corresponding to the target traffic directed graph, and generate a comparison array by combining the relationship array and the vertex connectivity array; wherein the vertex state array is a two-dimensional array generated according to the vertex representation information of each vertex in the target traffic directed graph, the vertex connectivity array is a two-dimensional array generated according to the degree of each vertex in the target traffic directed graph, and the relationship array is a two-dimensional array generated according to the connection between any two vertices in the target traffic directed graph;

[0015] Extracting representation information from the vertex state array and the reference array through the first message passing component in the graph recursive network model to obtain a representation information extraction result of the first message passing component;

[0016] Traversing m to complete the following operations: extracting the representation information of the m-1th message passing component and the reference array through the mth message passing component in the graph recursive network model, and obtaining the representation information extraction result of the mth message passing component; wherein 1<m≤X;

[0017] The characterization information extraction result of the Xth message transmission component is used as the road network data characterization vector.

[0018] In one implementation, the m-th message passing component includes a representation vector integration component and a nonlinear transformation component; the representation information extraction result of the m-1-th message passing component and the reference array is extracted through the m-th message passing component in the graph recursive network model to obtain the representation information extraction result of the m-th message passing component, including:

[0019] Obtaining an influence coefficient array of the mth message delivery component;

[0020] Based on the representation vector integration component of the mth message transmission component, performing a product operation on the representation information extraction result of the m-1th message transmission component, the reference array and the influence coefficient array to obtain an operational product;

[0021] Based on the nonlinear transformation component of the mth message passing component, a nonlinear transformation operation is performed on the operational product to obtain a characterization information extraction result of the mth message passing component.

[0022] In one implementation, the intervention prediction instruction representation vector includes sub-intervention representation vectors corresponding to each of the traffic elements, and the intervention prediction is performed on the road network data representation vector based on the intervention prediction instruction representation vector to obtain a target road network data representation vector, including:

[0023] If the number of the sub-intervention representation vectors in the intervention prediction instruction representation vector is 1, performing intervention prediction on the road network data representation vector based on the intervention prediction instruction representation vector through the representation information prediction component to obtain the target road network data representation vector;

[0024] If there are multiple sub-intervention representation vectors in the intervention prediction instruction representation vector, the representation information prediction component is used in sequence to perform intervention prediction on the road network data representation vector based on each sub-intervention representation vector to obtain the target road network data representation vector.

[0025] In one implementation, the step of sequentially performing intervention prediction on the road network data representation vector according to each of the sub-intervention representation vectors through the representation information prediction component to obtain the target road network data representation vector includes:

[0026] By means of the representation information prediction component, intervention prediction is performed on the road network data representation vector according to the first sub-intervention representation vector to obtain a first road network data representation vector;

[0027] Traverse n to complete the following operations: perform intervention prediction on the n-1th road network data representation vector according to the nth sub-intervention representation vector through the representation information prediction component to obtain the nth road network data representation vector; wherein 1<n≤X, X represents the number of the sub-intervention representation vectors in the intervention prediction instruction representation vector, and X>1;

[0028] Taking the Xth road network data representation vector as the target road network data representation vector;

[0029] Before performing intervention prediction on the road network data representation vector based on the intervention prediction instruction representation vector to obtain the target road network data representation vector, the method further includes:

[0030] Acquire an initial representation information prediction component, a first training database and a second training database, wherein the first training database includes a plurality of traffic network training data representation vectors, and the second training database includes a plurality of intervention instruction representation vectors;

[0031] Determining a first training error value corresponding to the initial representation information prediction component based on the first training database, and determining a second training error value corresponding to the initial representation information prediction component based on the second training database;

[0032] The first training error value and the second training error value are merged to obtain a target training error value, and based on the target training error value, the initial representation information prediction component is adjusted to obtain the representation information prediction component.

[0033] In one implementation, determining, based on the first training database, a first training error value corresponding to the initial representation information prediction component includes:

[0034] Obtaining the training intervention prediction instruction representation vectors respectively corresponding to the traffic network training data representation vectors in the first training database, and the priori representation vectors respectively corresponding to the traffic network training data representation vectors;

[0035] Through the initial representation information prediction component, based on each of the training intervention prediction instruction representation vectors, the corresponding traffic network training data representation vectors are intervened and predicted to obtain the predicted road network data representation vectors corresponding to each of the traffic network training data representation vectors;

[0036] For each of the traffic network training data representation vectors, taking the commonality measurement result between the corresponding prior representation vector and the corresponding predicted road network data representation vector as the first model training error value corresponding to the traffic network training data representation vector;

[0037] The training error values ​​of the first models are integrated to obtain a first training error value corresponding to the initial representation information prediction component;

[0038] The determining, based on the second training database, a second training error value corresponding to the initial representation information prediction component includes:

[0039] Obtaining standard prediction result representation vectors respectively corresponding to each of the intervention instruction representation vectors in the second training database;

[0040] Respectively through the initial representation information prediction component, based on each of the intervention instruction representation vectors, the training intervention prediction instructions corresponding to each of the intervention instruction representation vectors are inferred to represent the road network data after intervention, so as to obtain the inferred road network data after intervention corresponding to each of the intervention instruction representation vectors;

[0041] For each of the intervention instruction representation vectors, a commonality measurement result between the corresponding inferenced post-intervention road network data representation vector and the corresponding standard prediction result representation vector is used as a second model training error value corresponding to the intervention instruction representation vector;

[0042] The training error values ​​of each of the second models are added together to obtain a second training error value corresponding to the initial representation information prediction component.

[0043] In one implementation, the step of obtaining the prediction indication information input by the user during the process of intelligent interaction with the traffic management system includes:

[0044] Upon receiving the interaction trigger information input by the user, triggering an interface for intelligent interaction between the user and the traffic management system;

[0045] The interaction start information is displayed to the user, and the prediction indication information input by the user is received.

[0046] On the other hand, the present application provides a computer system, including a memory and a processor, wherein the memory stores a computer program executable on the processor, and the processor implements the steps in the above method when executing the program.

[0047] The technical effect of the present application is as follows: The present application obtains the prediction indication information input by the user in the process of intelligent interaction with the traffic management system, the prediction indication information includes the traffic network data to be predicted, and the intervention prediction instruction for the traffic network data to be predicted, extracts the instruction representation vector of the intervention prediction instruction, obtains the intervention prediction instruction representation vector, extracts the road network data representation vector of the traffic network data to be predicted, obtains the road network data representation vector, performs intervention prediction on the road network data representation vector based on the intervention prediction instruction representation vector, obtains the target road network data representation vector, performs road network information restoration mapping on the target road network data representation vector, and obtains the target traffic network data. In this way, based on the intervention prediction instruction representation vector, the road network data representation vector is intervened and predicted, and the target road network data representation vector is obtained. The road network information is restored and mapped to obtain the target traffic network data, so that each traffic element of the obtained target traffic network data is generated under the intervention mode, thereby based on the intelligent interaction between the user and the traffic management system, the interactive ability in the prediction of traffic network data is improved, the obtained target traffic network data conforms to the intervention mode, and the prediction accuracy of the traffic network data is increased. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the present application and are used together with the specification to illustrate the technical solution of the present application.

[0049] Figure 1 A schematic diagram of the implementation flow of a road network status prediction method combined with ETC gantry data provided in an embodiment of the present application.

[0050] Figure 2 A hardware entity diagram of a computer system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solutions and advantages of the present application clearer, the technical solutions of the present application are further elaborated in detail below in conjunction with the drawings and embodiments. The described embodiments should not be regarded as limiting the present application. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

[0052] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing this application and are not intended to limit this application.

[0053] The embodiment of the present application provides a road network state prediction method combined with ETC gantry data, which can be executed by a processor of a computer system. The computer system can refer to a device with data processing capabilities such as a server, a laptop, a tablet computer, a desktop computer, a mobile device, etc. Figure 1 A schematic diagram of the implementation flow of a road network status prediction method combined with ETC gantry data provided in an embodiment of the present application is shown in FIG. Figure 1 As shown, the method comprises the following steps:

[0054] Step S100: Obtain prediction indication information input by the user during the process of intelligent interaction with the traffic management system, the prediction indication information including the traffic network data to be predicted and the intervention prediction instructions for the traffic network data to be predicted; wherein the intervention prediction instructions are used to indicate the intervention method of the traffic network data to be predicted on one or more traffic elements; the traffic network data to be predicted includes multiple ETC gantry data in the monitored traffic area.

[0055] In this step, the computer system is responsible for obtaining the prediction indication information input by the user when the user interacts intelligently with the traffic management system (for example, an AI agent embedded in the computer system). This process not only involves the collection of data, but also includes the understanding and analysis of the user's intention, which provides a basis for the subsequent prediction of the road network status. The following is a detailed explanation of step S100:

[0056] In step S100, the computer system establishes an intelligent interaction channel with the user. This channel can be an application based on a graphical user interface, where the user can interact with the AI ​​agent by clicking, entering text, or selecting options; or it can be a voice interaction system, where the user communicates with the AI ​​agent through voice commands. Regardless of the form, the key is to ensure that the user can easily and accurately convey his predicted needs.

[0057] Once the intelligent interactive channel is established, the user begins to input prediction instructions. This information consists of two parts: the traffic network data to be predicted and the intervention prediction instructions for these data. The traffic network data to be predicted comes from multiple ETC gantries in the monitored traffic area. ETC gantries are devices used for non-stop toll collection on highways. They can not only record vehicle traffic information, but also monitor traffic data such as vehicle flow and speed in real time. The computer system collects data from these ETC gantries to build a traffic network data set containing multiple traffic elements. These traffic elements can be, but are not limited to, vehicle flow (such as the number of vehicles passing through a certain section of road per unit time), congestion (such as the length of the vehicle queue or the average driving speed of a certain section of road), and comfort (such as measuring the passenger experience during driving, such as road surface flatness, noise level, etc.). For example, an ETC gantry may record that the total number of vehicles passing through the gantry in a certain period of time is 1,000, and the average speed is 60 kilometers per hour. These data constitute part of the traffic network data.

[0058] Next, the user inputs the intervention prediction instruction for the traffic network data to be predicted. This instruction is input by the user according to his own needs or expectations of traffic conditions. The intervention prediction instruction is used to indicate the intervention method of the traffic network data to be predicted on one or more traffic elements. For example, the user may want to know how traffic elements such as traffic flow, congestion and road comfort in the traffic network will change if a new traffic control measure is added to a certain road (such as restricting the passage of large vehicles during a specific time period). The "intervention" here refers to the traffic control measure set by the user, and the prediction in this application refers to the prediction of specific traffic elements such as traffic flow, congestion and comfort. These intervention prediction instructions can exist in different forms, such as in the form of strategy tags, just like attaching specific labels to different intervention strategies (such as "large vehicle restriction strategy" labels) to facilitate computer system recognition and processing; they can also exist in text form, such as the user directly inputs a detailed text description such as "restricting the passage of large vehicles from [start time] to [end time] on [specific road name], predicting changes in traffic flow, congestion and comfort".

[0059] After obtaining the prediction instruction information, the computer system processes and analyzes the information. For the traffic network data to be predicted, the computer system can perform operations such as cleaning, integration and feature extraction to ensure the accuracy and consistency of the data. For intervention prediction instructions, the computer system uses natural language processing (NLP) technology or pattern matching algorithms to analyze the user's intentions. For example, if the intervention prediction instruction is entered in text form, the computer system first performs pre-processing operations such as word segmentation and part-of-speech tagging on the text, and then uses semantic analysis technology to understand the meaning of the text and convert it into a format that can be processed by the computer (such as vector representation).

[0060] In addition, in order to more accurately understand the user's intervention prediction instructions, the computer system can also use machine learning models for assistance. For example, a classification model can be trained to identify which intervention strategy label the text instruction entered by the user belongs to; or a generation model can be trained to generate the corresponding intervention strategy representation based on the user's descriptive text. These models can be trained and optimized with a large amount of historical data to improve the accuracy of identifying user intentions.

[0061] Finally, the prediction indication information obtained in step S100 is the basis of the entire road network state prediction method. Subsequent steps (such as step S200 and step S300) will extract the road network data representation vector and intervene in the prediction based on this information. Therefore, ensuring the accuracy and completeness of the prediction indication information obtained in step S100 is crucial to improving the performance of the entire prediction method.

[0062] As an implementation mode, step S100, obtaining prediction indication information input by a user during intelligent interaction with a traffic management system, includes:

[0063] Step S110: upon receiving the interaction trigger information input by the user, triggering the interface for intelligent interaction between the user and the traffic management system;

[0064] Step S120: displaying interaction start information to the user, and receiving prediction indication information input by the user.

[0065] In step S110, the computer system waits for and receives interaction trigger information input by the user. This trigger information can be sent by the user in some way (such as clicking a button, voice command, etc.) to start the intelligent interaction process with the traffic management system. Once the trigger information is received, the computer system will respond immediately and trigger the interface for the user to interact intelligently with the traffic management system.

[0066] This interactive interface is a bridge for users to communicate with the system. It provides a channel for users to input prediction instructions. The design of the interactive interface should be intuitive and easy to use, and can guide users to clearly express their prediction needs. For example, the interactive interface may contain multiple input fields or options, corresponding to different traffic elements (such as traffic flow, congestion, comfort, etc.) and intervention methods (such as speed limit regulation at specific locations or intervals, level settings, specific vehicle restrictions, vehicle diversion, etc.). Users can construct their own prediction instructions by filling in these fields or selecting corresponding options.

[0067] The triggering method of the interactive interface may vary depending on the specific application scenario. In some cases, the interactive interface may be an independent application or webpage that users need to actively open and log in to use; in other cases, the interactive interface may be embedded in other functional modules of the traffic management system, and the interactive interface will be automatically triggered when users use these functional modules. Regardless of the method used, the goal of the computer system is to ensure that users can easily and quickly start the intelligent interaction process with the system.

[0068] In step S120, once the interactive interface is triggered and displayed to the user, the computer system will immediately display interactive start information to the user. This start information may be a brief prompt or description to inform the user that the interactive process has begun and guide the user to the next step. For example, the start information may include prompts such as "Please enter your forecast indication information" or "Please select the traffic elements and intervention methods you are concerned about".

[0069] At the same time, while displaying the interactive start information, the computer system is also ready to receive the prediction instruction information input by the user. The prediction instruction information is the specific prediction requirement put forward by the user to the traffic management system. It contains the traffic network data that the user wants the system to predict and the corresponding intervention prediction instructions. This information is crucial for the subsequent road network status prediction because it defines the goal and scope of the prediction.

[0070] Users can enter forecast instructions through input fields or options provided by the interactive interface. For example, users may first select the traffic element of interest (such as traffic flow), then enter the forecast time range (such as the next week) and spatial range (such as the central area of ​​the city), and finally select or enter intervention forecast instructions (such as implementing flow control measures). During the input process, the computer system may provide real-time feedback or suggestions to help users express their forecast needs more accurately. Once the user completes the input of the forecast instruction information and submits it to the system, the computer system will immediately receive and process the information. The processing process may include operations such as data cleaning, verification, and formatting to ensure the accuracy and consistency of the input data. Subsequently, the information will be passed to subsequent steps for further processing and analysis.

[0071] In step S120, the computer system should maintain a high degree of sensitivity and accuracy when receiving the forecast indication information input by the user. This is because the quality of the forecast indication information directly affects the accuracy and reliability of the subsequent road network status prediction. If there are errors or omissions in the input information, the forecast result may deviate from the actual situation, thereby affecting the decision-making effect of the traffic management system. Therefore, the computer system should use a variety of means to ensure the accuracy and completeness of the input information, such as providing input verification, error prompts, and real-time feedback functions.

[0072] Step S200: extracting the instruction representation vector of the intervention prediction instruction to obtain the intervention prediction instruction representation vector, and extracting the road network data representation vector of the traffic road network data to be predicted to obtain the road network data representation vector.

[0073] In step S200, the computer system first processes the intervention prediction instructions. These instructions are input by the user during the intelligent interaction process and are used to instruct the AI ​​agent which traffic elements to focus on when making predictions and the expected changes in these elements. In order to convert these instructions into representation vectors that can be processed by the computer, the computer system performs an instruction representation vector extraction operation.

[0074] The process of extracting instruction representation vectors can be divided into two main steps: text processing and vector mapping. For intervention prediction instructions input in text form, the computer system will first perform text processing, including preprocessing operations such as word segmentation, part-of-speech tagging, and stop word removal. These operations help convert the original text into a format that is easier to analyze. For example, the intervention prediction instruction input by the user is "predict the traffic flow and congestion of the XX section after the implementation of one-way traffic on the XX section during peak hours." After word segmentation, the instruction is split into vocabulary units such as "prediction", "peak hour", "XX section", "implementation", "one-way traffic", "after", "this section", "traffic flow" and "congestion".

[0075] Next, the computer system performs vector mapping on the processed text. The purpose of this step is to convert text information into points in the vector space for subsequent calculations and analysis. Vector mapping can be achieved through a variety of methods, the most common of which is word embedding technology, such as Word2Vec, GloVe, etc. These technologies can map each word into a high-dimensional vector space, so that semantically similar words are also close in distance in the vector space. For example, in the above instructions, the two words "traffic flow" and "congestion" are semantically related, and their vector representations in the vector space should also have a high degree of similarity.

[0076] In addition to text processing and vector mapping, for intervention prediction instructions in the form of strategy labels, the computer system can directly perform label vector mapping. Label vector mapping is to map each strategy label to a vector space of fixed dimension, where each dimension represents a different feature or attribute. For example, a three-dimensional vector space can be defined to represent different intervention strategy labels, where the first dimension represents the presence or absence of flow control measures, the second dimension represents the degree of route optimization, and the third dimension represents other possible intervention measures. In this way, each strategy label can be mapped to a point in this vector space.

[0077] After obtaining the intervention prediction instruction representation vector, the computer system then processes the traffic network data to be predicted. These data come from multiple ETC gantries in the monitored traffic area and include various traffic elements such as traffic flow, congestion, and comfort. In order to convert these data into representation vectors that can be processed by the computer, the computer system performs a road network data representation vector extraction operation.

[0078] The process of extracting road network data representation vectors can be divided into three steps: data preprocessing, feature extraction, and vector integration. First, the computer system preprocesses the original road network data, including data cleaning, outlier processing, missing value filling, and other operations. These operations help ensure the accuracy and consistency of the data and lay the foundation for subsequent feature extraction. For example, if the data of an ETC gantry is missing traffic flow information for a certain period of time due to equipment failure, the computer system can estimate the traffic flow for that period of time through interpolation or other methods.

[0079] Next, the computer system will perform feature extraction on the preprocessed road network data. Feature extraction is the process of extracting useful information for the prediction task from the raw data. In road network status prediction, feature extraction may involve a variety of methods, such as time series analysis, spatial statistical analysis, machine learning feature selection, etc. For example, for traffic flow data, time series analysis methods can be used to extract features such as trend items, seasonal items, and random items; for congestion data, spatial statistical analysis methods can be used to extract features such as the shape, size, and location of the congested area.

[0080] After extracting the features, the computer system will perform vector integration on these features. Vector integration is the process of integrating multiple features into a vector space of fixed dimension. This can be achieved through a variety of methods, such as simple concatenation, weighted average, principal component analysis, etc. For example, if three features of traffic flow, congestion, and comfort are extracted, they can be concatenated into a three-dimensional vector, where each dimension represents the value of a feature. If there is correlation or redundancy between the features, principal component analysis can be used to reduce their dimensionality to a lower dimensional vector space.

[0081] In the process of extracting the road network data representation vector, the computer system can also use machine learning models to further improve the effect of feature extraction and vector integration. For example, the Graph Recursive Neural Network (GRNN) model can be used to process the spatial dependencies in the road network data. GRNN is a neural network model specifically used to process graph structure data. It can transfer information between nodes of the graph through a message passing mechanism to extract the deep features of the nodes. In the prediction of road network status, the traffic road network can be regarded as a graph structure, in which each ETC gantry is regarded as a node, and the connection relationship between the nodes represents the connection status of the road. By processing the road network data through the GRNN model, the deep feature representation of each node can be obtained, and then integrated into the road network data representation vector.

[0082] Step S300: Based on the intervention prediction instruction representation vector, perform intervention prediction on the road network data representation vector to obtain the target road network data representation vector, and perform road network information restoration mapping on the target road network data representation vector to obtain the target traffic network data corresponding to the traffic network data to be predicted; wherein the traffic elements presented in the target traffic network data are generated under the intervention mode.

[0083] At the beginning of step S300, the computer system combines the intervention prediction instruction representation vector and the road network data representation vector for prediction. The intervention prediction instruction representation vector contains the user's expectations or assumptions about the traffic network status, while the road network data representation vector represents the current or historical traffic network status. In order to effectively combine the two, the computer system uses a machine learning model that can process such composite inputs and make predictions.

[0084] One possible solution is to use a Representation Information Prediction Component (RIPC). RIPC is a machine learning model specifically designed to handle this type of prediction task. It receives two input vectors (intervention prediction instruction representation vector and road network data representation vector) and outputs a target road network data representation vector through a complex internal calculation process. The internal structure of RIPC may include components such as multi-layer neural networks, attention mechanisms, and recurrent neural networks, which work together to capture the complex relationships between input vectors and generate accurate prediction results.

[0085] Specifically, when RIPC receives the intervention prediction instruction representation vector and the road network data representation vector as input, it encodes the two vectors through a multi-layer neural network to extract their respective high-level feature representations. Then, RIPC uses the attention mechanism to dynamically adjust the weights of the two vectors in the prediction process, so that features that are more relevant to the prediction task can receive greater attention. Next, RIPC inputs the encoded vector into a recurrent neural network to capture dependencies in the time series. Finally, RIPC outputs a target road network data representation vector, which contains a high-level feature representation of the prediction results.

[0086] If the intervention prediction instruction representation vector contains multiple sub-intervention prediction instruction representation vectors (i.e., the user proposes multiple intervention strategies), then RIPC may need to make multiple predictions, each prediction considering only one sub-intervention prediction instruction representation vector. In this case, RIPC can first make a prediction based on the first sub-intervention prediction instruction representation vector and the road network data representation vector to obtain the first target road network data representation vector. Then, it will use the first target road network data representation vector as a new input and input it into RIPC together with the second sub-intervention prediction instruction representation vector for a second prediction, and so on, until all sub-intervention prediction instruction representation vectors are processed. Finally, the computer system can select the last target road network data representation vector as the final prediction result, or fuse multiple prediction results as needed.

[0087] After obtaining the target road network data representation vector, the computer system performs a road network information restoration mapping operation to convert it into a human-understandable traffic road network data form. The purpose of this step is to map the prediction results from the high-level feature representation space back to the original data space, so that the prediction results can be directly used in practical application scenarios such as traffic management, planning or decision-making.

[0088] The road network information restoration mapping operation can be implemented in many ways, one of the common methods is to use a decoder network. The decoder network is a neural network structure corresponding to the encoder network, which can decode the high-level feature representation output by the encoder network back to the original data space. In the road network state prediction, the decoder network can receive the target road network data representation vector as input, and gradually restore the various traffic elements in the road network data (such as traffic flow, congestion, comfort, etc.) through a multi-layer neural network. Each layer of the neural network may correspond to a specific traffic element or a group of related traffic elements. By decoding layer by layer, the complete target traffic road network data is finally obtained.

[0089] In addition to using the decoder network, the computer system can also combine other methods to restore the road network information. For example, it can use domain knowledge or expert experience to build some rules or constraints to ensure that the restored traffic network data conforms to the actual situation and physical laws. In addition, the computer system can also perform post-processing operations on the restored traffic network data, such as smoothing, outlier detection and correction, to improve the accuracy and reliability of the data.

[0090] After the road network information restoration and mapping is completed, the computer system obtains the target traffic network data. These data present the state of the traffic network under the intervention method specified by the user in the intervention prediction instruction, and can be used to evaluate the effects of different intervention strategies, optimize traffic management plans, and guide traffic planning decisions. For example, if the user wants to predict the traffic conditions after one-way traffic is implemented on a certain road section during peak hours, the computer system can generate the target traffic network data by executing step S300, and display the changes in key indicators such as traffic flow and congestion after the implementation of one-way traffic. This information is very valuable to traffic management departments and can help them better formulate and implement traffic management strategies.

[0091] As an implementation mode, the intervention prediction instruction includes one or more sub-intervention prediction instructions, and the instruction types of the sub-intervention prediction instructions include label instruction types and text instruction types. Based on this, in step S200, the intervention prediction instruction is subjected to instruction representation vector extraction to obtain the intervention prediction instruction representation vector, including:

[0092] Step S210: For each sub-intervention prediction instruction, if the instruction type of the sub-intervention prediction instruction is a label instruction type, the sub-intervention prediction instruction is mapped to a label vector to obtain a sub-intervention representation vector corresponding to the sub-intervention prediction instruction; if the instruction type of the sub-intervention prediction instruction is a text instruction type, the sub-intervention prediction instruction is encoded to obtain a sub-intervention representation vector corresponding to the sub-intervention prediction instruction;

[0093] Step S220: integrating the sub-intervention characterization vectors corresponding to each sub-intervention prediction instruction to obtain an intervention prediction instruction characterization vector; the sub-intervention prediction instruction corresponds to each traffic element, and the sub-intervention prediction instruction is used to indicate the intervention method of the traffic network data to be predicted on the corresponding traffic element.

[0094] In step S210, the computer system first processes each sub-intervention prediction instruction to obtain a corresponding sub-intervention characterization vector. This processing process depends on the type of sub-intervention prediction instruction. If the sub-intervention prediction instruction is a label instruction, the computer system will map a label vector to it. Label vector mapping is a technique for mapping discrete labels to continuous vector space, which can convert each label into a vector representation of a fixed dimension. This vector representation not only retains the semantic relationship between labels, but also enables labels to be operated and compared in a continuous space. For example, assuming there are three label instructions: "flow limiting measures", "optimize routes" and "improve comfort", the computer system may map them to three points in the three-dimensional vector space, such as (1,0,0), (0,1,0) and (0,0,1). In this way, each label instruction corresponds to a unique sub-intervention characterization vector.

[0095] If the sub-intervention prediction instruction is a text instruction, the computer system will perform text semantic encoding on it. Text semantic encoding is a technology that converts text information into vector representation, which can capture the semantic information in the text and present this information in the form of a vector. Text semantic encoding can be implemented in a variety of ways, such as word embedding, sentence embedding, or paragraph embedding. In the word embedding method, each word or phrase is mapped to a high-dimensional vector space, and the distance between vectors reflects the semantic similarity between words or phrases. In the sentence embedding or paragraph embedding method, the entire sentence or paragraph is mapped to a vector space, and the vector represents the overall semantics of the sentence or paragraph. For example, assuming that a text instruction is "predict the change in traffic flow after the implementation of one-way traffic on the XX road section during peak hours", the computer system may use the word embedding method to map each word in the instruction to the vector space, and integrate these vectors into a sub-intervention representation vector by weighted averaging or other methods.

[0096] Regardless of whether the sub-intervention prediction instruction is a label instruction or a text instruction, after processing in step S210, the computer system will obtain a corresponding sub-intervention representation vector. These vectors not only contain the semantic information of the instruction, but also enable the instruction to be calculated and compared in the subsequent prediction process.

[0097] In step S220, the computer system performs representation vector integration on the sub-intervention representation vectors corresponding to all sub-intervention prediction instructions to obtain the final intervention prediction instruction representation vector. Representation vector integration is a process of merging multiple vectors into one vector, which can be achieved through a variety of methods, such as simple splicing, weighted averaging, principal component analysis, etc. In the simple splicing method, the computer system will sequentially splice all sub-intervention representation vectors into a longer vector. In the weighted averaging method, the computer will assign a weight to each sub-intervention representation vector based on its importance or confidence, and then average the weighted vectors. In the principal component analysis method, the computer will perform dimensionality reduction processing on all sub-intervention representation vectors to extract the most important information components and combine these components into a new vector.

[0098] The choice of which representation vector integration method to use depends on the specific application scenario and requirements. For example, if all sub-intervention prediction instructions have the same importance in the prediction process, then the simple splicing method may be a good choice. If some sub-intervention prediction instructions are more reliable or more important than others, then the weighted average method may be more appropriate. If there is a high correlation or redundancy between the sub-intervention representation vectors, then the principal component analysis method may help reduce the data dimension and improve the prediction efficiency.

[0099] During the characterization vector integration process, the computer system also needs to consider the correspondence between the sub-intervention prediction instructions and the traffic elements. Each sub-intervention prediction instruction corresponds to one or more traffic elements (such as traffic flow, congestion, comfort, etc.) and indicates the intervention method on these elements. Therefore, when performing characterization vector integration, the computer system ensures that each sub-intervention characterization vector can correctly reflect its corresponding traffic element and intervention method. This can be achieved by introducing additional tags or weights during the vector integration process to distinguish the impact of different sub-intervention characterization vectors on the prediction results.

[0100] For example, suppose there are three sub-intervention prediction instructions: "flow limiting measures" (corresponding to traffic flow elements), "optimized routes" (corresponding to congestion elements) and "improved comfort" (corresponding to comfort elements). After processing in step S210, the computer system obtains three sub-intervention representation vectors v1, v2 and v3. In step S220, the computer system can select a simple splicing method to integrate them into an intervention prediction instruction representation vector V, that is, V = [v1, v2, v3]. At the same time, in order to distinguish the impact of different sub-intervention representation vectors on the prediction results, the computer system can assign a weight w1, w2 and w3 to each vector (these weights can be determined based on domain knowledge, expert experience or historical data), and average the weighted vectors, that is, V = (w1*v1 + w2*v2 + w3*v3) / (w1 + w2 + w3). In this way, the intervention prediction instruction representation vector V contains the information of all sub-intervention prediction instructions and correctly reflects their intervention methods on the corresponding traffic elements.

[0101] As an implementation method, the traffic network data to be predicted includes multiple traffic confluences and roads connecting the traffic confluences. Based on this, in step S200, the traffic network data to be predicted is subjected to a road network data characterization vector extraction to obtain a road network data characterization vector, including:

[0102] Step S230: performing traffic directed graph transformation on the traffic network data to be predicted, and obtaining a target traffic directed graph corresponding to the traffic network data to be predicted, wherein the graph vertices in the target traffic directed graph correspond to the traffic confluences in the traffic network data to be predicted, and the connecting lines in the target traffic directed graph correspond to the roads in the traffic network data to be predicted; wherein the graph vertex representation information of the graph vertices in the target traffic directed graph is used to represent the traffic information of the corresponding traffic confluences;

[0103] Step S240: extracting a road network data representation vector from the target traffic directed graph to obtain a road network data representation vector.

[0104] In step S230, the computer system first transforms the traffic network data to be predicted into a traffic directed graph. A traffic directed graph is a special graph structure that uses nodes (graph vertices) to represent traffic confluences and uses directed edges (connecting lines) to represent roads connecting these confluences. This representation method can clearly show the topological structure and traffic flow of the road network data, providing a basis for the subsequent extraction of the road network data representation vector.

[0105] When transforming the traffic directed graph, the computer system will map each traffic confluence in the traffic network data to be predicted to a vertex in the target traffic directed graph, and map the roads connecting these confluences to directed edges. At the same time, in order to more comprehensively capture the traffic information of the traffic confluence, the computer system will also assign graph vertex representation information to each graph vertex. This representation information may include data on various traffic elements such as traffic flow, congestion, and comfort at the traffic confluence, which together constitute the feature vector of the graph vertex.

[0106] For example, suppose the traffic network data to be predicted contains three traffic confluences A, B, and C, and two roads AB and BC connecting them. After the traffic directed graph transformation, the computer system will obtain a target traffic directed graph, which contains three graph vertices (corresponding to the three traffic confluences A, B, and C) and two directed edges (corresponding to the two roads AB and BC). At the same time, each graph vertex will have a feature vector. For example, the feature vector of vertex A may contain information such as the traffic volume, congestion index, and average speed of confluence A.

[0107] In step S240, the computer system extracts the road network data characterization vector from the target traffic directed graph to obtain the road network data characterization vector. This step is the core link in the road network state prediction method, which involves feature extraction and vector representation of nodes and edges in the target traffic directed graph.

[0108] When extracting road network data representation vectors, computer systems can use a variety of methods, such as Graph Recursive Neural Network (GRNN), Graph Convolutional Network (GCN), etc. These methods are deep learning models specially designed to process graph structured data, which can capture the global and local features of the graph by transferring information between graph nodes.

[0109] Taking the GRNN model as an example, the model gradually extracts deep features of graph nodes through multi-layer message passing components. In each message passing component, the model calculates a new feature vector based on the feature vector of the current node and the feature vector of the neighboring node. This process can be expressed as:

[0110] ;

[0111] in, represents the feature vector of node i in layer l, represents the set of neighbor nodes of node i, represents the message passing function of layer l, represents the weight matrix of the lth layer, represents the activation function, Represents a concatenation operation.

[0112] In the GRNN model, the message passing function Can be any differentiable function, which is used to calculate the information transfer between node i and its neighbor node j. A common choice is to use a simple linear transformation plus a nonlinear activation function, such as:

[0113] ;

[0114] in, represents the linear transformation matrix of the lth layer, express Activation function.

[0115] Through iterative calculations of multi-layer message passing components, the GRNN model can gradually extract the deep features of each node in the target traffic directed graph. Ultimately, the computer system can splice these feature vectors together to form a road network data representation vector that contains global and local feature information of the entire traffic road network.

[0116] In practical applications, the specific structure and parameters of the GRNN model may need to be adjusted and optimized according to the characteristics of the traffic network data and the requirements of the prediction task. For example, the optimal number of model layers, node feature dimensions, and message passing function parameters can be selected through methods such as cross-validation.

[0117] In summary, S230 and S240 in step S200 are key steps in processing the traffic network data to be predicted. Through traffic directed graph transformation, the computer system can transform complex traffic network data into a clear graph structure representation; through network data representation vector extraction, the computer system can capture the global and local feature information of the network data, providing strong data support for subsequent intervention prediction.

[0118] As an implementation method, the road network data representation vector extraction is completed based on a graph recursive network model, and the number of components of the message transmission component in the graph recursive network model is X, X>1. Based on this, step S240, extracting the road network data representation vector from the target traffic directed graph, and obtaining the road network data representation vector includes:

[0119] Step S241: obtaining the vertex state array, relationship array and vertex connectivity array corresponding to the target traffic directed graph, and combining the relationship array and the vertex connectivity array to generate a comparison array; wherein the vertex state array is a two-dimensional array generated based on the graph vertex representation information of each graph vertex in the target traffic directed graph, the vertex connectivity array is a two-dimensional array generated based on the degree of each graph vertex in the target traffic directed graph, and the relationship array is a two-dimensional array generated based on the connection status between any two graph vertices in the target traffic directed graph;

[0120] Step S242: extracting representation information from the vertex state array and the reference array through the first message passing component in the graph recursive network model to obtain a representation information extraction result of the first message passing component;

[0121] Step S243: traverse m to complete the following operations: extract the representation information of the m-1th message passing component and the reference array through the mth message passing component in the graph recursive network model, and obtain the representation information extraction result of the mth message passing component; wherein 1<m≤X;

[0122] Step S244: extracting the characterization information of the Xth message transmission component as a road network data characterization vector.

[0123] In step S240, the computer system extracts the road network data representation vector from the target traffic directed graph to obtain the road network data representation vector. This step is completed based on the Graph Recursive Neural Network (GRNN), which gradually extracts the deep features of the graph nodes through multi-layer message passing components.

[0124] In step S241, the computer system first needs to obtain the vertex state array, relationship array and vertex connectivity array corresponding to the target traffic directed graph, and generate a comparison array by combining the relationship array and vertex connectivity array. These arrays are the basic data for the graph recursive network model to extract characterization information.

[0125] The vertex state array is a two-dimensional array, which can be understood as a feature matrix, which is used to describe the state of each vertex in the target traffic directed graph at different time points. This state information may include data on various traffic elements such as traffic flow, congestion, and average speed at traffic confluences. For example, assuming that the target traffic directed graph contains three vertices A, B, and C, and each vertex has state data at five time points, then the vertex state array may be a 3×5 matrix, where each row corresponds to the state data of a vertex, and each column corresponds to the state data at a time point.

[0126] The relationship array is also a two-dimensional array, which can be understood as an adjacency matrix, which is used to describe whether there is a connection relationship between any two graph vertices in the target traffic directed graph and the type or strength of the relationship. This array is similar to the adjacency matrix in graph theory, but may contain richer information. For example, in a traffic road network, the connection relationship between two graph vertices may not only indicate that there is a road connecting them, but also the type of road (such as highways, urban main roads, etc.), the number of lanes, speed limits and other information. Therefore, the elements in the relationship array may not only be 0 or 1 (indicating whether there is a connection), but may also be a composite value containing multiple attributes.

[0127] The vertex connectivity array is also a two-dimensional array that represents the number of connections between each vertex in the target traffic directed graph and other vertices. This array can be understood as a degree matrix in graph theory, but it provides more specific information about the connectivity of graph vertices. In a traffic network, the connectivity of a graph vertex may reflect the importance of the traffic confluence in the network and the size of the traffic flow. For example, a graph vertex with high connectivity may be an important traffic hub, and its traffic flow and congestion have a greater impact on the entire network.

[0128] The reference array is a new array generated by combining the relationship array and the vertex connectivity array. It may contain additional information to guide the message passing components to exchange information. For example, the reference array may contain the relative importance or weight information between each graph vertex and its neighbor vertices, which can affect the information flow and feature update during the message passing process.

[0129] In step S242, the computer system extracts characterization information from the vertex state array and the reference array through the first message passing component in the graph recursive network model. This step is the initial stage of feature extraction in the graph recursive network model, which uses the specific structure and parameters of the first message passing component to process the input data to obtain preliminary characterization information extraction results.

[0130] The first message passing component may contain a variety of network structures and operations, such as linear transformation, nonlinear activation function, attention mechanism, etc. These structures and operations work together on the vertex state array and the reference array to extract the preliminary feature representation of the graph vertex. For example, the first message passing component may first perform a linear transformation on the vertex state array, then introduce nonlinear factors through the ReLU activation function, and finally use the attention mechanism to perform a weighted summation of the feature representations of neighbor vertices based on the weight information in the reference array. In this way, each graph vertex will get an updated feature vector after being processed by the first message passing component.

[0131] In step S243, the computer system traverses m (1<m≤X), and further extracts the representation information of the m-1th message passing component and the reference array through the mth message passing component in the graph recursive network model. This step is the key link for deep feature extraction of the graph recursive network model, which gradually extracts the deep features of the graph nodes through iterative calculations of multiple layers of message passing components.

[0132] In each layer of the message passing component, the computer system uses the specific structure and parameters of the layer to process the representation information extraction results of the previous layer. This processing may include a variety of network structures and operations such as linear transformation, nonlinear activation function, and pooling operation. At the same time, the reference array will also participate in the calculation process of each layer of the message passing component as additional input information to guide information exchange and feature update.

[0133] As the number of message passing components increases, the feature representation of graph vertices will gradually become more complex and abstract. For example, in shallower message passing components, the feature representation of graph vertices may mainly reflect the information of their local neighbors; while in deeper message passing components, the feature representation of graph vertices may incorporate more extensive road network information and more complex traffic patterns.

[0134] The output of each layer of message passing components is further processed as the input of the next layer of message passing components. This iterative calculation method enables the graph recursive network model to gradually capture the global and local feature information of the graph data.

[0135] In step S244, the computer system uses the characterization information extraction result of the Xth message passing component as the final road network data characterization vector. This step is the final stage of feature extraction of the graph recursive network model, which outputs the deep feature representation obtained after iterative calculation of the multi-layer message passing component.

[0136] The road network data representation vector is a high-dimensional vector that contains the deep feature information of all graph vertices in the target traffic directed graph. These feature information not only reflects the local neighbor information and road network structure information of the graph vertices, but also may contain more complex traffic patterns and dynamic change information. Therefore, the road network data representation vector can be regarded as a comprehensive and in-depth description of the target traffic directed graph.

[0137] In practical applications, the road network data representation vector can be used in a variety of traffic management and planning tasks. For example, it can be used as input data to be sent to subsequent prediction models for tasks such as traffic flow prediction and congestion prediction; it can also be used in real-time decision-making tasks such as traffic signal control and route planning to provide accurate road network status information and prediction result support.

[0138] In summary, S241 to S244 in step S240 are the key steps for the graph recursive network model to extract the road network data representation vector. Through the processing and analysis of these steps, the computer system can extract the deep feature representation of the target traffic directed graph and use it as the road network data representation vector for subsequent traffic management and planning tasks. This process not only makes full use of the advantages of the graph recursive network model in processing graph structure data, but also gradually captures the global and local feature information of the road network data through the iterative calculation of the multi-layer message passing components, providing strong support for the accurate prediction and effective management of traffic status.

[0139] As an implementation, the mth message passing component includes a representation vector integration component and a nonlinear transformation component. Based on this, step S243, through the mth message passing component in the graph recursive network model, the representation information extraction result of the m-1th message passing component and the reference array are extracted to obtain the representation information extraction result of the mth message passing component, including:

[0140] Step S2431: Obtain the influence coefficient array of the mth message delivery component;

[0141] Step S2432: Based on the representation vector integration component of the mth message delivery component, a product operation is performed on the representation information extraction result of the m-1th message delivery component, the reference array, and the influence coefficient array to obtain an operational product;

[0142] Step S2433: Based on the nonlinear transformation component of the mth message passing component, a nonlinear transformation operation is performed on the operational product to obtain a representation information extraction result of the mth message passing component.

[0143] In step S243, the computer system further extracts representation information from the representation information extraction result of the m-1th message passing component and the reference array through the mth message passing component in the graph recursive network model. This step involves the collaborative work of the representation vector integration component and the nonlinear transformation component.

[0144] In step S2431, the computer system first obtains the influence coefficient array of the m-th message transmission component. This array is a weight array, which contains weight information for guiding the representation vector integration component to perform information integration. These weight information may reflect the relative importance or contribution of different input data (such as the representation information extraction result of the m-1-th message transmission component, the reference array, etc.) in the representation information extraction process.

[0145] The specific form and calculation method of the influence coefficient array may vary depending on the specific implementation of the mth message passing component. For example, in some cases, the influence coefficient array may be a vector with the same dimension as the input data, where each element corresponds to the weight of one dimension of the input data; in other cases, the influence coefficient array may be a matrix or a higher-dimensional tensor to capture the complex relationship between the input data.

[0146] Regardless of the specific form of the influence coefficient array, its calculation may involve a variety of factors. For example, it may be determined based on the output features of the m-1th message passing component, the relationship information in the comparison array, and the prior knowledge or training data of the mth message passing component. In some implementations, the influence coefficient array may be represented by learnable parameters that are optimized during the model training process to maximize the prediction performance.

[0147] In step S2432, the computer system performs a product operation on the representation information extraction result, the comparison array and the influence coefficient array of the m-1th message transmission component based on the representation vector integration component of the mth message transmission component to obtain an operational product. This step is one of the core functions of the representation vector integration component, which integrates information by weighted summing of input data.

[0148] The specific implementation of the representation vector integration component may vary depending on the design of the m-th message passing component. However, a common implementation is to use a weighted sum operation. In this implementation, the representation vector integration component first weights the representation information extraction results of the m-1-th message passing component and the reference array according to the influence coefficient array, and then sums the weighted data to obtain the operational product.

[0149] For example, suppose the output feature of the m-1th message passing component is an N-dimensional vector h (m-1) , the comparison array is an N×N matrix A (representing the relationship between vertices), and the influence coefficient array is an N-dimensional vector w (representing the weight of each vertex). In this case, the representation vector integration component may perform the following operations to calculate the operational product h (m) :

[0150] ;

[0151] The dot multiplication operation represents element-wise multiplication, and the matrix multiplication represents linear transformation. This formula means that the feature vector h of the m-1th layer is firstly transformed by comparing the matrix A. (m-1) Perform a linear transformation, and then perform a weighted summation on the transformed results according to the influence coefficient vector w to obtain the feature vector h of the mth layer (m)This formula is just an example. In practice, the implementation of the representation vector integration component may be more complex. For example, it may contain multiple weighted sum operations, nonlinear activation functions, attention mechanisms, and other components to capture the complex relationships and dynamic changes between input data.

[0152] In step S2433, the computer system performs a nonlinear transformation operation on the operational product based on the nonlinear transformation component of the mth message passing component to obtain the representation information extraction result of the mth message passing component. This step is one of the core functions of the nonlinear transformation component, which enhances the expression ability of the model by introducing nonlinear factors.

[0153] The specific implementation of the nonlinear transformation component may vary depending on the design of the m-th message passing component. However, a common implementation is to use a nonlinear activation function. A nonlinear activation function can map the input data into a nonlinear space, thereby capturing the complex patterns and structures in the data.

[0154] For example, the ReLU (Rectified Linear Unit) activation function commonly used in deep learning is a nonlinear transformation component. The ReLU function sets the negative values ​​in the input data to zero, while keeping the positive values ​​unchanged. This operation can introduce sparsity and alleviate the gradient vanishing problem, thereby improving the training efficiency and generalization ability of the model.

[0155] In addition to the ReLU function, there are many other types of nonlinear activation functions to choose from, such as Sigmoid function, Tanh function, Leaky ReLU function, etc. These functions have different characteristics and applicable scenarios, and can be selected according to specific tasks and data characteristics.

[0156] In step S2433, the nonlinear transformation component may use the operational product as input data and perform a transformation operation through the selected nonlinear activation function. The transformed result is the representation information extraction result of the mth message passing component, which will be passed to the next layer of message passing components for further processing or directly used for subsequent tasks (such as classification, regression, etc.).

[0157] In summary, S2431 to S2433 in step S243 are the key steps for deep feature extraction of the graph recursive network model. By acquiring the influence coefficient array, the product operation of the representation vector integration component, and the nonlinear transformation operation of the nonlinear transformation component, the computer system can gradually extract the deep feature representation of the target traffic directed graph. These feature representations not only contain the local neighbor information and road network structure information of the graph vertices, but may also incorporate a wider range of traffic patterns and dynamic change information.

[0158] As an implementation method, the intervention prediction instruction representation vector includes sub-intervention representation vectors corresponding to each traffic element. Based on this, step S300 performs intervention prediction on the road network data representation vector based on the intervention prediction instruction representation vector to obtain the target road network data representation vector, including:

[0159] Step S310: if the number of sub-intervention representation vectors in the plurality of pre-prediction instruction representation vectors is 1, then the representation information prediction component performs intervention prediction on the road network data representation vector based on the intervention prediction instruction representation vector to obtain a target road network data representation vector;

[0160] Step S320: if there are multiple sub-intervention representation vectors in the pre-prediction instruction representation vectors, the representation information prediction component is used in sequence to perform intervention prediction on the road network data representation vector according to each sub-intervention representation vector to obtain a target road network data representation vector.

[0161] In step S300, the computer system performs intervention prediction on the road network data representation vector based on the intervention prediction instruction representation vector to obtain the target road network data representation vector. This step is the core link in the road network state prediction method, which involves the deep fusion and prediction of the intervention prediction instruction and the road network data.

[0162] In step S310, the computer system checks the number of sub-intervention representation vectors in the intervention prediction instruction representation vector. If the number is 1, that is, the intervention prediction instruction only involves the intervention mode of one traffic element, then the computer system will directly perform intervention prediction through the representation information prediction component. The representation information prediction component is a machine learning model specially designed to handle such prediction tasks. It can take the intervention prediction instruction representation vector and the road network data representation vector as input and output the target road network data representation vector.

[0163] The specific implementation of the representation information prediction component may vary depending on the application scenario and data characteristics. However, a common choice is to use a neural network model, such as a multi-layer perceptron (MLP), a convolutional neural network (CNN), or a recurrent neural network (RNN). These models have strong nonlinear fitting and feature extraction capabilities, and can learn complex mapping relationships from input data.

[0164] Taking a multi-layer perceptron as an example, it may contain multiple hidden layers, each of which is composed of multiple neurons. These neurons process the input data through operations such as weighted summation and activation functions, and transmit information layer by layer. In step S310, the representation information prediction component may first receive the intervention prediction instruction representation vector and the road network data representation vector as input, and perform feature extraction and information integration through the hidden layers of the multi-layer perceptron. Finally, at the output layer, the model may use an appropriate activation function (such as a Softmax function or a Sigmoid function) to map the feature vector to the space of the target road network data representation vector.

[0165] In step S310, since the intervention prediction instruction only involves the intervention method of one traffic element, the output of the representation information prediction component will be directly used as the target road network data representation vector. This vector contains the prediction results of the road network data on the corresponding traffic element under the specified intervention strategy.

[0166] In step S320, if there are multiple sub-intervention representation vectors in the intervention prediction instruction representation vector, that is, the intervention prediction instruction involves intervention methods of multiple traffic elements, then the computer system will perform multiple intervention predictions in sequence through the representation information prediction component. The processing strategy of this step is more complicated than that of step S310 because it needs to consider the mutual influence and superposition effect between multiple sub-intervention representation vectors.

[0167] At the beginning of step S320, the computer system determines the order of prediction according to the order of the sub-intervention characterization vectors. This order may be determined based on factors such as the importance, priority, or user-specified order of the sub-intervention characterization vectors. Then, the computer system will perform intervention predictions in sequence according to the determined order.

[0168] In each intervention prediction, the computer system passes the current sub-intervention representation vector and the road network data representation vector as input to the representation information prediction component. The representation information prediction component will perform feature extraction, information integration and prediction calculation based on these input data, and output an intermediate result. This intermediate result can be regarded as a partial prediction result of the road network data on the corresponding traffic elements under the current sub-intervention strategy.

[0169] After obtaining the first intermediate result, the computer system merges or updates it with the original network data representation vector to reflect the impact of the first sub-intervention strategy. Then, the computer system uses the updated network data representation vector and the next sub-intervention representation vector as input for the second intervention prediction. This process is repeated until all sub-intervention representation vectors have been processed.

[0170] After the last intervention prediction, the computer system will obtain the final target road network data representation vector. This vector contains the comprehensive prediction results of the road network data on multiple traffic elements under all specified intervention strategies. Due to the mutual influence and superposition effect between multiple sub-intervention representation vectors, the final target road network data representation vector may be different from the prediction results under any single sub-intervention strategy.

[0171] In practical applications, the specific implementation methods and prediction strategies of the representation information prediction components may need to be adjusted and optimized according to the data characteristics and task requirements. For example, more complex neural network structures (such as graph convolutional neural networks, graph attention networks, etc.) can be used to capture the spatial dependencies of road network data; ensemble learning methods (such as Bagging, Boosting, etc.) can be used to improve the stability and accuracy of predictions; and more reasonable prediction processes and parameter settings can be designed by combining domain knowledge and expert experience.

[0172] In summary, S310 and S320 in step S300 are key steps for the computer system to perform intervention prediction on the road network data representation vector based on the intervention prediction instruction representation vector. Through the processing and analysis of these two steps, the computer system can generate a target road network data representation vector that meets the user-specified intervention strategy and target traffic element intervention method.

[0173] As an implementation mode, in step S320, the road network data representation vector is intervened and predicted in sequence through the representation information prediction component according to each sub-intervention representation vector to obtain the target road network data representation vector, including:

[0174] Step S321: performing intervention prediction on the road network data representation vector according to the first sub-intervention representation vector through the representation information prediction component to obtain the first road network data representation vector;

[0175] Step S322: traverse n to complete the following operations: perform intervention prediction on the n-1th network data representation vector according to the nth sub-intervention representation vector through the representation information prediction component to obtain the nth network data representation vector; wherein 1<n≤X, X represents the number of sub-intervention representation vectors in the intervention prediction instruction representation vector, and X>1;

[0176] Step S323: taking the Xth road network data representation vector as the target road network data representation vector.

[0177] In step S320, when the intervention prediction instruction representation vector includes multiple sub-intervention representation vectors, the computer system sequentially performs multiple intervention predictions on the road network data representation vector through the representation information prediction component to obtain the final target road network data representation vector. This process involves processing multiple sub-intervention representation vectors one by one and gradually updating the road network data representation vector.

[0178] In step S321, the computer system performs intervention prediction on the road network data representation vector according to the first sub-intervention representation vector through the representation information prediction component. This step is the beginning of multiple intervention predictions, which determines the initial prediction direction and the initial changes of the road network data.

[0179] The representation information prediction component is a machine learning model specifically designed to handle such prediction tasks, which can take the sub-intervention representation vector and the road network data representation vector as input and output a new road network data representation vector. This new vector reflects the changes that may occur in the road network data under the first sub-intervention strategy.

[0180] Taking a multi-layer perceptron (MLP) as an example, the representation information prediction component may include multiple hidden layers, each of which is composed of multiple neurons. In step S321, the input layer of the MLP may receive the first sub-intervention representation vector and the road network data representation vector as input. Then, these input data are processed layer by layer through the hidden layers, and each layer performs nonlinear transformation on the input data through weighted summation, activation function and other operations. Finally, at the output layer, the MLP may use an appropriate activation function (such as a Sigmoid function or a Tanh function) to map the feature vector to the space of the new road network data representation vector.

[0181] Since the first sub-intervention representation vector is the first to be processed, the corresponding road network data representation vector is the original road network data representation vector without any intervention prediction. This vector contains the initial state of the road network data and is the basis for subsequent intervention predictions.

[0182] In step S322, the computer system traverses n (1<n≤X), and sequentially uses the representation information prediction component to perform intervention prediction on the n-1th road network data representation vector according to the nth sub-intervention representation vector. This step is the core link of multiple intervention predictions. It gradually updates the road network data representation vector by gradually superimposing the influence of multiple sub-intervention strategies to reflect more complex traffic state changes.

[0183] In each traversal, the computer system passes the current sub-intervention representation vector and the previous network data representation vector as input to the representation information prediction component. The representation information prediction component will perform feature extraction, information integration and prediction calculation based on these input data, and output a new network data representation vector. This new vector contains the further changes that may occur to the network data under the current sub-intervention strategy.

[0184] In step S322, the road network data representation vector is updated each time the traversal is performed. This means that when processing the nth sub-intervention representation vector, the input road network data representation vector is already the result of the influence of the first n-1 sub-intervention strategies. Therefore, as the traversal proceeds, the road network data representation vector will gradually reflect more complex traffic state changes, including the mutual influence and superposition effect between multiple sub-intervention strategies.

[0185] For example, suppose there are three sub-intervention representation vectors (corresponding to three different traffic element intervention methods), and they need to be processed in sequence. When processing the first sub-intervention representation vector, the first road network data representation vector is obtained; when processing the second sub-intervention representation vector, the first road network data representation vector is used as input to obtain the second road network data representation vector; when processing the third sub-intervention representation vector, the second road network data representation vector is used as input to obtain the final target road network data representation vector.

[0186] In step S323, the computer system uses the Xth road network data representation vector (i.e., the road network data representation vector obtained by the last traversal) as the target road network data representation vector. This step is the end of the multiple intervention prediction, and it outputs the final prediction result.

[0187] The target road network data representation vector contains the comprehensive prediction results of the road network data on multiple traffic elements under all specified sub-intervention strategies. This vector not only reflects the individual impact of each sub-intervention strategy on the road network data, but also reflects the mutual influence and superposition effect between them. Therefore, the target road network data representation vector is a comprehensive and accurate prediction result, which can provide strong support for subsequent traffic management and planning.

[0188] Since the target road network data representation vector is gradually obtained through multiple intervention predictions, it may contain richer information and more complex structures than the original road network data representation vector. This means that in subsequent tasks (such as traffic flow prediction, congestion assessment, etc.), the target road network data representation vector can be used more accurately for analysis and decision-making.

[0189] In summary, S321 to S323 in step S320 are key steps for the computer system to perform multiple intervention predictions on the road network data representation vector in sequence through the representation information prediction component. Through the processing and analysis of these three steps, the computer system can generate a target road network data representation vector that meets the user-specified multiple sub-intervention strategies and target traffic element intervention methods.

[0190] As an implementation manner, in step S300, before performing intervention prediction on the road network data representation vector based on the intervention prediction instruction representation vector to obtain the target road network data representation vector, the method further includes:

[0191] Step S10: acquiring an initial representation information prediction component, a first training database and a second training database, wherein the first training database includes a plurality of traffic network training data representation vectors, and the second training database includes a plurality of intervention instruction representation vectors;

[0192] Step S20: determining a first training error value corresponding to the initial representation information prediction component based on the first training database, and determining a second training error value corresponding to the initial representation information prediction component based on the second training database;

[0193] Step S30: fusing the first training error value and the second training error value to obtain a target training error value, and adjusting the initial representation information prediction component based on the target training error value to obtain the representation information prediction component.

[0194] Prior to step S300, the computer system performs a series of component training processes to ensure that the representation information prediction component can accurately perform intervention prediction on the road network data representation vector based on the intervention prediction instruction representation vector. This process includes obtaining an initial representation information prediction component, a first training database and a second training database, and performing model training and calibration based on these data.

[0195] In step S10, the computer system first obtains an initial characterization information prediction component, a first training database, and a second training database. The initial characterization information prediction component is a machine learning model that has not been fully trained, and it will be used in subsequent intervention prediction tasks. The first training database contains multiple traffic network training data characterization vectors, which are extracted from historical traffic network data and undergo appropriate preprocessing and feature engineering steps to form a data format suitable for model training. Each traffic network training data characterization vector corresponds to a specific traffic network state, including information on various traffic elements such as traffic flow, congestion, and road conditions. The second training database contains multiple intervention instruction characterization vectors, which are obtained by encoding and converting the intervention prediction instructions input by the user. They represent specific traffic element changes or intervention strategy effects that the user wants the system to predict.

[0196] For example, suppose there is an initial representation information prediction component, which is a multi-layer perceptron (MLP) model. The first training database may contain 1,000 traffic network training data representation vectors, each of which is an array of 100 features, covering multiple traffic elements such as traffic flow, average speed, and congestion index. The second training database may contain 500 intervention instruction representation vectors, each of which is an array of 10 features, representing the intervention strategy that the user wants to predict (such as flow control, optimized route, etc.) and its corresponding target traffic elements (such as reduced traffic flow, reduced congestion, etc.).

[0197] In step S20, the computer system determines the first training error value and the second training error value corresponding to the initial representation information prediction component based on the first training database and the second training database. The purpose of this step is to evaluate the performance of the initial representation information prediction component when processing traffic network data and intervention instructions alone, and to quantify its prediction error.

[0198] For the first training database, the computer system takes each traffic network training data representation vector as input, and predicts it through the initial representation information prediction component to obtain the predicted road network data representation vector. Then, the computer system compares the prediction result with the real road network data representation vector, calculates the difference between them (such as mean square error, cross entropy loss, etc.), and obtains the first training error value. This error value reflects the prediction accuracy of the initial representation information prediction component when processing traffic network data.

[0199] For the second training database, the computer system takes each intervention instruction representation vector as input, and predicts it through the initial representation information prediction component to obtain the predicted intervention effect representation vector (although the intervention effect may not be directly output in vector form in actual applications, here, in order to illustrate the training process, it is assumed that there is a corresponding prediction output). Then, the computer system compares the prediction result with the actual intervention effect (or the expected intervention effect representation vector), calculates the difference between them, and obtains the second training error value. This error value reflects the prediction accuracy of the initial representation information prediction component when processing the intervention instruction.

[0200] In step S30, the computer system fuses the first training error value and the second training error value to obtain a target training error value. The fusion process may involve weighted averaging, summing or other forms of combination of the two error values. The fused target training error value comprehensively considers the performance of the initial representation information prediction component in processing traffic network data and intervention instructions, and provides a comprehensive evaluation standard for subsequent model adjustment.

[0201] After obtaining the target training error value, the computer system adjusts the initial representation information prediction component based on the error value. The adjustment process usually involves adjusting the parameters of the model (such as weights and biases), optimizing the model structure (such as adding or reducing hidden layers, changing activation functions, etc.), or using more advanced training algorithms (such as Adam optimizer, learning rate decay, etc.). Through continuous iterative training and parameter adjustment, the computer system gradually reduces the target training error value and improves the prediction accuracy and generalization ability of the initial representation information prediction component.

[0202] During the tuning process, the computer system may also use strategies such as cross-validation and early stopping to prevent overfitting and underfitting. Cross-validation evaluates the stability and generalization ability of the model by dividing the training data into multiple subsets for multiple training and validation. The early stopping method monitors the error changes on the validation set during the training process and stops training in advance when the error no longer decreases significantly to avoid overfitting of the model on the training set.

[0203] Finally, after sufficient training and tuning, the initial representation information prediction component is optimized into a representation information prediction component, which can accurately output the target road network data representation vector under the given intervention prediction instruction representation vector and road network data representation vector input, providing strong support for subsequent traffic management and planning.

[0204] In summary, steps S10 to S30 are key steps in the component training process, which involve multiple links such as data acquisition, error calculation, model adjustment, etc. Through this process, the computer system can ensure that the representation information prediction component has high prediction accuracy and generalization ability when processing complex traffic network data and intervention instructions, providing a solid foundation for subsequent intervention prediction tasks.

[0205] As an implementation manner, step S20, based on the first training database, determines a first training error value corresponding to the initial representation information prediction component, including:

[0206] Step S21: obtaining the training intervention prediction instruction representation vectors corresponding to the respective traffic network training data representation vectors in the first training database, and the priori representation vectors corresponding to the respective traffic network training data representation vectors;

[0207] Step S22: performing intervention prediction on the corresponding traffic network training data representation vectors based on each training intervention prediction instruction representation vector through the initial representation information prediction component, and obtaining the predicted road network data representation vectors corresponding to each traffic network training data representation vector;

[0208] Step S23: for each traffic network training data representation vector, taking the commonality measurement result between the corresponding priori representation vector and the corresponding predicted road network data representation vector as the first model training error value corresponding to the traffic network training data representation vector;

[0209] Step S24: fusing the first model training error values ​​to obtain a first training error value corresponding to the initial representation information prediction component.

[0210] In step S20, the computer system determines the first training error value corresponding to the initial representation information prediction component based on the first training database. This step is a key link in the component training process, which involves multiple sub-steps such as data acquisition, model prediction, and error calculation.

[0211] In step S21, the computer system first obtains the training intervention prediction instruction representation vector and the priori representation vector corresponding to each traffic network training data representation vector from the first training database. These vectors are important components of the training data and provide a basis for subsequent model prediction and error calculation.

[0212] Traffic network training data representation vectors are extracted from historical traffic network data, and they contain information about various traffic elements such as traffic flow, congestion, road conditions, etc. These vectors are the main input data for model training, and they represent the feature representations under different traffic network conditions.

[0213] The training intervention prediction instruction representation vector is a vector representation obtained by encoding and converting the intervention prediction instructions input by the user. These vectors represent the specific traffic element changes or intervention strategy effects that the user wants the system to predict. During the training process, they are used as additional input information of the model to guide the model to learn how to predict changes in the traffic network state based on the intervention instructions.

[0214] Prior representation vectors are known, expected network data representation vectors corresponding to each network training data representation vector. These vectors represent the state that the network should be in without intervention instructions. They are usually obtained through expert knowledge, historical data or other reliable sources, and are used to evaluate the accuracy of model predictions.

[0215] For example, suppose there is a traffic network training data representation vector h1, which represents the traffic status of a certain area in a certain time period. The corresponding training intervention prediction instruction representation vector g1 may indicate that the user wants to predict the change in traffic status after the flow restriction measures are implemented in the area. The prior representation vector p1 may indicate the traffic status that the area should present in the time period if the flow restriction measures are not implemented.

[0216] In step S22, the computer system performs intervention prediction on the corresponding traffic network training data representation vector based on each training intervention prediction instruction representation vector through the initial representation information prediction component. This step is the core link of model prediction, which involves passing the input data (traffic network training data representation vector and training intervention prediction instruction representation vector) to the model and obtaining the prediction result (prediction network data representation vector) through the internal calculation process of the model.

[0217] During the prediction process, the initial representation information prediction component uses its internal network structure and parameters to process and analyze the input data. This process may include multiple computational steps such as forward propagation of multi-layer neural networks, convolution operations, pooling operations, and attention mechanisms. Ultimately, the model outputs a predicted road network data representation vector, which represents the possible state of the traffic network under a given intervention prediction instruction.

[0218] Continuing with the above example, assume that the initial representation information prediction component is a multi-layer perceptron (MLP) model. In step S22, the computer system passes the traffic network training data representation vector h1 and the training intervention prediction instruction representation vector g1 as input to the MLP model. The model will perform layer-by-layer calculations through its hidden layers and ultimately output a predicted network data representation vector h1'. This vector represents the traffic status that the area may present during the time period after the flow restriction measures are implemented.

[0219] In step S23, the computer system uses the commonality measurement result between the corresponding prior representation vector and the predicted road network data representation vector for each traffic network training data representation vector as the first model training error value corresponding to the traffic network training data representation vector. This step is a key step in error calculation, which involves comparing the difference between the model prediction result and the expected result and quantifying the size of the difference.

[0220] The commonality measurement results can be calculated in a variety of ways, such as mean square error (MSE), cross-entropy loss, cosine similarity, etc. These measurement methods can capture different aspects between the predicted results and the expected results (such as numerical differences, probability distribution differences, direction differences, etc.) and give corresponding error values.

[0221] Continuing with the above example, assuming that mean square error is used as the commonality metric, in step S23, the computer system calculates the mean square error between the predicted road network data representation vector h1' and the prior representation vector p1. This error value represents the error magnitude of the MLP model when predicting the traffic state after the flow restriction measures are implemented in the area.

[0222] For each traffic network training data representation vector in the first training database, the computer system will repeat the operations of steps S22 and S23 to obtain the corresponding first model training error value. These error values ​​will be used for subsequent model calibration and performance evaluation.

[0223] In step S24, the computer system fuses the first model training error values ​​corresponding to the characterization vectors of the traffic network training data to obtain the first training error value corresponding to the initial characterization information prediction component. This step is a key step in error fusion, which integrates multiple individual training error values ​​into an overall error value to reflect the performance of the model on the entire training data set.

[0224] The fusion process may involve weighted averaging, summing, or other forms of combination of multiple error values. The specific fusion method used depends on the task requirements and the characteristics of the error distribution. For example, if the differences between the error values ​​are large, they may need to be weighted averaged to balance the importance of different data points; if the differences between the error values ​​are small and evenly distributed, they can be directly summed to obtain the overall error value.

[0225] The first training error value obtained in the end is one of the important indicators for evaluating the performance of the initial representation information prediction component. It reflects the prediction accuracy and generalization ability of the model when processing traffic network data and intervention prediction instructions. In the subsequent model calibration process, the computer system will adjust the parameters and structure of the model based on this error value to improve its prediction performance.

[0226] As an implementation manner, in step S20, determining a second training error value corresponding to the initial representation information prediction component based on the second training database includes:

[0227] Step S25: obtaining standard prediction result representation vectors corresponding to each intervention instruction representation vector in the second training database;

[0228] Step S26: using the initial representation information prediction component, based on each intervention instruction representation vector, the training intervention prediction instruction corresponding to each intervention instruction representation vector is inferred as the representation vector of the road network data after intervention, so as to obtain the inferred representation vector of the road network data after intervention corresponding to each intervention instruction representation vector;

[0229] Step S27: for each intervention instruction representation vector, taking the commonality measurement result between the corresponding reasoned post-intervention road network data representation vector and the corresponding standard prediction result representation vector as the second model training error value corresponding to the intervention instruction representation vector;

[0230] Step S28: adding the training error values ​​of the second models to obtain the second training error value corresponding to the initial representation information prediction component.

[0231] In step S20, in addition to determining the first training error value of the initial representation information prediction component based on the first training database, the computer system also needs to determine the second training error value of the component based on the second training database. This step is another key link in the component training process, which involves multiple sub-steps such as processing the intervention instruction representation vector, model reasoning prediction, and error calculation.

[0232] In step S25, the computer system first obtains the standard prediction result representation vectors corresponding to each intervention instruction representation vector from the second training database. These vectors are important components of the training data and provide a benchmark for subsequent model reasoning and error calculation.

[0233] The intervention instruction representation vector is a vector representation obtained by encoding and converting the intervention prediction instruction input by the user. They represent the specific traffic element changes or intervention strategy effects that the user wants the system to predict. The standard prediction result representation vector is the known and expected post-intervention road network data representation vector corresponding to each intervention instruction representation vector. These vectors represent the state that the traffic network should present under a given intervention instruction, and they are usually obtained through expert knowledge, simulation experiments or other reliable sources.

[0234] For example, suppose there is an intervention instruction representation vector g1, which indicates that the user wants to predict the traffic state change after the flow restriction measures are implemented in a certain area. The corresponding standard prediction result representation vector p1' may indicate the traffic state that the area should present after the flow restriction measures are implemented, including the reduction of vehicle flow and the relief of congestion.

[0235] In step S26, the computer system uses the initial representation information prediction component to infer the representation vector of the road network data after intervention for the corresponding training intervention prediction instruction based on each intervention instruction representation vector. This step is the core link of model reasoning, which involves passing the input data (intervention instruction representation vector) to the model and obtaining the reasoning result (the representation vector of the road network data after intervention) through the internal calculation process of the model.

[0236] During the inference process, the initial representation information prediction component uses its internal network structure and parameters to process and analyze the input data. This process is similar to the prediction process in step S22, but the purpose is different. During the prediction process, the goal of the model is to generate a prediction result that is as close as possible to the actual traffic network state; while during the inference process, the goal of the model is to generate a reasonable representation vector of the post-intervention network data based on the given intervention instructions.

[0237] Continuing with the above example, in step S26, the computer system will pass the intervention instruction representation vector g1 as input to the initial representation information prediction component. The model will generate a reasoned intervention road network data representation vector h1'' through its internal calculation process. This vector represents the possible traffic state of the area after the flow restriction measures are implemented. It is a reasonable speculation made by the model based on the input intervention instruction and the knowledge learned by itself.

[0238] In step S27, for each intervention instruction representation vector, the computer system uses the commonality measurement result between the corresponding reasoned intervention network data representation vector and the standard prediction result representation vector as the second model training error value corresponding to the intervention instruction representation vector. This step is a key step in error calculation, which involves comparing the difference between the model reasoning result and the expected result and quantifying the size of the difference.

[0239] The commonality measurement results can be calculated in a variety of ways, such as mean square error (MSE), cross-entropy loss, cosine similarity, etc. These measurement methods can capture different aspects between the inference results and the expected results (such as numerical differences, probability distribution differences, direction differences, etc.) and give corresponding error values.

[0240] Continuing with the above example, assume that mean square error is used as the commonality measurement method. In step S27, the computer system calculates the mean square error between the road network data representation vector h1'' after the reasoning intervention and the standard prediction result representation vector p1'. This error value represents the error magnitude generated by the initial representation information prediction component when reasoning about the traffic state after the flow restriction measures are implemented in the area.

[0241] For each intervention instruction representation vector in the second training database, the computer system will repeat the operations of steps S26 and S27 to obtain the corresponding second model training error value. These error values ​​will be used for subsequent model adjustment and performance evaluation.

[0242] In step S28, the computer system adds the second model training error values ​​corresponding to each intervention instruction representation vector to obtain the second training error value corresponding to the initial representation information prediction component. This step is a key step in error fusion, which integrates multiple individual training error values ​​into an overall error value to reflect the performance of the model on the entire second training database.

[0243] The addition process is relatively simple and direct, and does not require complex calculations or conversions. The second training error value can be obtained by simply adding the second model training error values ​​corresponding to each intervention instruction representation vector one by one. This error value is one of the important indicators for evaluating the performance of the initial representation information prediction component in processing intervention instructions and performing inference prediction.

[0244] The final second training error value, like the first training error value, is a key indicator for evaluating the performance of the initial representation information prediction component. They reflect the prediction accuracy and reasoning ability of the model when processing traffic network data and intervention instructions, respectively. In the subsequent model calibration process, the computer system will adjust the parameters and structure of the model based on these two error values ​​to improve its overall performance.

[0245] Through the operations of steps S25 to S28, the computer system can comprehensively evaluate the performance of the initial characterization information prediction component when processing the second training database, and provide strong data support for subsequent model adjustment. This process not only helps to improve the prediction accuracy and reasoning ability of the model, but also provides a guarantee for the accuracy and reliability of the entire road network state prediction method.

[0246] Compared with the existing technology, the method of the present application has the following significant advantages: first, through intelligent interaction technology, two-way information flow between users and traffic management systems is realized, so that users can intervene and customize the prediction process according to their actual needs, thereby improving the flexibility and accuracy of the prediction; second, through machine learning technology, in-depth mining and analysis of traffic network data are realized, which can effectively extract key features, build accurate prediction models and process the complex relationships between different traffic elements, thereby improving the accuracy and reliability of prediction results; finally, by combining real-time traffic information sources such as ETC gantry data, real-time monitoring and dynamic prediction of traffic network status are realized, which provides strong support for applications such as real-time traffic monitoring, dynamic traffic scheduling and intelligent traffic signal control of traffic management systems.

[0247] In summary, the road network status prediction method proposed in this application has innovated and improved upon the deficiencies and challenges in the existing technology, aiming to improve the accuracy and flexibility of traffic network data prediction through intelligent interaction and machine learning technology, and to provide more scientific and effective decision support for urban traffic management and planning.

[0248] Figure 2 A hardware entity diagram of a computer system provided in an embodiment of the present application is shown in FIG. Figure 2 As shown, the hardware entity of the computer system 1000 includes: a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program that can be run on the processor 1001, and the processor 1001 implements the steps in the method of any of the above embodiments when executing the program.

[0249] The above is only an implementation method of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application.

Claims

1. A road network status prediction method combining ETC gantry data, characterized in that: The method comprises: Acquire prediction indication information input by a user during intelligent interaction with a traffic management system, wherein the prediction indication information includes traffic network data to be predicted and an intervention prediction instruction for the traffic network data to be predicted; wherein the intervention prediction instruction is used to indicate an intervention method for the traffic network data to be predicted on one or more traffic elements, and the traffic network data to be predicted includes data of multiple ETC gantries in a monitored traffic area; Extracting an instruction representation vector from the intervention prediction instruction to obtain an intervention prediction instruction representation vector, and extracting a road network data representation vector from the traffic road network data to be predicted to obtain a road network data representation vector; Based on the intervention prediction instruction representation vector, the road network data representation vector is intervened and predicted to obtain a target road network data representation vector, and the target road network data representation vector is mapped to restore the road network information to obtain the target traffic road network data corresponding to the traffic road network data to be predicted; wherein the traffic elements presented by the target traffic road network data are generated under the intervention mode; The intervention prediction instruction representation vector includes sub-intervention representation vectors corresponding to each of the traffic elements, and the intervention prediction is performed on the road network data representation vector based on the intervention prediction instruction representation vector to obtain the target road network data representation vector, including: If the number of the sub-intervention representation vectors in the intervention prediction instruction representation vector is 1, performing intervention prediction on the road network data representation vector based on the intervention prediction instruction representation vector through the representation information prediction component to obtain the target road network data representation vector; If there are multiple sub-intervention representation vectors in the intervention prediction instruction representation vector, the representation information prediction component is used in sequence to perform intervention prediction on the road network data representation vector according to each of the sub-intervention representation vectors to obtain the target road network data representation vector; The step of sequentially performing intervention prediction on the road network data representation vector according to each of the sub-intervention representation vectors through the representation information prediction component to obtain the target road network data representation vector includes: By means of the representation information prediction component, intervention prediction is performed on the road network data representation vector according to the first sub-intervention representation vector to obtain a first road network data representation vector; Traverse n to complete the following operations: perform intervention prediction on the n-1th road network data representation vector according to the nth sub-intervention representation vector through the representation information prediction component to obtain the nth road network data representation vector; wherein 1<n≤X, X represents the number of the sub-intervention representation vectors in the intervention prediction instruction representation vector, and X>1; Taking the Xth road network data representation vector as the target road network data representation vector; Before performing intervention prediction on the road network data representation vector based on the intervention prediction instruction representation vector to obtain the target road network data representation vector, the method further includes: Acquire an initial representation information prediction component, a first training database and a second training database, wherein the first training database includes a plurality of traffic network training data representation vectors, and the second training database includes a plurality of intervention instruction representation vectors; Determining a first training error value corresponding to the initial representation information prediction component based on the first training database, and determining a second training error value corresponding to the initial representation information prediction component based on the second training database; fusing the first training error value and the second training error value to obtain a target training error value, and adjusting the initial representation information prediction component based on the target training error value to obtain the representation information prediction component; The determining, based on the first training database, a first training error value corresponding to the initial representation information prediction component includes: Obtaining the training intervention prediction instruction representation vectors respectively corresponding to the traffic network training data representation vectors in the first training database, and the priori representation vectors respectively corresponding to the traffic network training data representation vectors; Through the initial representation information prediction component, based on each of the training intervention prediction instruction representation vectors, the corresponding traffic network training data representation vectors are intervened and predicted to obtain the predicted road network data representation vectors corresponding to each of the traffic network training data representation vectors; For each of the traffic network training data representation vectors, taking the commonality measurement result between the corresponding prior representation vector and the corresponding predicted road network data representation vector as the first model training error value corresponding to the traffic network training data representation vector; The training error values ​​of the first models are integrated to obtain a first training error value corresponding to the initial representation information prediction component; The determining, based on the second training database, a second training error value corresponding to the initial representation information prediction component includes: Obtaining standard prediction result representation vectors respectively corresponding to each of the intervention instruction representation vectors in the second training database; Respectively through the initial representation information prediction component, based on each of the intervention instruction representation vectors, the training intervention prediction instructions corresponding to each of the intervention instruction representation vectors are inferred to represent the road network data after intervention, so as to obtain the inferred road network data after intervention corresponding to each of the intervention instruction representation vectors; For each of the intervention instruction representation vectors, a commonality measurement result between the corresponding inferenced post-intervention road network data representation vector and the corresponding standard prediction result representation vector is used as a second model training error value corresponding to the intervention instruction representation vector; The training error values ​​of each of the second models are added together to obtain a second training error value corresponding to the initial representation information prediction component.

2. The method according to claim 1, characterized in that The intervention prediction instruction includes one or more sub-intervention prediction instructions, and the instruction types of the sub-intervention prediction instructions include label instruction types and text instruction types. The extracting the instruction representation vector of the intervention prediction instruction to obtain the intervention prediction instruction representation vector includes: For each of the sub-intervention prediction instructions, if the instruction type of the sub-intervention prediction instruction is the label instruction type, the sub-intervention prediction instruction is mapped to a label vector to obtain a sub-intervention representation vector corresponding to the sub-intervention prediction instruction; if the instruction type of the sub-intervention prediction instruction is the text instruction type, the sub-intervention prediction instruction is encoded with text semantics to obtain a sub-intervention representation vector corresponding to the sub-intervention prediction instruction; The sub-intervention representation vectors corresponding to each of the sub-intervention prediction instructions are integrated to obtain the intervention prediction instruction representation vector; the sub-intervention prediction instruction corresponds to each of the traffic elements, and the sub-intervention prediction instruction is used to indicate the intervention method of the traffic network data to be predicted on the corresponding traffic element.

3. The method according to claim 1, characterized in that The traffic network data to be predicted includes a plurality of traffic confluences and roads connecting the traffic confluences. The extracting of the network data characterization vector from the traffic network data to be predicted to obtain the network data characterization vector includes: Performing a traffic directed graph transformation on the traffic network data to be predicted, obtaining a target traffic directed graph corresponding to the traffic network data to be predicted, wherein the graph vertices in the target traffic directed graph correspond to the traffic confluences in the traffic network data to be predicted, and the connecting lines in the target traffic directed graph correspond to the roads in the traffic network data to be predicted; wherein the graph vertex representation information of the graph vertices in the target traffic directed graph is used to represent the traffic information of the corresponding traffic confluence; A road network data representation vector is extracted from the target traffic directed graph to obtain the road network data representation vector.

4. The method according to claim 3, characterized in that The road network data representation vector extraction is completed based on a graph recursive network model, the number of components of the message transmission component in the graph recursive network model is X, X>1; the road network data representation vector extraction is performed on the target traffic directed graph to obtain the road network data representation vector, including: Acquire the vertex state array, relationship array and vertex connectivity array corresponding to the target traffic directed graph, and generate a comparison array by combining the relationship array and the vertex connectivity array; wherein the vertex state array is a two-dimensional array generated according to the vertex representation information of each vertex in the target traffic directed graph, the vertex connectivity array is a two-dimensional array generated according to the degree of each vertex in the target traffic directed graph, and the relationship array is a two-dimensional array generated according to the connection between any two vertices in the target traffic directed graph; Extracting representation information from the vertex state array and the reference array through the first message passing component in the graph recursive network model to obtain a representation information extraction result of the first message passing component; Traversing m to complete the following operations: extracting the representation information of the m-1th message passing component and the reference array through the mth message passing component in the graph recursive network model, and obtaining the representation information extraction result of the mth message passing component; wherein 1<m≤X; The characterization information extraction result of the Xth message transmission component is used as the road network data characterization vector.

5. The method according to claim 4, characterized in that The m-th message passing component includes a representation vector integration component and a nonlinear transformation component; the representation information extraction result of the m-1-th message passing component and the reference array is extracted through the m-th message passing component in the graph recursive network model to obtain the representation information extraction result of the m-th message passing component, including: Obtaining an influence coefficient array of the mth message delivery component; Based on the representation vector integration component of the mth message transmission component, performing a product operation on the representation information extraction result of the m-1th message transmission component, the reference array and the influence coefficient array to obtain an operational product; Based on the nonlinear transformation component of the mth message passing component, a nonlinear transformation operation is performed on the operational product to obtain a characterization information extraction result of the mth message passing component.

6. The method according to claim 1, characterized in that The obtaining of prediction indication information input by the user during the process of intelligent interaction with the traffic management system includes: Upon receiving the interaction trigger information input by the user, triggering an interface for intelligent interaction between the user and the traffic management system; The interaction start information is displayed to the user, and the prediction indication information input by the user is received.

7. A computer system comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, wherein: When the processor executes the program, the steps in the method according to any one of claims 1 to 6 are implemented.

Images