Smart city management method and system based on vehicle-road-cloud collaboration

By employing a vehicle-road-cloud collaborative approach, correlation analysis and marginal contribution detection are performed to identify important data streams. Traffic flow prediction is then conducted using a basic inference network, addressing the issues of insufficient data utilization and inaccurate prediction in existing technologies. This approach enables more accurate traffic flow prediction and scientific decision support.

CN118736844BActive Publication Date: 2024-11-12SICHUAN HUATI LIGHTING TECH
View PDF 10 Cites 0 Cited by

Patent Information

Application Number
CN202411225696.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-03
Publication Date
2024-11-12
Estimated Expiration
2044-09-03

AI Technical Summary

Technical Problem

Existing traffic flow prediction methods lack a systematic evaluation mechanism when processing multimodal monitoring data streams, resulting in insufficient data utilization, inaccurate prediction results, and neglect of the dynamic relationships and complex changing characteristics between data streams.

Method used

By using a vehicle-road-cloud collaborative approach, we can obtain vehicle traffic flow evolution streams and multimodal monitoring data streams in the target monitoring area, perform correlation analysis and marginal contribution detection, screen out important data streams, and use a basic inference network for inference to obtain more accurate vehicle traffic flow predictions.

Benefits of technology

It improves the reliability and accuracy of traffic flow forecasting, provides scientific decision support, and provides a basis for developing traffic intervention strategies for smart city management.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118736844B_ABST
    Figure CN118736844B_ABST
Patent Text Reader

Abstract

The present application provides a smart city management method and system based on vehicle-road-cloud collaboration. Based on correlation analysis and marginal contribution evaluation, the obtained vehicle-road multimodal monitoring data stream is filtered based on the marginal contribution standard to make the obtained target vehicle-road multimodal monitoring data stream more reliable when using the basic reasoning network. At the same time, the influence of each target vehicle-road multimodal monitoring data stream on the vehicle flow evolution flow of past cycles is characterized, so that the reasons for determining the target vehicle-road multimodal monitoring data stream are more sufficient, and the inferred vehicle flow evolution flow obtained by reasoning is more accurate and reliable, which is conducive to the accurate formulation of traffic intervention strategies in smart management.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of data processing, and in particular to a smart city management method and system based on vehicle-road-cloud collaboration. Background Art

[0002] With the acceleration of urbanization, problems such as traffic congestion, environmental pollution and traffic accidents are becoming increasingly severe, posing a major challenge to the sustainable development of cities. As one of the effective ways to solve these problems, smart cities are gradually becoming the mainstream trend of global urban development. In smart city management, accurate prediction and effective regulation of traffic flow are key links to improve urban traffic operation efficiency, reduce congestion and ensure safety. Existing traffic flow prediction methods often lack a systematic evaluation mechanism when processing multimodal monitoring data streams, resulting in insufficient data utilization and inaccurate prediction results. Although some methods try to optimize data input through feature selection or dimensionality reduction techniques, they usually rely on empirical judgment or heuristic algorithms and lack objective evaluation criteria and scientific basis. In addition, these methods often ignore the dynamic relationship and mutual influence between data streams, making it difficult to fully capture the complex and changing characteristics of traffic flow. Summary of the invention

[0003] In view of this, the present application provides a smart city management method and system based on vehicle-road-cloud collaboration. The technical solution of the present application is implemented as follows:

[0004] On the one hand, an embodiment of the present application provides a smart city management method based on vehicle-road-cloud collaboration, the method comprising: obtaining a past period vehicle flow evolution stream and a plurality of vehicle-road multimodal monitoring data streams in a target monitoring area, wherein the unit root test statistic of the vehicle-road multimodal monitoring data stream is less than a statistical reference value; performing correlation analysis on each of the past period vehicle flow evolution streams and each of the vehicle-road multimodal monitoring data streams to obtain a correlation analysis result; adding each of the vehicle-road multimodal monitoring data streams to an important data stream set or a secondary data stream set according to the correlation analysis result. flow set; according to the important data flow set, the secondary data flow set and the past period vehicle flow evolution flow, perform marginal contribution detection on each of the vehicle-road multimodal monitoring data flows in the secondary data flow set to obtain a marginal contribution detection result; use the vehicle-road multimodal monitoring data flow whose marginal contribution detection result meets the marginal contribution standard as the target vehicle-road multimodal monitoring data flow; wherein the target vehicle-road multimodal monitoring data flow is used to input multiple basic inference networks for inference to obtain an inferred vehicle flow evolution flow corresponding to the past period vehicle flow evolution flow.

[0005] 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.

[0006] The beneficial effects of the present application include at least: the method provided by the present application, based on correlation analysis and marginal contribution evaluation, filters the obtained vehicle-road multimodal monitoring data stream based on the marginal contribution standard, so that the obtained target vehicle-road multimodal monitoring data stream is more reliable when using the basic reasoning network. At the same time, it characterizes the influence of each target vehicle-road multimodal monitoring data stream on the vehicle flow evolution flow of past cycles, so that the reasons for determining the target vehicle-road multimodal monitoring data stream are more sufficient, and the inferred vehicle flow evolution flow obtained by reasoning is more accurate and reliable, which is conducive to the accurate formulation of traffic intervention strategies in smart management. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] 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.

[0008] Figure 1 A schematic diagram of the implementation process of a smart city management method based on vehicle-road-cloud collaboration provided in an embodiment of the present application.

[0009] Figure 2 A hardware entity diagram of a computer system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0010] 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.

[0011] The embodiment of the present application provides a smart city management method based on vehicle-road-cloud collaboration, 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, and a desktop computer.

[0012] Figure 1 A schematic diagram of the implementation process of a smart city management method based on vehicle-road-cloud collaboration provided in an embodiment of the present application is shown in FIG. Figure 1 As shown, the method includes:

[0013] Step S100: Obtain the past period vehicle flow evolution flow and multiple vehicle-road multimodal monitoring data streams of the target monitoring area, wherein the vehicle-road multimodal monitoring data stream is the control variable of the past period vehicle flow evolution flow, and the unit root test statistic of the vehicle-road multimodal monitoring data stream is less than the statistical reference value.

[0014] In the embodiment of the present application, we first focus on a specific target monitoring area, such as a busy commercial district or transportation hub in a city. This area may have a large flow of people and complex traffic conditions, and is a key target for intelligent traffic management. Vehicle flow data over a period of time (such as the past month) is collected by installing traffic flow monitoring equipment (such as geomagnetic sensors, cameras, etc.) in the monitoring area. These data record the passage of vehicles at different time periods every day, forming a vehicle flow time series. This sequence is the evolution flow of vehicle flow in the past cycle, which intuitively shows the evolution trend of vehicle flow during the time period, such as a significant increase in traffic during peak hours in the morning and evening, and relatively less traffic during the midnight period.

[0015] In order to more fully understand the changes in vehicle flow and the reasons behind them, the computer system not only focuses on the vehicle flow itself, but also widely collects multimodal monitoring data streams from the vehicle side and the road side. These data streams include but are not limited to: vehicle-side data: such as GPS trajectory data, which records the vehicle's location, speed, driving direction and other information. By analyzing these data, you can understand the characteristics of the vehicle's driving path, speed changes, etc.; road-side data: such as traffic light status, road congestion index, weather conditions (obtained through roadside weather stations), etc. These data reflect the impact of road conditions and environmental factors on traffic flow.

[0016] After collecting these data, the computer system performs a unit root test on the vehicle-road multimodal monitoring data stream. Commonly used test methods include ADF and KPSS. The purpose is to determine whether the data series is stationary, that is, whether its statistical characteristics (such as mean and variance) change over time. Stationarity is an important prerequisite for time series analysis, because non-stationary data may lead to inaccurate model prediction results. Taking ADF as an example, the calculation of its statistic t is based on the following regression model:

[0017] ;

[0018] Among them, ∆ represents the first-order difference, y t is the time series to be tested, p is the lag order, ε tis the error term. The null hypothesis H0:γ=0 (i.e., there is a unit root, non-stationary), and the alternative hypothesis H1:γ<0 (stationary). The result is a statistic value and a corresponding p-value. When the p-value is less than the significance level (such as 0.05), the null hypothesis is rejected and it is considered to be stationary. If the unit root test statistic (ADF or KPSS value) of the vehicle-road multimodal monitoring data stream is less than the preset statistical reference value, it indicates that these data streams are stationary and suitable for subsequent correlation analysis and model construction.

[0019] Step S200: performing correlation analysis on each of the past period vehicle flow evolution streams and each of the vehicle-road multimodal monitoring data streams to obtain a correlation analysis result.

[0020] Specifically, the computer system compares and analyzes the multimodal monitoring data stream of each road (such as vehicle GPS trajectory data, road congestion index, weather conditions, etc.) with the evolution of vehicle flow in the past cycle. This analysis is not limited to simple linear correlation, but also considers more complex time series dynamic relationships, because changes in traffic flow are often affected by multiple factors in a non-instantaneous manner.

[0021] In order to perform such correlation analysis, the computer system can use a variety of statistical methods and machine learning techniques. For example, methods such as impulse response function analysis, error correction model, causal inference model, etc. can be used to analyze whether the vehicle-road monitoring data stream is helpful in predicting the future value of the vehicle flow evolution stream.

[0022] In addition, computer systems can also use correlation coefficient-based analysis (such as Pearson correlation coefficient, Spearman rank correlation coefficient, etc.) to quantify the degree of linear or nonlinear correlation between different data streams. The calculation formulas of these correlation coefficients provide quantitative indicators for understanding the relationship between data. For example, the calculation formula of Pearson correlation coefficient r is:

[0023] ;

[0024] Among them, x i and i are the observation values ​​of the two time series at time point i, and are the means of the two time series, and n is the number of observations. In the scenario of traffic flow prediction, more advanced machine learning models can also be used to capture these relationships, such as time series prediction models (such as ARIMA, SARIMA), neural networks (such as LSTM, GRU), or ensemble learning methods (such as random forests, gradient boosting trees). These models can predict future vehicle flow by learning patterns in historical data, while considering multiple vehicle-road multimodal monitoring data streams as input features. Take LSTM (Long Short-Term Memory Network) as an example, it is a neural network structure that is particularly suitable for processing sequence data and can capture long-term dependencies. In the application of smart city management, the LSTM network can receive the evolution stream of vehicle flow in the past cycle and multiple vehicle-road multimodal monitoring data streams as input sequences, update and forget information through its internal "gate" mechanism (forget gate, input gate, output gate), and finally output a prediction of future vehicle flow.

[0025] By executing step S200, the computer system can identify key factors that have a significant impact on vehicle flow changes, providing an important basis for subsequent model building and forecasting analysis. This correlation analysis not only deepens the understanding of the dynamic changes in traffic flow, but also provides data support for the formulation of more effective traffic management strategies.

[0026] Step S300: adding each of the vehicle-road multimodal monitoring data streams to an important data stream set or a secondary data stream set according to the correlation analysis result.

[0027] In step S300, the correlation indicators between each vehicle-road multimodal monitoring data stream and the vehicle flow evolution stream of the past cycle calculated in step S200 are reviewed. These indicators may include correlation coefficients, statistical significance levels of causal tests, etc. Then, the computer system determines which data streams should be considered important and which are secondary based on a preset threshold or an adaptive method based on the characteristics of the data itself. Taking the correlation coefficient as an example, the computer system can set a positive correlation coefficient threshold (such as 0.5). Only when the correlation coefficient between the data stream and the vehicle flow evolution stream exceeds this threshold, the data stream will be considered to have a significant impact on the prediction and thus be added to the set of important data streams. Conversely, if the correlation coefficient is lower than the threshold, the data stream will be classified as a secondary data stream and may not be given priority in the subsequent prediction model or its weight will be reduced. This classification process is not static, and it can be dynamically adjusted based on new data input and prediction result feedback. For example, if a minor data stream performs well in subsequent data analysis or model training and its influence gradually increases, the computer system can re-evaluate its importance and move it from the set of minor data streams to the set of important data streams. In addition, for some complex application scenarios, the computer system can also use more advanced machine learning techniques to assist in classification decisions. For example, feature selection algorithms (such as model-based feature selection, filter-based feature selection, etc.) can be used to automatically identify the most important features for the prediction task (here, the vehicle-road multimodal monitoring data stream). These algorithms usually work by evaluating the contribution of each feature to the model performance (such as accuracy, F1 score, etc.), thereby helping the computer system to more accurately construct a set of important data streams.

[0028] Specifically, assuming that the computer system adopts a model-based feature selection method, it can first use all the collected vehicle-road multimodal monitoring data streams to train a preliminary prediction model (such as linear regression, random forest or neural network). Then, the contribution of each data stream to the model performance is evaluated through cross-validation and other techniques. Finally, according to the contribution ranking, the data stream with the largest contribution is added to the important data stream set, and the rest are classified into the secondary data stream set.

[0029] It should be noted that the specific implementation details of this classification process (such as threshold setting, feature selection algorithm selection, etc.) will vary according to the actual application scenario, data characteristics, and prediction goals. Therefore, when the computer system is actually deployed, it needs to be adjusted and optimized according to the specific situation.

[0030] Step S400: Based on the important data stream set, the secondary data stream set and the past period vehicle flow evolution stream, perform marginal contribution detection on each of the vehicle-road multimodal monitoring data streams in the secondary data stream set to obtain a marginal contribution detection result.

[0031] Step S400 is a process of evaluating the marginal contribution of each vehicle-road multimodal monitoring data stream in the set of secondary data streams to the past cycle vehicle flow evolution flow prediction model. The computer system performs marginal contribution detection (or incremental test) to identify those secondary data streams that can significantly improve the prediction accuracy based on the consideration of important data streams.

[0032] First, based on the results of step S300, the computer system has identified which data streams belong to the important data stream set and have a significant impact on the evolution of vehicle traffic, and which belong to the secondary data stream set and whose direct contribution may not be so prominent. In step S400, the system will test each data stream in the secondary data stream set one by one to quantify the additional prediction capability improvement it can bring after adding it to the prediction model.

[0033] The specific implementation of marginal contribution detection may involve a variety of statistical methods and machine learning techniques. A typical example is to use cross-validation (such as K-fold cross-validation) to evaluate the incremental effect of data streams. In this process, the computer system first uses a set of important data streams and the evolution of vehicle traffic in past cycles as the input of the baseline model to train and evaluate the prediction performance of the model (such as mean square error MSE, coefficient of determination R², etc.). Then, for each data stream in the set of secondary data streams, the computer system adds it to the input of the baseline model, retrains the model, and evaluates the prediction performance again. By comparing the changes in model performance before and after the addition, the marginal contribution of the data stream can be calculated. This change can be quantified by the amount of improvement in performance indicators, such as the reduction in MSE or the increase in R².

[0034] In order to more accurately evaluate the marginal contribution, the computer system can also use more complex model selection criteria, such as Akaike Information Criterion (AIC) or Bayesian Information Criterion (BIC). These criteria not only consider the goodness of fit of the model, but also the complexity of the model, which helps to avoid overfitting and more accurately reflect the true contribution of the data stream. In the calculation process, if MSE is used as a performance indicator, its calculation formula is:

[0035] ;

[0036] Where n is the number of samples, y i is the actual observed value, is the model prediction value. By comparing the changes in MSE before and after adding the secondary data stream, we can intuitively see the degree of improvement of the model prediction performance by the data stream. Finally, the computer system screens out those secondary data streams that can significantly improve the prediction accuracy based on the marginal contribution test results. These data streams may contain some information that is not significant when analyzed alone, but can play an important role when used in combination with other data streams. Through this step, the computer system can further optimize the input feature set of the prediction model, improve the accuracy and reliability of the prediction, and provide more scientific decision-making support for smart city management.

[0037] Step S500: The vehicle-road multimodal monitoring data stream whose marginal contribution detection result meets the marginal contribution standard is used as the target vehicle-road multimodal monitoring data stream; wherein the target vehicle-road multimodal monitoring data stream is used to input multiple basic inference networks for inference to obtain an inferred vehicle flow evolution flow corresponding to the vehicle flow evolution flow of the past period.

[0038] In step S500, the computer system selects those vehicle-road multimodal monitoring data streams whose marginal contributions meet the preset standards from the secondary data stream set according to the marginal contribution detection results. These selected data streams are called target vehicle-road multimodal monitoring data streams. These target data streams will then be used as inputs to multiple basic reasoning networks to derive the inferred vehicle flow evolution flow corresponding to the past period vehicle flow evolution flow, that is, the prediction of future vehicle flow trends.

[0039] Specifically, the computer system first sets a set of marginal contribution standards, which may be based on statistical significance, percentage improvement in predictive performance, or other quantitative indicators. For example, a threshold can be set, and only when the data stream is added to the model and the prediction accuracy is improved by more than a certain percentage (such as 5%), its marginal contribution is considered to meet the standard. The computer system determines which data streams should be selected as target data streams by comparing the marginal contribution test results of each secondary data stream with these standards. Once the target data stream is determined, the computer system inputs these data and the data in the important data stream set into multiple pre-built basic reasoning networks. These basic reasoning networks may include different types of machine learning models, such as linear regression models, support vector machines (SVMs), random forests, gradient boosting trees (GBDTs), and deep learning models such as long short-term memory networks (LSTMs). Each model has its own unique advantages and applicable scenarios. By combining these models, their respective strengths can be fully utilized to improve the accuracy and robustness of predictions.

[0040] During the reasoning process, these basic reasoning networks will learn the inherent laws and trends of the vehicle flow evolution flow based on the input target vehicle-road multimodal monitoring data stream and the data in the important data stream set, and predict the future traffic flow situation accordingly. The reasoning result will form an inferred vehicle flow evolution flow, which not only reflects the evolution pattern of vehicle flow in the past cycle, but also predicts the trend of vehicle flow in the future.

[0041] It should be noted that different models may have different inference results for the same set of data. Therefore, after multiple basic inference networks output their own inference results, the computer system usually uses ensemble learning methods (such as voting, averaging, or weighted averaging) to fuse these results to obtain a more stable and accurate final prediction. This process helps to improve the reliability and generalization ability of the prediction results.

[0042] In summary, step S500 selects the multimodal monitoring data stream of vehicles and roads with significant marginal contributions as the target data stream, and inputs it into multiple basic reasoning networks for reasoning analysis, and finally obtains an accurate prediction of future vehicle flow trends. This process not only fully utilizes the information value of multi-source monitoring data, but also improves the accuracy and robustness of the prediction by integrating the prediction results of multiple models, providing strong data support for smart city management.

[0043] In one implementation, the step S200 of performing correlation analysis on each of the past period vehicle flow evolution streams and each of the vehicle-road multimodal monitoring data streams may specifically include:

[0044] For any of the vehicle-road multimodal monitoring data streams, complete the following steps:

[0045] Step S210: Determine the first time point corresponding to the past period vehicle flow evolution flow based on the set of observation values ​​of the past a time points of the vehicle-road multimodal monitoring data stream and the set of observation values ​​of the past b time points of the past period vehicle flow evolution flow; wherein the set of observation values ​​of the past a time points includes the values ​​of all time points in the interval from time point a to the current time point of the vehicle-road multimodal monitoring data stream, and the set of observation values ​​of the past b time points includes the values ​​of all time points in the interval from time point b to the current time point of the past period vehicle flow evolution flow, the time point a and the time point b are any time point before the current time point, and the first time point is any time point after the current time point.

[0046] As an implementation method, the first time point can be determined by: weighted summing each value in the set of observation values ​​of the past a time points of the vehicle-road multimodal monitoring data stream to obtain a first weighted summation result; weighted summing each value in the set of observation values ​​of the past b time points of the vehicle flow evolution stream of the past period to obtain a second weighted summation result; adding the first weighted summation result, the second weighted summation result, the second coordination parameter and the first coordination parameter to obtain the first time point corresponding to the vehicle flow evolution stream of the past period.

[0047] Step S210 aims to predict the vehicle flow situation at a certain moment in the future (i.e., the first time point) by integrating the vehicle-road multimodal monitoring data stream and the historical data of the vehicle flow evolution stream in the past cycles.

[0048] First, the computer system collects vehicle-road multimodal monitoring data streams from multiple data sources. These data streams may include changes in the state of traffic lights, road congestion index, weather conditions, etc., which reflect various external factors that affect vehicle flow. At the same time, the system also collects data on the evolution of vehicle flow in past cycles, that is, time series data on vehicle flow in the past period of time. In step S210, the computer system focuses on two key time series sets: one is the set of observations of the vehicle-road multimodal monitoring data stream at the past a time points, where "a time points" refers to all observations from an earlier time point a to the current time point (excluding the current time point). These observations constitute a sequence, which can be regarded as a historical window of the vehicle-road multimodal monitoring data stream. The second is the set of observations of the past b time points of the evolution of vehicle flow in past cycles. Similarly, this is also a sequence of vehicle flow observations from the historical time point b to the current time point (excluding the current time point).

[0049] In order to predict the vehicle flow at a certain moment in the future (the first time point), the computer system needs to process these historical data. Specifically, the system first performs weighted summation on the observations in the two time series sets respectively. For the vehicle-road multimodal monitoring data stream, the computer system assigns different weights to each observation according to its importance and timeliness (these weights can be pre-set based on domain knowledge or learned through machine learning algorithms), and then adds the weighted observations to obtain the first weighted summation result. Similarly, for the vehicle flow evolution stream of past cycles, the system also performs a weighted summation operation to obtain the second weighted summation result. Next, the computer system combines these two weighted summation results with two coordination parameters (the first coordination parameter and the second coordination parameter). The role of the coordination parameters is to fine-tune the weighted summation results to more accurately reflect the impact of historical data on future predictions. The values ​​of these coordination parameters can be empirical values ​​obtained based on historical data analysis, or they can be dynamically adjusted through optimization algorithms.

[0050] Finally, the system adds the first weighted summation result, the second weighted summation result and the two coordination parameters, and the result is the predicted value of the vehicle flow evolution flow of the past cycle at the first time point. This value reflects the expected situation of vehicle flow at a certain moment in the future, considering the influence of the multimodal monitoring data flow of the vehicle and the road. The specific formula for determining the first time point can be expressed as:

[0051] ;

[0052] Among them, Modal_Data i represents the i-th observation value of the vehicle-road multimodal monitoring data stream at the past a time points, w 1i is the corresponding weight; Traffic_Data j represents the jth observation value of the vehicle flow evolution flow in the past period at the past b time points, w 2j is the corresponding weight; 1 and λ 2 They are the first coordination parameter and the second coordination parameter respectively.

[0053] Step S220: determining a second time point corresponding to the past periodic vehicle flow evolution flow according to the observation value set of the past b time points of the past periodic vehicle flow evolution flow, wherein the second time point is consistent with the time point corresponding to the first time point;

[0054] The second time point may be determined by: performing weighted summation on each value in the set of observation values ​​of the past b time points of the past periodic vehicle flow evolution flow according to the corresponding weights, to obtain a third weighted summation result; adding the third weighted summation result, the first coordination parameter, and the second coordination parameter, to obtain the second time point corresponding to the past periodic vehicle flow evolution flow;

[0055] Specifically, in step S220, the computer system first retrieves from the database a set of observation values ​​of the past b time points of the past periodic vehicle flow evolution flow. This set contains the vehicle flow observation data from time point b (i.e., an earlier historical moment) to the current time point (excluding the current time point). These data reflect the evolution trend of vehicle flow over the past period of time.

[0056] Next, in order to predict the vehicle flow at a certain moment in the future, the computer system performs weighted summation on these historical observations. Unlike step S210, only the data of the vehicle flow evolution flow itself in the past period is considered here, and the multimodal monitoring data stream of the vehicle and road is not involved. In the process of weighted summation, each observation value will obtain a weight according to its corresponding time point. These weights can be set according to various factors such as the distance of the observation value from the current time point, the periodic law of flow change, holiday effects, etc. For example, more recent observations may obtain higher weights because they can better reflect the recent flow change trend.

[0057] The result of the weighted summation is the third weighted summation result, which represents a preliminary prediction of the vehicle flow at a certain moment in the future based only on the historical data of the vehicle flow evolution flow of the past period. However, this preliminary prediction may need further adjustment to more accurately reflect the actual situation. Therefore, the computer system combines the third weighted summation result with two coordination parameters (the first coordination parameter and the second coordination parameter). These two coordination parameters can be regarded as fine-tuning factors for the preliminary prediction results, which can be set based on the results of historical data analysis, domain knowledge or machine learning algorithms. Finally, the third weighted summation result is added to the two coordination parameters, and the result obtained is the predicted value of the second time point corresponding to the vehicle flow evolution flow of the past period. This predicted value reflects the expected situation of vehicle flow at a certain moment in the future without considering the influence of the multimodal monitoring data flow of the vehicle and the road.

[0058] The specific formula for determining the second time point can be expressed as:

[0059] ;

[0060] Traffic_Data j represents the jth observation value of the vehicle flow evolution flow in the past period at the past b time points, w 3jis the corresponding weight; 1 and λ 2 They are the first coordination parameter and the second coordination parameter respectively.

[0061] Step S230: performing variance consistency detection on the first time point and the second time point, and obtaining a variance consistency detection result as a correlation analysis result.

[0062] Step S230 evaluates the strength of correlation between the vehicle-road multimodal monitoring data stream and the vehicle flow evolution stream of past cycles by performing variance consistency detection on the two predicted time points (i.e., the first time point and the second time point) obtained in steps S210 and S220.

[0063] Specifically, in step S210, the vehicle flow at a certain moment in the future (the first time point) has been predicted based on the historical data of the vehicle-road multimodal monitoring data stream and the evolution flow of vehicle flow in past cycles. In step S220, the system predicts the vehicle flow at the same future moment (the second time point) based only on the historical data of the evolution flow of vehicle flow in past cycles. Although these two predicted values ​​are for the same time point, their prediction results may be different due to the different input data (one is multimodal data and the other is single flow data).

[0064] In order to quantify this difference and evaluate the contribution of the vehicle-road multimodal monitoring data stream to the prediction results, the computer system uses a variance consistency test method, such as a variance ratio test. The test results will be output as the results of the correlation analysis. If the variance consistency test result is close to 1, it means that the variability of the two predicted values ​​is similar, suggesting that the vehicle-road multimodal monitoring data stream has a significant impact on the prediction of the vehicle flow evolution flow in the past cycle; conversely, if the variance consistency test result deviates far from 1, it may indicate that the vehicle-road multimodal monitoring data stream has limited contribution to the prediction.

[0065] Through the variance consistency test in step S230, the computer system can objectively evaluate the correlation strength between the vehicle-road multimodal monitoring data stream and the vehicle flow evolution stream in past cycles, providing a scientific basis for subsequent model construction and feature selection.

[0066] In one implementation, the step S300, adding each of the vehicle-road multimodal monitoring data streams to the important data stream set or the secondary data stream set according to the correlation analysis result, may specifically include:

[0067] Step S310: if the correlation analysis result is not less than the correlation analysis reference value, adding the vehicle-road multimodal monitoring data stream to the important data stream set;

[0068] Step S320: If the correlation analysis result is less than the correlation analysis reference value, the vehicle-road multimodal monitoring data stream is added to the secondary data stream set.

[0069] In step S300, the computer system evaluates each vehicle-road multimodal monitoring data stream to determine its contribution to the prediction of the vehicle flow evolution flow in the past cycle. This judgment is based on the correlation analysis results obtained by the variance consistency test in step S200. Based on these results, the data streams are divided into two categories: an important data stream set and a secondary data stream set. The important data stream set contains data streams that have a significant impact on vehicle flow prediction, while the secondary data stream set contains data streams with relatively less impact.

[0070] In step S310, the computer system judges the correlation analysis results of each vehicle-road multimodal monitoring data stream. Specifically, the computer system sets a correlation analysis reference value (usually based on historical data analysis, domain knowledge or expert opinions) as the boundary for distinguishing important data streams from secondary data streams. When the correlation analysis result of a data stream is not less than this reference value, the system believes that the data stream contributes significantly to vehicle flow prediction, so it is added to the important data stream set. For example, assume that a variety of vehicle-road multimodal monitoring data streams are collected, including traffic light status, road congestion index, weather conditions, vehicle GPS trajectories, etc. In step S200, the correlation between these data streams and the vehicle flow evolution streams of past cycles has been evaluated through variance consistency detection. Now, these data streams need to be classified into important or secondary sets.

[0071] Taking the traffic light status data stream as an example, if its correlation analysis result shows a high degree of consistency with the vehicle flow evolution stream, and the result is not less than the set correlation analysis reference value, then the system will determine that the data stream has an important impact on the vehicle flow prediction. Therefore, in step S310, the traffic light status data stream will be added to the important data stream set. This means that in the subsequent prediction model construction, this data stream will be given priority to more accurately predict future changes in vehicle flow.

[0072] Corresponding to step S310, step S320 is responsible for processing those vehicle-road multimodal monitoring data streams whose correlation analysis results are less than the correlation analysis reference value. Although these data streams contribute relatively little to vehicle flow prediction, they may still have certain value in certain specific cases or when used in combination with other data streams. Therefore, the system does not completely ignore these data streams, but classifies them into a set of secondary data streams so that they can be used as alternative features in subsequent analysis or model optimization.

[0073] Assume that in step S200, the system evaluates the correlation between the vehicle GPS trajectory data stream and the vehicle flow evolution stream of the past period, but the result shows that the correlation is not strong (for example, the variance ratio deviates far from 1). If this result is less than the set correlation analysis reference value, the system will classify the data stream into the secondary data stream set. However, it should be noted that although the vehicle GPS trajectory data stream may have limited contribution when used alone, it can produce unexpected synergistic effects when combined with other data streams (such as traffic light status, road congestion index, etc.). Therefore, in the subsequent analysis or model optimization stage, the system will still retain these data streams as potential input features for further exploration and utilization when necessary. It is worth noting that the classification of important data streams and secondary data streams is not static. With the continuous collection of new data and the continuous optimization of analysis methods, data streams that were originally considered secondary can gradually show their importance; vice versa. Therefore, in the actual application of smart cities, computer systems need to regularly review and update the data stream classification results to ensure the accuracy and timeliness of the prediction model.

[0074] In addition, the setting of correlation analysis reference values ​​in different application scenarios may also be different. For example, the key factors affecting vehicle flow may be different during urban traffic peak hours and non-peak hours. Therefore, when setting correlation analysis reference values, the system needs to fully consider the specificity of the application scenario and formulate reasonable thresholds based on historical data analysis results and domain knowledge.

[0075] Once the data streams are classified into important or minor sets, they can play their respective roles in the subsequent prediction model construction. The data streams in the important data stream set will be given priority as core input features; while the data streams in the minor data stream set can be explored and utilized as alternative features in the model optimization stage. In this way, the computer system can maximize the information value of multi-source monitoring data while ensuring the accuracy of prediction.

[0076] By scientifically and rationally classifying vehicle-road multimodal monitoring data streams into important and secondary sets, computer systems can optimize resource allocation and improve the accuracy and efficiency of prediction models.

[0077] In one embodiment, after adding each of the vehicle-road multimodal monitoring data streams to the important data stream set or the secondary data stream set according to the correlation analysis result in step S300, the method may further include:

[0078] Step S301: adding the vehicle-road multimodal monitoring data stream of the same type in the important data stream set to the data stream set corresponding to the type;

[0079] Step S302: Add the vehicle-road multimodal monitoring data stream of the same type in the secondary data stream set to the data stream set corresponding to the type; wherein the data stream set includes the vehicle-road multimodal monitoring data stream in the important data stream set and the vehicle-road multimodal monitoring data stream in the secondary data stream set.

[0080] After step S300, the computer system will further classify each vehicle-road multimodal monitoring data stream in the important data stream set. This classification is based on the source or nature of the data stream, which can usually be divided into two categories: vehicle-side data stream and road-side data stream. The vehicle-side data stream mainly comes from the vehicle itself, such as the vehicle's GPS location information, speed, acceleration, etc.; while the road-side data stream mainly comes from road infrastructure, such as traffic light status, road congestion, road conditions, etc.

[0081] For each type of data flow, a corresponding data flow set is created, and all data flows belonging to that type in the important data flow set are added to the set. This is done to facilitate subsequent data management and analysis, so that the system can process and utilize the data more efficiently.

[0082] For example, suppose that in a smart city's traffic management system, the important data stream set includes a variety of vehicle-road multimodal monitoring data streams. Among them, there is a GPS trajectory data stream from the vehicle side, which records the vehicle's driving path and speed changes; there is a traffic light status data stream from the road side, which reflects the real-time status of the traffic light; and there is a road congestion index data stream, which provides a quantitative indicator of road congestion.

[0083] In step S301, the computer will first identify the types of these data streams. The GPS trajectory data stream is classified as a vehicle-side data stream because it comes from the vehicle itself; while the traffic light status data stream and the road congestion index data stream are derived from the road infrastructure and are therefore classified as road-side data streams.

[0084] Subsequently, the computer will create corresponding data stream sets for the vehicle-side data stream and the road-side data stream, and add the data streams belonging to these two categories in the important data stream set to the two sets respectively. For example, the GPS trajectory data streams of all vehicles will be added to the vehicle-side data stream set, and the traffic light status and road congestion index data streams of all roads will be added to the road-side data stream set.

[0085] Similar to step S301, step S302 also involves the classification and detailed classification of the types of vehicle-road multimodal monitoring data streams in the secondary data stream set. The computer system also divides the data streams into different types such as vehicle-side data streams and road-side data streams according to their sources or properties, and creates a corresponding data stream set for each type. Then, all data streams belonging to the type in the secondary data stream set are added to the corresponding set.

[0086] The secondary data stream set may include other factors that may affect vehicle flow, such as weather condition data streams (such as temperature, humidity, precipitation, etc.), holiday information data streams (such as holiday names, dates, etc.). Although these factors may not contribute significantly to vehicle flow prediction when used alone, they can produce significant synergistic effects when used in combination with vehicle-side and road-side data streams.

[0087] In step S302, the secondary data streams are first classified into categories. For example, the weather condition data stream may be regarded as an independent data stream category, while the holiday information data stream may be further subdivided according to the nature of the holiday (such as public holidays, school holidays, etc.).

[0088] Then, a corresponding data stream set is created for each category, and all data streams belonging to that category in the secondary data stream set are added to the corresponding set. For example, all data streams related to weather conditions will be added to the weather conditions data stream set; and all information related to holidays will be added to the holiday information data stream set.

[0089] By subdividing and classifying important data flow sets and secondary data flow sets, the computer system not only realizes efficient organization and management of data flows, but also provides more possibilities for subsequent data analysis and model optimization. Specifically, this subdivided classification method has the following application values:

[0090] Improve data analysis efficiency: Merging data streams of the same type into the same set can simplify the data analysis process and improve the efficiency of data analysis. For example, when analyzing road congestion, relevant data streams can be directly extracted from the road-side data stream set for analysis without the need for tedious screening and matching in the entire data set.

[0091] Optimize prediction models: When building prediction models, you can make targeted selections and optimizations based on the degree of impact of different types of data streams on the model. For example, in a traffic flow prediction model, you can prioritize key data streams from the vehicle and road ends as input features to improve the prediction accuracy of the model.

[0092] Support decision-making: The classified data stream collection can provide more comprehensive and detailed information support for decision makers. Decision makers can formulate more accurate and effective traffic management strategies based on the information in different types of data stream collections. For example, when dealing with severe weather or holiday traffic peaks, corresponding emergency plans and diversion measures can be formulated based on the information in the weather condition data stream collection and the holiday information data stream collection.

[0093] Promoting resource sharing and collaboration: In the multi-source heterogeneous data environment of smart city management, the refined and classified data stream collection helps promote resource sharing and collaboration between different data sources. For example, establishing an association between the vehicle-side data stream collection and the road-side data stream collection can achieve information interaction and collaborative control between vehicles and road infrastructure, thereby improving the operating efficiency and safety of the entire transportation system.

[0094] In summary, steps S301 and S302 not only improve the efficiency and accuracy of data processing and analysis by classifying and organizing the important data flow set and the secondary data flow set, but also provide more comprehensive and in-depth information support and decision-making basis for the traffic management of smart cities. The introduction of this derivative implementation method further enriches the technical means and methodology system of smart city management, and provides a strong guarantee for building a more intelligent and efficient urban transportation system.

[0095] In one embodiment, the step S400, based on the important data stream set, the secondary data stream set and the past period vehicle flow evolution stream, performs marginal contribution detection on each of the vehicle-road multimodal monitoring data streams in the secondary data stream set to obtain a marginal contribution detection result, which may specifically include:

[0096] For any of the vehicle-road multimodal monitoring data streams in the secondary data stream set, the following steps are performed:

[0097] Step S410: Determine the linear independence of the first detection array based on the past period vehicle flow evolution flow, the important data flow set and the vehicle-road multimodal monitoring data flow.

[0098] First, the computer system retrieves the time series data of the vehicle flow evolution flow in the past period, the time series data of all data flows in the important data flow set, and the time series data of the secondary vehicle-road multimodal monitoring data flow to be evaluated from the database. These data are usually a series of observations indexed by timestamps, and each timestamp corresponds to a set of one or more data points.

[0099] Next, the computer system integrates these time series data into a multidimensional matrix to form the first detection array. Each row of this matrix represents a time point, and each column represents a data stream (including vehicle flow evolution stream, each data stream in the important data stream set, and the secondary data stream to be evaluated). The elements in the matrix are the observation values ​​corresponding to the time point and data stream.

[0100] For example, suppose you have the following data:

[0101] The evolution flow of vehicle flow in the past period (Flow): F1, F2, …, Fn;

[0102] Important Data Streams:

[0103] Road congestion index (Congestion): C1, C2, …, Cn;

[0104] Traffic light status (Signal): S1, S2, …, Sn;

[0105] Minor Stream (MS) to be evaluated: MS1, MS2, …, MSn;

[0106] Where n is the number of time points. Then the first detection array A can be expressed as:

[0107] ;

[0108] After constructing the first detection array, the computer system calculates the linear independence of the matrix, usually by calculating the rank of the array. For example, based on the row reduced echelon form (RREF) transformation, singular value decomposition (SVD), etc. Assuming that the linear independence of the first detection array A is r1, this value represents the number of linearly independent components in the data set including the past cycle vehicle flow evolution flow, the important data flow set, and the secondary data flow to be evaluated.

[0109] The significance of step S410 is to evaluate whether the secondary vehicle-road multimodal monitoring data stream to be evaluated provides additional independent information when combined with other key data streams. If the linear independence r1 of the first detection array is significantly higher than the linear independence of the matrix constructed when only the vehicle flow evolution stream of the past cycle and the important data stream set are included, it indicates that the secondary data stream to be evaluated does provide valuable independent information for the prediction model, thereby helping to improve the prediction accuracy of the model. Through this step, the computer system can scientifically screen out secondary data streams that have significant marginal contributions to the prediction model, providing strong support for subsequent data processing and model optimization.

[0110] Step S420: Determine the linear independence of the second detection array according to the past period vehicle traffic evolution flow and the important data stream set.

[0111] Step S420 aims to evaluate the linear independence of the past cycle vehicle flow evolution stream when combined with the important data stream set. This step quantifies the information independence and contribution of these core data streams in the prediction model by constructing a second detection array that does not contain the secondary vehicle-road multimodal monitoring data stream to be evaluated and calculating its linear independence.

[0112] The computer system extracts the time series data of the evolution flow of vehicle flow in the past cycle and the time series data of all data flows in the important data flow set from the database. These data usually contain key factors affecting vehicle flow, such as road congestion, traffic light status, weather conditions, etc. The system then integrates these time series data into a new multidimensional matrix to form a second detection array. Unlike the first detection array, the second detection array does not contain the secondary vehicle-road multimodal monitoring data stream to be evaluated.

[0113] In the subsequent steps (such as step S430), the computer system compares the linear independence r2 of the second detection array with the linear independence r1 of the first detection array containing the secondary data stream to be evaluated, and evaluates the marginal contribution of the secondary data stream to the prediction model by calculating the independence difference (r1-r2). If the independence difference is significantly greater than 0, it means that the secondary data stream to be evaluated provides additional independent information for the prediction model, which has a positive effect on improving the prediction accuracy.

[0114] Through this step, the computer system can scientifically evaluate the role and contribution of core data streams and secondary data streams in the prediction model, providing strong support for subsequent data processing and model optimization.

[0115] Step S430: performing difference analysis on the linear independence of the first detection array and the linear independence of the second detection array to obtain the independence difference.

[0116] Step S430 aims to quantify the contribution of the secondary data stream to be evaluated to the information independence of the prediction model by comparing the linear independence of the first detection array (including the vehicle flow evolution flow of past cycles, the set of important data streams, and the secondary vehicle-road multimodal monitoring data stream to be evaluated) with the second detection array (including only the vehicle flow evolution flow of past cycles and the set of important data streams).

[0117] After completing steps S410 and S420, the computer system has calculated the ranks of the first detection array A and the second detection array B, that is, their linear independence, which are recorded as r1 and r2 respectively. Next, in step S430, the system will perform a difference analysis on the two ranks to calculate the independence difference (r1-r2).

[0118] The value of the independence difference reflects the amount of additional linear independent information that the secondary data stream to be evaluated brings to the prediction model. Specifically: If the independence difference is greater than 0, it means that after adding the secondary data stream to be evaluated, the linear independent information in the prediction model has increased. This means that the data stream provides independent information that is different from the existing data streams (the vehicle flow evolution streams of past periods and the set of important data streams), and has potential contributions to improving the accuracy and robustness of the prediction model. If the independence difference is equal to 0, it indicates that there is a linear correlation between the secondary data stream to be evaluated and the existing data streams, and no additional independent information is brought to the prediction model. Through the difference analysis of step S430, the computer system can scientifically evaluate the contribution of the secondary data streams to the information independence of the prediction model, thereby providing strong support for subsequent data processing and model optimization.

[0119] Step S440: Perform a fit test on the past period vehicle flow evolution flow, the important data flow set and the vehicle-road multimodal monitoring data flow to obtain the test fit and fit confidence.

[0120] Step S440 quantifies the statistical relationship strength and significance of the relationship between the vehicle flow evolution flow of past cycles, the important data flow set, and the vehicle-road multimodal monitoring data flow to be evaluated through a fit test.

[0121] Goodness of fit testing aims to evaluate the ability of a statistical model to explain a data set, that is, the degree of closeness between the model's predicted values ​​and the actual observed values. In the context of smart city management, this helps to determine whether the model's ability to predict the evolution of vehicle flow in past cycles has improved after the introduction of the multimodal vehicle-road monitoring data stream to be evaluated.

[0122] In step S440, the computer system uses a statistical modeling method to perform a goodness of fit test. A common method is to construct a regression model (linear regression, logistic regression, polynomial regression, etc., depending on the data characteristics and prediction objectives) containing all relevant data streams, taking the past cycle vehicle flow evolution stream as the dependent variable, and the important data stream set and the vehicle-road multimodal monitoring data stream to be evaluated as the independent variable. Assume that the constructed regression model is in the form of:

[0123] ;

[0124] Among them, Flow tRepresents the observed value of the vehicle flow evolution flow at time point t in the past period, Congestion t ,Signal t ,…,MS t represents the set of important data streams and the observed values ​​of the multimodal monitoring data streams to be evaluated at time point t, β 0 ,β 1 ,…,β n is the model parameter, ϵ t is the error term.

[0125] The coefficient of determination R is usually used to test the goodness of fit. 2 Or the adjusted coefficient of determination AdjustedR 2 To measure the fit of the model. 2 It reflects the ratio of the variability explained by the model to the total variability, and its calculation formula is:

[0126] ;

[0127] Among them, SS res is the residual sum of squares, SS tot is the total sum of squares. However, as the number of independent variables in the model increases, R 2 Usually tends to increase, even if the new independent variable does not contribute substantially to the model's predictive power. Therefore, the adjusted coefficient of determination (Adjusted R2) is sometimes used to avoid this bias.

[0128] Fit confidence: In order to judge the significance of model fit, statistical tests such as F test are usually performed. The F test judges the overall significance of the model by comparing the variability explained by the model with the residual variability. If the value of the F statistic is large enough (that is, the corresponding p value is less than the significance level, such as 0.05), the model is considered significant, that is, the independent variables in the model have significant explanatory power for the dependent variable.

[0129] Assume that a computer system is evaluating the marginal contribution of a multimodal vehicle-road monitoring data stream representing wind speed to predicting traffic flow. By constructing a regression model that includes the past cycle vehicle flow evolution stream, road congestion index, traffic light status, and wind speed data stream, and calculating the model's R2 value and performing an F test, the system can obtain the model's detection fit and fit confidence.

[0130] If the R2 value of the model is significantly improved after adding the wind speed data stream, and the p-value of the F test is less than the significance level, then it can be considered that the wind speed data stream has a significant marginal contribution to the predicted traffic flow. On the contrary, if the R2 value does not change much or the F test is not significant, it may indicate that the wind speed data stream has limited contribution to the prediction model. Through the fit test of step S440, the computer system can scientifically evaluate the contribution of the multimodal monitoring data stream of the vehicle-road to be evaluated to the prediction model, providing strong support for subsequent data processing and model optimization.

[0131] Step S450: taking the independence difference, the detection fit and the fit confidence as the marginal contribution detection result.

[0132] In step S450, the computer system calculates the previously calculated independence difference and test fit (such as the determination coefficient R 2 ) and the confidence of fit (such as the p-value of the F test) are combined to form the final marginal contribution test result. These indicators jointly reflect the contribution of the multimodal monitoring data stream of the vehicle-road to be evaluated to the prediction model of the vehicle flow evolution flow in the past cycle. The independence difference measures the increase in the amount of linear independent information in the prediction model after adding the data stream to be evaluated. The larger the independence difference, the more independent information the data stream to be evaluated provides, and the greater the potential for improving the model's prediction ability. The test fit reflects the degree of fit of the model to the data, R 2 The closer the value is to 1, the better the model fits, that is, the smaller the deviation between the model prediction and the actual observation. Fit confidence is used to evaluate the significance of model fit, which is often judged by the p-value of a statistical test (such as the F test). The smaller the p-value (usually less than the set significance level, such as 0.05), the higher the significance of the model fit, that is, the independent variable in the model has a significant impact on the dependent variable.

[0133] In one embodiment, the step S500, taking the vehicle-road multimodal monitoring data stream whose marginal contribution detection results meet the marginal contribution standard as the target vehicle-road multimodal monitoring data stream, may include: if the independence difference is less than the independence reference value, the detection fit is less than the detection fit reference value, and the fitting confidence is less than the fitting confidence reference value, taking the vehicle-road multimodal monitoring data stream as the target vehicle-road multimodal monitoring data stream.

[0134] In step S500, the computer system sets a set of strict screening criteria, including reference values ​​of independence difference, detection fit and fit confidence. Only when the marginal contribution detection result of a certain vehicle-road multimodal monitoring data stream meets the following three conditions at the same time, the data stream will be selected as the target data stream.

[0135] Through the screening process of step S500, the computer system can scientifically identify the vehicle-road multimodal monitoring data stream that has a significant contribution to the prediction model, and use it as the target data stream for subsequent analysis. This process not only optimizes the input feature set of the prediction model, but also improves the scientificity and accuracy of traffic flow prediction in smart city management.

[0136] In one embodiment, the step S410, determining the linear independence of the first detection array according to the past period vehicle flow evolution stream, the important data stream set and the vehicle-road multimodal monitoring data stream, may include:

[0137] Step S411: Determine the observation value set of the past c time points of the vehicle flow evolution flow of the past period, the observation value set of the past d time points of each of the vehicle-road multimodal monitoring data flows in the important data flow set, and the observation value set of the past e time points of the vehicle-road multimodal monitoring data flow; wherein, the observation value set of the past c time points includes the values ​​of all time points in the interval from time point c to the current time point of the vehicle flow evolution flow of the past period, the observation value set of the past d time points includes the values ​​of all time points in the interval from time point d to the current time point of the vehicle-road multimodal monitoring data flow in the important data flow set, and the observation value set of the past e time points includes the values ​​of all time points in the interval from time point e to the current time point of the vehicle-road multimodal monitoring data flow.

[0138] In step S411, the computer system is responsible for collecting and organizing time series data from different data sources, including the vehicle flow evolution flow of past cycles, each vehicle-road multimodal monitoring data flow in the important data flow set, and the vehicle-road multimodal monitoring data flow to be evaluated.

[0139] The computer system first determines the set of observations of the past c time points of the vehicle flow evolution stream of the past cycle. Here, the past cycle refers to a specific time period, such as the past month, quarter, or year, depending on the analysis requirements and data availability. Time point c is the selected starting time point, which is located before the current time point and is used to review and analyze the historical evolution trend of vehicle flow. For example, assume that the current time point is May 1, 2023, and the past cycle is set to the past month (i.e., April 1 to April 30, 2023). If c is set to April 1, then the set of observations of the past c time points of the vehicle flow evolution stream of the past cycle contains the vehicle flow data for each day from April 1 to April 30. Next, the system determines the set of observations of the past d time points for each vehicle-road multimodal monitoring data stream in the set of important data streams. These data streams are considered to be key factors affecting vehicle flow, such as road congestion index, traffic light status, weather conditions, etc. The value of d may vary from data stream to data stream, reflecting the time lag and persistence of the impact of different factors on vehicle flow. For example, continuing with the current time of May 1, 2023, for the important data stream of road congestion index, if d is set to 7 days, then the set of observations at the past d time points includes the road congestion index data from April 24 (that is, 7 days before May 1) to April 30.

[0140] Finally, the system determines the set of observations of the past e time points of the vehicle-road multimodal monitoring data stream to be evaluated. This data stream is the data stream that is currently being evaluated for its contribution to the vehicle flow prediction model, such as new traffic monitoring camera data, the number of traffic-related posts on social media, etc. The value of e is also set according to the characteristics of the data stream and the analysis requirements. For example, suppose that a new traffic monitoring camera data stream is being evaluated to capture vehicle behavior on a specific road section. If e is set to 14 days, then its set of observations at the past e time points includes monitoring data from April 18 (14 days before May 1) to April 30.

[0141] Through step S411, the computer system collects observation sets from different data sources and different time scales, providing rich input data for subsequent data analysis and model building. These data not only reflect the historical evolution trend of vehicle flow, but also cover a variety of external factors that may affect vehicle flow. By comprehensively analyzing these data, the system can more accurately predict future changes in vehicle flow and provide a scientific basis for traffic management in smart cities.

[0142] Step S412: performing difference analysis on the values ​​of two random different time points in the set of observation values ​​at the past c time points to obtain a first difference analysis value.

[0143] Step S412 aims to explore the changing trend and dynamic characteristics of the data by calculating the difference between two randomly selected time points in the time series data. Difference analysis is a simple and effective data analysis method that reveals the rate and direction of data change by calculating the numerical differences between different time points in the time series data. In the traffic flow prediction scenario of a smart city, the change in vehicle flow is often affected by various factors such as road conditions, weather conditions, holidays, etc. Through difference analysis, the flow fluctuations caused by these factors can be captured, providing valuable features for subsequent model construction.

[0144] In step S412, the computer system first focuses on the set of observations of the past c time points of the vehicle flow evolution stream in the past cycle. This set contains the historical data of vehicle flow from time point c (an earlier time point) to the current time point.

[0145] Randomly select time points: The computer system randomly selects two different time points from these time points, denoted as t1 and t2, where t1 < t2. The selection of these two time points is random to ensure the universality and representativeness of the analysis results.

[0146] Calculate the difference: Next, the system calculates the difference between the vehicle flow observations corresponding to these two time points, which reflects the change in vehicle flow during the period from t1 to t2.

[0147] The calculated difference is the first difference analysis value. This value itself may be a specific number representing the absolute change in flow; but in some cases, the ratio or percentage change of this difference relative to a reference value (such as the average flow of the interval between the two time points) may also be concerned to more accurately evaluate the relative change degree of the flow.

[0148] Suppose in a certain traffic hot spot area of a smart city, a computer system is collecting and analyzing data on vehicle flow in past cycles. The past cycle is set as the past month, that is, it includes the vehicle flow observations in the past 30 days. In step S412, the system randomly selects two time points, such as the 10th day and the 25th day, and calculates the difference in vehicle flow between these two days. If the difference is positive, it indicates that the vehicle flow has increased from the 10th day to the 25th day, which may reflect the growth of traffic demand in this area or the improvement of surrounding road conditions. If the difference is negative, it means that the vehicle flow has decreased, which may be affected by factors such as bad weather, traffic accidents, or traffic control. Through a large number of random difference analyses, the computer system can accumulate a large number of first difference analysis values, which will serve as important inputs for subsequent data analysis, feature engineering, and model construction. For example, these difference analysis values can be used to calculate statistical indicators such as the average rate of flow change and the fluctuation range, or can be directly input as features into a machine learning model to predict the future trend of vehicle flow changes.

[0149] Perform the following steps for any set of observations at the past d time points:

[0150] Step S413: Analyze the difference between the values of two different randomly selected time points in the set of observations at the past d time points to obtain a second difference analysis value.

[0151] In step S413, the computer system first focuses on a specific vehicle-road multi-modal monitoring data stream in the important data stream set. This data stream may represent one of the various factors affecting vehicle flow, such as the road congestion index, the change frequency of traffic signal states, the change of weather conditions, etc.

[0152] Selection of the set of observations: The system starts the analysis from the set of observations at the past d time points of this data stream. This set contains all the observations of this data stream from time point d (an earlier starting time point) to the current time point. The value of d is determined according to the characteristics of the data stream and the analysis requirements, reflecting the size of the time window considered for the analysis.

[0153] Selection of random time points: Next, the system randomly selects two different time points in this set of observations, denoted as t3 and t4, where t3 < t4. The selection of these two time points is random to ensure the diversity and representativeness of the analysis results.

[0154] Then, the system calculates the difference between the observed values corresponding to these two time points, and this difference reflects the change of the factor represented by this data stream during the period from t3 to t4. Finally, the system records the calculated second difference analysis value as a quantitative representation of the dynamic change characteristics of this data stream. This process will be repeated for each data stream in the set of important data streams to obtain the difference analysis results of each data stream. By executing step S413, the computer system can systematically analyze the dynamic change characteristics of each vehicle-road multi-modal monitoring data stream in the set of important data streams, providing rich data support for the traffic management of smart cities. These difference analysis results not only help to deeply understand various factors affecting traffic flow, but also lay a solid foundation for subsequent data modeling and decision-making.

[0155] Step S414: Perform difference analysis on the values of two different randomly selected time points in the set of observed values of the past e time points to obtain a third difference analysis value.

[0156] By calculating the numerical difference between two randomly selected time points in the set of observed values of the past e time points, step S414 aims to reveal the dynamic change characteristics within this data stream and provide key information for subsequent data processing and model optimization. In step S414, the computer system focuses on the vehicle-road multi-modal monitoring data stream to be evaluated, which may be newly introduced or not fully evaluated before, but may be of great significance for traffic flow prediction.

[0157] First, the system selects data from the set of observed values of the past e time points of this data stream for analysis. This set contains all the recorded observed values of this data stream from time point e (an earlier starting time point) to the current time point. The value of e is determined according to the update frequency of the data stream, the analysis requirements, and the data availability. Then, the system randomly selects two different time points in this set of observed values, denoted as t5 and t6, where t5 < t6. This random selection method helps to capture the dynamic changes of the data stream in different time periods and avoid over-reliance on a specific time period. Then, the system calculates the difference between the observed values corresponding to these two time points, and this difference reflects the change amount of the factor represented by this data stream during the period from t5 to t6. Finally, the system records the calculated third difference analysis value and may compare and analyze it with the difference analysis results of other data streams. This process helps to evaluate the actual contribution of the data stream to be evaluated to the traffic flow prediction model and provides a basis for the weight allocation of the data stream in the model.

[0158] By executing step S414, the computer system can systematically evaluate the dynamic change characteristics of the vehicle-road multimodal monitoring data stream to be evaluated, and provide strong data support for the traffic management of smart cities. This not only helps to improve the accuracy of traffic flow prediction, but also provides a scientific basis for traffic planning and policy formulation.

[0159] Step S415: Determine the linear independence of the first detection array according to the first difference analysis value, the second difference analysis value and the third difference analysis value.

[0160] Step S415 integrates the first, second and third difference analysis values ​​calculated in the previous steps to evaluate the linear independence of the first detection array. Although this process does not directly calculate the linear independence through the difference analysis values ​​(because the difference analysis itself does not directly reflect linear independence), these differences can be used as features to indirectly infer or assist in determining the linear relationship of the data stream in the first detection array.

[0161] Since the difference analysis values ​​themselves are not directly equivalent to the measure of linear independence, the computer system needs to use these difference analysis values ​​through other means. One feasible method is to use these difference analysis values ​​as features to construct a feature vector or matrix, and then use linear algebra, statistics or machine learning methods to evaluate the linear relationship between these features, thereby indirectly inferring the linear independence of the first detection array.

[0162] First, the computer system combines the first, second, and third differential analysis values ​​into a feature vector or matrix. For example, assuming there is a scenario with n important data streams and 1 data stream to be evaluated, then for each data stream (including the vehicle flow evolution stream in the past period), one or more differential analysis values ​​are calculated (depending on the values ​​of c, d, e and the number of random selections).

[0163] If only one difference analysis value is calculated for each data stream (for simplicity), the feature vector may be as follows (here three data streams are taken as an example, including the vehicle flow evolution stream of the past cycle, an important data stream and a data stream to be evaluated):

[0164] ;

[0165] The second difference analysis value 1 to the second difference analysis value n respectively correspond to the difference analysis values ​​of n important data streams.

[0166] After constructing the feature vectors or matrices, the computer system can use a variety of methods to evaluate the linear relationship between these features, thereby indirectly inferring the linear independence of the first detection array.

[0167] Method 1: Calculate the condition number: By calculating the condition number of the feature matrix, the degree of ill-conditioning of the matrix can be evaluated. The larger the condition number, the more ill-conditioned the matrix is, that is, the larger the ratio between the eigenvalues, and the more likely the column vectors (i.e., features) of the matrix are linearly correlated. However, this method requires the feature matrix to be a square matrix, and may not be intuitive enough in practical applications.

[0168] Method 2: Principal Component Analysis (PCA): PCA can be used to identify the principal components in the data, that is, the directions in which the data varies the most. If the PCA results show that most of the variation can be explained by a few principal components, then it means that there is a strong linear correlation between the original features. However, PCA itself does not directly give a quantitative value of linear independence.

[0169] Method 3: Statistical test: Use statistical tests (such as the variance inflation factor (VIF) test in multiple linear regression) to assess multicollinearity between independent variables. Although this is not a direct way to assess linear independence, it can reveal the degree of linear dependence between features.

[0170] Method 4: Feature selection of machine learning models: When building a prediction model, the importance of each feature is evaluated through feature selection algorithms (such as recursive feature elimination (RFE) and model-based feature selection). Although these algorithms are usually used to optimize model performance rather than directly evaluate linear independence, they can help identify the features that contribute most to model predictions and are relatively independent.

[0171] In one implementation, the step S420, determining the linear independence of the second detection array according to the past period vehicle flow evolution stream and the important data stream set, may specifically include:

[0172] Step S421: Determine the set of observation values ​​of the past f time points of the past periodic vehicle flow evolution flow, and the set of observation values ​​of the past g time points of each of the vehicle-road multimodal monitoring data flows in the important data flow set; wherein the set of observation values ​​of the past f time points includes the values ​​of all time points in the interval from time point f to the current time point of the past periodic vehicle flow evolution flow, and the set of observation values ​​of the past g time points includes the values ​​of all time points in the interval from time point g to the current time point of the vehicle-road multimodal monitoring data flows in the important data flow set;

[0173] Step S422: performing difference analysis on the values ​​of two random time points in the set of observation values ​​at the past f time points to obtain a fourth difference analysis value corresponding to the values ​​of the two random time points;

[0174] For any set of observations at the past g time points, complete the following steps:

[0175] Step S423: performing difference analysis on the values ​​of two random time points in the set of observation values ​​at the past g time points to obtain a fifth difference analysis value corresponding to the values ​​of the two random time points;

[0176] Step S424: superimposing each of the fourth difference analysis values ​​to obtain a fourth superimposition result, and superimposing each of the fifth difference analysis values ​​to obtain a fifth superimposition result;

[0177] Step S425: Determine the linear independence of the second detection array according to the fourth superposition result and the fifth superposition result.

[0178] In step S421, the computer system first determines the observation value set of a specific time period of each vehicle-road multimodal monitoring data flow in the past cycle vehicle flow evolution flow and the important data flow set. This step is the basis of data processing and provides necessary data input for subsequent analysis.

[0179] The computer system retrieves the observations of the past cycle vehicle flow evolution flow at the past f time points from the database to form an observation set. The "past cycle" here usually refers to a longer period of time, such as the past month, a quarter, or a year, depending on the analysis requirements. f, as the number of time points, determines the length of the observation set, which should be long enough to reflect the periodicity and trend of traffic evolution while avoiding introducing too much noise. For example, assuming that the current time is June 1, 2023, the past cycle is set to the past month (i.e., May 1 to May 31, 2023). If f is set to 30 (representing one observation per day), the observation set of the past cycle vehicle flow evolution flow will contain vehicle flow data for each day from May 1 to May 31.

[0180] For each vehicle-road multimodal monitoring data stream in the important data stream set (such as road congestion index, traffic light status, weather conditions, etc.), the computer system also retrieves its observations at the past g time points to form their own observation sets. The value of g may be different from f, reflecting that different data streams may have different time sensitivities and update frequencies. For example, continuing with the above scenario, assume that the update frequency of the road congestion index is once an hour, and the data range to be analyzed is consistent with the vehicle flow evolution stream (i.e., the past month). Then, for the road congestion index data stream, if g is set to 720 (representing 24 hours a day, one observation per hour, and a total of 30 days), its observation set will contain hourly road congestion index data from May 1 to May 31.

[0181] In step S422, the computer system performs a difference analysis on the set of observed values of the vehicle flow evolution flow in past cycles to explore the dynamic changes in the flow data between different time points.

[0182] Randomly select time points: Randomly select two different time points (denoted as t7 and t8, where t7 < t8) from the set of observed values of the past f time points. These two time points represent any two sampling moments in the flow data. Calculate the difference between the flow observed values corresponding to these two time points to obtain the fourth difference analysis value (denoted as ΔFlow) (ΔFlow = flow(t8) - flow(t7)). To obtain more comprehensive change information, this step may need to be repeated multiple times, each time randomly selecting different pairs of time points and recording all the calculated fourth difference analysis values.

[0183] For each data flow in the set of important data flows, step S423 performs an operation similar to step S422, but for the set of observed values of their respective data flows.

[0184] Randomly select time points: Randomly select two different time points (denoted as t9 and t10, where t9 < t10) from the set of observed values of the past g time points. Calculate the difference between the observed values corresponding to these two time points to obtain the fifth difference analysis value (for example, for the road congestion index data flow (i.e., ΔCongestion), ΔCongestion = congestion index(t10) - congestion index(t9)). Similar to step S422, to obtain more comprehensive change information, the difference analysis needs to be repeated multiple times for each data flow, and all the calculated fifth difference analysis values are recorded.

[0185] In step S424, the computer system performs a superposition process on all the difference analysis values calculated in steps S422 and S423 to form two comprehensive difference results: the fourth superposition result and the fifth superposition result.

[0186] Add up all the fourth difference analysis values (i.e., the difference analysis results of the vehicle flow evolution flow in past cycles) to obtain the fourth superposition result (∑ΔFlow). This result reflects the overall change amount of the vehicle flow between different time points. For each data flow in the set of important data flows, first add up all its fifth difference analysis values to obtain their respective superposition results, and then further add up these results to obtain the final fifth superposition result. This result reflects the overall change amount of all important data flows between different time points.

[0187] In step S425, the computer system uses the fourth superposition result and the fifth superposition result obtained in step S424 to indirectly evaluate the linear independence degree of the second detection array.

[0188] Since the difference superposition result essentially reflects the amount of change or volatility of the data, it can be assumed that this amount of change is associated with the linear independence of the data to some extent. Specifically, if the data streams in the second detection array are highly linearly correlated, then their changing trends should be similar, resulting in a certain consistency or regularity in the difference superposition results. Conversely, if the linear independence between the data streams is high, their difference superposition results may be more random or scattered. Obtaining linear independence by indirect inference can be achieved in the following way.

[0189] Principal Component Analysis (PCA): Combine the observations of the past cycle vehicle flow evolution stream and the important data stream collection into a multidimensional data set, and apply PCA to identify the main directions of change in the data (i.e., principal components). If the number of principal components is much smaller than the number of original data streams, and most of the changes can be explained by a few principal components, it may indicate that there is a strong linear correlation between the original data streams.

[0190] Singular Value Decomposition (SVD): Perform SVD decomposition on the second detection array and infer the rank of the matrix (i.e., the number of linearly independent components) by observing the size and distribution of the singular values. Smaller singular values ​​may correspond to noise or redundant information, while larger singular values ​​correspond to truly independent components in the data.

[0191] Statistical test: Statistical test methods (such as multicollinearity test) are used to directly evaluate the linear relationship between data streams. However, these methods usually require the assumption that the data meets certain distribution conditions or model forms, so they may be subject to certain limitations in practical applications.

[0192] Machine learning feature selection: When building a predictive model, feature selection algorithms such as recursive feature elimination (RFE), model-based feature selection, etc. are used to evaluate the importance of each data stream and its contribution to the model's predictive performance. Although these methods do not directly evaluate linear independence, they can help identify the most useful and relatively independent features for the prediction task.

[0193] As an implementation mode, the method may further include:

[0194] Step S600: performing the reasoning through a plurality of the basic reasoning networks according to the target vehicle-road multimodal monitoring data stream, the important data stream set and the past period vehicle flow evolution stream, to obtain a basic reasoning set corresponding to each basic reasoning network;

[0195] Step S700: Fusing each of the basic reasoning sets to obtain the inferred vehicle traffic evolution flow corresponding to the past period vehicle traffic evolution flow.

[0196] In step S600, the computer system uses the target vehicle-road multimodal monitoring data stream, the important data stream set, and the past cycle vehicle flow evolution stream as input, and performs reasoning analysis through multiple pre-trained basic reasoning networks. Each basic reasoning network is built based on different algorithms or models, aiming to capture potential patterns and laws in the data from different angles.

[0197] In order to make full use of the characteristics of different data streams, the system can select multiple types of basic inference networks, including but not limited to:

[0198] Statistical regression models: such as linear regression and polynomial regression, are suitable for capturing linear or nonlinear relationships between variables.

[0199] Time series analysis models: such as ARIMA (autoregressive integrated moving average model), seasonal decomposition (STL decomposition), etc., are particularly suitable for processing trends, seasonality and random components in time series data.

[0200] Machine learning models: such as random forests, gradient boosted tree (GBDT), support vector machines (SVM), etc., can process high-dimensional data and automatically learn complex relationships between features.

[0201] Deep learning models: such as recurrent neural networks (RNNs), long short-term memory networks (LSTMs), gated recurrent units (GRUs), etc., are good at capturing long-term dependencies in sequence data.

[0202] For each base inference network, the system performs the following steps:

[0203] According to the needs of the network, the input data is preprocessed as necessary, such as normalization, missing value filling, outlier processing, etc. Load the pre-trained basic inference network model. These models may have been trained on similar data sets and have shown good generalization ability. Input the preprocessed data into the basic inference network, perform the forward propagation process, and obtain the inference result of the network. The inference result may be a sequence of predicted values, indicating the evolution trend of vehicle flow in the future. Save the inference result of each basic inference network as a basic inference set for use in subsequent steps. For example, suppose that in a traffic hotspot area in a smart city, the system is using the vehicle flow evolution stream of past cycles, a set of important data streams (including road congestion index, traffic light status, weather conditions, etc.) and the target vehicle-road multimodal monitoring data stream (such as newly installed traffic monitoring camera data) to predict vehicle flow. The system selects linear regression model, LSTM network and random forest as the basic inference network.

[0204] For the linear regression model, it is applied to capture the linear relationship between vehicle flow and road congestion index. The LSTM network is used to learn the long-term dependencies in the evolution flow of vehicle flow in past cycles, especially the periodic changes in traffic flow. The random forest takes advantage of its integrated learning to capture nonlinear relationships in the data from multiple perspectives and automatically handle feature selection issues. Each network outputs a predicted vehicle flow evolution flow sequence as its own basic reasoning set.

[0205] After obtaining the inference results of multiple basic inference networks, step S700 aims to obtain a more accurate and stable inference vehicle flow evolution flow by fusing these results. The fusion process can adopt a variety of ensemble learning methods, such as averaging, weighted averaging, voting, stacking, etc.

[0206] As an implementation mode, the basic reasoning network includes an independent basic reasoning network, a domain knowledge reasoning network and a comprehensive basic reasoning network; the step S600, according to the target vehicle-road multimodal monitoring data stream, the important data stream set and the past period vehicle flow evolution stream, performs the reasoning through a plurality of the basic reasoning networks to obtain a basic reasoning set corresponding to each of the basic reasoning networks, which may specifically include:

[0207] Step S610: performing independent reasoning through the plurality of independent basic reasoning networks one by one according to the vehicle flow evolution flow of the past period, and obtaining independent basic reasoning sets corresponding to the plurality of independent basic reasoning networks respectively;

[0208] Step S620: performing reasoning through the domain knowledge reasoning network according to the past period vehicle flow evolution flow and road prior data, and obtaining an expert basic reasoning set corresponding to the domain knowledge reasoning network;

[0209] Step S630: According to the target vehicle-road multimodal monitoring data stream, the important data stream set and the past period vehicle flow evolution stream, comprehensive reasoning is performed one by one through the multiple comprehensive basic reasoning networks to obtain the comprehensive basic reasoning sets corresponding to the multiple comprehensive basic reasoning networks. The independent basic reasoning network mainly relies on the data characteristics of the past period vehicle flow evolution stream itself for reasoning, and does not directly rely on external multimodal monitoring data streams or domain knowledge. This type of network usually includes some classic time series analysis models, statistical models and simple machine learning models, which can capture the inherent laws of periodicity, trend, seasonality, etc. in traffic data. Time series analysis models: such as ARIMA (autoregressive integrated moving average model), SARIMA (seasonal ARIMA model), etc.

[0210] Statistical models: such as exponential smoothing (including simple exponential smoothing, Holt two-parameter exponential smoothing, Holt-Winters three-parameter exponential smoothing, etc.), which can smooth random fluctuations in time series data while retaining long-term trends and seasonal changes.

[0211] Simple machine learning models: such as linear regression and decision trees, although they are not specifically designed for time series data, they can capture linear relationships or simple rules in traffic data in some cases.

[0212] In step S610, the computer system inputs the past period vehicle flow evolution flow into each independent basic reasoning network one by one, and performs the following reasoning steps:

[0213] Data preprocessing: Preprocess the traffic evolution flow data according to the needs of each network, such as removing outliers, filling missing values, standardizing or normalizing, etc.

[0214] Model loading and configuration: Load a pre-trained independent basic inference network model and adjust the model parameters (such as the p, d, q parameters of the ARIMA model, the smoothing coefficient of the Holt-Winters model, etc.) according to the actual situation.

[0215] Reasoning execution: The pre-processed traffic evolution flow data is input into the model to execute the reasoning process. The model will calculate the predicted value of vehicle flow in the future based on the internal algorithm.

[0216] Result recording: The inference results of each independent basic inference network are saved as independent basic inference sets. These sets contain the prediction sequences of future vehicle traffic by each network.

[0217] Assume that in a busy traffic section of a smart city, the system collects daily vehicle flow data for the past three years as the past cycle vehicle flow evolution flow. In order to predict the vehicle flow in the next week, the system selects the ARIMA model and the Holt-Winters three-parameter exponential smoothing model as independent basic inference networks.

[0218] For the ARIMA model, the system first determines the model's p, d, and q parameters through ACF (autocorrelation function) and PACF (partial autocorrelation function) graphs, and then uses these parameters to train the model. In the inference phase, the system inputs the traffic data of the past three years into the trained ARIMA model to obtain the traffic forecast sequence for the next week.

[0219] For the Holt-Winters model, the system sets trend smoothing coefficients, seasonal smoothing coefficients, and error smoothing coefficients, and trains the model based on historical data. During inference, the flow data of the past three years is also input into the model to obtain another set of flow forecast sequences for the next week.

[0220] In this way, the system obtains two sets of prediction results based on different independent basic inference networks, providing diverse options for subsequent fusion.

[0221] The domain knowledge reasoning network combines road prior data and past cycle vehicle flow evolution flow, and uses expert knowledge or domain rules for reasoning. This type of network usually embeds professional knowledge in fields such as traffic engineering and urban planning, and can more accurately reflect the changing laws of actual traffic conditions.

[0222] Expert system: A rule-based system that simulates the expert decision-making process through a predefined rule base and an inference engine.

[0223] Hybrid model: A model that combines statistical models, machine learning models and expert rules, such as Bayesian networks, fuzzy logic systems, etc.

[0224] In step S620, the computer system uses the road prior data (such as road grade, number of lanes, number of intersections, surrounding land use conditions, etc.) and the past cycle vehicle flow evolution flow as input, and performs reasoning through the domain knowledge reasoning network:

[0225] Data integration: Integrate road prior data with past cycle vehicle flow evolution streams to form an input dataset containing rich contextual information.

[0226] Rule definition: Based on the professional knowledge and historical experience in the field of traffic engineering, a series of inference rules are defined. These rules may involve multiple aspects such as traffic distribution, congestion propagation, and holiday effects.

[0227] Reasoning execution: The integrated data is input into the domain knowledge reasoning network and reasoning is performed according to predefined rules. The network may use forward reasoning, backward reasoning, or hybrid reasoning to generate traffic prediction results in combination with expert knowledge.

[0228] Result recording: The reasoning results of the domain knowledge reasoning network are saved as an expert basic reasoning set. This set reflects the contribution of expert knowledge to traffic prediction.

[0229] Continuing with the example of a busy road section in a smart city, in addition to collecting the evolution of vehicle flow in past cycles, the system also obtains prior road data for the road section, including road grade, number of lanes, number of intersections, etc. In order to more accurately predict vehicle flow in the next week, the system uses a rule-based expert system as the domain knowledge reasoning network.

[0230] The system first defines multiple inference rules based on road prior data, such as "high-grade roads with more than four lanes have a larger increase in traffic during peak hours in the morning and evening" and "road sections with dense intersections have larger fluctuations in traffic during holidays".

[0231] The system then embeds these rules into the expert system and performs reasoning based on the evolution of vehicle traffic in past cycles. During the reasoning process, the computer system selects applicable rules for traffic forecasting based on the current time (such as whether it is approaching a holiday), historical traffic patterns, and prior road data.

[0232] Finally, the system obtains a traffic prediction sequence based on expert knowledge as part of the expert basic reasoning set. This set not only reflects the inherent laws of traffic data, but also incorporates the expert's understanding and judgment of traffic conditions.

[0233] The comprehensive basic reasoning network combines multiple data sources such as the target vehicle-road multimodal monitoring data stream, important data stream collection, and past cycle vehicle flow evolution stream, and performs comprehensive reasoning through complex machine learning models or deep learning models. This type of network can capture the inherent connections and mutual influences between different data sources and generate more comprehensive and accurate traffic prediction results.

[0234] Complex machine learning models: such as ensemble learning models such as random forest, gradient boosted tree (GBDT), XGBoost, and kernel method models such as support vector machine (SVM).

[0235] Deep learning models: such as recurrent neural networks (RNNs), long short-term memory networks (LSTMs), gated recurrent units (GRUs), and attention mechanism models, are particularly suitable for processing sequence data and capturing long-term dependencies.

[0236] In step S630, the computer system uses the target vehicle-road multimodal monitoring data stream, the important data stream set and the past period vehicle flow evolution stream as comprehensive inputs, and performs reasoning one by one through multiple comprehensive basic reasoning networks:

[0237] Feature engineering: Perform feature engineering on the data according to the needs of each network, including operations such as feature selection, feature extraction, and feature transformation. For example, different sensor data in the multimodal monitoring data stream can be fused into new feature variables; the evolution stream of vehicle flow in the past period can be differentiated and logarithmically transformed to extract trend information.

[0238] Model loading and configuration: Load the pre-trained comprehensive basic inference network model and adjust the model parameters (such as learning rate, batch size, number of iterations, etc.) according to actual conditions.

[0239] Reasoning execution: The processed comprehensive input data is input into the model to perform the reasoning process. The model will calculate the predicted value of vehicle flow in the future period based on the internal algorithm.

[0240] Result recording: The inference results of each comprehensive basic inference network are saved as comprehensive basic inference sets. These sets contain the comprehensive prediction sequences of the future vehicle flow of each network.

[0241] Taking the busy traffic section of the smart city as an example, we not only collected the evolution flow of vehicle flow in the past cycle and the road prior data, but also obtained the multimodal monitoring data flow of the target vehicle and road (such as real-time traffic monitoring video, vehicle GPS trajectory data) and important data flow collection (such as road congestion index, traffic light status, weather conditions, etc.). In order to make full use of the advantages of these data sources, the system selected random forest and LSTM networks as the comprehensive basic inference network.

[0242] For the random forest model, the system first performs feature engineering to convert the relevant data in the multimodal monitoring data stream and the important data stream set into feature vectors suitable for model input. Then the trained random forest model is loaded for inference. Random forest reduces the risk of overfitting and improves prediction accuracy by integrating the results of multiple decision trees. For the LSTM network, the system divides the data sequence by time step and uses the vehicle flow evolution flow of the past cycle, the important data stream set, and the processed multimodal monitoring data stream as input sequences. The LSTM network captures the long-term dependencies in the sequence data through its internal gating mechanism and generates a flow prediction sequence. Finally, the system obtains two sets of comprehensive prediction results based on random forest and LSTM network as part of the comprehensive basic reasoning set. These sets not only integrate information from multiple data sources but also use the powerful reasoning ability of complex models to generate more accurate and comprehensive flow prediction results.

[0243] As an implementation method, the types of the independent basic inference networks include Holt-Winters smoothing network, trend elimination network and data stream component splitting network. Based on this, the step S610, according to the past periodic vehicle flow evolution flow, performs independent inference through multiple independent basic inference networks one by one, and obtains independent basic inference sets corresponding to multiple independent basic inference networks respectively. Specifically, it may include: according to the past periodic vehicle flow evolution flow, performs independent inference through the Holt-Winters smoothing network, the trend elimination network and the data stream component splitting network one by one, and obtains independent basic inference sets corresponding to the Holt-Winters smoothing network, the trend elimination network and the data stream component splitting network respectively.

[0244] In step S610, the computer system uses various types of independent basic reasoning networks, such as the Holt-Winters smoothing network, the trend elimination network (such as the differential autoregressive model), and the data stream component splitting network, to perform reasoning analysis on the past cycle vehicle flow evolution flow one by one to obtain the corresponding independent basic reasoning sets. The Holt-Winters smoothing network is a classic method widely used in time series prediction, especially for data with obvious seasonality and trend. In the implementation of the computer system, the Holt-Winters smoothing network generates a future traffic forecast sequence by smoothing the past cycle vehicle flow evolution flow and taking into account the long-term trend and seasonal changes in the traffic data.

[0245] Specifically, the computer system first determines the parameters of the Holt-Winters smoothing network based on historical data, including the smoothing coefficients of the level parameter (Level), trend parameter (Trend) and seasonality parameter (Seasonality). Then, the vehicle flow evolution flow of the past period is used as input, and the Holt-Winters smoothing algorithm is applied for iterative calculation, gradually updating the level, trend and seasonal components until the entire forecast period is covered. Finally, the Holt-Winters smoothing network outputs a smooth flow forecast sequence that contains trends and seasonality as part of an independent basic reasoning set.

[0246] Trend elimination networks are designed to remove trend components from time series data in order to more accurately analyze the volatility and periodicity of the remaining parts. The difference autoregression model (a simplified form of the ARIMA model, especially when the difference order d is greater than 0) is a common implementation of this type of network. Specifically, the computer system first eliminates the trend component in the past cycle vehicle flow evolution stream through a difference operation and determines the appropriate difference order. Then, the autoregression (AR) and moving average (MA) models are applied to the differenced sequence for modeling, and the model parameters are estimated by minimizing the prediction error. The difference autoregression model generates a traffic forecast sequence with trend components removed, which reflects the volatility and potential periodic patterns of traffic data and provides another perspective for the independent basic reasoning set. The data stream component separation network focuses on decomposing the past cycle vehicle flow evolution stream into multiple interpretable components, such as trend, seasonality, and random noise. Although the specific model may vary depending on the application scenario, the basic idea is to separate these components through mathematical transformation or optimization algorithm.

[0247] The computer system can use algorithms such as STL (Seasonal Trend Decomposition) to split the data flow components. The STL algorithm iteratively decomposes the traffic data into three parts: trend, seasonality and residual, and allows the user to smooth each component as needed. The data flow component splitting network outputs three independent component sequences: trend component, seasonal component and random noise component. These components not only provide an in-depth understanding of the structure of the traffic data, but also can be used as feature variables in an independent basic reasoning set for subsequent analysis. In step S610, the computer system independently infers the past cycle vehicle flow evolution flow through the Holt-Winters smoothing network, the trend elimination network and the data flow component splitting network one by one. Each network generates a corresponding independent basic reasoning set based on its unique algorithm and assumptions. These sets not only reflect different aspects and characteristics of traffic data, but also provide rich prediction options for subsequent fusion and integration. In this way, the computer system can more comprehensively capture the complex dynamics and pattern changes in the past cycle vehicle flow evolution flow, and provide a solid data support and analysis basis for traffic flow prediction in smart city management.

[0248] As an implementation mode, the types of the comprehensive basic reasoning network include a multivariate time series regression network and a LSTM network; based on this, the step S630, according to the target vehicle-road multimodal monitoring data stream, the important data stream set and the past period vehicle flow evolution stream, performs comprehensive reasoning through multiple comprehensive basic reasoning networks one by one, and obtains the comprehensive basic reasoning sets corresponding to the multiple comprehensive basic reasoning networks respectively. Specifically, it may include: according to the target vehicle-road multimodal monitoring data stream, the important data stream set and the past period vehicle flow evolution stream, performs comprehensive reasoning through the multiple time series regression network and the LSTM network one by one, and obtains the comprehensive basic reasoning sets corresponding to the multiple time series regression network and the LSTM network respectively.

[0249] The comprehensive basic reasoning network can integrate information from multiple data sources, including multimodal monitoring data streams of target vehicles and roads, important data stream sets, and vehicle flow evolution streams in past cycles, and conduct comprehensive analysis through complex algorithms and models to generate more accurate prediction results. In step S630, the computer system uses a multivariate time series regression network (Vector Autoregression, VAR) and a long short-term memory network (LSTM) as representatives of the comprehensive basic reasoning network to conduct comprehensive reasoning on the input data and generate their corresponding comprehensive basic reasoning sets. The multivariate time series regression network is a statistical model used to analyze the relationship between multiple time series variables. In the traffic flow prediction of smart city management, the VAR model can capture the dynamic interaction between different data streams and their joint impact on future vehicle flow.

[0250] Specifically, the computer system first converts the target vehicle-road multimodal monitoring data stream, the important data stream set and the vehicle flow evolution stream of the past cycle into a format suitable for VAR model input. This usually involves aligning, standardizing and possibly differentially processing the time series data of each data stream to ensure that the model can handle stationary time series. When constructing the VAR model, the computer system determines the lag order of the model, that is, how many past time points of data are considered to affect the current and future data, for example, by optimizing the selection through information criteria (such as AIC, BIC). After determining the lag order, the model will use the historical observations of these data streams to estimate parameters, which reflect the dynamic relationship between the variables. In the inference stage, the computer system inputs the latest data stream observations into the trained VAR model, and the model will calculate the predicted value of future vehicle flow based on the estimated parameters and input data. This predicted value takes into account the mutual influence and dynamic changes between all input data streams. After the inference is completed, the computer system saves the prediction results of the VAR model as part of the comprehensive basic inference set for use in subsequent steps. The LSTM network is a special recurrent neural network (RNN) that is particularly suitable for processing long-term dependencies in sequence data. In traffic flow prediction, the LSTM network can capture the periodicity and trend in the evolution flow of vehicle flow in past cycles and the instant change information in the multimodal monitoring data stream.

[0251] The computer system first performs feature engineering on the input data, converting the target vehicle-road multimodal monitoring data stream, the important data stream set, and the past cycle vehicle flow evolution stream into feature vectors or sequences that can be processed by the LSTM network. This may include window division, feature selection, feature scaling, and other operations of the time series. In the inference stage, the computer system inputs the latest data stream observations as input sequences into the trained LSTM network. The network will use internal states (such as hidden states and cell states) to capture long-term dependencies in the sequence and generate a predicted sequence of future vehicle flow. The prediction results of the LSTM network are also saved as part of the comprehensive basic reasoning set, which complements the prediction results of the VAR model. In step S630, the computer system performs comprehensive reasoning on the input data through the VAR model and the LSTM network, and generates two independent comprehensive basic reasoning sets. These two sets reflect the understanding and prediction capabilities of different models for the input data. The VAR model focuses on capturing the dynamic relationship and mutual influence between variables, while the LSTM network is better at processing long-term dependencies and immediate changes in sequence data. By fusing and integrating these two inference sets (such as weighted averaging, stacking, etc.), the computer system can fully utilize the advantages of different models to generate more accurate and comprehensive vehicle flow prediction results. This process not only improves the accuracy and robustness of the prediction, but also provides a strong data foundation for traffic flow control and decision support in smart city management.

[0252] An embodiment of the present application provides a computer system, including a memory and a processor, wherein the memory stores a computer program that can be executed on the processor, and when the processor executes the program, some or all of the steps in the above method are implemented.

[0253] 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.

[0254] 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 smart city management method based on vehicle-road-cloud collaboration, characterized in that: The method comprises: Acquire a past period vehicle flow evolution stream and a plurality of vehicle-road multimodal monitoring data streams in a target monitoring area, wherein a unit root test statistic of the vehicle-road multimodal monitoring data stream is less than a statistical reference value; Performing correlation analysis on each of the past period vehicle flow evolution streams and each of the vehicle-road multimodal monitoring data streams to obtain a correlation analysis result; Adding each of the vehicle-road multimodal monitoring data streams to an important data stream set or a secondary data stream set according to the correlation analysis result; According to the important data stream set, the secondary data stream set and the past period vehicle flow evolution stream, a marginal contribution detection is performed on each of the vehicle-road multimodal monitoring data streams in the secondary data stream set to obtain a marginal contribution detection result; The vehicle-road multimodal monitoring data stream whose marginal contribution detection result meets the marginal contribution standard is used as the target vehicle-road multimodal monitoring data stream; wherein the target vehicle-road multimodal monitoring data stream is used to input multiple basic reasoning networks for reasoning to obtain the reasoned vehicle flow evolution stream corresponding to the vehicle flow evolution stream of the past period; Wherein, after adding each of the vehicle-road multimodal monitoring data streams to the important data stream set or the secondary data stream set according to the correlation analysis result, the method further includes: Adding the vehicle-road multimodal monitoring data stream of the same type in the important data stream set to the data stream set corresponding to the type; Adding the vehicle-road multimodal monitoring data stream of the same type in the secondary data stream set to the data stream set corresponding to the type; The data stream set includes the vehicle-road multimodal monitoring data stream in the important data stream set and the vehicle-road multimodal monitoring data stream in the secondary data stream set; The performing of marginal contribution detection on each of the vehicle-road multimodal monitoring data streams in the secondary data stream set according to the important data stream set, the secondary data stream set and the past period vehicle flow evolution stream to obtain a marginal contribution detection result includes: For any of the vehicle-road multimodal monitoring data streams in the secondary data stream set, the following steps are performed: Determining the linear independence of the first detection array according to the past period vehicle flow evolution stream, the important data stream set and the vehicle-road multimodal monitoring data stream; Determining the linear independence of the second detection array according to the past period vehicle flow evolution stream and the important data stream set; Performing difference analysis on the linear independence of the first detection array and the linear independence of the second detection array to obtain the independence difference; Performing a fit test on the past period vehicle flow evolution stream, the important data stream set, and the vehicle-road multimodal monitoring data stream to obtain a test fit and a fit confidence; The independence difference, the detection fit and the fit confidence are used as marginal contribution detection results; The method of taking the vehicle-road multimodal monitoring data stream whose marginal contribution detection result meets the marginal contribution standard as the target vehicle-road multimodal monitoring data stream includes: If the independence difference is less than the independence reference value, the detection fit is less than the detection fit reference value, and the fit confidence is less than the fit confidence reference value, the vehicle-road multimodal monitoring data stream is used as the target vehicle-road multimodal monitoring data stream; The determining of the linear independence of the first detection array according to the past period vehicle flow evolution flow, the important data flow set and the vehicle-road multimodal monitoring data flow includes: Determine the observation value set of the past c time points of the past period vehicle flow evolution flow, the observation value set of the past d time points of each of the vehicle-road multimodal monitoring data flows in the important data flow set, and the observation value set of the past e time points of the vehicle-road multimodal monitoring data flow; wherein the observation value set of the past c time points includes the values ​​of all time points in the interval from time point c to the current time point of the past period vehicle flow evolution flow, the observation value set of the past d time points includes the values ​​of all time points in the interval from time point d to the current time point of the vehicle-road multimodal monitoring data flow in the important data flow set, and the observation value set of the past e time points includes the values ​​of all time points in the interval from time point e to the current time point of the vehicle-road multimodal monitoring data flow; Performing difference analysis on the values ​​of two random different time points in the set of observation values ​​at the past c time points to obtain a first difference analysis value; For any set of observations at the past d time points, complete the following steps: Performing difference analysis on the values ​​of two random different time points in the set of observation values ​​at the past d time points to obtain a second difference analysis value; Performing difference analysis on the values ​​of two random different time points in the set of observation values ​​at the past e time points to obtain a third difference analysis value; Determining the linear independence of the first detection array according to the first difference analysis value, the second difference analysis value and the third difference analysis value; Determining the linear independence of the second detection array according to the past period vehicle flow evolution flow and the important data flow set includes: Determine the set of observation values ​​of the past f time points of the past periodic vehicle flow evolution flow, and the set of observation values ​​of the past g time points of each of the vehicle-road multimodal monitoring data flows in the important data flow set; wherein the set of observation values ​​of the past f time points includes the values ​​of all time points in the interval from time point f to the current time point of the past periodic vehicle flow evolution flow, and the set of observation values ​​of the past g time points includes the values ​​of all time points in the interval from time point g to the current time point of the vehicle-road multimodal monitoring data flows in the important data flow set; Performing difference analysis on the values ​​of two random time points in the set of observation values ​​at the past f time points to obtain a fourth difference analysis value corresponding to the values ​​of the two random time points; For any set of observations at the past g time points, complete the following steps: Performing difference analysis on the values ​​of two random time points in the set of observation values ​​at the past g time points to obtain a fifth difference analysis value corresponding to the values ​​of the two random time points; Superimposing each of the fourth difference analysis values ​​to obtain a fourth superimposed result, and superimposing each of the fifth difference analysis values ​​to obtain a fifth superimposed result; The linear independence of the second detection array is determined according to the fourth superposition result and the fifth superposition result.

2. The method according to claim 1, characterized in that: The step of performing correlation analysis on each of the past period vehicle flow evolution streams and each of the vehicle-road multimodal monitoring data streams includes: For any of the vehicle-road multimodal monitoring data streams, complete the following steps: Determine the first time point corresponding to the past periodic vehicle flow evolution flow according to the observation value set of the past a time points of the vehicle-road multimodal monitoring data stream and the observation value set of the past b time points of the past periodic vehicle flow evolution flow; The set of observation values ​​at the past a time points includes the values ​​of all time points in the interval from time point a to the current time point of the vehicle-road multimodal monitoring data stream, and the set of observation values ​​at the past b time points includes the values ​​of all time points in the interval from time point b to the current time point of the past period vehicle flow evolution stream, the time point a and the time point b are any time points before the current time point, and the first time point is any time point after the current time point; Determine a second time point corresponding to the past periodic vehicle flow evolution flow according to a set of observation values ​​of the past b time points of the past periodic vehicle flow evolution flow, wherein the second time point is consistent with the time point corresponding to the first time point; The first time point and the second time point are subjected to variance consistency detection, and the variance consistency detection result obtained is used as the correlation analysis result.

3. The method according to claim 2, characterized in that The adding each of the vehicle-road multimodal monitoring data streams to an important data stream set or a secondary data stream set according to the correlation analysis result includes: If the correlation analysis result is not less than the correlation analysis reference value, adding the vehicle-road multimodal monitoring data stream to the important data stream set; If the correlation analysis result is less than the correlation analysis reference value, the vehicle-road multimodal monitoring data stream is added to the secondary data stream set.

4. The method according to claim 1, characterized in that The method further comprises: According to the target vehicle-road multimodal monitoring data stream, the important data stream set and the past period vehicle flow evolution stream, the reasoning is performed through a plurality of the basic reasoning networks to obtain a basic reasoning set corresponding to each of the basic reasoning networks; Each of the basic reasoning sets is fused to obtain the inferred vehicle traffic evolution flow corresponding to the past period vehicle traffic evolution flow.

5. The method according to claim 4, characterized in that The basic reasoning network includes an independent basic reasoning network, a domain knowledge reasoning network and a comprehensive basic reasoning network; the reasoning is performed through a plurality of the basic reasoning networks according to the target vehicle-road multimodal monitoring data stream, the important data stream set and the past period vehicle flow evolution stream, to obtain a basic reasoning set corresponding to each basic reasoning network, including: According to the vehicle flow evolution flow of the past period, independent reasoning is performed one by one through the multiple independent basic reasoning networks to obtain independent basic reasoning sets corresponding to the multiple independent basic reasoning networks respectively; According to the past period vehicle flow evolution flow and road prior data, reasoning is performed through the domain knowledge reasoning network to obtain an expert basic reasoning set corresponding to the domain knowledge reasoning network; According to the target vehicle-road multimodal monitoring data stream, the important data stream set and the past period vehicle flow evolution stream, comprehensive reasoning is performed one by one through multiple comprehensive basic reasoning networks to obtain comprehensive basic reasoning sets corresponding to multiple comprehensive basic reasoning networks.

6. The method according to claim 5, characterized in that The types of independent basic inference networks include Holt-Winters smoothing networks, trend elimination networks, and data stream component splitting networks; The independent reasoning is performed one by one through the multiple independent basic reasoning networks according to the vehicle flow evolution flow of the past period, and the independent basic reasoning sets corresponding to the multiple independent basic reasoning networks are obtained, including: According to the vehicle flow evolution flow of the past period, independent reasoning is performed one by one through the Holt-Winters smoothing network, the trend elimination network and the data stream component splitting network to obtain independent basic reasoning sets corresponding to the Holt-Winters smoothing network, the trend elimination network and the data stream component splitting network respectively.

7. The method according to claim 5, characterized in that The types of the comprehensive basic reasoning network include a multivariate time series regression network and an LSTM network. According to the target vehicle-road multimodal monitoring data stream, the set of important data streams and the past period vehicle flow evolution stream, comprehensive reasoning is performed one by one through multiple comprehensive basic reasoning networks to obtain comprehensive basic reasoning sets corresponding to multiple comprehensive basic reasoning networks, including: according to the target vehicle-road multimodal monitoring data stream, the set of important data streams and the past period vehicle flow evolution stream, comprehensive reasoning is performed one by one through the multiple time series regression network and the LSTM network to obtain comprehensive basic reasoning sets corresponding to the multiple time series regression network and the LSTM network.

8. 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 7 are implemented.

Citation Information

Patent Citations

  • Network flow anomaly detection method

    CN110868431A

  • Road network mixed flow diurnal variation prediction method considering unmanned vehicle

    CN110942626A

  • Vehicle management platform flow real-time monitoring and management method fusing multiple modes

    CN115689847A

  • Method, system and equipment for identifying indoor charging behavior of electric bicycle and medium

    CN117194924A

  • Intelligent traffic management method for adjustment based on traffic flow

    CN117351702A