Intelligent dynamic road management system for smart park

Abstract

The invention discloses a smart park intelligent dynamic road management system, and relates to the technical field of smart park road management, and the system comprises a collection and feedback module which collects original data, constructs a traffic track matrix to carry out road flow prediction, obtains a flow prediction value, carries out the real-time feedback, and obtains a final prediction value; the construction and detection module is used for constructing a simulation model, recognizing a road bottleneck in combination with the final predicted value, and then calculating the queuing amount for optimization detection to obtain a detection result; the clustering and optimizing module is used for carrying out trajectory clustering based on the detection result and then optimizing total delay in combination with a BPR function and a Frank-Wolfe algorithm to obtain a final optimization result; according to the method, the flow distribution proportion of each road can be dynamically adjusted, so that the overall traffic delay of the park roads is reduced on the premise of meeting the traffic demand.

Description

Technical Field

[0001] The present invention relates to the technical field of smart park road management, and in particular to a smart park intelligent dynamic road management system. Background Art

[0002] With the deepening of smart city construction, smart parks, as typical scenarios for refined urban management, have become a hot topic in the field of intelligent transportation. Current park road management technology mainly relies on the Internet of Things perception layer to collect traffic flow data, combined with traditional time series models or shallow machine learning algorithms for short-term traffic forecasting. In recent years, deep learning technology has gradually been applied to the field of traffic forecasting. For example, convolutional neural networks are used to extract spatial features and long-short-term memory networks are used to capture temporal dependencies to form a hybrid model to improve prediction accuracy.

[0003] However, the existing intelligent dynamic road management systems for smart parks find it difficult to fully capture the dynamic correlations between vehicle trajectories, resulting in insufficient generalization capabilities of the prediction models. Most of these systems remain at the static optimization level, lacking real-time response to traffic flow changes and closed-loop feedback mechanisms, making it impossible to dynamically intervene in bottleneck sections. Furthermore, most optimization methods are based on heuristic or linear algorithms, making it difficult to balance global optimality and local convergence speed. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides an intelligent dynamic road management system for a smart park, which solves the problems of insufficient generalization ability of the prediction model, inability to dynamically intervene in bottleneck sections, and difficulty in balancing global optimality and local convergence speed.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a smart park intelligent dynamic road management system, which includes:

[0008] The collection and feedback module collects raw data, constructs a traffic trajectory matrix to predict road flow, and provides real-time feedback after obtaining the traffic prediction value to obtain the final prediction value;

[0009] The real-time feedback refers to re-optimizing the parameters of the prediction model or enhancing the residual of the flow prediction value using the ICDEBOA algorithm;

[0010] The construction and detection module builds a simulation model, identifies road bottlenecks based on the final prediction value, calculates the queue volume, performs optimization detection, and obtains the detection results;

[0011] The clustering and optimization module clusters the trajectories based on the detection results, and then optimizes the total delay by combining the BPR function and the Frank-Wolfe algorithm to obtain the final optimization result.

[0012] The display and storage module displays and stores the final optimization results.

[0013] As a preferred solution of the smart park intelligent dynamic road management system of the present invention, wherein: the collecting of raw data, constructing a traffic trajectory matrix to predict road flow, and providing real-time feedback after obtaining the flow prediction value to obtain the final prediction value include the following steps:

[0014] Equip each vehicle in the park with an RFID tag, and install an RFID reader at each road entrance in the park to collect traffic data and road flow data, and obtain vehicle trajectory data from the map API through the API interface;

[0015] Use sliding window technology to divide the traffic data and traffic flow data by time point and construct a traffic trajectory matrix;

[0016] A prediction model is constructed using a convolutional neural network and a long short-term memory network architecture. The traffic trajectory matrix is ​​input into the final prediction model, and the traffic flow prediction vector is output.

[0017] The traffic prediction vector includes the traffic prediction value of each road at each time point;

[0018] The absolute value is used to calculate the absolute error between the traffic prediction value and the current actual value, and the mean formula is used to calculate the average absolute error of all roads, which is defined as the current feedback value. The absolute error of each road is compared with the current feedback value. When the absolute error is greater than the current feedback value, the ICDEBOA algorithm is used to re-optimize the parameters of the prediction model. Otherwise, the residual of the traffic prediction value is enhanced.

[0019] As a preferred solution of the smart park intelligent dynamic road management system of the present invention, the residual enhancement of the traffic prediction value includes the following steps:

[0020] After collecting the historical traffic data and road flow data of the park and performing time alignment and normalization, a historical traffic trajectory matrix was constructed;

[0021] Obtain regional road network data of the park roads through the API interface and build an adjacency matrix;

[0022] Using subtraction operation, the difference between the road flow data and the flow prediction value is calculated, which is defined as the flow residual;

[0023] The roads within the park are regarded as nodes, the traffic residuals are used as node features, and the weights of the edges between nodes are defined through the adjacency matrix to construct a road topology graph.

[0024] Build correction models using graph convolutional networks and gated temporal convolutional network architectures;

[0025] Input the historical traffic trajectory matrix and road topology map into the correction model and output the final correction residual;

[0026] Using the addition operation, the sum of the corrected residual and the flow prediction value is calculated and defined as the final prediction value.

[0027] As a preferred solution of the smart park intelligent dynamic road management system of the present invention, the re-optimization of the parameters of the prediction model using the ICDEBOA algorithm includes the following steps:

[0028] Use the parameter combination of the current prediction model as an individual, randomly generate a population, traverse the individuals in the population, and use SineMap to perturb each individual and initialize the population;

[0029] The parameter combination includes the convolution kernel size, the number of hidden units in the LSTM layer, and the learning rate;

[0030] The mean square error is further used as the objective function, and the objective function value is calculated as the individual fitness value, and it is iteratively updated. The scaling factor is set using the network search method, and different individuals are randomly selected from the individual fitness values ​​to generate mutant individuals. A random number generator is used to generate uniformly distributed random numbers, and the random numbers are iteratively mapped using the Sine mapping formula to obtain the disturbance value. The disturbance value is input into the Sine function to obtain the disturbance term. The exponential decay function is used to combine the disturbance term to calculate the perception coefficient and update the position of each individual.

[0031] Use the domain method to set the crossover probability value, and use the crossover probability value to cross the mutant individual with the original individual to generate offspring, calculate the fitness value of the offspring and the parent generation respectively, when the offspring's fitness value is less than the parent's fitness value, the offspring and the parent generation are replaced, otherwise the parent generation is retained;

[0032] According to the replacement results, the scaling factors and crossover probabilities corresponding to the offspring are integrated to obtain the scaling factor set and the crossover probability set. The Lehmer mean formula is used to calculate the mean of the scaling factors and crossover probabilities in the scaling factor and crossover probability sets, respectively, and define them as the new scaling factors and new crossover probabilities. The iteration is continued. When the number of iterations reaches the maximum number, the optimal model parameter combination is output.

[0033] The optimal model parameter combination includes the convolution kernel size, the number of hidden units in the LSTM layer, and the learning rate;

[0034] The optimal model parameter combination is replaced with the original parameters of the prediction model, and the traffic prediction vector is re-output through the prediction model.

[0035] As a preferred solution of the smart park intelligent dynamic road management system of the present invention, wherein: the construction of the simulation model, combining the final prediction value to identify the road bottleneck, calculating the queue volume for optimization detection, and obtaining the detection result includes the following steps:

[0036] Use the VISSIM tool to build a network simulation model of the campus roads and simulate the traffic flow of each road in the campus under different traffic densities. During the simulation process, when a critical point of traffic flow appears on a road, this critical point is defined as the traffic bottleneck of that road. Using the proportional formula, the ratio of the final predicted value to the traffic bottleneck is calculated and defined as the bottleneck warning value.

[0037] Use the domain method to set the judgment threshold, compare the bottleneck warning value with the judgment threshold, and when the bottleneck warning value is greater than or equal to the judgment threshold, the road is judged to be an entrance and exit bottleneck; otherwise, the road is judged to be an intersection bottleneck.

[0038] The VISSIM tool was further used to construct road sections with gates or intersections within the park. After repeated simulations of gate traffic flow and intersection weaving flow, the maximum values ​​were taken as the minimum critical flow of the gate and intersection bottlenecks, respectively. The objective function was defined to minimize the bottleneck flow.

[0039] Calculate the objective function value and iterate. When the number of iterations reaches the maximum, output the optimal predicted traffic vector.

[0040] The optimal predicted flow vector includes the optimal predicted flow value of each road;

[0041] The disturbance term and error term are set using the regression analysis method, and the road queue volume is calculated in combination with the optimal predicted flow value. When the queue volume is less than the critical flow, it means that the current road queue volume is acceptable and no optimization is required. Otherwise, it means that road flow optimization is required.

[0042] As a preferred solution of the smart park intelligent dynamic road management system of the present invention, wherein: after performing trajectory clustering based on the detection results, optimizing the total delay by combining the BPR function and the Frank-Wolfe algorithm to obtain the final optimization result includes the following steps:

[0043] Based on the vehicle trajectory data of each vehicle, the starting and ending coordinates of the vehicle in the park are extracted, and OD demand pairs are generated. The OD demand pairs are used as points, and the distance between the current point and all adjacent points is calculated using Euclidean distance, and the points are arranged in ascending order. A curve graph is drawn based on the arrangement results, and the distance to the inflection point in the curve graph is selected as the neighborhood radius. The neighborhood radius is compared with the distance to obtain the clustering result and construct the OD demand matrix;

[0044] The OD demand refers to the demand of each vehicle from the park entrance to the desired location or from the current location to the exit;

[0045] Based on the optimal predicted traffic value, the BPR delay function is used to calculate the road time consumption and obtain the expected delay time of each road. The objective function is defined to minimize the total delay time, the objective function value is calculated, and iteration is performed. During the iteration process, the weight coefficient is set using the cross-validation method, and the BPR delay differential formula is used to combine the optimal predicted traffic value and critical traffic in the current iteration process to calculate the delay change rate of each road, which is defined as the load change gradient, and calculate the direction vector;

[0046] Further, based on the final prediction value, the partial derivative of each road is calculated using the total delay function, and the final prediction value of each road is expanded using the chain rule to obtain the delay gradient of each road;

[0047] Use the second-order differential approximation method combined with the final predicted value to approximate the second-order derivative of each road, integrate it, obtain the Hessian matrix, and calculate the step size;

[0048] The update logic formula of the Frank-Wolfe algorithm is combined with the direction vector and step size to calculate the new road flow distribution vector. The convergence threshold is set using the empirical rule and the L2 norm is used to calculate the difference between the new road flow distribution vector and the original road flow distribution vector. The difference is then compared with the convergence threshold. When the difference is less than the convergence threshold, the iteration is stopped and the new road flow distribution vector is used as the optimal distribution solution. Otherwise, the direction vector is recalculated.

[0049] The optimal allocation solution refers to the optimal available traffic flow and minimum delay time for each road.

[0050] As a preferred solution of the smart park intelligent dynamic road management system described in the present invention, the recalculation of the direction vector refers to recalculating the direction vector when the difference is greater than or equal to the convergence threshold, and performing the step size, new road flow distribution and difference calculation in turn, and then comparing it with the convergence threshold again. If the difference is still greater than the convergence threshold after the maximum number of iterations, all differences in the iteration process are arranged in ascending order, and the new road flow distribution corresponding to the minimum difference in the arrangement result is taken as the optimal allocation solution.

[0051] As a preferred solution of the smart park intelligent dynamic road management system described in the present invention, the display based on the final optimization results refers to pushing the optimal allocation plan in real time through the park's large-screen display, and displaying the current park road status and specific traffic information through a path map.

[0052] As a preferred solution of the smart park intelligent dynamic road management system described in the present invention, the storage refers to storing the optimal allocation plan through a database, and after storage, adding a timestamp to each data and encrypting it using a symmetric encryption algorithm.

[0053] As a preferred solution of the smart park intelligent dynamic road management system described in the present invention, after collecting the traffic data, road flow data and vehicle trajectory data, preprocessing operations are first performed, including denoising and normalization.

[0054] The present invention has the following beneficial effects: by introducing the ICDEBOA algorithm, the Sine mapping perturbation mechanism, the exponential decay perception coefficient, and the Lehmer mean strategy, the present invention has stronger global convergence ability and stability against falling into local extreme values ​​in complex nonlinear spaces. By adaptively iteratively optimizing the key parameter combination in the convolutional neural network, the generalization performance and prediction accuracy of the traffic flow prediction model under diverse campus traffic patterns are significantly improved. In addition, by utilizing a residual correction model based on a graph convolutional network and a gated temporal convolutional network, the spatial-temporal dependency characteristics of historical trajectory data and road topology are fully exploited. Feature extraction and correction are performed through a multi-level temporal convolution structure, thereby deeply correcting the initial prediction results and significantly reducing prediction errors. Secondly, the nonlinear relationship between road flow and travel time is characterized based on the BPR delay function. The optimal flow distribution vector under multiple OD paths is solved with the help of the Frank-Wolfe algorithm. The direction vector, step size, and Hessian matrix are then iteratively calculated, enabling the present invention to dynamically adjust the flow distribution ratio of each road, thereby reducing the overall traffic delay of campus roads while meeting traffic demand. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0056] Figure 1 This is a structural diagram of the smart park intelligent dynamic road management system in Example 1.

[0057] Figure 2 This is a flow chart of the collection and feedback module in Example 1.

[0058] Figure 3 This is a flow chart of the clustering and optimization module in Example 1. DETAILED DESCRIPTION

[0059] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0060] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0061] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0062] Example 1, with reference to Figures 1 to 3 , which is the first embodiment of the present invention, provides a smart park intelligent dynamic road management system, including the following steps:

[0063] S1, the acquisition and feedback module, collects raw data, constructs a traffic trajectory matrix to predict road flow, and provides real-time feedback after obtaining the flow prediction value to obtain the final prediction value; the real-time feedback refers to re-optimizing the parameters of the prediction model using the ICDEBOA algorithm or enhancing the residual of the flow prediction value;

[0064] Specifically, collecting raw data, constructing a traffic trajectory matrix to predict road flow, and providing real-time feedback after obtaining the traffic prediction value to obtain the final prediction value include the following steps:

[0065] Equip every vehicle in the park with an RFID tag, and install an RFID reader at every road entrance in the park to collect traffic data and road flow data;

[0066] Obtain vehicle trajectory data from the map API through the API interface;

[0067] Use sliding window technology to divide the traffic data and traffic flow data by time point and construct a traffic trajectory matrix;

[0068] Build predictive models using convolutional neural network and long short-term memory network architectures;

[0069] The prediction model includes an input layer, a convolutional layer, an LSTM layer, a fully connected layer and an output layer;

[0070] The convolutional layer includes a ReLU activation function;

[0071] The LSTM layer includes a gating mechanism;

[0072] Use mean square error as the loss function and use Adam optimizer to optimize and iterate the model parameters. During the iteration process, when the loss function decreases no longer significantly, stop the iteration and output the final prediction model.

[0073] The traffic trajectory matrix is ​​input into the final prediction model, convolution is performed through the convolution layer, and nonlinear transformation is performed through the ReLU activation function to obtain the spatial feature tensor. The spatial feature tensors are combined in time series, and the LSTM layer and gating mechanism are used to capture the temporal evolution of traffic flow in the spatial feature tensor to obtain the hidden state. After mapping through the fully connected layer, the traffic prediction vector is output through the output layer.

[0074] The traffic prediction vector includes the traffic prediction value of each road at each time point;

[0075] The absolute value is used to calculate the absolute error between the traffic prediction value and the current actual value, and the mean formula is used to calculate the average absolute error of all roads, which is defined as the current feedback value;

[0076] The absolute error of each road is compared with the current feedback value. When the absolute error is greater than the current feedback value, the ICDEBOA algorithm is used to re-optimize the parameters of the prediction model. Otherwise, the residual enhancement of the traffic prediction value is performed.

[0077] The ICDEBOA algorithm is a fusion optimization algorithm that integrates differential evolution algorithm, chaotic mapping and butterfly swarm optimization algorithm, namely "chaotic differential evolution butterfly swarm optimization algorithm". In the present invention, by using chaotic mapping to generate individual positions in the population initialization stage (rather than traditional random generation), the initial diversity of the population is improved, the search starting point is made more dispersed, which is conducive to global exploration. In the individual update stage, differential mutation and crossover mechanisms are introduced to strengthen and update the candidate solutions, improve the information interaction ability between individuals, thereby accelerating the convergence speed and enhancing the ability to jump out of the local optimum. Secondly, an adaptive parameter adjustment mechanism is introduced to make the perception coefficient, new scaling factor and new crossover probability change dynamically with iteration, so as to adapt to the search requirements of different stages (early stage of global exploration and late stage of local fine search);

[0078] By installing RFID tags on each vehicle in the park and RFID readers at each road entrance, the present invention can obtain vehicle entry and exit data in real time, providing raw data support for subsequent traffic flow monitoring and prediction. By obtaining vehicle trajectory data from the map API through the API interface, the present invention can further capture the specific driving trajectory of each vehicle and analyze the spatiotemporal changes of road flow and traffic density through these trajectory data. In addition, by combining convolutional neural networks with long short-term memory networks, the present invention can simultaneously capture the spatial characteristics and temporal evolution laws in traffic flow data. The convolutional neural network layer can efficiently extract the spatial characteristics of road flow data. , while the long short-term memory network layer can capture the changing trend of traffic flow over time. The two complement each other and effectively improve the spatiotemporal accuracy of the traffic prediction model, thereby providing more accurate traffic prediction in a dynamically changing traffic environment. Secondly, by providing feedback on the absolute error between the prediction result and the actual traffic data, the accuracy of the prediction model is further enhanced. First, by calculating the absolute error between the predicted value and the actual traffic value and comparing it with the current feedback value, when the error is large, the parameters of the prediction model will be further optimized. Through this real-time feedback mechanism, the present invention can automatically adjust and optimize after each prediction, thereby continuously improving the accuracy of traffic prediction.

[0079] Furthermore, residual enhancement of the traffic prediction value includes the following steps:

[0080] After collecting the historical traffic data and road flow data of the park and performing time alignment and normalization, a historical traffic trajectory matrix was constructed;

[0081] Obtain regional road network data of the park roads through the API interface and build an adjacency matrix;

[0082] Using subtraction operation, the difference between the road flow data and the flow prediction value is calculated, which is defined as the flow residual;

[0083] The roads within the park are regarded as nodes, the traffic residuals are used as node features, and the weights of the edges between nodes are defined through the adjacency matrix to construct a road topology graph.

[0084] Build the correction model using graph convolutional network and gated temporal convolutional network architecture, including the first, second, third and final layers;

[0085] The first, second and third layers all contain graph convolutional networks and gated temporal convolutional networks;

[0086] The mean square error is further used as the loss function, and the Adam optimizer is used to optimize and iterate the model parameters. During the iteration process, when the loss function no longer decreases significantly, the iteration is stopped and the final correction model is output.

[0087] The historical traffic trajectory matrix and the road topology map are input into the correction model. The causal convolution operation of the gated temporal convolutional network in the first layer is used to perform causal convolution on the historical traffic trajectory matrix to obtain the temporal features. The temporal features are combined with the road topology map using the graph convolutional network in the first layer to obtain the spatial topological features. The spatial topological features are then repeated in the second and third layers to obtain the final spatial topological features.

[0088] The final layer is used to perform weighted fusion on the spatial topological features output by each layer to obtain the final corrected residual;

[0089] Using the addition operation, the sum of the corrected residual and the flow prediction value is calculated and defined as the final prediction value.

[0090] By constructing a historical traffic trajectory matrix and combining it with a road topology map, deep learning of spatial and temporal features is performed using graph convolutional networks and gated temporal convolutional networks. This can capture the complex dynamic relationship of traffic flows between roads within the park. The causal convolution operation of the gated temporal convolutional network enables the present invention to process time series data and improve the temporal precision and accuracy of traffic prediction. This makes the present invention particularly suitable for complex park road networks and can provide real-time and accurate traffic prediction in actual park traffic environments. In addition, the present invention dynamically adjusts the prediction error through a multi-level deep learning model, and the correction of the traffic residual effectively reduces the deviation of the prediction result by weighted fusion of the output results of different layers, and enhances adaptability and stability. This enables the present invention to provide more accurate traffic prediction values ​​in real-time park traffic management, and quickly respond to sudden traffic flow changes, thereby optimizing the park's traffic scheduling. Through the deep fusion of graph convolutional networks and gated temporal convolutional networks, the present invention can fully explore the spatial-temporal dependence of park traffic flow changes, thereby optimizing traffic prediction and road capacity.

[0091] Furthermore, re-optimizing the parameters of the prediction model using the ICDEBOA algorithm includes the following steps:

[0092] Use the parameter combination of the current prediction model as an individual, randomly generate a population, traverse the individuals in the population, and use SineMap to perturb each individual and initialize the population;

[0093] The parameter combination includes the convolution kernel size, the number of hidden units in the LSTM layer, and the learning rate;

[0094] Further, the mean square error is used as the objective function, the objective function value is calculated as the individual fitness value, and it is iteratively updated;

[0095] Use the network search method to set the scaling factor, and randomly select different individuals from the individual fitness values ​​to generate mutant individuals. The formula is:

[0096] v i =x i1 +F×(x i2 -x i3 )

[0097] Where, v i represents the generated i-th mutant individual, x i1 represents the i1th individual randomly selected from the population, F represents the scaling factor, x i2 and x i3 They represent the i2th and i3th individuals randomly selected from the population respectively;

[0098] Use a random number generator to generate uniformly distributed random numbers, and use the Sine mapping formula to iteratively map the random numbers to obtain a disturbance value, which is then input into the Sine function to obtain a disturbance term.

[0099] The perceptual coefficient is calculated using an exponential decay function combined with a disturbance term. The formula is:

[0100]

[0101] In the formula, c represents the perception coefficient, e represents the base of the natural logarithm, G represents the current number of iterations, and A max represents the maximum number of iterations, sin(·) represents the Sine function, and r represents the disturbance value;

[0102] Combined with the perception coefficient, the position of each individual is updated using the formula:

[0103] p' i =p i +c×(v i -p i )

[0104] Where p′ i represents the updated position of the i-th individual, p i Indicates the current position of the i-th individual;

[0105] Use the domain method to set the crossover probability value, and use the crossover probability value to cross the mutant individual with the original individual to generate offspring, calculate the fitness value of the offspring and the parent generation respectively, when the offspring's fitness value is less than the parent's fitness value, the offspring and the parent generation are replaced, otherwise the parent generation is retained;

[0106] According to the replacement result, the scaling factors and crossover probabilities corresponding to the offspring are integrated respectively to obtain a scaling factor set and a crossover probability set;

[0107] Use the Lehmer mean formula to calculate the mean of the scaling factor and crossover probability in the scaling factor and crossover probability sets, respectively, and define them as the new scaling factor and the new crossover probability;

[0108] Continue to iterate, and when the number of iterations reaches the maximum, output the optimal model parameter combination;

[0109] The optimal model parameter combination includes the convolution kernel size, the number of hidden units in the LSTM layer, and the learning rate;

[0110] The optimal model parameter combination is replaced with the original parameters of the prediction model, and the traffic prediction vector is re-output through the prediction model.

[0111] By using Sine mapping to perturb individuals in the population, the limitations of traditional optimization algorithm initialization are broken. Sine mapping can increase the diversity of the population, thereby improving the coverage of the search space. This method helps the present invention avoid premature convergence to the local optimal solution during the search process, allowing the algorithm to explore a wider solution space during the optimization process, improving the global optimization ability, and taking the convolution kernel size, the number of LSTM hidden units and the learning rate as optimization targets. The algorithm automatically selects the most appropriate parameter combination. The appropriate convolution kernel size can better extract spatial features, and the reasonable number of LSTM layers and hidden units ensures the accuracy and generalization ability of time series modeling, while the optimized learning rate ensures the stability and efficiency of the training process. This automated optimization method reduces the complexity of manual parameter adjustment and can improve the prediction accuracy and reliability of the present invention. The introduction of the perception coefficient enables the exploration process of individuals in the search space to gradually converge according to the progress of the algorithm iteration, avoiding the phenomenon of over-exploration. Combined with the exponential decay function, the gradual reduction of the perception coefficient can enable the algorithm to converge stably in the later iteration process, ensuring that the global optimal solution is found. This mechanism effectively improves the adaptability of the algorithm in complex high-dimensional problems, optimizes the convergence speed and stability of the traffic prediction model, and further enhances the diversity of the population by performing mutation and crossover operations on individuals, enabling the algorithm to jump out of the local optimal solution and find a better solution. The mutation operation ensures that the algorithm will not be restricted in the global search by randomly adjusting the individual parameters, while the crossover operation helps different individuals to exchange information and improves the propagation efficiency of excellent features in the population. These operations jointly enhance the effect of the optimization process and significantly improve the adaptability and overall performance of the prediction model parameters in the present invention. Secondly, the use of the Lehmer mean ensures the trade-off in the update process of multiple parameters, so that some more important parameters can obtain a larger optimization range. The weighted average method optimizes the update strategy of multiple parameters, thereby improving the accuracy and stability of model training, and can automatically output the optimal model parameter combination (the optimal convolution kernel size, the number of hidden units in the LSTM layer, and the learning rate combination), so that the traffic prediction model performs well in various traffic scenarios, significantly improving the prediction accuracy and reducing the prediction error.

[0112] After the prediction result is generated, the present invention calculates the absolute error between the predicted value and the actual value through the feedback mechanism, and compares it with the mean error feedback value to trigger the adaptive optimization module or residual enhancement module of the present invention, which enables the present invention to form a dynamic closed loop of prediction-feedback-correction, so that the present invention can respond quickly according to the real-time traffic status of the park. If the error exceeds the threshold, the present invention will call the ICDEBOA evolutionary optimization algorithm to reconstruct the prediction model parameters online. This algorithm combines the Sine mapping perturbation and exponential decay perception mechanism to dynamically generate hyperparameter combinations such as convolution kernel size, number of LSTM units and learning rate. Through the Lehmer mean optimization scaling factor and crossover probability, the balance between the stability of parameter update and the exploration ability is achieved. ICDEBOA not only fine-tunes the prediction model parameters, but also realizes the original prediction model. Collaborative iteration enables the present invention to have the ability to continuously evolve and significantly improve its adaptability to sudden traffic changes. When the error is small, the present invention enters the residual enhancement path, constructs a topological graph with roads as nodes and flow residuals as features, and combines historical flow data to form a residual correction graph model. The graph convolutional network is responsible for capturing the spatial dependencies between roads, and the gated time convolutional network models the transmission law of flow residuals in the time dimension. The two alternately extract features through a multi-level structure and perform weighted fusion in the final layer to output accurate residual correction values. The residual correction module not only improves the prediction accuracy, but also complements the ICDEBOA module - one optimizes the prediction ability from the "model parameter space", and the other enhances the prediction results from the "residual space". A cross-compensation linkage is formed between the two, which enables the present invention to effectively avoid the limitations of single-path optimization.

[0113] S2, the construction and detection module, builds a simulation model, and after identifying the road bottleneck based on the final prediction value, calculates the queue volume for optimization detection and obtains the detection results;

[0114] Specifically, after building a simulation model and identifying road bottlenecks based on the final prediction value, the queue volume is calculated for optimization detection. Obtaining the detection results includes the following steps:

[0115] Use the VISSIM tool to build a network simulation model of the campus roads and simulate the traffic flow of each road in the campus under different traffic densities. During the simulation process, when a critical point of traffic flow appears on a road, this critical point is defined as the traffic bottleneck of that road.

[0116] Use the proportional formula to calculate the ratio of the final predicted value to the traffic bottleneck, which is defined as the bottleneck warning value;

[0117] Use the domain method to set the judgment threshold, compare the bottleneck warning value with the judgment threshold, and when the bottleneck warning value is greater than or equal to the judgment threshold, the road is judged to be an entrance and exit bottleneck; otherwise, the road is judged to be an intersection bottleneck.

[0118] The VISSIM tool was further used to construct road sections with gates or intersections within the park, and after repeated simulations of gate traffic flow and intersection weaving flow, the maximum values ​​were taken as the minimum critical flow of the gate and intersection bottlenecks respectively;

[0119] Define the objective function to minimize the bottleneck flow, the formula is:

[0120] f=∑ j∈B (q j -b j ) 2

[0121] Where f represents the objective function value, B represents the bottleneck path set, and q j represents the final predicted value of the jth road, b j represents the minimum critical flow of the jth road, (·) 2 represents the square of the error;

[0122] Calculate the objective function value and iterate. When the number of iterations reaches the maximum, output the optimal predicted traffic vector.

[0123] The optimal predicted flow vector includes the optimal predicted flow value of each road;

[0124] The disturbance term and error term are set using regression analysis, and combined with the optimal predicted traffic value, the road queue volume is calculated using the formula:

[0125]

[0126] Where u j (t) represents the optimal predicted flow value of the jth road at time t, k j (t) represents the disturbance term of the j-th road at time t, represents the minimum critical flow of the jth road, Δt represents the time interval, δ j (t) represents the error term of the j-th road at time t;

[0127] When the queue volume is less than the critical flow rate, it means that the current road queue volume is acceptable and no optimization is required; otherwise, it means that road flow optimization is required.

[0128] By using the VISSIM tool to create a simulation model of the campus road network and simulating traffic conditions under different traffic density conditions, the traffic flow changes during peak and trough periods can be accurately simulated. The simulation model can provide data support based on real scenarios for traffic flow management, enhance the foresight and response capabilities to traffic fluctuations, and by accurately simulating traffic density and bottlenecks, managers can effectively prevent congestion and make timely road optimization adjustments. The ratio of the final predicted value to the traffic bottleneck is calculated using a proportional formula, which is defined as the bottleneck warning value. The comparison is combined with the judgment threshold set by the domain method to accurately determine different types of road bottlenecks. This process distinguishes between entrance and exit gate bottlenecks and intersection bottlenecks, providing more detailed decision-making basis for subsequent optimization steps, realizing accurate classification of road bottlenecks, and avoiding misjudgments caused by excessive or insufficient traffic volume. By distinguishing different types of bottlenecks, the present invention can adopt customized optimization strategies for specific problems, thereby effectively reducing campus The traffic congestion phenomenon in the park is analyzed by setting the disturbance term and error term through regression analysis method, and combining with the optimal predicted flow value to calculate the queue volume of the road, so that the present invention can accurately predict the actual queue situation of each road and ensure that the traffic flow does not exceed the maximum carrying capacity of the bottleneck. The use of regression analysis method improves the prediction accuracy of the queue volume and provides a more refined control means for road flow optimization. Secondly, through real-time monitoring and dynamic adjustment, the present invention can efficiently handle the critical queue volume and avoid traffic congestion or ineffective waiting. When the queue volume exceeds the critical flow, it is determined that flow optimization is required. This judgment mechanism enables the present invention to make flow optimization decisions based on real-time data and predicted information, thereby quickly responding to changes in traffic flow and ensuring stable traffic order in the park. Through the interactive judgment logic of queue volume and critical flow, the present invention realizes the extraction of decision-making basis from predicted values. Only when real congestion is about to form or has formed will the optimization process be entered, which greatly improves the resource scheduling efficiency and algorithm execution efficiency.

[0129] S3, clustering and optimization module, after clustering the trajectories based on the detection results, combines the BPR function and the Frank-Wolfe algorithm to optimize the total delay and obtain the final optimization result;

[0130] Specifically, after clustering the trajectories based on the detection results, the BPR function and the Frank-Wolfe algorithm are combined to optimize the total delay. The final optimization result includes the following steps:

[0131] Based on the vehicle trajectory data of each vehicle, the starting and ending coordinates of the vehicle in the park are extracted, and OD demand pairs are generated. The OD demand pairs are used as points, and the distance between the current point and all adjacent points is calculated using Euclidean distance, and the points are arranged in ascending order. A curve graph is drawn based on the arrangement results, and the distance to the inflection point in the curve graph is selected as the neighborhood radius. The neighborhood radius is compared with the distance to obtain the clustering result;

[0132] Based on the clustering results, the OD demand matrix is ​​constructed;

[0133] The OD demand refers to the demand of each vehicle from the park entrance to the desired location or from the current location to the exit;

[0134] According to the optimal predicted traffic value, the BPR delay function is used to calculate the road time consumption and obtain the expected delay time of each road;

[0135] Define the objective function to minimize the total delay time. The formula is:

[0136] M=∑ j∈B u j (t)·h j (t)

[0137] Where M represents the objective function value, h j (t) represents the expected delay time of the jth road at time t;

[0138] Calculate the objective function value and iterate. During the iteration process, use the cross-validation method to set the weight coefficient, and use the BPR delay differential formula combined with the optimal predicted flow value and critical flow in the current iteration process to calculate the delay change rate of each road, which is defined as the load change gradient.

[0139] The direction vector is calculated based on the weight coefficient, the expected delay time, and the load change gradient. The formula is:

[0140]

[0141] Where d′(z) represents the direction vector at the zth iteration, represents the weight coefficient, s(z) represents the expected delay time at the zth iteration, o(z) represents the traffic distribution vector at the zth iteration, and g(z) represents the load change gradient at the zth iteration;

[0142] Further, based on the final prediction value, the partial derivative of each road is calculated using the total delay function, and the final prediction value of each road is expanded using the chain rule to obtain the delay gradient of each road;

[0143] The second-order derivative of each road is approximately calculated using the second-order differential approximation method combined with the final predicted value, and then integrated to obtain the Hessian matrix;

[0144] According to the direction vector, the step length is calculated as follows:

[0145]

[0146] In the formula, γ(z) represents the step size of the zth iteration, represents the delayed gradient of the flow distribution vector at the zth iteration, H(o(z)) represents the Hessian matrix of the flow distribution vector at the zth iteration, and <·> represents the vector inner product;

[0147] The Frank-Wolfe algorithm's update logic formula is combined with the direction vector and step size to calculate the new road flow distribution vector.

[0148] The convergence threshold is set using empirical rules, and the L2 norm is used to calculate the difference between the new road flow distribution vector and the original road flow distribution vector. The difference is then compared with the convergence threshold. When the difference is less than the convergence threshold, the iteration is stopped and the new road flow distribution vector is used as the optimal distribution solution. Otherwise, the direction vector is recalculated.

[0149] The optimal allocation solution refers to the optimal available traffic flow and minimum delay time for each road.

[0150] By clustering the trajectory data of each vehicle, the OD demand pair is first calculated according to the starting and ending positions of each vehicle, and the neighborhood radius is determined using the Euclidean distance. This can effectively cluster the traffic demand patterns within the park and provide accurate demand data for subsequent traffic optimization. By generating OD demand pairs, it can ensure that the subsequent optimization process is more accurate, avoiding overly rough traffic predictions, thereby improving prediction accuracy. The application of the BPR delay function can calculate the expected delay time based on the traffic prediction value of each road, thereby providing benchmark data for traffic flow optimization. By considering the relationship between traffic and capacity of different roads, the present invention can avoid the common low-precision delay estimation problem and improve the performance of the present invention during peak hours. Or adaptability under special conditions, and optimize road flow distribution through the Frank-Wolfe algorithm. Combined with the direction vector, step size and Hessian matrix, the present invention can gradually adjust the road flow in each iteration to minimize the total delay time. The global optimal solution capability of the Frank-Wolfe algorithm makes the final flow distribution plan more stable and can cope with traffic flow changes in complex road network environments. Secondly, by setting the convergence threshold and using the L2 norm for difference calculation, the present invention can accurately find the optimal flow distribution plan after multiple rounds of iterations. Compared with the simple single-round optimization method, this process can gradually correct the errors in the initial calculation, and ultimately achieve lower delay time and more efficient flow scheduling.

[0151] Furthermore, when the recalculation of the direction vector indicates that the difference is greater than or equal to the convergence threshold, the direction vector is recalculated, and the step size, new road flow distribution and difference are calculated in sequence, and then compared with the convergence threshold again. If the difference is still greater than the convergence threshold after the maximum number of iterations, all differences in the iterative process are arranged in ascending order, and the new road flow distribution corresponding to the smallest difference in the arrangement result is taken as the optimal allocation plan.

[0152] By comparing the difference with the convergence threshold in real time and dynamically updating the direction vector, the optimized path is made more stable. It is especially suitable for complex campus traffic networks with multiple OD paths. Once the direction vector is adjusted, the subsequent step size and flow updates are also dynamically optimized, so that each round of iteration can move towards a new and better minimum value direction, thereby avoiding falling into the local optimal solution. If the convergence standard is still not met after the maximum number of iterations, the traditional algorithm often directly abandons the current result. In this case, the present invention introduces a full-process difference record and minimum difference selection mechanism, sorts all intermediate results in ascending order, and selects the "minimum difference corresponding solution" as the final optimal approximate solution, effectively making up for the global suboptimal solution that may be missed in the traditional optimization process. Through the dual limit method of the maximum number of iterations and the difference judgment, the usage boundary of computing resources is effectively controlled, and the algorithm is avoided from ineffective iteration for a long time in complex networks, thereby improving the actual deployment efficiency.

[0153] S4, display and storage module, displays and stores the final optimization results;

[0154] Specifically, displaying the final optimization results means pushing the optimal allocation plan in real time through the park's large-screen display, and displaying the current park road status and specific traffic information through a path map;

[0155] By synchronizing the optimal plan and route map to the campus screen through the communication network, clear and intuitive information guidance is provided to managers and drivers, and the uniformity of management and scheduling is improved.

[0156] Furthermore, storage refers to storing the optimal allocation plan in a database, and after storage, adding a timestamp to each data and encrypting it using a symmetric encryption algorithm;

[0157] By storing the optimal allocation plan in a database, structured management of the optimization results is achieved, which facilitates subsequent calls and rapid deployment. The timestamp operation further enhances the capabilities of the present invention in data traceability and spatiotemporal dimension analysis, and the embedding of the symmetric encryption mechanism ensures the confidentiality and security of the data.

[0158] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. Smart Park Intelligent Dynamic Road Management System, characterized by: include, The collection and feedback module collects raw data, constructs a traffic trajectory matrix to predict road flow, and provides real-time feedback after obtaining the traffic prediction value to obtain the final prediction value; The real-time feedback refers to re-optimizing the parameters of the prediction model or enhancing the residual of the flow prediction value using the ICDEBOA algorithm; The construction and detection module builds a simulation model, identifies road bottlenecks based on the final prediction value, calculates the queue volume, performs optimization detection, and obtains the detection results; The clustering and optimization module clusters the trajectories based on the detection results, and then optimizes the total delay by combining the BPR function and the Frank-Wolfe algorithm to obtain the final optimization result. The display and storage module displays and stores the final optimization results.

2. The intelligent dynamic road management system for a smart park according to claim 1, characterized in that: The steps of collecting raw data, constructing a traffic trajectory matrix to predict road flow, and providing real-time feedback after obtaining the traffic prediction value to obtain the final prediction value include the following steps: Equip each vehicle in the park with an RFID tag, and install an RFID reader at each road entrance in the park to collect traffic data and road flow data, and obtain vehicle trajectory data from the map API through the API interface; Use sliding window technology to divide the traffic data and traffic flow data by time point and construct a traffic trajectory matrix; A prediction model is constructed using a convolutional neural network and a long short-term memory network architecture. The traffic trajectory matrix is ​​input into the final prediction model, and the traffic flow prediction vector is output. The traffic prediction vector includes the traffic prediction value of each road at each time point; The absolute value is used to calculate the absolute error between the traffic prediction value and the current actual value, and the mean formula is used to calculate the average absolute error of all roads, which is defined as the current feedback value. The absolute error of each road is compared with the current feedback value. When the absolute error is greater than the current feedback value, the ICDEBOA algorithm is used to re-optimize the parameters of the prediction model. Otherwise, the residual of the traffic prediction value is enhanced.

3. The intelligent dynamic road management system for a smart park according to claim 2, characterized in that: The residual enhancement of the flow prediction value comprises the following steps: After collecting the historical traffic data and road flow data of the park and performing time alignment and normalization, a historical traffic trajectory matrix was constructed; Obtain regional road network data of the park roads through the API interface and build an adjacency matrix; Using subtraction operation, the difference between the road flow data and the flow prediction value is calculated, which is defined as the flow residual; The roads within the park are regarded as nodes, the traffic residuals are used as node features, and the weights of the edges between nodes are defined through the adjacency matrix to construct a road topology graph. Build correction models using graph convolutional networks and gated temporal convolutional network architectures; Input the historical traffic trajectory matrix and road topology map into the correction model and output the final correction residual; Using the addition operation, the sum of the corrected residual and the flow prediction value is calculated and defined as the final prediction value.

4. The intelligent dynamic road management system for a smart park according to claim 2, characterized in that: The method of re-optimizing the parameters of the prediction model using the ICDEBOA algorithm comprises the following steps: Use the parameter combination of the current prediction model as an individual, randomly generate a population, traverse the individuals in the population, and use SineMap to perturb each individual and initialize the population; The parameter combination includes the convolution kernel size, the number of hidden units in the LSTM layer, and the learning rate; The mean square error is further used as the objective function, and the objective function value is calculated as the individual fitness value, and it is iteratively updated. The scaling factor is set using the network search method, and different individuals are randomly selected from the individual fitness values ​​to generate mutant individuals. A random number generator is used to generate uniformly distributed random numbers, and the random numbers are iteratively mapped using the Sine mapping formula to obtain the disturbance value. The disturbance value is input into the Sine function to obtain the disturbance term. The exponential decay function is used to combine the disturbance term to calculate the perception coefficient and update the position of each individual. Use the domain method to set the crossover probability value, and use the crossover probability value to cross the mutant individual with the original individual to generate offspring, calculate the fitness value of the offspring and the parent generation respectively, when the offspring's fitness value is less than the parent's fitness value, the offspring and the parent generation are replaced, otherwise the parent generation is retained; According to the replacement results, the scaling factors and crossover probabilities corresponding to the offspring are integrated to obtain the scaling factor set and the crossover probability set. The Lehmer mean formula is used to calculate the mean of the scaling factors and crossover probabilities in the scaling factor and crossover probability sets, respectively, and define them as the new scaling factors and new crossover probabilities. The iteration is continued. When the number of iterations reaches the maximum number, the optimal model parameter combination is output. The optimal model parameter combination includes the convolution kernel size, the number of hidden units in the LSTM layer, and the learning rate; The optimal model parameter combination is replaced with the original parameters of the prediction model, and the traffic prediction vector is re-output through the prediction model.

5. The intelligent dynamic road management system for a smart park according to claim 3, characterized in that: The construction of the simulation model, identification of the road bottleneck based on the final prediction value, calculation of the queue volume for optimization detection, and obtaining the detection results include the following steps: Use the VISSIM tool to build a network simulation model of the campus roads and simulate the traffic flow of each road in the campus under different traffic densities. During the simulation process, when a critical point of traffic flow appears on a road, this critical point is defined as the traffic bottleneck of that road. Using the proportional formula, the ratio of the final predicted value to the traffic bottleneck is calculated and defined as the bottleneck warning value. Use the domain method to set the judgment threshold, compare the bottleneck warning value with the judgment threshold, and when the bottleneck warning value is greater than or equal to the judgment threshold, the road is judged to be an entrance and exit bottleneck; otherwise, the road is judged to be an intersection bottleneck. The VISSIM tool was further used to construct road sections with gates or intersections within the park. After repeated simulations of gate traffic flow and intersection weaving flow, the maximum values ​​were taken as the minimum critical flow of the gate and intersection bottlenecks, respectively. The objective function was defined to minimize the bottleneck flow. Calculate the objective function value and iterate. When the number of iterations reaches the maximum, output the optimal predicted traffic vector. The optimal predicted flow vector includes the optimal predicted flow value of each road; The disturbance term and error term are set using the regression analysis method, and the road queue volume is calculated in combination with the optimal predicted flow value. When the queue volume is less than the critical flow, it means that the current road queue volume is acceptable and no optimization is required. Otherwise, it means that road flow optimization is required.

6. The intelligent dynamic road management system for a smart park according to claim 5, characterized in that: After performing trajectory clustering based on the detection results, optimizing the total delay by combining the BPR function with the Frank-Wolfe algorithm to obtain the final optimization result includes the following steps: Based on the vehicle trajectory data of each vehicle, the starting and ending coordinates of the vehicle in the park are extracted, and OD demand pairs are generated. The OD demand pairs are used as points, and the distance between the current point and all adjacent points is calculated using Euclidean distance, and the points are arranged in ascending order. A curve graph is drawn based on the arrangement results, and the distance to the inflection point in the curve graph is selected as the neighborhood radius. The neighborhood radius is compared with the distance to obtain the clustering result and construct the OD demand matrix; The OD demand refers to the demand of each vehicle from the park entrance to the desired location or from the current location to the exit; Based on the optimal predicted traffic value, the BPR delay function is used to calculate the road time consumption and obtain the expected delay time of each road. The objective function is defined to minimize the total delay time, the objective function value is calculated, and iteration is performed. During the iteration process, the weight coefficient is set using the cross-validation method, and the BPR delay differential formula is used to combine the optimal predicted traffic value and critical traffic in the current iteration process to calculate the delay change rate of each road, which is defined as the load change gradient, and calculate the direction vector; Further, based on the final prediction value, the partial derivative of each road is calculated using the total delay function, and the final prediction value of each road is expanded using the chain rule to obtain the delay gradient of each road; Use the second-order differential approximation method combined with the final predicted value to approximate the second-order derivative of each road, integrate it, obtain the Hessian matrix, and calculate the step size; The update logic formula of the Frank-Wolfe algorithm is combined with the direction vector and step size to calculate the new road flow distribution vector. The convergence threshold is set using the empirical rule and the L2 norm is used to calculate the difference between the new road flow distribution vector and the original road flow distribution vector. The difference is then compared with the convergence threshold. When the difference is less than the convergence threshold, the iteration is stopped and the new road flow distribution vector is used as the optimal distribution solution. Otherwise, the direction vector is recalculated. The optimal allocation solution refers to the optimal available traffic flow and minimum delay time for each road.

7. The intelligent dynamic road management system for a smart park according to claim 6, characterized in that: The recalculation of the direction vector means that when the difference is greater than or equal to the convergence threshold, the direction vector is recalculated, and the step size, new road flow distribution and difference are calculated in sequence, and then compared with the convergence threshold again. If the difference is still greater than the convergence threshold after the maximum number of iterations, all differences in the iterative process are arranged in ascending order, and the new road flow distribution corresponding to the smallest difference in the arrangement result is taken as the optimal distribution plan.

8. The intelligent dynamic road management system for a smart park according to claim 7, characterized in that: The display according to the final optimization result refers to pushing the optimal allocation plan in real time through the large-screen display of the park, and displaying the current park road status and specific traffic information through the path map.

9. The intelligent dynamic road management system for a smart park according to claim 8, characterized in that: The storage refers to storing the optimal allocation plan in a database, and after storage, adding a timestamp to each data and encrypting it using a symmetric encryption algorithm.

10. The smart park intelligent dynamic road management system according to claim 1, characterized in that: After the traffic data, road flow data and vehicle trajectory data are collected, pre-processing operations are first performed, including denoising and normalization.

Images