Smart city and intelligent traffic collaborative management method and system based on artificial intelligence

Abstract

The invention discloses a smart city and intelligent traffic collaborative management method based on artificial intelligence, and relates to the technical field of traffic collaborative management, and the method comprises the following steps: S1, data processing; s2, performing congestion assessment and road section division; s3, all-weather congestion management is carried out; step S4, peak period management and information optimization; and S5, carrying out comprehensive evaluation and optimization. The invention also discloses a smart city and intelligent traffic collaborative management system based on artificial intelligence. The system comprises a data acquisition unit, a historical database, a big data processing unit, a traffic light control module, an information display module, an early warning information sending module and a communication unit. According to the invention, abundant and accurate traffic data are acquired through multiple channels, and a wCGwNN improved algorithm is introduced into all-weather congestion road section management, so that traffic signal lamp intelligent agents cooperate with each other, and an optimal traffic management decision which comprehensively considers multiple factors is obtained.

Description

Technical Field

[0001] The present invention relates to the technical field of traffic collaborative management, and specifically to a method and system for collaborative management of smart cities and intelligent transportation based on artificial intelligence. Background Art

[0002] With the increase in the number of vehicles, urban traffic congestion will cause serious problems such as great economic losses and environmental pollution. The intelligent traffic management system is an effective method to reduce urban traffic congestion. Through modern communication information technology and optimization control theory, real-time traffic management and control are carried out according to the obtained traffic environment information, so as to improve the driving experience, reduce traffic congestion, reduce traffic accidents and reduce vehicle pollution emissions, which are important goals of intelligent traffic management.

[0003] The invention with the publication number CN110503834A discloses an intelligent traffic management decision-making method for realizing multi-intersection collaboration based on GAN. By using the generative adversarial network technology, global traffic state information is generated at each intersection according to its respective local state information, and then based on the generated global traffic state information, an intelligent traffic management decision-making for multi-intersection collaboration is realized. However, this method often focuses on a few conventional data sources to obtain traffic data, ignoring diverse data sources rich in travel information such as online car-hailing platforms, logistics vehicle positioning systems, and bus and subway card swiping records, resulting in incomplete traffic data collection, making it difficult to fully present the actual traffic situation, missing many traffic behavior details such as different travel purposes and different vehicle types, affecting the subsequent accurate judgment of the traffic situation, and lacking in-depth mining and analysis of the correlation between different types of data (such as road construction information, weather impact data, special event information, and conventional traffic flow data), unable to make full use of the mutual relationship between these data to more comprehensively grasp the traffic scenario, resulting in the inability to comprehensively consider the influence of various factors when formulating traffic management strategies. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention provides a method and system for collaborative management of smart cities and intelligent transportation based on artificial intelligence, and solves the existing problems.

[0005] To achieve the above objectives, the present invention is realized through the following technical solutions: A method for collaborative management of smart cities and intelligent transportation based on artificial intelligence includes the following steps:

[0006] Step S1, data processing: Integrate multi-source data such as Baidu Maps, Amap Navigation, the official system of the traffic management department, and intelligent traffic sensing devices, comprehensively collect road conditions and traffic flow information, and at the same time cover road construction, weather impact, and special event data, construct a complete traffic data set, clean and fill and correct the data, and extract spatio-temporal features for later use;

[0007] Step S2, Congestion Assessment and Road Section Division: Evaluate congestion using multiple indicators, build a machine learning model to optimize parameters according to the regular patterns of time periods and regions, subdivide by road grade and region, and classify the congested road sections throughout the day into mild, moderate, and severe levels;

[0008] Step S3, All-Weather Congestion Management

[0009] S3.1, Intelligent Decision-Making: Introduce a multi-agent reinforcement learning framework and the wCGwNN model, regard traffic lights as agents, and obtain the optimal decision in combination with the reward function;

[0010] S3.2, Collaborative Application: Push the optimal decision information to vehicles in real time through the vehicle networking, enabling the vehicles to adaptively adjust their driving states;

[0011] Step S4, Peak Period Management and Information Optimization

[0012] S4.1, Congestion Prediction and Diversion: Monitor the queue at intersections, predict congestion using a microscopic model, and set signs in advance for diversion;

[0013] S4.2, Traffic Light Timing Adjustment: According to the congestion and queue conditions, use an algorithm to adjust the traffic light duration ratio during peak hours, and set a differentiated green light plan;

[0014] S4.3, Route and Information Upgrade: Use deep learning to optimize the route prediction model, introduce probability prediction to accurately judge whether a vehicle passes through a congested road section, enrich the pushed information and optimize it according to the feedback;

[0015] Step S5, Comprehensive Evaluation and Optimization: Build a multi-dimensional index system to measure the effectiveness of the strategy, conduct regular quantitative evaluations, adjust the strategy parameters according to the results, form a closed-loop feedback mechanism, shorten the review cycle according to congestion, and timely verify and correct the data conclusions.

[0016] Preferably, in step S1, the collected data is cleaned, the interpolation method is used to fill in the missing values, and the outlier detection algorithm is used to correct the abnormal data to ensure the accuracy and integrity of the data. Through spatio-temporal correlation analysis, the periodic characteristics of the data (such as weekdays and weekends, different seasons, etc.) and the traffic flow correlation characteristics of the road sections are extracted, providing strong support for subsequent analysis.

[0017] Preferably, in step S2, a real-time model is constructed using machine learning algorithms. According to the traffic rules of different time periods (such as morning and evening rush hours, off-peak hours) and regions (commercial areas, residential areas, etc.), the data is dynamically analyzed, the congestion judgment parameters are automatically optimized, the changing trend of congestion is captured in real time, the roads are further subdivided according to the road grade (high-speed roads, urban arterial roads, etc.) and regions, and the all-weather congested road sections are refined into mild, moderate, and severe congestion levels, providing a basis for formulating targeted traffic management strategies.

[0018] Preferably, after the processed traffic data is structured and organized in step S3, it is input into the wCGwNN model, and sequentially passes through the convolution layer to extract local features, the pooling layer to reduce the data dimension, the GRUw module to capture time series features using wavelet functions, and the fully connected layer to integrate features. Combined with the reward function including vehicle waiting time, travel speed, exhaust emissions, energy consumption, and bus priority, it drives the intelligent agent to collaboratively optimize regional traffic flow and obtain the optimal decision.

[0019] Preferably, the calculation steps of the wCGwNN model include:

[0020] 【1】Prediction from images to indicators

[0021] 【1.1】Calculation formula from image to indicator

[0022] 1) First, the processed traffic-related data is converted into a 3D pixel structure and fed into the input layer of the wCGwNN. The input data consists of N images showing different information distributions in the same area. Each image corresponds to three color channels, for a total of N*3 channels. The N images can be: a real-time distribution map of traffic flow in the same area, a real-time distribution map of population in the same area, a real-time distribution map of other information in the same area, etc.

[0023] It can also be map information, population distribution information, and data mapping information (such as the impact of important changes or major changes on traffic flow). These images include traffic flow maps, population distribution maps, or traffic flow changes caused by special events. Each image corresponds to m labels, including regions, key points of attention, etc., which are used to guide the training and output of the neural network;

[0024] 2) Perform forward propagation calculations of the convolutional layer and pooling layer in the CPNN (convolutional neural network part) of the wCGwNN model. The calculation process is shown in the following formula:

[0025]

[0026] In the above formula, is the value of the position index (m,n) in the feature map output by the l-th layer neuron, where m and n are the position indexes before pooling; l-1 is the output of the l-1 layer; core l Is a two-dimensional matrix, as the convolution kernel of the lth layer; Conv(O l-1 , core l , m, n) is the convolution layer pair O l-1 Use convolution kernel core l Perform convolution operation; c is the number of channels in the photo; size l is the size of the convolution kernel; is the value at position (m + i, n + j) and channel number c in the feature map output by the (l - 1)-th layer; is the two-dimensional kernel part corresponding to the input channel c in the convolutional kernel of the l-th layer; b l , ε l is the bias value of the l-th layer, used to adjust the activation level of the output feature map;

[0027] The output of the convolutional layer needs to be calculated through the activation function ReLU, and the calculation process is as follows:

[0028]

[0029] There are two convolutional processes in the above formula, that is, each process contains two parts: convolution and activation. Func(Z) represents the activation function, and ReLU(Z) is defined as ReLU(z) = max(0, z). Let it be the initial input, is the result after the first convolutional process, is the pooling process after the first convolutional process, is the result after the second convolutional process, is the pooling process after the second convolution, is to output the data converted into 1D data;

[0030] Extract local features in traffic data through convolutional operations, and use pooling operations to reduce the data dimension, reduce the amount of calculation, and retain key features at the same time. The pooling formula is as follows:

[0031]

[0032] In the above formula, is the value at position index (m′, n′) in the feature map output by the l-th layer neuron, is the value at position index (m, n) in the feature map output by the (l - 1)-th layer neuron, Pooling 2×2 (Z) represents the pooling calculation; m′, n′ are the position indices of the feature map after pooling;

[0033] 3) Introduce a gated recurrent unit based on wavelet transform in the GRUw module for forward propagation calculation. The calculation process is as follows:

[0034]

[0035] The above formula includes two layers of GRU layers based on wavelet transform. Among them, x i is the input of the GRUw module, is the output of the last layer of the CPNN part, size 5 is the size of the feature map output by the 5th layer; Linear intermediate variables for the reset gate, update gate, and candidate state respectively; W x，r , W x，z , W x，h Weight matrices for the reset gate, update gate, and candidate state respectively; ε r , ε z , Bias terms for the reset gate, update gate, and candidate state respectively; z i Is the output of the update gate, controlling the fusion ratio of the current state and the previous state; W h，h Is the historical information (r i ·h i-1 ) after being adjusted by the reset gate to the weight matrix of the candidate state. r i Is the output of the reset gate, controlling the degree of forgetting of the previous hidden state; Is the candidate hidden state; h i Is the current hidden state, h i-1 Is the hidden state at the previous time step; σ is the Sigmoid activation function; tanh is the hyperbolic tangent function, and i represents the index of the time step;

[0036] 4) Perform forward propagation calculation of the fully connected layer in the FCNN. The FCNN part has a total of four layers. The last layer of the FCNN is the output layer of the CNN model, integrating the features extracted by the previous layers and mapping them to the dimension finally used for decision-making. The total number of neurons in the first layer of the FCNN is size -4 ×1 is equal to the number of neurons in the last layer of the CPNN part, size -5 ×size -5 . The output of the first layer of the FCNN And the output of the last layer of the CPNN Are converted. The conversion relationship between And Is shown by the following formula:

[0037]

[0038] The calculation process of the final output y n Of the forward calculation is as follows:

[0039]

[0040] In the above formula, O n Is the output of the first fully connected layer, where n represents n neurons in this layer; O k Is the output of the second fully connected layer, where k represents k neurons in this layer; w i,n , w n，k , w n，kis the fully-connected weight matrix, where i and n represent the weight from the i-th neuron to the n-th neuron, where n and k represent the weight from the n-th neuron to the k-th neuron, and where k and j represent the weight from the k-th neuron to the j-th neuron; ε n , ε k , ε j are the biases of the first, second, and third fully-connected layers, respectively; is the output of the first layer of the FCNN; is the output of the first wavelet transform, -4 indicating that this value is calculated at the fourth-to-last layer, i.e., the output of the fourth-to-last layer of the neural network; is the output of the second wavelet transform, -3 indicating that this value is calculated at the third-to-last layer, i.e., the output of the third-to-last layer of the neural network; is the output of the second-to-last output layer; is the output of the last layer of the CPNN; is the output of the wavelet transform activation; b n , b k correspond to the dynamic translation parameters of the first and second wavelet transforms; a n , a k correspond to the dynamic scale parameters of the first and second wavelet transforms; σ is the Sigmoid activation function; size l is the number of neurons in the l-th layer of the neural network; size -1 , size -2 , size -3 are the number of neurons in the last, second-to-last, and third-to-last layers of the neural network, respectively; softmax is the softmax function; is the output of the fully-connected layer is the exponential transformation of n -1 represents the final output of the neural network, -1 indicating the last layer of the neural network, i.e., the last layer of the neural network, and n representing the n-th label; the meaning of the n-th label of the output can be: the overall traffic index, the traffic index of a certain block;

[0041] 【1.2】Training Samples and Training Process

[0042] Each training sample is There are a total of M training samples, which form a training sample set, and the network is trained using the M training samples;

[0043] 【2】Neural Network Prediction

[0044] After training the network, for the real-time at future time t is input into the network as a prediction sample, and the network calculates the corresponding y n -1 , yn -1 The data provided for decision-making can obtain a total of S traffic control strategies Ps = f(y n -1 ), s = 1, 2,..., S;

[0045] 【3】Data processing and large model training

[0046] Data literalization: Each group of n y n -1 is converted into the corresponding literal description Ai. For example, a series of indicators output by the neural network (such as 0.9 (overall index), 0.1 (regional index 1), etc.) are converted into literal descriptions such as "The overall area is relatively congested (0.9), the southwest is not congested (0.1)..."

[0047] Construction of large model training samples: Ai and the corresponding output Ao of the large language model (traffic optimization suggestions) are formed into a pair of training samples S for the large language model. There are a total of S m samples;

[0048] Fine-tuning of the large model: These samples are used as fine-tuning training samples for the large language model (DeepSeek) to fine-tune the large model so that it can master the ability to obtain Ao based on Ai;

[0049] 【4】Strategies obtained by large model prediction

[0050] When the large model encounters the newly input y n -1 and the converted Ai, it can give the traffic optimization suggestion text Pn, that is, the traffic strategy;

[0051] Assume that the neural network outputs y n -1 as: 0.8 (overall index), 0.2 (regional index 1), 0.3 (regional index 2), 0.6 (regional index 3), 0.7 (regional index 4);

[0052] Converted to Ai: The overall area is relatively congested (0.8), the northeast is not congested (0.2), the northwest is not congested (0.3), the southeast is congested (0.6), the southwest is congested (0.7);

[0053] The traffic optimization suggestion text Pn (Ao) output by the large model can be: The current overall area is congested. It is recommended that if there is no urgent matter, try to avoid traveling. If you must travel, you can choose to drive in the northeast or northwest directions to avoid the congested areas in the southeast and southwest;

[0054] 5) Set up a warning index range, determine the warning index according to the output result, send the warning index to the corresponding department, and make a comprehensive, overall and sustainable optimal traffic management decision based on the warning index. This index is the traffic index for the global area or the traffic index for each regional block;

[0055] 6) Combine vehicle networking technology to push the optimal traffic management decision to vehicles in real time. Vehicles perform adaptive cruise or route adjustment based on the decision information (such as recommended vehicle speed, estimated travel time, etc.).

[0056] Preferably, the above model needs to be trained before working. After training is completed, data is input into the model for prediction calculation. Its training steps include:

[0057] S1: Give the network structure

[0058] S2: Forward calculation

[0059] S3: Give hyperparameters:

[0060] batch_size = 64;

[0061] epochs = 6000;

[0062] learning_rate = 0.0005;

[0063] EarlyStopper sets patience = 500, delta = 0.0001 S4;

[0064] The above batch_size refers to the number of samples used to calculate the loss and update the model parameters in each training iteration; epochs refers to the number of times the entire training dataset is completely traversed by the model; learning_rate refers to the amplitude of model parameter update in each training iteration; EarlyStopper is a tool used for early stopping during model training; where patience is the patience value, indicating the number of rounds that training will continue when the loss on the validation set no longer improves; delta is the minimum threshold for loss improvement;

[0065] S4: Start training

[0066] S5: Obtain the training result.

[0067] Preferably, in step S4:

[0068] Real-time monitor the vehicle queuing status on the road sections between intersections, combine microscopic traffic flow models (such as car-following models, cellular automata models) to predict the congestion propagation direction and trend, and take early diversion measures, such as setting up temporary traffic signs, guiding vehicle diversion, etc.;

[0069] According to the real-time congestion level and vehicle queuing situation, use dynamic programming or distributed collaborative optimization algorithms to flexibly adjust the traffic light duration ratio at intersections during peak hours. Considering the differences in traffic demands in different directions (such as main roads and secondary roads, in and out of the city directions), set differentiated green light schemes to enhance the traffic capacity at intersections;

[0070] Use deep learning algorithms combined with real-time and historical travel data to optimize the vehicle driving route prediction model, and introduce probability prediction to accurately judge whether a vehicle passes through a congested section. Enrich the congestion information pushed to vehicles (including estimated travel time, congestion reasons, real-time traffic condition videos or pictures) and the details of the planned route, and optimize the route in real time according to the driver's feedback to improve the practicality and reliability of information services.

[0071] Preferably, in step S5:

[0072] Establish a comprehensive evaluation index system covering multiple dimensions such as traffic flow, congestion level, traffic safety, public satisfaction, and economic cost to comprehensively measure the implementation effect of traffic management strategies;

[0073] Regularly use statistical analysis and simulation technologies to quantitatively evaluate the effect of traffic management decisions. According to the evaluation results, adjust and optimize the strategies and algorithm parameters in a timely manner to form a closed-loop feedback mechanism to ensure that the management strategies are continuously adapted to the actual traffic conditions;

[0074] Flexibly shorten the review cycle according to the dynamic congestion, establish a rapid feedback channel. Once it is found that the congestion situation is misjudged or there are new changes, timely verify and correct the congestion evaluation data and analysis conclusions to ensure the accuracy and timeliness of the decision-making basis, provide a solid support for traffic management decisions, and promote the healthy development of the urban traffic system.

[0075] The present invention also discloses an intelligent city and intelligent transportation collaborative management system based on artificial intelligence, including:

[0076] A data collection unit for collecting multi-source data, comprehensively collecting road conditions and traffic flow information, and also covering road construction, weather impacts, and special event data;

[0077] A historical database for storing historical data;

[0078] A big data processing unit for processing the collected multi-source data and the historical data in the historical database and generating prediction results;

[0079] A traffic light control module for adjusting the traffic light duration ratio according to the prediction results processed by the big data processing unit;

[0080] An information display module for presenting the processed information to the staff;

[0081] An early warning information sending module, configured to send the processed congestion information to vehicles that have not entered the congested section in the form of an early warning;

[0082] A communication unit, configured for communication connection between the data acquisition unit and the big data processing unit.

[0083] Beneficial effects

[0084] The present invention provides a method and system for collaborative management of a smart city and intelligent transportation based on artificial intelligence.

[0085] Compared with the prior art, the following beneficial effects are achieved:

[0086] 1. The method and system for collaborative management of a smart city and intelligent transportation based on artificial intelligence collect rich and accurate traffic data through multiple channels, analyze them using advanced machine learning and deep learning algorithms, introduce an improved wCGwNN model algorithm in all-weather congested section management, and combine a multi-agent reinforcement learning framework. It can fully explore the spatio-temporal features and complex pattern information in traffic data, enable traffic signal agents to cooperate with each other, and obtain the optimal traffic management decision considering multiple factors. This not only improves the intelligent level of traffic management but also can more accurately respond to traffic conditions at different times, in different regions, and with different congestion degrees, effectively improving road traffic efficiency, reducing vehicle waiting time and exhaust emissions, realizing the optimal allocation of traffic resources, and promoting the development of the urban traffic system towards high efficiency and sustainability.

[0087] 2. The method and system for collaborative management of a smart city and intelligent transportation based on artificial intelligence, on the one hand, adopt diverse and targeted management measures for different types of congested sections (all-weather congested sections and peak-hour congested sections), such as increasing the monitoring of the vehicle queue length between intersections and accurately adjusting traffic light times during peak hours, which can prevent and relieve congestion in advance, reduce the risk of congestion propagation, and improve the intersection traffic capacity; on the other hand, by improving the vehicle driving route prediction model and enriching the information service content, it provides accurate congested section information and reasonable route planning for drivers, and optimizes in real time according to driver feedback, improving the driver's travel experience and decision-making accuracy, reducing time waste and anxiety caused by congestion, enhancing the public's satisfaction with traffic management, and promoting the harmonious development of urban traffic. Description of the drawings

[0088] Figure 1 It is a schematic flowchart of the method of the present invention;

[0089] Figure 2 It is a schematic connection diagram of the system of the present invention;

[0090] Figure 3This is a schematic diagram of the wCGwNN model structure of the present invention. DETAILED DESCRIPTION

[0091] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0092] See for example Figures 1-3 , the present invention provides the following two technical solutions:

[0093] The first implementation method: a method and system for collaborative management of smart cities and intelligent transportation based on artificial intelligence, comprising the following steps:

[0094] Step S1: Data collection and integration

[0095] Multi-channel data collection: Road conditions are collected from software such as Baidu Maps and AutoNavi Navigation. Data is also obtained from the official traffic flow monitoring systems of traffic management departments and intelligent traffic sensing equipment (such as geomagnetic sensors and electronic tag readers on the road). This includes collecting road construction information (including the start and end locations of construction sections, estimated construction duration, lane occupation by construction, etc.), data related to the impact of weather on roads (such as changes in the traffic capacity of various sections under rain, snow, and fog), and special event information (such as the location, severity, and estimated processing time of traffic accidents). Multi-source data is then integrated to obtain comprehensive traffic-related information.

[0096] Data preprocessing: Deeply clean and evaluate the quality of all types of collected data, use data interpolation to fill missing values, and adopt outlier detection algorithms (such as methods based on statistical distribution, clustering algorithms, etc.) to identify and correct abnormal data to ensure the integrity and reliability of the data. Through spatiotemporal correlation analysis, perform feature engineering on the data to extract periodic features such as weekly, monthly, and seasonal features, as well as traffic correlation features upstream and downstream of road sections, in preparation for subsequent analysis.

[0097] Step S2: Congestion analysis and road section classification

[0098] Multi-dimensional congestion assessment: Use indicators from multiple dimensions (such as vehicle speed, vehicle density, queue length, travel time index, etc.) to comprehensively assess road congestion, rather than relying solely on simple road condition signs (such as congestion, slow traffic, etc. prompts on map software).

[0099] Real-time dynamic analysis: Establish a real-time data analysis model, and use big data analysis and machine learning algorithms (such as online support vector machines, real-time decision trees, etc.) to dynamically analyze the massive real-time data collected. The model automatically adjusts the thresholds and weights for congestion judgment according to traffic patterns in different time periods (morning and evening rush hours on weekdays, holidays, etc.) and different regions (commercial areas, residential areas, industrial areas, etc.), more realistically identifying the changing trends of road congestion and promptly capturing sudden congestion situations.

[0100] Refined classification and hierarchical comparison of road segments: Classify and layer roads according to different attributes. First, classify them according to road grades (expressways, urban expressways, main roads, secondary roads, etc.), and then further subdivide them within the same grade according to regions (different urban areas, different business districts, etc.). For all-day congested road segments, further classify them into different congestion levels such as mild congestion (vehicle speed slightly lower than the normal level within a certain range), moderate congestion (vehicle speed significantly reduced, with short pauses during vehicle driving), and severe congestion (extremely low vehicle speed, with vehicles queuing for a long time), facilitating subsequent targeted management. Through hierarchical classification and comparison, clearly see the differences in congestion situations of different types and regions of roads, providing a basis for formulating traffic guidance strategies.

[0101] Step S3, Formulation and implementation of traffic management strategies

[0102] Management of all-day congested road segments: Introduce a multi-agent reinforcement learning framework and the wCGwNN model. Regard the traffic lights at each intersection as an agent. After structuring and organizing the processed traffic data, input it into the wCGwNN model, and successively pass through the convolutional layer to extract local features, the pooling layer to reduce the data dimension, the GRUw module to capture time series features using wavelet functions, and the fully connected layer to integrate features. Combine the reward function that includes factors such as vehicle waiting time, passing speed, exhaust emissions, energy consumption, and bus priority passage to drive the agents to collaboratively optimize the regional traffic flow and obtain the optimal decision.

[0103] Collaborative application: Combine vehicle networking technology to push the optimal traffic management decision to vehicles in real time. Vehicles perform adaptive cruise or route adjustment based on the decision information (such as recommended vehicle speed, estimated passing time, etc.). At the same time, set up an intelligent guidance system at intersections to provide visual driving guidance for drivers, reducing decision-making errors and unnecessary parking waits.

[0104] Step S4, Peak period management and information optimization

[0105] Peak-hour congestion section management: Increase the monitoring of the vehicle queue lengths between intersections. By means of video image processing technology or geomagnetic sensor data, accurately measure the actual lengths and distributions of vehicle queues. Combine with microscopic traffic flow models (such as car-following models, cellular automata models, etc.) to more accurately predict the congestion propagation trends of sections and take diversion measures in advance.

[0106] Apply dynamic programming algorithms or distributed collaborative optimization algorithms. According to the congestion levels and vehicle queue situations monitored in real time, dynamically adjust the proportion of the green light time lengths at each intersection during peak hours in real time. Consider the traffic demand differences in different directions of intersections (such as main roads and secondary roads, inbound and outbound directions, etc.) and set differential green light time allocation schemes to ensure the balanced distribution of traffic flow at intersections and improve the traffic capacity of intersections.

[0107] Congestion section prediction and information service: Apply deep learning algorithms (such as long short-term memory network LSTM, convolutional neural network CNN, etc.) to improve the vehicle driving route prediction model by combining real-time traffic data and users' historical travel data. Introduce probability prediction methods to assign probability values to each possible driving route and accurately judge whether a vehicle will pass through a congestion section.

[0108] Enrich the congestion information and planned route content sent to vehicles. In addition to informing the locations of congestion sections and new planned routes, also provide detailed information such as the estimated travel time of congestion sections, reasons for congestion (such as accidents, construction, traffic control, etc.), real-time traffic condition videos or pictures, etc. to help drivers better understand the traffic conditions and make decisions. At the same time, based on drivers' feedback and actual driving trajectories, optimize and adjust the planned routes in real time to improve the practicality and reliability.

[0109] Step S5, Comprehensive evaluation and optimization

[0110] Establish an evaluation index system: Construct a comprehensive evaluation index system for traffic management measures, covering multiple factors such as traffic flow, congestion level, traffic safety, public satisfaction, economic cost, etc.

[0111] Effect evaluation and feedback: Regularly use statistical analysis methods (such as comparative analysis, trend analysis, etc.) and simulation technologies (such as microscopic traffic simulation software) to evaluate the effects of implemented traffic management decisions and quantify the implementation effects of various measures.

[0112] Dynamic adjustment and optimization: According to the evaluation results, timely adjust and optimize traffic management strategies and algorithm parameters to form a closed-loop feedback adjustment mechanism. At the same time, publicly disclose the evaluation results to the public, accept public supervision and suggestions, ensure the scientificity and transparency of traffic management, and promote the sustainable development of urban traffic.

[0113] Recheck mechanism enhancement: Shorten the recheck cycle, dynamically adjust the recheck frequency according to the congestion situation, establish a fast feedback channel for recheck results. Once it is found during recheck that the previous congestion situation judgment is incorrect or there are new changes, immediately feedback the information to the analysis system. The system automatically corrects the previous congestion assessment data and analysis conclusions, and conducts a new comparison analysis to ensure the accuracy and timeliness of the entire process, and better serve traffic management decisions.

[0114] In the above embodiments, the calculation steps of the wCGwNN model include:

[0115] 【1】 Prediction from pictures to metrics

[0116] 【1.1】 Calculation formula from pictures to metrics

[0117] 1) First, convert the processed traffic-related data into an image three-dimensional pixel structure and input it into the input layer of wCGwNN. The input data includes N pictures with different information distributions in the same area. Each picture corresponds to 3 color channels, for a total of N*3 channels: The N pictures can be: real-time traffic flow distribution maps in the same area, real-time population distribution maps in the same area, real-time distribution maps of other information amounts in the same area, etc.

[0118] It can also be map information, population distribution information, and data mapping information (such as the impact of important changes or major fluctuations on traffic flow). These pictures include traffic flow maps, population distribution maps, or flow change maps caused by special events, etc. Each picture corresponds to m labels, including regions, key focus points, etc., which are used to guide the training and output of the neural network.

[0119] 2) Perform forward propagation calculations for the convolutional layer and pooling layer in the CPNN (convolutional neural network part) of the wCGwNN model. The calculation process is shown in the following formula:

[0120]

[0121] In the above formula, is the value at the position index (m, n) in the feature map output by the neurons in the l-th layer, where m and n are the position indices before pooling; O l-1 is the output of the (l - 1)-th layer; core l is a two-dimensional matrix, serving as the convolutional kernel of the l-th layer; Conv(O l-1 , core l , m, n) is the convolution operation performed by the convolutional layer on O l-1 using the convolutional kernel core l ; c is the number of channels in the photo; size l is the size of the convolutional kernel; It is the value at position (m + i, n + j) and channel number c in the feature map output by the (l - 1)-th layer; It is the two-dimensional kernel part corresponding to the input channel c in the convolutional kernel of the l-th layer; b l , ε l It is the bias value of the l-th layer, used to adjust the activation level of the output feature map;

[0122] The output of the convolutional layer Needs to be calculated through the activation function ReLU, and the calculation process is as follows:

[0123]

[0124] There are two convolutional processes in the above formula, that is, each process contains two parts: convolution and activation. Func(Z) represents the activation function, and ReLU(Z) is defined as ReLU(z) = max(0, z). Let it be the initial input, It is the result after the first convolutional process, It is the pooling process after the first convolutional process, It is the result after the second convolutional process, It is the pooling process after the second convolution, It is to output the data converted into 1D data;

[0125] Extract local features in traffic data through convolutional operations, use pooling operations to reduce the data dimension, reduce the amount of calculation while retaining key features. The pooling formula is as follows:

[0126]

[0127] In the above formula, It is the value at position index (m′, n′) in the feature map output by the neurons of the l-th layer, It is the value at position index (m, n) in the feature map output by the neurons of the (l - 1)-th layer, Pooling 2×2 (Z) represents the pooling calculation; m′, n′ are the position indices of the feature map after pooling;

[0128] 3) Introduce a gated recurrent unit based on wavelet transform in the GRUw module for forward propagation calculation. The calculation process is as follows:

[0129]

[0130]

[0131] The above formula includes two GRU layers based on wavelet transform. Among them, x i Is the input of the GRUw module, is the output of the last layer of the CPNN part, size 5 is the size of the feature map output by the 5th layer; are the linear intermediate variables of the reset gate, update gate, and candidate state respectively; W x，r 、W x，z 、W x，h are the weight matrices of the reset gate, update gate, and candidate state respectively; ε r ,ε z , are the bias terms of the reset gate, update gate, and candidate state respectively; z i is the output of the update gate, controlling the fusion ratio of the current state and the previous state; W h，h is the historical information (r i ·h i-1 ) after being adjusted by the reset gate to the weight matrix of the candidate state. r i is the output of the reset gate, controlling the degree of forgetting of the previous hidden state; is the candidate hidden state; h i is the current hidden state, h i-1 is the hidden state at the previous time step; σ is the Sigmoid activation function; tanh is the hyperbolic tangent function, and i represents the index of the time step;

[0132] 4) Perform forward propagation calculation of the fully connected layer in the FCNN. There are four layers in the FCNN part. The last layer of the FCNN is the output layer of the CNN model, which integrates the features extracted by the previous layers and maps them to the dimension finally used for decision-making. The total number of neurons in the first layer of the FCNN size- 4 ×1 is equal to the number of neurons in the last layer of the CPNN part size-5×size-5. The output of the first layer of the FCNN and the output of the last layer of the CPNN are converted. The conversion relationship between and is shown in the following formula:

[0133]

[0134] The calculation process of the final output y n of the forward calculation is as follows:

[0135]

[0136] In the above formula, O n is the output of the first fully connected layer, where n represents n neurons in this layer; O k is the output of the second fully connected layer, where k represents k neurons in this layer; w i,n ,w n，k ,wn，k is the fully connected weight matrix, where i and n represent the weight from the i-th neuron to the n-th neuron, where n and k represent the weight from the n-th neuron to the k-th neuron, and where k and j represent the weight from the k-th neuron to the j-th neuron; ε n , ε k , ε j are the biases of the first, second, and third fully connected layers respectively; is the output of the first layer of the FCNN; is the output of the first wavelet transform, -4 indicating that this value is calculated at the fourth layer from the bottom, i.e., the output of the fourth layer from the bottom of the neural network; is the output of the second wavelet transform, -3 indicating that this value is calculated at the third layer from the bottom, i.e., the output of the third layer from the bottom of the neural network; is the output of the second last output layer; is the output of the last layer of the CPNN; is the output activated by the wavelet transform; b n , b k correspond to the dynamic translation parameters of the first and second wavelet transforms; a n , a k correspond to the dynamic scale parameters of the first and second wavelet transforms; σ is the Sigmoid activation function; size l is the number of neurons in the l-th layer of the neural network; size -1 , size -2 , size -3 are the number of neurons in the last, second last, and third last layers of the neural network respectively; softmax is the normalized exponential function; is the output of the fully connected layer of the exponential transformation; y n -1 represents the final output of the neural network, -1 indicating the last layer of the neural network, i.e., the last layer of the neural network, and n representing the n-th label; the meaning of the n-th label of the output can be: the overall traffic index, the traffic index of a certain block;

[0137] 【1.2】Training Samples and Training Process

[0138] Each training sample is There are a total of M training samples, which form a training sample set, and the network is trained using the M training samples;

[0139] 【2】Neural Network Prediction

[0140] After training the network, for the real-time at future time t is input into the network as a prediction sample, and the network calculates the corresponding y n-1 , y n -1 is the data provided for decision-making. A total of S traffic control strategies Ps = f(y n -1 ), s = 1, 2,..., S;

[0141] 【3】Data Processing and Large Model Training

[0142] Data textification: Each group of n y n -1 is converted into the corresponding text description Ai. For example, a series of indicators output by the neural network (such as 0.9 (overall index), 0.1 (regional index 1), etc.) are converted into text descriptions like "The overall area is relatively congested (0.9), the southwest is not congested (0.1)..."

[0143] Construction of large model training samples: Ai and the corresponding output Ao of the large language model (traffic optimization suggestions) form a pair of training samples S for the large language model. There are a total of S m samples;

[0144] Fine-tuning of the large model: These samples are used as the fine-tuning training samples of the large language model (DeepSeek) to fine-tune the large model so that it can master the ability to obtain Ao based on Ai;

[0145] 【4】Strategies Obtained by Large Model Prediction

[0146] When the large model encounters the newly input y n -1 and the converted Ai, it can give the traffic optimization suggestion text Pn, that is, the traffic strategy.

[0147] Suppose the neural network outputs y n -1 as: 0.8 (overall index), 0.2 (regional index 1), 0.3 (regional index 2), 0.6 (regional index 3), 0.7 (regional index 4).

[0148] Converted to Ai: The overall area is relatively congested (0.8), the northeast is not congested (0.2), the northwest is not congested (0.3), the southeast is congested (0.6), and the southwest is congested (0.7).

[0149] The traffic optimization suggestion text Pn (Ao) output by the large model can be: The current overall area is congested. It is recommended that if there is no urgent matter, try to avoid traveling. If you must travel, you can choose to drive in the northeast or northwest directions to avoid the congested areas in the southeast and southwest.

[0150] Based on the traffic data and relevant strategies processed by the large model through neural networks, decision-making suggestions can be provided in aspects such as travel planning, traffic control, and public transportation optimization:

[0151] Travel planning suggestions

[0152] Personal travel: When the overall traffic index is high, indicating congestion in the overall area, it is recommended that individuals try to travel during off-peak hours. If travel must be carried out during the current period, according to the index of each area, give priority to choosing routes in areas with smooth traffic. For example, if the neural network output shows that the indexes of the northeast and northwest regions are low, that is, the traffic is not congested, the large model can suggest traveling in these directions and avoiding congested areas such as the southeast and southwest. For short-distance travel, cycling or walking can also be recommended, which is not only environmentally friendly but also avoids traffic congestion.

[0153] Goods transportation: For logistics and transportation enterprises, when traffic is congested in certain areas, it is recommended to adjust the transportation plan, such as changing the delivery route, avoiding congested areas, and choosing relatively smooth roads around to ensure the timely delivery of goods. Or adjust the transportation time, choose periods with less traffic flow such as at night for transportation, reducing transportation costs and time costs.

[0154] Traffic control suggestions

[0155] Traffic police deployment: According to the traffic index of each area, in areas with severe congestion (such as the southeast and southwest regions with higher area indexes), it is recommended to increase the traffic police presence, strengthen traffic command and guidance, maintain traffic order, promptly handle traffic accidents and emergencies, and relieve traffic congestion.

[0156] Traffic signal optimization: For intersections with large traffic flow and congestion, it is recommended to adjust the timing plan of traffic signals. For example, extend the green light time, increase the vehicle throughput, and reduce the vehicle waiting time; or adopt an intelligent traffic signal control system to dynamically adjust the signal timing according to the real-time traffic flow and improve the road traffic efficiency.

[0157] Temporary traffic control: In the case of particularly severe traffic congestion, it can be recommended to implement temporary traffic control measures, such as restricting the passage of certain vehicles (such as restricting large trucks) and one-way traffic, etc., to reduce the number of vehicles on the road and relieve traffic pressure.

[0158] Public transportation optimization suggestions

[0159] Bus route adjustment: If traffic congestion in certain areas affects the normal operation of bus routes for a long time, the large model can recommend that the bus company adjust the bus routes, avoid congested sections, and choose other relatively smooth routes to ensure the on-time rate and operation efficiency of buses and facilitate passengers' travel.

[0160] Increase transportation capacity: For areas and time periods with large traffic flow and strong passenger demand, it is recommended to increase the transportation capacity of public transportation, such as increasing the number of bus trips and putting more subway trains into operation, to meet the travel needs of passengers, reduce the use of private vehicles, and thus alleviate traffic congestion.

[0161] Optimize station settings: Based on traffic data and passenger travel needs, evaluate whether the existing bus station settings are reasonable. It is recommended to adjust stations that are too dense or unreasonable, reduce vehicle stopping time, and improve the operating speed of buses.

[0162] Urban planning suggestions

[0163] Road construction and widening: Through the analysis of long-term traffic data, for sections with continuously large traffic flow and severe congestion, it is recommended that urban planning departments consider carrying out road construction or widening projects to increase road capacity to meet the needs of future traffic development.

[0164] Parking lot planning: In densely populated areas with frequent commercial activities, based on traffic flow and parking demand, it is recommended to reasonably plan the construction location and scale of parking lots, increase the supply of parking facilities, and reduce traffic congestion caused by vehicles searching for parking spaces;

[0165] Optimize the layout of functional areas: Considering from a long-term perspective, it is recommended to optimize the layout of urban functional areas, avoid excessive separation of work and residence, reduce the long-distance commuting needs of residents, thereby reducing traffic flow and alleviating traffic pressure.

[0166] 5) Set the warning index range, determine the warning index according to the output result, send the warning index to the corresponding department, and make comprehensive, comprehensive and sustainable optimal traffic management decisions based on the warning index. This index is the traffic index of the global area or the traffic index of each regional block;

[0167] 6) Combine vehicle networking technology to push the optimal traffic management decision to vehicles in real time. Vehicles perform adaptive cruise or route adjustment based on decision-making information (such as recommended vehicle speed, estimated travel time, etc.). At the same time, set up an intelligent guidance system at intersections to provide visual driving guidance for drivers, reducing decision-making errors and unnecessary parking waits.

[0168] The above model needs to be trained before working. After training is completed, data is input into the model for prediction calculation. Its training steps include:

[0169] S1: Give the network structure

[0170] S2: Forward calculation

[0171] S3: Give hyperparameters:

[0172] batch_size = 64;

[0173] epochs = 6000;

[0174] learning_rate = 0.0005;

[0175] The EarlyStopper sets patience = 500, delta = 0.0001S4;

[0176] The above batch_size refers to the number of samples used to calculate the loss and update the model parameters in each training iteration; epochs refers to the number of times the entire training dataset is completely traversed by the model; learning_rate refers to the magnitude of the model parameter update in each training iteration; EarlyStopper is a tool for early stopping during model training; where patience is the patience value, indicating the number of rounds that training will continue when the loss on the validation set no longer improves; delta is the minimum threshold for loss improvement.

[0177] S4: Start training

[0178] S5: Obtain the training results.

[0179] The second implementation manner: The present invention also discloses an intelligent city and intelligent transportation collaborative management system based on artificial intelligence, including:

[0180] A data acquisition unit, used to collect multi-source data, comprehensively collect road conditions and traffic flow information, and at the same time cover road construction, weather impacts and special event data;

[0181] A historical database, used to store historical data;

[0182] A big data processing unit, used to process the collected multi-source data and historical data in the historical database and generate prediction results;

[0183] A traffic light control module, used to adjust the traffic light duration ratio according to the prediction results processed by the big data processing unit;

[0184] An information display module, used to display the processed information to the staff;

[0185] A warning information sending module, used to send the processed congestion information to vehicles that have not entered the congested section in the form of a warning;

[0186] A communication unit, used for the communication connection between the data acquisition unit and the big data processing unit.

[0187] In the above embodiments, the data acquisition unit includes an intersection information acquisition module that uses intersection monitoring probes to collect information, a data acquisition module that uses third-party data sources to collect traffic data, and a historical data acquisition module that uses a historical database to collect historical data.

[0188] In the above embodiments, the big data processing unit includes a data collection module, a data analysis module, a data storage module, a model construction module, and a road planning module. The big data processing unit is electrically connected to the traffic light control module and the information display module, and sends information to vehicles that have not entered the congested section through the communication unit and the warning information sending module.

[0189] The present invention adopts diversified and targeted management measures for different types of congested sections (all-weather congested sections and peak-hour congested sections). For example, during peak hours, it increases the monitoring of the vehicle queue length between intersections and makes precise traffic light time adjustments, which can prevent congestion in advance and relieve congestion in a timely manner, reduce the risk of congestion spread, and improve the traffic capacity of intersections.

[0190] At the same time, the content not described in detail in this specification belongs to the prior art well-known to those skilled in the art, and the model parameters of each electrical appliance are not specifically limited, and conventional equipment can be used.

[0191] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device.

[0192] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An artificial intelligence-based collaborative management method for smart cities and intelligent transportation, characterized in that, The following steps are involved: Step S1, data processing: Integrate multi-source traffic data, comprehensively collect road conditions and traffic flow information, and also include road construction, weather impacts, and special event data to construct a complete traffic data set, clean and fill in the corrected data, and extract spatiotemporal features for future use; Step S2, congestion assessment and road segmentation: Use multiple indicators to assess congestion, build a machine learning model to optimize parameters based on time period and regional patterns, and classify congested sections throughout the day into light, moderate, and heavy levels; Step S3: All-weather congestion management: S3.

1. Intelligent Decision-Making: Introducing a multi-agent reinforcement learning framework and the wCGwNN model, treating traffic lights as agents and combining them with reward functions to achieve optimal decisions. S3.

2. Collaborative Application: Optimal decision-making information is pushed to vehicles in real time through the Internet of Vehicles, enabling vehicles to adaptively adjust their driving status. Step S4: Peak period management and information optimization S4.

1. Congestion prediction and diversion: Monitor queues at intersections, use micro-models to predict congestion, and set up traffic signs in advance; S4.2 Traffic Light Timing Adjustment: Based on congestion and queuing conditions, an algorithm is used to adjust the ratio of traffic light durations during peak hours and set differentiated green light plans; S4.

3. Route and Information Upgrade: Utilize deep learning to optimize route prediction models, introduce probabilistic prediction to accurately determine whether a vehicle will pass through congested sections, enrich push information, and optimize based on feedback; Step S5, comprehensive evaluation and optimization: Build a multi-dimensional indicator system to measure the effectiveness of the strategy, conduct regular quantitative evaluations, adjust strategy parameters based on the results, form a closed-loop feedback mechanism, shorten the review cycle based on congestion, and promptly verify and correct data conclusions.

2. The collaborative management method of smart city and intelligent transportation based on artificial intelligence according to claim 1, characterized in that: In step S1, the collected data is cleaned, missing values are filled using interpolation, and abnormal data are corrected using an outlier detection algorithm to ensure the accuracy and completeness of the data. The periodic characteristics of the data and the road section traffic correlation characteristics are extracted through spatiotemporal correlation analysis to provide strong support for subsequent analysis.

3. The collaborative management method of smart city and intelligent transportation based on artificial intelligence according to claim 1, characterized in that: In step S2, a real-time model is constructed using a machine learning algorithm. Based on traffic patterns in different time periods and regions, data is dynamically analyzed, congestion judgment parameters are automatically optimized, and congestion trends are captured in real time. Roads are further subdivided according to road grades and regions, and all-weather congested sections are refined into light, medium, and heavy congestion levels, providing a basis for formulating targeted traffic management strategies.

4. The collaborative management method of smart city and intelligent transportation based on artificial intelligence according to claim 1, characterized in that: After the processed traffic data is structured and organized in step S3, it is input into the wCGwNN model, which sequentially passes through the convolution layer to extract local features, the pooling layer to reduce the data dimension, the GRUw module to capture time series features using wavelet functions, and the fully connected layer to integrate features. Combined with the reward function including vehicle waiting time, travel speed, exhaust emissions, energy consumption and bus priority factors, it drives the intelligent agent to collaboratively optimize regional traffic flow and reach the optimal decision.

5. The collaborative management method of smart city and intelligent transportation based on artificial intelligence according to claim 4, characterized in that: The calculation steps of the wCGwNN model include: 【1】Prediction from images to indicators 【1.1】Calculation formula from image to indicator 1) First, the processed traffic-related data is converted into a three-dimensional pixel structure of an image and input into the input layer of the wCGwNN; 2) Perform forward propagation calculations for the convolutional layer and pooling layer in the CPNN of the wCGwNN model. The calculation process is shown in the following formula: In the above formula, is the value at the position index (m, n) in the feature map output by the l-th layer neuron, where m and n are the position indices before pooling; O l-1 is the output of the (l - 1)-th layer; core l is a two-dimensional matrix, serving as the convolutional kernel of the l-th layer; Conv(O l-1 , core l , m, n) represents the convolution operation performed by the convolutional layer on O l-1 using the convolutional kernel core l ; c is the number of channels in the photo; size l is the size of the convolutional kernel; is the value at the position (m + i, n + j) and with the number of channels c in the feature map output by the (l - 1)-th layer; is the two-dimensional kernel part corresponding to the input channel c in the l-th layer convolutional kernel; b l , ε l is the bias value of the l-th layer, used to adjust the activation level of the output feature map; Output of the convolutional layer It needs to be calculated through the activation function ReLU, and the calculation process is as follows: There are two convolution processes in the above formula, that is, each process contains two parts: convolution and activation. Func(Z) represents the activation function, and ReLU(Z) is defined as ReLU(z) = max(0, z). is set as the initial input, is the result after the first convolution process, is the pooling process after the first convolution process, is the result after the second convolution process, is the pooling process after the second convolution, is to output the data converted into 1D data; Extract local features from traffic data through convolution operations, and use pooling operations to reduce the data dimension, reduce the computational amount, and retain key features at the same time. The pooling formula is as follows: In the above formula, is the value at the position index (m′, n′) in the feature map output by the neurons in the l-th layer, is the value at the position index (m, n) in the feature map output by the neurons in the (l - 1)-th layer, and Pooling 2×2 (Z) represents the pooling calculation; m′ and n′ are the position indices of the feature map after pooling; 3) Introduce a gated recurrent unit based on wavelet transform in the GRUw module to perform forward propagation calculations. The calculation process is as follows: The above formula includes two GRU layers based on wavelet transform, where x i is the input of the GRUw module, is the output of the last layer of the CPNN part, and size 5 is the feature map size of the output of the 5th layer; are the linear intermediate variables of the reset gate, update gate, and candidate hidden state respectively; W x，r 、W x，z 、W x，h are the weight matrices of the reset gate, update gate, and candidate hidden state respectively; ε r , ε z , are the bias amounts of the reset gate, update gate, and candidate hidden state respectively; r i is the output of the reset gate, controlling the degree of forgetting of the previous hidden state; z i is the output of the update gate, controlling the fusion ratio of the current state and the previous state; W h，h is the weight matrix from the historical information (r i ·h i-1 ) adjusted by the reset gate to the candidate state. is the candidate hidden state; h i is the current hidden state, and h i-1 is the hidden state of the previous time step; σ is the Sigmoid activation function; tanh is the hyperbolic tangent function, and i represents the index of the time step; 4) Perform the forward propagation calculation of the fully connected layer in the FCNN. The FCNN part has a total of four layers. The last layer of the FCNN is the output layer of the CNN model, which integrates the features extracted from the previous layers and maps them to the dimension finally used for decision-making. The total number of neurons in the first layer of the FCNN is size -4 × 1 is equal to the number of neurons in the last layer of the CPNN part, size -5 × size -5 . The output of the first layer of the FCNN and the output of the last layer of the CPNN are converted. The conversion relationship between is shown in the following formula: Forward calculation of the final output y n The calculation process is as follows: In the above formula, O n is the output of the first fully connected layer, where n represents n neurons in this layer; is the output of the first layer of the FCNN; w i，n , w n，k , w n，k are the fully connected weight matrices, where i, n represent the weight from the i-th neuron to the n-th neuron, where n, k represent the weight from the n-th neuron to the k-th neuron, and where k, j represent the weight from the k-th neuron to the j-th neuron; ε n , ε k , ε j are the biases of the first, second, and third fully connected layers respectively; is the output of the first wavelet transform, -4 indicates that this value is calculated at the fourth-to-last layer, i.e., the output of the fourth-to-last layer of the neural network; O k is the output of the second fully connected layer, where k represents k neurons in this layer; is the output of the second wavelet transform, -3 indicates that this value is calculated at the third-to-last layer, i.e., the output of the third-to-last layer of the neural network; an, a k correspond to the dynamic scale parameters of the first and second wavelet transforms; b n , b k correspond to the dynamic translation parameters of the first and second wavelet transforms; is the output of the second-to-last output layer; is the output of the last layer of the CPNN; is the output activated by the wavelet transform; σ is the Sigmoid activation function; size l is the spatial dimension of the feature map output by the l-th layer of the neural network; size -1 , size -2 , size -3 are the numbers of neurons in the last, second-to-last, and third-to-last layers of the neural network respectively; spftmax is the softmax function; is the exponential transformation of the output of the fully connected layer ; y n -1 represents the final output of the neural network, -1 represents the last layer of the neural network, i.e., the last layer of the neural network, and n represents the n-th label; the meaning of the n-th label of the output can be: the overall traffic index, the traffic index of a certain block; 【1.2】Training Samples and Training Process Each training sample is There are a total of M training samples, which form a training sample set, and the network is trained using the M training samples; 【2】Neural Network Prediction After the network is trained, for the real-time situation at future time t is input into the network as a prediction sample, and the network calculates the corresponding y n -1 , y n -1 is the data provided for decision-making. A total of S traffic control strategies Ps = f(y n -1 ), s = 1, 2,..., S; 【3】Data Processing and Large Model Training Data literalization: n y's in each group n -1 Convert to the corresponding literal description Ai. For example, convert a series of metrics output by the neural network into a literal description such as "The overall area is relatively congested, and the southwest is not congested..." Large model training sample construction: Pair the AI with the corresponding output Ao of the large language model to form a training sample S for the large language model. There are a total of S m samples; Fine-tuning of the large model: Use these samples as the fine-tuning training samples of the large language model to fine-tune the large model so that it can master the ability to obtain Ao according to Ai; 【4】The large model predicts and obtains the strategy When the large model encounters the newly input y n -1 When it comes to the transformed Ai, it can give the traffic optimization suggestion text Pn, that is, the traffic strategy; 5) Set the warning index interval, determine the warning index according to the output result, send the warning index to the corresponding department, and make comprehensive, comprehensive and sustainable optimal traffic management decisions according to the warning index. This index is the traffic index of the global area or the traffic index of each regional block; 6) Combine vehicle networking technology to push the optimal traffic management decision to the vehicle in real time, and the vehicle makes adaptive cruise or route adjustment according to the decision information.

6. The collaborative management method of smart city and intelligent transportation based on artificial intelligence according to claim 1, characterized in that: The above-mentioned model needs to be trained before working. After training is completed, data is input into the model for prediction calculations. Its training steps include: S1: Give the network structure S2: Forward calculation S3: Give hyperparameters: batch_size = 64; epochs = 6000; learning_rate = 0.0005; EarlyStopper sets patience = 500, delta = 0.0001S4; The above batch_size refers to the number of samples used to calculate the loss and update the model parameters in each training iteration; epochs refers to the number of times the entire training dataset is completely traversed by the model; learning_rate refers to the amplitude of model parameter update in each training iteration; EarlyStopper is a tool used for early stopping during model training; where patience is the patience value, indicating the number of rounds that training will continue when the loss on the validation set no longer improves; delta is the minimum threshold for loss improvement; S4: Start training S5: Obtain the training result.

7. The collaborative management method of smart city and intelligent transportation based on artificial intelligence according to claim 1, characterized in that: In the said step 4): Real-time monitor the vehicle queuing status between intersections, combine the microscopic traffic flow model to predict the congestion propagation direction and trend, and take dredging measures in advance; According to the real-time congestion degree and vehicle queuing situation, use dynamic programming or distributed cooperative optimization algorithms to flexibly adjust the traffic light duration ratio at intersections during peak hours, consider the traffic demand differences in different directions, set differentiated green light schemes, and enhance the intersection passing capacity; Use deep learning algorithms to optimize the vehicle driving route prediction model combined with real-time and historical travel data, introduce probability prediction to accurately judge whether the vehicle passes through congested sections, enrich the congestion information and planned route details pushed to the vehicle, and optimize the route in real time according to the driver's feedback to improve the practicality and reliability of information services.

8. The collaborative management method of smart city and intelligent transportation based on artificial intelligence according to claim 1, characterized in that: In step 5): Establish a multi-dimensional comprehensive evaluation index system covering traffic flow, congestion level, traffic safety, public satisfaction and economic cost to comprehensively measure the implementation effect of traffic management strategies; Regularly use statistical analysis and simulation technologies to quantitatively evaluate the effect of traffic management decisions, and timely adjust and optimize strategies and algorithm parameters according to the evaluation results to form a closed-loop feedback mechanism to ensure the continuous adaptation of management strategies to the actual traffic conditions; Flexibly shorten the review cycle according to the dynamic congestion, establish a rapid feedback channel, and once it is found that the congestion situation is misjudged or there are new changes, timely verify and correct the congestion evaluation data and analysis conclusions to ensure the accuracy and timeliness of the decision-making basis, provide strong support for traffic management decisions, and promote the healthy development of the urban traffic system.

9. The collaborative management system of smart city and intelligent transportation based on artificial intelligence is characterized in that Including: A data acquisition unit for collecting multi-source data, comprehensively collecting road conditions and traffic flow information, and at the same time covering road construction, weather impacts and special event data; A historical database for storing historical data; A big data processing unit for processing the collected multi-source data and historical data in the historical database and generating prediction results; A traffic light control module for adjusting the traffic light duration ratio according to the prediction results processed by the big data processing unit; An information display module for displaying the processed information to the staff; A warning information sending module for sending the processed congestion information to the vehicles that have not entered the congested section in the form of a warning; A communication unit for the communication connection between the data acquisition unit and the big data processing unit.

Images