Traffic spatio-temporal data calculation optimization method based on double-layer multi-round-trip operation model

Abstract

The invention discloses a traffic spatio-temporal data calculation optimization method based on a double-layer multi-round-trip operation model, and the method comprises the following steps: collecting multi-item multi-source data, and carrying out the unified representation through a multi-modal feature vector, the multi-source data at least comprising vehicle state parameters, user behavior data and environment data; performing feature extraction on the multi-modal feature vector by adopting an ST-GCN model; a personalized recommendation strategy is generated based on deep reinforcement learning and in combination with multiple optimization targets of a vehicle system; a personalized recommendation strategy is optimized by adopting a deep network updating formula; a Pareto optimal solution is used for balancing the recommendation strategy; executing multiple rounds of feedback and adaptive optimization; and outputting the personalized recommendation in a multi-dimensional output form.

Description

Technical field:

[0001] The invention relates to a traffic spatiotemporal data calculation optimization method based on a double-layer multi-round trip operation model. Background technology:

[0002] In modern road transport systems, with the continuous expansion of traffic networks and the rapid growth of data traffic, the system faces many challenges in multi-source data processing, real-time decision-making and personalized recommendations. The input volume of multi-source heterogeneous data (such as vehicle-mounted sensors, traffic cameras, etc.) increases sharply during peak hours, making it difficult for existing computing systems to achieve millisecond-level responses in high-concurrency scenarios, and the overall data processing efficiency is low. This directly restricts the system's real-time detection of abnormal events, dynamic traffic status prediction and path planning accuracy, which in turn affects traffic safety and operational efficiency.

[0003] At the level of data fusion and intelligent decision-making, the system has multi-dimensional defects: the differences in the spatiotemporal structure and the inconsistent formats of data sources make it difficult to coordinate the processing of structured and unstructured data, resulting in insufficient data continuity and correlation, which weakens the practical application value of decision-making results; traditional recommendation algorithms cannot adapt to the complex needs of dynamic traffic scenarios, especially when dealing with emergencies such as severe weather and sudden accidents. Their response speed is significantly delayed, and there is a lack of personalized recommendation strategies for different drivers, vehicles, and road conditions, which affects the flexibility of the traffic management system and user experience; as the amount of data and network scale continue to grow, the existing system also has bottlenecks in resource scheduling flexibility and computing resource utilization, making it difficult to meet the requirements of future intelligent transportation systems for scalability and efficiency. Summary of the invention:

[0004] The embodiment of the present invention provides a traffic spatiotemporal data calculation optimization method based on a double-layer multi-round-trip operation model. The method is reasonably designed and adopts a multi-round iterative calculation mechanism combining stream processing and batch processing to improve the real-time, accuracy and decision-making reliability of data processing. The method adopts ST-GCN and GNN models to achieve unified modeling and deep analysis, fusion and collaborative analysis of multi-source heterogeneous data, and constructs an adaptive optimization recommendation strategy based on DRL to enhance the real-time and adaptability of personalized recommendation strategies, providing users with accurate personalized path planning and safety warnings; at the same time, through containerized deployment and microservice architecture, the system's elastic expansion and dynamic resource scheduling are realized, which improves the system's scalability and overall operating efficiency, and solves the problems existing in the prior art.

[0005] The technical solution adopted by the present invention to solve the above technical problems is:

[0006] A traffic spatiotemporal data calculation optimization method based on a double-layer multi-round trip operation model, the calculation optimization method comprising the following steps:

[0007] S1, Collect multiple multi-source data and represent them uniformly through multi-modal feature vectors. The multi-source data includes at least vehicle state parameters, user behavior data, and environmental data;

[0008] S2, Use the ST-GCN model to extract features from the multi-modal feature vectors;

[0009] S3, Generate personalized recommendation strategies based on deep reinforcement learning and combined with multiple optimization objectives of the vehicle system;

[0010] S4, Optimize the personalized recommendation strategy using the deep network update formula;

[0011] S5, Use the Pareto optimal solution to balance the recommendation strategy;

[0012] S6, Perform multiple rounds of feedback and adaptive optimization;

[0013] S7, Output personalized recommendations in the form of multi-dimensional outputs.

[0014] Collecting multiple multi-source data and representing them uniformly through multi-modal feature vectors includes the following steps:

[0015] S1.1, Construct multi-modal feature vectors. The multi-modal feature vectors are:

[0016]

[0017] where, X t v is the vehicle state feature, including position, speed, and fuel consumption, used to reflect the running state of the vehicle and evaluate the current performance and requirements of the vehicle; X t u is the user feature, including historical routes and driving styles, used to reflect the driving habits and preferences of the user; X t θ is the environmental feature, including road conditions and weather, used to reflect external conditions and adjust the personalized recommendation strategy to adapt to changes in environmental conditions;

[0018] S1.2, Standardize data from different sources and at different scales:

[0019]

[0020] where, X t is the original data point that needs to be standardized, such as the speed parameter of the vehicle, the age parameter of the user, and the temperature parameter of the environment;

[0021] μ is the average value of the original data set;

[0022] ε is the standard deviation of the original data, used to represent the degree of data dispersion.

[0023] The ST-GCN model is as follows:

[0024] Ht = σ(W1·Xt + W2·A·Xt)

[0025] The fused feature representation is:

[0026]

[0027] Among them, W1 and W2 are network weights used to capture the spatio-temporal correlation of data;

[0028] A is the adjacency matrix used to represent the topological relationship between different features;

[0029] H t is the extracted feature, and σ is the activation function used to introduce non-linearity;

[0030] Z t is the fused feature, which combines vehicle features, user features, and environmental features to provide comprehensive input for subsequent decision-making.

[0031] Generating a personalized recommendation strategy based on deep reinforcement learning and combining multiple optimization objectives of the vehicle system includes the following steps:

[0032] S3.1, Define the state space:

[0033] St = f(Zt)

[0034] The state space is used to reflect the comprehensive state of the current system and provide a basis for decision-making;

[0035] S3.2, Define the action space:

[0036] At = (a1, a2,..., an)

[0037] including path planning, driving behavior adjustment, and energy consumption optimization, used to adjust the operation strategy of the vehicle;

[0038] S3.3, Define the reward function:

[0039] R(St, at) = a1·Rtime + a2·Rfuel + a3·Rsafe

[0040] The reward function needs to comprehensively consider three dimensions of time, energy consumption, and safety, and balance the importance of different dimensions by setting different weights a1, a2, and a3 to guide policy optimization.

[0041] The deep network update formula is:

[0042]

[0043] Among them, Q(S t , a t ) is the state-action value function, which is used to represent the expected value of selecting different actions in the state space;

[0044] η is the learning rate, which is used to control the update speed;

[0045] γ is the discount factor, which is used to balance short-term and long-term rewards;

[0046]

[0047] is expressed as the optimal reward estimate value of the future state, which is used to guide the policy to develop in the long-term optimal direction. The Pareto optimal solution is solved by the weighted sum method as:

[0048] F(x) = ∑λifi(x)

[0049] where λ i is the target weight, which is used to adjust the priorities of different targets.

[0050] Performing multiple rounds of feedback and adaptive optimization includes the following steps:

[0051] S6.1, setting the short-term feedback is expressed as:

[0052]

[0053] where θ t represents the model parameters at the current time, which are used to determine the behavior of the recommendation policy and define the mapping relationship from state to action;

[0054] η is the learning rate, which is used to control the step size of model parameter update and determines the amplitude of parameter adjustment in each iteration;

[0055] represents the gradient of the loss function L with respect to the model parameters θ, which reflects the sensitivity of the loss function to each parameter in the current state;

[0056] L(Zt, at, Rt) is the loss function, which is used to measure the performance of the model in the current state and is usually a function of features, actions, and rewards, reflecting the difference between the model prediction value and the true value;

[0057] The short-term feedback can adjust the model parameters through gradient ascent to adapt to the immediate observed data and enhance the adaptability of the policy;

[0058] S6.2, setting the long-term feedback is expressed as:

[0059]

[0060] Among them, L(θ) is the loss function, and N is the number of samples;

[0061] The long-term feedback optimizes the overall performance of the model by minimizing the squared difference between the predicted value and the true value, ensuring the stability and consistency of the strategy.

[0062] The multi-dimensional output is as follows:

[0063] Output = {Route, ETA, Risk Level, Fuel Consumption, Driving Advice};

[0064] It includes route planning, estimated time of arrival, risk level, fuel consumption, and driving advice, providing comprehensive personalized recommendations for users.

[0065] The computing optimization architecture corresponding to the computing optimization method includes a data collection and access layer, a stream processing layer, a batch processing layer, a unified scheduling and resource management layer, and an output and service layer;

[0066] In the data collection and access layer, real-time data is obtained from multi-source devices such as in-vehicle terminal devices, in-vehicle sensors, and monitoring cameras, and duplicate removal, format conversion, data cleaning, and standardization are performed to generate a high-quality input data stream;

[0067] In the stream processing layer, a spatio-temporal convolutional network is used to extract spatio-temporal features, combined with a graph neural network to identify traffic accidents, congestion, and abnormal events, and multi-round calculations and optimizations are performed on real-time data to improve data processing efficiency and decision-making accuracy;

[0068] In the batch processing layer, historical traffic data is analyzed and predicted based on a deep learning model, and a traffic flow prediction and recommendation model is trained to optimize traffic pattern recognition and decision rule generation;

[0069] In the unified scheduling and resource management layer, resource management is carried out to achieve elastic expansion and load balancing of computing resources, and resource allocation is dynamically adjusted according to traffic prediction and load data to ensure the efficient operation of the system;

[0070] In the output and service layer, it is used to provide personalized recommendations, safety warnings, and optimization suggestions to in-vehicle terminals and management platforms, and detailed optimization strategies and implementation suggestions to users.

[0071] With the above structure, the present invention obtains real-time data from multi-source devices such as in-vehicle terminal devices, in-vehicle sensors, and monitoring cameras through the data acquisition and access layer, and performs deduplication, format conversion, data cleaning, and standardization to generate a high-quality input data stream; through the stream processing layer, a spatio-temporal convolutional network is used to extract spatio-temporal features, combined with a graph neural network to identify traffic accidents, congestion, and abnormal events, and multiple rounds of calculations and optimizations are performed on the real-time data to improve data processing efficiency and decision-making accuracy; through the batch processing layer, historical traffic data is analyzed and predicted, and a traffic flow prediction and recommendation model is trained to optimize traffic pattern recognition and decision rule generation; through the unified scheduling and resource management layer, resource management is carried out to achieve elastic expansion and load balancing of computing resources, and according to traffic prediction and load data, resource allocation is dynamically adjusted to ensure the efficient operation of the system; through the output and service layer, personalized recommendations, safety warnings, and optimization suggestions are provided to in-vehicle terminals and management platforms, and detailed optimization strategies and implementation suggestions are provided to users, having the advantages of precision, efficiency, practicality, and reliability. BRIEF DESCRIPTION OF THE DRAWINGS:

[0072] Figure 1 It is a schematic flowchart of the present invention.

[0073] Figure 2 It is a schematic architecture diagram of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS:

[0074] To clearly illustrate the technical features of the present solution, the present invention will be elaborated in detail below through specific embodiments and in conjunction with its accompanying drawings.

[0075] As Figure 1-2 shown in

[0076] S1, collect multiple multi-source data and represent them uniformly through multi-modal feature vectors, where the multi-source data at least includes vehicle state parameters, user behavior data, and environmental data;

[0077] S2, use the ST-GCN model to extract features from the multi-modal feature vectors;

[0078] S3, generate a personalized recommendation strategy based on deep reinforcement learning and in combination with multiple optimization objectives of the vehicle system;

[0079] S4, use the deep network update formula to optimize the personalized recommendation strategy;

[0080] S5, use the Pareto optimal solution to balance the recommendation strategy;

[0081] S6, perform multiple rounds of feedback and adaptive optimization;

[0082] S7 outputs personalized recommendations in a multi-dimensional output form.

[0083] Collecting multiple multi-source data and uniformly representing them through multi-modal feature vectors includes the following steps:

[0084] S1.1, constructing a multi-modal feature vector, the multi-modal feature vector is:

[0085]

[0086] where X t v is the vehicle state feature, including position, speed, and fuel consumption, used to reflect the running state of the vehicle and evaluate the current performance and requirements of the vehicle; X t u is the user feature, including historical paths and driving styles, used to reflect the driving habits and preferences of the user; X t θ is the environmental feature, including road conditions and weather, used to reflect external conditions and adjust the personalized recommendation strategy to adapt to changes in environmental conditions;

[0087] S1.2, performing standardization processing on data from different sources and at different scales:

[0088]

[0089] where X t is the original data point that needs to be standardized, such as the speed parameter of the vehicle, the age parameter of the user, and the temperature parameter of the environment;

[0090] μ is the average value of the original data set;

[0091] ε is the standard deviation of the original data, used to represent the degree of data dispersion.

[0092] The ST-GCN model is:

[0093] Ht = σ(W1·Xt + W2·A·Xt)

[0094] The fused feature representation is:

[0095]

[0096] where W1 and W2 are network weights, used to capture the spatio-temporal correlation of the data;

[0097] A is the adjacency matrix, used to represent the topological relationship between different features;

[0098] H t is the extracted feature, and σ is the activation function, used to introduce non-linearity;

[0099] Z t The fused features are obtained by integrating vehicle features, user features, and environmental features, providing comprehensive inputs for subsequent decision-making.

[0100] Generating a personalized recommendation strategy based on deep reinforcement learning and combined with multiple optimization objectives of the vehicle system includes the following steps:

[0101] S3.1, Define the state space:

[0102] St = f(Zt)

[0103] The state space is used to reflect the comprehensive state of the current system and provide a basis for decision-making;

[0104] S3.2, Define the action space:

[0105] At = (a1, a2,..., an)

[0106] It includes path planning, driving behavior adjustment, and energy consumption optimization, and is used to adjust the operation strategy of the vehicle;

[0107] S3.3, Define the reward function:

[0108] R(St, at) = a1·Rtime + a2·Rfuel + a3·Rsafe

[0109] The reward function should comprehensively consider three dimensions: time, energy consumption, and safety, and balance the importance of different dimensions by setting different weights a1, a2, and a3 to guide policy optimization.

[0110] The update formula for the deep network is:

[0111]

[0112] Among them, Q(S t , a t ) is the state-action value function, which is used to represent the expected value of choosing different actions in the state space;

[0113] η is the learning rate, which is used to control the update speed;

[0114] γ is the discount factor, which is used to balance short-term and long-term rewards;

[0115]

[0116] represents the optimal reward estimate value of the future state and is used to guide the policy to develop in the direction of long-term optimality. Solving the Pareto optimal solution by the weighted sum method is:

[0117] F(x) = ∑λifi(x)

[0118] Among them, λ i is the target weight, which is used to adjust the priorities of different targets.

[0119] Performing multiple rounds of feedback and adaptive optimization includes the following steps:

[0120] S6.1, Set the short-term feedback as:

[0121]

[0122] Among them, θ t represents the model parameters at the current time, which are used to determine the behavior of the recommendation strategy and define the mapping relationship from state to action;

[0123] η is the learning rate, which is used to control the step size of model parameter update and determines the amplitude of parameter adjustment in each iteration;

[0124] represents the gradient of the loss function L with respect to the model parameters θ, which reflects the sensitivity of the loss function to each parameter in the current state;

[0125] L(Zt, at, Rt) is the loss function, which is used to measure the performance of the model in the current state. It is usually a function of features, actions, and rewards, reflecting the difference between the model prediction value and the true value;

[0126] The short-term feedback can adjust the model parameters through gradient ascent to adapt to the immediate observed data and enhance the adaptability of the strategy;

[0127] S6.2, Set the long-term feedback as:

[0128]

[0129] Among them, L(θ) is the loss function and N is the number of samples;

[0130] The long-term feedback optimizes the overall performance of the model by minimizing the squared difference between the predicted value and the true value, ensuring the stability and consistency of the strategy.

[0131] The multi-dimensional output is:

[0132] Output = {Route, ETA, Risk Level, Fuel Consumption, Driving Advice};

[0133] including route planning, estimated time of arrival, risk level, fuel consumption, and driving advice, providing comprehensive personalized recommendations for users.

[0134] The computing optimization architecture corresponding to the computing optimization method includes a data collection and access layer, a stream processing layer, a batch processing layer, a unified scheduling and resource management layer, and an output and service layer;

[0135] In the data collection and access layer, real-time data is obtained from multi-source devices such as in-vehicle terminal devices, in-vehicle sensors, and monitoring cameras, and duplicate removal, format conversion, data cleaning, and standardization are performed to generate a high-quality input data stream;

[0136] In the stream processing layer, a spatio-temporal convolutional network is used to extract spatio-temporal features, combined with a graph neural network to identify traffic accidents, congestion, and abnormal events, and multiple rounds of calculations and optimizations are performed on the real-time data to improve data processing efficiency and decision-making accuracy;

[0137] In the batch processing layer, historical traffic data is analyzed and predicted based on a deep learning model, and a traffic flow prediction and recommendation model is trained to optimize traffic pattern recognition and decision rule generation;

[0138] In the unified scheduling and resource management layer, resource management is performed to achieve elastic expansion and load balancing of computing resources, and resource allocation is dynamically adjusted according to traffic prediction and load data to ensure the efficient operation of the system;

[0139] In the output and service layer, it is used to provide personalized recommendations, safety warnings, and optimization suggestions to in-vehicle terminals and management platforms, and detailed optimization strategies and implementation suggestions to users.

[0140] The working principle of the traffic spatio-temporal data computing optimization method based on the double-layer multi-round operation model in the embodiments of the present invention is as follows: A multi-round iterative computing mechanism combining stream processing and batch processing is adopted to improve the real-time performance, accuracy, and decision-making reliability of data processing. This method uses ST-GCN and GNN models to achieve unified modeling, in-depth analysis, fusion, and collaborative analysis of multi-source heterogeneous data, and constructs an adaptive optimization recommendation strategy based on DRL to strengthen the real-time performance and adaptability of personalized recommendation strategies, providing users with accurate personalized route planning and safety warnings; at the same time, elastic expansion of the system and dynamic resource scheduling are achieved through containerized deployment and microservice architecture, improving the scalability and overall operation efficiency of the system.

[0141] Furthermore, the present application can construct a high-density spatio-temporal data computing and recommendation optimization architecture according to the characteristics of intelligent traffic data to achieve the fusion and collaborative analysis of multi-source heterogeneous data, and strengthen the real-time performance and adaptability of personalized recommendation strategies.

[0142] In the overall solution, the calculation optimization method includes the following steps: collecting multiple multi-source data and uniformly representing them through multi-modal feature vectors, where the multi-source data at least includes vehicle state parameters, user behavior data, and environmental data; using the ST-GCN model to extract features from the multi-modal feature vectors; generating personalized recommendation strategies based on deep reinforcement learning and combining multiple optimization objectives of the vehicle system; using the deep network update formula to update the personalized recommendation strategies; using the Pareto optimal solution to balance the recommendation strategies; performing multiple rounds of feedback and adaptive optimization; and outputting personalized recommendations in the form of multi-dimensional outputs.

[0143] Specifically, collecting multiple multi-source data and uniformly representing them through multi-modal feature vectors includes the following steps:

[0144] Construct a multi-modal feature vector, where the multi-modal feature vector is:

[0145]

[0146] Among them, X t v is the vehicle state feature, including position, speed, and fuel consumption, which is used to reflect the running state of the vehicle, evaluate the current performance and requirements of the vehicle; X t u is the user feature, including historical routes and driving styles, which is used to reflect the driving habits and preferences of the user; X t θ is the environmental feature, including road conditions and weather, which is used to reflect external conditions and adjust the personalized recommendation strategy to adapt to changes in environmental conditions;

[0147] Perform standardization processing on data from different sources and at different scales:

[0148]

[0149] Among them, X t is the original data point that needs to be standardized, such as the speed parameter of the vehicle, the age parameter of the user, and the temperature parameter of the environment;

[0150] μ is the average value of the original data set;

[0151] ε is the standard deviation of the original data, which is used to represent the degree of data dispersion.

[0152] Through the construction of the multi-modal feature vector, the vehicle state feature, user feature, and environmental feature are fused to achieve the unified modeling of multi-source heterogeneous data.

[0153] For the ST-GCN model, it is set as:

[0154] Ht=σ(W1·Xt+W2·A·Xt)

[0155] The fused features are expressed as:

[0156]

[0157] Among them, W1 and W2 are network weights used to capture the spatio-temporal correlation of data;

[0158] A is the adjacency matrix used to represent the topological relationship between different features;

[0159] H t is the extracted feature, and σ is the activation function used to introduce non-linearity;

[0160] Z t is the fused feature, which synthesizes vehicle features, user features, and environmental features to provide comprehensive input for subsequent decision-making.

[0161] Furthermore, generating a personalized recommendation strategy based on deep reinforcement learning and combined with multiple optimization objectives of the vehicle system includes the following steps:

[0162] S3.1, defining the state space:

[0163] St = f(Zt)

[0164] The state space is used to reflect the comprehensive state of the current system and provide a basis for decision-making;

[0165] S3.2, defining the action space:

[0166] At = (a1, a2,..., an)

[0167] including path planning, driving behavior adjustment, and energy consumption optimization, used to adjust the operation strategy of the vehicle;

[0168] S3.3, defining the reward function:

[0169] R(St, at) = a1·Rtime + a2·Rfuel + a3·Rsafe

[0170] The reward function needs to comprehensively consider three dimensions of time, energy consumption, and safety, and balance the importance of different dimensions by setting different weights a1, a2, and a3 to guide policy optimization.

[0171] For the deep network update formula, it is set as:

[0172]

[0173] Among them, Q(S t , a t) is the state - action value function, which is used to represent the expected value of choosing different actions in the state space;

[0174] η is the learning rate, which is used to control the update speed;

[0175] γ is the discount factor, which is used to balance short - term and long - term rewards;

[0176]

[0177] is expressed as the optimal reward estimate value of the future state, which is used to guide the policy towards the long - term optimal direction. The Pareto optimal solution is obtained by the weighted sum method as:

[0178] F(x) = ∑λifi(x)

[0179] Among them, λ i is the target weight, which is used to adjust the priority of different targets.

[0180] For adaptive optimization and multi - round feedback, it specifically includes the following steps:

[0181] S6.1, set the short - term feedback as:

[0182]

[0183] Among them, θ t represents the model parameters at the current time, which are used to determine the behavior of the recommendation policy and define the mapping relationship from state to action;

[0184] η is the learning rate, which is used to control the step size of model parameter update and determines the amplitude of parameter adjustment in each iteration;

[0185] represents the gradient of the loss function L with respect to the model parameters θ, which reflects the sensitivity of the loss function to each parameter in the current state;

[0186] L(Zt,at,Rt) is the loss function, which is used to measure the performance of the model in the current state. It is usually a function of features, actions, and rewards, reflecting the difference between the model prediction value and the true value;

[0187] The short - term feedback can adjust the model parameters through gradient ascent to adapt to the immediate observed data and enhance the adaptability of the policy;

[0188] S6.2, set the long - term feedback as:

[0189]

[0190] Among them, L(θ) is the loss function and N is the number of samples;

[0191] The long-term feedback optimizes the overall performance of the model by minimizing the square difference between the predicted value and the true value, ensuring the stability and consistency of the strategy, and then obtains the multi-dimensional output as follows:

[0192] Output={Route,ETA,Risk Level,Fuel Consumption,Driving Advice};

[0193] Including route planning, estimated arrival time, risk level, fuel consumption and driving advice, providing users with comprehensive personalized recommendations.

[0194] The computing optimization architecture corresponding to the computing optimization method includes data collection and access layer, stream processing layer, batch processing layer, unified scheduling and resource management layer, and output and service layer.

[0195] Specifically, in the data collection and access layer, real-time data is obtained from multiple source devices such as vehicle-mounted terminal devices, vehicle-mounted sensors, and surveillance cameras, and deduplication, format conversion, data cleaning and standardization are performed to generate high-quality input data streams; in the stream processing layer, spatiotemporal convolutional networks are used to extract spatiotemporal features, combined with graph neural networks to identify traffic accidents, congestion and abnormal events, and multiple rounds of calculations and optimizations are performed on real-time data to improve data processing efficiency and decision accuracy; in the batch processing layer, historical traffic data is analyzed and predicted based on deep learning models, and traffic flow prediction and recommendation models are trained to optimize traffic pattern recognition and decision rule generation; in the unified scheduling and resource management layer, resource management is performed to achieve elastic expansion and load balancing of computing resources, and resource allocation is dynamically adjusted according to traffic predictions and load data to ensure efficient operation of the system; in the output and service layer, it is used to provide personalized recommendations, safety warnings and optimization suggestions to vehicle-mounted terminals and management platforms, as well as provide users with detailed optimization strategies and implementation suggestions.

[0196] During daily driving and commuting, drivers will frequently engage in bad driving behaviors such as sudden acceleration and sudden braking. At the same time, when the vehicle's fuel consumption is high and the tires wear quickly, the system can generate personalized driving suggestions based on the driver's driving behavior and real-time vehicle data to help the driver optimize his driving habits and reduce fuel consumption and maintenance costs.

[0197] The architecture system can collect vehicle status characteristics, user characteristics and environmental characteristics from on-board terminals, sensors and other devices; specifically including the current vehicle location, speed, fuel consumption, tire wear; user driving behavior preferences, historical fuel consumption records; current weather, road conditions, and traffic flow.

[0198] Fuse and unify the above features, and then use a spatio-temporal convolutional network to process the feature vectors to capture the spatio-temporal correlation between the vehicle state and driving behavior.

[0199] Furthermore, perform deep reinforcement learning. Based on the extracted features, define the state space, action space, and reward function to generate a personalized recommendation strategy. The state space comprehensively reflects the current state of the system, including vehicle fuel consumption, tire wear, driving behavior (such as the frequency of hard acceleration and hard braking), etc. The action space includes driving behavior adjustments (such as reducing hard acceleration and hard braking), energy consumption optimization (such as maintaining a constant speed), and maintenance suggestions (such as regularly checking the tires). For the reward function, comprehensively consider the three objectives of fuel consumption, tire wear, and driving safety.

[0200] Update and optimize the recommendation strategy through the deep network update formula. Assume that in the current state, the best driving behavior predicted by the system is "reduce hard acceleration and hard braking", but it is actually observed that this behavior does not significantly reduce fuel consumption. Adjust the model parameters through short-term feedback to enhance the adaptability of the strategy.

[0201] Furthermore, solve the Pareto optimal solution through the weighted sum method, comprehensively considering the three objectives of fuel consumption, tire wear, and driving safety, and perform multi-round feedback optimization according to the actual driving behavior of the driver. When it is detected that the driver still has hard acceleration behavior, the system immediately reminds him to maintain a constant speed.

[0202] In the feedback optimization process, a large amount of historical data can be accumulated to continuously optimize the model parameters and improve the accuracy and robustness of the recommendation strategy.

[0203] The final output of the personalized recommendation results includes: The finally generated recommendation strategy is: driving behavior adjustment (it is recommended to reduce hard acceleration and hard braking and maintain a constant speed); fuel consumption optimization; tire maintenance suggestions (regularly check the tire pressure to avoid excessive wear); safety improvement (reduce the frequency of hard braking and reduce the accident risk).

[0204] It should be specifically noted that this application can effectively analyze the driving behavior of the driver and the real-time vehicle data, and provide accurate and personalized driving suggestions for users. This method not only helps drivers optimize their driving habits, reduce fuel consumption and maintenance costs, but also improves driving safety, providing strong support for the realization of intelligent transportation systems.

[0205] In summary, the traffic spatio-temporal data calculation optimization method based on the double-layer multi-round operation model in the embodiments of the present invention adopts a multi-round iterative calculation mechanism that combines stream processing and batch processing, improving the real-time performance, accuracy, and decision-making reliability of data processing. This method uses the ST-GCN and GNN models to achieve unified modeling, in-depth analysis, fusion, and collaborative analysis of multi-source heterogeneous data, and constructs an adaptive optimization recommendation strategy based on DRL to strengthen the real-time performance and adaptability of the personalized recommendation strategy, providing users with accurate personalized route planning and safety warnings. At the same time, through containerized deployment and microservice architecture, the elastic expansion and dynamic resource scheduling of the system are realized, improving the scalability and overall operation efficiency of the system.

[0206] The above specific embodiments cannot be used to limit the protection scope of the present invention. For those skilled in the art of this technology, any alternative improvement or transformation made to the embodiments of the present invention falls within the protection scope of the present invention.

[0207] Where the present invention is not described in detail, it is all well-known technology to those skilled in the art of this technology.

Claims

1. An optimization method for traffic spatio-temporal data calculation based on a double-layer multi-round-trip operation model, characterized in that, The described computing optimization method includes the following steps: S1. Collect multiple multi-source data and represent them uniformly through multi-modal feature vectors. The multi-source data includes at least vehicle state parameters, user behavior data, and environmental data; S2. Use the ST-GCN model to extract features from the multi-modal feature vectors; S3. Generate personalized recommendation strategies based on deep reinforcement learning and combined with multiple optimization objectives of the vehicle system; S4. Optimize the personalized recommendation strategies using the deep network update formula; S5. Use the Pareto optimal solution to balance the recommendation strategies; S6. Perform multiple rounds of feedback and adaptive optimization; S7. Output personalized recommendations in the form of multi-dimensional outputs.

2. The traffic spatio-temporal data calculation optimization method based on the double-layer multi-round-trip operation model according to claim 1, characterized in that Collecting multiple multi-source data and representing them uniformly through multi-modal feature vectors includes the following steps: S1.

1. Construct multi-modal feature vectors, and the multi-modal feature vectors are: Among them, X t v is a vehicle state feature, including position, speed, and fuel consumption, which is used to reflect the operating state of the vehicle and evaluate the current performance and requirements of the vehicle; X t u is a user feature, including historical routes and driving styles, which is used to reflect the driving habits and preferences of the user; X t θ is an environmental feature, including road conditions and weather, which is used to reflect external conditions and adjust personalized recommendation strategies to adapt to changes in environmental conditions; S1.

2. Standardize data from different sources and at different scales: Among them, X t is the original data point that needs to be normalized, such as the speed parameter of a vehicle, the age parameter of a user, and the temperature parameter of the environment; μ is the average value of the original data set; ε is the standard deviation of the original data, which is used to represent the degree of data dispersion.

3. The traffic spatio-temporal data calculation optimization method based on the double-layer multi-round trip operation model according to claim 1, characterized in that The ST-GCN model is: Ht=σ(W1·Xt+W2·A·Xt) The fused feature representation is: Among them, W1 and W2 are network weights, which are used to capture the spatio-temporal correlation of the data; A is the adjacency matrix, which is used to represent the topological relationship between different features; H t is the extracted feature, and σ is the activation function used to introduce non-linearity; Z t The fused features are obtained by integrating vehicle features, user features, and environmental features, providing comprehensive inputs for subsequent decision-making.

4. The traffic spatio-temporal data calculation optimization method based on the double-layer multi-round-trip operation model according to claim 3, wherein Generating personalized recommendation strategies based on deep reinforcement learning and combined with multiple optimization objectives of the vehicle system includes the following steps: S3.

1. Define the state space: St=f(Zt) The state space is used to reflect the comprehensive state of the current system and provide a basis for decision-making; S3.

2. Define the action space: At=(a1,a2,...,an) It includes path planning, driving behavior adjustment, and energy consumption optimization, which are used to adjust the operation strategy of the vehicle; S3.

3. Define the reward function: R(St,at)=a1·Rtime+a2·Rfuel+a3·Rsafe The reward function should comprehensively consider three dimensions of time, energy consumption, and safety, and balance the importance of different dimensions by setting different weights a1, a2, and a3 to guide strategy optimization.

5. The traffic spatio-temporal data calculation optimization method based on the double-layer multi-round-trip operation model according to claim 4, characterized in that The deep network update formula is: Among them, Q(S t , a t ) is the state-action value function, which is used to represent the expected value of selecting different actions in the state space; η is the learning rate, which is used to control the update speed; γ is the discount factor, which is used to balance short-term and long-term rewards; Represents the optimal reward estimate value of the future state, which is used to guide the strategy to develop in the long-term optimal direction.

6. The traffic spatio-temporal data calculation optimization method based on the double-layer multi-round trip operation model according to claim 1, wherein The Pareto optimal solution is solved by the weighted sum method as: F(x)=∑λifi(x) Among them, λ i is the target weight, which is used to adjust the priorities of different targets.

7. The traffic spatio-temporal data calculation optimization method based on the double-layer multi-round-trip operation model according to claim 5, wherein Performing multiple rounds of feedback and adaptive optimization includes the following steps: S6.

1. Set the short-term feedback, which is expressed as: θt←θt+η▽θL(Zt,at,Rt) Among them, θ t represents the model parameters of the current time, which are used to determine the behavior of the recommendation strategy and define the mapping relationship from state to action; η is the learning rate, which is used to control the step size of model parameter updates and determines the amplitude of parameter adjustment in each iteration; ▽θL(Zt,at,Rt) represents the gradient of the loss function L with respect to the model parameter θ, which reflects the sensitivity of the loss function to each parameter in the current state; $L(Z_t, a_t, R_t)$ is a loss function used to measure the performance of the model in the current state. It is usually a function of features, actions, and rewards, reflecting the difference between the predicted value and the true value of the model; The short-term feedback can adjust the model parameters through gradient ascent to adapt to the immediate observed data and enhance the adaptability of the policy; S6.2, set the long-term feedback to be expressed as: where $L(\theta)$ is the loss function and $N$ is the number of samples; The long-term feedback optimizes the overall performance of the model by minimizing the squared difference between the predicted value and the true value, ensuring the stability and consistency of the policy.

8. The optimized method for calculating traffic spatio-temporal data based on the double-layer multi-round operation model according to claim 1, wherein The multi-dimensional output is: Output = {Route, ETA, Risk Level, Fuel Consumption, Driving Advice}; It includes route planning, estimated time of arrival, risk level, fuel consumption, and driving advice, providing comprehensive personalized recommendations for users.

9. The traffic spatio-temporal data calculation optimization method based on the double-layer multi-round operation model according to claim 1, characterized in that The computing optimization architecture corresponding to the computing optimization method includes a data collection and access layer, a stream processing layer, a batch processing layer, a unified scheduling and resource management layer, and an output and service layer; In the data collection and access layer, real-time data is obtained from multi-source devices such as in-vehicle terminal devices, in-vehicle sensors, and monitoring cameras, and duplicate removal, format conversion, data cleaning, and standardization are performed to generate a high-quality input data stream; In the stream processing layer, a spatio-temporal convolutional network is used to extract spatio-temporal features, combined with a graph neural network to identify traffic accidents, congestion, and abnormal events, and multiple rounds of calculations and optimizations are performed on the real-time data to improve data processing efficiency and decision-making accuracy; In the batch processing layer, historical traffic data is analyzed and predicted based on a deep learning model, and a traffic flow prediction and recommendation model is trained to optimize traffic pattern recognition and decision rule generation; In the unified scheduling and resource management layer, resource management is performed to achieve elastic expansion and load balancing of computing resources, and resource allocation is dynamically adjusted according to traffic prediction and load data to ensure the efficient operation of the system; In the output and service layer, it is used to provide personalized recommendations, safety warnings, and optimization suggestions to in-vehicle terminals and management platforms, and detailed optimization strategies and implementation suggestions to users.

Images