Intelligent traffic flow prediction system and method based on large model

Abstract

The invention relates to the technical field of computer software development, in particular to an intelligent traffic flow prediction system and method based on a large model, and the system comprises a multi-source data collection module, a data preprocessing module, a deep learning large model, a prediction output module, and a visual display module. The method has the beneficial effects that resource utilization is optimized while code generation is assisted. Through an advanced algorithm and mechanism, the system not only can efficiently generate code snippets according to the input of a developer, but also can intelligently monitor and optimize resource consumption in the code generation process, including but not limited to the utilization efficiency of resources such as a CPU, a memory and network bandwidth, so that the development efficiency of software is improved, and the development cost is reduced. And reasonable utilization and high-performance operation of system resources are ensured. The method is suitable for the field of software development, and is particularly suitable for automatic code writing and optimization of resource-sensitive development environments and large complex projects.

Description

Technical Field

[0001] This invention relates to the field of computer software development technology, specifically to an intelligent traffic flow prediction system and method based on a large model. Background Technology

[0002] Traditional traffic flow prediction methods mainly include statistical methods such as historical averaging, Kalman filtering, and ARIMA models, as well as machine learning methods based on support vector machines and neural networks. However, these methods have the following technical limitations when dealing with complex urban traffic systems: 1. Insufficient prediction accuracy: Traditional methods struggle to capture the nonlinear variation patterns and complex spatiotemporal dependencies of traffic flow, and prediction errors increase significantly under abnormal conditions such as emergencies and severe weather.

[0003] 2. Limited data processing capabilities: It is unable to effectively integrate and process multi-source heterogeneous traffic big data, including sensor data, GPS trajectory data, social media data, meteorological data, etc., resulting in insufficient utilization of information.

[0004] 3. Poor real-time performance: The computational complexity is high, making it difficult to meet the real-time prediction needs of large-scale urban traffic networks and unable to support dynamic traffic management decisions.

[0005] 4. Weak generalization ability: The model parameters are fixed and cannot adapt to changes in traffic characteristics in different regions and time periods, resulting in a sharp decline in performance in new scenarios.

[0006] 5. Insufficient scalability: It is difficult to scale to large-scale urban transportation networks and cannot support collaborative forecasting across regions and cities.

[0007] In recent years, deep learning technology has made groundbreaking progress in fields such as computer vision and natural language processing. In particular, the emergence of large-scale models such as Transformer and BERT has provided new technical pathways for processing complex sequence data. However, existing deep learning methods still face the following challenges in the application of traffic flow prediction: 1. The lack of specific design tailored to the characteristics of traffic data means that directly applying general models has limited effectiveness. 2. The graph structure characteristics and spatial dependencies of the transportation network were not fully considered. 3. Lack of effective multi-scale time modeling mechanisms 4. The efficiency of model training and inference needs to be improved. Therefore, there is an urgent need to develop an intelligent traffic flow prediction system based on deep learning large models that can efficiently process multi-source traffic big data, accurately model complex spatiotemporal dependencies, and achieve high-precision, real-time traffic flow prediction. Summary of the Invention

[0008] The purpose of this invention is to provide an intelligent traffic flow prediction system and method based on a large model, so as to solve the technical problems mentioned in the background art, such as low prediction accuracy, poor real-time performance, and weak generalization ability, and to achieve high-precision real-time prediction of urban traffic flow.

[0009] To achieve the above objectives, the present invention provides the following technical solution: an intelligent traffic flow prediction system based on a large model, comprising: The multi-source data acquisition module is used to collect traffic sensor data, GPS trajectory data, meteorological data, and traffic event data in real time. The data preprocessing module is used to clean, standardize, spatiotemporally align, and extract features from the collected multi-source data. Large-scale deep learning models, including Transformer encoders, graph neural networks, and multi-scale temporal modeling units, are used to learn the complex spatiotemporal dependencies of traffic data. The prediction output module is used to generate traffic flow prediction results at multiple time scales; The visualization module is used to present the prediction results to users in the form of charts.

[0010] Preferably, the Transformer encoder includes: Multi-head self-attention mechanism is used to capture long-distance temporal dependencies in traffic data; Spatiotemporal location coding unit, used to encode the sequential information of time series and the location information of spatial networks; Feedforward neural networks are used for nonlinear feature transformation. Residual connections and layer normalization are used to improve the stability of model training.

[0011] Preferably, the graph neural network includes: The graph construction module is used to model the road network as a graph structure, with road segments as nodes and adjacent relationships as edges. Graph convolutional layers are used to aggregate feature information from neighboring nodes; An adaptive adjacency matrix learning mechanism is used to automatically discover implicit spatial dependencies. Graph pooling layers are used to reduce the complexity of graphs.

[0012] Preferably, the multi-scale time modeling unit includes: Multi-resolution convolutional networks are used to capture short-term, medium-term, and long-term temporal patterns simultaneously. Dilated convolutional layers are used to expand the temporal receptive field; A time-attention mechanism is used to adaptively focus on important historical moments; The feature fusion layer is used to integrate features from different time scales.

[0013] Preferably, the data preprocessing module includes: The data cleaning unit is used to identify and process outliers, missing values, and duplicate values. Data standardization units are used to unify the temporal granularity and spatial resolution of different data sources; Feature engineering units are used to extract temporal, spatial, and statistical features; The data augmentation unit is used to augment the training samples through data transformation.

[0014] A method for intelligent traffic flow prediction based on a large model includes the following steps: S1: Real-time acquisition of traffic sensor data, GPS trajectory data, meteorological data, and traffic event data through a multi-source data acquisition module; S2: Cleaning, standardizing, spatiotemporally aligning, and feature extraction of the collected multi-source data; S3: Construct a road network graph structure, using road segments as nodes and adjacent relationships as edges; S4: Use the Transformer encoder to learn the temporal dependencies of traffic data; S5: Using graph neural networks to learn spatial dependencies in road networks; S6: Use multi-scale temporal modeling units to capture traffic patterns at different time scales; S7: Integrates spatiotemporal features to generate traffic flow prediction results; S8: Display the prediction results through a visual interface.

[0015] Preferably, the data preprocessing in step S2 includes: Use statistical methods to detect and handle outliers; Use interpolation methods to fill in missing values; Use Z-score to standardize and unify the data distribution; Extract time features such as hours, days of the week, and holidays; Extract spatial features such as road grade and number of lanes.

[0016] Preferably, the Transformer encoder processing in step S4 includes: Calculate the query matrix, key matrix, and value matrix; The attention weights are calculated using a multi-head self-attention mechanism. Add spatiotemporal location encoding information; Feature transformation is performed using a feedforward neural network; Apply residual connections and layer normalization.

[0017] Preferably, the graph neural network processing in step S5 includes: Construct an adjacency matrix based on the road network topology; Aggregate features of neighboring nodes through graph convolution operations; Learn to use adaptive adjacency matrices to discover implicit relationships; Applying graph pooling reduces computational complexity.

[0018] Preferably, the multi-scale time modeling in step S6 includes: Capture multi-scale temporal patterns using different kernel sizes; Dilated convolution is used to expand the temporal receptive field; Apply time-attention mechanisms to focus on important moments; Integrate time features at different scales.

[0019] Compared with the prior art, the beneficial effects of the present invention are: This invention proposes a large-model-based intelligent traffic flow prediction system and method that optimizes resource utilization while assisting in code generation. Through advanced algorithms and mechanisms, the system can not only efficiently generate code snippets based on developer input, but also intelligently monitor and optimize resource consumption during the code generation process, including but not limited to the utilization efficiency of resources such as CPU, memory, and network bandwidth. This improves software development efficiency while ensuring the rational use of system resources and high-performance operation. This invention is applicable to the software development field, especially to the automated writing and optimization of code in resource-sensitive development environments and large, complex projects. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the present invention clear and complete, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only some, not all, embodiments of the present invention, and are merely illustrative of the embodiments of the present invention. They are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Example 1: This invention provides a technical solution: an intelligent traffic flow prediction system based on a large model, comprising the following core modules: 1. Multi-source data acquisition and preprocessing module: The data acquisition submodule is responsible for collecting multi-source heterogeneous traffic data in real time. Sensor data: This includes data such as traffic flow, vehicle speed, occupancy rate, and vehicle type classification collected by loop detectors, microwave detectors, and video detectors.

[0023] Floating car data: Location and speed information generated by mobile devices such as GPS trajectory data, taxi trajectory data, and bus GPS data.

[0024] Mobile signaling data: Population movement pattern data based on mobile phone signaling.

[0025] Meteorological data: environmental factors affecting traffic such as temperature, humidity, rainfall, visibility, and wind speed; Event data: Information on emergencies such as traffic accidents, road construction, and large-scale events; POI data: static geographic information such as point of interest distribution and land use types; Social media data: Traffic-related information from social media platforms such as Weibo and WeChat.

[0026] The data preprocessing submodule employs advanced big data processing technologies: Data cleaning: Detecting and processing outliers, missing values, and duplicate values ​​using statistical methods and machine learning algorithms; Data standardization: Use methods such as Z-score standardization and Min-Max standardization to unify the units and distribution of different data sources; Spatiotemporal alignment: unifying data with different temporal and spatial granularities into the same spatiotemporal grid; Feature engineering: Extracting time features (hours, days of the week, months, holidays, etc.), spatial features (road grade, number of lanes, speed limits, etc.), and traffic features (historical mean, variance, trend, etc.). Data augmentation: Expanding training data through methods such as temporal sliding windowing, spatial interpolation, and noise injection.

[0027] 2. Core Architecture of Large-Scale Deep Learning Models Spatiotemporal Transformer Encoder: It adopts a multi-layer Transformer architecture, with each layer containing a multi-head self-attention mechanism, a feedforward neural network, residual connections, and layer normalization; it designs a dedicated spatiotemporal position encoding, which simultaneously models the absolute and relative positional relationships in time and space; it introduces a learnable spatiotemporal masking mechanism to adaptively focus on important spatiotemporal regions; and it supports variable-length sequence inputs to adapt to the needs of different prediction time windows.

[0028] Graph Neural Network Spatial Modeling: Construct a multi-layer Graph Convolutional Neural Network (GCN) to model the topology of the road network; design an adaptive graph learning mechanism to automatically discover hidden spatial dependencies; adopt a Graph Attention Network (GAT) to assign different weights to different neighbor nodes; support dynamic graph structures to adapt to changes in the road network.

[0029] Multi-scale temporal modeling: Design a multi-resolution temporal convolutional network to capture short-term, medium-term, and long-term temporal dependencies; employ causal convolution and dilated convolution to expand the receptive field while maintaining computational efficiency; introduce a temporal attention mechanism to adaptively focus on important historical moments.

[0030] Multi-task learning framework: Simultaneously predicts multiple traffic indicators (flow, speed, density, occupancy, etc.); adopts an architecture of shared encoder and task-specific decoder; designs an adaptive weight adjustment mechanism to balance the learning of different tasks; 3. Distributed training and inference module.

[0031] Distributed training framework: It adopts a distributed training strategy that combines data parallelism and model parallelism; it implements gradient compression and communication optimization to reduce the communication overhead of distributed training; it supports both asynchronous and synchronous update modes; and it integrates automatic mixed precision training to improve training efficiency.

[0032] Model compression and optimization: Knowledge distillation is used to transfer knowledge from large models to lightweight models; model pruning and quantization are implemented to reduce model parameters and computational load; a dynamic inference mechanism is designed to adjust the model depth according to the input complexity.

[0033] Online learning and incremental updates: Implement an online learning algorithm based on a sliding window; adopt meta-learning technology to quickly adapt to new traffic patterns; design a model version management mechanism to support smooth model updates. 4. Prediction output and visualization module.

[0034] Multi-timescale prediction: Short-term forecasts: high-frequency forecasts for 5-minute, 15-minute, and 30-minute periods; Medium-term forecasts: 1-hour, 3-hour, and 6-hour mid-frequency forecasts; Long-term forecasts: Low-frequency forecasts for 1 day, 3 days, and 7 days.

[0035] Uncertainty Quantification: Bayesian deep learning methods are used to quantify prediction uncertainty; prediction confidence intervals and probability distributions are provided; and risk-aware decision-making is supported.

[0036] Visualization features include: real-time traffic flow heatmap; time series plot of prediction results; traffic network topology visualization; and model attention weight visualization.

[0037] Core algorithm flow: 1. Multi-source data fusion: 2: Spatiotemporal Feature Extraction: Time characteristics: periodicity (hours, weeks, months), trend characteristics, and holiday characteristics; Spatial characteristics: road attributes, POI distribution, network topology characteristics; Traffic characteristics: historical statistical characteristics, real-time status characteristics; 3: Graph Structure Construction: Construct an adjacency matrix A∈R^(N×N) A_ij = 1 if node i and node j are adjacent, otherwise 0. It supports multiple graph construction strategies: distance graph, similarity graph, and adaptive graph.

[0038] Training large deep learning models 1. Model Architecture Design: Encoder:Multi-layerTransformer+GCN Decoder:Multi-taskpredictionheads Loss:L=α·L_mse+β·L_mae+γ·L_uncertainty 2. Training Strategy: The TeacherForcing training strategy is adopted; the course learning is implemented, and training is carried out step by step from simple to complex; learning rate warm-up and cosine annealing scheduling are used; early stopping and model checkpointing mechanisms are integrated.

[0039] 3. Hyperparameter optimization: Bayesian optimization is used to automatically search for the optimal hyperparameters; multi-objective optimization is supported, balancing accuracy and efficiency.

[0040] Real-time forecasts and online updates 1. Real-time reasoning: Input: Feature x_t at the current time step; Output: Predictions for the next H time steps ŷ_{t+1:t+H}; Inference time: <100ms (single prediction); 2. Online learning: Incremental learning based on sliding window; Concept drift detection and model adaptation; Federated learning supports multi-regional collaborative training; 3. Performance monitoring: Real-time monitoring of prediction accuracy and system performance; Automatically triggers model retraining mechanism; Provides a detailed performance analysis report.

[0041] Example 2, based on Example 1, proposes an intelligent traffic flow prediction method based on a large model, applying the system described in claim 5, including the following steps: 1: Multi-source data acquisition and preprocessing The multi-source data acquisition module of this invention collects traffic information from different data sources in real time through a standardized interface protocol. The specific implementation process is as follows: Sensor data acquisition: Loop detectors, microwave detectors, and video detectors are deployed on major urban roads to collect basic traffic parameters such as traffic flow, vehicle speed, and vehicle occupancy. The detector sampling frequency is set to 30 seconds, and data transmission uses a 4G / 5G wireless network to ensure real-time requirements.

[0042] GPS trajectory data processing: GPS trajectory data from commercial vehicles such as taxis, buses, and trucks are collected. A map matching algorithm is used to match GPS points to the road network, extracting traffic status information at the road segment level. For trajectories with poor data quality, Kalman filtering is used for smoothing.

[0043] Meteorological data fusion: Obtain meteorological data such as temperature, humidity, rainfall, and visibility from meteorological departments to analyze the impact of meteorological conditions on traffic flow. Establish a meteorological-traffic impact model to quantify the magnitude of traffic flow changes under different meteorological conditions.

[0044] Data preprocessing process: First, data cleaning is performed to identify and handle outliers and missing values; then, data standardization is performed to unify the time granularity and spatial resolution of different data sources; finally, feature engineering is performed to extract time features (hours, days of the week, holidays), spatial features (road class, number of lanes), and historical statistical features.

[0045] 2: Design of large-scale deep learning models The deep learning large-scale model of this invention adopts a multi-layered architecture design, and the specific implementation is as follows: Transformer encoder design: A 12-layer Transformer encoder is constructed, with each layer containing a multi-head self-attention mechanism and a feedforward neural network. The number of attention heads is set to 16, and the hidden layer dimension is 1024. A specialized spatiotemporal location encoding is designed to model both the sequential relationships of the time series and the topological relationships of the spatial network.

[0046] Graph Neural Network Modeling: A graph convolutional neural network is used to model the road network structure. Road segments are treated as graph nodes, and connections between adjacent road segments are treated as graph edges. An adaptive adjacency matrix learning mechanism is designed to automatically discover implicit spatial dependencies. The number of graph convolutional layers is set to 4, with each layer having an output dimension of 256.

[0047] Multi-scale temporal modeling: A multi-resolution temporal convolutional network was designed to simultaneously capture short-term (5-30 minutes), medium-term (1-6 hours), and long-term (1-7 days) temporal dependency patterns. Dilated convolutions were used to expand the receptive field, with kernel sizes set to 3, 5, and 7, and dilation rates of 1, 2, and 4, respectively.

[0048] Loss function design: A multi-task learning framework is adopted to simultaneously predict multiple traffic indicators such as flow rate, speed, and density. A weighted loss function is designed to balance the importance of different tasks. The loss function includes mean squared error loss, mean absolute error loss, and quantile loss, with a weight ratio of 0.4:0.4:0.2.

[0049] 3: Model Training and Optimization The model training employs a distributed training strategy, and the specific implementation process is as follows: Data preparation: Historical traffic data were divided into training, validation, and test sets in a 7:2:1 ratio. Training samples were constructed using a sliding window method, with an input window length of 60 time steps (5 hours) and a prediction window length of 12 time steps (1 hour).

[0050] Model initialization: Model parameters were initialized using the Xavier initialization method. The learning rate was set to 0.001, the Adam optimizer was used, and the batch size was set to 64. A learning rate decay strategy was implemented, decreasing by a factor of 0.1 every 20 epochs.

[0051] Training strategy: An early stopping mechanism is used to prevent overfitting; training stops when the validation set loss does not decrease for 10 consecutive epochs. Gradient clipping is used to limit the gradient norm to within 1.0. Model performance is evaluated on the validation set after each epoch.

[0052] Distributed training: A data-parallel strategy was employed, distributing training data across 8 GPUs for parallel training. The AllReduce algorithm was used to synchronize gradients, ensuring training consistency. The total training time was approximately 24 hours, and the converged model achieved a mean absolute error of 8.5% on the test set.

[0053] 4: Deployment of Real-Time Prediction System The real-time prediction system adopts a cloud-edge collaborative architecture, and the specific deployment scheme is as follows: Cloud-based training platform: Deploy the model training platform in the cloud, configuring a high-performance GPU cluster for large-scale model training. Establish an automated training pipeline and regularly update the model with the latest data. The cloud platform is responsible for model version management, performance monitoring, and A / B testing.

[0054] Edge inference nodes: Edge inference nodes are deployed in the traffic management center and configured with lightweight prediction models for real-time inference. Edge nodes have local data caching, hot model updates, and fault recovery capabilities. The response time for a single prediction is controlled within 50 milliseconds.

[0055] Data transmission optimization: Data compression and incremental transmission technologies are employed to reduce data transmission volume between the cloud and the edge. A multi-level caching mechanism is established to improve data access efficiency. A network outage resumption function is designed to ensure normal system operation during network anomalies.

[0056] Load balancing: Deploy multiple edge inference nodes and use a load balancing algorithm to distribute prediction requests. The request distribution strategy is dynamically adjusted based on the node load to ensure optimal overall system performance.

[0057] 5: Application of Prediction Results The prediction results are displayed to traffic management personnel through a visualization interface, with specific applications as follows: Real-time monitoring: A real-time traffic flow heat map is displayed on a large screen in the traffic management center, with different colors representing different levels of congestion. Managers can intuitively understand the city's traffic conditions and promptly detect any anomalies.

[0058] Forecast Display: Provides traffic flow forecast curves for the next 1 hour, 3 hours, and 6 hours to help managers formulate traffic control measures in advance. The forecast results include confidence intervals, quantifying the uncertainty of the forecast.

[0059] Intelligent dispatching: Based on prediction results, the system automatically generates traffic signal optimization plans, route guidance suggestions, and emergency plans. It can automatically adjust traffic light timings to optimize traffic flow distribution.

[0060] Decision support: Providing data support for transportation planning and infrastructure construction. Through long-term predictive analysis, identifying traffic bottlenecks and guiding road expansion and new construction projects.

[0061] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart traffic flow prediction system based on a large model, characterized in that: include: The multi-source data acquisition module is used to collect traffic sensor data, GPS trajectory data, meteorological data, and traffic event data in real time. The data preprocessing module is used to clean, standardize, spatiotemporally align, and extract features from the collected multi-source data. Large-scale deep learning models, including Transformer encoders, graph neural networks, and multi-scale temporal modeling units, are used to learn the complex spatiotemporal dependencies of traffic data. The prediction output module is used to generate traffic flow prediction results at multiple time scales; The visualization module is used to present the prediction results to users in the form of charts.

2. The intelligent traffic flow prediction system based on a large model according to claim 1, characterized in that: The Transformer encoder includes: Multi-head self-attention mechanism is used to capture long-distance temporal dependencies in traffic data; Spatiotemporal location coding unit, used to encode the sequential information of time series and the location information of spatial networks; Feedforward neural networks are used for nonlinear feature transformation. Residual connections and layer normalization are used to improve the stability of model training.

3. The intelligent traffic flow prediction system based on a large model according to claim 2, characterized in that: The graph neural network includes: The graph construction module is used to model the road network as a graph structure, with road segments as nodes and adjacent relationships as edges. Graph convolutional layers are used to aggregate feature information from neighboring nodes; An adaptive adjacency matrix learning mechanism is used to automatically discover implicit spatial dependencies. Graph pooling layers are used to reduce the complexity of graphs.

4. The intelligent traffic flow prediction system based on a large model according to claim 3, characterized in that: The multi-scale time modeling unit includes: Multi-resolution convolutional networks are used to capture short-term, medium-term, and long-term temporal patterns simultaneously. Dilated convolutional layers are used to expand the temporal receptive field; A time-attention mechanism is used to adaptively focus on important historical moments; The feature fusion layer is used to integrate features from different time scales.

5. The intelligent traffic flow prediction system based on a large model according to claim 4, characterized in that: The data preprocessing module includes: The data cleaning unit is used to identify and process outliers, missing values, and duplicate values. Data standardization units are used to unify the temporal granularity and spatial resolution of different data sources; Feature engineering units are used to extract temporal, spatial, and statistical features; The data augmentation unit is used to augment the training samples through data transformation.

6. A method for intelligent traffic flow prediction based on a large model, using the system described in claim 5, characterized in that: Includes the following steps: S1: Real-time acquisition of traffic sensor data, GPS trajectory data, meteorological data, and traffic event data through a multi-source data acquisition module; S2: Cleaning, standardizing, spatiotemporally aligning, and feature extraction of the collected multi-source data; S3: Construct a road network graph structure, using road segments as nodes and adjacent relationships as edges; S4: Use the Transformer encoder to learn the temporal dependencies of traffic data; S5: Use graph neural networks to learn spatial dependencies in road networks; S6: Use multi-scale temporal modeling units to capture traffic patterns at different time scales; S7: Integrates spatiotemporal features to generate traffic flow prediction results; S8: Display the prediction results through a visual interface.

7. The intelligent traffic flow prediction method based on a large model according to claim 6, characterized in that: The data preprocessing in step S2 includes: Use statistical methods to detect and handle outliers; Use interpolation methods to fill in missing values; Use Z-score to standardize and unify the data distribution; Extract time features such as hours, days of the week, and holidays; Extract spatial features such as road grade and number of lanes.

8. The intelligent traffic flow prediction method based on a large model according to claim 7, characterized in that: The Transformer encoder processing in step S4 includes: Calculate the query matrix, key matrix, and value matrix; The attention weights are calculated using a multi-head self-attention mechanism. Add spatiotemporal location encoding information; Feature transformation is performed using a feedforward neural network; Apply residual connections and layer normalization.

9. The intelligent traffic flow prediction method based on a large model according to claim 8, characterized in that: The graph neural network processing in step S5 includes: Construct an adjacency matrix based on the road network topology; Aggregate features of neighboring nodes through graph convolution operations; Learn to use adaptive adjacency matrices to discover implicit relationships; Applying graph pooling reduces computational complexity.

10. The intelligent traffic flow prediction method based on a large model according to claim 9, characterized in that: The multi-scale time modeling in step S6 includes: Capture multi-scale temporal patterns using different kernel sizes; Dilated convolution is used to expand the temporal receptive field; Use time-attention mechanisms to focus on important moments; Integrate time features at different scales.

Images