A traffic flow dynamic prediction method based on spatio-temporal correlation feature fusion of multi-source heterogeneous data

Abstract

A kind of traffic flow dynamic prediction method based on multi-source heterogeneous data spatio-temporal correlation feature fusion, including multi-source heterogeneous data acquisition module, multi-source data standardization processing module, heterogeneous data feature fusion module, traffic flow prediction scheduling module, decision support and service module, system management and monitoring module, multi-source heterogeneous data acquisition module is used to real-time acquisition multi-source heterogeneous data, multi-source data standardization processing module is used for data pre-processing, heterogeneous data feature fusion module is used for data feature extraction and fusion, traffic flow prediction scheduling module is used for traffic flow prediction, decision support and service module is used for report and decision generation, system management and monitoring module is used for system monitoring and management, a kind of traffic flow dynamic prediction method based on multi-source heterogeneous data spatio-temporal correlation feature fusion of the present application, propose multi-source heterogeneous data fusion algorithm based on neural network to carry out feature fusion to data, propose deep learning based traffic flow prediction algorithm to traffic flow dynamic prediction.

Description

Technical Field

[0001] This invention relates to the fields of neural networks, deep learning, and dynamic prediction technology, specifically to a method for dynamic traffic flow prediction based on the fusion of spatiotemporal correlation features of multi-source heterogeneous data. Background Technology

[0002] Neural network technology is a computational model based on a multi-layered neuron structure, designed to address the challenges of direct alignment between multi-source heterogeneous data, complex correlations, and high nonlinearity. In a traffic flow dynamic prediction method based on the fusion of spatiotemporal correlation features of multi-source heterogeneous data, multi-source heterogeneous traffic data exhibit significant differences in temporal scale, spatial granularity, and semantic expression. Traditional linear methods struggle to characterize their intrinsic correlations. Neural network technology, by constructing a multi-layered mapping structure, performs unified representation learning on features from different data sources. It can calculate cross-source feature similarity and achieve feature space mapping alignment, thereby completing deep fusion in the same semantic space and generating high-value spatiotemporal correlation feature vectors with consistent semantics and spatiotemporal meaning. This provides a more complete and reliable feature foundation for traffic flow dynamic prediction.

[0003] Deep learning and dynamic prediction technology is a technical system based on deep network structures to automatically learn and model complex spatiotemporal patterns. It aims to solve the problems of strong nonlinearity, strong temporal dependence, and difficulty in accurately depicting sudden changes in traffic flow during its temporal evolution. In a traffic flow dynamic prediction method based on the fusion of spatiotemporal correlation features of multi-source heterogeneous data, traffic status is not only affected by short-term proximity changes, but also exhibits obvious daily, weekly, and event-driven characteristics. A single static model is difficult to simultaneously take into account changes at multiple time scales. Deep learning and dynamic prediction technology constructs a spatiotemporal deep model to hierarchically model the fused multi-source heterogeneous features. While learning spatial correlations, it depicts the dynamic evolution law of time, realizes the dynamic calculation and alignment of cross-source feature similarity, and continuously updates the state representation during the prediction process, thereby generating high-value spatiotemporal correlation feature vectors that can reflect the real-time traffic operation status and long-term evolution trend.

[0004] Existing traffic flow dynamic prediction methods based on the fusion of spatiotemporal correlation features of multi-source heterogeneous data suffer from several drawbacks. They rely on rule-driven and shallow models to process multi-source data, making it difficult to effectively characterize the nonlinear correlation between heterogeneous data. This results in insufficient cross-source feature alignment capabilities and low information density in the fusion results. Furthermore, existing methods employ static and fixed-frequency prediction strategies, which cannot adjust prediction behavior in real time according to changes in traffic conditions. Their responsiveness to sudden events and complex scenarios is limited, and their overall prediction accuracy, stability, and engineering applicability are all insufficient, making it difficult to meet the actual needs of complex urban traffic systems for high-precision and robust dynamic prediction. Summary of the Invention

[0005] The purpose of this invention is to provide a traffic flow dynamic prediction method based on the fusion of spatiotemporal correlation features of multi-source heterogeneous data. This addresses the problems of existing traffic flow dynamic prediction methods based on the fusion of spatiotemporal correlation features of multi-source heterogeneous data mentioned in the background art. These methods rely on rule-driven and shallow models to process multi-source data, making it difficult to effectively characterize the nonlinear correlation between heterogeneous data. This results in insufficient cross-source feature alignment capability and low information density of the fusion results. Furthermore, existing methods employ static prediction and fixed-frequency prediction strategies, which cannot adjust the prediction behavior in real time according to changes in traffic conditions. Their response capability to sudden events and complex scenarios is limited, and their overall prediction accuracy, stability, and engineering applicability are all insufficient, making it difficult to meet the actual needs of complex urban traffic systems for high-precision and robust dynamic prediction.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a traffic flow dynamic prediction method based on the fusion of spatiotemporal correlation features of multi-source heterogeneous data, comprising a multi-source heterogeneous data acquisition module, a multi-source data standardization processing module, a heterogeneous data feature fusion module, a traffic flow prediction and scheduling module, a decision support and service module, and a system management and monitoring module, characterized in that: the multi-source heterogeneous data acquisition module is used to collect raw traffic data in real time from heterogeneous data infrastructure including fixed traffic detectors, floating car trajectory data streams, urban event semantic data sources, GPS data, and meteorological monitoring networks; the multi-source data standardization processing module is used to clean, denoise, imput missing values, and standardize the collected raw data to eliminate noise and scale differences; the heterogeneous data feature fusion module includes a feature layering extraction unit and a cross-source feature fusion unit, wherein the feature layering extraction unit is used for layering feature extraction... The system extracts the spatiotemporal and semantic features of traffic flow to obtain multi-dimensional spatiotemporal patterns. The cross-source feature fusion unit proposes a neural network-based multi-source heterogeneous data fusion algorithm to calculate cross-source feature similarity and achieve mapping alignment, deeply fusing heterogeneous features to generate high-value spatiotemporal correlation feature vectors. The traffic flow prediction and scheduling module proposes a deep learning-based traffic flow prediction algorithm for short-term and medium-to-long-term dynamic prediction of the fused features, and manages the real-time scheduling, priority allocation, and resource optimization of prediction tasks. The decision support and service module transforms the prediction results into visual reports, traffic flow management strategies, and early warning signals, providing operable decision-making basis and external service interfaces for urban traffic management. The system management and monitoring module provides users with an interactive interface for real-time monitoring of data flow, model performance, and system resources, and automatically detects anomalies, triggers alarms, and performs maintenance operations.

[0007] Preferably, the multi-source heterogeneous data acquisition module integrates multiple protocol access interfaces, including MQTT protocol to interface with fixed traffic detectors, HTTP API interface to call urban event semantic data sources, Kafka stream processing interface to receive floating car trajectory and GPS data streams, and real-time data from meteorological monitoring networks, to achieve full-dimensional coverage of traffic multi-source heterogeneous data acquisition. The data sources include road network sensors for traffic flow and speed, mobile terminals for taxi and ride-hailing vehicle trajectories, unstructured event text from social media and traffic management reports, satellite positioning data, and meteorological parameters such as temperature and rainfall, providing highly timely input for subsequent processing.

[0008] Preferably, the multi-source data standardization processing module performs data cleaning by applying an outlier filtering threshold based on historical statistics, performs noise reduction using wavelet transform noise reduction technology, performs spatiotemporal context-aware missing value filling by combining KNN interpolation with road network topology, and performs multi-scale data standardization processing using Z-score normalization and quantile transformation. This achieves noise suppression, missing value repair, and dimensional unification of the original multi-source heterogeneous data, eliminates the differences in accuracy, inconsistent sampling frequencies, and unit scale conflicts among multi-source data, and outputs a structured, highly consistent, high-quality dataset, laying a robust foundation for subsequent feature extraction.

[0009] Preferably, the heterogeneous data feature fusion module includes a feature layering extraction unit. The feature layering extraction unit integrates the time series processing library built into Apache Flink, the spatial relationship modeling module of PyTorch Geometric, and the text parsing pipeline of Hugging Face Transformers to build a layered parallel computing architecture. This enables automated feature extraction of multi-dimensional spatiotemporal patterns of traffic flow, generating an independent feature vector set that includes historical traffic evolution patterns, road network spatial propagation effects, and the impact of external event disturbances. This fully explores the inherent complementarity between multi-source heterogeneous data.

[0010] Preferably, the heterogeneous data feature fusion module includes a cross-source feature fusion unit. The cross-source feature fusion unit proposes a multi-source heterogeneous data fusion algorithm based on neural networks. By designing a cross-source alignment mechanism based on neural networks, the backpropagation neural network establishes a basic mapping model, projects data from different sources onto a unified feature space, and then combines particle swarm optimization technology to form a hybrid fusion strategy. This achieves deep fusion generation of high-value spatiotemporal correlation feature vectors, improves the spatiotemporal consistency and prediction reliability of cross-source traffic data, and provides strong representational input for subsequent traffic flow prediction.

[0011] Preferably, the neural network-based multi-source heterogeneous data fusion algorithm is as follows: First, a three-layer backpropagation neural network is constructed as the basic mapping model for multi-source heterogeneous data fusion. The model projects traffic data from the multi-source data standardization processing module onto a unified feature space. In the constructed neural network, the input layer receives standardized multi-source traffic features, the hidden layer performs nonlinear transformations, and the output layer generates the fused traffic state estimate, as expressed by the following formula: in, The input feature vector is represented by a dimension of 5 and includes floating car trajectory data, GPS data, fixed traffic detector data, urban event semantic data, and meteorological monitoring network data. This is represented as the weight matrix from the input layer to the hidden layer, with dimensions [5,4]. It is represented as a hidden layer bias vector with dimensions [5,1]; It is represented as the weight matrix from the hidden layer to the output layer, with dimensions [1,5]. This represents the output layer bias vector, with dimensions [1,1]. It is represented as an activation function, using the hyperbolic tangent function, with a range of [-1, 1], which can effectively handle nonlinear mappings and avoid the gradient vanishing problem; The output of the neural network represents the fused traffic state estimate. The constructed neural network uses a 5-5-1 network structure: the input layer contains 5 neurons corresponding to 5 input feature parameters, the hidden layer contains 5 neurons responsible for nonlinear feature transformation and pattern extraction, and the output layer contains 1 neuron producing a single fused output. This approach ensures model expressiveness while avoiding overfitting. Then, a particle swarm optimization mechanism combined with genetic optimization is designed to improve the training effect of the neural network. First, a fitness function is defined to evaluate network performance. Then, the population is divided into three subpopulations based on fitness values, and different optimization strategies are applied to each subpopulation to ensure a balance between global exploration and local exploitation of the search space. The specific formula is as follows: in, This is represented as an individual fitness value; a higher value indicates better individual performance. This is represented as the fitness adjustment factor, with a value of 1.5. The value has been verified through numerous experiments to produce fitness values ​​with strong discriminative power within different error ranges; This represents the total number of training samples, with 100-200 samples selected to ensure sufficient training. Represented as the first The network prediction output for each sample; Indicates the first For each sample's actual traffic state value, the fitness function is designed as the reciprocal of the sum of squared errors, so that individuals with smaller errors receive higher fitness. The coefficient ensures that the fitness value remains stable within a reasonable range, avoiding computational accuracy issues caused by values ​​that are too large or too small, and the population size... Setting the threshold to 100 balances computational efficiency and search quality, and is the experimentally verified optimal size. By dynamically adjusting the neural network weights and thresholds, the problems of traditional backpropagation neural networks easily getting trapped in local optima and slow convergence speed are solved. This enables the neural network to adaptively learn complex nonlinear relationships between multi-source data, improving the accuracy and stability of feature fusion. Secondly, a population hierarchical optimization strategy is constructed. By sorting the particle swarm according to its fitness value from high to low and dividing it into three subpopulations, customized optimization operations are applied to different subpopulations. High-fitness individuals are retained, medium-fitness individuals generate better offspring through crossover, and low-fitness individuals explore new solution spaces through mutation. The specific formula is expressed as follows: in, This represents an optimized particle velocity vector, with an initial velocity range of [-1, 1], and a maximum velocity limit. minimum speed This prevents particles from jumping out of the optimal region too quickly; This represents the optimized particle position vector, corresponding to the encoding of neural network weights and biases, limiting the maximum position. minimum position This ensures the range of feasible solutions. It is represented as inertia weight, with a value range of [0.4, 0.9]. It adopts a linear decreasing strategy from 0.9 to 0.4. In the early stage, a large inertia is conducive to global exploration, and in the later stage, a small inertia is conducive to local development. This is represented as an individual learning factor, set to 1.49, which controls the intensity of particle movement towards the individual's historical best position; Represented as the global learning factor, set to 1.69, it controls the intensity of particle movement towards the global optimal position of the swarm; and It is represented as a random number uniformly distributed in the interval [0,1], and randomness is introduced to enhance the diversity of algorithms; This represents the individual's historical best position, recording the corresponding particle's historical best solution; This represents the global optimal position of the population and records the best solution throughout the entire history of the population. Represented as the current iteration number, maximum iteration number Experiments revealed that, while ensuring convergence accuracy, the computational complexity was controlled. The particle velocity and position update mechanism achieved a balance between global search and local exploitation, effectively avoiding premature mastery. Then, the optimal parameters of the neural network were determined through an iterative optimization process, and the individual fitness values ​​were adjusted accordingly. As an evaluation criterion, the first The particle position vector of the next iteration This is transformed into a neural network weight and threshold matrix, enabling the algorithm to adaptively learn complex mapping relationships between multiple data sources, overcoming the limitations of a single data source. After each iteration, based on the fitness value... The optimal individual is selected, and the global optimal position is determined when the convergence condition and the maximum number of iterations are met. The corresponding parameters are assigned to the neural network to construct the final fusion model, as shown in the following formula: in, This is represented as the optimized input-to-hidden-layer weight matrix, with dimensions [5,4] and element values ​​ranging from [-5,5]. This is represented as the optimized hidden layer bias vector, with dimensions [5,1] and element values ​​ranging from [-3,3]. This is represented as the optimized weight matrix from the hidden layer to the output layer, with dimensions [1,5] and element values ​​ranging from [-5,5]. This is represented as the optimized output layer bias vector, with dimensions [1,1] and values ​​ranging from [-2,2]. These represent the number of neurons in the input layer, hidden layer, and output layer, respectively, corresponding to the 5-5-1 network structure; This represents the number of weight parameters from the input layer to the hidden layer; This is expressed as the number of hidden layer bias parameters; This indicates the end position of the segment for the first three parameters; Represented as a matrix reconstruction function, it transforms a vector into a matrix of a specified dimension. By converting the optimized particle position vector into network parameters, a high-precision multi-source data fusion model is established. Finally, high-value spatiotemporal correlation feature vectors are generated. Utilizing the nonlinear mapping capability of the constructed neural network, heterogeneous data is projected onto a unified semantic space. A weighted fusion mechanism is used to highlight the contributions of key features. The specific formula is expressed as follows: in, It is represented as a fused high-value spatiotemporal correlation feature vector with a dimension of 7, containing the activation values ​​of 5 hidden layer neurons and 2 features derived from the optimization process; It is represented as a feature normalization function, which uses Min-Max standardization to map the feature values ​​to the [0,1] interval; Represented as the first The importance weights of each hidden neuron are calculated using the Softmax function to ensure that important features receive higher weights. Represented as the index of the number of hidden neurons; and These represent the optimized inputs to the first... The weight vectors and biases of each hidden neuron; This is represented as a vector concatenation operation; It is represented as a feature weight vector of the optimization process, which is determined by cross-validation to balance the convergence characteristics of the algorithm and the data fitting characteristics. Represented as the Hadamard product; It is represented as the deviation vector between the particle's current position and the global optimal position, with a dimension of 2. By applying the optimized neural network parameters to multi-source traffic data, time series features, spatial topological features, and semantic association features are deeply integrated to extract feature representations containing spatiotemporal patterns, overcoming the limitations of a single data source and improving the prediction accuracy and robustness of subsequent modules.

[0012] Preferably, the traffic flow prediction and scheduling module proposes a traffic flow prediction algorithm based on deep learning. By constructing a spatiotemporal deep learning model and a dynamic task scheduling strategy, it intelligently manages the triggering timing of prediction tasks, the allocation of computing resources, and emergency event responses, thereby achieving dynamic prediction of short-term and medium-to-long-term traffic flow status.

[0013] Preferably, the deep learning-based traffic flow prediction algorithm is as follows: First, a sparse attention graph convolutional network is constructed as the core model for dynamic traffic flow prediction. By designing a graph convolution operator that expands the receptive field, local and non-local spatial dependencies in the traffic network are obtained, enabling the model to simultaneously perceive the complex spatial relationships between neighboring road segments and distant nodes. The traffic network represents a graph structure, nodes represent road segments and intersections, and edges represent spatial connections. The weights between different nodes are dynamically adjusted through a sparse attention mechanism, enabling the model to adaptively learn the strength of spatial dependencies. The specific formula is expressed as follows: in, Represented as nodes The output feature vector; Represented as the ReLU activation function; Represented as nodes of Neighbor set; Expressed as the expansion rate, it has been experimentally verified that it can achieve the best balance between computational efficiency and receptive field range; Represented as the neighbor order, with a value range of 1. It covers spatial dependencies from local to global. Represented as a similarity function, It is represented as a two-layer fully connected network that maps input features to a 32-dimensional latent space; Represented as a learnable weight matrix, Represented as the real number field; and Represented as nodes and Input features, Represented as a node, This is represented as the node count index. Then, a multi-period temporal feature fusion mechanism is designed. First, the spatial features from different historical periods are concatenated along the channel dimension. Then, feature transformation is performed through a single-layer sparse attention map convolution. The specific formula is as follows: in, Represented as the fused spatiotemporal feature matrix, Represented as the total number of road network nodes; These represent the spatial features extracted from historical data for the current day, the previous day, and the two days prior, respectively, with each dimension being [dimension number missing]. This is represented as a channel-dimensional concatenation operation; This is represented as a single-layer sparse attention map convolution operation, with the weight matrix dimension adjusted to... This maps a 192-dimensional input to a 48-dimensional output. Represented as a traffic network map structure, the number of historical periods is fixed at 3, representing the current day and the two days prior. This avoids the curse of dimensionality while covering sufficient historical information. By integrating short-term proximity traffic patterns with long-term historical patterns, the complex temporal dependencies of traffic flow are obtained, enabling the algorithm to handle both sudden events and periodic patterns simultaneously, improving short-term and medium-to-long-term prediction accuracy. Secondly, a dynamic task scheduling strategy is constructed. Through a dual-buffered queue architecture and priority allocation strategy, the execution order and resource allocation of prediction tasks are intelligently managed. This allows for real-time response to sudden traffic events while maintaining efficient execution of regular prediction tasks. The strategy establishes two task queues, with a priority threshold for the regular prediction queue. and emergency event queue priority threshold Task priorities are dynamically calculated based on the severity and scope of the event, and resource allocation ratios are determined accordingly. It is directly proportional to the priority, and the specific formula is as follows: in, This represents the proportion of computing resources allocated to the current task; This is represented as a task priority score; This indicates the area of ​​the affected region, expressed in square kilometers. Represented as the total area of ​​the road network; This represents the severity score of the event, with a value ranging from 0 to 1; and It is represented as a weighting coefficient, which is determined through statistical analysis of historical event response effects, and can optimally balance the spatial impact range and the severity of the event. Represented as the index of the number of tasks currently pending. This represents the total number of tasks currently pending. By establishing a dynamic mapping relationship between computing resources and task urgency, the algorithm gains rapid response capabilities under unexpected events. Secondly, it allocates computing resources according to the proportion... Dynamically adjust the computational depth and frequency of the prediction task for For high-priority tasks, a complete deep learning model is used for high-frequency prediction; for For low-priority tasks, reduce the model's prediction frequency. The specific formula is as follows: in, This represents the actual predicted result. This represents the predicted output of a complete deep learning model, which has high computational complexity but optimal accuracy. This represents the prediction output of a lightweight prediction model, indicating that the number of hidden layer neurons is reduced to 24 compared to the full deep learning model, retaining 70% of the prediction capability but improving computational efficiency by 2.5 times. Represented as a prediction result based on historical similarity retrieval, it has the lowest computational complexity and is suitable for routine predictions in non-critical areas. Through a dynamic balance between computational resources and prediction accuracy, the algorithm maintains prediction quality for critical areas even under high load conditions. Finally, it generates short-term (5-30 minutes) and medium-to-long-term (1-24 hours) traffic flow prediction results. First, external environmental factors are embedded as feature vectors, which are then concatenated with spatiotemporal features and used to generate predictions through a two-layer fully connected network. The specific formula is as follows: in, This represents the predicted output of the complete deep learning model. and Represented as a learnable weight matrix, and Represented as a bias vector; This represents the predicted output of the lightweight model, with the number of hidden layer neurons reduced to 24. It retains 70% of the predictive power but improves computational efficiency by 2.5 times; This is represented as the predicted output of the historical similarity model. The cosine similarity function is used. Represented as the first The spatiotemporal characteristics of similar historical periods It is represented as a weighted coefficient based on similarity. For the capacity of the historical sample database, Represented as a historical sample database capacity index; Represents the hyperbolic tangent activation function; Represented as a feature vector of external environmental factors, It is represented as a combination of spatiotemporal characteristics and external factors. Represented as external factor dimensions, it includes 5-dimensional weather information, 1-dimensional holiday markers, and 2-dimensional POI distribution. The three models share the same input features, but by differentiating computational complexity and model capacity, the algorithm can achieve optimal allocation of system resources and maximize overall prediction efficiency while ensuring the prediction quality of key areas.

[0014] Preferably, the decision support and service module automatically generates road network heat maps and trend line charts through multimodal result transformation technology, calls the configurable Drools rule base to transform the predicted data into traffic light timing optimization strategies, and pushes structured prediction results and early warning signals to external traffic management platforms in real time through a standardized RESTful API interface. This enables cross-system collaborative decision-making, avoids delays caused by manual intervention, transforms prediction results into actionable decision-making basis, and strengthens the city's collaborative traffic management capabilities.

[0015] Preferably, the system management and monitoring module provides a user-interactive intelligent control interface, allowing users to customize monitoring rules and configure multi-level approval workflows through drag-and-drop interface, and push interactive intervention pop-ups when anomalies are detected, allowing users to override automatic processing strategies and adjust parameters in real time, thereby achieving user-led full lifecycle management of the system.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. A cross-source feature fusion unit proposes a multi-source heterogeneous data fusion algorithm based on neural networks. First, by constructing a multi-layer backpropagation neural network, heterogeneous information from multiple sources—floating car trajectories, GPS, fixed detectors, urban event semantics, and meteorological monitoring—from the multi-source data standardization processing module is mapped to a unified semantic feature space. This achieves semantic alignment of heterogeneous signals and coupling of spatiotemporal information, avoiding information loss and temporal disconnection problems caused by traditional source-by-source processing. Second, a hybrid optimization mechanism of particle swarm optimization and genetics is used to jointly search network weights and thresholds. Through a strategy that emphasizes population hierarchies, selective retention, and crossover mutation, a balance between global exploration and local development is achieved, reducing the risk of getting trapped in local optima and improving convergence speed. This results in a more robust parameter solution with limited computational overhead. Third, the algorithm highlights the contributions of key hidden units and optimized derived features through importance weighting and normalization processing. It can adaptively identify and amplify the temporal, spatial, and semantic features most sensitive to traffic state prediction, thus generating… The spatiotemporal correlation feature vectors possess high information content and strong noise tolerance, which is beneficial for subsequent prediction models to achieve higher accuracy and robustness. The algorithm sets constraints on parameter range, speed, and position update mechanisms to ensure the numerical stability and interpretability of the model training process, facilitating engineering implementation and optimization. At the same time, the weight ratio determined by the hybrid optimization strategy and experimental cross-validation helps to balance fitting ability and generalization performance, reducing the risk of overfitting. Overall, this fusion algorithm not only improves the ability to characterize complex nonlinear relationships between multi-source data and the quality of feature extraction, but also improves the traffic prediction effect at short and medium time scales. This enables the algorithm to provide more timely and accurate decision support in application scenarios such as congestion warning, signal optimization, and traffic resource scheduling. Furthermore, the modular design and controlled computational complexity of the algorithm facilitate deployment and expansion in actual urban traffic information platforms, supporting online updates and gradual iterations. This allows the entire method to improve prediction performance while taking into account engineering feasibility and promotional value, demonstrating good technological innovation and application benefits.

[0017] 2. The traffic flow prediction and scheduling module proposes a deep learning-based traffic flow prediction algorithm. This algorithm constructs a sparse attention graph convolutional network as the core prediction model to achieve a deep characterization of local and non-local spatial dependencies in the traffic network, thereby enhancing the representation and recognition capabilities of deep learning in traffic flow prediction scenarios. The model adaptively adjusts the weights between nodes based on graph convolution operators with expanded receptive fields, strengthening the capture of complex relationships between distant nodes and improving the detection sensitivity of cascading congestion propagation paths. The algorithm constructs a multi-period temporal feature fusion mechanism, deeply integrating the spatial features of the current day and the previous two days in the channel dimension, and uniformly transforming them through a single-layer sparse attention graph convolution. This allows the algorithm to simultaneously consider short-term proximity dynamics and long-term patterns within the deep learning framework, improving short-term prediction accuracy and medium- to long-term trend stability. To meet real-time requirements, the algorithm designs a dynamic task scheduling strategy, using a double-buffered queue architecture to differentiate between routine prediction tasks and emergency event responses, combined with priority evaluation based on regional impact and event severity. This algorithm achieves on-demand allocation of computing resources, ensuring high-frequency, high-precision prediction output for critical areas during emergencies while maintaining continuous service for routine prediction tasks. It optimizes resource utilization through a three-level execution strategy: using a full deep learning model for high-priority tasks to guarantee accuracy, a lightweight model for medium-priority tasks to balance efficiency, and historical similarity retrieval for low-priority tasks to conserve resources. These three approaches work collaboratively under a unified input to achieve a dynamic balance between system-level prediction quality and computational overhead. External environmental information is embedded into spatiotemporal features and used as deep learning input, enhancing the model's responsiveness to weather and event impacts, thereby improving the robustness and adaptability of prediction results. The overall solution's architecture balances scalability and engineering feasibility, facilitating deployment in large-scale urban road networks. It supports online scheduling and rapid iteration, providing timely and accurate congestion warnings and scheduling suggestions to traffic management departments, improving urban traffic efficiency and emergency response capabilities, demonstrating significant technological advancements and application value. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the structure of the present invention; Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on 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.

[0020] Please see Figure 1This invention provides a method for dynamic traffic flow prediction based on the fusion of spatiotemporal correlation features of multi-source heterogeneous data. The method includes a multi-source heterogeneous data acquisition module, a multi-source data standardization processing module, a heterogeneous data feature fusion module, a traffic flow prediction and scheduling module, a decision support and service module, and a system management and monitoring module. The multi-source heterogeneous data acquisition module is used to collect raw traffic data in real time from heterogeneous data infrastructure including fixed traffic detectors, floating car trajectory data streams, urban event semantic data sources, GPS data, and meteorological monitoring networks. The multi-source data standardization processing module is used to clean, denoise, imput missing values, and standardize the collected raw data to eliminate noise and scale differences. The heterogeneous data feature fusion module includes a feature layering extraction unit and a cross-source feature fusion unit. The feature layering extraction unit is used to extract the spatiotemporal correlation features of traffic flow in layers. The system integrates feature and semantic association features to obtain multi-dimensional spatiotemporal patterns. A cross-source feature fusion unit proposes a neural network-based multi-source heterogeneous data fusion algorithm to calculate cross-source feature similarity and achieve mapping alignment, deeply fusing heterogeneous features to generate high-value spatiotemporal associated feature vectors. A traffic flow prediction and scheduling module proposes a deep learning-based traffic flow prediction algorithm for short-term and medium-to-long-term dynamic prediction of fused features, and manages the real-time scheduling, priority allocation, and resource optimization of prediction tasks. A decision support and service module transforms prediction results into visual reports, traffic flow management strategies, and early warning signals, providing actionable decision-making basis and external service interfaces for urban traffic management. A system management and monitoring module provides users with an interactive interface for real-time monitoring of data flow, model performance, and system resources, and automatically detects anomalies, triggers alarms, and performs maintenance operations.

[0021] See Figure 1 Furthermore, the multi-source heterogeneous data acquisition module integrates multiple protocol access interfaces, including MQTT protocol to interface with fixed traffic detectors, HTTP API interface to call urban event semantic data sources, Kafka stream processing interface to receive floating car trajectory and GPS data streams, and real-time data from meteorological monitoring networks. This enables full-dimensional coverage acquisition of multi-source heterogeneous traffic data. Data sources include road network sensors for traffic flow and speed, mobile terminals for taxi and ride-hailing vehicle trajectories, unstructured event text from social media and traffic management reports, satellite positioning data, and meteorological parameters such as temperature and rainfall, providing highly timely input for subsequent processing.

[0022] See Figure 1Furthermore, the multi-source data standardization processing module performs data cleaning by applying an outlier filtering threshold based on historical statistics, performs noise reduction using wavelet transform noise reduction technology, performs spatiotemporal context-aware missing value filling by combining KNN interpolation with road network topology, and performs multi-scale data standardization using Z-score normalization and quantile transformation. This achieves noise suppression, missing value repair, and dimensional unification of the original multi-source heterogeneous data, eliminates the differences in accuracy, inconsistent sampling frequencies, and unit scale conflicts among multi-source data, and outputs a structured, highly consistent, high-quality dataset, laying a robust foundation for subsequent feature extraction.

[0023] See Figure 1 Furthermore, the heterogeneous data feature fusion module includes a feature layering extraction unit. This feature layering extraction unit integrates the Apache Flink built-in time series processing library, the PyTorch Geometric spatial relationship modeling module, and the Hugging Face Transformers text parsing pipeline to construct a layered parallel computing architecture. This enables automated feature extraction of multi-dimensional spatiotemporal patterns of traffic flow, generating an independent feature vector set that includes historical traffic evolution patterns, road network spatial propagation effects, and the impact of external event disturbances. This fully leverages the inherent complementarity between multi-source heterogeneous data.

[0024] See Figure 1 Furthermore, the heterogeneous data feature fusion module includes a cross-source feature fusion unit, which proposes a multi-source heterogeneous data fusion algorithm based on neural networks. By designing a cross-source alignment mechanism based on neural networks, the backpropagation neural network establishes a basic mapping model, projects data from different sources onto a unified feature space, and then combines particle swarm optimization technology to form a hybrid fusion strategy. This achieves deep fusion generation of high-value spatiotemporal correlation feature vectors, improves the spatiotemporal consistency and prediction reliability of cross-source traffic data, and provides strong representational input for subsequent traffic flow prediction.

[0025] See Figure 1 Furthermore, the specific details of the neural network-based multi-source heterogeneous data fusion algorithm are as follows: First, a three-layer backpropagation neural network is constructed as the basic mapping model for multi-source heterogeneous data fusion. The model projects traffic data from the multi-source data standardization processing module onto a unified feature space. In the constructed neural network, the input layer receives standardized multi-source traffic features, the hidden layer performs nonlinear transformations, and the output layer generates the fused traffic state estimate, as expressed in the following formula: in, The input feature vector is represented by a dimension of 5 and includes floating car trajectory data, GPS data, fixed traffic detector data, urban event semantic data, and meteorological monitoring network data. This is represented as the weight matrix from the input layer to the hidden layer, with dimensions [5,4]. It is represented as a hidden layer bias vector with dimensions [5,1]; It is represented as the weight matrix from the hidden layer to the output layer, with dimensions [1,5]. This represents the output layer bias vector, with dimensions [1,1]. It is represented as an activation function, using the hyperbolic tangent function, with a range of [-1, 1], which can effectively handle nonlinear mappings and avoid the gradient vanishing problem; The output of the neural network represents the fused traffic state estimate. The constructed neural network uses a 5-5-1 network structure: the input layer contains 5 neurons corresponding to 5 input feature parameters, the hidden layer contains 5 neurons responsible for nonlinear feature transformation and pattern extraction, and the output layer contains 1 neuron producing a single fused output. This approach ensures model expressiveness while avoiding overfitting. Then, a particle swarm optimization mechanism combined with genetic optimization is designed to improve the training effect of the neural network. First, a fitness function is defined to evaluate network performance. Then, the population is divided into three subpopulations based on fitness values, and different optimization strategies are applied to each subpopulation to ensure a balance between global exploration and local exploitation of the search space. The specific formula is as follows: in, This is represented as an individual fitness value; a higher value indicates better individual performance. This is represented as the fitness adjustment factor, with a value of 1.5. The value has been verified through numerous experiments to produce fitness values ​​with strong discriminative power within different error ranges; This represents the total number of training samples, with 100-200 samples selected to ensure sufficient training. Represented as the first The network prediction output for each sample; Indicates the first For each sample's actual traffic state value, the fitness function is designed as the reciprocal of the sum of squared errors, so that individuals with smaller errors receive higher fitness. The coefficient ensures that the fitness value remains stable within a reasonable range, avoiding computational accuracy issues caused by values ​​that are too large or too small, and the population size... Setting the threshold to 100 balances computational efficiency and search quality, and is the experimentally verified optimal size. By dynamically adjusting the neural network weights and thresholds, the problems of traditional backpropagation neural networks easily getting trapped in local optima and slow convergence speed are solved. This enables the neural network to adaptively learn complex nonlinear relationships between multi-source data, improving the accuracy and stability of feature fusion. Secondly, a population hierarchical optimization strategy is constructed. By sorting the particle swarm according to its fitness value from high to low and dividing it into three subpopulations, customized optimization operations are applied to different subpopulations. High-fitness individuals are retained, medium-fitness individuals generate better offspring through crossover, and low-fitness individuals explore new solution spaces through mutation. The specific formula is expressed as follows: in, This represents an optimized particle velocity vector, with an initial velocity range of [-1, 1], and a maximum velocity limit. minimum speed This prevents particles from jumping out of the optimal region too quickly; This represents the optimized particle position vector, corresponding to the encoding of neural network weights and biases, limiting the maximum position. minimum position This ensures the range of feasible solutions. It is represented as inertia weight, with a value range of [0.4, 0.9]. It adopts a linear decreasing strategy from 0.9 to 0.4. In the early stage, a large inertia is conducive to global exploration, and in the later stage, a small inertia is conducive to local development. This is represented as an individual learning factor, set to 1.49, which controls the intensity of particle movement towards the individual's historical best position; Represented as the global learning factor, set to 1.69, it controls the intensity of particle movement towards the global optimal position of the swarm; and It is represented as a random number uniformly distributed in the interval [0,1], and randomness is introduced to enhance the diversity of algorithms; This represents the individual's historical best position, recording the corresponding particle's historical best solution; This represents the global optimal position of the population and records the best solution throughout the entire history of the population. Represented as the current iteration number, maximum iteration number Experiments revealed that, while ensuring convergence accuracy, the computational complexity was controlled. The particle velocity and position update mechanism achieved a balance between global search and local exploitation, effectively avoiding premature mastery. Then, the optimal parameters of the neural network were determined through an iterative optimization process, and the individual fitness values ​​were adjusted accordingly. As an evaluation criterion, the first The particle position vector of the next iteration This is transformed into a neural network weight and threshold matrix, enabling the algorithm to adaptively learn complex mapping relationships between multiple data sources, overcoming the limitations of a single data source. After each iteration, based on the fitness value... The optimal individual is selected, and the global optimal position is determined when the convergence condition and the maximum number of iterations are met. The corresponding parameters are assigned to the neural network to construct the final fusion model, as shown in the following formula: in, This is represented as the optimized input-to-hidden-layer weight matrix, with dimensions [5,4] and element values ​​ranging from [-5,5]. This is represented as the optimized hidden layer bias vector, with dimensions [5,1] and element values ​​ranging from [-3,3]. This is represented as the optimized weight matrix from the hidden layer to the output layer, with dimensions [1,5] and element values ​​ranging from [-5,5]. This is represented as the optimized output layer bias vector, with dimensions [1,1] and values ​​ranging from [-2,2]. These represent the number of neurons in the input layer, hidden layer, and output layer, respectively, corresponding to the 5-5-1 network structure; This represents the number of weight parameters from the input layer to the hidden layer; This is expressed as the number of hidden layer bias parameters; This indicates the end position of the segment for the first three parameters; Represented as a matrix reconstruction function, it transforms a vector into a matrix of a specified dimension. By converting the optimized particle position vector into network parameters, a high-precision multi-source data fusion model is established. Finally, high-value spatiotemporal correlation feature vectors are generated. Utilizing the nonlinear mapping capability of the constructed neural network, heterogeneous data is projected onto a unified semantic space. A weighted fusion mechanism is used to highlight the contributions of key features. The specific formula is expressed as follows: in, It is represented as a fused high-value spatiotemporal correlation feature vector with a dimension of 7, containing the activation values ​​of 5 hidden layer neurons and 2 features derived from the optimization process; It is represented as a feature normalization function, which uses Min-Max standardization to map the feature values ​​to the [0,1] interval; Represented as the first The importance weights of each hidden neuron are calculated using the Softmax function to ensure that important features receive higher weights. Represented as the index of the number of hidden neurons; and These represent the optimized inputs to the first... The weight vectors and biases of each hidden neuron; This is represented as a vector concatenation operation; It is represented as a feature weight vector of the optimization process, which is determined by cross-validation to balance the convergence characteristics of the algorithm and the data fitting characteristics. Represented as the Hadamard product; It is represented as the deviation vector between the particle's current position and the global optimal position, with a dimension of 2. By applying the optimized neural network parameters to multi-source traffic data, time series features, spatial topological features, and semantic association features are deeply integrated to extract feature representations containing spatiotemporal patterns, overcoming the limitations of a single data source and improving the prediction accuracy and robustness of subsequent modules.

[0026] See Figure 1 Furthermore, the traffic flow prediction and scheduling module proposes a traffic flow prediction algorithm based on deep learning. By constructing a spatiotemporal deep learning model and a dynamic task scheduling strategy, it integrates short-term proximity and long-term traffic patterns and external environmental factors to predict traffic flow. It adopts a dual-buffer queue architecture and a priority allocation strategy to differentiate resource allocation for routine prediction tasks and emergency event responses. It intelligently manages the triggering timing of prediction tasks, the allocation of computing resources, and emergency event responses to achieve dynamic prediction of short-term and medium-to-long-term traffic flow status.

[0027] See Figure 1 Furthermore, the deep learning-based traffic flow prediction algorithm is as follows: First, a sparse attention graph convolutional network is constructed as the core model for dynamic traffic flow prediction. By designing a graph convolution operator that expands the receptive field, local and non-local spatial dependencies in the traffic network are obtained, enabling the model to simultaneously perceive the complex spatial relationships between neighboring road segments and distant nodes. The traffic network represents a graph structure, nodes represent road segments and intersections, and edges represent spatial connections. The weights between different nodes are dynamically adjusted through a sparse attention mechanism, allowing the model to adaptively learn the strength of spatial dependencies. The specific formula is expressed as follows: in, Represented as nodes The output feature vector; Represented as the ReLU activation function; Represented as nodes of Neighbor set; Expressed as the expansion rate, it has been experimentally verified that it can achieve the best balance between computational efficiency and receptive field range; Represented as the neighbor order, with a value range of 1. It covers spatial dependencies from local to global. Represented as a similarity function, It is represented as a two-layer fully connected network that maps input features to a 32-dimensional latent space; Represented as a learnable weight matrix, Represented as the real number field; and Represented as nodes and Input features, Represented as a node, This is represented as the node count index. Then, a multi-period temporal feature fusion mechanism is designed. First, the spatial features from different historical periods are concatenated along the channel dimension. Then, feature transformation is performed through a single-layer sparse attention map convolution. The specific formula is as follows: in, Represented as the fused spatiotemporal feature matrix, Represented as the total number of road network nodes; These represent the spatial features extracted from historical data for the current day, the previous day, and the two days prior, respectively, with each dimension being [dimension number missing]. This is represented as a channel-dimensional concatenation operation; This is represented as a single-layer sparse attention map convolution operation, with the weight matrix dimension adjusted to... This maps a 192-dimensional input to a 48-dimensional output. Represented as a traffic network map structure, the number of historical periods is fixed at 3, representing the current day and the two days prior. This avoids the curse of dimensionality while covering sufficient historical information. By integrating short-term proximity traffic patterns with long-term historical patterns, the complex temporal dependencies of traffic flow are obtained, enabling the algorithm to handle both sudden events and periodic patterns simultaneously, improving short-term and medium-to-long-term prediction accuracy. Secondly, a dynamic task scheduling strategy is constructed. Through a dual-buffered queue architecture and priority allocation strategy, the execution order and resource allocation of prediction tasks are intelligently managed. This allows for real-time response to sudden traffic events while maintaining efficient execution of regular prediction tasks. The strategy establishes two task queues, with a priority threshold for the regular prediction queue. and emergency event queue priority threshold Task priorities are dynamically calculated based on the severity and scope of the event, and resource allocation ratios are determined accordingly. It is directly proportional to the priority, and the specific formula is as follows: in, This represents the proportion of computing resources allocated to the current task; This is represented as a task priority score; This indicates the area of ​​the affected region, expressed in square kilometers. Represented as the total area of ​​the road network; This represents the severity score of the event, with a value ranging from 0 to 1; and It is represented as a weighting coefficient, which is determined through statistical analysis of historical event response effects, and can optimally balance the spatial impact range and the severity of the event. Represented as the index of the number of tasks currently pending. This represents the total number of tasks currently pending. By establishing a dynamic mapping relationship between computing resources and task urgency, the algorithm gains rapid response capabilities under unexpected events. Secondly, it allocates computing resources according to the proportion... Dynamically adjust the computational depth and frequency of the prediction task for For high-priority tasks, a complete deep learning model is used for high-frequency prediction; for For low-priority tasks, reduce the model's prediction frequency. The specific formula is as follows: in, This represents the actual predicted result. This represents the predicted output of a complete deep learning model, which has high computational complexity but optimal accuracy. This represents the prediction output of a lightweight prediction model, indicating that the number of hidden layer neurons is reduced to 24 compared to the full deep learning model, retaining 70% of the prediction capability but improving computational efficiency by 2.5 times. Represented as a prediction result based on historical similarity retrieval, it has the lowest computational complexity and is suitable for routine predictions in non-critical areas. Through a dynamic balance between computational resources and prediction accuracy, the algorithm maintains prediction quality for critical areas even under high load conditions. Finally, it generates short-term (5-30 minutes) and medium-to-long-term (1-24 hours) traffic flow prediction results. First, external environmental factors are embedded as feature vectors, which are then concatenated with spatiotemporal features and used to generate predictions through a two-layer fully connected network. The specific formula is as follows: in, This represents the predicted output of the complete deep learning model. and Represented as a learnable weight matrix, and Represented as a bias vector; This represents the predicted output of the lightweight model, with the number of hidden layer neurons reduced to 24. It retains 70% of the predictive power but improves computational efficiency by 2.5 times; This is represented as the predicted output of the historical similarity model. The cosine similarity function is used. Represented as the first The spatiotemporal characteristics of similar historical periods It is represented as a weighted coefficient based on similarity. For the capacity of the historical sample database, Represented as a historical sample database capacity index; Represents the hyperbolic tangent activation function; Represented as a feature vector of external environmental factors, It is represented as a combination of spatiotemporal characteristics and external factors. Represented as external factor dimensions, it includes 5-dimensional weather information, 1-dimensional holiday markers, and 2-dimensional POI distribution. The three models share the same input features, but by differentiating computational complexity and model capacity, the algorithm can achieve optimal allocation of system resources and maximize overall prediction efficiency while ensuring the prediction quality of key areas.

[0028] See Figure 1 Furthermore, the decision support and service module automatically generates road network heat maps and trend line charts through multimodal result transformation technology, and calls the configurable Drools rule base to transform the predicted data into traffic light timing optimization strategies. It also pushes structured prediction results and early warning signals to external traffic management platforms in real time through a standardized RESTful API interface, realizing cross-system collaborative decision-making, avoiding delays caused by manual intervention, and transforming prediction results into actionable decision-making basis, thereby strengthening the city's collaborative traffic management capabilities.

[0029] See Figure 1 Furthermore, the system management and monitoring module provides a user-interactive intelligent control interface, allowing users to customize monitoring rules and configure multi-level approval workflows through drag-and-drop interface, and push interactive intervention pop-ups when anomalies are detected, allowing users to override automatic processing strategies and adjust parameters in real time, thereby achieving user-led full lifecycle management of the system.

[0030] In practical use, firstly, the multi-source heterogeneous data acquisition module is used to collect raw traffic data in real time from heterogeneous data infrastructure including fixed traffic detectors, floating car trajectory data streams, urban event semantic data sources, GPS data, and meteorological monitoring networks. Secondly, the multi-source data standardization processing module is used to clean, denoise, imput missing values, and standardize the collected raw data to eliminate noise and scale differences. Then, the heterogeneous data feature fusion module includes a feature hierarchical extraction unit and a cross-source feature fusion unit. The feature hierarchical extraction unit is used to extract the spatiotemporal features and semantic association features of traffic flow hierarchically to obtain multi-dimensional spatiotemporal patterns. The cross-source feature fusion unit proposes a multi-source heterogeneous data fusion algorithm based on neural networks for... The system calculates cross-source feature similarity and performs mapping alignment, deeply fusing heterogeneous features to generate high-value spatiotemporal correlation feature vectors. Secondly, the traffic flow prediction and scheduling module proposes a deep learning-based traffic flow prediction algorithm for short-term and medium-to-long-term dynamic prediction of the fused features, and manages the real-time scheduling, priority allocation, and resource optimization of prediction tasks. Then, the decision support and service module transforms the prediction results into visual reports, traffic flow management strategies, and early warning signals, providing actionable decision-making basis and external service interfaces for urban traffic management. Finally, the system management and monitoring module provides users with an interactive interface for real-time monitoring of data flow, model performance, and system resources, and automatically detects anomalies, triggers alarms, and performs maintenance operations.

[0031] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for dynamic traffic flow prediction based on the fusion of spatiotemporal correlation features of multi-source heterogeneous data, comprising a multi-source heterogeneous data acquisition module, a multi-source data standardization processing module, a heterogeneous data feature fusion module, a traffic flow prediction and scheduling module, a decision support and service module, and a system management and monitoring module, characterized in that: The multi-source heterogeneous data acquisition module is used to collect raw traffic data in real time from heterogeneous data infrastructure including fixed traffic detectors, floating car trajectory data streams, urban event semantic data sources, GPS data, and meteorological monitoring networks. The multi-source data standardization processing module is used to clean, denoise, imput missing values, and standardize the collected raw data to eliminate noise and scale differences. The heterogeneous data feature fusion module includes a feature layering extraction unit and a cross-source feature fusion unit. The feature layering extraction unit is used to extract the spatiotemporal features and semantic association features of traffic flow in layers to obtain multi-dimensional spatiotemporal patterns. The cross-source feature fusion unit proposes a neural network-based multi-source heterogeneous data fusion algorithm to calculate cross-source feature similarity and achieve mapping alignment, deeply fusing heterogeneous features to generate high-value spatiotemporal association feature vectors. The traffic flow prediction and scheduling module proposes a deep learning-based traffic flow prediction algorithm to perform short-term and medium-to-long-term dynamic predictions on the fused features and manage the real-time scheduling, priority allocation, and resource optimization of prediction tasks. The decision support and service module is used to transform the prediction results into visual reports, traffic flow management strategies, and early warning signals, providing operable decision-making basis and external service interfaces for urban traffic management. The system management and monitoring module provides users with an interactive interface for real-time monitoring of data flow, model performance, and system resources, and automatically detects anomalies, triggers alarms, and performs maintenance operations.

2. The traffic flow dynamic prediction method based on the fusion of spatiotemporal correlation features of multi-source heterogeneous data according to claim 1, characterized in that: The multi-source heterogeneous data acquisition module integrates multiple protocol access interfaces to achieve full-dimensional coverage of multi-source heterogeneous traffic data. The data sources include heterogeneous data from road network sensors, mobile terminals, unstructured event text, satellite positioning, and meteorological parameters, providing highly timely multi-source heterogeneous data input for subsequent processing.

3. The traffic flow dynamic prediction method based on the fusion of spatiotemporal correlation features of multi-source heterogeneous data according to claim 1, characterized in that: The multi-source data standardization processing module performs outlier filtering thresholding, wavelet transform denoising, missing value imputation, and multi-scale standardization strategies on the original multi-source heterogeneous data. This achieves noise suppression, missing value repair, and dimensional unification of the original data, eliminates differences in sensor accuracy, inconsistent sampling frequencies, and unit scale conflicts, and outputs a structured, highly consistent, high-quality multi-source heterogeneous dataset, laying a robust foundation for subsequent feature extraction.

4. The traffic flow dynamic prediction method based on the fusion of spatiotemporal correlation features of multi-source heterogeneous data according to claim 1, characterized in that: The heterogeneous data feature fusion module includes a feature hierarchical extraction unit and a cross-source feature fusion unit. The feature hierarchical extraction unit uses a hierarchical parallel computing architecture to achieve refined capture of multi-dimensional spatiotemporal patterns of traffic flow, generating an independent feature vector set that includes historical traffic evolution patterns, road network spatial propagation effects, and the impact of external event disturbances, fully exploring the inherent complementarity of multi-source heterogeneous data. The cross-source feature fusion unit proposes a multi-source heterogeneous data fusion algorithm based on neural networks. By designing a cross-source alignment mechanism based on neural networks, it achieves deep fusion generation of high-value spatiotemporal correlated feature vectors, improving feature information density and noise resistance, and providing strong representational input for subsequent traffic flow prediction.

5. The traffic flow dynamic prediction method based on the fusion of spatiotemporal correlation features of multi-source heterogeneous data according to claim 4, characterized in that, First, a three-layer backpropagation neural network is constructed as the basic mapping model for multi-source heterogeneous data fusion. This model projects traffic data from the multi-source data standardization processing module onto a unified feature space. In the constructed neural network, the input layer receives standardized multi-source traffic features, the hidden layer performs nonlinear transformations, and the output layer generates the fused traffic state estimate. The specific formula is as follows: in, The input feature vector is represented by a dimension of 5 and includes floating car trajectory data, GPS data, fixed traffic detector data, urban event semantic data, and meteorological monitoring network data. This is represented as the weight matrix from the input layer to the hidden layer, with dimensions [5,4]. It is represented as a hidden layer bias vector with dimensions [5,1]; It is represented as the weight matrix from the hidden layer to the output layer, with dimensions [1,5]. This represents the output layer bias vector, with dimensions [1,1]. It is represented as an activation function, using the hyperbolic tangent function, with a range of [-1, 1], which can effectively handle nonlinear mappings and avoid the gradient vanishing problem; The output of the neural network represents the fused traffic state estimate. The constructed neural network uses a 5-5-1 network structure: the input layer contains 5 neurons corresponding to 5 input feature parameters, the hidden layer contains 5 neurons responsible for nonlinear feature transformation and pattern extraction, and the output layer contains 1 neuron producing a single fused output. This approach ensures both model expressive power and avoids overfitting. Then, a particle swarm optimization mechanism combined with genetic optimization is designed to improve the training effect of the neural network. First, a fitness function is defined to evaluate network performance. Then, the population is divided into three subpopulations based on fitness values, and different optimization strategies are applied to each subpopulation to ensure a balance between global exploration and local exploitation of the search space. The specific formula is as follows: in, This is represented as an individual fitness value; a higher value indicates better individual performance. This is represented as the fitness adjustment factor, with a value of 1.

5. The value has been verified through numerous experiments to produce fitness values ​​with strong discriminative power within different error ranges; This represents the total number of training samples, with 100-200 samples selected to ensure sufficient training. Represented as the first The network prediction output for each sample; Indicates the first For each sample's actual traffic state value, the fitness function is designed as the reciprocal of the sum of squared errors, so that individuals with smaller errors receive higher fitness. The coefficient ensures that the fitness value remains stable within a reasonable range, avoiding computational accuracy issues caused by values ​​that are too large or too small, and the population size... Setting the threshold to 100, to balance computational efficiency and search quality, is the experimentally verified optimal size. By dynamically adjusting the neural network weights and thresholds, the problems of traditional backpropagation neural networks easily getting trapped in local optima and having slow convergence speed are solved. This allows the neural network to adaptively learn complex nonlinear relationships between multi-source data, improving the accuracy and stability of feature fusion. Secondly, a population hierarchical optimization strategy is constructed. By sorting the particle swarm according to its fitness value from high to low and dividing it into three subpopulations, customized optimization operations are applied to different subpopulations. High-fitness individuals are retained, medium-fitness individuals generate better offspring through crossover, and low-fitness individuals explore new solution spaces through mutation. The specific formula is expressed as: in, This represents an optimized particle velocity vector, with an initial velocity range of [-1, 1] and a maximum velocity limit. minimum speed This prevents particles from jumping out of the optimal region too quickly; This represents the optimized particle position vector, corresponding to the encoding of neural network weights and biases, limiting the maximum position. minimum position This ensures the range of feasible solutions. It is represented as inertia weight, with a value range of [0.4, 0.9]. It adopts a linear decreasing strategy from 0.9 to 0.

4. In the early stage, a large inertia is conducive to global exploration, and in the later stage, a small inertia is conducive to local development. This is represented as an individual learning factor, set to 1.49, which controls the intensity of particle movement towards the individual's historical best position; Represented as the global learning factor, set to 1.69, it controls the intensity of particle movement towards the global optimal position of the swarm; and It is represented as a random number uniformly distributed in the interval [0,1], and randomness is introduced to enhance the diversity of algorithms; This represents the individual's historical best position, recording the corresponding particle's historical best solution; This represents the global optimal position of the population and records the best solution throughout the entire history of the population. Represented as the current iteration number, maximum iteration number Experiments revealed that, while ensuring convergence accuracy, the computational complexity was controlled. The particle velocity and position update mechanism achieved a balance between global search and local exploitation, effectively avoiding premature mastery. Then, the optimal parameters of the neural network were determined through an iterative optimization process, and the individual fitness values ​​were adjusted accordingly. As an evaluation criterion, the first The particle position vector of the next iteration This is transformed into a neural network weight and threshold matrix, enabling the algorithm to adaptively learn complex mapping relationships between multiple data sources, overcoming the limitations of a single data source. After each iteration, based on the fitness value... The optimal individual is selected, and the global optimal position is determined when the convergence condition and the maximum number of iterations are met. The corresponding parameters are assigned to the neural network to construct the final fusion model, as shown in the following formula: in, This is represented as the optimized input-to-hidden-layer weight matrix, with dimensions [5,4] and element values ​​ranging from [-5,5]. This is represented as the optimized hidden layer bias vector, with dimensions [5,1] and element values ​​ranging from [-3,3]. This is represented as the optimized weight matrix from the hidden layer to the output layer, with dimensions [1,5] and element values ​​ranging from [-5,5]. This is represented as the optimized output layer bias vector, with dimensions [1,1] and values ​​ranging from [-2,2]. These represent the number of neurons in the input layer, hidden layer, and output layer, respectively, corresponding to the 5-5-1 network structure; This represents the number of weight parameters from the input layer to the hidden layer; This is expressed as the number of hidden layer bias parameters; This indicates the end position of the segment for the first three parameters; Represented as a matrix reconstruction function, it transforms a vector into a matrix of a specified dimension. By converting the optimized particle position vector into network parameters, a high-precision multi-source data fusion model is established. Finally, high-value spatiotemporal correlation feature vectors are generated. Utilizing the nonlinear mapping capability of the constructed neural network, heterogeneous data is projected onto a unified semantic space. A weighted fusion mechanism is used to highlight the contributions of key features. The specific formula is expressed as follows: in, It is represented as a fused high-value spatiotemporal correlation feature vector with a dimension of 7, containing the activation values ​​of 5 hidden layer neurons and 2 features derived from the optimization process; It is represented as a feature normalization function, which uses Min-Max standardization to map the feature values ​​to the [0,1] interval; Represented as the first The importance weights of each hidden neuron are calculated using the Softmax function to ensure that important features receive higher weights. Represented as the index of the number of hidden neurons; and These represent the optimized inputs to the first... The weight vector and bias value of each hidden neuron; This is represented as a vector concatenation operation; It is represented as a feature weight vector of the optimization process, which is determined by cross-validation to balance the convergence characteristics of the algorithm and the data fitting characteristics. Represented as the Hadamard product; It is represented as the deviation vector between the particle's current position and the global optimal position, with a dimension of 2. By applying the optimized neural network parameters to multi-source traffic data, time series features, spatial topological features, and semantic association features are deeply integrated to extract feature representations containing spatiotemporal patterns, overcoming the limitations of a single data source and improving the prediction accuracy and robustness of subsequent modules.

6. The traffic flow dynamic prediction method based on the fusion of spatiotemporal correlation features of multi-source heterogeneous data according to claim 1, characterized in that: The traffic flow prediction and scheduling module proposes a traffic flow prediction algorithm based on deep learning. By constructing a spatiotemporal deep learning model and a dynamic task scheduling strategy, it intelligently manages the triggering timing of prediction tasks, the allocation of computing resources, and emergency event responses, thereby achieving dynamic prediction of short-term and medium-to-long-term traffic flow conditions.

7. The traffic flow dynamic prediction method based on the fusion of spatiotemporal correlation features of multi-source heterogeneous data according to claim 6, characterized in that, First, a sparse attention graph convolutional network is constructed as the core model for traffic flow dynamic prediction. By designing a graph convolution operator that expands the receptive field, local and non-local spatial dependencies in the traffic network are obtained, enabling the model to simultaneously perceive the complex spatial relationships between neighboring road segments and distant nodes. The traffic network represents a graph structure, nodes represent road segments and intersections, and edges represent spatial connections. The weights between different nodes are dynamically adjusted through a sparse attention mechanism, allowing the model to adaptively learn the strength of spatial dependencies. The specific formula is expressed as follows: in, Represented as nodes The output feature vector; Represented as the ReLU activation function; Represented as nodes of Neighbor set; Expressed as the expansion rate, it has been experimentally verified that it can achieve the best balance between computational efficiency and receptive field range; Represented as the neighbor order, with a value range of 1. It covers spatial dependencies from local to global. Represented as a similarity function, It is represented as a two-layer fully connected network that maps input features to a 32-dimensional latent space; Represented as a learnable weight matrix, Represented as the real number field; and Represented as nodes and Input features, Represented as a node, This is represented as the node count index; then, a multi-period temporal feature fusion mechanism is designed. First, the spatial features of different historical periods are concatenated along the channel dimension, and then feature transformation is performed through a single-layer sparse attention map convolution. The specific formula is as follows: in, Represented as the fused spatiotemporal feature matrix, Represented as the total number of road network nodes; These represent the spatial features extracted from historical data for the current day, the previous day, and the two days prior, respectively, with each dimension being [dimension number missing]. This is represented as a channel-dimensional concatenation operation; This is represented as a single-layer sparse attention map convolution operation, with the weight matrix dimension adjusted to... This maps a 192-dimensional input to a 48-dimensional output. Represented as a traffic network map structure, the number of historical periods is fixed at 3, representing the current day and the two days prior. This avoids the curse of dimensionality while covering sufficient historical information. By integrating short-term proximity traffic patterns with long-term historical patterns, the complex temporal dependencies of traffic flow are obtained, enabling the algorithm to handle both sudden events and periodic patterns simultaneously, improving short-term and medium-to-long-term prediction accuracy. Secondly, a dynamic task scheduling strategy is constructed. Through a dual-buffered queue architecture and priority allocation strategy, the execution order and resource allocation of prediction tasks are intelligently managed. This allows for real-time response to sudden traffic events while maintaining efficient execution of regular prediction tasks. The strategy establishes two task queues, with a priority threshold for the regular prediction queue. and emergency event queue priority threshold Task priorities are dynamically calculated based on the severity and scope of the event, and resource allocation ratios are determined accordingly. It is directly proportional to the priority, and the specific formula is as follows: in, This represents the proportion of computing resources allocated to the current task; This is represented as a task priority score; This indicates the area of ​​the affected region, expressed in square kilometers. Represented as the total area of ​​the road network; This represents a severity score for the event, with a value ranging from 0 to 1. and It is represented as a weighting coefficient, which is determined through statistical analysis of historical event response effects, and can optimally balance the spatial impact range and the severity of the event. Represented as the index of the number of tasks currently pending. This represents the total number of tasks currently pending. By establishing a dynamic mapping relationship between computing resources and task urgency, the algorithm gains rapid response capabilities under unexpected events. Secondly, it allocates computing resources according to the proportion... Dynamically adjust the computational depth and frequency of the prediction task for For high-priority tasks, a complete deep learning model is used for high-frequency prediction; for For low-priority tasks, reduce the model's prediction frequency. The specific formula is as follows: in, This represents the actual predicted result. This represents the predicted output of a complete deep learning model, which has high computational complexity but optimal accuracy. This represents the prediction output of a lightweight prediction model, indicating that the number of hidden layer neurons is reduced to 24 compared to the full deep learning model, retaining 70% of the prediction capability but improving computational efficiency by 2.5 times. Represented as a prediction result based on historical similarity retrieval, it has the lowest computational complexity and is suitable for routine predictions in non-critical areas. Through a dynamic balance between computational resources and prediction accuracy, the algorithm maintains prediction quality for critical areas even under high load conditions. Finally, it generates short-term (5-30 minutes) and medium-to-long-term (1-24 hours) traffic flow prediction results. First, external environmental factors are embedded as feature vectors, which are then concatenated with spatiotemporal features and used to generate predictions through a two-layer fully connected network. The specific formula is as follows: in, This represents the predicted output of the complete deep learning model. and Represented as a learnable weight matrix, and Represented as a bias vector; This represents the predicted output of the lightweight model, with the number of hidden layer neurons reduced to 24. It retains 70% of the predictive power but improves computational efficiency by 2.5 times; This is represented as the predicted output of the historical similarity model. The cosine similarity function is used. Represented as the first The spatiotemporal characteristics of similar historical periods It is represented as a weighted coefficient based on similarity. For the capacity of the historical sample database, Represented as a historical sample database capacity index; Represents the hyperbolic tangent activation function; Represented as a feature vector of external environmental factors, It is represented as a combination of spatiotemporal characteristics and external factors. Represented as external factor dimensions, it includes 5-dimensional weather information, 1-dimensional holiday markers, and 2-dimensional POI distribution. The three models share the same input features, but by differentiating computational complexity and model capacity, the algorithm can achieve optimal allocation of system resources and maximize overall prediction efficiency while ensuring the prediction quality of key areas.

8. The traffic flow dynamic prediction method based on the fusion of spatiotemporal correlation features of multi-source heterogeneous data according to claim 1, characterized in that: The decision support and service module uses multimodal result transformation technology to convert the predicted scheduling results into visual reports, traffic flow management strategies, and early warning signals, realizing the transformation of predicted results into actionable decision-making basis. It also outputs real-time prediction services through external interfaces, enhancing the city's traffic collaborative management capabilities.

9. The traffic flow dynamic prediction method based on the fusion of spatiotemporal correlation features of multi-source heterogeneous data according to claim 1, characterized in that: The system management and monitoring module provides a user-interactive intelligent control interface, supporting users to customize monitoring rules, approve key operation and maintenance operations, and intervene in anomaly handling processes in real time, thereby achieving user-led full lifecycle management of the system.

Images