Traffic flow prediction optimization method based on intelligent optimization algorithm

Abstract

The invention relates to a traffic flow prediction optimization method based on an intelligent optimization algorithm, which is applied to the technical field of traffic flow prediction. The system comprises a multi-modal data acquisition and fusion module, a hybrid deep learning prediction module, an intelligent parameter optimization module, an integrated learning module, a multi-scale error correction module, a prediction result analysis and visualization module, a distributed learning and privacy protection module and the like. Wherein the hybrid deep learning prediction module adopts an attention mechanism to fuse models such as CNN, Transform and LSTM, and introduces an adaptive error correction mechanism, so that the prediction precision is remarkably improved; the intelligent parameter optimization module accelerates the model search efficiency by using technologies such as quantum computing and swarm intelligence optimization; the integrated learning module and the adversarial training strategy enhance the robustness and generalization of the model; and the multi-scale error correction module adopts a time sequence decomposition and self-adaptive online learning method to improve the accuracy of long-term and short-term prediction.

Description

Technical Field

[0001] The present invention relates to a traffic vehicle flow prediction optimization method based on an intelligent optimization algorithm, and in particular to a traffic vehicle flow prediction optimization method based on an intelligent optimization algorithm applied in the technical field of vehicle flow prediction. Background Art

[0002] With the acceleration of urbanization and the rapid growth of motor vehicle ownership, urban traffic congestion is becoming increasingly serious, bringing huge challenges to people's travel and urban management. Accurately predicting the trend of traffic flow changes in the future is of great significance for alleviating traffic congestion, optimizing traffic induction, and rationally allocating traffic resources.

[0003] In recent years, the vigorous development of artificial intelligence technology has brought new opportunities for traffic flow prediction. More and more studies have begun to explore the use of intelligent algorithms such as machine learning and deep learning to build traffic flow prediction models. However, the existing technology still has some limitations and shortcomings, which are mainly reflected in the following aspects:

[0004] Insufficient data utilization. Most methods rely on a single data source such as historical traffic flow data, ignoring the integration of multi-source heterogeneous data such as weather, road network structure, and driving behavior, making it difficult to fully describe the complex laws of traffic flow evolution.

[0005] The model is not general enough. Most existing methods are targeted at specific scenarios and data distributions, lack effective processing of extreme events and outliers, and the extrapolation performance and robustness in practical applications need to be improved.

[0006] Poor interpretability. Many deep learning models are black box models, and the prediction results are difficult to explain and cannot provide a basis for decision-making, which is not conducive to the analysis and intervention of traffic management personnel.

[0007] The optimization efficiency is not high. As the data scale and model complexity increase, the computational overhead of parameter optimization and model training increases significantly, making it difficult to meet the needs of real-time prediction.

[0008] Chinese invention patent CN117994986B proposed a BP neural network traffic flow prediction method based on an improved electric eel foraging optimization algorithm. This method has made useful attempts in data preprocessing and key factor analysis, but it still mainly relies on a single neural network model, lacks multi-model fusion, and has limited generalization ability. In addition, the BP network is a shallow network and has insufficient modeling capabilities for complex nonlinear relationships.

[0009] Chinese invention patent CN116110234B proposed a traffic flow prediction method based on CNN and LSTM. This method explored data enhancement, anomaly detection, feature extraction, etc., but lacked in-depth research and improvement on the attention mechanism, and the spatiotemporal correlation modeling was not sophisticated enough. In addition, this method did not address the interpretability of the prediction results and the optimization efficiency issues.

[0010] In summary, the shortcomings of existing traffic flow prediction technology are mainly: insufficient use of data fusion, the generalization and robustness of the model need to be strengthened, the predictive results are not interpretable, and the optimization efficiency needs to be improved. These problems restrict the widespread application of traffic flow prediction in the field of smart transportation. Summary of the invention

[0011] In view of the above-mentioned prior art, the technical problem to be solved by the present invention is the problems of insufficient data utilization, insufficient model generalization, weak interpretability and low optimization efficiency in the prior art.

[0012] To solve the above problems, the present invention provides a traffic flow prediction and optimization method based on an intelligent optimization algorithm, including a central processing unit for executing computing tasks and controlling system operation, a graphics processing unit for accelerating the training and reasoning of deep learning models, a memory for storing system software, model parameters and historical data, and a communication interface for exchanging data with external devices and networks. The number of cores of the graphics processing unit is not less than 1000, and the capacity of the memory is not less than 1TB. The method also includes a data acquisition and preprocessing module, a multimodal data fusion module, a hybrid deep learning prediction module, an intelligent parameter optimization module, an integrated learning module, a multi-scale error correction module, a prediction result evaluation and feedback module, an explainability and visualization module, an edge computing and distributed learning module, and a system management and interface module that are sequentially connected;

[0013] The data collection and preprocessing module is used to collect traditional traffic data and non-traditional data, and perform denoising, anomaly detection, time series segmentation and dimensionality reduction processing;

[0014] The multimodal data fusion module is used to integrate and fuse multi-source data through the attention mechanism and knowledge graph;

[0015] The hybrid deep learning prediction module is used to extract spatiotemporal features and build a prediction model using CNN and Transformer structures;

[0016] The intelligent parameter optimization module is used to optimize model parameters by combining quantum computing and improved bat algorithm;

[0017] The ensemble learning module is used to dynamically integrate multiple base learners and enhance the model robustness through adversarial samples;

[0018] The multi-scale error correction module is used to perform short-, medium- and long-term error correction using Kalman filtering and residual learning;

[0019] The prediction result evaluation and feedback module is used to verify the prediction performance in multiple scenarios and achieve human-machine collaborative optimization;

[0020] The explainability and visualization module is used to generate multi-level explanation reports and interactive visualization interfaces;

[0021] The edge computing and distributed learning modules are used to achieve distributed prediction with edge-cloud collaboration and privacy protection;

[0022] The system management and interface module is used to uniformly schedule system resources and provide external service interfaces;

[0023] The hybrid deep learning prediction module uses the following fusion algorithm formula for prediction:

[0024] Y=α*FCNN(X1)+β*FTransformer(X2)+γ*FLSTM(X3)+δ*G(E)

[0025] Among them, Y is the final prediction result, X1, X2, and X3 represent the input data of different modalities, FCNN, FTransformer, and FLSTM represent the prediction functions of CNN, Transformer, and LSTM, respectively, G(E) is the error correction function, and α, β, γ, and δ are dynamic weights determined by the attention mechanism;

[0026] The dynamic weights α, β, γ, δ are calculated through the following attention mechanism:

[0027] [α,β,γ,δ]=softmax(W*[h1;h2;h3;hE]+b)

[0028] Where W is the weight matrix, b is the bias vector, h1, h2, h3, and hE are the hidden states of the CNN, Transformer, LSTM, and error correction models, respectively.

[0029] In the above-mentioned traffic flow prediction optimization method based on intelligent optimization algorithm, the accuracy, robustness and real-time performance of traffic flow prediction can be improved, and a reliable basis can be provided for decision-making such as road traffic status analysis, signal light duration optimization, and path planning guidance, effectively alleviating traffic congestion and improving travel experience. The present invention can also save the cost of traffic infrastructure construction, extend the service life of roads, reduce exhaust emissions, and create considerable economic and social benefits.

[0030] As a further improvement of the present application, the data acquisition and preprocessing module includes a traditional traffic data acquisition unit, a non-traditional data acquisition unit, a wavelet transform denoising unit, a graph neural network anomaly detection unit, a dynamic time series segmentation unit and a principal component analysis dimensionality reduction unit, which are sequentially connected;

[0031] Traditional traffic data collection units are used to collect conventional traffic data such as vehicle flow and speed;

[0032] Non-traditional data collection units are used to collect auxiliary data such as satellite images and social media;

[0033] The wavelet transform denoising unit is used to perform multi-scale denoising on the original data;

[0034] The graph neural network anomaly detection unit is used to identify abnormal data based on the road network topology;

[0035] The dynamic time series segmentation unit is used to adaptively divide the time window;

[0036] The principal component analysis dimensionality reduction unit is used to reduce the data dimension and retain the main features.

[0037] As a further improvement of the present application, the multimodal data fusion module includes a data alignment and standardization unit, a knowledge graph integration unit and an attention mechanism fusion unit which are sequentially connected;

[0038] The attention mechanism fusion unit is used to adaptively assign weights to different data sources;

[0039] The knowledge graph integration unit is used to integrate prior knowledge such as urban planning and transportation policies;

[0040] The data alignment and standardization unit is used to unify the format and scale of data from different sources.

[0041] As a further improvement of the present application, the hybrid deep learning prediction module includes a CNN spatial feature extraction unit, a Transformer temporal dependency capture unit, a spatiotemporal attention mechanism unit, and a multi-layer perceptron prediction unit connected in sequence;

[0042] The CNN spatial feature extraction unit is used to capture the spatial dependencies of the road network;

[0043] Transformer temporal dependency capture unit is used to learn long-term temporal dependencies;

[0044] The spatiotemporal attention mechanism unit is used to dynamically adjust the importance of spatiotemporal features;

[0045] The multi-layer perceptron prediction unit is used to finally generate the prediction results.

[0046] As another improvement of the present application, the intelligent parameter optimization module includes a quantum computing optimization unit, an improved bat algorithm unit, a dynamic neural architecture search unit, and an adaptive search strategy unit connected in sequence;

[0047] The quantum computing optimization unit is used to accelerate the search for the global optimal solution;

[0048] Improved bat algorithm unit for adaptive parameter adjustment;

[0049] Dynamic neural architecture search unit is used to automatically optimize the network structure;

[0050] The adaptive search strategy unit is used to dynamically adjust the search behavior of the optimization algorithm.

[0051] As another improved supplement of the present application, the integrated learning module includes a base learner training unit, a dynamic weighted integration unit, an adversarial sample generation unit and a model robustness verification unit connected in sequence;

[0052] The base learner training unit is used to train various models such as STAN, GBDT, and XGBoost;

[0053] The dynamic weighted integration unit is used to adjust the model weights according to different spatiotemporal conditions;

[0054] The adversarial sample generation unit is used to generate challenging test samples;

[0055] The model robustness verification unit is used to evaluate the performance of the model under adversarial samples.

[0056] As another improved supplement of the present application, the multi-scale error correction module includes a Kalman filter short-term correction unit, a multi-scale residual learning unit and an adaptive online learning unit connected in sequence;

[0057] The Kalman filter short-term correction unit is used to correct the short-term prediction error in real time;

[0058] Multi-scale residual learning units are used to simultaneously consider short-, medium- and long-term prediction errors;

[0059] The adaptive online learning unit is used to update the model in real time to adapt to traffic pattern changes.

[0060] As another improvement of the present application, the prediction result evaluation and feedback module includes a multi-index evaluation unit, a multi-scenario simulation verification unit, a human-machine collaborative feedback unit and a model dynamic adjustment unit which are sequentially connected;

[0061] The multi-index evaluation unit is used to comprehensively evaluate the accuracy, stability and timeliness of the forecast;

[0062] The multi-scenario simulation verification unit is used to test the model performance under extreme conditions in a virtual environment;

[0063] The human-machine collaborative feedback unit is used to combine expert experience and machine learning for model tuning;

[0064] The model dynamic adjustment unit is used to automatically optimize the model structure and parameters according to the evaluation results.

[0065] The edge computing and distributed learning module includes an edge-cloud collaborative learning scheduling unit, a federated learning protocol unit, a privacy-preserving computing unit, and a distributed model aggregation unit that are sequentially connected;

[0066] The edge-cloud collaborative learning scheduling unit is used to optimize the allocation of computing tasks on edge devices and the cloud;

[0067] The federated learning protocol unit is used to achieve secure sharing and collaborative modeling of multi-region data;

[0068] The privacy-preserving computing unit is used to perform distributed computing while protecting data privacy;

[0069] The distributed model aggregation unit is used to integrate multiple local models to generate a global model;

[0070] The privacy-preserving computing unit uses homomorphic encryption technology, which specifically includes: a data encryption subunit, a secure computing subunit, and a result decryption subunit connected in sequence;

[0071] The data encryption subunit is used to encrypt the original data;

[0072] A secure computing subunit for data analysis and model training in an encrypted domain;

[0073] A result decryption subunit, used to securely decrypt the calculation result;

[0074] Homomorphic encryption uses the Paillier encryption algorithm, and the key length is no less than 2048 bits.

[0075] A traffic flow prediction optimization method based on an intelligent optimization algorithm comprises the following steps:

[0076] S1. Collect traditional traffic data and non-traditional data, and perform denoising, anomaly detection, time series segmentation and dimensionality reduction preprocessing;

[0077] S2, using attention mechanism and knowledge graph integration method to fuse multi-source data;

[0078] S3. Build a hybrid deep learning prediction model combining CNN and Transformer to extract spatiotemporal features; use the following fusion algorithm formula for prediction:

[0079] Y=α*FCNN(X1)+β*FTransformer(X2)+γ*FLSTM(X3)+δ*G(E)

[0080] Among them, Y is the final prediction result, X1, X2, and X3 represent the input data of different modalities, FCNN, FTransformer, and FLSTM represent the prediction functions of CNN, Transformer, and LSTM, respectively, G(E) is the error correction function, and α, β, γ, and δ are dynamic weights determined by the attention mechanism;

[0081] S4. Use quantum computing and improved bat algorithm to optimize model parameters and structure;

[0082] Among them, the number of quantum bits optimized for quantum computing is set between 8 and 16, the population size of the improved bat algorithm is set between 20 and 50, and the number of iterations is 100 to 500;

[0083] S5, dynamically integrate multiple base learners and enhance model robustness by generating adversarial samples;

[0084] S6, use Kalman filtering and multi-scale residual learning to perform short-, medium- and long-term error correction;

[0085] Among them, the time window for short-term error correction is set to 1-3 hours, medium-term to 6-12 hours, and long-term to 24-72 hours;

[0086] S7. Verify the prediction performance in multiple scenarios and optimize the model through human-computer collaboration;

[0087] S8, generate multi-level interpretation reports and interactive visualization results;

[0088] S9, achieving edge-cloud collaboration and privacy-preserving distributed prediction based on federated learning;

[0089] S10. Provide prediction services through a unified system management interface.

[0090] In summary, this application has the following beneficial effects:

[0091] 1. To improve the prediction accuracy, the present invention adopts a multimodal data fusion strategy, comprehensively utilizes multi-source heterogeneous data such as traffic flow data, satellite image data, and road network structure data, and realizes deep fusion and feature enhancement of data through attention mechanism and knowledge graph, thereby significantly improving the accuracy of traffic flow prediction. In addition, the fusion algorithm further optimizes the prediction accuracy by weighted integration of multiple deep learning models such as CNN, Transformer, LSTM, and introducing an adaptive error correction mechanism.

[0092] 2. Enhanced robustness and generalization. The present invention introduces a variety of optimization strategies to enhance the robustness and generalization performance of the prediction model. Through adversarial training, deceptive adversarial samples are generated to improve the model's ability to cope with extreme scenarios; data enhancement technology is used to expand the diversity of training samples and reduce the risk of overfitting; transfer learning is used to draw on prior knowledge from other cities and transportation systems to accelerate the model learning and optimization process. The prediction model of the present invention can still maintain stable and reliable prediction performance in the presence of interference factors and data loss, and its adaptability and extrapolation capabilities are enhanced.

[0093] 3. Provide rich explanations and analyses. The present invention specially designs an interpretable analysis and visualization module, which uses technologies such as mean and causal inference to explore key influencing factors and reveal the internal mechanism and law of traffic flow changes; through multi-level visualization and interactive interface, the prediction results are intuitively presented to facilitate user understanding and analysis. The rich explanation and analysis functions provide powerful auxiliary decision-making support for traffic management departments, help timely gain insight into traffic conditions, and optimize traffic scheduling and control measures.

[0094] 4. Achieve efficient optimization and distributed learning. The present invention adopts advanced optimization technologies such as quantum computing and swarm intelligence optimization, which greatly improves the search efficiency of model parameters and structures and reduces computing resource consumption. At the same time, the present invention introduces edge computing and federated learning paradigms. Through distributed collaborative learning and model aggregation, while protecting data privacy, the computing and storage resources of edge nodes such as roadside equipment and vehicle-mounted terminals are fully utilized to achieve low latency and high efficiency in model training and reasoning.

[0095] 5. Easy to integrate, apply and promote. The present invention provides a unified system architecture and management interface. Each functional module is decoupled, independent and organically integrated, and has strong modularity, configurability and scalability. The system supports flexible parameter configuration and dynamic function combination, and can be easily integrated into various intelligent transportation systems and application platforms. The standardized API interface facilitates interconnection with third-party systems, improving the ease of use and scalability of the technical solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0096] Figure 1 This is the overall structure diagram of this application;

[0097] Figure 2 This is a flow chart of the interactions between the various modules of this application;

[0098] Figure 3 This is a diagram of data flow and processing sequence within the data acquisition and preprocessing module of this application;

[0099] Figure 4 This is a timing diagram of the hybrid deep learning prediction module of this application. DETAILED DESCRIPTION

[0100] Five implementation modes of the present application are described in detail below with reference to the accompanying drawings.

[0101] Example 1

[0102] See also Figure 1 , Figure 2 and Figure 3 , this embodiment provides a traffic flow prediction system based on an intelligent optimization algorithm, including a data acquisition and preprocessing module, a multimodal data fusion module and a hybrid deep learning prediction module connected in sequence.

[0103] The data acquisition and preprocessing module further includes a traditional traffic data acquisition unit, a non-traditional data acquisition unit, a wavelet transform denoising unit, a graph neural network anomaly detection unit, a dynamic time series segmentation unit and a principal component analysis dimensionality reduction unit which are connected in sequence.

[0104] Among them, the traditional traffic data collection unit is used to collect conventional traffic parameter data such as vehicle flow and speed. The non-traditional data collection unit further collects multi-source heterogeneous data such as satellite images and social media information to enrich the data dimension and information volume.

[0105] The wavelet transform denoising unit uses the multi-scale characteristics of wavelet transform to perform denoising on the collected raw data, remove high-frequency noise interference, and improve data quality. The graph neural network anomaly detection unit makes full use of the topological information of the road network structure to build a graph neural network model to automatically identify abnormal data and ensure the accuracy and reliability of the data.

[0106] The dynamic time series segmentation unit adopts an adaptive windowing method to dynamically adjust the time window according to data characteristics and task requirements, which can fully utilize time series information and reduce computational complexity. The principal component analysis dimensionality reduction unit uses a linear transformation method to reduce data dimensions and highlight key factors while minimizing information loss, thus preparing for subsequent fusion modeling.

[0107] The multimodal data fusion module further includes a data alignment and normalization unit, a knowledge graph integration unit and an attention mechanism fusion unit which are connected in sequence.

[0108] The data alignment and standardization unit solves the heterogeneity problem of data from different sources and formats. Through operations such as interpolation and normalization, it achieves spatiotemporal alignment and scale unification. The knowledge graph integration unit introduces prior knowledge such as road network topology and traffic rules, constructs a knowledge graph that integrates multimodal data and rule constraints, and gives the data more semantic information.

[0109] The attention mechanism fusion unit uses the attention mechanism to dynamically assign weights to data of different modalities, adaptively focusing on key sections and key time periods to capture important features. Through the organic fusion of multi-dimensional features, the spatiotemporal evolution of road network traffic flow is comprehensively described.

[0110] The hybrid deep learning prediction module contains a CNN spatial feature extraction unit, a Transformer temporal dependency capture unit, a spatiotemporal attention mechanism unit, and a multi-layer perceptron prediction unit that are connected in sequence.

[0111] The CNN spatial feature extraction unit uses the local receptive field and weight sharing characteristics of the convolutional neural network to automatically extract multi-scale spatial correlation features of the road network. The Transformer temporal dependency capture unit introduces a self-attention mechanism to mine long-range dependencies in the time series of road traffic flow. The spatiotemporal attention mechanism unit calculates attention weights in both time and space dimensions and dynamically adjusts the importance of spatiotemporal features.

[0112] Finally, the multi-layer perceptron prediction unit integrates the hidden layer states that represent the spatiotemporal characteristics and achieves accurate prediction of future traffic flow through a fully connected network structure.

[0113] In addition, the hybrid deep learning prediction module also designs a fusion prediction algorithm to give full play to the respective strengths of different deep learning models:

[0114] Y=α*FCNN(X1)+β*FTransformer(X2)+γ*FLSTM(X3)+δ*G(E)

[0115] In the formula, Y is the final predicted traffic flow, X1, X2, and X3 represent different modal input data, FCNN, FTransformer, and FLSTM represent the feature representations extracted by CNN, Transformer, and LSTM networks, respectively, and G(E) is the error correction function. α, β, γ, and δ are weight coefficients dynamically assigned according to the attention mechanism, reflecting the importance of different models.

[0116] By weighted integration of CNN, Transformer, LSTM and other models, integrating the ability to capture local spatial features and global temporal features, and introducing an error correction mechanism, the accuracy and robustness of traffic flow prediction can be significantly improved.

[0117] In summary, the traffic flow prediction system proposed in this embodiment comprehensively considers the spatiotemporal characteristics and multi-source heterogeneous characteristics of traffic flow data, integrates prior knowledge, and combines multiple deep learning models to build a complete set of intelligent traffic flow prediction solutions. The system has high prediction accuracy and strong adaptability, and can effectively support the construction of smart transportation.

[0118] Example 2

[0119] See also Figures 1 to 4 Based on Example 1, this example further improves the model optimization and integration strategy of the traffic flow prediction system, and introduces an intelligent parameter optimization module, an integrated learning module and a multi-scale error correction module.

[0120] The intelligent parameter optimization module includes a quantum computing optimization unit, an improved bat algorithm unit, a dynamic neural architecture search unit and an adaptive search strategy unit which are connected in sequence.

[0121] The quantum computing optimization unit uses quantum bits to represent the solution space and accelerates the optimization process through quantum parallel computing. The improved bat algorithm unit imitates the echolocation behavior of bats and introduces adaptive frequency, loudness and pulse emission rate adjustment strategies to balance global exploration and local development. The dynamic neural architecture search unit automatically optimizes the hyperparameters such as the number of layers and kernel size of CNN, Transformer and other models. The adaptive search strategy unit dynamically adjusts the trade-off between exploration and utilization according to the model performance to improve search efficiency.

[0122] Through intelligent parameter optimization, the model training speed and generalization ability are significantly improved. The number of quantum bits optimized for quantum computing is set to 8-16, the population of the improved bat algorithm is 20-50, and the number of iterations is 100-500, which greatly shortens the optimization time while ensuring the quality of search.

[0123] The integrated learning module includes a base learner training unit, a dynamic weighted integration unit, an adversarial sample generation unit and a model robustness verification unit which are connected in sequence.

[0124] The base learner training unit uses cross-validation and other methods to train multiple heterogeneous models, such as LSTM, GRU, and decision trees. The dynamic weighted integration unit adaptively adjusts the weights of each model according to background factors such as time and weather to highlight the role of key models. The adversarial sample generation unit uses the generative adversarial network to automatically generate samples that are confusing to the model. The model robustness verification unit uses adversarial samples to test the generalization ability of the model and screen out models with high robustness.

[0125] By dynamically integrating multiple heterogeneous models and combining the prediction capabilities of different models, the prediction bias can be further reduced. At the same time, the adversarial learning strategy effectively improves the robustness of the model, enabling it to cope with extreme abnormal scenarios.

[0126] The multi-scale error correction module includes a Kalman filter short-term correction unit, a multi-scale residual learning unit and an adaptive online learning unit which are connected in sequence.

[0127] The Kalman filter short-term correction unit uses the Kalman filter algorithm to correct the prediction error in real time using new observation data. The multi-scale residual learning unit establishes a residual model of traffic flow changes at short, medium and long time scales to capture the fluctuation pattern of traffic flow. The adaptive online learning unit dynamically adjusts the model based on the prediction error feedback to adapt to the changing traffic conditions.

[0128] Combined with the characteristics of multi-scale traffic flow, short-term prediction deviations are corrected in real time, and medium- and long-term deviations are modeled using residuals, which can significantly reduce cumulative errors. The short-term correction time window is set to 1-3 hours, the medium-term is 6-12 hours, and the long-term is 24-72 hours, taking into account both computational efficiency and prediction accuracy.

[0129] In addition, the adaptive online learning strategy enables the model to continuously update and improve itself according to the latest observation data, with continuous learning capabilities and adaptability to dynamic traffic environments.

[0130] In summary, based on Example 1, this embodiment further enhances the performance of the prediction model from aspects such as parameter optimization, model integration, and error correction. While improving the prediction accuracy, it significantly improves the robustness, generalization ability, and self-learning ability of the model, making the intelligent traffic prediction system more practical and complete.

[0131] Example 3

[0132] See also Figures 1 to 4 Based on Example 1 and Example 2, this example further improves the practical design of the traffic flow prediction system, and introduces a prediction result evaluation and feedback module, an explainability and visualization module, and an edge computing and distributed learning module.

[0133] The prediction result evaluation and feedback module includes a multi-index evaluation unit, a multi-scenario simulation verification unit, a human-computer collaborative feedback unit and a model dynamic adjustment unit which are connected in sequence.

[0134] The multi-index evaluation unit comprehensively evaluates the prediction accuracy and stability based on multiple indicators such as mean absolute error, root mean square error, and mean relative error. The multi-scenario simulation verification unit verifies the model prediction effect in a simulation environment for extreme situations such as holidays and bad weather. The human-machine collaborative feedback unit provides an expert feedback interface, which comprehensively utilizes human experience and knowledge and machine analysis capabilities to continuously optimize the model. The model dynamic adjustment unit adaptively corrects model parameters and structure based on multi-dimensional evaluation results.

[0135] Through a closed-loop evaluation and feedback mechanism, we can improve model performance in a targeted manner and enhance the applicability of the algorithm in complex and ever-changing real-world scenarios. At the same time, we can introduce the collaboration between artificial intelligence and human experts to give full play to their complementary advantages and improve the effectiveness of optimization.

[0136] The explainability and visualization module includes a key feature analysis unit, a causal relationship mining unit, a hierarchical visualization unit and an interactive interface unit which are connected in sequence.

[0137] The key feature analysis unit uses game theory methods such as mean and Nash equilibrium to measure the impact of key factors on the prediction results. The causal relationship mining unit uses a causal inference graph model to reveal the internal mechanism of traffic flow changes. The layered visualization unit intuitively presents the prediction results from multiple levels, including macro, meso and micro. The interactive interface unit provides an interactive operation interface and supports user-defined query analysis.

[0138] Increase interpretability and visualization design to make complex deep learning models and prediction processes transparent, easy for traffic management personnel to understand and interpret, and enhance credibility. Through key feature analysis and causal relationship mining, the law of traffic flow changes is deeply revealed. With multi-level visualization design and humanized interaction, the prediction effect is intuitively displayed, improving the ease of use of the system.

[0139] The edge computing and distributed learning module includes an edge-cloud collaborative learning scheduling unit, a federated learning protocol unit, a privacy-preserving computing unit, and a distributed model aggregation unit, which are connected in sequence.

[0140] The edge-cloud collaborative learning scheduling unit coordinates the scheduling of edge devices and cloud resources to optimize model training and reasoning efficiency. The federated learning protocol unit implements encrypted communication and parameter aggregation between multiple edge nodes. The privacy-preserving computing unit uses technologies such as homomorphic encryption and secure multi-party computing to carry out distributed learning while protecting data privacy. The distributed model aggregation unit uses weighted averaging, voting and other mechanisms to integrate and aggregate multiple local models to form a global model.

[0141] Introducing edge computing and distributed learning paradigms, fully utilizing the computing and storage resources of edge nodes such as roadside cameras and vehicle-mounted equipment to achieve low latency and high efficiency in model training and reasoning. Through the federated learning architecture, distributed data is used to improve model performance while protecting data privacy. Using homomorphic encryption technology, encrypted data participates in calculations, and encrypted results can be obtained without decryption, greatly reducing the risk of data leakage.

[0142] The homomorphic encryption process is as follows: the data encryption unit encrypts the original data, sends the encrypted data to the secure computing unit for computational processing, and the result decryption unit decrypts the encrypted computational result. The entire process uses the Paillier encryption algorithm, supports addition and multiplication operations, and has a key length of no less than 2048 bits, taking into account both security and efficiency.

[0143] In summary, based on Examples 1 and 2, this embodiment further strengthens the practical design of the system. While improving prediction accuracy and operating efficiency, it fully considers the interpretability of the algorithm, visual interactivity, and data security and privacy protection, laying a solid foundation for the actual deployment and application of the intelligent traffic prediction system.

[0144] Example 4

[0145] See also Figures 1 to 4 This embodiment provides a traffic flow prediction optimization method based on an intelligent optimization algorithm. Combined with the system modules in Embodiments 1-3, the entire prediction optimization process is described in detail, including the following steps:

[0146] S1. Data collection and preprocessing. Multi-source heterogeneous data such as vehicle flow, speed, and satellite images are collected through traditional traffic data collection units and non-traditional data collection units. The original data is denoised using the wavelet transform denoising unit, the graph neural network anomaly detection unit identifies abnormal data, the dynamic time series segmentation unit performs adaptive time window division, and the principal component analysis dimensionality reduction unit achieves data dimensionality reduction.

[0147] S2. Multimodal data fusion. The data alignment and standardization unit realizes the spatiotemporal alignment and scale unification of data from different sources. The knowledge graph integration unit integrates prior knowledge to construct a semantic association graph. The attention mechanism fusion unit uses the attention mechanism to adaptively empower and highlight key sections and time features.

[0148] S3, hybrid deep learning prediction. The CNN spatial feature extraction unit is used to extract multi-scale spatial features, the Transformer temporal dependency capture unit is used to model long-range temporal dependencies, and the spatiotemporal attention mechanism unit dynamically adjusts the spatiotemporal dimension feature weights. Finally, the multi-layer perceptron prediction unit is used to generate the future traffic flow prediction value. The following fusion prediction algorithm is used:

[0149] Y=α*FCNN(X1)+β*FTransformer(X2)+γ*FLSTM(X3)+δ*G(E)

[0150] Among them, Y is the predicted traffic flow, X1, X2, and X3 are inputs of different modes, FCNN, FTransformer, and FLSTM are CNN, Transformer, and LSTM features, respectively, G(E) is the error correction function, and α, β, γ, and δ are attention weight coefficients.

[0151] S4, intelligent parameter optimization. The quantum computing optimization unit uses quantum parallel computing to accelerate optimization, with 8-16 quantum bits. The improved bat algorithm unit balances exploration and exploitation through adaptive parameter adjustment strategies, with 20-50 populations and 100-500 iterations. The dynamic neural architecture search unit optimizes network hyperparameters, and the adaptive search strategy unit dynamically balances exploration and exploitation to improve search efficiency.

[0152] S5. Ensemble learning prediction. The base learner training unit trains multiple heterogeneous models such as LSTM, GRU, etc. The dynamic weighted integration unit dynamically adjusts the model weights according to background factors. The adversarial sample generation unit uses GAN to generate adversarial samples, and the model robustness verification unit selects models with good generalization performance for integration.

[0153] S6, multi-scale error correction. The Kalman filter short-term correction unit corrects the prediction deviation in real time through Kalman filtering, with a time window of 1-3 hours. The multi-scale residual learning unit builds a medium- and long-term residual model to capture the traffic flow fluctuation pattern, with a medium-term time window of 6-12 hours and a long-term time window of 24-72 hours. The adaptive online learning unit dynamically adjusts the model according to feedback and continuously optimizes adaptively.

[0154] S7. Prediction performance evaluation and feedback. The multi-index evaluation unit evaluates the prediction accuracy and stability from multiple evaluation indicators. The multi-scenario simulation verification unit simulates and tests the model performance in extreme scenarios. The human-machine collaborative feedback unit optimizes the model by combining expert experience and machine analysis capabilities. The model dynamic adjustment unit adaptively modifies the model parameters and structure accordingly.

[0155] S8. Explainable analysis and visualization. The key feature analysis unit uses SHAP and other methods to analyze key influencing factors. The causal relationship mining unit infers the causal mechanism of traffic flow changes. The hierarchical visualization unit presents the prediction results from macro, meso and micro levels. The interactive interface unit provides a user-friendly interactive analysis operation interface.

[0156] S9. Federated learning and privacy protection. The edge-cloud collaborative learning scheduling unit coordinates the scheduling of edge and cloud resources to optimize the efficiency of model training and inference. The federated learning protocol unit implements secure communication and parameter aggregation between distributed nodes. The privacy protection computing unit uses homomorphic encryption, encrypted data participates in the calculation, and the result decryption unit decrypts the calculation result. The distributed model aggregation unit uses weighted average and other mechanisms to fuse local models into a global model.

[0157] S10, prediction service interface. Through the unified system management interface module, various functional modules are integrated to provide a high-availability, high-concurrency vehicle flow prediction service API interface, supporting flexible calls to third-party systems.

[0158] The above steps collect multi-source heterogeneous traffic big data, deeply integrate semantic information, mix multiple advanced deep learning models, and use quantum computing and swarm intelligence optimization algorithms to achieve parameter optimization. Dynamically integrate multiple heterogeneous weak learners to build a traffic flow prediction model with spatiotemporal attention and error correction capabilities. After continuous self-optimization through multi-dimensional evaluation feedback, federated learning and privacy protection mechanisms are introduced, ultimately forming a complete, efficient, secure, and explainable traffic flow prediction optimization method, providing accurate and reliable decision-making basis for smart transportation systems.

[0159] Example 5

[0160] See also Figures 1 to 4 This embodiment proposes an improved hybrid deep learning traffic flow prediction model, which further refines and improves the model's fusion algorithm and optimization strategy based on embodiments 1-4.

[0161] The fusion algorithm formula of the prediction model is:

[0162] Y=α*FCNN(X1)+β*FTransformer(X2)+γ*FLSTM(X3)+δ*G(E,X4)

[0163] Among them, Y is the final traffic flow prediction result. X1, X2, and X3 represent the input features of three different modes respectively:

[0164] X1 represents the time series characteristics, statistical characteristics, and frequency domain characteristics of traffic flow data;

[0165] X2 represents the high-level semantic features extracted from satellite image data;

[0166] X3 represents the features obtained by modeling the road network topology using graph neural network.

[0167] FCNN, FTransformer, and FLSTM are optimized CNN, Transformer, and LSTM prediction models respectively:

[0168] FCNN enhances the ability to capture local and global spatial dependencies by introducing multi-scale convolution and attention mechanisms;

[0169] FTransformer enhances the modeling capability of long-distance temporal dependencies by introducing position encoding and multi-head attention mechanism;

[0170] FLSTM adopts the bidirectional LSTM structure and attention mechanism to better mine bidirectional temporal correlation information.

[0171] G(E,X4) is an adaptive error correction function that not only considers the historical forecast error E, but also introduces external environmental factors X4 (such as weather, holidays, etc.) to dynamically adjust the correction strategy. G(E,X4) uses time series decomposition technology to model and correct errors of different time scales separately.

[0172] The dynamic weights α, β, γ, δ in the model are calculated through an improved attention mechanism:

[0173] [α,β,γ,δ]=softmax(W*[h1+e1;h2+e2;h3+e3;hE+eE]+b)

[0174] Among them, W and b are the weight matrix and bias vector of the attention mechanism, respectively, and h1, h2, h3, and hE are the hidden state outputs of FCNN, FTransformer, FLSTM, and the error correction model, respectively.

[0175] e1, e2, e3, and eE are embedding vectors of external environmental factors, obtained through end-to-end learning, and used to assist in adjusting the distribution of attention weights:

[0176] e1 is the time step embedding, which allows the model to give different weights according to different time steps;

[0177] e2 is the weather condition embedding, which allows the model to dynamically adjust its focus according to different weather conditions;

[0178] e3 is the holiday embedding, which allows the model to pay more attention to special periods such as holidays;

[0179] eE is the historical error distribution embedding, which allows the model to adjust the weight distribution according to the dynamic changes of the error distribution.

[0180] In order to further improve the model performance, this embodiment also adopts the following optimization strategies:

[0181] Using the stacking ensemble learning method, a two-layer stacking structure is designed: the first layer uses the outputs of FCNN, FTransformer, and FLSTM as independent features, and the second layer uses the attention-enhanced meta-model Ymeta=Attention(Q,K,V) to learn the interactions and dependencies between different base models and output the final prediction results.

[0182] During the model training process, the Cyclical Learning Rates learning rate scheduling strategy is used to accelerate convergence, the elastic network regularization technology combining L1 and L2 is used to reduce overfitting, adversarial training is introduced to enhance robustness, and data enhancement techniques such as data perturbation and interpolation are applied to improve generalization performance.

[0183] By leveraging the idea of ​​transfer learning, we learn prior knowledge from other cities and transportation systems, and design adaptive transfer strategies and multi-task loss functions to avoid negative transfer while improving model performance.

[0184] In summary, this embodiment has comprehensively refined and improved the hybrid deep learning traffic flow prediction model, starting from multimodal data fusion, model structure optimization, attention mechanism improvement, training strategy enhancement, transfer learning application and other aspects, forming a set of effective technical solutions, which provides a theoretical basis and implementation path for subsequent engineering practice.

[0185] In view of current practical needs, the above-mentioned implementation mode adopted in this application is not limited to the scope of protection. Various changes made within the knowledge scope of technical personnel in this field without departing from the concept of this application still fall within the scope of protection of the present invention.

Claims

1. A traffic flow prediction and optimization system based on an intelligent optimization algorithm, comprising a central processing unit for executing computing tasks and controlling system operation, a graphics processing unit for accelerating the training and reasoning of deep learning models, a memory for storing system software, model parameters and historical data, and a communication interface for exchanging data with external devices and networks, wherein the number of cores of the graphics processing unit is not less than 1000, and the capacity of the memory is not less than 1TB, and is characterized in that: It also includes a data acquisition and preprocessing module, a multimodal data fusion module, a hybrid deep learning prediction module, an intelligent parameter optimization module, an integrated learning module, a multi-scale error correction module, a prediction result evaluation and feedback module, an explainability and visualization module, an edge computing and distributed learning module, and a system management and interface module, which are connected in sequence; The data acquisition and preprocessing module is used to collect traditional traffic data and non-traditional data, and perform denoising, anomaly detection, time series segmentation and dimensionality reduction processing; The multimodal data fusion module is used to integrate and fuse multi-source data through the attention mechanism and knowledge graph; The hybrid deep learning prediction module is used to extract spatiotemporal features and build a prediction model using CNN and Transformer structures; The intelligent parameter optimization module is used to optimize model parameters by combining quantum computing and improved bat algorithm; The integrated learning module is used to dynamically integrate multiple base learners and enhance the model robustness through adversarial samples; The multi-scale error correction module is used to perform short-, medium- and long-term error correction using Kalman filtering and residual learning; The prediction result evaluation and feedback module is used to verify the prediction performance in multiple scenarios and achieve human-machine collaborative optimization; The explainability and visualization module is used to generate a multi-level explanation report and an interactive visualization interface; The edge computing and distributed learning module is used to achieve distributed prediction with edge-cloud collaboration and privacy protection; The system management and interface module is used to uniformly schedule system resources and provide external service interfaces; The hybrid deep learning prediction module uses the following fusion algorithm formula for prediction: Y=α*FCNN(X1)+β*FTransformer(X2)+γ*FLSTM(X3)+δ*G(E) Among them, Y is the final prediction result, X1, X2, and X3 represent the input data of different modalities, FCNN, FTransformer, and FLSTM represent the prediction functions of CNN, Transformer, and LSTM, respectively, G(E) is the error correction function, and α, β, γ, and δ are dynamic weights determined by the attention mechanism; The dynamic weights α, β, γ, δ are calculated by the following attention mechanism: [α,β,γ,δ]=softmax(W*[h1;h2;h3;hE]+b) Where W is the weight matrix, b is the bias vector, h1, h2, h3, and hE are the hidden states of the CNN, Transformer, LSTM, and error correction models, respectively.

2. The traffic flow prediction and optimization system based on intelligent optimization algorithm according to claim 1 is characterized by: The data acquisition and preprocessing module includes a traditional traffic data acquisition unit, a non-traditional data acquisition unit, a wavelet transform denoising unit, a graph neural network anomaly detection unit, a dynamic time series segmentation unit and a principal component analysis dimensionality reduction unit connected in sequence; The conventional traffic data collection unit is used to collect conventional traffic data such as vehicle flow and speed; The non-traditional data acquisition unit is used to collect auxiliary data such as satellite images and social media; The wavelet transform denoising unit is used to perform multi-scale denoising on the original data; The graph neural network anomaly detection unit is used to identify abnormal data based on the road network topology structure; The dynamic time series segmentation unit is used to adaptively divide the time window; The principal component analysis dimension reduction unit is used to reduce the data dimension and retain the main features.

3. The traffic flow prediction and optimization system based on intelligent optimization algorithm according to claim 1 is characterized by: The multimodal data fusion module includes a data alignment and standardization unit, a knowledge graph integration unit and an attention mechanism fusion unit which are sequentially connected; The attention mechanism fusion unit is used to adaptively assign weights to different data sources; The knowledge graph integration unit is used to integrate prior knowledge such as urban planning and transportation policies; The data alignment and standardization unit is used to unify the format and scale of data from different sources.

4. The traffic flow prediction and optimization system based on intelligent optimization algorithm according to claim 1 is characterized by: The hybrid deep learning prediction module includes a CNN spatial feature extraction unit, a Transformer temporal dependency capture unit, a spatiotemporal attention mechanism unit and a multi-layer perceptron prediction unit connected in sequence; The CNN spatial feature extraction unit is used to capture the spatial dependency of the road network; The Transformer temporal dependency capture unit is used to learn long-term temporal dependencies; The spatiotemporal attention mechanism unit is used to dynamically adjust the importance of spatiotemporal features; The multi-layer perceptron prediction unit is used to finally generate a prediction result.

5. The traffic flow prediction and optimization system based on intelligent optimization algorithm according to claim 1 is characterized by: The intelligent parameter optimization module includes a quantum computing optimization unit, an improved bat algorithm unit, a dynamic neural architecture search unit and an adaptive search strategy unit which are sequentially connected; The quantum computing optimization unit is used to accelerate the search for the global optimal solution; The improved bat algorithm unit is used to implement adaptive parameter adjustment; The dynamic neural architecture search unit is used to automatically optimize the network structure; The adaptive search strategy unit is used to dynamically adjust the search behavior of the optimization algorithm.

6. The traffic flow prediction and optimization system based on intelligent optimization algorithm according to claim 1 is characterized by: The integrated learning module includes a base learner training unit, a dynamic weighted integration unit, an adversarial sample generation unit and a model robustness verification unit which are sequentially connected; The base learner training unit is used to train various models such as STAN, GBDT, and XGBoost; The dynamic weighted integration unit is used to adjust the model weight according to different spatiotemporal conditions; The adversarial sample generation unit is used to generate challenging test samples; The model robustness verification unit is used to evaluate the performance of the model under adversarial samples.

7. The traffic flow prediction and optimization system based on intelligent optimization algorithm according to claim 1 is characterized by: The multi-scale error correction module includes a Kalman filter short-term correction unit, a multi-scale residual learning unit and an adaptive online learning unit connected in sequence; The Kalman filter short-term correction unit is used to correct the short-term prediction error in real time; The multi-scale residual learning unit is used to simultaneously consider short, medium and long-term prediction errors; The adaptive online learning unit is used to update the model in real time to adapt to changes in traffic patterns.

8. The traffic flow prediction and optimization system based on intelligent optimization algorithm according to claim 1 is characterized by: The prediction result evaluation and feedback module includes a multi-index evaluation unit, a multi-scenario simulation verification unit, a human-machine collaborative feedback unit and a model dynamic adjustment unit connected in sequence; The multi-index evaluation unit is used to comprehensively evaluate the accuracy, stability and timeliness of the prediction; The multi-scenario simulation verification unit is used to test the model performance under extreme conditions in a virtual environment; The human-machine collaborative feedback unit is used to combine expert experience and machine learning to perform model tuning; The model dynamic adjustment unit is used to automatically optimize the model structure and parameters according to the evaluation results.

9. The traffic flow prediction and optimization system based on intelligent optimization algorithm according to claim 1 is characterized by: The edge computing and distributed learning module includes an edge-cloud collaborative learning scheduling unit, a federated learning protocol unit, a privacy protection computing unit, and a distributed model aggregation unit that are sequentially connected; The edge-cloud collaborative learning scheduling unit is used to optimize the distribution of computing tasks on edge devices and the cloud; The federated learning protocol unit is used to achieve secure sharing and collaborative modeling of multi-region data; The privacy protection computing unit is used to perform distributed computing under the premise of protecting data privacy; The distributed model aggregation unit is used to integrate multiple local models to generate a global model; The privacy protection computing unit adopts homomorphic encryption technology, and specifically includes: a data encryption subunit, a security computing subunit and a result decryption subunit connected in sequence; The data encryption subunit is used to encrypt the original data; The secure computing subunit is used to perform data analysis and model training on the encrypted domain; The result decryption subunit is used to securely decrypt the calculation result; The homomorphic encryption adopts the Paillier encryption algorithm, and the key length is not less than 2048 bits.

10. The method for traffic flow prediction and optimization based on intelligent optimization algorithm according to claim 1, characterized in that: The steps include: S1, collect traditional traffic data and non-traditional data, and perform denoising, anomaly detection, time series segmentation and dimensionality reduction preprocessing; S2, using attention mechanism and knowledge graph integration method to fuse multi-source data; S3. Build a hybrid deep learning prediction model combining CNN and Transformer to extract spatiotemporal features; use the following fusion algorithm formula for prediction: Y=α*FCNN(X1)+β*FTransformer(X2)+γ*FLSTM(X3)+δ*G(E) Among them, Y is the final prediction result, X1, X2, and X3 represent the input data of different modalities, FCNN, FTransformer, and FLSTM represent the prediction functions of CNN, Transformer, and LSTM, respectively, G(E) is the error correction function, and α, β, γ, and δ are dynamic weights determined by the attention mechanism; S4. Use quantum computing and improved bat algorithm to optimize model parameters and structure; Among them, the number of quantum bits optimized for quantum computing is set between 8 and 16, the population size of the improved bat algorithm is set between 20 and 50, and the number of iterations is 100 to 500; S5, dynamically integrate multiple base learners and enhance model robustness by generating adversarial samples; S6, use Kalman filtering and multi-scale residual learning to perform short, medium and long-term error correction; Among them, the time window for short-term error correction is set to 1-3 hours, medium-term to 6-12 hours, and long-term to 24-72 hours; S7. Verify the prediction performance in multiple scenarios and optimize the model through human-computer collaboration; S8, generate multi-level interpretation reports and interactive visualization results; S9, achieving edge-cloud collaboration and privacy-preserving distributed prediction based on federated learning; S10. Provide prediction services through a unified system management interface.

Images