Traffic prediction method, device, storage medium and electronic device

Abstract

The present invention relates to the field of traffic prediction technology, and specifically to a traffic prediction method, device, storage medium and electronic device. The method includes: obtaining source data from multiple data sources, the source data including traffic data and traffic-related data; stratifying and extracting features from the source data according to the time dimension and the space dimension, respectively, to obtain data features at multiple levels; using a cross-attention mechanism to align and complement the features of different source data, to obtain data after feature interaction; using a reinforcement learning algorithm to adjust the fusion weights of different source data, and fusing the source data based on the fusion weights to obtain fused data; based on multiple levels of data features, data after feature interaction and fused data, using a pre-trained traffic prediction model to perform traffic prediction. The present invention solves the problems of poor scene adaptability caused by simple stacking of multi-source data and static fusion rules, and improves the accuracy and reliability of prediction results.

Description

Technical Field

[0001] The present invention relates to the technical field of traffic prediction, and in particular to a traffic prediction method, device, storage medium and electronic equipment. Background Art

[0002] Existing traffic forecasting systems mostly rely on a single data source and traditional prediction models, making them difficult to cope with the complexity of multi-source, heterogeneous data and the dynamic changes in traffic. Specifically, a single data source often fails to capture the multidimensional characteristics of traffic fluctuations, resulting in low prediction accuracy. Traditional models also have limited performance when faced with volatile external factors.

[0003] In order to address the shortcomings of a single data source, multi-source data fusion has gradually been applied to traffic prediction. However, current multi-source data fusion is usually simply stacked, or static fusion rules are used to achieve data fusion. This fusion method has disadvantages such as poor scenario adaptability, which makes it difficult to achieve accurate traffic prediction in different scenarios. Summary of the Invention

[0004] In view of this, the present invention provides a traffic prediction method, device, storage medium and electronic device to solve the problem of poor scene adaptability of the data fusion method in the prior art.

[0005] In a first aspect, the present invention provides a traffic prediction method, which includes: obtaining source data from multiple data sources, the source data including traffic data and traffic-related data; stratifying and extracting features from the source data according to the time dimension and the spatial dimension, respectively, to obtain data features at multiple levels; using a cross-attention mechanism to align and complement the features of different source data to obtain data after feature interaction; using a reinforcement learning algorithm to adjust the fusion weights of different source data, and fusing the source data based on the adjusted fusion weights to obtain fused data; based on the data features at multiple levels, the data after feature interaction, and the fused data, using a pre-trained traffic prediction model to perform traffic prediction.

[0006] In the present invention, for the source data obtained from multiple data sources, the spatiotemporal feature decoupling can be used to deeply analyze the data from the time and space dimensions respectively; the advantages of different data sources can be complemented by feature alignment and complementation to enhance data quality; dynamic weight allocation can adapt to complex scene changes, solving the problems of poor scene adaptability caused by simple stacking of multi-source data and static fusion rules in traditional methods, and improving the accuracy and reliability of prediction results.

[0007] In an optional embodiment, the source data is stratified and feature extracted according to the time dimension and the spatial dimension respectively to obtain data features at multiple levels, including: dividing the source data into three time granularity layers according to second-level real-time data, minute-level dynamic data, and hourly-level trend data; using a long short-term memory network-temporal convolutional network to perform multi-scale feature extraction on the data of the three time granularity layers to obtain data features at multiple levels; dividing the source data into three levels of spatial data according to the road unit level, regional road network level, and urban macro level; using a graph attention mechanism to dynamically assign influence weights of different spatial data, and fusing features in the spatial data based on the influence weights.

[0008] In the present invention, by dividing the source data into three time granularity layers according to second-level real-time data, minute-level dynamic data and hour-level trend data, it is possible to accurately capture short-term changes in the source data and effectively grasp long-term trends, so that the model has stronger adaptability and robustness when facing changes in traffic conditions at different time scales; dividing the source data into three levels of spatial data according to the road unit level, regional road network level and urban macro level, it is possible to accurately analyze at the micro level, grasp regional coordination at the meso level, and optimize urban layout at the macro level. At the same time, through the dynamic weight distribution mechanism, data at different spatial levels can be more effectively integrated, thereby improving the accuracy and reliability of traffic flow prediction.

[0009] In an optional embodiment, a pre-trained traffic prediction model is used to perform traffic prediction based on data features at multiple levels, data after feature interaction, and data after fusion, including: extracting basic statistical features from traffic data; extracting time series features from traffic data, and aggregating, reducing dimensionality, and performing feature selection processing on the extracted time series features; using recurrent neural networks, convolutional neural networks, self-attention mechanisms, adversarial training, and autoencoders to extract deep-level features from traffic data; based on data features at multiple levels, data after feature interaction, basic statistical features, time series features, and deep-level features of fused data, a dynamic regression model is used to perform traffic prediction, and the parameters of the dynamic regression model are updated using a Kalman filter.

[0010] By extracting basic statistical features, time series features, and deep-level features, this method fully utilizes traffic data, improving the comprehensiveness and accuracy of traffic forecasts. The traffic forecast module utilizes a dynamic regression model, dynamically adjusting prediction parameters based on real-time data changes, and exhibits strong environmental adaptability. Furthermore, this dynamic regression model, combined with techniques such as Kalman filtering, enables adaptive parameter adjustment, ensuring high-precision forecasts in complex environments.

[0011] In an optional embodiment, when the flow prediction is traffic flow prediction, the pre-trained flow prediction model is a lightweight model deployed in the roadside unit. The lightweight model serves as a student model and is trained using a teacher model deployed in the cloud combined with a knowledge distillation algorithm.

[0012] In this paper, a lightweight model is deployed in the roadside unit and trained using a cloud-based teacher model combined with a knowledge distillation algorithm. This lightweight model can significantly reduce model size and computational complexity while retaining the key features and performance of a large model. Furthermore, its deployment in the roadside unit enables edge computing, improving prediction speed.

[0013] In an optional embodiment, the traffic prediction model also includes a global model obtained after multiple parameter aggregation and optimization using a cross-regional federated learning platform, and the parameters of the global model are updated in the following manner: obtaining the encrypted model parameters uploaded by each node; dynamically weighting and aggregating the model parameters to obtain aggregated parameters; and updating the parameters of the global model based on the aggregated parameters.

[0014] In the present invention, by establishing a cross-regional federated learning platform and constructing a global model, the model generalization ability can be improved while ensuring data privacy.

[0015] In an optional embodiment, when the flow prediction is a traffic flow prediction, the method further includes: constructing a digital twin of the area to be predicted; injecting source data from multiple data sources into the digital twin to simulate the scene and obtain simulation results; and adjusting the fusion weights and parameters of the flow prediction model based on the comparison between the simulation results and the flow prediction results of the corresponding scene.

[0016] In the present invention, a digital twin is constructed for simulation, and the fusion weights and parameters of the traffic prediction model are adjusted based on the comparison between the simulation results and the traffic prediction results of the corresponding scenarios, thereby significantly reducing the prediction error rate and improving the accuracy of traffic prediction.

[0017] In an optional embodiment, the method also includes: using a Bayesian deep learning model to process the parameters of the traffic prediction model to determine the confidence interval of the traffic prediction results output by the traffic prediction model; when there are contradictions in the source data of different data sources, using a Bayesian deep learning model in combination with a traffic result model to determine the confidence intervals of the prediction results corresponding to different data sources; and determining the credible data in the contradictory data source based on the confidence intervals of the prediction results corresponding to different data sources.

[0018] In the present invention, by providing a confidence interval, the reliability of the prediction results can be more intuitively understood. At the same time, the confidence interval can also assist in judging the credibility of the data source, thereby improving the accuracy of the data and further improving the accuracy of the model prediction.

[0019] In a second aspect, the present invention provides a traffic prediction device, which includes: a data acquisition module for acquiring source data from multiple data sources, the source data including traffic data and traffic-related data; a spatiotemporal stratification module for stratifying and extracting features from the source data according to the time dimension and the spatial dimension, respectively, to obtain data features at multiple levels; a feature interaction module for aligning and complementing features of different source data using a cross-attention mechanism to obtain data after feature interaction; a weight adjustment module for adjusting the fusion weights of different source data using a reinforcement learning algorithm, and fusing the source data based on the adjusted fusion weights to obtain fused data; a prediction module for performing traffic prediction using a pre-trained traffic prediction model based on data features at multiple levels, data after feature interaction, and fused data.

[0020] In a third aspect, the present invention provides a computer device comprising: a memory and a processor, the memory and the processor being communicatively connected to each other, computer instructions being stored in the memory, and the processor executing the traffic prediction method of the first aspect or any corresponding embodiment thereof by executing the computer instructions.

[0021] In a fourth aspect, the present invention provides a computer-readable storage medium having computer instructions stored thereon, the computer instructions being used to enable a computer to execute the traffic prediction method of the first aspect or any corresponding embodiment thereof.

[0022] In a fifth aspect, the present invention provides a computer program product, comprising computer instructions for causing a computer to execute the traffic prediction method of the first aspect or any corresponding embodiment thereof. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0024] Figure 1 is a flow chart of a flow prediction method according to an embodiment of the present invention;

[0025] Figure 2is a structural block diagram of a flow prediction device according to an embodiment of the present invention;

[0026] Figure 3 Schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0027] To make the purpose, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative efforts shall fall within the scope of protection of the present invention.

[0028] According to an embodiment of the present invention, an embodiment of a traffic prediction method is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0029] In this embodiment, a flow prediction method is provided, which can be used in electronic devices such as computers, mobile phones, tablet computers, etc. Figure 1 FIG. 1 is a flow chart of a method for traffic prediction according to an embodiment of the present invention. Figure 1 As shown, the process includes the following steps:

[0030] Step S101, source data from multiple data sources are obtained, the source data including flow data and flow-related data. In this embodiment, the flow prediction method is applied to the prediction of traffic flow. The multiple data sources include flow data, traffic signal data, traffic road data, and traffic accident data obtained by roadside sensors, cameras, radar devices, traffic lights, traffic management systems, mobile devices (such as smartphones or navigation devices), remote sensing satellite equipment, etc.; in addition, data affecting traffic, such as weather, and traffic-related data mentioned in social media, such as traffic accidents or large-scale events, can also be obtained.

[0031] The acquired source data can be preliminarily processed. For example, image data collected by cameras can be processed using techniques such as image processing to obtain data such as the traffic flow it monitors. Data acquired by sensors, radar devices, and other equipment can also be analyzed to determine the traffic flow it monitors, such as by converting it into a numerical vector or performing feature engineering (including adding timestamps or geographic locations). In addition, data from social media can be encoded using the BERT model, converting the data into a high-dimensional vector representation that contains rich semantic features. For example, feature vectors related to keywords such as traffic accidents and road construction mentioned in the data can be extracted.

[0032] Step S102: The source data is stratified and feature extracted according to the time dimension and the spatial dimension, respectively, to obtain data features at multiple levels. When stratifying the source data according to the time dimension, data can be divided into data of different time granularities according to different time levels, and then feature extraction is performed on the different time granularities. When stratifying the source data according to the spatial dimension, data can be divided into data of different spatial structures according to the size of the space in which the source data resides.

[0033] In step S103, a cross-attention mechanism is used to align and complement the features of different source data to obtain data after feature interaction. Specifically, for different source data, different methods can be used to process them according to whether their data types are structured data or unstructured data to obtain corresponding feature vectors. For example, for structured data such as traffic flow and weather forecasts, corresponding numerical vectors can be obtained through conversion; while for data in social media, which are usually in text form, corresponding feature vectors can be obtained through encoding and other methods. The process of obtaining the feature vector can be performed after the source data is obtained in step S101. When performing feature interaction on the acquired feature vectors, a cross-attention mechanism is used to calculate the attention scores between the feature vectors of different source data, which are then converted into attention weights, and feature alignment and complementation are performed by weighting or splicing.

[0034] For example, for sensor data and social media data, operations such as dot products and linear transformations are used to calculate the degree of correlation between each feature vector in the sensor data and each feature vector in the social media text data, resulting in an attention matrix. The attention matrix is normalized using the softmax function, converting the attention scores into attention weights. These weights range from 0 to 1, and for each element in the sensor data, the sum of its attention weights with all elements in the social media text data is 1. Thus, each weight represents the degree of attention a particular element in the sensor data pays to the corresponding element in the social media text data. Based on the calculated attention weights, a weighted sum is performed on the feature vectors of the social media text data. For example, for a particular element in the sensor data, the feature vectors of each token in the corresponding social media text data are weighted and added together according to the attention weights to produce a new feature vector aligned with the sensor data element. This new feature vector incorporates key information related to the element in the social media text data, achieving feature-level alignment between the two modalities. The weighted social media text feature vector is then concatenated or added to the original traffic sensor data feature vector. For example, concatenation involves concatenating the two vectors by dimension to form a new fused feature vector. In this way, the new feature vector not only contains the information of the traffic sensor data itself, but also incorporates the related information in the social media text data, achieving feature complementarity.

[0035] For example, when social media mentions a traffic accident in a certain area, the cross-attention mechanism is used to integrate the semantic features related to the accident into the real-time traffic speed data features of the corresponding area. This allows the fused data to not only reflect the current traffic speed, but also reflect the accident factors that may affect the speed, providing richer and more accurate feature information for subsequent tasks such as traffic flow prediction.

[0036] In step S104, a reinforcement learning algorithm is used to adjust the fusion weights of the different source data, and the source data is fused based on the adjusted fusion weights to obtain fused data. When adjusting the fusion weights of the different source data, the reinforcement learning algorithm can be used to automatically adjust the weights based on environmental sensitivity, using the scene parameters in the source data as the state, the adjustment of the fusion weights of each data source as the action, and the accuracy of the traffic flow prediction as the reward.

[0037] Specifically, weather conditions (coding values for different weather conditions such as rain, snow, fog, and haze), event types (classification identifiers for traffic accidents, large-scale events, etc.), time period characteristics (identification of morning and evening peak or off-peak hours), and the current fusion weights of each data source are used as components of the state. The adjustment operation of the fusion weight of each data source, such as increasing or decreasing a fixed step value for each data source or adjusting according to a preset comparison method, is used as the action. The accuracy of traffic flow prediction, such as the error between the predicted value and the actual value, is used as the reward. After that, a reinforcement learning network is selected and trained based on the state, action, and reward. As the training continues, the reinforcement learning network will continue to learn and optimize, and gradually find the optimal data source fusion weight strategy in different scenarios. After determining the fusion weight, the data can be weighted fused based on the fusion weight.

[0038] For example, in heavy rain, through continuous trial and learning, the algorithm will gradually increase the weight of radar data, while reducing the weight of other data sources that may be less reliable in bad weather; during large-scale events, the weights of traffic sensor data, social media data, etc. in areas related to the event will be automatically adjusted based on factors such as the scale and location of the event to adapt to complex and changing traffic scenarios and continuously improve the accuracy of traffic flow forecasts.

[0039] Step S105: Based on multiple levels of data features, data after feature interaction, and fused data, a pre-trained traffic prediction model is used to perform traffic flow prediction. Specifically, the traffic prediction model can be trained using multiple levels of data features, data after feature interaction, and fused data obtained from historically acquired source data. This pre-trained traffic prediction model is then used to predict traffic flow at the next moment after obtaining the current moment's data.

[0040] The traffic prediction method provided by the embodiment of the present invention can deeply analyze the source data obtained from multiple data sources from the time and space dimensions through spatiotemporal feature decoupling; it can complement the advantages of different data sources and enhance data quality through feature alignment and complementarity; it can adapt to complex scene changes through dynamic weight allocation, solving the problems of poor scene adaptability caused by simple stacking of multi-source data and static fusion rules in traditional methods, and improving the accuracy and reliability of prediction results.

[0041] In this embodiment, a flow prediction method is provided, which includes the following steps:

[0042] Step S201: Acquire source data from multiple data sources, including flow data and flow-related data. Figure 1 Step S101 of the illustrated embodiment will not be described in detail here.

[0043] In step S202 , the source data is layered and features are extracted according to the time dimension and the space dimension to obtain data features at multiple levels.

[0044] Specifically, the above step S202 includes:

[0045] Step S2021, divide the source data into three time granularity layers according to second-level real-time data, minute-level dynamic data, and hour-level trend data; among them, second-level real-time data may include data such as traffic light status obtained by traffic lights; minute-level dynamic data includes data such as vehicle trajectories obtained by GPS devices; hour-level trend data includes data such as historical traffic patterns, for example, the traffic volume is large in a certain period of time and the traffic volume is small in other periods of time.

[0046] In step S2022, a long short-term memory network-temporal convolutional network is used to perform multi-scale feature extraction on the data of the three time granularity layers to obtain data features at multiple levels. Specifically, the data of the three time granularity layers obtained by stratification are first normalized and then input into a long short-term memory network (LSTM). LSTM has memory units and a gating mechanism that can effectively capture long-term dependencies in time series. For real-time data at the second level, LSTM can remember the rapid changes in the state of traffic lights, such as multiple switching of traffic lights in a short period of time, and the immediate impact of these changes on vehicle flow. When processing minute-level dynamic data, LSTM can integrate the movement trajectory information of floating vehicles within a few minutes, analyze the vehicle's driving direction, speed change trend, etc., and then infer the real-time congestion status of the road. For hourly trend data, LSTM can learn the changing patterns of traffic flow over a longer period of time, such as the peak and trough patterns of traffic flow during specific periods of time on weekdays. When processing data at different time granularities, LSTM updates memory cells based on the data's chronological order, effectively integrating information from different time steps and outputting results that include long-term dependency features. These results contain key information about traffic flow changes at different time granularities, such as the growth or decline in traffic flow on a particular road section over several consecutive hours, and the similarity of traffic flow within the same time period on different dates.

[0047] The LSTM output is input into the TCN (Temporal Convolutional Network) layer. Based on a convolutional neural network architecture, the TCN applies convolution operations to time series data, effectively extracting local time series features. The TCN performs convolution operations by sliding convolution kernels of varying sizes across the time dimension. For second- and minute-level data, smaller kernels can capture local feature changes over short periods of time, such as subtle fluctuations in vehicle speed within seconds after a traffic light status change, or a sharp change in the speed of a floating vehicle within a minute. For hourly data, larger kernels can focus on local trends over longer periods of time, such as local peaks and valleys in traffic flow over several hours. Through convolution operations, the TCN automatically extracts local features at different time granularities. These features reflect short-term changes and fluctuations in traffic flow at different time scales. The TCN also uses causal convolution to ensure that future information is not used in predictions, consistent with practical time series prediction logic.

[0048] The local features at different time granularities output by the TCN layer are fused with the long-term dependency features output by the LSTM. Through a simple concatenation operation, the feature vectors output by the LSTM and TCN are concatenated dimensionally to form a comprehensive feature vector containing features at multiple time scales. This comprehensive feature vector encompasses both long-term trends in traffic flow (such as hourly traffic patterns) and short-term local fluctuations (such as the effects of traffic lights and vehicle speed changes at the second and minute levels). The fused feature vector can be used as input to subsequent traffic flow prediction models for traffic flow forecasting. Because these features integrate information at multiple time scales, they can more comprehensively reflect the changing patterns of traffic flow, thereby improving the accuracy and reliability of predictions.

[0049] In step S2023, the source data is divided into three levels of spatial data according to the road unit level, regional road network level, and urban macro level; among them, the road unit level data includes data such as sensors, cameras, and radars at a single intersection; the regional road network level focuses on the road network within a certain area, and the road topology structure is modeled through a graph neural network. Roads are abstracted as nodes of the graph, and the connection relationships between roads (such as intersection connections and road section connectivity) are used as edges to construct a graph structure. The graph neural network uses the data of nodes and edges (including source data from various data sources) to learn the traffic flow transmission rules and mutual influence relationships between roads in the region. For example, the flow distribution pattern of main roads and secondary roads during busy traffic hours, and the scope and degree of impact of congestion on surrounding roads when a certain road section is congested.

[0050] The urban macro-level analyzes the overall city landscape and primarily includes satellite remote sensing and mobile phone signaling data. Satellite remote sensing data can capture information such as the macro layout of urban road networks and land use types, while mobile phone signaling data can reveal urban residents' travel activities and population mobility trends. Combining these two data allows analysis of the intensity of traffic connections between different urban areas, tidal traffic phenomena, and the impact of urban functional zones (such as commercial, residential, and industrial areas) on traffic flow.

[0051] In step S2024, a graph attention mechanism is used to dynamically assign influence weights to different spatial data sets and fuse the features within the spatial data based on the influence weights. Specifically, the graph attention mechanism calculates the weights of different spatial data sets and, based on the calculated attention weights, performs a weighted fusion of features at different levels. For road unit-level nodes, their own features are weightedly added to those of regional road network-level and city-level nodes to generate the fused features.

[0052] Step S203: Use the cross-attention mechanism to align and complement the features of different source data to obtain data after feature interaction. Figure 1 Step S103 of the illustrated embodiment will not be described in detail here.

[0053] Step S204: Use a reinforcement learning algorithm to adjust the fusion weights of different source data, and fuse the source data based on the adjusted fusion weights to obtain fused data. Figure 1 Step S104 of the illustrated embodiment will not be described in detail here.

[0054] In step S205, traffic prediction is performed using a pre-trained traffic prediction model based on the data features at multiple levels, the data after feature interaction, and the fused data. In addition to fusing the traffic data and traffic-related data in the aforementioned steps, relevant features can be further extracted from the acquired traffic data. This traffic data includes not only current traffic data but also historical traffic data.

[0055] Furthermore, before extracting features from traffic data, the acquired traffic data can be preprocessed. This preprocessing process specifically involves cleaning the data to remove outliers (sudden increases in traffic, empty changes), missing values (data that only contains time but no actual values), and noisy data to ensure data reliability. For example, traffic bursts caused by system failures or sensor errors can be eliminated. To facilitate subsequent analysis and feature extraction, the data can be normalized or standardized to convert data of varying magnitudes to a unified scale. This means scaling all data to the [0, 1] range. Traffic data can also be divided into specific time windows (such as per second, per minute, or per hour) to facilitate subsequent statistics and analysis. For example, traffic data can be divided by minute to calculate the sum or average of traffic per minute.

[0056] Specifically, the above step S205 includes:

[0057] Step S2051 extracts basic statistical features from the traffic data. Basic statistical analysis can reveal traffic trends, fluctuations, and periodicity, providing a preliminary understanding of traffic variation patterns. In this embodiment, the basic statistical features extracted include mean, variance, standard deviation, extreme values, and median.

[0058] Specifically, the mean of traffic data It represents the average traffic level over a period of time and is calculated using the following formula:

[0059]

[0060] Where, Reflects the overall traffic trend of the system, Q i Represents the flow at each moment, and n represents the time.

[0061] Variance D(Q) and standard deviation reflect the volatility and dispersion of traffic data. Variance can measure the severity of traffic changes, and standard deviation is the square root of variance, which is used to describe the dispersion of traffic data. Variance is calculated using the following formula:

[0062]

[0063] Min / Max extracts the minimum and maximum values of traffic data, reflecting the range of traffic. Median is the middle value of sorted data and is particularly suitable for traffic data with extreme values, as it can prevent them from affecting the average.

[0064] Step S2052 extracts time series features from the traffic data and performs aggregation, dimensionality reduction, and feature selection on the extracted time series features. Since traffic data is typically represented as time series data, time series extraction can be used to extract temporal features of the traffic data. In this embodiment, the extracted time series features include trend, seasonality, autocorrelation, volatility, kurtosis, and skewness.

[0065] Trend features include extracting long-term traffic trends using sliding averages or time series decomposition. Trend features help identify the direction of traffic changes over time. Traffic often exhibits periodicity, such as daily, weekly, or monthly fluctuations. Using Fourier transforms or time series decomposition methods, periodic (seasonal) characteristics can be identified to help predict peak and trough traffic periods. For autocorrelation, the autocorrelation function can be used to measure the correlation between different time points in the traffic series. The ACF plot can be used to determine whether traffic exhibits significant periodicity or lag effects.

[0066]

[0067] Where k represents the lag time step.

[0068] Volatility can be measured using variance or moving standard deviation. Larger volatility indicates frequent changes in traffic, and the system's response strategy can be adjusted accordingly. Kurtosis reflects the peakedness of the traffic distribution, while skewness reflects the asymmetry of the traffic distribution. These characteristics can further understand the distribution pattern of traffic data. In other words, the rate of change of traffic can be determined based on the slope. Large kurtosis (large slope) indicates extreme fluctuations in the traffic data. Positive skewness indicates a long-tailed distribution, while negative skewness indicates a left-leaning distribution.

[0069] Alternatively, you can use the Z-score method to detect traffic peaks using a threshold method or an algorithm based on local extreme values. These abnormal traffic flows may correspond to certain key events. The Z-score is specifically expressed using the following formula:

[0070]

[0071] Where Z t is the Z-score value of the traffic data at time t; x t is the actual flow data at time t; μ is the mean of the flow data; σ is the standard deviation of the flow data. Through the Z-score value, the flow values of different time periods can be standardized so that they can be directly compared. According to different time characteristics, the corresponding Z-score value is adjusted. When the Z-score value is greater than a certain threshold, it is regarded as an abnormal point or flow peak. For example, when Zt >3, it can be considered that the traffic at this moment t has an abnormal peak. t When <-3, it can be considered that the t flow at that moment has an abnormally low value.

[0072] If the extracted time series features contain a large number of dimensions and features, more representative features can be extracted through aggregation and dimensionality reduction techniques. For example, principal component analysis can be used to reduce the dimensionality of high-dimensional traffic feature data to a low-dimensional space. By reducing the number of features, the model's training time and computing resource requirements are reduced; important information is retained, redundant features are removed, and the analysis process is simplified. Alternatively, recursive feature elimination can be used to select the most helpful features for prediction from a large number of features, reducing feature redundancy. Recursive feature elimination (RFE) is an algorithm that selects the optimal feature subset by gradually removing unimportant features. It is applicable to a variety of models and application scenarios.

[0073] Among them, the recursive feature elimination algorithm specifically adopts the following steps to perform feature selection:

[0074] 1: Train the model.

[0075] Using all features, a specified model is trained through a support vector machine, and the importance score or weight of each feature is obtained through the model.

[0076] 2: Evaluate the importance of features.

[0077] Based on the model's output, evaluate the importance of each feature. For example, in a support vector machine model, the feature weight coefficient can be used as a measure of feature importance. Step 1 directly outputs the feature weight coefficient.

[0078] 3: Remove the feature with the smallest contribution.

[0079] Based on the evaluation results, remove the features with the lowest importance score (usually the features with the smallest weight). This step will remove the features that have the least impact on the model, thereby gradually reducing the number of features.

[0080] 4: Repeat recursion.

[0081] After each feature is removed, the above steps are repeated using the remaining features, retraining the model and evaluating feature importance. This process is repeated recursively until the preset number of features is reached.

[0082] 5: Select the optimal feature set.

[0083] When RFE stops, the features that remain are the subset of features that the model considers most important. These features can be used to build a simplified model, improving computational efficiency and generalization capabilities.

[0084] Step S2053 uses recurrent neural networks, convolutional neural networks, self-attention mechanisms, adversarial training, and autoencoders to extract deep features from the traffic data. After completing basic statistical analysis and extracting time series features, deep learning methods can be used to further explore the complex characteristics of the traffic data, including nonlinear relationships, long-term dependencies, and underlying patterns. Deep features can be extracted by processing the extracted time series features.

[0085] Specifically, the extraction of the deep features can be achieved using the following process:

[0086] 1. Data preprocessing and feature engineering.

[0087] Normalize or standardize time series features to improve model training stability. Use a sliding window approach to construct time series inputs to enhance the model's ability to learn long-term dependencies. Perform feature selection to remove redundant features, ensuring the model focuses on the most critical variables. This process can be achieved through recursive feature elimination in step S204 and will not be further described here.

[0088] 2. Model selection and training.

[0089] Recurrent neural networks (RNNs) and their variants (such as LSTMs and GRUs) are used to extract features from processed time series features. They can learn long-term temporal dependencies and capture the dynamic characteristics of traffic changes. Convolutional neural networks (CNNs) extract local temporal features from time series features. This convolutional neural network can be combined with a recurrent neural network to form a CNN-LSTM model, improving feature extraction capabilities. The self-attention mechanism (Transformer) is used to extract deep features and efficiently capture long-term dependencies. Adversarial training and autoencoders can identify extreme traffic spikes or abnormal patterns.

[0090] 3. Model optimization and tuning.

[0091] For models used to extract deep features, optimization algorithms (such as Adam and SGD) can be used to adjust model weights and improve convergence speed. Model performance can also be optimized through hyperparameter tuning (such as learning rate, number of hidden layers, and activation function selection). Furthermore, data augmentation and regularization (such as Dropout) can be used to prevent overfitting and improve generalization.

[0092] 4. Model application and feature interpretation.

[0093] When using the above model to extract deep features, the model output can be analyzed through explainable AI methods (such as SHAP (SHapley Additive exPlanations) and LIME) to understand the key influencing factors.

[0094] In step S2054, a dynamic regression model is used to predict traffic based on data features at multiple levels, data after feature interaction, basic statistical features of the fused data, time series features, and deep-level features. The parameters of the dynamic regression model are updated using a Kalman filter. The dynamic regression model can adjust the regression parameters in real time according to environmental factors and changes in historical data to adapt to traffic fluctuations. The model can use dynamic parameter estimation with a feedback mechanism, such as a Kalman filter, to update the model parameters in real time to ensure high accuracy of the prediction results. The specific prediction and parameter adjustment process of the dynamic regression model can be implemented with reference to relevant technologies and will not be described in detail here.

[0095] In this embodiment, a flow prediction method is provided, which includes the following steps:

[0096] Step S301: Acquire source data from multiple data sources. The source data includes flow data and flow-related data. Figure 1 Step S101 of the illustrated embodiment will not be described in detail here.

[0097] Step S302: stratify and extract features from the source data according to the time dimension and the spatial dimension to obtain data features at multiple levels. Figure 1 Step S102 of the illustrated embodiment will not be described in detail here.

[0098] Step S303: Use the cross-attention mechanism to align and complement the features of different source data to obtain data after feature interaction. Figure 1 Step S103 of the illustrated embodiment will not be described in detail here.

[0099] Step S304: Use the reinforcement learning algorithm to adjust the fusion weights of different source data, and fuse the source data based on the adjusted fusion weights to obtain fused data. Figure 1 Step S104 of the illustrated embodiment will not be described in detail here.

[0100] Step S305 , based on the data features of multiple levels, the data after feature interaction and the fused data, a pre-trained traffic prediction model is used to perform traffic prediction.

[0101] In an optional implementation, when traffic flow prediction is used, the pre-trained traffic prediction model is a lightweight model deployed in a roadside unit (RSU). This lightweight model serves as a student model and is trained using a teacher model deployed in the cloud combined with a knowledge distillation algorithm. Specifically, the RSU, as a frontier node for traffic data collection, can acquire local traffic information such as traffic sensor data and camera image data in real time. These lightweight models are designed to run on resource-constrained edge devices and feature efficient computing and storage.

[0102] The knowledge distillation algorithm transfers knowledge learned by a complex large cloud model (the teacher model) to a simpler, lightweight model (the student model). This allows the lightweight model to retain the key features and performance of the large model while significantly reducing its size and computational complexity. For example, a lightweight model can be compressed to a size of 5MB.

[0103] The lightweight model of the edge computing layer (roadside unit) can achieve localized real-time prediction within 200ms. Taking traffic scenarios as an example, through real-time collection of intersection vehicle flow, speed and other data, the model can quickly predict traffic flow changes in the next time period, providing timely decision support for the intelligent control of traffic lights. For example, when it is discovered that a certain intersection is about to be congested, the traffic light duration can be adjusted in time to alleviate traffic pressure. In addition, MobileNetv2 can also be deployed in the roadside unit to optimize the YOLO detection network to process image data collected by the camera. This detection network uses the lightweight architecture of MobileNetv2 to improve the YOLO detection network, while maintaining 95% accuracy while reducing 60% computing power consumption, making the model more suitable for running on edge devices, ensuring fast and accurate detection of traffic targets (such as vehicles, pedestrians, etc.), and providing reliable data for real-time prediction.

[0104] In an optional embodiment, the traffic prediction model also includes a global model obtained after multiple parameter aggregation and optimization using a cross-regional federated learning platform, and the parameters of the global model are updated in the following manner: obtaining the encrypted model parameters uploaded by each node; dynamically weighting and aggregating the model parameters to obtain aggregated parameters; and updating the parameters of the global model based on the aggregated parameters.

[0105] Specifically, a cross-regional federated learning platform can be set up in the cloud, which has functions such as encrypted communication, parameter aggregation, and model management. The platform connects multiple city nodes, and each city node conducts joint model training locally without sharing the original data. The parameters of the training model are shared on the platform through encrypted sharing technology. The platform dynamically weights and aggregates the shared parameters and obtains a global model based on the aggregated parameters. This global model is then fed back to each participant based on the platform to provide them with prediction services. At the same time, the participants continue to train and optimize the model based on local data, upload the new parameters back to the platform, and further optimize the global model, forming a continuous cycle of optimization.

[0106] As a result, the global model can adapt to traffic conditions in different cities, improving its prediction accuracy across different regions. Tests have shown that this mechanism can improve cold-start prediction accuracy in newly connected cities by 40%. When new cities lack sufficient local data for full training, they can quickly obtain a relatively accurate initial model by leveraging the global model on the federated learning platform and the parameters shared with other cities. This reduces cold-start prediction errors and enables more effective prediction and management of local traffic flows.

[0107] Step S306: construct a digital twin of the area to be predicted; inject source data from multiple data sources into the digital twin to simulate the scene and obtain simulation results; adjust the fusion weights and parameters of the traffic prediction model based on the comparison between the simulation results and the traffic prediction results of the corresponding scene.

[0108] The area to be predicted can be part of the roads in the city or the entire city, and this embodiment does not specifically limit this. Specifically, when building a digital twin, it is necessary to accurately draw the road network, covering the shape of each street and intersection, the number of lanes, speed limit and other information; simulate traffic facilities, such as the timing rules of traffic lights, the location and operating hours of bus stops; reflect the driving characteristics of vehicles, including parameters such as speed and acceleration of different types of vehicles; and even consider the travel behavior and impact of pedestrians. Through these precise modeling, a virtual environment that is highly similar to real urban traffic is created in the virtual space.

[0109] Next, source data from multiple data sources, including traffic data and traffic-related data, is injected into the constructed digital twin. Various traffic scenarios are then simulated within the digital twin using the injected historical data. Furthermore, data from the simulated traffic scenarios is fed into the traffic prediction model to perform traffic forecasts. The forecast results are compared with the simulation results to evaluate the accuracy of the traffic prediction model. The accuracy of the fusion mechanism and the prediction model is evaluated, and model parameters and fusion strategies are continuously adjusted. For example, if a data source is found to be insufficiently contributing to the prediction results in a specific scenario, its weight in the fusion mechanism can be adjusted. For the prediction model, the number of layers or nodes in the neural network can be changed, or hyperparameters such as the learning rate can be adjusted. Through repeated adjustments and testing, the prediction results are gradually optimized. For example, for a specific city, by constructing a digital twin and injecting data for simulation, the fusion weights and traffic prediction model parameters were adjusted, resulting in a reduction in the prediction error rate from 12.7% to 7.3%.

[0110] Step S307: Use the Bayesian deep learning model to process the parameters of the traffic prediction model to determine the confidence interval of the traffic prediction results output by the traffic prediction model; when there are contradictions in the source data of different data sources, use the Bayesian deep learning model in combination with the traffic result model to determine the confidence intervals of the prediction results corresponding to the different data sources; determine the credible data in the contradictory data source based on the confidence intervals of the prediction results corresponding to the different data sources.

[0111] Specifically, Bayesian deep learning combines Bayesian methods and deep learning. It can describe the uncertainty of model parameters through probability distribution and output confidence intervals for prediction results. When determining the confidence interval, the Bayesian deep learning module first determines the prior distribution of model parameters through probability distribution estimation. It then calculates the posterior distribution using Bayes' theorem based on the prior distribution. Monte Carlo sampling is then used to perform multiple sampling from the posterior distribution of the parameters. The traffic prediction model then predicts traffic based on the parameter values from each sampling. Finally, the confidence interval is calculated based on the distribution of prediction results corresponding to the parameter values from the multiple samplings. The confidence interval of the prediction results can help decision makers understand the reliability of the prediction results. For example, when formulating traffic control strategies, if the confidence interval of the predicted traffic is wide, decision makers will need to consider more response options.

[0112] Furthermore, when data from multiple data sources conflict, such as when sensors indicate unimpeded traffic but social media reports indicate an incident, the data from each source can be fed into a traffic prediction model equipped with a Bayesian deep learning module. The confidence intervals of the predictions from different data sources are observed. The credibility of the data source is determined based on these confidence intervals. In addition to using the Bayesian deep learning module, manual verification of the credibility of the data source and processing of the data source can also be conducted based on practical experience.

[0113] This embodiment also provides a flow prediction device for implementing the above-mentioned embodiments and preferred implementations. Details already described will not be repeated. As used below, the term "module" may refer to a combination of software and / or hardware that implements a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, implementation using hardware, or a combination of software and hardware, is also possible and contemplated.

[0114] This embodiment provides a flow prediction device, such as Figure 2 As shown, including:

[0115] A data acquisition module 21 is used to acquire source data from multiple data sources, where the source data includes flow data and flow-related data;

[0116] The spatiotemporal stratification module 22 is used to stratify and extract features from the source data according to the time dimension and the spatial dimension, and obtain data features at multiple levels;

[0117] A feature interaction module 23 is used to align and complement the features of different source data using a cross-attention mechanism to obtain feature-interacted data;

[0118] The weight adjustment module 24 is used to adjust the fusion weights of different source data using a reinforcement learning algorithm, and fuse the source data based on the adjusted fusion weights to obtain fused data;

[0119] The prediction module 25 is used to perform traffic prediction using a pre-trained traffic prediction model based on data features at multiple levels, data after feature interaction, and fused data.

[0120] The further functional description of each of the above modules is the same as that of the above corresponding embodiments and will not be repeated here.

[0121] The embodiment of the present invention also provides a computer device having the above Figure 2 The flow prediction device shown.

[0122] See also Figure 3 , Figure 3 is a structural diagram of a computer device provided by an optional embodiment of the present invention, such as Figure 3As shown, the computer device includes: one or more processors 10, memory 20, and interfaces for connecting various components, including high-speed interfaces and low-speed interfaces. Various components utilize different buses to communicate with each other and can be installed on a common mainboard or installed in other ways as needed. The processor can process the instructions executed in the computer device, including instructions stored in the memory or on the memory to display the graphical information of the GUI on an external input / output device (such as, a display device coupled to the interface). In some optional embodiments, if necessary, multiple processors and / or multiple buses can be used together with multiple memories and multiple memories. Equally, multiple computer devices can be connected, and each device provides part of the necessary operations (for example, as a server array, a group of blade servers, or a multi-processor system). Figure 3 A processor 10 is taken as an example.

[0123] The processor 10 may be a central processing unit, a network processor, or a combination thereof. The processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit, a programmable logic device, or a combination thereof. The programmable logic device may be a complex programmable logic device, a field programmable gate array, a general purpose array logic, or any combination thereof.

[0124] The memory 20 stores instructions that can be executed by at least one processor 10, so as to enable at least one processor 10 to execute the method shown in the above embodiment.

[0125] The memory 20 may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created based on the use of a computer device for displaying a small program landing page, etc. In addition, the memory 20 may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some optional embodiments, the memory 20 may optionally include a memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0126] The memory 20 may include a volatile memory, such as a random access memory; the memory may also include a non-volatile memory, such as a flash memory, a hard disk or a solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0127] The computer device further includes a communication interface 30 for the computer device to communicate with other devices or a communication network.

[0128] The embodiment of the present invention also provides a computer-readable storage medium. The above-mentioned method according to the embodiment of the present invention can be implemented in hardware, firmware, or implemented as a computer code that can be recorded in a storage medium, or implemented as a computer code that is originally stored in a remote storage medium or a non-temporary machine-readable storage medium and downloaded through a network and will be stored in a local storage medium, so that the method described herein can be stored in such software processing on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only storage memory, a random access memory, a flash memory, a hard disk or a solid-state drive, etc.; further, the storage medium can also include a combination of the above-mentioned types of memory. It can be understood that a computer, a processor, a microprocessor controller or programmable hardware includes a storage component that can store or receive software or computer code. When the software or computer code is accessed and executed by a computer, a processor or hardware, the method shown in the above embodiment is implemented.

[0129] A portion of the present invention may be applied as a computer program product, such as a computer program instruction, which, when executed by a computer, can call or provide the method and / or technical solution according to the present invention through the operation of the computer. Those skilled in the art should understand that the form in which the computer program instruction exists in a computer-readable medium includes, but is not limited to, a source file, an executable file, an installation package file, etc. Accordingly, the way in which the computer program instruction is executed by the computer includes, but is not limited to: the computer directly executes the instruction, or the computer compiles the instruction and then executes the corresponding compiled program, or the computer reads and executes the instruction, or the computer reads and installs the instruction and then executes the corresponding installed program. Here, the computer-readable medium may be any available computer-readable storage medium or communication medium that can be accessed by the computer.

[0130] Although the embodiments of the present invention have been described with reference to the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention. Such modifications and variations are all within the scope defined by the appended claims.

Claims

1. A flow prediction method, characterized in that: The method comprises: Acquire source data from multiple data sources, the source data including flow data and flow-related data; The source data is layered and features are extracted according to the time dimension and the space dimension respectively to obtain data features at multiple levels; A cross-attention mechanism is used to align and complement the features of different source data to obtain data after feature interaction; The reinforcement learning algorithm is used to adjust the fusion weights of different source data, and the source data is fused based on the adjusted fusion weights to obtain fused data; Based on data features at multiple levels, data after feature interaction, and fused data, a pre-trained traffic prediction model is used to perform traffic prediction. The source data is layered and feature extracted according to the time dimension and the space dimension to obtain data features at multiple levels, including: Divide the source data into three time granularity layers according to second-level real-time data, minute-level dynamic data, and hour-level trend data; Long short-term memory network-temporal convolutional network is used to extract multi-scale features from data at three time granularity layers to obtain data features at multiple levels; Dividing the source data into three levels of spatial data according to road unit level, regional road network level and city macro level; A graph attention mechanism is used to dynamically assign influence weights of different spatial data, and features in the spatial data are fused based on the influence weights.

2. The method according to claim 1, characterized in that Based on data features at multiple levels, data after feature interaction, and fused data, a pre-trained traffic prediction model is used to perform traffic prediction, including: Extract basic statistical features from traffic data; Extract time series features from traffic data, and perform aggregation, dimensionality reduction, and feature selection on the extracted time series features; Use recurrent neural networks, convolutional neural networks, self-attention mechanisms, adversarial training, and autoencoders to extract deep features from traffic data; Based on data features at multiple levels, data after feature interaction, basic statistical features of fused data, time series features, and deep-level features, a dynamic regression model is used to predict traffic flow, and the parameters of the dynamic regression model are updated using a Kalman filter.

3. The method according to claim 1, characterized in that When the flow prediction is traffic flow prediction, the pre-trained flow prediction model is a lightweight model deployed in the roadside unit. The lightweight model serves as a student model and is trained using a teacher model deployed in the cloud combined with a knowledge distillation algorithm.

4. The method according to claim 1, wherein The traffic prediction model also includes a global model obtained through multiple parameter aggregation and optimization using a cross-regional federated learning platform. The parameters of the global model are updated in the following manner: Obtain the encrypted model parameters uploaded by each node; Performing dynamic weighted aggregation on the model parameters to obtain aggregated parameters; The parameters of the global model are updated based on the aggregated parameters.

5. The method according to claim 1, wherein When the flow prediction is a traffic flow prediction, the method further includes: Build a digital twin of the area to be predicted; Injecting source data from multiple data sources into the digital twin to perform scene simulation and obtain simulation results; Based on the comparison between the simulation results and the traffic prediction results of the corresponding scenarios, the fusion weights and the parameters of the traffic prediction model are adjusted.

6. The method according to claim 1, characterized in that The method further comprises: The Bayesian deep learning model is used to process the parameters of the traffic prediction model and determine the confidence interval of the traffic prediction results output by the traffic prediction model; When there are contradictions in the source data from different data sources, the Bayesian deep learning model is combined with the traffic result model to determine the confidence intervals of the prediction results corresponding to different data sources; The credible data in the conflicting data sources is determined based on the confidence intervals of the prediction results corresponding to different data sources.

7. A flow prediction device, characterized in that: The device comprises: A data acquisition module, configured to acquire source data from multiple data sources, wherein the source data includes flow data and flow-related data; A spatiotemporal stratification module is used to stratify and extract features from the source data according to the time dimension and the spatial dimension, respectively, to obtain data features at multiple levels; The feature interaction module is used to align and complement the features of different source data using a cross-attention mechanism to obtain data after feature interaction; A weight adjustment module is used to adjust the fusion weights of different source data using a reinforcement learning algorithm, and to fuse the source data based on the adjusted fusion weights to obtain fused data; The prediction module is used to predict traffic flow using a pre-trained traffic prediction model based on multiple layers of data features, data after feature interaction, and fused data; The source data is layered and feature extracted according to the time dimension and the space dimension to obtain data features at multiple levels, including: Divide the source data into three time granularity layers according to second-level real-time data, minute-level dynamic data, and hour-level trend data; Long short-term memory network-temporal convolutional network is used to extract multi-scale features from data at three time granularity layers to obtain data features at multiple levels; Dividing the source data into three levels of spatial data according to road unit level, regional road network level and city macro level; A graph attention mechanism is used to dynamically assign influence weights of different spatial data, and features in the spatial data are fused based on the influence weights.

8. A computer device, characterized in that: include: A memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the traffic prediction method according to any one of claims 1 to 6 by executing the computer instructions.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a computer to execute the traffic prediction method according to any one of claims 1 to 6.

Images