Traffic flow prediction method and device based on feature embedding and high-frequency filtering, and storage medium

Abstract

The embodiment of the invention relates to the field of artificial intelligence, and provides a traffic flow prediction method and device based on feature embedding and high-frequency filtering, and a storage medium, and the method comprises the steps: obtaining the original data of a to-be-predicted traffic flow, splicing and fusing the native fluctuation embedded data, the periodic embedded data and the dynamic similarity embedded data corresponding to the original data to obtain fused embedded data; performing filtering processing on the fused data according to a high-frequency filtering network, performing spatial feature extraction on the original data according to a graph convolutional neural network, performing spatial-temporal feature fusion on the time sequence feature data and the spatial feature data based on linear transformation and a residual connection principle, and performing spatial-temporal feature fusion on the time sequence feature data and the spatial feature data according to the fused time feature data. And predicting to obtain traffic flow prediction data associated with the original data. By adopting the method, the accuracy of traffic flow prediction can be improved, and the model prediction effect is improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a traffic flow prediction method, apparatus and storage medium based on feature embedding and high-frequency filtering. Background Technology

[0002] In the field of artificial intelligence technology, with the acceleration of global urbanization, the number of urban populations and motor vehicles has surged, leading to increasingly severe problems such as traffic congestion, safety hazards, and environmental pollution. Intelligent Transportation Systems (ITS), as a core technological approach to solving urban traffic problems, have become a key support for improving traffic efficiency and ensuring travel safety through real-time perception, intelligent decision-making, and dynamic control. Traffic flow prediction, as a core foundational technology of ITS, analyzes historical traffic data and establishes predictive models to provide a scientific basis for dynamic route planning, intelligent signal control, and traffic management decisions.

[0003] However, traditional traffic flow prediction models only represent time features through time-periodic embedding, resulting in a relatively simple data embedding and insufficient data representation. The models struggle to fully capture the complex patterns of the time dimension. Furthermore, traditional prediction methods have limited responsiveness to high-frequency fluctuations such as short-term traffic surges, leading to a significant increase in errors in short-term prediction scenarios. Secondly, traditional prediction methods often fuse spatiotemporal features through simple splicing or overlay, resulting in the loss of feature information during the fusion process and affecting the final prediction accuracy.

[0004] In summary, traditional traffic flow prediction techniques suffer from low accuracy and poor prediction performance. Summary of the Invention

[0005] This application provides a traffic flow prediction method, apparatus, and storage medium based on feature embedding and high-frequency filtering, which can improve the accuracy of traffic flow prediction and enhance the model prediction effect.

[0006] In a first aspect, embodiments of this application provide a traffic flow prediction method based on feature embedding and high-frequency filtering, comprising:

[0007] Obtain the original traffic flow data to be predicted, and then splice and fuse the original volatility embedded data, periodic embedded data, and dynamic similarity embedded data corresponding to the original data to obtain the fused embedded data.

[0008] The fused data is filtered using a high-frequency filtering network to obtain filtered time-series feature data.

[0009] Spatial features are extracted from the original data using a graph convolutional neural network to obtain spatial feature data associated with the original data;

[0010] Based on the principles of linear transformation and residual connection, the temporal feature data and the spatial feature data are fused to obtain the fused spatiotemporal feature data.

[0011] Based on the fused time feature data, predicted traffic flow data associated with the original data is obtained.

[0012] Secondly, embodiments of this application provide a traffic flow prediction device based on feature embedding and high-frequency filtering, which has the function of implementing the traffic flow prediction method based on feature embedding and high-frequency filtering provided in the first aspect above. The function can be implemented in hardware or by hardware executing corresponding software. The hardware or software includes one or more modules corresponding to the above function, and the modules can be software and / or hardware.

[0013] In one possible design, the device includes:

[0014] The embedded data fusion module is used to acquire the original data of traffic flow to be predicted, and to splice and fuse the original fluctuation embedded data, periodic embedded data and dynamic similarity embedded data corresponding to the original data to obtain the fused embedded data.

[0015] The filtering module is used to filter the fused data according to the high-frequency filtering network to obtain filtered time-series feature data.

[0016] The spatial feature extraction module is used to extract spatial features from the original data based on the graph convolutional neural network to obtain spatial feature data associated with the original data.

[0017] The spatiotemporal feature fusion module is used to perform spatiotemporal feature fusion on the time-series feature data and the spatial feature data based on the principles of linear transformation and residual connection, so as to obtain fused spatiotemporal feature data.

[0018] The prediction output module is used to predict traffic flow data associated with the original data based on the fused time feature data.

[0019] Another aspect of this application provides a traffic flow prediction device based on feature embedding and high-frequency filtering, which includes at least one connected processor and memory, wherein the memory is used to store program code, and the processor is used to call the program code in the memory to execute the methods described in the above aspects.

[0020] In another aspect, this application provides a computer storage medium including instructions that, when executed on a computer, cause the computer to perform the methods described in the above aspects.

[0021] In another aspect, this application provides a computer program product containing instructions that, when run on a computer, cause the computer to perform the methods described in the above aspects.

[0022] Compared to traditional traffic flow prediction methods, this approach combines the dynamic and inherent fluctuation characteristics of the original data to fully represent the data, facilitating the model's comprehensive capture of time-dimensional features. It enhances the response to high-frequency fluctuations such as short-term traffic surges, thereby improving accuracy in short-term prediction scenarios. It abandons the traditional method of simply splicing or overlaying spatiotemporal features, designing a targeted fusion mechanism to avoid the loss of feature information during the fusion process, thus improving prediction accuracy. This ultimately improves the accuracy of traffic flow prediction and enhances the model's prediction performance. Attached Figure Description

[0023] Figure 1 This is an application environment diagram from one embodiment;

[0024] Figure 2 This is a flowchart illustrating a traffic flow prediction method based on feature embedding and high-frequency filtering in one embodiment.

[0025] Figure 3 This is a schematic diagram of the time information embedding module in one embodiment;

[0026] Figure 4 This is a schematic diagram of the high-frequency filtering module in one embodiment;

[0027] Figure 5 This is a schematic diagram of the spatiotemporal feature fusion module in one embodiment;

[0028] Figure 6 This is a flowchart illustrating a traffic flow prediction method based on feature embedding and high-frequency filtering in another embodiment.

[0029] Figure 7 This is a schematic diagram of the overall process principle in one embodiment;

[0030] Figure 8 This is a block diagram of a traffic flow prediction device based on feature embedding and high-frequency filtering in one embodiment;

[0031] Figure 9 This is an internal structure diagram of a traffic flow prediction device based on feature embedding and high-frequency filtering in one embodiment;

[0032] Figure 10 This is an internal structure diagram of a traffic flow prediction device based on feature embedding and high-frequency filtering in another embodiment. Detailed Implementation

[0033] The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or modules is not necessarily limited to those explicitly listed, but may include other steps or modules not explicitly listed or inherent to these processes, methods, products, or devices. The division of modules in the embodiments of this application is merely a logical division; in actual applications, there may be other division methods. For example, multiple modules may be combined into or integrated into another system, or some features may be ignored or not performed. Additionally, the shown or discussed mutual coupling or direct coupling or communication connection may be through some interface, and the indirect coupling or communication connection between modules may be electrical or other similar forms, none of which are limited in the embodiments of this application. Furthermore, the modules or sub-modules described as separate components may or may not be physically separated, may or may not be physical modules, or may be distributed among multiple circuit modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the embodiments of this application.

[0034] Figure 1 As shown in the application environment diagram of one embodiment, this application provides a traffic flow prediction method based on feature embedding and high-frequency filtering, which can be applied to, for example... Figure 1 In the application scenario shown, terminal 102 communicates with server 104 via a network.

[0035] The terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, etc. The server 104 can be implemented using a standalone server or a server cluster composed of multiple servers.

[0036] It should be noted that the terminal 102 involved in this application embodiment can be a wired terminal or a wireless terminal. It can be a device that provides voice and / or data connectivity to a user, a handheld device with wireless connectivity, or other processing devices connected to a wireless modem. The wireless terminal can communicate with one or more core networks via a wireless access network. The wireless terminal can be a mobile terminal, such as a mobile phone (or "cellular" phone) or a computer with a mobile terminal. For example, it can be a portable, pocket-sized, handheld, computer-embedded, or vehicle-mounted mobile device that exchanges voice and / or data with the wireless access network. Examples include personal communication service telephones, cordless phones, Session Initiation Protocol (SIP) phones, wireless local loop stations, personal digital assistants, and other devices. Wireless terminals can also be referred to as systems, subscriber units, subscriber stations, mobile stations, mobile stations, remote stations, access points, remote terminals, access terminals, user terminals, terminal equipment, user agents, user devices, or user equipment.

[0037] Figure 2 This is a flowchart illustrating a traffic flow prediction method based on feature embedding and high-frequency filtering in one embodiment. The following refers to... Figure 2 This application introduces a traffic flow prediction method based on feature embedding and high-frequency filtering, which includes:

[0038] S201. Obtain the original traffic flow data to be predicted, and then splice and fuse the original data with the original volatility embedded data, periodic embedded data, and dynamic similarity embedded data to obtain the fused embedded data.

[0039] The original data is historical data reflecting the characteristics of traffic flow that is to be predicted. The original volatility embedded data, periodic embedded data, and dynamic similarity embedded data are three types of embedded data designed in this application in terms of time, which respectively represent the temporal characteristics of traffic flow in the original data from different dimensions. The fused embedded data is the data obtained by fusing these three types of embedded data.

[0040] Specifically, native volatility embeddings are used to preserve the original information in the original data, periodic embeddings are used to represent the periodic patterns of the data, and dynamic similarity embeddings are used to represent the similarity in the changing trends and volatility characteristics between adjacent data. More specifically, native volatility embeddings can be represented by Edata, periodic embeddings by Ep, and dynamic similarity embeddings by Ea.

[0041] S202, The fused data is filtered using a high-frequency filtering network to obtain filtered time-series feature data.

[0042] Among them, the high-frequency filtering network is a composite network model designed to perform high-frequency filtering. The filtered time-series feature data refers to the data obtained after filtering that contains the time-series features of the original data.

[0043] S203, Spatial features are extracted from the original data using a graph convolutional neural network to obtain spatial feature data associated with the original data.

[0044] Among them, Graph Convolutional Network (GCN) is a type of neural network used for feature extraction. Spatial feature data is the extracted data containing spatial features from the original data.

[0045] S204, based on the principles of linear transformation and residual connection, performs spatiotemporal feature fusion on time-series feature data and spatial feature data to obtain fused spatiotemporal feature data.

[0046] Specifically, linear transformation is used to adjust the dimension and numerical distribution of data features to adapt the data to the model's processing requirements; residual connection allows the data to skip some network layers and be directly added to the output of subsequent layers, avoiding gradient vanishing during deep network training, while preserving the original features and improving the model's learning efficiency and expressive power; the fused spatiotemporal feature data is the data obtained by fusing temporal feature data and spatial feature data.

[0047] S205, based on the fused time characteristic data, predictive data of traffic flow associated with the original data is obtained.

[0048] Traffic flow prediction data refers to data output through a network with prediction output capabilities, such as traffic flow predictions for future time steps.

[0049] Compared to traditional traffic flow prediction methods, this embodiment first acquires the original traffic flow data to be predicted, then concatenates and fuses native fluctuation-embedded data, periodic embedding data, and dynamic similarity embedding data. Next, it filters the fused data using a high-frequency filtering network, and then extracts spatial features from the original data using a graph convolutional neural network to obtain spatial feature data. Finally, it fuses temporal feature data and spatial feature data to obtain fused spatiotemporal feature data, which is then used to predict traffic flow. This embodiment combines the dynamic features and native fluctuation features of the original data, providing a comprehensive representation of the data and facilitating the model's capture of time-dimensional features. It improves the response to high-frequency fluctuations such as short-term traffic surges, thereby enhancing accuracy in short-term prediction scenarios. It abandons the traditional method of simply concatenating or superimposing spatiotemporal features, designing a targeted fusion mechanism to avoid the loss of feature information during the fusion process, thus improving prediction accuracy. This enhances the accuracy of traffic flow prediction and improves the model's prediction performance.

[0050] Optionally, in some embodiments of this application, the method further includes: obtaining data corresponding to a preset time step interval in the original data; performing a linear transformation on the data corresponding to the preset time step interval based on a linear layer to obtain native volatility embedded data.

[0051] The preset time step interval is the time step interval from the starting time step of historical data to the starting time step of prediction; the linear layer is a network layer used for linear transformation, which can be denoted as Linear.

[0052] For example, in order to preserve the native information in the original data, this application utilizes a linear layer to obtain the native volatility embedding data, which can be denoted as: The mathematical expression of its calculation process can be: Where df represents the dimension of feature embedding, Linear represents linear transformation, X represents the data at multiple time steps of the input; t represents the current time step, which is the prediction start time point. Based on historical data at time t and before, future traffic flow is predicted; t-T+1 represents the starting time step of the historical data, which is the first time step of the input sequence; T represents the time step length of the historical data used for prediction, which is the length of the input sequence. This represents traffic flow data for T consecutive time steps from time step t-T+1 to t.

[0053] In this embodiment, the original information of the original data is preserved by linear layer transformation, which effectively overcomes the shortcomings of traditional feature processing that easily loses the original fluctuation characteristics, improves the accuracy of data embedding in representing the original fluctuations, and provides basic features that are more in line with the characteristics of the original data for subsequent prediction tasks.

[0054] Optionally, in some embodiments of this application, the method further includes: determining the initial embedded data as learnable day-number embedded data and time-step embedded data within a day; determining the index data as day-number sequence data and time-step sequence data within a day of traffic flow sequence; using the index data as an index to extract the extracted embedded data from the initial embedded data; and connecting and expanding the extracted embedded data to obtain periodic embedded data.

[0055] Among them, the day sequence data in the weekly traffic flow sequence can be the data of which day in the weekly traffic flow sequence; the time step sequence data in a day can be the data of which time step in a day.

[0056] For example, the number of learnable days in a week is embedded in the data and represented as follows: Where Nw represents the number of days in a week, Nw=7; the time steps within a day are embedded in the data as follows: Where Nd refers to the total number of time steps in a day.

[0057] Then, It represents the data of which day in a week's traffic flow sequence and the data of which time step within a day; the data of which day in a week and the data of which time step need to correspond one-to-one with the length T of the input time series in order to achieve accurate embedding of time features.

[0058] Furthermore, in each Within each time step, index data is used as an index to extract the corresponding extracted post-embedded data from the feature embedding data, specifically including: weekly embedding. and time step embedding .

[0059] Finally, through concatenation and expansion, periodic embedded data can be obtained, namely, the periodic embedding of traffic flow sequences. .

[0060] The connection and expansion process specifically includes: concatenating the two along the df dimension to obtain the merged periodic feature matrix [Ewt,Edt]∈R. T×2df Then, since traffic data usually contains N monitoring nodes, it is necessary to associate the time period features with the spatial nodes. Therefore, the T×2df matrix is ​​expanded into a three-dimensional tensor of T×N×2df through the repeat operation to obtain the periodic embedding.

[0061] In this embodiment, by introducing periodic time features, the shortcomings of traditional embedding methods that ignore the periodicity of traffic data are effectively overcome, the accuracy of the representation of the time periodicity of traffic flow is improved, and a basis for subsequent prediction that fits the spatiotemporal characteristics of traffic data is provided.

[0062] Optionally, in some embodiments of this application, the method further includes: constructing time-adaptive embedded data based on the similarity principle of the data change trends and fluctuation characteristics of adjacent time series; and using the time-adaptive embedded data as dynamic similarity embedded data.

[0063] For example, since adjacent time series data tend to have high similarities in terms of trends and fluctuation characteristics, this application designs a time-adaptive embedding. It adaptively captures complex temporal correlations based on the changing trends of adjacent time series data.

[0064] In this embodiment, dynamic embeddings are generated by mining the inherent similarity between adjacent time series, which effectively overcomes the shortcomings of traditional embeddings in characterizing the dynamic correlation of time series, improves the accuracy of characterizing complex correlations in the time dimension, provides dynamic features that are more in line with the characteristics of time series data for subsequent processing, and enhances the model's ability to capture the patterns of time series changes.

[0065] In addition, in some embodiments, the process of concatenating and fusing the original volatility embedded data, periodic embedded data, and dynamic similarity embedded data corresponding to the original data to obtain the fused embedded data specifically includes:

[0066] Hidden temporal information can be embedded and represented by splicing and fusion. The mathematical expression for the splicing and fusion process can be: Where the hidden dimension dh is equal to 3df + da.

[0067] It should be noted that the process of constructing and splicing the embedded data in the above embodiments is the time information embedding process of this application, which can be implemented by a time information embedding module. Figure 3 This is a schematic diagram of the time information embedding module in one embodiment. Figure 3 As shown, the input raw data is processed through a linear layer to obtain the original fluctuating embedded data Edata, which is then combined with the periodic embedded data Ep and the dynamic similarity embedded data Ea to obtain the fused embedded data Fusion.

[0068] According to the above embodiments It can be seen that da refers to the output feature dimension of the dynamic similarity embedding data Ea. The three-branch features (Edata, Ep, Ea) are concatenated and fused into the hidden temporal feature Z. Since Edata contributes the df dimension, Ep contributes the 2df dimension, and Ea contributes the da dimension, the total hidden dimension is... .

[0069] The following specific embodiment demonstrates the form of embedded data in this application and illustrates how to represent hidden information in normal data by embedding vectors.

[0070] In this embodiment, embedding time information into a vector can be understood as a process of "labeling" and recording changes in traffic flow data. For example, during the morning rush hour on a Monday from 7:00 to 9:00 (recorded every 5 minutes, for a total of 24 data points), the traffic flow changes on one road segment are as follows:

[0071] Native fluctuation-type embedded data is like a "real-time change recorder". If the traffic suddenly increases from 1,000 vehicles to 1,500 vehicles at 7:30, it will transform this sudden increase in raw fluctuation into a series of 16-bit vectors (such as [0.72, 0.35, ...]) and indicate the model. The larger the number, the more obvious the fluctuation.

[0072] Periodically embedded data is similar to a "calendar tagger". For example, Monday's tag is [0.89, 0.21, ...], Sunday's tag is [-0.7, -0.4, 0.2, ...], and each day of the week has its own tag. In this way, the model captures the time cycle pattern by recording the tags of each day of the week.

[0073] Dynamic similarity embedding data acts like a "trend comparator." For example, if the traffic volume increases continuously from 800 to 1000 between 7:00 and 7:10, and from 1000 to 1200 between 7:10 and 7:20, two highly similar 8-bit vectors (such as [0.61, 0.19, ...] and [0.59, 0.21, ...]) are generated. However, if the traffic volume suddenly drops to 900 at 7:30 due to temporary traffic control, an 8-bit vector with a significant difference from the vectors generated in the previous two time periods will be generated to indicate that the model trend has changed.

[0074] Optionally, in some embodiments of this application, filtering the fused data using a high-frequency filtering network to obtain filtered time-series feature data includes: normalizing the fused data using an instance normalization layer to obtain instance-normalized data; filtering the instance-normalized data using a high-frequency filtering layer to obtain filter layer output data; performing linear transformation and nonlinear activation on the filter layer output data using a feedforward network layer to obtain feedforward network output data; and performing inverse instance normalization on the feedforward network output data using an inverse instance normalization layer to obtain filtered time-series feature data.

[0075] The high-frequency filtering network includes an instance normalization layer, a high-frequency filtering layer, a feedforward network layer, and an inverse instance normalization layer. The instance-normalized data refers to the data after instance normalization, the filter layer output data refers to the data output by the high-frequency filtering layer, and the feedforward network output data refers to the data output by the feedforward network layer.

[0076] The filtering process in this embodiment can be implemented by a high-frequency filtering network (also known as a high-frequency filtering module). Figure 4 This is a schematic diagram of the structure of a high-frequency filtering module in one embodiment.

[0077] For example, given that time series data are typically collected over long periods, these non-stationary series inevitably expose changes in the distribution of the predictive model over time. To address this issue, such as... Figure 4 As shown, this application adds an instance normalization layer before the frequency filter. After capturing the time dependencies between time series data, it uses a feedforward network to project them back to the time series data and then performs an inverse instance normalization operation.

[0078] In this embodiment, by combining instance normalization and inverse instance normalization, the problem of model distribution changing over time due to the non-stationarity of long-term series data is effectively overcome, the ability to capture the time dependence of time series data is improved, the effectiveness of feature transformation is enhanced by the projection processing of the feedforward network, the quality of filtered time series feature data is improved, and a more stable feature foundation is provided for subsequent prediction and other tasks.

[0079] Optionally, in some embodiments of this application, filtering the instance-normalized data according to the high-frequency filtering layer to obtain the filter layer output data includes: performing a Fourier transform on the instance-normalized data to obtain the Fourier transform result; constructing a context shaping filter based on a neural network; performing linear dense operations and multiplication operations on the Fourier transform result according to the context shaping filter to output the context shaping result; and performing an inverse Fourier transform on the product of the context shaping result and the Fourier transform result to obtain the filter layer output data.

[0080] Among them, such as Figure 4 As shown, the Fourier transform is FFT, and the Fourier transform result is the output of FFT. The context shaping filter is Contextual Shaping Filter, and the context shaping result is the output of linear dense operation and multiplication operation. The inverse Fourier transform is inverse FFT.

[0081] For example, the calculation formula for the filtering process is as follows:

[0082]

[0083] in, Indicates instance normalization, Indicates Fourier transform, This refers to the input of the context shaping filter in the diagram. This represents the feedforward network layer. To enable the high-frequency filtering layer to learn parameters generated from the input data and thus better adapt to the data, this application uses a neural network. It flexibly adjusts the frequency filter to respond to the input data, utilizing a network. The context-shaping filter is derived. First, through linear dense operations... Embedding raw data enhances the ability to model complex data. Then, it is combined with... Learnable parameters Performing multiplication in dimension D produces... ,in It is an activation function, and the final output is... . This refers to the original data embedded in a new data space after linear intensive operations.

[0084] Additionally, it should be noted that, Figure 4 In this context, LayerNorm (layer normalization) refers to normalizing the output Q of the high-frequency filtering layer at the layer level, further stabilizing the data distribution and aiding the learning of subsequent network layers; Dropout is a regularization technique that randomly discards some neurons during training to prevent overfitting and enhance the model's generalization ability; FFN (feedforward network) is a feedforward neural network composed of fully connected layers, etc., which performs linear transformations and non-linear activations on the data processed by LayerNorm and Dropout to extract more abstract features, with the output being M; Dense refers to dense layers or fully connected layers.

[0085] In this embodiment, an adaptive context shaping filter is generated by a neural network, which can flexibly adjust the frequency filtering to adapt to the characteristics of the input data. This effectively overcomes the shortcomings of fixed filters in dealing with complex data changes, improves the ability to accurately control frequency domain features, and enhances the pertinence and effectiveness of high-frequency filtering by combining Fourier transform and inverse transform. It also improves the retention of key features in the filtered data.

[0086] Optionally, in some embodiments of this application, spatial feature extraction is performed on the original data using a graph convolutional neural network to obtain spatial feature data associated with the original data, including: acquiring a static road network map, a dynamic road network map, and a similar map; while keeping the original dimensions of the tensors unchanged, a fusion method of content superposition of multimodal feature tensors is used to fuse the static road network map, the dynamic road network map, and the similar map to obtain a multi-graph fusion network; and spatial feature data associated with the original data is extracted based on the multi-graph fusion network and a two-layer spatial graph convolutional network.

[0087] In this embodiment, multi-map fusion in the multi-map fusion network refers to the fusion of static road network maps, dynamic road network maps, and similar maps.

[0088] Among them, static road network map Used to describe the inherent topology of road networks, where the attributes of nodes and edges do not change over time, it is mainly used to capture long-term stable spatial dependencies in transportation systems.

[0089] Among them, dynamic road network map This is used to describe changes in road network correlation, and its dynamic characteristics primarily reflect short-term traffic environment dynamics. To construct the dynamic map, this application uses traffic speed... As auxiliary features, a 1x1 convolutional layer is used as the embedding function. The embedded feature tensor is rearranged, multiplied by its transpose matrix, and then normalized to generate a dynamic road network map. The formula is as follows:

[0090]

[0091] in Represents embedded functions The parameters are dynamically updated through backpropagation based on traffic flow prediction results. This represents the matrix transpose operation. This indicates normalization processing.

[0092] Among them, similar graphs This similarity map is used to represent the similarity of traffic flow sequences between different roads over time. It is calculated based on the DTW algorithm. The DTW algorithm is an algorithm used to measure the similarity between two time series. Its essence is to calculate the optimal matching path between two time series signals. It allows the time axis to be scaled to find the best correspondence. The calculation steps are as follows: (1) Construct the distance matrix: calculate the distance between each pair of points in the two sequences to form a distance matrix. (2) Initialize the cumulative distance matrix: set the cumulative distance of the first row and the first column of the matrix to the values ​​of the corresponding subscripts of the first row and the first column of the distance matrix. (3) Fill the cumulative distance matrix: calculate the minimum cumulative distance of each point step by step from the upper left to the lower right according to the distance matrix constructed between them. (4) Backtrack the optimal path: backtrack from the lower right corner of the cumulative distance matrix to the upper left corner to find the optimal matching path.

[0093] Among them, the similarity matrix The calculation formula is as follows:

[0094]

[0095] in, This represents the numerical result obtained after applying the DTW algorithm to a pair of points vi and vj. To set the threshold, exp is an exponential function raised to the power of the natural logarithm base e.

[0096] Finally, a fusion method that combines the contents of multimodal feature tensors while keeping the original dimensions of the tensors unchanged is adopted to fuse static road network maps, dynamic road network maps, and similar maps.

[0097] Spatial graph convolution adopts the Chebyshev graph convolution method, which innovatively introduces a K-order polynomial approximation strategy. By truncating the expanded approximate convolution kernel function, it fundamentally avoids the matrix spectral decomposition process compared with the traditional spectral domain convolution method, which can significantly reduce the amount of computation. It only needs to calculate the coefficients of the Chebyshev polynomial and the related matrix multiplications.

[0098] In addition, in some embodiments, the spatiotemporal feature fusion process corresponds to a spatiotemporal feature fusion module. Figure 5 A schematic diagram of the spatiotemporal feature fusion module in one embodiment is provided.

[0099] like Figure 5 As shown, after the spatial and temporal feature extraction modules generate high-level features respectively, in order to preserve the original features of the data as much as possible, this application separately extracts temporal features. and spatial features Linear transformations and residual connections are used to further adjust the dependencies between sequences to improve the expressive power of the model.

[0100] exist Figure 5 In this process, the FFN (Feedforward Network) performs a linear transformation operation; residual connections are achieved by element-wise summing of the features processed by the FFN with the original input features. Furthermore, a fusion method is employed that combines the contents of multimodal feature tensors while maintaining the original tensor dimensions, summing the temporal and spatial features to obtain fused spatiotemporal features. The specific formula is as follows:

[0101]

[0102] in, All are learnable parameters. The number of layers in the spatial and temporal feature extraction module is designed as one layer in this application. Presentation layer normalization processing, This represents the spatial feature tensor after linear transformation, residual connection, and layer normalization. This represents the time feature tensor after linear transformation, residual connection, and layer normalization. For the first The spatial feature output of the layer, For the first The temporal feature output of the layer, Represents 0 to The sum of the spatial feature outputs of the layers, This represents the sum of the time feature outputs from layers 0 to K. Furthermore, the output layer is used to perform multi-step prediction of traffic flow.

[0103] The purpose of using LayerNorm (layer normalization) before stacking is to stabilize the distribution of input data in each layer, make the gradient smoother during backpropagation, reduce training oscillations caused by feature scale fluctuations, thereby enhancing gradient backpropagation and improving training stability.

[0104] Figure 6 This is a flowchart illustrating a traffic flow prediction method based on feature embedding and high-frequency filtering in another embodiment. In another embodiment, Figure 6 A traffic flow prediction method based on feature embedding and high-frequency filtering is provided, including the following steps:

[0105] S601, Obtain the data corresponding to the preset time step interval in the original data; Perform linear transformation on the data corresponding to the preset time step interval based on the linear layer to obtain the native volatility embedded data.

[0106] S602, determine the initial embedding data as the embedding data of the number of days in a week that can be learned and the embedding data of the time steps in a day; determine the index data as the day order data and the time step order data in the traffic flow sequence of a week.

[0107] S603: Using the index data as an index, extract the extracted embedded data from the initial embedded data; connect and expand the extracted embedded data to obtain periodic embedded data.

[0108] S604, based on the principle of similarity between the changing trends and fluctuation characteristics of adjacent time series data, constructs time-adaptive embedded data; and uses the time-adaptive embedded data as dynamic similarity embedded data.

[0109] S605, the original data is spliced ​​and fused together with the original volatility embedded data, periodic embedded data and dynamic similarity embedded data to obtain the fused embedded data.

[0110] S606, the fused data is normalized according to the instance normalization layer to obtain the instance normalized data; the instance normalized data is filtered according to the high-frequency filtering layer to obtain the output data of the filtering layer.

[0111] S607: The feedforward network output data is obtained by performing linear transformation and nonlinear activation on the output data of the filter layer according to the feedforward network layer; the inverse instance normalization layer is used to perform inverse instance normalization on the output data of the feedforward network to obtain the filtered time series feature data.

[0112] S608: Obtain static road network map, dynamic road network map, and similar map; while keeping the original dimensions of the tensor unchanged, use the content superposition fusion method of multimodal feature tensor to fuse the static road network map, dynamic road network map, and similar map to obtain a multi-map fusion network.

[0113] S609 extracts spatial feature data associated with the original data based on a multi-graph fusion network and a two-layer spatial graph convolutional network.

[0114] S610, based on the principles of linear transformation and residual connection, performs spatiotemporal feature fusion on temporal feature data and spatial feature data to obtain fused spatiotemporal feature data; based on the fused temporal feature data, predictive data of traffic flow associated with the original data is obtained.

[0115] It should be noted that the specific limitations of the above steps can be found in the above description of the specific limitations of a traffic flow prediction method based on feature embedding and high-frequency filtering, and will not be repeated here.

[0116] Figure 7 This is a schematic diagram of the overall process principle in one embodiment. See below for reference. Figure 7 The present application describes the research process and other technical details of the traffic flow prediction method based on feature embedding and high-frequency filtering provided in this application using a specific embodiment.

[0117] In traditional technologies, the core objective of traffic flow prediction is to accurately capture the spatiotemporal dependencies in traffic data, thereby achieving reliable predictions of future traffic conditions. However, with the exponential growth in the scale of traffic data, the limitations of traditional methods and early deep learning models in time-dimensional modeling have become increasingly apparent, failing to meet the demands of intelligent transportation systems for high-precision, real-time predictions. Therefore, optimizing time-series feature extraction capabilities has become a key breakthrough direction for improving traffic flow prediction performance.

[0118] The temporal characteristics of traffic flow data pose multidimensional challenges to predictive models. These challenges manifest in several ways: First, long-term dependence, such as the periodic traffic flow changes during morning and evening rush hours requiring models to capture long-term trends across time steps. Second, high-frequency fluctuations, such as sudden accidents or temporary traffic control measures causing short-term traffic surges within 5 minutes, contain crucial predictive clues but require models with sensitive response capabilities. Third, periodic patterns, such as weekday and weekend traffic differences and congestion patterns at different times of the day, require models to effectively utilize daily / weekly periodic characteristics. Fourth, proximity similarity, the continuity of road conditions, weather, and other factors within a short period leads to highly similar traffic flow trends in adjacent time steps, requiring models to capture this local similarity. These complex temporal characteristics collectively determine the high-precision requirement for time-dimensional modeling in traffic flow prediction.

[0119] Traditional techniques have significant limitations in modeling the temporal dimension. Traditional recurrent networks, due to the vanishing gradient problem, struggle to effectively capture early information in long sequences, leading to insufficient modeling of long-term dependencies. While Transformer-based models alleviate long-term dependency issues through self-attention mechanisms, their globally weighted summation characteristic focuses more on low-frequency long-term trends and is insufficiently responsive to high-frequency fluctuations. Furthermore, existing methods often focus on improving model structure, neglecting effective representation of the input data itself and failing to fully utilize explicit embedding of features such as temporal periodicity and the similarity of adjacent data, resulting in insufficient understanding of data patterns. Simultaneously, the fusion of temporal and spatial features often employs simple concatenation or superposition without a targeted fusion mechanism, leading to the loss of feature information during the fusion process and further limiting the improvement of prediction accuracy. These issues collectively constitute the key bottlenecks hindering breakthroughs in traffic flow prediction performance.

[0120] In traditional techniques, an improved spatiotemporal graph neural network (STGNN) for spatiotemporal sequence prediction focuses on enhancing the model's ability to capture spatiotemporal patterns by explicitly embedding spatiotemporal identity features. Another Transformer-based time series prediction model uses a self-attention mechanism to capture global temporal dependencies in long sequences and explicitly incorporates positional information from time steps using external knowledge such as positional encoding. However, these techniques still suffer from the following problems:

[0121] (1) Existing models only represent time features through time periodic embedding, without combining the dynamic features and original fluctuation features of the original traffic data, resulting in insufficient data representation and difficulty in fully capturing the complex patterns of the time dimension; (2) The ability to respond to high-frequency fluctuations such as short-term traffic surges is limited, resulting in a significant increase in error in short-term prediction scenarios; (3) Spatiotemporal features are mostly fused by simple splicing or superposition, without designing a targeted fusion mechanism, resulting in the loss of feature information during the fusion process, affecting the final prediction accuracy.

[0122] To address the aforementioned issues, this application provides a traffic flow prediction method based on feature embedding and high-frequency filtering, corresponding to a traffic flow prediction model that combines feature embedding and high-frequency filtering to achieve high-precision, real-time traffic flow prediction. Details are as follows.

[0123] like Figure 7 As shown, the traffic flow prediction method based on feature embedding and high-frequency filtering proposed in this application takes historical traffic flow data as input and outputs the traffic flow prediction value for the future time step through the process of "input preprocessing → multi-dimensional time feature embedding → high-frequency filtering feature enhancement → dynamic spatiotemporal interaction fusion → prediction output".

[0124] This application proposes a time information embedding module, which addresses the problem of limited time information representation in existing models by using a three-branch embedding sub-module based on "periodicity, dynamic similarity, and native fluctuation." It extracts native fluctuation information Edata through a linear layer, and then fuses the three branches of information—Edata, periodicity information Ep, and dynamic similarity information Ea—to improve the sufficiency of time information.

[0125] This application proposes a high-frequency filtering network that addresses the problem of weak high-frequency information extraction capability through adaptive filtering technology. The network constructs a learnable context-shaping filter to dynamically adjust frequency filtering and performs frequency domain decomposition and reconstruction of the input data, achieving accurate capture of high-frequency information and effective suppression of low-frequency background.

[0126] This application proposes a spatiotemporal feature fusion module, which employs a dual-path mechanism to extract temporal and spatial features through two separate paths, thereby addressing the information loss issue during spatiotemporal feature fusion. Specifically, spatial features are extracted using multi-graph fusion combined with two layers of spatial graph convolution.

[0127] This application strengthens the representation of data features and enhances the ability to model time correlation by designing a time information embedding module. It solves the problem that existing methods have poor information integration capabilities and lack memory mechanisms when extracting time correlations, due to the traditional attention mechanism or Transformer module.

[0128] This application proposes a high-frequency filtering network that extracts the time correlation of input data in the frequency domain. By flexibly adjusting the frequency filtering through a learnable context-shaping filter, it enhances the ability to capture high-frequency information and solves the problem that existing methods have insufficient information extraction capabilities when processing high-frequency signals and cannot fully utilize the full spectrum information.

[0129] This application proposes a spatiotemporal feature fusion module, which adjusts sequence dependencies through linear transformation and residual connection, and sums temporal and spatial features by superimposing the contents of multimodal feature tensors while keeping the original tensor dimensions unchanged, so as to reduce feature loss during the fusion process and achieve more efficient spatiotemporal dependency fusion.

[0130] exist Figure 7 In this model, the input data for the temporal feature extraction module is a graph structure describing changes over time, where the connections between nodes and edges are dynamically updated with each time step. This is achieved through time slicing, dividing continuous time into discrete time steps, with each time step constructing a static graph.

[0131] exist Figure 7 In the spatial feature extraction module, the input data "X_speed" represents traffic speed feature parameters, which are used as auxiliary data to construct the dynamic road network map. In the multi-map fusion module, the generation of the dynamic road network map (Ad) depends on the temporal variation pattern of traffic speed data.

[0132] Based on a multi-dimensional time feature embedding mechanism, this application proposes a three-branch time feature embedding method of "periodicity-dynamic similarity-original fluctuation". It integrates three types of features: the periodicity of time, the similarity of the data change trends of adjacent time series, and the fluctuation pattern of the original flow data, and outputs fused time features.

[0133] This application utilizes feature enhancement techniques based on learnable high-frequency filtering and low-frequency suppression to design a frequency domain filtering mechanism based on Fourier transform. This mechanism can decompose and reassemble the input data at the frequency domain level and dynamically adjust the frequency filtering process through a learnable context-shaping filter. This design enables the model to accurately capture high-frequency information while effectively suppressing low-frequency background.

[0134] This application proposes a "space-time" dual-path fusion strategy, which extracts spatial features through multi-graph fusion and two-layer graph convolution, extracts temporal features through temporal information embedding and high-frequency filtering, and finally achieves spatiotemporal feature fusion through tensor superposition.

[0135] The advantage of this application lies in its more comprehensive representation of time information. Compared with traditional models STID and PDFormer, the three-branch time information embedding module of "periodicity-dynamic similarity-native fluctuation" integrates the time features. The time features not only include the external periodic knowledge contained in STID and PDFormer, but also embed information about the dynamic similarity and fluctuation of the data itself that is not covered by STID and PDFormer. This significantly improves the sufficiency of time information and its adaptability to actual traffic scenarios.

[0136] Furthermore, this application has the advantage of more accurate high-frequency fluctuation capture, taking into account both long-term trends and short-term changes. STID relies on traditional gated convolution to extract temporal features, which is insufficient for responding to high-frequency fluctuations such as short-term traffic surges; while PDFormer's self-attention mechanism, due to its global weighted summation characteristics, focuses more on low-frequency long-term trends and easily ignores high-frequency details. This application, however, achieves accurate capture of high-frequency details through a high-frequency filtering module, making up for the shortcomings of STID and PDFormer in high-frequency information extraction.

[0137] Figures 1 to 7 Any technical feature in the embodiments corresponding to any of the above items is also applicable to the embodiments of this application. Figures 8 to 10 The corresponding implementation examples will not be repeated hereafter.

[0138] The above describes a traffic flow prediction method based on feature embedding and high-frequency filtering in the embodiments of this application. The apparatus for performing the above method is described below.

[0139] Figure 8 Here is a structural block diagram of a traffic flow prediction device based on feature embedding and high-frequency filtering in one embodiment. The following refers to... Figure 8 A traffic flow prediction device based on feature embedding and high-frequency filtering is described, the device comprising:

[0140] The embedded data fusion module 801 is used to acquire the original data of traffic flow to be predicted, and to splice and fuse the original data corresponding to the original data, the original fluctuation embedded data, the periodic embedded data, and the dynamic similarity embedded data to obtain the fused embedded data.

[0141] The filtering module 802 is used to filter the fused data according to the high-frequency filtering network to obtain filtered time-series feature data.

[0142] The spatial feature extraction module 803 is used to extract spatial features from the original data based on the graph convolutional neural network to obtain spatial feature data associated with the original data.

[0143] The spatiotemporal feature fusion module 804 is used to perform spatiotemporal feature fusion on time-series feature data and spatial feature data based on the principles of linear transformation and residual connection, so as to obtain fused spatiotemporal feature data.

[0144] The prediction output module 805 is used to predict traffic flow data associated with the original data based on the fused time feature data.

[0145] In this embodiment of the application, based on, as follows Figure 8 The connections between the modules shown in the diagram demonstrate how the cooperation between these modules can improve the accuracy of traffic flow prediction and enhance the model's prediction performance.

[0146] In another embodiment, a traffic flow prediction device based on feature embedding and high-frequency filtering is provided. This device can be a computer device, such as a server, and its internal structure diagram can be as follows: Figure 9 As shown, the device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The device's database stores relevant data. The I / O interfaces are used for exchanging information between the processor and external devices. The device's communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements the various methods described in the above embodiments.

[0147] In yet another embodiment, a traffic flow prediction device based on feature embedding and high-frequency filtering is provided. This device can be a computer device, such as a terminal, and its internal structure diagram can be as follows: Figure 10As shown, the device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements the various methods described in the above embodiments. The display unit is used to form a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device.

[0148] Those skilled in the art will understand that Figure 9 and Figure 10 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the traffic flow prediction device based on feature embedding and high-frequency filtering applied thereto. Specifically, the device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements, in order to achieve the functions of computer equipment such as terminals or servers.

[0149] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0150] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0151] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, apparatuses, or modules, and may be electrical, mechanical, or other forms.

[0152] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0153] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium.

[0154] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.

[0155] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., a solid-state disk (SSD)).

[0156] The technical solutions provided in the embodiments of this application have been described in detail above. Specific examples have been used in the embodiments of this application to illustrate the principles and implementation methods of the embodiments of this application. The description of the above embodiments is only for the purpose of helping to understand the methods and core ideas of the embodiments of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the embodiments of this application. Therefore, the content of this specification should not be construed as a limitation on the embodiments of this application.

Claims

1. A traffic flow prediction method based on feature embedding and high-frequency filtering, characterized in that, The method includes: Obtain the original traffic flow data to be predicted, and then splice and fuse the original volatility embedded data, periodic embedded data, and dynamic similarity embedded data corresponding to the original data to obtain the fused embedded data. The fused data is filtered using a high-frequency filtering network to obtain filtered time-series feature data. Spatial features are extracted from the original data using a graph convolutional neural network to obtain spatial feature data associated with the original data; Based on the principles of linear transformation and residual connection, the temporal feature data and the spatial feature data are fused to obtain the fused spatiotemporal feature data. Based on the fused time feature data, predicted traffic flow data associated with the original data is obtained.

2. The method according to claim 1, characterized in that, The method further includes: Obtain the data corresponding to the preset time step interval in the original data; the preset time step interval is the time step interval from the start time step of the historical data to the start time step of the prediction. The data corresponding to the preset time step interval is linearly transformed based on the linear layer to obtain the native volatility embedded data.

3. The method according to claim 1, characterized in that, The method further includes: The initial embedding data was determined to be the embedding data of the number of days in a week that can be learned and the embedding data of time steps in a day; The index data is determined to be the day-order data and the time-step order data within a day of a week's traffic flow sequence; Using the index data as an index, the extracted embedded data is obtained from the initial embedded data; The extracted and embedded data are concatenated and expanded to obtain the periodic embedded data.

4. The method according to claim 1, characterized in that, The method further includes: Based on the principle of similarity in the changing trends and fluctuation characteristics of adjacent time series data, time-adaptive embedded data is constructed. The time-adaptive embedded data is used as the dynamic similarity embedded data.

5. The method according to claim 1, characterized in that, The high-frequency filtering network includes an instance normalization layer, a high-frequency filtering layer, a feedforward network layer, and an inverse instance normalization layer. The step of filtering the fused data using the high-frequency filtering network to obtain filtered time-series feature data includes: The fused data is normalized according to the instance normalization layer to obtain the instance-normalized data. The normalized data of the instance is filtered by the high-frequency filtering layer to obtain the output data of the filtering layer. The feedforward network output data is obtained by performing linear transformation and nonlinear activation on the output data of the filter layer based on the feedforward network layer. The inverse instance normalization layer is used to perform inverse instance normalization on the output data of the feedforward network to obtain the filtered time-series feature data.

6. The method according to claim 5, characterized in that, The step of filtering the instance-normalized data according to the high-frequency filtering layer to obtain the output data of the filtering layer includes: Perform a Fourier transform on the normalized data of the instance to obtain the Fourier transform result; A context shaping filter is constructed based on a neural network. Linear dense operations and multiplication operations are performed on the Fourier transform result according to the context shaping filter to output the context shaping result. Perform an inverse Fourier transform on the product of the context shaping result and the Fourier transform result to obtain the output data of the filter layer.

7. The method according to claim 1, characterized in that, The step of extracting spatial features from the original data using a graph convolutional neural network to obtain spatial feature data associated with the original data includes: Obtain static road network maps, dynamic road network maps, and similar maps; While keeping the original dimensions of the tensors unchanged, a fusion method of content superposition of multimodal feature tensors is adopted to fuse the static road network map, the dynamic road network map and the similar map to obtain a multi-graph fusion network. Spatial feature data associated with the original data are extracted based on the multi-graph fusion network and the two-layer spatial graph convolutional network.

8. A traffic flow prediction device based on feature embedding and high-frequency filtering, characterized in that, The device includes: The embedded data fusion module is used to acquire the original data of traffic flow to be predicted, and to splice and fuse the original fluctuation embedded data, periodic embedded data and dynamic similarity embedded data corresponding to the original data to obtain the fused embedded data. The filtering module is used to filter the fused data according to the high-frequency filtering network to obtain filtered time-series feature data. The spatial feature extraction module is used to extract spatial features from the original data based on the graph convolutional neural network to obtain spatial feature data associated with the original data. The spatiotemporal feature fusion module is used to perform spatiotemporal feature fusion on the time-series feature data and the spatial feature data based on the principles of linear transformation and residual connection, so as to obtain fused spatiotemporal feature data. The prediction output module is used to predict traffic flow data associated with the original data based on the fused time feature data.

9. A traffic flow prediction device based on feature embedding and high-frequency filtering, characterized in that, The device includes: At least one processor and memory; The memory is used to store program code, and the processor is used to call the program code stored in the memory to execute the method as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that, It includes instructions that, when executed on a computer, cause the computer to perform the method as described in any one of claims 1 to 7.

Images