A Traffic Flow Prediction Method Based on Adaptive Lightweight Temporal Convolutional Networks

Abstract

This invention discloses a traffic flow prediction method based on an adaptive lightweight temporal convolutional network. In the context-aware embedding layer, historical traffic flow data is downsampled, and contextual information is provided for the traffic flow at each downsampled time step. This information is then aggregated with the temporal embedding to obtain processed data. Spatial embedding, the processed data, and temporal embedding are used as inputs to a stacked spatiotemporal layer. In the output layer, the outputs of the adaptive lightweight temporal convolutional module in each spatiotemporal layer are skip-connected to obtain aggregated features. These features are then obtained through convolution operations to obtain the predicted traffic flow. This invention employs an adaptive lightweight temporal convolutional module to deeply extract temporal features, enabling adaptive updating of model parameters in a lightweight and fine-grained manner, thereby improving the ability to handle multiple temporal patterns of traffic flow. It requires no prior knowledge and can be easily extended to other spatiotemporal prediction tasks.

Description

Technical Field

[0001] This application belongs to the field of intelligent transportation technology, and in particular relates to a traffic flow prediction method based on an adaptive lightweight temporal convolutional network. Background Technology

[0002] Traffic flow prediction is an important component of intelligent transportation systems. Accurate traffic flow prediction provides fundamental support for many applications, not only helping traffic managers to take measures in advance to alleviate congestion, but also facilitating route decisions for travelers.

[0003] However, the complex spatiotemporal characteristics inherent in traffic data make accurate and reliable traffic flow prediction extremely challenging. Traffic flow possesses inherent time-series attributes (such as periodicity and trends) and spatial attributes (such as geographic location and traffic topology). To improve the accuracy of traffic flow prediction, researchers have conducted extensive studies to extract features from the temporal and spatial dimensions of traffic data. Meanwhile, encouraging progress has been made in theoretical and experimental research on graph convolutional networks and attention mechanisms. Nevertheless, traffic flow prediction on large-scale datasets remains challenging.

[0004] First, existing high-performance, complex spatiotemporal models for traffic flow prediction rely on expensive computations. When faced with large-scale real-world datasets, these high-performance models struggle to be trained with limited hardware resources. Furthermore, when deployed on edge devices with limited computing power, they fail to meet the real-time requirements of inference.

[0005] Secondly, traffic flow throughout the day often exhibits multiple temporal patterns (e.g., stable, increasing, fluctuating, decreasing). However, existing research often assumes that traffic flow throughout the day has a common pattern and attempts to extract temporal features in a coarse-grained manner (i.e., an average representation of multiple temporal patterns of traffic flow), which hinders the in-depth extraction of such features. Summary of the Invention

[0006] The purpose of this application is to provide a traffic flow prediction method based on an adaptive lightweight temporal convolutional network, which overcomes the two key problems mentioned in the background technology, can effectively extract the spatiotemporal features of traffic flow, and its lightweight nature allows it to be used under limited hardware resources.

[0007] To achieve the above objectives, the technical solution of this application is as follows:

[0008] A traffic flow prediction method based on an adaptive lightweight temporal convolutional network, wherein the adaptive lightweight temporal convolutional network includes a context-aware embedding layer, stacked spatiotemporal layers, and an output layer, and the traffic flow prediction method based on the adaptive lightweight temporal convolutional network includes:

[0009] In the context-aware embedding layer, the input historical traffic flow is downsampled, and then contextual information is provided for the traffic flow at each time step after downsampling. This information is then aggregated with the time embedding to obtain the processed data.

[0010] Spatial embedding, processed data, and temporal embedding are used as inputs to stacked spatiotemporal layers. For any spatiotemporal layer, the output of the previous spatiotemporal layer is aggregated with the spatial embedding and then passed through an adaptive lightweight temporal convolution module with the temporal embedding. Then, convolution and residual connections are performed, and finally batch normalization is used to obtain the output features of the current spatiotemporal layer.

[0011] In the output layer, the outputs of the adaptive lightweight temporal convolutional modules in each spatiotemporal layer are skip-connected to obtain aggregated features, and then the predicted traffic flow is obtained through convolution operations.

[0012] Furthermore, the provision of contextual information for traffic flow at each downsampled time step is expressed as the following formula:

[0013] X context =X int ||X fd ||X bd

[0014] Among them, X int X is the downsampled traffic flow sequence. context It is a traffic flow sequence with contextual information; || is the join function; X fd and X bd These represent the forward and backward differences in traffic flow at each time step.

[0015] Furthermore, the time-embedded aggregation yields the processed data, expressed as follows:

[0016] X P =Γ P *(X context ||E temporal )+b P

[0017] Where X P For the processed data, Γ P For the convolution kernel, b P For bias, * represents standard convolution operation, || represents the concatenation function, and X represents the bias. context It is a traffic flow sequence with contextual information, Etemporal Embedded for time.

[0018] Furthermore, the time embedding includes daily embedding and weekly embedding.

[0019] Furthermore, the adaptive lightweight temporal convolution module in the l-th spatiotemporal layer performs the following operations:

[0020]

[0021]

[0022] Among them, A (l) and B (l) It represents the feature of equal size; Dropout[.] indicates the dropout operation; σ(.) is the sigmoid function; ⊙ is the Hadamard product. and These are the convolution kernel and bias generated for the input traffic flow at time t0, respectively. and These are learnable parameters. and It is a bias; The output features of the adaptive lightweight temporal convolutional module in the l-th spatiotemporal layer are... E represents the output feature of the (l-1)th spatiotemporal layer. spatial Indicates spatial embedding, Indicates residual connection, E temporal This indicates time embedding.

[0023] Furthermore, the aggregated features are obtained by skip-connecting the outputs of the adaptive lightweight temporal convolutional modules in each spatiotemporal layer, as expressed by the following formula:

[0024]

[0025] Among them, H sc These are aggregated features following skip connections; For convolution kernel; For bias, This represents the output features of the adaptive lightweight temporal convolution module in the l-th spatiotemporal layer.

[0026] This application proposes a traffic flow prediction method based on an adaptive lightweight temporal convolutional network. It downsamples traffic flow in the temporal dimension while providing contextual information for traffic flow at each time step. Then, an adaptive lightweight temporal convolutional module is employed to deeply extract temporal features. This module can adaptively update model parameters in a lightweight and fine-grained manner, thereby improving the ability to handle multiple temporal patterns of traffic flow. This application effectively learns the latent spatiotemporal patterns of traffic flow through data-driven spatiotemporal embedding technology, requiring no prior knowledge and easily extending to other spatiotemporal prediction tasks. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the adaptive lightweight temporal convolutional network framework of this application.

[0028] Figure 2 This is a flowchart of the traffic flow prediction method in this application. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0030] In traffic flow prediction, for N sensors in a given area, the traffic data at time step t can be represented as: Where d represents the dimension of the traffic data recorded by the sensor. In this application, d is set to 1, focusing on traffic flow prediction.

[0031] The goal of traffic flow prediction is to learn a mapping function. Predicting traffic flow for the next T time steps based on traffic flow in the previous T′ time steps can be expressed as formula (1):

[0032]

[0033] To address the challenge of traffic flow prediction on large-scale datasets, this application proposes an Adaptive Lightweight Temporal Convolutional Network (ALTCN-CD model) based on a context-aware downsampling strategy. This model adheres to lightweight design principles to efficiently extract spatiotemporal features. Furthermore, it is entirely data-driven and requires no prior knowledge, thus allowing for convenient application to other spatiotemporal prediction tasks. Figure 1As shown, the ALTCN-CD model consists of three parts: a context-aware embedding layer, stacked spatiotemporal layers, and an output layer. The context-aware embedding layer processes the input data to obtain spatiotemporal embeddings. Its built-in context-aware downsampling strategy downsamples the input data to reduce sequence length and lighten the model, while providing contextual information for traffic flow at each time step to avoid information loss. The stacked spatiotemporal layers fuse an adaptive lightweight temporal convolution module (ALTCM module) and spatiotemporal embeddings to efficiently extract spatiotemporal features. Simultaneously, the ALTCM module adaptively updates model parameters in a lightweight and fine-grained manner to handle various temporal patterns of traffic flow. Finally, the output layer aggregates features through skip connections and generates prediction results.

[0034] One embodiment of this application, such as Figure 2 As shown, a traffic flow prediction method based on an adaptive lightweight temporal convolutional network is proposed, including:

[0035] Step S1: In the context-aware embedding layer, the input historical traffic flow is downsampled, and then context information is provided for the traffic flow at each time step after downsampling. This information is then aggregated with the time embedding to obtain the processed data.

[0036] In recent years, advanced traffic flow prediction methods, especially those based on graph convolutional networks (GCNs) and attention mechanisms, often require significant computational resources to achieve marginal performance improvements. Therefore, these methods are difficult to deploy on large-scale datasets with limited hardware resources.

[0037] This application proposes a context-aware embedding layer with a built-in context-aware downsampling strategy, which significantly reduces the computational cost of the model while retaining the model's ability to extract spatiotemporal features.

[0038] The context-aware downsampling strategy proposed in this application refers to the context-aware embedding layer downsampling the original input data. The preprocessing process performed. For example, if the original input data sequence length is T′, the context-aware embedding layer downsamples X using fixed-interval sampling. Assume the downsampled subsequence is... (That is, the sequence length of X is reduced from T′ to L) int If the downsampling process can be formalized as formula (2):

[0039]

[0040] While naive downsampling reduces sequence length and lowers the computational cost of the model, it inevitably leads to information loss, thereby reducing the model's predictive performance.

[0041] To address the aforementioned issues, this embodiment provides contextual information for traffic flow at each time step within the context-aware embedding layer. This avoids information loss and preserves the model's ability to extract spatiotemporal features. The process of providing contextual information can be expressed as formula (3):

[0042] X context =X int ||X fd ||X bd (3)

[0043] in It is an X with contextual information. int || is the join function; Forward difference value, representing the difference between traffic flow at each time step and traffic flow at the previous adjacent time step; The backward difference value represents the difference between the traffic flow at each time step and the traffic flow at the next adjacent time step. To avoid information disclosure, this application will use X... fd The value is set to 0 at the (tT′+1)th time step.

[0044] It should be understood that although this embodiment uses the forward and backward difference values ​​of traffic flow at each time step to represent context information, there are other alternative ways to implement context information. For example, context information could also be the trend relationship between traffic flow at each time step and traffic flow at adjacent steps or multiple adjacent steps.

[0045] Subsequently, the context-aware embedding layer employs spatiotemporal embedding to achieve efficient extraction of spatiotemporal features. Based on the research findings of those skilled in the art, this embodiment introduces known technical terms, namely: spatiotemporal embedding. Zhou Embedded and spatial embedding Its parameters are learnable during model training; daily embedding represents the recognition sequence at each time step of a day; weekly embedding represents the recognition sequence for each day of a week; spatial embedding represents the recognition sequence for each sensor; N day (In this embodiment, it is set to 288) represents the number of time steps in a day; N week =7 represents the number of days in a week; D represents the hidden dimension of the ALTCN-CD model proposed in this embodiment.

[0046] Based on X int middle The time attribute tp (e.g., the 88th time step on Wednesday) is used to obtain the time embedding in this embodiment. (For example, the identification sequence representation at the 88th time step on Wednesday), as shown in Equation (4):

[0047]

[0048] in (For example, the identification sequence at the 88th time step of the day) and (For example, the identification sequence on Wednesday) are E day and E week This corresponds to a subset of the time attribute tp. This embodiment uses formula (5) to represent X. context and E temporal Put together:

[0049] X P =Γ P *(X context ||E temporal )+b P (5)

[0050] in These represent the processed data, the convolution kernel, and the bias, respectively. * represents the standard convolution operation, and || represents the connection function.

[0051] Step S2: Use the spatial embedding, processed data, and temporal embedding as inputs to the stacked spatiotemporal layers. For any spatiotemporal layer, the output of the previous spatiotemporal layer is aggregated with the spatial embedding and then passed through an adaptive lightweight temporal convolution module with the temporal embedding. Then, convolution and residual connections are performed, and finally batch normalization is performed to obtain the output features of the current spatiotemporal layer.

[0052] In this embodiment, each of the stacked spatiotemporal layers includes an ALTCM module, a convolutional module, and a batch normalization module. The stacked spatiotemporal layers embed space into E. spatial Processed data X P and time embedding E temporal As input, it enables efficient learning of the underlying spatiotemporal patterns of traffic flow.

[0053] Specifically, in this embodiment, L spatiotemporal layers are stacked, and their propagation rules can be expressed as formulas (6) and (7):

[0054]

[0055]

[0056] in, and These represent the output features and operation set of the ALTCM module in the l-th spatiotemporal layer, respectively; l∈{1,2,...,L}; M (l) for The sequence length; and These represent the output features, convolution kernel, and bias of the l-th spatiotemporal layer, respectively; BatchNorm(.) and These represent batch normalization and residual join operations, respectively. Note: In the first spatiotemporal layer,

[0057] Although real-world traffic flow exhibits various temporal patterns throughout the day (e.g., stable, increasing, fluctuating, decreasing), existing models often extract temporal features in a coarse-grained manner. These models assume a common pattern in traffic flow throughout the day and make common estimates of traffic flow at different times, making it difficult to achieve deep extraction of temporal features. Therefore, this application proposes an ALTCM module capable of modeling multiple temporal patterns of traffic flow in a fine-grained manner.

[0058] The ALTCM module generates different model parameters for different times of day to facilitate the model's handling of various temporal patterns of traffic flow. Generally, a trade-off needs to be struck between fine-grained and coarse-grained modeling of traffic flow across multiple temporal patterns, as the former implies higher computational costs. To address this issue, the ALTCM module proposed in this application employs a lightweight design.

[0059] Specifically, the ALTCM module is based on time embedding E temporal Model parameters were generated, and the temporal embedding contained the temporal attributes of the input traffic flow. This process in the l-th spatiotemporal layer can be expressed by equations (8) and (9):

[0060]

[0061]

[0062] in, and These are the time-specific convolution kernel and bias generated for the input traffic flow at time t0, where t0 represents the time of the input data and K is the kernel size. and These are learnable parameters. and It is a bias.

[0063] Next, this application uses a temporal convolutional network to illustrate the main idea of ​​the ALTCM module. The ALTCM module uses a gating mechanism and convolutional operations in the temporal dimension to extract temporal features. This process in the l-th spatiotemporal layer can be expressed as formulas (10) and (11):

[0064]

[0065]

[0066] in, and is a feature of equal size; Dropout[.] represents the dropout operation; σ(.) is the sigmoid function; ⊙ is the Hadamard product.

[0067] It is important to note that in actual prediction scenarios, traffic flow data at different times needs to be input into the model in batches. Therefore, the ALTCM module needs to use grouped convolution operations to replace the standard convolution operation in formula (10) in order to improve efficiency through parallel processing.

[0068] Step S3: In the output layer, the outputs of the adaptive lightweight temporal convolutional modules in each spatiotemporal layer are skip-connected to obtain aggregated features, and then the predicted traffic flow is obtained through convolution operations.

[0069] In the output layer, the output features of the ALTCM module are skip-connected to obtain aggregated features, and then a prediction result is generated through convolution. The above operations can be expressed as formulas (12) and (13):

[0070]

[0071]

[0072] in, These are aggregated features following skip connections; and These are the convolution kernels for the skip connection operation and the convolution kernels for the first and second convolution operations in the output layer, respectively. and It is a bias; It is the predicted traffic flow.

[0073] To verify the technical solution of this application, experiments were conducted on four public datasets: PEMS03, PEMS04, PEMS07, and PEMS08. These four public datasets were all extracted from the California Transportation Authority Performance Measurement System and were summarized and recorded every 5 minutes, which means that there are 288 time steps in a day.

[0074] The following baseline methods are compared with the ALTCN-CD model of this application. The baseline methods include:

[0075] DCRNN: Diffusion Convolutional Recurrent Neural Network uses a recurrent neural network to perform temporal modeling and a GCN with a geographic distance-based graph to perform spatial modeling.

[0076] Graph WaveNet: Graph WaveNet uses dilated convolutions for temporal modeling and GCNs and data-driven graphs for spatial modeling.

[0077] STSGCN: Spatiotemporal Synchronization Graph Convolutional Network designs a spatiotemporal synchronization graph and uses GCN for spatiotemporal modeling.

[0078] STFGNN: Spatiotemporal Fusion Graph Neural Network aggregates multiple temporal and spatial graphs together, then uses dilated convolution to extract long-term temporal features, and uses GCN for spatiotemporal modeling.

[0079] STGODE: The Spatiotemporal Graph Ordinary Differential Equation Network uses dilated convolutions to extract long-term temporal features. Furthermore, it constructs distance-based and semantic-based graphs and uses ordinary differential equations for spatiotemporal modeling.

[0080] ASTGNN: Attention-based Spatiotemporal Graph Neural Network uses an attention mechanism for temporal modeling and graph update operations, and a GCN for spatial modeling.

[0081] SCINet: Sampling Convolution and Interaction Networks. A special recursive structure is designed based on the interaction of downsampled subsequences to model time.

[0082] Auto-DSTSGN: Automatically dilated spatiotemporal synchronization graph network constructs a graph using a neural architecture search method and performs spatial modeling using GCN, then performs temporal modeling using dilated convolution.

[0083] For the baseline methods described above, the experiments used the metrics published in their original literature to ensure a fair comparison. Specifically, three commonly used metrics were employed to evaluate the model's predictive performance: Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Root Mean Square Error (RMSE). Lower values ​​for MAE, MAPE, and RMSE indicate better predictive performance.

[0084] The experiment uses traffic flow data from the past 12 time steps (i.e., 1 hour) to predict traffic flow data for the next 12 time steps, i.e., T′ = T = 12. The sequence length L of the downsampled subsequence is... int The value was set to 6. Before model training, this experiment pre-calculated the forward difference value X of traffic flow at each time step. fd and backward difference value X bd The model was normalized using the Z-Score method. The hidden dimension D of the ALTCN-CD model was set to 32, the number of stacked spatiotemporal layers L was set to 5, the kernel size K for convolution operations in the ALTCM module was set to 2, and the batch size was set to 64. The learning rate was set to 0.001, the number of training generations was set to 150, and the early stopping tolerance was set to 25.

[0085] Based on the raw metrics provided by existing benchmarks and baseline methods, Table 1 lists the average prediction performance of various methods over 12 time steps. For fairness, this application conducted 5 experiments on the ALTCN-CD model to record the average results. Based on the comparison results, the following conclusions can be drawn:

[0086] Table 1

[0087]

[0088] (1) Overall, the indicators of each method are the worst on PEMS07 and the best on PEMS08, which indicates that the spatiotemporal pattern of traffic flow in PEMS07 is more complex than that in PEMS08.

[0089] (2) Compared with the latest traffic flow prediction models, the ALTCN-CD model proposed in this application shows comparable prediction performance on PEMS03, but shows the best prediction performance on PEMS04, PEMS07 and PEMS08, thus verifying the effectiveness and robustness of the ALTCN-CD model.

[0090] (3) DCRNN performed poorly on all four public datasets, suggesting that relying on predefined information to learn the spatiotemporal patterns of traffic flow may mislead the model.

[0091] (4) Graph WaveNet outperforms some recent methods (STSGCN, STFGNN, STGODE) on four public datasets, validating the effectiveness of spatiotemporal modeling in a data-driven manner.

[0092] (5) The state-of-the-art models based on GCN (Auto-DSTSGN) and attention mechanism (ASTGNN) perform well because they are capable of effective spatiotemporal modeling.

[0093] (6) SCINet, which only performs time modeling, shows competitive prediction performance, verifying the effectiveness of downsampled subsequence interaction learning.

[0094] Experiments compared the training and inference efficiencies of different methods on PEMS08. Specifically, this application compared the computational costs of six methods on the same server, and the results are shown in Table 2. It can be seen that although state-of-the-art traffic flow prediction models (i.e., ASTGNN and Auto-DSTSGN) improve prediction performance compared to Graph WaveNet, they all incur relatively high computational costs. The ALTCN-CD model proposed in this application has a training time of only 4.56 seconds per generation and an inference time of only 0.37 seconds, which is significantly more efficient than other methods and can be deployed on edge devices with limited hardware resources.

[0095] Table 2

[0096]

[0097] This application also conducted ablation experiments. In order to study the role of each component in the proposed ALTCN-CD model, the following variants of the ALTCN-CD model were designed for comparative analysis:

[0098] • w / o E-day: The ALTCN-CD model does not include daily embeddings; its time embeddings only include weekly embeddings.

[0099] • w / o E-week: An ALTCN-CD model that does not include week embeddings; its time embeddings only include day embeddings.

[0100] • w / o E-spatial: ALTCN-CD model without spatial embedding.

[0101] • w / o Context: The ALTCN-CD model does not include context information and does not take the forward and backward differences in traffic flow at each time step as input.

[0102] • w / o Adaptive: An ALTCN-CD model that does not include adaptive updates to model parameters, extracting temporal features in a coarse-grained manner.

[0103] • w / o Downsampling: The ALTCN-CD model does not downsample the input data and does not use interval sampling to reduce the sequence length. To process the original sequence of length T′ (set to 12), it stacks 11 spatiotemporal layers.

[0104] Five experiments were conducted for each variant of the ALTCN-CD model, and the mean and standard deviation of the experimental results are shown in Table 3.

[0105] Table 3

[0106]

[0107] The following conclusions can be drawn from the results in Table 3:

[0108] (1) All three metrics of the ALTCN-CD model are optimal, indicating that each component of the ALTCN-CD model contributes to the prediction of the ALTCN-CD model.

[0109] (2) The prediction performance of ALTCN-CD is significantly better than w / o E-day, w / o E-week and w / o E-spatial, indicating that data-driven spatiotemporal embedding can help the model to perform efficient spatiotemporal modeling.

[0110] (3) The prediction performance of w / o Context is significantly worse than that of ALTCN-CD, indicating that the naive downsampling strategy will cause information loss, thereby reducing the prediction performance of the model.

[0111] (4) The prediction performance data of w / o Adaptive show that fine-grained extraction of temporal features can further improve the prediction performance of the model.

[0112] (5) The prediction performance data with no downsampling shows only a slight difference compared to ALTCN-CD, indicating that the context-aware embedding layer proposed in this paper can shorten the sequence length while avoiding information loss. It is worth noting that providing contextual information without downsampling may lead to information redundancy, which may be misleading in some cases and reduce the model's prediction performance.

[0113] Unlike existing models that slightly improve traffic flow prediction accuracy at a relatively high computational cost, the proposed ALTCN-CD model follows a lightweight design principle, enabling its application to large-scale datasets with lower computational costs. Specifically, it proposes a context-aware downsampling strategy to downsample the input sequence, reducing its length and thus lowering computational costs. Simultaneously, it uses the forward and backward differences in traffic flow at each time step as contextual information, preserving the model's feature extraction capabilities while avoiding information loss. Furthermore, the proposed ALTCM module updates model parameters in a lightweight and fine-grained manner to handle various temporal patterns of traffic flow, facilitating deeper extraction of temporal features. Finally, the ALTCN-CD model effectively models the spatiotemporal patterns of traffic flow using spatiotemporal embedding. Notably, the ALTCN-CD model is purely data-driven, requiring no specific domain knowledge, and can therefore be easily extended to other spatiotemporal prediction tasks. Extensive experimental results on four public datasets validate the effectiveness and robustness of the ALTCN-CD model.

[0114] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A traffic flow prediction method based on an adaptive lightweight temporal convolutional network, characterized in that, The adaptive lightweight temporal convolutional network includes a context-aware embedding layer, stacked spatiotemporal layers, and an output layer. The traffic flow prediction method based on the adaptive lightweight temporal convolutional network includes: In the context-aware embedding layer, the input historical traffic flow is downsampled, and then contextual information is provided for the traffic flow at each time step after downsampling. This information is then aggregated with the time embedding to obtain the processed data. Spatial embedding, processed data, and temporal embedding are used as inputs to stacked spatiotemporal layers. For any spatiotemporal layer, the output of the previous spatiotemporal layer is aggregated with the spatial embedding and then passed through an adaptive lightweight temporal convolution module with the temporal embedding. Then, convolution and residual connections are performed, and finally batch normalization is used to obtain the output features of the current spatiotemporal layer. In the output layer, the outputs of the adaptive lightweight temporal convolutional modules in each spatiotemporal layer are skip-connected to obtain aggregated features, and then the predicted traffic flow is obtained through convolution operations. Among them, the l The adaptive lightweight temporal convolution module in each spatiotemporal layer performs the following operations: ； in, and It is a feature of equal dimensions; This indicates a dropout operation; It is the sigmoid function; It is the product of Hadamah. and They are respectively targeting The convolution kernel and bias generated by the input traffic flow at each time step. and These are learnable parameters. and It is a bias; For the first l Output features of the adaptive lightweight temporal convolution module in each spatiotemporal layer For the first l -1 spatiotemporal layer output features Indicates spatial embedding, Indicates residual connection, This indicates time embedding.

2. The traffic flow prediction method based on adaptive lightweight temporal convolutional networks as described in claim 1, characterized in that, The contextual information for traffic flow at each time step after downsampling is expressed as the following formula: ； in, This is the downsampled traffic flow sequence. It is a traffic flow sequence with contextual information; For connection functions; and These represent the forward and backward differences in traffic flow at each time step.

3. The traffic flow prediction method based on adaptive lightweight temporal convolutional networks as described in claim 1, characterized in that, The time-embedded aggregation yields the processed data, represented by the following formula: ； in For the processed data, For convolution kernel, For bias, For standard convolution operations, For the join function, It is a traffic flow sequence with contextual information. Embedded for time.

4. The traffic flow prediction method based on adaptive lightweight temporal convolutional networks as described in claim 3, characterized in that, The time embedding includes daily embedding and weekly embedding.

5. The traffic flow prediction method based on adaptive lightweight temporal convolutional networks as described in claim 1, characterized in that, The aggregated features are obtained by skip-connecting the outputs of the adaptive lightweight temporal convolutional modules in each spatiotemporal layer, as expressed by the following formula: ； in, These are aggregated features following skip connections; For convolution kernel; For bias, For the first l Output features of the adaptive lightweight temporal convolution module in each spatiotemporal layer.

Images