A spatiotemporal data prediction method and a data acquisition monitoring system

Abstract

The application provides a kind of spatio-temporal data prediction method and data acquisition monitoring system, the method comprises: according to the spatio-temporal data collected, establish topological space graph, record the graph signal sequence of each node of topological space graph at different time;Construct the spatio-temporal data prediction model based on spatio-temporal generative adversarial network, use historical graph signal sequence as training set, train spatio-temporal data prediction model;Spatio-temporal generative adversarial network includes generator and discriminator, generator is used to model the spatio-temporal dependence in spatio-temporal data, discriminator is used to regularize spatio-temporal generative adversarial network;Using the trained model to predict spatio-temporal data, obtain the graph signal sequence of each node at future preset time step.The application introduces the adversarial loss in the objective function of prediction model, to model the uncertainty in data, and learn the real spatio-temporal data distribution through the process of confrontation, strengthen the ability of prediction model to learn data representation, so as to alleviate the problem that multi-step prediction error grows too fast.

Description

Technical Field

[0001] This invention relates to the field of spatiotemporal data prediction technology, and in particular to a spatiotemporal data prediction method and data acquisition and monitoring system based on spatiotemporal generative adversarial networks. Background Technology

[0002] The real world generates a vast amount of spatiotemporal data. Utilizing this data to develop efficient urban applications, such as traffic prediction, has significant practical value and research implications. Spatiotemporal data is characterized by its large scale, high dimensionality, complex structure, heterogeneity, and high degree of nonlinearity. To address the problem of large-scale spatiotemporal data prediction, spatiotemporal data prediction models based on advanced deep learning technologies such as graph neural networks have been proposed and successfully applied to modeling complex spatiotemporal data. Among these models, the more advanced ones include those based on graph attention mechanisms and graph convolutional neural networks. Both belong to the message passing mechanism, connecting the network output with historical spatiotemporal information by aggregating information from relevant nodes.

[0003] While existing graph neural network (Graph Neural Network) techniques have achieved good results in spatiotemporal data prediction, they still have the following drawbacks: First, most existing spatiotemporal data prediction models use loss functions such as mean squared error (MSE) as objective functions to optimize model parameters. However, real-world spatiotemporal data has strong randomness, making it difficult to model this uncertainty using a specific, inflexible single objective function. Second, another problem with existing methods is the rapid growth of multi-step prediction errors. In multi-step prediction problems, existing methods are generally divided into two approaches: iterative multi-step prediction and direct multi-step prediction. Iterative multi-step prediction is more flexible and requires less data, so most spatiotemporal data prediction models adopt this approach, but it suffers from severe error accumulation problems in multi-step prediction tasks. Although spatiotemporal data prediction models using direct multi-step prediction strategies can alleviate the error accumulation problem in models based on iterative multi-step prediction strategies to some extent, it poses a significant challenge to the model's learning ability. Different models exhibit very large error variations as the prediction time steps increase in multi-step prediction tasks. Therefore, how to enable models to learn data features more effectively is a problem that needs in-depth research in spatiotemporal data prediction tasks.

[0004] Generative Adversarial Networks (GANs) are a class of neural networks based on adversarial learning processes. They can learn probability distributions from given real data and generate new data using the learned distributions. Leveraging the powerful modeling capabilities of GANs, recent research has considered incorporating GAN modeling ideas into prediction problems. For time-series data recorded by a single spatial location sensor, Koochali et al. introduced Conditional GANs to model the data distribution, overcoming the limitation of previous mean-regression strategies in modeling complex variable relationships. Wu et al. proposed a time-series prediction model based on GANs, improving multi-step prediction performance through the regularization of the discriminator. For spatiotemporal data prediction scenarios involving multiple spatial location sensors, Wang et al. proposed a sequence-to-sequence GAN for predicting urban pedestrian traffic data; however, this model uses traditional CNNs to model spatial dependencies, limiting its application to Euclidean structure data and preventing direct application to topological graph structure data. Summary of the Invention

[0005] To address the aforementioned problems, the present invention aims to provide a spatiotemporal data prediction method and a data acquisition and monitoring system. The method proposes an adversarial graph neural network model for spatiotemporal data prediction—Spatial Temporal GAN ​​(STGAN). This forms a framework combining a denoising spatiotemporal graph attention network and a generative adversarial network. Adversarial loss is introduced into the objective function of the prediction model to model the uncertainty in the data. Through the adversarial process, the model learns the real spatiotemporal data distribution, enhancing its ability to learn data representations, thereby alleviating the problem of excessively rapid growth in multi-step prediction errors.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] On the one hand, a spatiotemporal data prediction method is provided, including the following steps:

[0008] Based on the collected spatiotemporal data, a topological spatial graph is established, and the graph signal sequence of each node in the topological spatial graph at different times is recorded.

[0009] A spatiotemporal data prediction model based on a spatiotemporal generative adversarial network is constructed, and the model is trained using historical graph signal sequences as a training set.

[0010] The spatiotemporal generative adversarial network includes a generator and a discriminator. The generator is used to model the spatiotemporal dependencies in spatiotemporal data and predict generated data for a given input data. The discriminator is used to regularize the spatiotemporal generative adversarial network by sampling the generated data output by the generator and the real data and inputting them into the discriminator. When the discriminator cannot distinguish between the two, the spatiotemporal data prediction model is considered to have converged.

[0011] The trained spatiotemporal data prediction model is used to predict spatiotemporal data and obtain the graph signal sequence of each node at a preset time step in the future.

[0012] Preferably, the spatiotemporal data is traffic data collected from the urban road network.

[0013] Preferably, in the spatiotemporal generative adversarial network, the Wasserstein distance is used as the optimization objective to measure the difference between the real data distribution and the generated data distribution.

[0014] Preferably, in the spatiotemporal generative adversarial network, a Huber loss term is added to the generator's loss function as the generator's prediction loss.

[0015] Preferably, in the generator, a localized spatiotemporal graph is obtained through random walk theory to simulate the spatiotemporal dependencies in spatiotemporal data. Based on the localized spatiotemporal graph, the spatiotemporal dependencies between nodes are synchronously captured through a spatiotemporal graph attention module.

[0016] Preferably, the generator's input includes two time-period historical graph signal sequences: recent input and periodic input;

[0017] For recent input data, spatiotemporal features are first extracted through the spatiotemporal graph attention module, then periodic feature information from periodic input data is introduced using the self-attention module, and finally the predicted future graph signal sequence is generated through a fully connected layer.

[0018] Preferably, when training the spatiotemporal generative adversarial network, the generator parameters are updated once per iteration, and the corresponding discriminator parameters are updated multiple times per iteration.

[0019] On the other hand, a data acquisition and monitoring system is provided, which includes a data acquisition module, a data analysis and prediction module, a data service center, and a monitoring APP;

[0020] The data acquisition module is used to collect spatiotemporal data; the data analysis and prediction module is used to predict the collected spatiotemporal data according to the spatiotemporal data prediction method; the data service center stores historical spatiotemporal data, real-time spatiotemporal data, and business process data, and provides retrieval services; the monitoring APP is used to provide data query, display, online update, and modification, facilitating real-time monitoring by management personnel.

[0021] Preferably, the data acquisition module includes: a traffic measurement instrument, a data acquisition front-end, a serial port server, an industrial control computer, a display, and data acquisition software.

[0022] Preferably, the monitoring APP includes: a client, a server, and a system management backend;

[0023] The client is used for user registration and login, online query, modification, and logout; the server is used for registration and login verification, as well as data transmission, addition, modification, and deletion functions; the system management backend is used for database management.

[0024] Compared with the prior art, the technical solution provided by the present invention has the following beneficial effects:

[0025] This invention proposes a spatiotemporal data prediction method and a data acquisition and monitoring system based on spatiotemporal generative adversarial networks (GANs). A spatiotemporal GAN ​​is a framework combining graph attention networks and GANs, consisting of a generator and a discriminator. The generator models the spatiotemporal dependencies in the spatiotemporal data, while the discriminator regularizes the prediction model, enabling it to better learn the representation of the spatiotemporal data.

[0026] This invention introduces adversarial loss into the objective function of the prediction model to model the uncertainty in the data. By learning the real spatiotemporal data distribution through an adversarial process, it enhances the prediction model's ability to learn data representations, thereby alleviating the problem of excessively rapid growth in multi-step prediction errors. The effectiveness of the proposed method is verified on two real-world public transportation datasets.

[0027] This invention constructs a public transportation data acquisition and monitoring system, which consists of a data acquisition module, a data analysis and prediction module, and a monitoring APP that are interconnected to form a complete system that facilitates real-time monitoring. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a flowchart of the spatiotemporal data prediction method provided in the embodiments of the present invention;

[0030] Figure 2 This is a topological spatial diagram of traffic data provided in the embodiments of the present invention;

[0031] Figure 3 This is a schematic diagram of the STGAN adversarial learning process provided in an embodiment of the present invention;

[0032] Figure 4 This is a schematic diagram of the generator network structure provided in an embodiment of the present invention;

[0033] Figure 5 This is a localized spatiotemporal graph provided in the embodiments of the present invention;

[0034] Figure 6 This is a schematic diagram of the discriminator network structure provided in an embodiment of the present invention;

[0035] Figure 7 This is a comparison chart of the changes in the traffic speed prediction result MAE along the time step provided in the embodiments of the present invention;

[0036] Figure 8 This is a comparison chart of the changes in the traffic speed prediction result RMSE along the time step provided in the embodiments of the present invention;

[0037] Figure 9 This is a comparison chart of the changes in the traffic flow prediction result MAE along the time step provided in the embodiments of the present invention;

[0038] Figure 10 This is a comparison chart of the changes in the traffic flow prediction result RMSE along the time step provided in the embodiments of the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0040] Embodiments of the present invention provide a spatiotemporal data prediction method, such as... Figure 1 As shown, the method includes the following steps:

[0041] Based on the collected spatiotemporal data, a topological spatial graph is established, and the graph signal sequence of each node in the topological spatial graph at different times is recorded.

[0042] A spatiotemporal data prediction model based on a spatiotemporal generative adversarial network is constructed, and the model is trained using historical graph signal sequences as a training set.

[0043] The spatiotemporal generative adversarial network includes a generator and a discriminator. The generator is used to model the spatiotemporal dependencies in spatiotemporal data and predict generated data for a given input data. The discriminator is used to regularize the spatiotemporal generative adversarial network by sampling the generated data output by the generator and the real data and inputting them into the discriminator. When the discriminator cannot distinguish between the two, the spatiotemporal data prediction model is considered to have converged.

[0044] The trained spatiotemporal data prediction model is used to predict spatiotemporal data and obtain the graph signal sequence of each node at a preset time step in the future.

[0045] Specifically, the spatiotemporal data refers to traffic data collected from the urban road network. Given traffic data collected from the urban road network (road network), its spatial road network structure is defined as a topological spatial graph. in, Let be the set of nodes in the graph. Let A be the number of nodes, corresponding to N road segments in the road network; ε is the set of edges, representing the connectivity between nodes (road segments); A∈R N×N Let G be the adjacency matrix of graph G. If two nodes n i and n j If there is an edge between two road segments (e.g., two road segments are connected), then A ij =1, otherwise A ij =0, A ij The subscript ij represents the row and column index of matrix A; R is the set of real numbers.

[0046] like Figure 2 As shown, the road network is represented as a topological space graph G, and the graph signal sequence of each node in the topological space graph is recorded at different times. Figure 2 In the middle (a), the graph signal sequence of traffic data is represented. Represents the graphical signal observed in the road network at time t, where For node n i The signal value at time t, where N represents the number of nodes in the road network topology space, and F represents the number of variables observed by each node. Figure 2 In the diagram (b), the time series of three variables recorded by the sensor at a certain node is shown.

[0047] Assuming past T h At this point in time... (Time-Space-Feature 3D Tensor Data) represents a graph signal sequence of F variables recorded by N nodes, with its corresponding data distribution as follows: The goal here is to learn a function f that can predict the future T given an input. p Traffic flow at each node in the road network at each time step in Represents node n i The future target variable values ​​predicted from time t+1.

[0048] To improve the accuracy of spatiotemporal data prediction, this invention introduces adversarial learning to enhance the prediction model's ability to learn data representations. Here, let the real T... p The data distribution of traffic flow Y at each node in the road network at each time step is as follows: And the forecast The corresponding data distribution is Through adversarial learning, we hope to find a generator. To minimize and Difference between the two distributions (where Div(·,·) represents divergence, used to measure the difference between two distributions), and it is used as a spatiotemporal data prediction model for practical prediction tasks.

[0049] The spatiotemporal generative adversarial network (GAN) designed in this invention employs the following adversarial learning process: Figure 3 As shown, the model consists of two parts: a generator and a discriminator. The generator part contains a graph neural network module to model the spatiotemporal dependencies in the spatiotemporal data. During training, the generator and discriminator are computed iteratively in sequence. In each iteration, the input data is first fed into the generator, which then produces the output. The generator's output, i.e., the generated data and the real data, is sampled and input into the discriminator. The model converges when the spatiotemporal graph sequence generated by the generator is so similar to the real spatiotemporal graph sequence that the discriminator has difficulty distinguishing whether its input comes from the generator's output or the real data.

[0050] The adversarial learning process of the Spatiotemporal Generative Adversarial Network (STGAN) will be explained in detail below.

[0051] First, the principle of Generative Adversarial Networks (GANs) is as follows: given a dataset, a generator in the GAN is generated. and discriminator Learning is conducted through an adversarial process. The task is to distinguish whether the input comes from real data or not. The generated data;

[0052] The task is to generate data with the same distribution as the real data, so that... It is difficult to distinguish between generated data and real data, so a discriminator needs to be added. Error rate.

[0053] The objective function of GAN is defined as:

[0054]

[0055] in, This indicates that the input Y of the discriminator comes from the real data distribution. The mathematical expectation, Indicates the input of the discriminator Output from generator That is, generating data distribution The mathematical expectation of X (where X represents the generator's input). For the two different inputs mentioned above, the discriminator's outputs are respectively... or

[0056] and loss function and They are defined as follows:

[0057]

[0058]

[0059] The definitions of the two loss functions mentioned above raise a significant problem, namely... The stronger the discrimination ability, the better. The more severe the gradient vanishing phenomenon, the more difficult it becomes for the algorithm to effectively learn the distribution of real data. Assuming a fixed generator Consider the discriminator in the optimal case Integrating the objective function of the above GAN yields:

[0060]

[0061] With fixed generator network parameters, the discriminator is input with a specific data point Y, which may come from a real data distribution. It may also come from the distribution of generated data.

[0062] In equation (5) right Differentiation yields:

[0063]

[0064] The optimal discriminant, after simplification, is: In the generator's loss function Add irrelevant items Minimize equation (7) is equivalent to minimize equation (3):

[0065]

[0066] The optimal discriminator Substituting into equation (7), we get:

[0067]

[0068] Kullback-Leibler (KL) divergence and Jensen-Shannon (JS) divergence are widely used metrics for measuring the difference in data distributions. Given two distributions... and Their KL divergence and JS divergence are defined as follows:

[0069]

[0070]

[0071] Substituting equations (9) and (10) into equation (8), we get:

[0072]

[0073] During the learning process of the original GAN ​​model, the closer the discriminator is to the optimal state, the closer minimizing the generator's loss is to minimizing the loss of the original GAN ​​model. and The JS divergence between distributions is used. However, the JS divergence cannot effectively measure the differences between distributions, causing the generator loss to become a fixed value of 2log2, leading to the vanishing gradient phenomenon and reducing the generator's ability to learn the real data distribution.

[0074] To address the problems existing in the original GAN ​​mentioned above, this invention uses Wasserstein GAN (WGAN) as the adversarial learning framework for STGAN, and uses Wasserstein distance instead of JS divergence as the optimization objective, which can more effectively measure the distribution of real data. With the distribution of generated data The differences between them. The objective function of WGAN is defined as:

[0075]

[0076] Figure 3 A schematic diagram of the adversarial learning of the STGAN model proposed in this invention is given in the context of a spatiotemporal data prediction task. As stated in the problem definition, the spatiotemporal data prediction task addressed by this invention is to predict the future T. P The target variable values ​​corresponding to N spatial location sensors at each time step in Represents t+i (i = 1, 2, ..., T) p The actual data values ​​recorded by each spatial sensor at any given time. Generator Generated prediction data Dependent on its own input graph signal sequence

[0077] Given a set of real data and generator-generated data in The adversarial loss component in the STGAN objective function, representing the generator's output, is defined as:

[0078]

[0079] Given that GANs and WGANs are designed to generate new data, rather than predict the future, therefore, in order for the generator to generate data... While adhering to the true data distribution, and to get as close as possible to the true data Y, a Huber loss term is added to the generator's loss function as the generator's prediction loss:

[0080]

[0081] Combining equations (13) and (14), the final objective function of STGAN is obtained:

[0082]

[0083] Among them, parameters Used to control the importance of the Huber loss in the overall objective function. This applies to the spatiotemporal data generated by the generator. When the model converges because it is so similar to real spatiotemporal data Y that the discriminator has difficulty distinguishing whether its input comes from the generator's output or real data.

[0084] As shown in Algorithm 1, during the training of the STGAN model, the gradient is calculated using the mini-batch stochastic gradient descent method (batch size set to m). Given θ, which represents the generator of the STGAN... The model parameters, where ω represents the discriminator. The model parameters are updated using the RMSProp optimization algorithm. The gradients of the generator and discriminator are shown below. θ and ▽ ω They are defined as follows:

[0085]

[0086]

[0087]

[0088] Furthermore, the structure of the generator provided in this embodiment of the invention is described as follows:

[0089] Given a historical graph signal sequence X in As input, generator The task is to generate the future T p At each time step, the target data of each node in graph G Furthermore, the generated data It can approximate the real data Y as closely as possible, making the discriminator It is impossible to correctly determine whether the input data comes from real data Y or generated data. This achieves a "deceptively realistic" effect. The generator in the STGAN framework proposed in this invention is the DSTGAT model, and the model construction is as follows: Figure 4 As shown. By inputting a large amount of historical spatiotemporal data, the generator... It can learn the distribution of real spatiotemporal data And use the learned generation distribution Predict the value of variables at future moments.

[0090] In order for the predictive model to more effectively extract temporal and spatial dependencies from the data, the input spatiotemporal data... First, apply a convolution operation with a kernel size of 1×1 to the signal vector of each node in the topological graph G at each time step. 1×1 (·) mapped to higher-dimensional space F emb This represents the mapped feature dimension. Unlike sequence learning models such as LSTM, which can preserve the temporal order information in the data, the attention mechanism cannot distinguish the relative order of the data itself and directly assigns the encoded data X to the input data. emb Inputting data into the attention module fails to effectively capture temporal causal information. To obtain the order information of the input sequence, positional embeddings are added to the data. For input data X... in The learnable temporal encoding matrix Add to X emb In the process, the output of the input layer is obtained.

[0091] X * =Conv 1×1 (X in )+P (18)

[0092] The spatiotemporal dependencies of nodes in a topological graph encompass various types: the influence of neighboring nodes at the current moment, the influence of a node's historical state on itself, and the influence of the historical state of neighboring nodes on the node. These complex spatiotemporal dependencies are crucial for accurate prediction. To more accurately capture the spatiotemporal dependencies in the aforementioned data, this invention establishes a localized spatiotemporal graph.

[0093] Information propagation in a topological graph can be represented as a diffusion process, which exhibits the Markov property. Therefore, the diffusion convolution used in DCRNN, with the aid of a state transition matrix, represents the information propagation in the graph across consecutive time steps as a process with restart probabilities of... A random walk. Given a graph signal X. t ∈R N×F and convolution kernel f θ The diffusion convolution operation is defined as:

[0094]

[0095] Where * denotes diffraction convolution, τ is the defined diffraction stride, and θ∈R τ×2 For the convolution kernel f θ Learnable parameters; the out-degree diagonal matrix D of the directed graph. O =diag(A), in-degree diagonal matrix D I =diag(A T ). and Let represent the state transition matrix and its inverse transition matrix for the diffusion process, respectively. Although the above diffusion convolution operation is defined on directed graphs, it can be extended to undirected graphs. When applied to undirected graphs, it is equivalent to the spectral graph convolution operation approximated by Chebyshev polynomials.

[0096] From the definition of diffusion convolution above, it can be seen that during the information diffusion process, the graph signal data X at the same moment is used continuously. t Unlike other types of data, the information of each node in a graph remains constant. Therefore, when using graph convolution operations such as diffraction convolution, different diffraction steps can use the same graph signal data at the same time step. For spatiotemporal data, the node states at each time step change dynamically over time. For each diffraction step, the graph signal at its corresponding time step needs to be considered. Therefore, this invention utilizes random walk theory to obtain a localized spatiotemporal graph A. ST ∈R τN×τN To simulate the spatiotemporal dependencies in spatiotemporal data, τ represents the diffusion step size, A ST It contains τN spatiotemporal nodes, each of which has both spatial and temporal attributes. Figure 5(a) presents previous research such as DCRNN, which models temporal and spatial dependencies separately, artificially severing the spatiotemporal correlation in the data; Figure 5 (b) shows the spatiotemporal dependency of the localized spatiotemporal graph modeling proposed in this invention, which extracts graph signal information at different times rather than at a single time through random walks; Figure 5 When (c) is τ=3, the localized spatiotemporal graph A ST In the form of, multiple adjacency matrices A of different orders are used. (τ) Composition A ST Among them, A (τ) Let A represent an adjacency matrix of order τ. (1) =A, for spatial node n in the topological graph G i Its τ-order neighbors are those that can be reached by passing through τ spatial nodes, and n is the nearest neighbor. i A set of spatial nodes.

[0097] Self-attention, as a type of dot-product attention mechanism, has been widely applied to various natural language processing and computer vision tasks in recent years. Classical self-attention involves three types of data: query, key, and... F d These are the dimensions of three data vectors. Attention can be calculated as follows:

[0098]

[0099] in, These represent the matrix forms obtained by stacking the query, key, and value, respectively, with SoftMax being the activation function.

[0100] The classic self-attention mechanism described above cannot effectively extract spatiotemporal dependencies from data. Therefore, this invention, based on localized spatiotemporal graphs, further proposes spatiotemporal graph attention to synchronously capture complex spatiotemporal dependencies between nodes. This is achieved by considering the input of the spatiotemporal graph attention module in the l-th sub-network layer of different components. (where T) (l-1) Where N is the time dimension of the input data, and F is the spatial dimension, i.e., the number of nodes. (l-1) (As a feature dimension), the spatiotemporal graph attention first uses three different fully connected layers to capture the long-sequence temporal dependencies of each node in the input data, resulting in Q = W. Q H (l-1) K = W K H (l-1) and Among them, W Q W K and Considering local spatiotemporal correlations, a multi-head attention mechanism is used to integrate Q, K, and... Divide into multiple heads along the time dimension, and for each head Computational attention to extract local spatiotemporal dependent features:

[0101] Z * =Concat(head1,…,head) s ,…,head S ) (twenty one)

[0102] in, The output of multi-head attention is represented by matrix transformation: Restore the time dimension to T (l-1) =Sτ, the spatial dimension is restored to N. For Z * The output of the spatiotemporal graph attention module is obtained using a convolutional layer with a 1×3 kernel. The `Concat()` operator concatenates multiple headers, each... This corresponds to a set of graph signal data with τ consecutive time steps. The corresponding input data for spatiotemporal graph attention is... and The attention coefficient matrix M∈R is obtained by performing matrix multiplication between Q and K. τN×τN The attention coefficient vector M in the i-th row of matrix M i ∈R τN Corresponding to q i With K = {k1,…,k j ,…,k τN The dot product calculation result of}; where M ij For vector q i With k j The product value represents the attention coefficient of spatiotemporal node i to spatiotemporal node j.

[0103]

[0104] When capturing dynamic relationships in data, attention mechanisms offer greater flexibility compared to graph convolution operations. However, simply introducing attention can introduce significant noise, increasing the difficulty of model learning. Therefore, the localized spatiotemporal graph A... ST The mask matrix for attention, i.e., Mask = A ST This can effectively reduce the learning difficulty of the model. Specifically, the element Mask in the i-th row and j-th column of the matrix Mask... ij Indicates whether there is a dependency relationship between spatiotemporal node j (j = 1, 2, ..., τN) and spatiotemporal node i. Mask ij =0 indicates no dependency, Mask ij ≠0 indicates the existence of a dependency. In the spatiotemporal graph attention module, the information of spatiotemporal node i is aggregated by spatiotemporal node j (Mask) that has a dependency on it. ijInformation v (≠0) j To obtain the update, the weight α is calculated during aggregation using the SoftMax function. ij .

[0105]

[0106]

[0107]

[0108] in, This represents the state of spatiotemporal node i after the spatiotemporal graph attention module aggregates and updates the relevant spatiotemporal node information; α ij This represents the state vector v corresponding to spatiotemporal node j when spatiotemporal node i updates its own state. j The weight of its influence, when Mask ij When α = 0, the corresponding α ij =0.

[0109] Generator The input contains two time periods of historical graph signal sequences: recent input and periodic input For input data from a recent time period, X is first extracted using the spatiotemporal graph attention module. recent The spatiotemporal features are then used to introduce periodic input X using a self-attention module. week The data contains periodic characteristics. Finally, a predicted future graph signal is generated through a fully connected layer.

[0110] Specific generators The network architecture consists of three stacked sub-network layers (pattern transfer layers). (Recent input: X) recent The output is obtained after processing by the input layer. This data is used as input to the first pattern transfer layer. In the pattern transfer layer, the input data first passes through a spatiotemporal graph attention module to extract complex spatiotemporal dependencies from the data. The output of the spatiotemporal graph attention module... After residual and layer normalization, the input is fed into the self-attention module and mapped through a fully connected layer to obtain Q. s At the same time, X week The result obtained from the input layer The input is fed into the self-attention module and mapped through two different fully connected layers to obtain K. s and V s Using the obtained Q s K s and V s Calculate self-attention to obtain the output After residual and layer normalization, the input is fed into the feedforward layer. The output of the feedforward layer is then subjected to residual and layer normalization to obtain output O. (1) The output will be O (1) As input to the second pattern transfer layer, the corresponding output O is obtained. (2) Similarly, the output O of the third mode transfer layer is obtained. (3) At the end of this component, O (3) The generator's output is obtained through a fully connected layer.

[0111] Furthermore, the structure of the discriminator provided in the embodiments of the present invention is described as follows:

[0112] In adversarial learning, given real data As input, the discriminator's output Approximate 1 as closely as possible; given the data generated by the generator When used as input, the discriminator output The generator tries to approximate the output of the discriminator as close to 0 as possible. The goal is to approximate 1 as closely as possible. The discriminator's error is backpropagated into the generator to guide it in learning features from the spatiotemporal data, resulting in more realistic data. A good discriminator is essential for regularization. Therefore, during STGAN training, for every iteration of the generator parameters, the corresponding discriminator parameters need to be updated multiple times (n). critic .

[0113] In designing the discriminator network structure, the two sub-network modules mentioned above were also referenced and used: the input layer and the spatiotemporal graph attention module. Specifically, the discriminator... Network architecture such as Figure 6 As shown. For the input, the discriminator extracts spatiotemporal feature information from the data using a spatiotemporal graph attention module, and combines it with three fully connected layers to obtain the discriminator output.

[0114] Accordingly, embodiments of the present invention also provide a data acquisition and monitoring system, which includes a data acquisition module, a data analysis and prediction module, a data service center, and a monitoring APP;

[0115] The data acquisition module is used to collect spatiotemporal data; the data analysis and prediction module is used to predict the collected spatiotemporal data according to the spatiotemporal data prediction method; the data service center stores historical spatiotemporal data, real-time spatiotemporal data, and business process data, and provides retrieval services; the monitoring APP is used to provide data query, display, online update, and modification, facilitating real-time monitoring by management personnel.

[0116] The data acquisition module includes: traffic measurement instruments, data acquisition front-end, serial port server, industrial control computer, monitor, data acquisition software, etc.

[0117] The data acquisition module is a support system responsible for traffic data collection. Developed in C++, it incorporates multiple data communication protocols from IEC 60870-5, including 101, 102, 103, 104, Modbus, CDT, and DISA. Its modeling conforms to the interface reference model, common information model (CIM), and component interface specification (CIS) requirements of IEC 61970, meeting international standards and allowing for seamless integration with various systems as middleware. It enables the access of data from monitoring systems, integrated energy management systems, metering, fault analysis, and alarm push notifications. The system supports the access of various devices and possesses the ability to parse multiple protocols.

[0118] The data acquisition and monitoring system adopts a two-tier architecture: a single integrated energy management system and a cloud platform centralized monitoring system. The integrated energy management system collects real-time traffic operation monitoring data, enabling local data monitoring, historical data sampling and storage, and uploading of key real-time data to the cloud platform centralized monitoring system. The cloud platform centralized monitoring system retrieves real-time traffic monitoring data for monitoring traffic conditions. Communication between the two systems can adopt the IEC104 power standard or other protocols. The real-time data acquisition frequency supports second-level accuracy as required by the protocol, and supports modes such as variable transmission, cyclic transmission, and call-to-action.

[0119] The data service center uses a real-time historical database to store historical and real-time spatiotemporal data of the production process and provides retrieval services; the business SQL database (Oracle or MySQL) stores static data of business processes and provides retrieval services. The application method of offline data analysis using a big data analysis platform can support the real-time monitoring needs of photovoltaic building operation status and meet various application-oriented and theme-oriented analysis needs.

[0120] The data service center is designed with a real-time database system. The real-time database system is a new type of database management system software. The high-speed database engine developed based on the 64-bit system and the advanced distributed cluster architecture make the present invention suitable for the collection, storage, retrieval and publication of massive real-time / historical data. It has good horizontal scalability and high availability, and can handle dynamic data that changes rapidly over time, thereby improving the speed and efficiency of data retrieval and searching.

[0121] The monitoring app is mainly used to query and display relevant data information, and to update and modify it online, so as to facilitate real-time monitoring by management personnel.

[0122] The monitoring app mainly consists of a client, a server, and a system management backend. The client is developed using the MUI front-end framework and HTML5, CSS, and JavaScript for user registration and login, online querying, modification, and logout. The server uses the ThinkJS server-side framework and a MySQL database for registration and login verification, as well as data transmission, addition, modification, and deletion. The system management backend is developed using HTML5, CSS, and JavaScript for database management.

[0123] The monitoring app is simple and convenient to operate, with a clean and aesthetically pleasing interface. It offers real-time monitoring, allowing registered users to log in from anywhere via their mobile phones. The system provides automatic query and display functions, as well as user registration information management capabilities. The monitoring app can query the distribution of monitoring points in a given area, utilizing the GPS positioning function of an Android phone to load the distribution of each area onto a map online and obtain the coordinates of inflection points within that area.

[0124] The implementation process of the method of the present invention will be described in detail below through more specific embodiments.

[0125] The required traffic dataset is obtained through the data acquisition module, and the model performance is evaluated on two real public transportation datasets: PeMS08 and METR-LA.

[0126] (1) The PeMS08 dataset is traffic data collected from highways at a sampling frequency of once every 30 seconds.

[0127] (2) The traffic speed dataset METR-LA contains traffic speed data for a total of four months recorded by 207 sensors on highways.

[0128] The original data was re-aggregated at 5-minute intervals, resulting in 288 records from each sampling time point per day. The data recorded at each node in both datasets was standardized using the following method before being input into the neural network.

[0129]

[0130] In the above formula, mean(x) and std(x) are the operations for calculating the mean and standard deviation of the data x, respectively.

[0131] In the experimental section of this invention, LSTM, GCRN, ASTGCN, and generator methods are used as benchmark models for comparative experiments.

[0132] The STGAN model proposed in this invention is implemented using the PyTorch deep learning framework. The input and output time ranges are T, respectively. w =12, T r =24 and T p =12; δ in the generator prediction loss is set to 1; diffusion step size τ = 3 in the localized spatiotemporal graph. The mapping dimension F of the periodic input, recent input, and discriminator input layer in the generator is... emb Both datasets were set to 36. The METR-LA traffic speed dataset and the PeMS08 traffic flow dataset were each divided into three parts in a 6:2:2 ratio for training, validation, and testing, respectively. The batch size was set to 16, the RMSProp optimization algorithm was used, and the learning rate α was set to 0.0008. One iteration of the generator corresponds to n iterations of the discriminator. critic =3; Model training for 40 epochs.

[0133] Analysis of experimental results:

[0134] (1) Analysis of Traffic Speed ​​Prediction Results

[0135] Table 1 lists the average prediction results of each experimental model for the METR-LA traffic speed dataset over the next 15 minutes (3 time steps), 30 minutes (6 time steps), 60 minutes (12 time steps), and 12 time steps. The experimental results show that the proposed STGAN achieves relatively better prediction results compared to other benchmark models. Specifically, compared to the DSTGAT model, STGAN improves the average MAE and RMSE by 6.43% and 3.27% respectively over 12 time steps on this dataset.

[0136] Table 1 Comparison of Traffic Speed ​​Prediction Accuracy

[0137]

[0138] further, Figure 7 and Figure 8The paper presents the MAE and RMSE of each model's prediction results on the METR-LA dataset, showing their variation over time steps. In the short-term (first 6 time steps) prediction task, the STGAN model improves MAE and RMSE by 5.90% and 1.25% respectively compared to DSTGAT. In contrast, in the long-term (last 6 time steps) prediction task, STGAN improves MAE and RMSE by an average of 7.17% and 2.71% respectively compared to DSTGAT. As the prediction time steps extend, the increase in MAE and RMSE of the STGAN model is relatively slow compared to other methods. This reflects that the STGAN model proposed in this invention effectively mitigates the phenomenon of multi-step prediction error growth, verifying the superior performance of the adversarial learning strategy proposed in this invention.

[0139] LSTM employs an iterative multi-step prediction strategy, achieving MAEs of 5.15, 6.87, and 9.18 at 15-minute, 30-minute, and 60-minute prediction levels, respectively. The error increases by 1.72 from 15 to 30 minutes and by 2.31 from 30 to 60 minutes, showing a very rapid increase along the time step. This is partly due to its modeling only temporal correlations and neglecting spatial correlations. In contrast, GCRN, also using an iterative multi-step prediction strategy, exhibits a significantly smaller overall prediction error compared to LSTM due to its graph convolution operation modeling spatial correlations. Its error increases by 1.36 from 15 to 30 minutes and by 2.12 from 30 to 60 minutes, showing a reduced rate of error increase. Meanwhile, the ASTGCN and DSTGAT models, employing a direct multi-step prediction approach, show relative increases in MAE of 1.08 and 1.11 from 15 to 30 minutes, respectively, and relative increases of 1.69 and 1.86 from 30 to 60 minutes, respectively. This verifies that the direct multi-step prediction approach can alleviate the error accumulation problem to some extent. Finally, compared with ASTGCN and DSTGAT, the spatiotemporal generative adversarial network STGAN proposed in this invention achieved relatively better prediction results. The MAE value of its prediction results increased by 0.89 from 15 minutes to 30 minutes and by 1.48 from 30 minutes to 60 minutes.

[0140] (2) Traffic flow forecast results analysis

[0141] Table 2 lists the traffic flow prediction results of each experimental model on the PeMS08 dataset. Similar to the experimental results on the METR-LA dataset, the proposed STGAN achieves better prediction results compared to other benchmark models on the PeMS08 dataset. Compared to the DSTGAT model, STGAN improves the average MAE and RMSE by 2.39% and 3.42% respectively over 12 time steps on this dataset. Compared to ASTGCN, STGAN improves the average MAE and RMSE by 9.19% and 5.18% respectively.

[0142] Table 2 Comparison of Traffic Flow Forecast Accuracy

[0143]

[0144] Figure 9 and 10 The paper presents the time-step variation curves of MAE and RMSE of the prediction results of each experimental model on the PeMS08 dataset. Compared with the METR-LA dataset, the MAE and RMSE variation curves of each model on the PeMS08 dataset are more dramatic. In the short-term (first 6 time steps) prediction task, the STGAN model improves MAE and RMSE by 4.37% and 3.08% respectively compared to DSTGAT. In the long-term (last 6 time steps) prediction task, STGAN improves MAE and RMSE by an average of 4.04% and 3.83% respectively compared to DSTGAT. The experimental results of traffic flow prediction on the PeMS08 dataset further demonstrate the effectiveness of the proposed spatiotemporal generative adversarial network STGAN. By utilizing adversarial learning strategies, it can alleviate the phenomenon of multi-step prediction error growth and further improve the accuracy of spatiotemporal data prediction tasks.

[0145] (3) Analysis of weight parameters of generator loss function

[0146] This invention experimentally studies the generator loss function of STGAN. The impact of the value on its prediction accuracy This represents the proportion of the adversarial loss and Huber loss (prediction loss) that play a role in generator parameter updates in STGAN. The data shown in Table 3 represent four different... The values ​​are the MAE and RMSE values ​​of the STGAN model's average prediction results over 60 minutes (12 time steps).

[0147] As can be seen, on two different datasets, the prediction accuracy of STGAN varies with the prediction loss weight value. The decrease is initially accompanied by an increase followed by a decrease. STGAN's prediction of loss weights... The prediction error is greatest when the prediction loss weight is 0.00003. As the prediction loss weight increases, the STGAN prediction error decreases. The model performs optimally when the weight value is 0.0003. As the weight values ​​continue to increase, the prediction accuracy of STGAN decreases slightly. Experimental results show that when the generation loss weights are too large, the model struggles to accurately capture spatiotemporal dependency features, leading to a decline in model performance. When the prediction loss weights... If the size is too small, it becomes more difficult for the model to learn spatiotemporal dependent features, which will also affect the model's prediction accuracy.

[0148] Table 3. Impact of Generator Prediction Loss Term Weights on Prediction Results

[0149]

[0150] In summary, this invention addresses the challenges of using a single loss function to model uncertainties in spatiotemporal data and the rapid increase in multi-step prediction errors. It proposes a spatiotemporal generative adversarial network (STGAN) model for spatiotemporal data prediction, forming a framework combining a simplified denoising spatiotemporal graph attention network and a generative adversarial network, consisting of a generator and a discriminator. The generator is a spatiotemporal data prediction model used to model spatiotemporal dependencies in the data, while the discriminator regularizes the graph neural network to better learn the spatiotemporal data representation. An adversarial loss is introduced into the objective function of the prediction model to model uncertainties in the data, and the adversarial process learns the real spatiotemporal data distribution, enhancing the accuracy of the prediction model and thus mitigating the problem of excessively rapid growth in multi-step prediction errors. Experimental results applying STGAN to two types of traffic data validate its effectiveness.

[0151] This invention also constructs a data acquisition and monitoring system. The entire invention constitutes a complete traffic data analysis and monitoring system, providing a useful reference for research and development in related fields.

[0152] Embodiments of the present invention also provide an electronic device, which may vary considerably due to different configurations or performance. It may include one or more central processing units (CPUs) and one or more memories, wherein the memory stores at least one instruction, which is loaded and executed by the processor to implement the steps of the above-described spatiotemporal data prediction method.

[0153] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including instructions that can be executed by a processor in a terminal to perform the spatiotemporal data prediction method described above. For example, the computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, or optical data storage device.

[0154] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0155] The use of terms such as "an embodiment," "an embodiment," "an exemplary embodiment," and "some embodiments" in the specification indicates that the described embodiment may include a specific feature, structure, or characteristic, but not every embodiment necessarily includes that specific feature, structure, or characteristic. Furthermore, when a specific feature, structure, or characteristic is described in connection with an embodiment, implementing such a feature, structure, or characteristic in conjunction with other embodiments (whether explicitly described or not) should be within the knowledge of those skilled in the art.

[0156] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0157] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc.

[0158] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A spatiotemporal data prediction method, characterized in that, Includes the following steps: Based on the collected spatiotemporal data, a topological spatial graph is established, and the graph signal sequence of each node in the topological spatial graph at different times is recorded; the spatiotemporal data is traffic data collected in the urban road network; A spatiotemporal data prediction model based on a spatiotemporal generative adversarial network is constructed, and the model is trained using historical graph signal sequences as a training set. The spatiotemporal generative adversarial network includes a generator and a discriminator. The generator is used to model the spatiotemporal dependencies in spatiotemporal data and predict generated data for a given input data. The discriminator is used to regularize the spatiotemporal generative adversarial network by sampling the generated data output by the generator and the real data and inputting them into the discriminator. When the discriminator cannot distinguish between the two, the spatiotemporal data prediction model is considered to have converged. The generator is described as follows: Given a historical graph signal sequence X in As input, the generator G's task is to generate the future T. p At each time step, the target data of each node in graph G And the generated data Approximating the real data Y; For the input spatiotemporal data First, apply a convolution operation with a kernel size of 1×1 to the signal vector of each node in the topological graph G at each time step. 1×1 (·) mapped to higher-dimensional space F emb Indicates the feature dimension after mapping; To obtain the order information of the input sequence, positional embedding is added to the data; for the input data X in The learnable temporal encoding matrix Add to X emb In the process, the output of the input layer is obtained. X * =Conv 1×1 (X in )+P Information propagation in a topological spatial graph can be represented as a diffusion process, which exhibits Markov property; localized spatiotemporal graph A is obtained using random walk theory. ST ∈R τN×τN To simulate the spatiotemporal dependencies in spatiotemporal data, τ represents the diffusion step size, A ST It contains τN spatiotemporal nodes, each of which has both spatial and temporal attributes; Building upon the localized spatiotemporal graph, a spatiotemporal graph attention module is further proposed to synchronously capture complex spatiotemporal dependencies between nodes; the input of the spatiotemporal graph attention module in the l-th sub-network layer of different components is addressed. Among them, T (l-1) Where N is the time dimension of the input data, and F is the spatial dimension, i.e., the number of nodes. (l-1) As a feature dimension, the spatiotemporal graph attention module first uses three different fully connected layers to capture the long-sequence temporal dependencies of each node in the input data, obtaining Q = W. Q H (l-1) K = W K H (l-1) and Among them, W Q W K and Considering local spatiotemporal correlations, a multi-head attention mechanism is used to integrate Q, K, and... Divide into multiple heads along the time dimension, and for each head Computational attention to extract local spatiotemporal dependent features: Z * =Concat(head1,…,head s ,…,head S ) in, The output of multi-head attention is represented by matrix transformation: Restore the time dimension to T (l-1) =Sτ, spatial dimension restored to N; for Z * The output of the spatiotemporal graph attention module is obtained using a convolutional layer with a 1×3 kernel. The `Concat()` operator concatenates multiple headers, each... The graph signal data corresponds to a set of τ consecutive time steps; the input data for its corresponding spatiotemporal graph attention is... and The attention coefficient matrix M∈R is obtained by performing matrix multiplication between Q and K. τN×τN The attention coefficient vector M in the i-th row of matrix M i ∈R τN Corresponding to q i With K = {k1,…,k j ,…,k τN The dot product calculation result of}; where M ij This represents the attention coefficient of spatiotemporal node i to spatiotemporal node j; Localized spatiotemporal graph A ST The mask matrix for attention, i.e., Mask = A ST This reduces the learning difficulty of the model; where the element in the i-th row and j-th column of the matrix Mask is Mask. ij Indicate whether there is a dependency relationship between spatiotemporal node j, j = 1, 2, ..., τN, and spatiotemporal node i. ij =0 indicates no dependency, Mask ij ≠0 indicates the existence of a dependency; in the spatiotemporal graph attention module, the information of spatiotemporal node i is aggregated by aggregating the information v of spatiotemporal node j that has a dependency on it. j Get updates, Mask ij ≠0; The input to generator G consists of two time periods of historical graph signal sequences: recent input. and periodic input For input data from a recent time period, X is first extracted using the spatiotemporal graph attention module. recent The spatiotemporal features are then used to introduce periodic input X using a self-attention module. week The data contains periodic features; finally, a predicted future graph signal is generated through a fully connected layer. The specific generator G network architecture consists of three stacked sub-network layers, with the recent input X... recent The output is obtained after processing by the input layer. This is used as the input to the first pattern transfer layer; in the pattern transfer layer, the input data first passes through the spatiotemporal graph attention module to extract complex spatiotemporal dependencies from the data; the output of the spatiotemporal graph attention module... After residual and layer normalization, the input is fed into the self-attention module and mapped through a fully connected layer to obtain Q. s At the same time, X week The result obtained from the input layer The input is fed into the self-attention module and mapped through two different fully connected layers to obtain K. s and V s； Using the obtained Q s K s and V s Calculate self-attention to obtain the output ; After residual and layer normalization, the input is fed into the feedforward layer. The output of the feedforward layer is then subjected to residual and layer normalization to obtain output O. (1) The output will be O (1) As input to the second pattern transfer layer, the corresponding output O is obtained. (2) Similarly, the output O of the third mode transfer layer is obtained. (3) ; at the end of this component, O (3) The generator's output is obtained through a fully connected layer. The trained spatiotemporal data prediction model is used to predict spatiotemporal data and obtain the graph signal sequence of each node at a preset time step in the future.

2. The spatiotemporal data prediction method according to claim 1, characterized in that, In the spatiotemporal generative adversarial network, the Wasserstein distance is used as the optimization objective to measure the difference between the real data distribution and the generated data distribution.

3. The spatiotemporal data prediction method according to claim 1, characterized in that, In the spatiotemporal generative adversarial network, a Huber loss term is added to the generator's loss function as the generator's prediction loss.

4. The spatiotemporal data prediction method according to claim 1, characterized in that, The generator's input includes two time-period historical graph signal sequences: recent input and periodic input; For recent input data, spatiotemporal features are first extracted through the spatiotemporal graph attention module, then periodic feature information from periodic input data is introduced using the self-attention module, and finally, a predicted future graph signal sequence is generated through a fully connected layer.

5. The spatiotemporal data prediction method according to claim 1, characterized in that, During the training of the spatiotemporal generative adversarial network, the generator parameters are updated once per iteration, while the corresponding discriminator parameters are updated multiple times per iteration.

6. A data acquisition and monitoring system, characterized in that, The data acquisition and monitoring system includes a data acquisition module, a data analysis and prediction module, a data service center, and a monitoring APP; The data acquisition module is used to collect spatiotemporal data; the data analysis and prediction module is used to predict the collected spatiotemporal data using the spatiotemporal data prediction method according to any one of claims 1-5; the data service center stores historical spatiotemporal data, real-time spatiotemporal data, and business process data, and provides retrieval services; the monitoring APP is used to provide data query, display, online update, and modification, facilitating real-time monitoring by management personnel.

7. The data acquisition and monitoring system according to claim 6, characterized in that, The data acquisition module includes: traffic measurement instrument, data acquisition front end, serial port server, industrial control computer, display, and data acquisition software.

8. The data acquisition and monitoring system according to claim 6, characterized in that, The monitoring app includes: a client, a server, and a system management backend; The client is used for user registration and login, online query, modification, and logout; the server is used for registration and login verification, as well as data transmission, addition, modification, and deletion functions; the system management backend is used for database management.

Images