A matrix-guided interpretable spatiotemporal integrated prediction method and system

Abstract

The present application provides a matrix-guided interpretable spatiotemporal integrated prediction method, system and electronic device, which belongs to the field of geographic information technology. In the integration strategy, a spatiotemporal embedding module is set for each base learner, and a spatial embedding branch and a temporal embedding branch are set in each spatiotemporal embedding module. First, each spatiotemporal embedding module uses a non-negative matrix decomposition method to decompose the spatiotemporal heterogeneity integration weights of the base learners into the product of the spatial factor matrix and the temporal factor matrix, and then the mapping relationship between the spatial correlation information and the spatial factor matrix is ​​modeled through the spatial embedding branch, and the mapping relationship between the temporal attribute information and the temporal factor matrix is ​​modeled through the temporal embedding branch, and then the output features of each base learner are integrated to obtain the integrated prediction result. By integrating matrix decomposition and integrated learning, guided by the solution process of the two factor matrices, the coordinated and precise expression of spatial heterogeneity and temporal non-stationarity is achieved.

Description

Technical Field

[0001] The present application relates to the field of geographic information technology, and in particular to a matrix-guided interpretable spatiotemporal integrated prediction method, system and electronic device. Background Art

[0002] Spatio-temporal prediction is a hot research topic in Geo-AI. This method combines time series prediction and spatial analysis techniques to infer the attributes or characteristics of a future spatiotemporal location by modeling the dependencies of spatiotemporal data. It has a wide range of applications in many fields, such as environment and climate (e.g. weather forecasting), public safety (e.g. crime prediction), intelligent transportation (e.g. traffic flow prediction), and human activities (e.g. human trajectory pattern mining).

[0003] In order to improve the prediction performance and generalization ability of spatiotemporal prediction models, the industry has introduced ensemble learning into spatiotemporal data modeling. By constructing and combining multiple heterogeneous base learners, the advantages of different single methods are combined to complete the prediction task. This ensemble modeling strategy has been proven by a large number of literatures to achieve stronger generalization ability and higher prediction accuracy than single modeling strategies. However, due to the dynamic randomness of geographic processes and the spatiotemporal heterogeneous distribution of geospatial data, the application of ensemble learning in spatiotemporal prediction still faces many challenges, which limits its application in the field of geoscience.

[0004] The design of ensemble strategies mainly includes statistical and learning-based methods, which aim to solve the ensemble weights of different base learners. Existing studies usually design ensemble strategies based on the assumption of sample independent and identical distribution, so that the weights of different base learners are globally fixed in the entire spatiotemporal range. However, since geographical phenomena vary significantly in different geographical locations and change dynamically over time, spatiotemporal data show obvious spatial heterogeneity and temporal non-stationarity, which may make global-based ensemble strategies perform poorly in some local spatiotemporal locations. Although some existing studies use geographically weighted regression models to obtain a set of weights based on spatial locations to integrate the spatial prediction results of different base learners, the weights are shared for all moments and the temporal non-stationarity caused by time-varying patterns cannot be captured. In addition, some studies use the K nearest neighbor method to find K similar historical samples for weighted prediction of the prediction samples at each moment, but the weights are shared for all spatial locations and spatial heterogeneity is ignored, which limits the prediction accuracy of the model.

[0005] Therefore, it is necessary to provide an improved technical solution to address the above-mentioned deficiencies in the prior art. Summary of the invention

[0006] The purpose of the present application is to provide a matrix-guided interpretable spatiotemporal integrated prediction method, system and electronic device to solve or alleviate the problem of insufficient coordinated expression of spatial heterogeneity and temporal non-stationarity existing in the above-mentioned prior art.

[0007] In order to achieve the above objectives, this application provides the following technical solutions:

[0008] In a first aspect, the present application provides a matrix-guided interpretable spatiotemporal integrated prediction method, comprising:

[0009] Obtain the spatiotemporal state matrix of the samples consisting of the monitoring values ​​of all spatial monitoring stations;

[0010] Using the first spatiotemporal prediction model, the second spatiotemporal prediction model, and the third spatiotemporal prediction model as base learners of the ensemble learning method, inputting the spatiotemporal state matrix into the base learners respectively, and obtaining the first output feature, the second output feature, and the third output feature correspondingly;

[0011] Integrating the first output feature, the second output feature, and the third output feature through an integration strategy to obtain a final integrated prediction result;

[0012] Among them, the integration strategy is provided with parallel spatiotemporal embedding modules corresponding to each base learner, and each of the spatiotemporal embedding modules decomposes the spatiotemporal heterogeneity integration weights of each base learner into the product of a spatial factor matrix and a temporal factor matrix by means of non-negative matrix decomposition. Each of the spatiotemporal embedding modules is provided with a spatial embedding branch and a temporal embedding branch; the spatial embedding branch is used to model the mapping relationship between the spatial correlation information of the monitoring site and the spatial factor matrix, and the temporal embedding branch is used to model the mapping relationship between the temporal attribute information of the monitoring site and the temporal factor matrix, thereby solving the spatiotemporal heterogeneity integration weights.

[0013] In combination with the first aspect, optionally, the first spatiotemporal prediction model is a spatiotemporal difference moving average autoregressive model;

[0014] The spatiotemporal difference moving average autoregressive model introduces a spatial weight matrix on the basis of the integrated moving average autoregressive model, transforms the non-stationary spatiotemporal state into a stationary spatiotemporal state through differentiation, and regresses the historical values ​​of the predicted variables and the present values ​​and lagged values ​​of the random error terms to obtain the first output feature.

[0015] In combination with the first aspect, optionally, the second spatiotemporal prediction model is a spatiotemporal K nearest neighbor model;

[0016] The spatiotemporal K nearest neighbor model introduces spatial information on the basis of the traditional K nearest neighbor method, maps the spatiotemporal state matrix to a new feature space, selects K spatial neighbors with the most similar features within the entire historical time range for weighted prediction, and obtains the second output feature.

[0017] In combination with the first aspect, optionally, the third spatiotemporal prediction model is a spatiotemporal graph convolutional neural network model;

[0018] The spatiotemporal graph convolutional neural network model uses a graph model to model the non-Euclidean structure of the spatial monitoring site, and adopts graph convolution operations and gated one-dimensional temporal convolution Gated TCN to model the complex association of spatial and temporal dimension features to obtain the third output feature.

[0019] In combination with the first aspect, optionally, the integration strategy adopts an integration strategy of multivariate linear regression.

[0020] In combination with the first aspect, optionally, the spatial embedding branch includes a first encoding unit, a first concatenation unit, and a first learning unit, the first learning unit includes two learnable network layers, each network layer includes a linear fully connected layer with Dropout and a ReLU activation function;

[0021] Accordingly, modeling the mapping relationship between the spatial correlation information of the monitoring sites and the spatial factor matrix includes the following steps:

[0022] The first encoding unit constructs a graph proximity matrix A based on the distance and attribute similarity between the monitoring sites, wherein the graph proximity matrix A includes a graph proximity matrix based on the spatial Euclidean distance and the graph proximity matrix based on the similarity of the monitored values

[0023] The first encoding unit calculates a Laplace normalized matrix according to the graph proximity matrix A, and performs eigendecomposition on the Laplace normalized matrix to obtain an eigenvalue matrix and an eigenvector;

[0024] The first encoding unit constructs a spatial embedding based on the spatial Euclidean distance based on the eigenvalue matrix and the eigenvector. and spatial embedding based on similarity of monitoring values

[0025] The first splicing unit embeds the space based on the spatial Euclidean distance and spatial embedding based on similarity of monitoring values Perform channel superposition to obtain the spatial embedding E′ s ; and embed the space into E′ sInput into the first learning unit to establish a mapping relationship between sample space association information and spatial factor matrix.

[0026] In combination with the first aspect, optionally, the time embedding branch includes a second encoding unit, a third encoding unit, a second concatenation unit, and a second learning unit, the second learning unit includes two learnable network layers, each network layer includes a linear fully connected layer with Dropout and a ReLU activation function;

[0027] Accordingly, modeling the mapping relationship between the time attribute information of the monitoring site and the time factor matrix includes the following steps:

[0028] The second encoding unit uses trigonometric function position encoding to relatively encode the time position of the sample at time t in the global range to obtain the time position embedding

[0029] The third encoding unit selects m time attributes of the sample at time t that are related to the integrated weight, and uses a one-hot encoding method (one-hot method) to encode each of the time attributes into a corresponding time attribute embedding 1≤j≤m;

[0030] The second splicing unit embeds the time position into With each of the time attributes embedded Splice and get the time embedding E′ t ; and embed the time into E′ t Input into the second learning unit to establish a mapping relationship between sample time information and time factor matrix.

[0031] In combination with the first aspect, optionally, the method further includes:

[0032] Calculate the spatial score vector and the temporal score vector of each base learner at different spatiotemporal positions according to the spatial factor matrix and the temporal factor matrix;

[0033] Based on the spatial score vector and the temporal score vector, response rules of different base learners under the conditions of spatial heterogeneity and temporal non-stationarity are analyzed.

[0034] In a second aspect, this embodiment provides a matrix-guided interpretable spatiotemporal integrated prediction system, including:

[0035] A data acquisition module configured to acquire a spatiotemporal state matrix of samples consisting of monitoring values ​​of all spatial monitoring sites;

[0036] A base learner module, configured to use the first spatiotemporal prediction model, the second spatiotemporal prediction model, and the third spatiotemporal prediction model as base learners of the ensemble learning method, input the spatiotemporal state matrix into the base learner respectively, and obtain the first output feature, the second output feature, and the third output feature accordingly;

[0037] An integration module, configured to integrate the first output feature, the second output feature, and the third output feature through an integration strategy to obtain a final integrated prediction result;

[0038] Among them, the integration strategy is provided with parallel spatiotemporal embedding modules corresponding to each base learner, and each of the spatiotemporal embedding modules decomposes the spatiotemporal heterogeneity integration weights of each base learner into the product of a spatial factor matrix and a temporal factor matrix by means of non-negative matrix decomposition. Each of the spatiotemporal embedding modules is provided with a spatial embedding branch and a temporal embedding branch; the spatial embedding branch is used to model the mapping relationship between the spatial correlation information of the monitoring site and the spatial factor matrix, and the temporal embedding branch is used to model the mapping relationship between the temporal attribute information of the monitoring site and the temporal factor matrix, thereby solving the spatiotemporal heterogeneity integration weights.

[0039] In a third aspect, this embodiment provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the matrix-guided interpretable spatiotemporal integrated prediction method as described in any of the above embodiments.

[0040] In a fourth aspect, this embodiment provides an electronic device, comprising: a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the matrix-guided interpretable spatiotemporal integrated prediction method as described in any of the above embodiments is implemented.

[0041] The technical solution of the embodiment of the present application has the following beneficial effects:

[0042] In the technical solution of this embodiment, the base learners of the integrated learning method include three base learners. In the integrated strategy, parallel spatiotemporal embedding modules corresponding to each base learner are set, and a spatial embedding branch and a temporal embedding branch are set in each spatiotemporal embedding module. First, each spatiotemporal embedding module uses a non-negative matrix decomposition method to decompose the spatiotemporal heterogeneity integration weights of each base learner into the product of a spatial factor matrix and a temporal factor matrix. Then, the spatial embedding branch is used to model the mapping relationship between the spatial correlation information of the monitoring site and the spatial factor matrix, and the value of each element in the spatial factor matrix is ​​obtained. The temporal embedding branch is used to model the mapping relationship between the temporal attribute information of the monitoring site and the temporal factor matrix, and the value of each element in the temporal factor matrix is ​​obtained. Then, the spatial factor matrix and the temporal factor matrix are multiplied to solve the spatiotemporal heterogeneity integration weight, and then the spatiotemporal heterogeneity integration weight is used to integrate the output features of the three base learners to obtain the final integrated prediction result. This scheme integrates matrix decomposition and ensemble learning, guided by the solution process of two factor matrices (i.e., spatial factor matrix and temporal factor matrix), to achieve the coordinated and precise expression of the spatial heterogeneity and temporal non-stationarity of the integration process. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The drawings constituting part of the present application are used to provide a further understanding of the present application. The exemplary embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. Among them:

[0044] Figure 1 A schematic diagram of a process of a matrix-guided interpretable spatiotemporal integrated prediction method provided according to some embodiments of the present application Figure 1 .

[0045] Figure 2 An overall block diagram of a matrix-guided interpretable spatiotemporal integrated prediction method provided according to some embodiments of the present application.

[0046] Figure 3 A logical block diagram of a spatiotemporal embedding module provided according to some embodiments of the present application.

[0047] Figure 4 A schematic diagram of the structure of a matrix-guided interpretable spatiotemporal integrated prediction system provided according to some embodiments of the present application.

[0048] Figure 5 A schematic diagram of the structure of an electronic device provided according to some embodiments of the present application. DETAILED DESCRIPTION

[0049] The present application will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments. Each example is provided by way of explanation of the present application and does not limit the present application. In fact, it will be clear to those skilled in the art that modifications and variations may be made in the present application without departing from the scope or spirit of the present application. For example, a feature shown or described as a part of an embodiment may be used in another embodiment to produce yet another embodiment. Therefore, it is desired that the present application includes such modifications and variations within the scope of the appended claims and their equivalents.

[0050] In the following description, the terms "first / second / third" involved are merely used to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that "first / second / third" can be interchanged with a specific order or sequence where permitted, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein.

[0051] The solution provided in the embodiments of the present application can be applied to electronic devices such as mobile phones, tablet computers, wearable devices, vehicle-mounted devices, augmented reality (AR) / virtual reality (VR) devices, laptop computers, ultra-mobile personal computers (UMPCs), netbooks, personal digital assistants (PDAs), etc. The embodiments of the present application do not impose any restrictions on the specific types of electronic devices.

[0052] This embodiment provides a matrix-guided interpretable spatiotemporal integrated prediction method. Figure 1 A schematic diagram of the process of the method is shown. Figure 2 This is the overall block diagram of the method.

[0053] like Figure 1 , Figure 2 As shown, the method may include steps S101 to S103:

[0054] Step S101: Obtain a spatiotemporal state matrix of samples consisting of monitoring values ​​of all spatial monitoring sites.

[0055] In this embodiment, a spatial monitoring station refers to a fixed location set up in geographic space for observing and monitoring specific environmental, ecological, meteorological, geological or other spatial phenomena. These stations collect and record data to obtain monitoring values ​​of various phenomena at specific locations, helping relevant personnel analyze and understand various natural and man-made phenomena. It should be noted that the geographical location of the spatial monitoring station is usually fixed, and each monitoring station has clear geographic coordinates to ensure the spatial consistency and comparability of the data. The monitoring station is equipped with various sensors and instruments for real-time or regular data collection. These data may include temperature, humidity, rainfall, wind speed, air quality, water quality, soil moisture, seismic activity, etc.

[0056] In this embodiment, there are multiple spatial monitoring sites, and the multiple spatial monitoring sites collect data at different time points to obtain monitoring values, and each sample corresponds to a spatiotemporal state matrix.

[0057] Specifically, define the spatial-temporal state: define is the monitoring value of the spatial monitoring site v at the tth time step, and the spatiotemporal state matrix X of the sample (t-H):t ∈R N×H It is defined as the monitoring values ​​of all spatial monitoring stations in the past H time steps, where N is the number of spatial monitoring stations, t is the time corresponding to the sample, H is the total number of historical time steps of the sample, and R is a set of real numbers.

[0058] Step S102: Using the first spatiotemporal prediction model, the second spatiotemporal prediction model, and the third spatiotemporal prediction model as base learners of the ensemble learning method, inputting the spatiotemporal state matrix into the base learners respectively, and obtaining the first output feature, the second output feature, and the third output feature accordingly.

[0059] This embodiment aims to integrate the spatiotemporal output features of different single methods to predict the future state of the target variable. Therefore, in this embodiment, three spatiotemporal prediction models are first designed as base learners, namely the first spatiotemporal prediction model, the second spatiotemporal prediction model, and the third spatiotemporal prediction model, and then the spatiotemporal state matrix X of the sample is used. (t-H):t , get the base learner output feature F for T time steps in the future at time t l (t+1):(t+T) ,in After that, an integrated model is constructed with the output features of the three base learners as input to predict the spatiotemporal state of the next T time steps. in

[0060] It should be noted that, in this embodiment, any spatio-temporal prediction model can be selected for the first spatio-temporal prediction model, the second spatio-temporal prediction model, and the third spatio-temporal prediction model. For example, the spatio-temporal Kriging model (Spatio-Temporal Kriging), the spatio-temporal autoregressive model (Spatio-Temporal Autoregressive Model, STAR), and the spatio-temporal convolutional neural network (Spatio-Temporal Convolutional Neural Network, ST-CNN) can be selected as the base learner of the ensemble learning. Different selections of spatio-temporal prediction models should not constitute a limitation on this embodiment.

[0061] Preferably, in the geographic spatiotemporal relationship modeling, the state of a spatial monitoring site at a future moment is not only related to the state of the site at a previous historical moment, but also affected by the historical state of the surrounding sites. Therefore, considering the local spatiotemporal dependence of spatiotemporal data, in some possible implementations, the first spatiotemporal prediction model is a spatio-temporal autoregressive integrated moving average model (ST-ARIMA).

[0062] In this embodiment, the ST-ARIMA model is one of the main means of modeling the local spatiotemporal dependence of spatiotemporal data. This model introduces a spatial weight matrix based on the integrated moving average autoregressive (ARIMA) model, transforms the non-stationary spatiotemporal state into a stationary spatiotemporal state through differentiation, and regresses the historical values ​​of the prediction variables and the present value of the random error term with the lagged value to obtain the first output feature.

[0063] In this embodiment, in order to solve the base learner ST-ARIMA output feature (first output feature) of the sample at time t, the ST-ARIMA model sequentially solves the state of each future prediction time step at each spatial monitoring site of the prediction sample. The next step I-order difference feature of the current time step t The calculation method is:

[0064]

[0065] Where p and q are the orders of the autoregressive term and the moving average term, respectively, I is the difference order, N is the number of spatial monitoring sites, and W is n is the spatial vector of the spatial proximity matrix at site n, and They represent the I-order difference state (historical value of the predicted variable) and the normal distribution random error (random error term) of the spatial monitoring station v at the historical time step ti, respectively. is the normally distributed random error of the spatial monitoring station v at the tth time step, represents the autoregressive parameters at time lag i and spatial lag n, θ in represents the moving average parameter at time lag i and spatial lag n.

[0066] Among them, the order p of the autoregressive term and the order q of the moving average term of the ST-ARMA model can be determined by referring to the order identification method of the ARIMA model, or by testing the spatio-temporal autocorrelation function (Spatio-Temporal Autocorrelation Function, STACF) and the spatio-temporal partial autocorrelation function (Spatio-Temporal Partial Autocorrelation Function, STPACF). They can also be identified by the Akaike Information Criterion (AIC) method. In addition, with the support of historical sample data, grid search can be used to determine the optimal order.

[0067] After determining the order of the ST-ARMA model, the least squares estimation (LSE) and maximum likelihood estimation (MLE) can be used to estimate the model parameters in equation (1), and then calculate the predicted state of the next time step. According to the above steps, each future prediction time step at each spatial monitoring station is solved in turn to obtain the spatiotemporal output characteristics of the base learner ST-ARMA model of all samples at time t (First output feature).

[0068] Since the modeling scope of the ST-ARIMA model is limited to the historical time step of the sample, only the states of the first few time steps of the current moment are selected to participate in the model operation. This ignores the information of past historical samples and limits the expression of the global relationship of spatiotemporal features. Therefore, considering the global feature similarity of spatiotemporal data, in some possible implementations, the second spatiotemporal prediction model is the spatio-temporal K-nearest Neighbours model (Spatio-Temporal K-Nearest Neighbours, ST-KNN). The spatio-temporal K-nearest neighbor model introduces spatial information on the basis of the traditional K-nearest neighbor method, maps the spatio-temporal state matrix to a new feature space, selects the K spatial neighbors with the most similar features in the entire historical time range for weighted prediction, and obtains the second output feature.

[0069] In this embodiment, the ST-KNN model is selected as the second spatiotemporal prediction model, which can not only realize the feature similarity modeling of spatiotemporal data in a global scope, but also realize the coupling of local spatiotemporal correlation and global feature similarity through the combination of the ST-KNN model and the ST-ARIMA model.

[0070] Specifically, in order to solve the base learner ST-KNN output feature (second output feature) of the sample at time t, the ST-KNN model solves the predicted state of each spatial monitoring site of the sample in turn. First, determine the time window τ of the site v of the predicted sample at time t v and spatial neighbors N v,ρ , and then construct a spatiotemporal neighbor matrix for the samples of each spatial monitoring site, which is the new feature space.

[0071] Among them, the time window τ v represents the history of τ time steps related to the current moment, τ v ∈[1,H], H is the historical time step of the entire sample.

[0072] For the spatial neighbors N v,ρ , ST-KNN calculates the distance Δ(v,v′) between the predicted sample site v and other sites v′. Sites whose distance is less than the spatial threshold ρ participate in forming the spatiotemporal neighbor matrix. The calculation method of Δ(v,v′) is defined as:

[0073] Δ(v,v′)=Dist(v,v′) 1-Corr(v,v′) (2)

[0074]

[0075] Where Dist(v,v′) is the spatial Euclidean distance between the predicted sample site v and site v′, Corr(v,v′) is the correlation coefficient between the predicted sample site v and site v′, To predict the covariance of the states of sample site v and site v′ at time t, are the variances of the predicted sample sites v and v′ at time t respectively.

[0076] Therefore, the original spatiotemporal neighbor matrix of the predicted sample site v is It can be expressed as N v,ρ The relevant nodes and τ v The spatial weight of the predicted sample site v is defined as The time weight is For the original space-time neighbor matrix M v Perform time-space weighting, and the weighted time-space neighbor matrix Γ v The calculation is as follows:

[0077]

[0078] Similarly, the weighted historical spatiotemporal neighbor matrix of all historical samples at site v is calculated in the same way as above: Then calculate the historical space-time neighbor matrix and the current weighted spatiotemporal neighbor matrix Γ v The distance between v , used for the selection of candidate neighbors. v The calculation formula is as follows:

[0079]

[0080] Where trac represents the trace of the matrix.

[0081] According to the result calculated by formula (7), select the space-time neighbor matrix Γ v The most similar (i.e., distance d v The K historical spatiotemporal neighbor matrices with the shortest value are used as spatial neighbors, and the characteristics of the base learner ST-KNN model at the site u at time t are predicted by Gaussian weights. The calculation formula is as follows:

[0082]

[0083] In the formula, is the predicted value of the kth neighbor of site v, ω v,k is the Gaussian weight of the kth neighbor, a v is the Gaussian kernel hyperparameter, d v,k is the distance from site v to its kth neighbor.

[0084] According to the above steps, the predicted state of each sample at each spatial monitoring station is solved in turn to obtain the output feature of the base learner ST-KNN

[0085] In real geographical processes, geographical relationships usually show complex nonlinear characteristics. Compared with the ST-ARIMA model that models the local linear relationship between historical states and future states, the convolutional neural network sets hidden neurons and nonlinear activation functions, which has a good mapping of the nonlinear relationship in the data. Therefore, in order to better express the characteristic nonlinear relationship of spatiotemporal data, in some possible implementations, the third spatiotemporal prediction model is the spatiotemporal graph convolutional neural network model (ST-GCN); the spatiotemporal graph convolutional neural network model uses a graph model to model the non-Euclidean structure of the spatial monitoring site, and uses graph convolution operations and gated one-dimensional temporal convolution Gated TCN to model the complex association of spatial and temporal dimension features to obtain the third output feature. By selecting the ST-GCN model as the third spatiotemporal prediction model for ensemble learning, the nonlinear characteristics of geographical relationships are further considered.

[0086] The graph model is used to represent and process relational data consisting of nodes and edges. The nodes in the graph are defined as spatial monitoring sites, and the edges are the relationships or connections between sites. The edges can be defined based on geographic location, functional similarity, etc. The node features are the attribute information of the spatial monitoring sites, such as monitoring data, geographic coordinates, etc.

[0087] In this embodiment, ST-GCN uses a graph model to establish complex relationships and dependencies between spatial monitoring sites. Specifically, ST-GCN consists of two superimposed spatiotemporal convolutional modules STblock, each of which is designed as a sandwich structure of a temporal convolutional layer TCOnn and a spatial graph convolutional layer GConv. The output feature of the base learner ST-GCN of a sample at time t is can be calculated as:

[0088]

[0089] STblock(X (t-H):t )=TConv{GConv[TConv(X (t-H):t )]} (11)

[0090] The original graph convolution calculation introduces the Laplacian matrix in spectral graph theory to express the properties of the graph. ST-GCN uses ChebNet to replace the Fourier transform operation in the graph convolution layer GConv, and uses the truncated expansion of the Chebyshev polynomial to improve the calculation method, effectively avoiding the high calculation cost of the Laplacian eigenvector. The calculation formula is as follows:

[0091]

[0092] Where φ is the input feature of the current graph convolutional layer, is the normalized graph Laplacian matrix, λ max represent The largest eigenvalue of Cheb x (·) represents the xth order Chebyshev polynomial, χ represents the total number of polynomials, ω x is the Chebyshev parameter, N is the number of space monitoring stations, I N is the N-order identity matrix, and L is the graph Laplacian matrix before normalization.

[0093] The temporal convolution layer TConv of ST-GCN uses the gated 1D Temporal Convolution (Gated TCN), which consists of one-dimensional causal convolution and gated linear units (Gated Linear Units, GLU). It uses the Sigmoid activation function to model the nonlinear relationship of data according to the causal time sequence, and the output result is equal to the dot product of the nonlinear feature and the residual linear feature. The calculation formula is as follows:

[0094] TConv(φ)=Sigmoid(CasualConv(φ))⊙(CasualConv(φ)+φ) (14)

[0095] Where φ is the input feature of the current temporal convolutional layer, and CasualConv is the causal convolutional layer.

[0096] Following the above steps, we can solve the spatiotemporal output features of the ST-GCN model.

[0097] This embodiment uses gated one-dimensional time series convolution technology to more effectively capture the dependencies in time series data and improve the prediction effect.

[0098] Step S103: Integrate the first output feature, the second output feature, and the third output feature through an integration strategy to obtain a final integrated prediction result.

[0099] Among them, the integration strategy is set up with parallel spatiotemporal embedding modules corresponding to each base learner. Each spatiotemporal embedding module decomposes the spatiotemporal heterogeneity integration weights of each base learner into the product of a spatial factor matrix and a temporal factor matrix by means of non-negative matrix decomposition. Each spatiotemporal embedding module sets a spatial embedding branch and a temporal embedding branch; the spatial embedding branch is used to model the mapping relationship between the spatial correlation information of the monitoring site and the spatial factor matrix, and the temporal embedding branch is used to model the mapping relationship between the temporal attribute information of the monitoring site and the temporal factor matrix, thereby solving the spatiotemporal heterogeneity integration weights.

[0100] To facilitate understanding, the integrated strategy design in existing research is first explained.

[0101] When designing ensemble strategies, existing studies usually assume that samples are independent and identically distributed, so that different base learners have globally fixed weights. Taking ordinary linear regression (OLR) as an ensemble strategy as an example, the model calculates the output features of samples at time t for three base learners: ST-ARIMA, ST-KNN, and ST-GCN. and Perform global regression to obtain integrated results The integration method is as shown in formula 15:

[0102]

[0103] In the formula, β 0 is the constant term coefficient, β l is the global average regression coefficient of the output feature of the lth base learner, l = 1, 2, 3, which can be calculated by ordinary least squares (OLS).

[0104] Due to the spatial heterogeneity of geographical elements, the relationship between the output of each base learner and the final integration result is different in different spatial locations. (t+1):(t+T) ∈R 3×N×T In this embodiment, multivariate linear regression is used as the integration strategy. Correspondingly, the spatiotemporal heterogeneity integration weight matrix is ​​defined as Each spatiotemporal position of each base learner corresponds to a different ensemble weight. The final ensemble prediction result It can be expressed as the sum of the output features of the base learner and the dot product of the spatiotemporal heterogeneity integration weights, as shown in the following formula:

[0105]

[0106] In the formula, l represents the serial number of the base learner, l = 1, 2, 3, F l (t+1):(t+T) is the output feature of the lth base learner. When l is 1, 2, or 3, F l (t+1):(t+T) Corresponding to and is the spatiotemporal heterogeneity ensemble weight matrix of the lth base learner.

[0107] Solving the spatiotemporal heterogeneity integrated weight matrix W (t+1):(t+H)The key is to construct a complex relationship between sample spatiotemporal information and integration weights, so that the base learners can be adaptively integrated according to the spatiotemporal pattern. However, the spatial heterogeneity characteristics of spatiotemporal data are intertwined with the temporal non-stationary characteristics, making the spatiotemporal heterogeneity integration weight matrix structure complex and difficult to accurately solve, and bringing great challenges to the interpretability analysis of the integration process. To this end, in this embodiment, three parallel spatiotemporal embedding modules (Spatio-temporal weighted block, STW-block) corresponding to the base learners are designed, and the spatiotemporal heterogeneity integration weight matrix W is solved by the spatiotemporal embedding module. (t+1):(t+H) .

[0108] Figure 3 FIG. 1 is a logic block diagram of a spatiotemporal embedding module provided according to some embodiments of the present application. Figure 3 (a) The spatiotemporal embedding module in the figure has three spatiotemporal embedding modules STW-block. Each spatiotemporal embedding module is provided with a spatial embedding branch and a temporal embedding branch. After the spatial information and temporal information are input into each STW-block, the lth STW-block (l is the serial number of the base learner, l = 1, 2, 3) uses non-negative matrix decomposition to integrate the spatiotemporal heterogeneity weight matrix W of the lth base learner at time t l (t+1):(t+T) ∈R N×T Decompose into Spatial Factor Matrix (SFM l ∈R N ×r , referred to as SFM) and the temporal factor matrix (Temporal Factor Matrix, The product of TFM, where W l (t+1):(t+T) In the figure, it is abbreviated as W l , r is the matrix decomposition dimension. Then, the spatial embedding branch and the temporal embedding branch of the l-th STW-block respectively model the relationship between spatial information, temporal information and the spatial factor matrix and the temporal factor matrix, thereby realizing the coordinated expression of spatial heterogeneity and temporal non-stationarity. The formal expression is as follows:

[0109]

[0110] SFM l =BranchS l (A) (18)

[0111]

[0112] In the formula, l represents the lth base learner, BranchSl is the spatial embedding branch of the l-th base learner of STW-block, and A is the graph adjacency matrix containing sample space information. l is the temporal embedding branch of the l-th base learner of STW-block, Ω t Represents the time attribute information of the sample at time t.

[0113] Furthermore, in some possible implementations, the spatial embedding branch includes a first encoding unit, a first concatenation unit, and a first learning unit, the first learning unit includes two learnable network layers, each network layer includes a linear fully connected layer with Dropout and a ReLU activation function; accordingly, the mapping relationship between the spatial association information of the monitoring sites and the spatial factor matrix is ​​modeled, including the following steps: the first encoding unit constructs a graph proximity matrix A based on the distance and attribute similarity between the monitoring sites, and the graph proximity matrix A includes a graph proximity matrix based on the spatial Euclidean distance and the graph proximity matrix based on the similarity of the monitored values The first coding unit calculates the Laplace normalized matrix according to the graph proximity matrix A, and performs eigendecomposition on the Laplace normalized matrix to obtain an eigenvalue matrix and an eigenvector; the first coding unit constructs a spatial embedding based on the spatial Euclidean distance based on the eigenvalue matrix and the eigenvector. and spatial embedding based on similarity of monitoring values The first concatenation unit embeds the space based on the spatial Euclidean distance and spatial embedding based on similarity of monitoring values Perform channel superposition to obtain the spatial embedding E′ s ; and embed the space into E′ s Input into the first learning unit to establish a mapping relationship between sample space association information and spatial factor matrix.

[0114] The design goal of the spatial embedding branch is to accurately construct the mapping relationship between the spatial correlation information of monitoring sites and the spatial factor matrix.

[0115] Reference Figure 3 (b) The spatial embedding branch in Fig. 3 contains various relationships between monitoring sites, such as distance and attribute similarity, which can be represented by different graph proximity matrices.

[0116] Inspired by the graph embedding algorithm, this application adopts the Eigenmaps graph embedding method to integrate the spatial information of the graph structure into the spatial embedding branch. Specifically, the first encoding unit constructs a spatial embedding based on the spatial Euclidean distance. and spatial embedding based on similarity of monitoring values This is accomplished by the following steps:

[0117] First, the first coding unit constructs a graph proximity matrix based on spatial Euclidean distance (also known as the spatial distance adjacency matrix) and the graph adjacency matrix based on the similarity of the monitored values (also known as the similarity adjacency matrix), and It is collectively called the graph proximity matrix and is represented by A. Define the Laplace normalization matrix The calculation is as follows:

[0118]

[0119] In the formula, Dist(v i ,v j ) indicates monitoring site v i and v j The Euclidean distance between the monitoring sites v i and v j The correlation coefficient of the monitoring value Sim(v i ,v j ) represents the similarity, σ 1 and σ 2 are the standard deviations of distance and similarity, ξ 1 and 2 is the distance threshold and similarity threshold, the diagonal matrix D∈R N×N is the degree matrix of the graph adjacency matrix A, A ij is the element in the i-th row and j-th column of the graph adjacency matrix A, I is the identity matrix, and N is the number of spatial monitoring sites.

[0120] Then, the first encoding unit performs eigendecomposition on the Laplace normalized matrix to obtain an eigenvalue matrix and an eigenvector. The eigendecomposition of the Laplace matrix is ​​implemented as:

[0121]

[0122] In the formula, the diagonal matrix Λ=diag(λ 0 ,λ 1 ,…,λ N-1 )∈R N×N is the eigenvalue matrix, satisfying 0 = λ 0 <λ 1 ≤…≤λ N-1 ,U=(u 0 ,u 1 ,…u N-1 )∈R 1×N is the feature vector.

[0123] Finally, the first coding unit converts the eigenvalue matrix and eigenvector calculated in the above steps into the e of each monitoring station.s Embedding in e-dimensional space s is constructed as Correspondingly, for the graph adjacency matrix the corresponding eigenvalue matrix and eigenvector, the constructed space embedding is a space embedding based on spatial Euclidean distance (dimension: N×e s ); for the graph adjacency matrix the corresponding eigenvalue matrix and eigenvector, the constructed space embedding is a space embedding based on the similarity of monitoring values (dimension: N×e s ).

[0124] Next, the first splicing unit stacks the space embedding based on spatial Euclidean distance and the space embedding based on the similarity of monitoring values in channels (dimension splicing) to obtain the space embedding and inputs the space embedding E′ s into the first learning unit to establish the mapping relationship between the sample space association information and the spatial factor matrix, and obtain the spatial factor matrix SFM.

[0125] The first learning unit of each space embedding branch includes two learnable network layers. Each network layer contains a linear fully connected layer (linear layer) with Dropout and a ReLU activation function. By the above settings, the neuron features are controlled to be positive values to ensure that the spatial factor matrix SFM has non-negative additivity.

[0126] In some possible implementation manners, the time embedding branch includes a second encoding unit, a third encoding unit, a second splicing unit, and a second learning unit. The second learning unit includes two learnable network layers. Each network layer contains a linear fully connected layer with Dropout and a ReLU activation function; correspondingly, modeling the mapping relationship between the time attribute information of the monitoring site and the time factor matrix includes the following steps: The second encoding unit uses trigonometric position encoding to perform relative encoding on the time position of the sample at time t in the global range to obtain the time position embedding The third encoding unit selects m time attributes of the sample at time t related to the integration weights, and encodes each time attribute into the corresponding time attribute embedding using the one-hot method The second splicing unit splices the time position embedding with each time attribute embedding to obtain the time embedding E′ t ; and inputs the time embedding E′ t into the second learning unit to establish the mapping relationship between the sample time information and the time factor matrix.

[0127] The design goal of the time embedding branch is to accurately construct the mapping relationship between the sample time features containing time position and time attribute information and the time factor matrix. In the integration strategy, the temporal non-stationary change characteristics of the data should be reflected. Not only the temporal position of the sample in the global scope should be considered, but also the temporal attribute information of the sample at the moment, so as to more effectively mine the change trend and periodicity in the time series data.

[0128] Specifically, see Figure 3 In the time embedding branch (c), first, the second encoding unit (not shown in the figure) encodes the time position information. Specifically, the trigonometric function position encoding (Sin-cos) is used to relatively encode the time position of the sample at time t in the global range to obtain the time position embedding The expression is as follows:

[0129]

[0130] In the formula, e t for The embedding dimension of , i is the dimension index of the position encoding vector, 1≤i≤e t .

[0131] The trigonometric position coding technology is adopted, and the continuity and monotonicity of the sine and cosine functions are used to ensure that there is a certain regularity between the coding vectors. For example, the difference between adjacent positions is small, while the difference between positions that are far away is large. The periodicity and phase difference of the sine and cosine functions are used to ensure that the coding vector can remain unique in any dimension, that is, the values ​​at different times in the same dimension are different.

[0132] Next, the third encoding unit encodes the time attribute information. Specifically, m time attributes related to the integrated weight of the sample at time t are selected. These time attributes can be, for example, months, days of the week, dates, and times, and each time attribute is encoded into a corresponding time attribute embedding using a one-hot encoding method (one-hot method). 1≤j≤m,e t is the embedding dimension.

[0133] Then, the second splicing unit embeds the time position into Embedded with various time attributes Concatenate to get temporal embedding And embed time into E′ t Input into the second learning unit to establish a mapping relationship between sample time information and the time factor matrix. Then, the feature dimension output by the second learning unit is transformed (reshaped) to obtain the time factor matrix TFM.

[0134] The second learning unit includes two learnable network layers, each of which contains a linear fully connected layer with Dropout and a ReLU activation function. Through the above settings, the neuron features are controlled to be positive values ​​to ensure that the time factor matrix TFM has non-negative additivity.

[0135] The above is an implementation scheme of a matrix-informed ensemble learning method (MI-EL) provided in this embodiment. In some examples, the method is used for spatiotemporal ensemble prediction tasks. Figure 2 The overall framework diagram of the method is shown in FIG. Figure 2 As shown in the figure, this method first takes into account the local spatiotemporal dependency, global feature similarity and feature nonlinearity of spatiotemporal data, and selects the spatiotemporal differential moving average autoregressive model (ST-ARIMA), the spatiotemporal K nearest neighbor model (ST-KNN) and the spatiotemporal graph convolutional neural network model (ST-GCN) as the base learners of the integration framework to obtain the output feature F (3×N×T). Then, three parallel spatiotemporal embedding modules (Spatio-temporal weighted block) are designed in the integration strategy to solve the spatiotemporal heterogeneity integration weights (i.e., heterogeneity integration weights W (3×N×T)) of each base learner respectively. In each spatiotemporal embedding module, the spatial features containing spatial correlation information (spatial information s (N×N)) are input into the spatial embedding branch to solve the spatial factor matrix, and the temporal features containing temporal attribute information are input into the temporal embedding branch to solve the temporal factor matrix. Then, the spatial factor matrix and the temporal factor matrix are multiplied to obtain the spatiotemporal heterogeneity integration weights, which are used to integrate the outputs of different base learners to obtain the final integrated prediction results. From the perspective of matrix decomposition, this method decomposes the complex spatiotemporal heterogeneity integrated weight matrix into a spatiotemporal factor matrix. By constructing a spatiotemporal embedding learning module, the mapping relationship between sample spatiotemporal information and the spatiotemporal factor matrix is ​​established, providing a new idea for the expression of spatiotemporal heterogeneity.

[0136] In addition to the problem of the co-expression of spatial heterogeneity and temporal non-stationarity in the ensemble strategy, the application of ensemble learning in spatiotemporal prediction also faces the challenge of interpretability. Specifically, after obtaining the ensemble prediction results, the existing ensemble learning methods still face the problem of difficulty in explaining the mechanism of spatial heterogeneity and temporal non-stationarity on the model integration process. In other words, the interpretability research of the existing ensemble prediction methods focuses on exploring the mechanism of the original input variables on the integration results, while ignoring the interpretability of the integration process, that is, the intensity and response rules of different base learners on the final results. Since different base learners have different advantages and disadvantages, they usually have different effects on the final prediction results. This difference in action mechanism comes from the fact that the base learners capture the unique spatiotemporal dependencies of geographical phenomena and change with the changes in geographical processes or patterns. However, since the existing research ignores the co-expression of spatial heterogeneity and temporal non-stationarity when designing the ensemble strategy, it is impossible to explain how different geographical spatial change processes affect the integration of base learners, and it is impossible to clarify the rules of this influence, which limits the in-depth understanding of the integration process.

[0137] In view of this, in some possible implementations, the matrix-guided interpretable spatiotemporal integrated prediction method also includes: calculating the spatial score vector and the temporal score vector of each base learner at different spatiotemporal positions according to the spatial factor matrix and the temporal factor matrix; based on the spatial score vector and the temporal score vector, analyzing the response rules of different base learners under the conditions of spatial heterogeneity and temporal non-stationarity.

[0138] Since MI-EL uses non-negative matrix decomposition to decompose the complex spatiotemporal heterogeneity integration weight matrix into a spatial factor matrix and a temporal factor matrix, while realizing the collaborative modeling of spatial heterogeneity and temporal non-stationarity of the integration process, the spatiotemporal heterogeneity integration weights are mapped to the two dimensions of space and time, and the non-negative additivity of the two factor matrices is ensured, which makes it possible to explore the interpretability of the integration process. Based on the spatial factor matrix and the temporal factor matrix, this embodiment calculates the spatial score vector and the temporal score vector of each base learner at different spatiotemporal positions according to the spatial factor matrix and the temporal factor matrix.

[0139] Reference Figure 2 After decomposing the spatiotemporal heterogeneity integrated weight matrix into a spatial factor matrix and a temporal factor matrix, the spatial factor matrix and the temporal factor matrix are summed up according to the embedding dimension r, and then the spatial score vector S_Score of the sample at time t of the l-th base learner is calculated respectively l ∈R N×1 and the time score vector The calculation expression is as follows:

[0140]

[0141] Where l is the lth base learner, l = 1, 2, 3, and k is the embedding dimension number.

[0142] Spatial score vector S_Score l Each element in represents the score of the lth base learner at each monitoring site of the sample, N is the number of spatial monitoring sites, and the time score vector Each element in represents the score of each future prediction time step of the l-th base learner at time t.

[0143] In addition, the time score vector The vector modulus of can represent the total score of the sample of the lth base learner at time t. l and It can be regarded as the feature map of the spatiotemporal heterogeneity integrated weight matrix in the spatial and temporal dimensions, S_Score l It is used to explain the mechanism of spatial distribution heterogeneity on the model integration process; Used to explain the mechanism of time non-stationarity on the model integration process. l and The higher the score, the stronger the effect of the base learner on the ensemble result.

[0144] In some embodiments, based on the calculation of the spatial score vector and the time score vector, the matrix-guided interpretable spatiotemporal integrated prediction method further includes: visualizing the spatial score vector and the time score vector to intuitively display the spatiotemporal variation of the score. By visualizing the spatiotemporal variation of the score, it is possible to further explore the spatial distribution and temporal evolution of the contribution of each base learner to the integrated result, thereby revealing the response rules of different base learners under the conditions of spatial heterogeneity and temporal non-stationarity.

[0145] In this embodiment, the spatial score vector and the temporal score vector of each base learner at different spatiotemporal positions are calculated according to the spatial factor matrix and the temporal factor matrix, which are used to measure the effect of the base learner on the integration result, while ensuring the spatiotemporal prediction accuracy, further improving the interpretability of the spatiotemporal integrated prediction model, which helps to understand the decision-making process of the model. Especially in the environment of sudden data, this interpretability enables humans to understand the reasons for the changes in model performance and make adaptive adjustments.

[0146] In summary, the matrix-guided interpretable spatiotemporal integrated prediction method provided by the present application is used for interpretable spatiotemporal prediction by decomposing the spatiotemporal heterogeneity of data through matrix decomposition. In response to the problem of ignoring the coordinated expression of spatial heterogeneity and temporal non-stationarity in the design of existing integration strategies, the present application decomposes the spatiotemporal heterogeneity integration weight matrix into the product of the spatial factor matrix and the temporal factor matrix, and uses the solution process of the two factor matrices as a guide to achieve the coordinated and precise expression of the spatial heterogeneity and temporal non-stationarity of the integration process. In response to the problem of the interpretability of the action mechanism of the model integration process, the present application constructs the spatial score vector and the temporal score vector with the help of the spatial factor matrix and the temporal factor matrix of each sample, respectively, to explicitly express the action strength and response rules of different basis learners under the conditions of spatial heterogeneity and temporal non-stationarity. In summary, the contributions of this application to the prior art are as follows:

[0147] (1) MI-EL: This embodiment of the present application combines matrix decomposition and ensemble learning to propose a novel interpretable spatiotemporal ensemble prediction method, which improves the interpretability of the model while ensuring the spatiotemporal prediction accuracy.

[0148] (2) Expression of spatiotemporal heterogeneity: From the perspective of matrix decomposition, the complex spatiotemporal heterogeneity integrated weight matrix is ​​decomposed into a spatiotemporal factor matrix. By constructing a spatiotemporal embedding learning module, the mapping relationship between sample spatiotemporal information and the spatiotemporal factor matrix is ​​established, which provides a new idea for realizing the expression of spatiotemporal heterogeneity.

[0149] (3) Interpretability of the integration process: The spatiotemporal score vector is introduced to map the spatiotemporal heterogeneity integration weight matrix into the spatial and temporal dimensions while ensuring non-negative additivity, which provides new possibilities for exploring the interpretability of the integration process.

[0150] (4) Empirical evidence: The proposed method is fully validated using real geographic spatiotemporal datasets, which yield prediction results that are more accurate than those of the current mainstream ensemble strategies. The method also clearly explains the spatial distribution pattern and temporal evolution of the influence of base learners on the prediction results during the ensemble process.

[0151] The present application also provides a matrix-guided interpretable spatiotemporal integrated prediction system, such as Figure 4 As shown, the system includes: a data acquisition module 401, a base learner module 402 and an integration module 403.

[0152] in:

[0153] The data acquisition module 401 is configured to acquire a spatiotemporal state matrix of samples consisting of monitoring values ​​of all spatial monitoring sites;

[0154] A base learner module 402 is configured to use the first spatiotemporal prediction model, the second spatiotemporal prediction model, and the third spatiotemporal prediction model as a base learner of an ensemble learning method, and input the spatiotemporal state matrix into the base learner to obtain a first output feature, a second output feature, and a third output feature respectively;

[0155] An integration module 403 is configured to integrate the first output feature, the second output feature, and the third output feature through an integration strategy to obtain a final integrated prediction result;

[0156] Among them, the integration strategy is set up with parallel spatiotemporal embedding modules corresponding to each base learner. Each spatiotemporal embedding module decomposes the spatiotemporal heterogeneity integration weights of each base learner into the product of a spatial factor matrix and a temporal factor matrix by means of non-negative matrix decomposition. Each spatiotemporal embedding module sets a spatial embedding branch and a temporal embedding branch; the spatial embedding branch is used to model the mapping relationship between the spatial correlation information of the monitoring site and the spatial factor matrix, and the temporal embedding branch is used to model the mapping relationship between the temporal attribute information of the monitoring site and the temporal factor matrix, thereby solving the spatiotemporal heterogeneity integration weights.

[0157] The matrix-guided interpretable spatiotemporal integrated prediction system provided in this embodiment can implement the steps and processes of the matrix-guided interpretable spatiotemporal integrated prediction method provided in any of the above embodiments and achieve the same technical effects, which will not be repeated here one by one.

[0158] Figure 5 is a schematic diagram of the structure of an electronic device provided according to some embodiments of the present application. Figure 5 As shown, the electronic device includes:

[0159] One or more processors 501;

[0160] The computer-readable storage medium (memory) may be configured to store one or more programs 502. When one or more processors 501 execute one or more programs 502, the following steps are implemented:

[0161] Obtain the spatiotemporal state matrix of the samples consisting of the monitoring values ​​of all spatial monitoring stations;

[0162] Using the first spatiotemporal prediction model, the second spatiotemporal prediction model, and the third spatiotemporal prediction model as base learners of the ensemble learning method, inputting the spatiotemporal state matrix into the base learners respectively, and obtaining the first output feature, the second output feature, and the third output feature correspondingly;

[0163] Integrating the first output feature, the second output feature, and the third output feature through an integration strategy to obtain a final integrated prediction result;

[0164] Among them, the integration strategy is provided with parallel spatiotemporal embedding modules corresponding to each base learner, and each of the spatiotemporal embedding modules decomposes the spatiotemporal heterogeneity integration weights of each base learner into the product of a spatial factor matrix and a temporal factor matrix by means of non-negative matrix decomposition. Each of the spatiotemporal embedding modules is provided with a spatial embedding branch and a temporal embedding branch; the spatial embedding branch is used to model the mapping relationship between the spatial correlation information of the monitoring site and the spatial factor matrix, and the temporal embedding branch is used to model the mapping relationship between the temporal attribute information of the monitoring site and the temporal factor matrix, thereby solving the spatiotemporal heterogeneity integration weights.

[0165] Optionally, the electronic device may further include a communication interface and a communication bus. Optionally, the communication interface may be an interface of a communication module, such as an interface of a GSM module.

[0166] The processor 501 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc. It may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The methods, steps and logic block diagrams disclosed in the embodiments of the present application may be implemented or executed. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0167] The electronic device of the embodiment of the present application exists in various forms, including but not limited to:

[0168] (1) Mobile communication devices: These devices are characterized by their mobile communication functions and their main purpose is to provide voice and data communications. These terminals include: smart phones (e.g., iPhone), multimedia phones, functional phones, and low-end phones.

[0169] (2) Ultra-mobile personal computer devices: These devices fall into the category of personal computers, have computing and processing capabilities, and generally also have mobile Internet access features. These terminals include: PDA, MID and UMPC devices, such as iPad.

[0170] (3) Portable entertainment devices: These devices can display and play multimedia content. They include audio and video players (such as iPods), handheld game consoles, e-books, smart toys, and portable car navigation devices.

[0171] (4) Server: A device that provides computing services. The server consists of a processor, hard disk, memory, system bus, etc. The server is similar to a general computer architecture, but because it needs to provide highly reliable services, it has higher requirements in terms of processing power, stability, reliability, security, scalability, and manageability.

[0172] (5) Other electronic devices with data interaction functions.

[0173] It should be pointed out that, according to the needs of implementation, the various components / steps described in the embodiments of the present application can be split into more components / steps, or two or more components / steps or partial operations of components / steps can be combined into new components / steps to achieve the purpose of the embodiments of the present application.

[0174] The above-mentioned method according to the embodiment of the present application can be implemented in hardware, firmware, or implemented as software or computer code that can be stored in a recording medium (such as a CD ROM, RAM, floppy disk, hard disk or magneto-optical disk), or implemented as a computer code originally stored in a remote recording medium or a non-temporary machine storage medium downloaded through a network and to be stored in a local recording medium, so that the method described herein can be stored in such software processing on a recording medium using a general-purpose computer, a special-purpose processor or programmable or special-purpose hardware (such as an ASIC or FPGA). It can be understood that a computer, a processor, a microprocessor controller or programmable hardware includes a storage component (e.g., RAM, ROM, flash memory, etc.) that can store or receive software or computer code, and when the software or computer code is accessed and executed by a computer, a processor or hardware, the matrix-guided interpretable spatiotemporal integrated prediction method described herein is implemented. In addition, when a general-purpose computer accesses the code for implementing the method shown here, the execution of the code converts the general-purpose computer into a special-purpose computer for executing the method shown here.

[0175] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A matrix-guided interpretable spatiotemporal integrated prediction method, characterized in that: include: Obtain the spatiotemporal state matrix of the samples consisting of the monitoring values ​​of all spatial monitoring stations; Using the first spatiotemporal prediction model, the second spatiotemporal prediction model, and the third spatiotemporal prediction model as base learners of the ensemble learning method, inputting the spatiotemporal state matrix into the base learners respectively, and obtaining the first output feature, the second output feature, and the third output feature correspondingly; Integrating the first output feature, the second output feature, and the third output feature through an integration strategy to obtain a final integrated prediction result; The integration strategy is provided with parallel spatiotemporal embedding modules corresponding to each base learner, each of which decomposes the spatiotemporal heterogeneity integration weights of each base learner into the product of a spatial factor matrix and a temporal factor matrix by means of non-negative matrix decomposition, and each of the spatiotemporal embedding modules is provided with a spatial embedding branch and a temporal embedding branch; the spatial embedding branch is used to model the mapping relationship between the spatial correlation information of the monitoring site and the spatial factor matrix, and the temporal embedding branch is used to model the mapping relationship between the temporal attribute information of the monitoring site and the temporal factor matrix, thereby solving the spatiotemporal heterogeneity integration weights; definition is the monitoring value of the spatial monitoring site v at the tth time step, and the spatiotemporal state matrix X of the sample (t-H):t ∈R N×H It is defined as the monitoring values ​​of all spatial monitoring stations in the past H time steps, where N is the number of spatial monitoring stations, t is the time corresponding to the sample, H is the total number of historical time steps of the sample, and R is a set of real numbers; The first spatiotemporal prediction model is a spatiotemporal difference moving average autoregressive model; The spatiotemporal difference moving average autoregressive model introduces a spatial weight matrix on the basis of the integrated moving average autoregressive model, transforms the non-stationary spatiotemporal state into a stationary spatiotemporal state by difference, and regresses the historical values ​​of the prediction variables and the present values ​​and lagged values ​​of the random error terms to obtain the first output feature; The second spatiotemporal prediction model is a spatiotemporal K nearest neighbor model; The spatiotemporal K nearest neighbor model introduces spatial information on the basis of the traditional K nearest neighbor method, maps the spatiotemporal state matrix to a new feature space, selects K spatial neighbors with the most similar features within the entire historical time range for weighted prediction, and obtains the second output feature; The third spatiotemporal prediction model is a spatiotemporal graph convolutional neural network model; The spatiotemporal graph convolutional neural network model uses a graph model to model the non-Euclidean structure of the spatial monitoring site, and adopts a graph convolution operation and a gated one-dimensional temporal convolution Gated TCN to model the complex association of spatial and temporal dimension features to obtain the third output feature; The integration strategy adopts the integration strategy of multiple linear regression; The spatial embedding branch includes a first encoding unit, a first concatenation unit and a first learning unit, wherein the first learning unit includes two learnable network layers, each of which includes a linear fully connected layer with Dropout and a ReLU activation function; Accordingly, modeling the mapping relationship between the spatial correlation information of the monitoring sites and the spatial factor matrix includes the following steps: The first encoding unit constructs a graph proximity matrix A based on the distance and attribute similarity between the monitoring sites, wherein the graph proximity matrix A includes a graph proximity matrix based on the spatial Euclidean distance and the graph proximity matrix based on the similarity of the monitored values The first encoding unit calculates a Laplace normalized matrix according to the graph proximity matrix A, and performs eigendecomposition on the Laplace normalized matrix to obtain an eigenvalue matrix and an eigenvector; The first encoding unit constructs a spatial embedding based on the spatial Euclidean distance based on the eigenvalue matrix and the eigenvector. and spatial embedding based on similarity of monitoring values The first splicing unit embeds the space based on the spatial Euclidean distance and spatial embedding based on similarity of monitoring values Perform channel superposition to obtain the spatial embedding E′ s ; and embed the space into E′ s Input into the first learning unit to establish a mapping relationship between sample space association information and space factor matrix; The time embedding branch includes a second encoding unit, a third encoding unit, a second concatenation unit, and a second learning unit, wherein the second learning unit includes two learnable network layers, each of which includes a linear fully connected layer with Dropout and a ReLU activation function; Accordingly, modeling the mapping relationship between the time attribute information of the monitoring site and the time factor matrix includes the following steps: The second encoding unit uses trigonometric function position encoding to relatively encode the time position of the sample at time t in the global range to obtain the time position embedding The third encoding unit selects m time attributes of the sample at time t that are related to the integrated weight, and encodes each of the time attributes into a corresponding time attribute embedding using a one-bit efficient encoding method The second splicing unit embeds the time position into With each of the time attributes embedded Splice and get the time embedding E t ′ ; and embed the time into E t ′ Input into the second learning unit to establish a mapping relationship between sample time information and time factor matrix.

2. The method according to claim 1, characterized in that Also includes: Calculate the spatial score vector and the temporal score vector of each base learner at different spatiotemporal positions according to the spatial factor matrix and the temporal factor matrix; Based on the spatial score vector and the temporal score vector, response rules of different base learners under the conditions of spatial heterogeneity and temporal non-stationarity are analyzed.

3. A matrix-guided interpretable spatiotemporal integrated prediction system, characterized in that: include: A data acquisition module configured to acquire a spatiotemporal state matrix of samples consisting of monitoring values ​​of all spatial monitoring sites; A base learner module, configured to use the first spatiotemporal prediction model, the second spatiotemporal prediction model, and the third spatiotemporal prediction model as base learners of the ensemble learning method, input the spatiotemporal state matrix into the base learner respectively, and obtain the first output feature, the second output feature, and the third output feature accordingly; An integration module, configured to integrate the first output feature, the second output feature, and the third output feature through an integration strategy to obtain a final integrated prediction result; The integration strategy is provided with parallel spatiotemporal embedding modules corresponding to each base learner, each of which decomposes the spatiotemporal heterogeneity integration weights of each base learner into the product of a spatial factor matrix and a temporal factor matrix by means of non-negative matrix decomposition, and each of the spatiotemporal embedding modules is provided with a spatial embedding branch and a temporal embedding branch; the spatial embedding branch is used to model the mapping relationship between the spatial correlation information of the monitoring site and the spatial factor matrix, and the temporal embedding branch is used to model the mapping relationship between the temporal attribute information of the monitoring site and the temporal factor matrix, thereby solving the spatiotemporal heterogeneity integration weights; definition is the monitoring value of the spatial monitoring site v at the tth time step, and the spatiotemporal state matrix X of the sample (t-H):t ∈R N×H It is defined as the monitoring values ​​of all spatial monitoring stations in the past H time steps, where N is the number of spatial monitoring stations, t is the time corresponding to the sample, H is the total number of historical time steps of the sample, and R is a set of real numbers; The first spatiotemporal prediction model is a spatiotemporal difference moving average autoregressive model; The spatiotemporal difference moving average autoregressive model introduces a spatial weight matrix on the basis of the integrated moving average autoregressive model, transforms the non-stationary spatiotemporal state into a stationary spatiotemporal state by difference, and regresses the historical values ​​of the prediction variables and the present values ​​and lagged values ​​of the random error terms to obtain the first output feature; The second spatiotemporal prediction model is a spatiotemporal K nearest neighbor model; The spatiotemporal K nearest neighbor model introduces spatial information on the basis of the traditional K nearest neighbor method, maps the spatiotemporal state matrix to a new feature space, selects K spatial neighbors with the most similar features within the entire historical time range for weighted prediction, and obtains the second output feature; The third spatiotemporal prediction model is a spatiotemporal graph convolutional neural network model; The spatiotemporal graph convolutional neural network model uses a graph model to model the non-Euclidean structure of the spatial monitoring site, and adopts a graph convolution operation and a gated one-dimensional temporal convolution Gated TCN to model the complex association of spatial and temporal dimension features to obtain the third output feature; The integration strategy adopts the integration strategy of multiple linear regression; The spatial embedding branch includes a first encoding unit, a first concatenation unit and a first learning unit, wherein the first learning unit includes two learnable network layers, each of which includes a linear fully connected layer with Dropout and a ReLU activation function; Accordingly, modeling the mapping relationship between the spatial correlation information of the monitoring sites and the spatial factor matrix includes the following steps: The first encoding unit constructs a graph proximity matrix A based on the distance and attribute similarity between the monitoring sites, wherein the graph proximity matrix A includes a graph proximity matrix based on the spatial Euclidean distance and the graph proximity matrix based on the similarity of the monitored values The first encoding unit calculates a Laplace normalized matrix according to the graph proximity matrix A, and performs eigendecomposition on the Laplace normalized matrix to obtain an eigenvalue matrix and an eigenvector; The first encoding unit constructs a spatial embedding based on the spatial Euclidean distance based on the eigenvalue matrix and the eigenvector. and spatial embedding based on similarity of monitoring values The first splicing unit embeds the space based on the spatial Euclidean distance and spatial embedding based on similarity of monitoring values Perform channel superposition to obtain the spatial embedding E′ s ; and embed the space into Ev s Input into the first learning unit to establish a mapping relationship between sample space association information and space factor matrix; The time embedding branch includes a second encoding unit, a third encoding unit, a second concatenation unit, and a second learning unit, wherein the second learning unit includes two learnable network layers, each of which includes a linear fully connected layer with Dropout and a ReLU activation function; Accordingly, modeling the mapping relationship between the time attribute information of the monitoring site and the time factor matrix includes the following steps: The second encoding unit uses trigonometric function position encoding to relatively encode the time position of the sample at time t in the global range to obtain the time position embedding The third encoding unit selects m time attributes of the sample at time t that are related to the integrated weight, and encodes each of the time attributes into a corresponding time attribute embedding using a one-bit efficient encoding method The second splicing unit embeds the time position into With each of the time attributes embedded Splice and get the time embedding E t ′ ; and embed the time into E t ′ Input into the second learning unit to establish a mapping relationship between sample time information and time factor matrix.

4. An electronic device, characterized in that: include: A memory, a processor, and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the matrix-guided interpretable spatiotemporal integrated prediction method according to any one of claims 1 to 2 is implemented.

Images