Population thermodynamic data prediction method based on spatio-temporal feature attention mechanism

Abstract

The invention belongs to the technical field of artificial intelligence, and particularly relates to a spatial-temporal feature attention mechanism-based population thermodynamic data prediction method, which comprises the following steps of: acquiring a target area map, and performing node division on the target area map to generate a target area node network; obtaining population data information of each node in the target area node network; calculating the population data information of each node by using a neural network to obtain time sequence features and spatial features of each node; generating a time feature vector and a space feature vector of each node according to the time sequence features and the space features; and generating predicted population thermodynamic data information according to the time feature vector and the space feature vector, thereby solving the problem that a traditional model is insufficient in spatial and temporal feature capture capability, solving the problem of spatial and temporal feature calculation redundancy in a deep learning algorithm, reducing model parameters by introducing spatial and temporal feature attention, and improving the prediction accuracy of the population thermodynamic data information. And the calculation efficiency of large-range and big-data population thermodynamic data can be well improved.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and specifically is a method for predicting population thermal data based on a spatiotemporal feature attention mechanism. Background Art

[0002] When faced with sudden natural disasters or man-made disasters (such as earthquakes, tsunamis, fires, viruses, etc.), the data on the affected population is often statistical information estimated according to the affected area, which cannot reflect the spatiotemporal distribution characteristics of the affected population. Due to the huge flow of people after the disaster, such data often lacks real-time effectiveness. The population density of a region is an important factor affecting disaster risk. When conducting risk assessment and post-earthquake reconstruction assessment for disasters such as earthquakes, the distribution of the population helps to consider the region's risk resistance and formulate urban disaster risk reduction plans. When distributing disaster relief supplies, the number and distribution of the affected population are the first factors to be considered during disaster relief. In the past, when counting the affected population, data mostly came from census data. Although these data have made important contributions in previous disaster relief, they underestimate the need for dynamic population prediction in disaster relief work. There are many related algorithms for obtaining dynamic population distribution after a disaster, which can be roughly divided into time series (statistical domain) analysis and prediction models and spatial analysis and prediction models.

[0003] Structural equation modeling with maximum likelihood estimation of normal distribution is the most widely used algorithm in population estimation research. Effective and accurate parameter estimation enables it to truly reflect the characteristics of population distribution. With the rise of time series regression models, time series of different dimensions are integrated together to establish a general regression framework, and the data are classified according to the type of time characteristic data. This is the key component of the general framework that can handle each time series model separately. The induction of population growth patterns and the proposal of prediction methods provide formal support for the autoregressive (AR) model of population time series with covariates. On this basis, the autoregressive moving average model (ARMA) was generated and applied to the solution of urban budget problems related to population. The autoregressive moving average model (ARMA) for multi-model selection is used to regulate the variability of the characteristic yield and stationary distribution of the long-term population and obtains a good fitting effect. Compared with ARMA, the moving average autoregressive model (ARIMA) and the seasonal moving average autoregressive model (SARIMA) have greatly improved the prediction accuracy and efficiency.

[0004] Crowdsourced geographic models with population spatial density functions as the core provide baseline information for the changing trends of urban population distribution driven by environmental changes and technological updates. Entangled models of population migration and road networks on complex topography are used to address the problem of how population networks are embedded in distribution space. The population density of densely populated coastal cities is often overestimated, and the use of zoning models based on urban information footprints can greatly reduce such errors. In traditional spatial domain distribution models, the mobility of time and space is often ignored, which may lead to huge differences between actual needs and distribution models, and the solution loses its optimality. A more sophisticated and maximized spatial-temporal stereoscopic framework can better reflect the real-time data distribution dynamics. Spatial population forecasts must take into account changes in the total population, urbanization, and development spatial patterns, while taking into account the uncertainty of each aspect. Although there are more and more forecasts, there are few forecasts with relatively high resolution.

[0005] Heatmap is a spatiotemporal data feature distribution model that has received widespread attention in recent years and is applied to various fields, such as weather forecasting, traffic congestion prediction, public transportation, and disease prevention. Heatmap is a powerful tool for representing spatiotemporal sequence distribution. It can transform massive data into information-rich and assumption-free visual summary data. The relationship and network structure of the data are mined to the maximum extent without destroying the original data. Heatmap has an overall clustering visualization effect on geographic data, especially population data, showing the concentration of data in a unit area, and is widely used to study the problem of urban spatiotemporal dynamic distribution. Baidu population heatmap is one of the important sources of population distribution feature data. Using Baidu heatmap big data, we can explore the dynamic changes of the flow of people in urban squares and its regularity at different times. The flow of people in different time periods and regions presents different characteristics. The classification of heatmap intensity can better help us analyze the spatiotemporal distribution of population aggregation.

[0006] In the past, the characteristic analysis and prediction of population heat mainly focused on the cluster analysis of heat values ​​and the regression prediction of time and space domains. The traditional method not only has a slow recognition speed, but also has a low accuracy rate. Since the population spatiotemporal distribution model is related to the efficiency of evacuation, rescue, and material transportation of natural disaster victims, in order to protect the safety of people's lives and property, the timeliness of emergency rescue determines that the algorithm needs to produce high-quality population spatiotemporal distribution prediction data in a short time. In recent years, with the rapid rise of artificial intelligence algorithms, many excellent algorithms have participated in spatiotemporal distribution prediction, and have achieved good results in the prediction of time and space domains. However, one of the most important challenges in the study of spatiotemporal series prediction based on deep algorithms is the fusion of spatial correlation and time dependence. Traditional methods either analyze time characteristics in order to achieve time dependence. Summary of the invention

[0007] To solve the problems in the above-mentioned background technology, the present invention provides a population thermal data prediction method based on the spatiotemporal feature attention mechanism, which extracts the population distribution characteristics in the time (statistical) domain and the spatial domain respectively through compression and convolution, and introduces the attention mechanism to convert different feature values ​​into weight values ​​and integrate them with the spatiotemporal data to obtain the prediction results.

[0008] To achieve the above object, the present invention provides the following technical solution: a method for predicting population thermal data based on a spatiotemporal feature attention mechanism, comprising:

[0009] Step S01, obtaining a target area map, and dividing the target area map into nodes to generate a target area node network;

[0010] Step S02, obtaining population data information of each node in the target area node network;

[0011] Step S03, using a neural network to calculate the population data information of each node to obtain the time series characteristics and spatial characteristics of each node;

[0012] Step S04, generating a time feature vector and a space feature vector of each node according to the time series feature and the space feature;

[0013] Step S05: generating predicted population thermal data information according to the time feature vector and the space feature vector.

[0014] Preferably, the step S03 includes:

[0015] Inputting the population data information of each node into a neural network to add a spatiotemporal feature attention weight to each node based on a spatiotemporal feature attention mechanism;

[0016] Performing data processing on the features corresponding to each node according to the spatiotemporal feature attention weights;

[0017] A convolution operation is performed on the corresponding features of each node after data processing to obtain the spatial features of each node, and the obtained spatial features are transformed to obtain the time series features of each node.

[0018] Preferably, the step S04 includes:

[0019] The time series features and spatial features corresponding to each node are input into a neural network to obtain a time feature vector and a spatial feature vector of each node.

[0020] Preferably, it also includes constructing a spatiotemporal attention fusion population thermal prediction framework, including:

[0021] Obtaining temporal and spatial channel context information;

[0022] The temporal channel attention and the spatial channel attention are integrated to form a spatiotemporal attention population thermal prediction framework;

[0023] The data consists of population heat point data containing longitude and latitude information. After time domain feature compression, spatial data filling, spatial domain feature compression and spatiotemporal attention feature fusion, it enters the Encoder-Decoder model. Through spatiotemporal feature fusion attention, the encoder learns the spatiotemporal feature vector representation of the input population thermal data information.

[0024] Preferably, the step S05 includes:

[0025] The time feature vector and the space feature vector are respectively fused with the population thermal feature time and space vector in sequence according to feature weights, and the fused time and space features are the output values ​​of the convolutional neural network.

[0026] Preferably, the spatiotemporal feature fusion attention is reintegrated with the backbone network to capture the spatiotemporal correlation between population thermal tensors at different times, thereby generating new features associated with spatiotemporal contexts.

[0027] Preferably, the starting point features and the end point features of each node are calculated to obtain a final representation reflecting the spatiotemporal contextual information features, and the output of the spatiotemporal attention transfer mechanism model layer is regressed through a fully connected layer and a tanh activation function to obtain the start / end flow value at the next moment, so as to complete the spatiotemporal attention fusion population thermal prediction.

[0028] Preferably, zero padding is performed on nodes that cannot generate data, and the zero padding logic formula is as follows:

[0029]

[0030] Among them, M is the spatial data, and m is the true value of M.

[0031] Preferably, the time domain feature compression includes:

[0032] A time compression framework is established, taking the population density tensor matrix in the window of a given time lt as input, and predicting the population density tensor matrix at the output time t+k. The formula is as follows:

[0033]

[0034] Among them, K is the number of time output windows, Input frame for spacetime, Output framework for space-time;

[0035] x t1 =F sq(x)=Avg[x(i,j)],i∈H,j∈W

[0036] x t2 =F sq (x)=Max[x(i,j)],i∈H,j∈W

[0037] x tn =F sq (x)=Sf[x(i,j)],i∈H,j∈W

[0038] Among them, the above formulas correspond to time domain mean compression, time domain maximum value compression, and other time domain statistical compression from top to bottom, respectively. sq is the compression function.

[0039] Preferably, the spatial domain feature compression includes performing spatial domain mean compression, spatial domain maximum value compression and other spatial domain statistical compression on the T dimension, and the corresponding formula is as follows:

[0040] x s1 =F sq (x)=Avg[x(k)],k∈T

[0041] x s2 =F sq (x)=Max[x(k)],k∈T

[0042] x sn =F sq (x)=Sf[x(k)],k∈T

[0043] Among them, the above formulas correspond to spatial domain mean compression, spatial domain maximum compression, and other statistical compressions in spatial domain from top to bottom. sq is the compression function.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] The present invention solves the problem of insufficient ability of traditional models to capture spatiotemporal features, and solves the problem of redundancy in spatiotemporal feature calculation in deep learning algorithms. By introducing spatiotemporal feature attention and reducing model parameters, the computational efficiency of large-scale, big data population thermal data can be greatly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a schematic diagram of the process of the present invention;

[0047] Figure 2 It is a schematic diagram of time domain data compression of the present invention;

[0048] Figure 3 It is a schematic diagram of spatial domain data compression of the present invention;

[0049] Figure 4 It is a schematic diagram of the fusion of spatiotemporal attention and backbone network of the present invention;

[0050] Figure 5 This is a flow chart of an embodiment of the present invention; DETAILED DESCRIPTION

[0051] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0052] A spatiotemporal prediction model was established based on population thermal data. Compared with other algorithm baselines, the model parameters dropped significantly and the efficiency was significantly improved. The experimental results show that it is necessary to simplify the steps in the spatiotemporal correlation of population heat maps based on spatiotemporal series prediction and improve efficiency. Compared with similar studies, while maintaining a certain accuracy, it greatly saves the prediction time, so that the prediction results can be obtained in a very short time after the disaster occurs, which greatly helps disaster relief work. Based on the population thermal data, the attention weight vectors in the time domain and the space domain are provided as the hyperparameters of the fully connected layer, which greatly improves the interpretability between the time series and the space series, making the analysis of spatiotemporal features more in-depth. This model framework is suitable for various spatiotemporal thermal data or graph data that can be converted into thermal data.

[0053] Different network structures are suitable for processing different types of data. The more similar the network structure is to the data, the higher the processing efficiency is and the more obvious the advantage is. The spatiotemporal feature prediction network constructed in this paper better simulates the correlation of sequence-related data, and the features affect the spatiotemporal distribution data according to the weight value. Therefore, when processing data with spatiotemporal distribution feature-related properties, especially spatial gridded data, this model is more likely to discover the laws of spatiotemporal features than other fully connected networks and obtain better classification or fitting results. The present invention is further described below based on specific embodiments.

[0054] like Figures 1 to 5 As shown, the present invention provides a method for predicting population thermal data based on a spatiotemporal feature attention mechanism, comprising:

[0055] Step S01, obtaining a target area map, and dividing the target area map into nodes to generate a target area node network;

[0056] In this embodiment, taking a certain city A as an example, the server is used to obtain the map data of city A, the city map is divided into nodes, and the city areas in the map of city A are used as nodes to generate a city node network. The map can add map descriptions, such as administrative areas, tourist areas, traffic areas, etc., and the areas in the map of city A are used as nodes to build a node network. Therefore, the area can be of any shape and is not limited. It is understandable that there are areas with large traffic flow and areas with small traffic flow or areas with basically no people in city A, such as mountainous areas, water areas, etc. (This part will be explained in the zero-filling operation below)

[0057] Step S02, obtaining population data information of each node in the target area node network;

[0058] This part can be obtained using real-time population data information, or it can be obtained from various historical time periods.

[0059] Step S03, using a neural network to calculate the population data information of each node to obtain the time series characteristics and spatial characteristics of each node;

[0060] Specifically, the population data information of each node is input into a neural network to add a spatiotemporal feature attention weight to each node based on a spatiotemporal feature attention mechanism;

[0061] Performing data processing on the features corresponding to each node according to the spatiotemporal feature attention weights;

[0062] A convolution operation is performed on the corresponding features of each node after data processing to obtain the spatial features of each node, and the obtained spatial features are transformed to obtain the time series features of each node.

[0063] Step S04, generating a time feature vector and a space feature vector of each node according to the time series feature and the space feature; inputting the time series feature and the space feature corresponding to each node into a neural network to obtain a time feature vector and a space feature vector of each node.

[0064] Step S05, generating predicted population thermal data information according to the time feature vector and the space feature vector. The time feature vector and the space feature vector are respectively fused with the population thermal feature time and space vector in sequence, and the fused time and space features are the output values ​​of the convolutional neural network.

[0065] Time domain feature compression includes:

[0066] Establish a time compression framework. Generally, spatiotemporal data prediction (including population thermal data) is carried out in the form of seq2seq spatiotemporal sequence. Since spatiotemporal data is three-dimensional data, it is necessary to compress the three-dimensional data along the time axis to obtain the time series. The compression framework takes the population density tensor matrix in the window of a given time lt as input, and calculates the predicted output population density tensor matrix at time t+k. The formula is as follows:

[0067]

[0068] Among them, K is the number of time output windows, Input frame for spacetime, Output framework for space-time;

[0069] x t1 =F sq (x)=Avg[x(i,j)],i∈H,j∈W

[0070] x t2 =F sq (x)=Max[x(i,j)],i∈H,j∈W

[0071] x tn =F sq (x)=Sf[x(i,j)],i∈H,j∈W

[0072] Among them, the above formulas correspond to time domain mean compression, time domain maximum value compression, and other time domain statistical compression from top to bottom, respectively. sq is the compression function. After compression, the three-dimensional space-time data is converted into one-dimensional data of 1×1×k.

[0073] Since the research area is planned as a rectangular area, many data collection points involve water areas and other factors that cannot generate data, so these data collection points are filled with zeros to make the data of each time period a 30x19 matrix. Therefore, the nodes that cannot generate data are filled with zeros, and the zero-filling logic formula is as follows:

[0074]

[0075] Among them, M is the spatial data, and m is the true value of M.

[0076] There are many spatiotemporal sequence prediction algorithms. For different spatiotemporal sequence prediction problems, there are different ways to deal with the relationship between spatiotemporal sequences, such as Markov chain-based models, Akaike and Bayesian information criteria, extracting spatiotemporal features by constructing extended three-dimensional convolutions, and density distribution methods based on clustering and spatiotemporal graph convolutional neural networks. These algorithms are based on different spatiotemporal sequence problems. Often, due to the need for a large amount of spatiotemporal fusion calculations and the investment of too much computing power in the static or dynamic relationship between spatiotemporal sequences, they are very redundant and inefficient when calculating population heat prediction problems. Therefore, a feature extraction and prediction algorithm for the special spatiotemporal data structure of population heat maps is constructed, which focuses on the two channels of time and space respectively, and achieves a great performance improvement while maintaining a small overhead.

[0077] The spatial domain feature compression includes spatial domain mean compression, spatial domain maximum compression and other spatial domain statistical compression on the T dimension. The corresponding formula is as follows:

[0078] x s1 =F sq (x)=Avg[x(k)],k∈T

[0079] x s2 =F sq (x)=Max[x(k)],k∈T

[0080] x sn =F sq (x)=Sf[x(k)],k∈T

[0081] Among them, the above formulas correspond to spatial domain mean compression, spatial domain maximum compression, and other statistical compressions in spatial domain from top to bottom. sq is the compression function.

[0082] It should be noted that the present invention also includes the extraction of attention to time series features, and improves the channel domain of CBAM (Convolutional Block Attention Module). Two channels of median and standard deviation are added on the basis of maxpool and avgpool, and the four channels of spatiotemporal data (total, average, median, standard deviation) are respectively sent to a multi-layer perceptron (MLP). The number of neurons in the first layer is C / r (r is the reduction rate), the activation function is Relu, and the number of neurons in the second layer is C. The two-layer neural network is shared. Then, the features output by MLP are summed up, and then the sigmoid activation operation is performed to generate the final time domain (statistical domain) weight feature vector. The time window is set to 1 week (7 days) in accordance with the time series period.

[0083] It should be noted that the present invention also includes the extraction of spatial multi-scale attention features. Specifically, the spatial data is respectively subjected to spatial feature extraction using convolution kernels of 7×7, 5×5, 3×3, and 1×1, and the four spatial feature channels are compressed into a 1×1×C vector. The vector is converted into the weight between multi-scale channels through the Softmax activation function, the spatial feature weights of each convolution kernel are learned, and then the weight vector is multiplied by the four spatial domain feature channels and then added to obtain the spatial domain feature output.

[0084] It should be noted that the present invention also includes constructing a spatiotemporal attention fusion population thermal prediction framework, including: obtaining time and space channel context information; fusing the time channel attention and the space channel attention to form a spatiotemporal attention population thermal prediction framework; the data is composed of population thermal point data containing longitude and latitude information, and after time domain feature compression, spatial data filling, spatial domain feature compression and spatiotemporal attention feature fusion, it enters the Encoder-Decoder model through spatiotemporal feature fusion attention, and the encoder learns the spatiotemporal feature vector representation of the input population thermal data information.

[0085] The spatiotemporal feature fusion attention is reintegrated with the backbone network to capture the spatiotemporal correlation between population heat tensors at different times, thereby generating new features associated with spatiotemporal context.

[0086] The starting point features and end point features of each node are calculated to obtain the final representation reflecting the spatiotemporal contextual information features. The output of the spatiotemporal attention transfer mechanism model layer is regressed through the fully connected layer and the tanh activation function to obtain the start / end flow value at the next moment, so as to complete the spatiotemporal attention fusion population thermal prediction.

[0087] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting population thermal data based on spatiotemporal feature attention mechanism, characterized in that: include: Step S01, obtaining a target area map, and dividing the target area map into nodes to generate a target area node network; Step S02, obtaining population data information of each node in the target area node network; Step S03, using a neural network to calculate the population data information of each node to obtain the time series characteristics and spatial characteristics of each node; Step S04, generating a time feature vector and a space feature vector of each node according to the time series feature and the space feature; Step S05: generating predicted population thermal data information according to the time feature vector and the space feature vector.

2. According to claim 1, a method for predicting population thermal data based on spatiotemporal feature attention mechanism is characterized in that: The step S03 comprises: Inputting the population data information of each node into a neural network to add a spatiotemporal feature attention weight to each node based on a spatiotemporal feature attention mechanism; Performing data processing on the features corresponding to each node according to the spatiotemporal feature attention weights; A convolution operation is performed on the corresponding features of each node after data processing to obtain the spatial features of each node, and the obtained spatial features are transformed to obtain the time series features of each node.

3. According to claim 1, a method for predicting population thermal data based on spatiotemporal feature attention mechanism is characterized in that: The step S04 comprises: The time series features and spatial features corresponding to each node are input into a neural network to obtain a time feature vector and a spatial feature vector of each node.

4. According to claim 1, a method for predicting population thermal data based on spatiotemporal feature attention mechanism is characterized in that: It also includes the construction of a spatiotemporal attention fusion population thermal prediction framework, including: Obtaining temporal and spatial channel context information; The temporal channel attention and the spatial channel attention are integrated to form a spatiotemporal attention population thermal prediction framework; The data consists of population heat point data containing longitude and latitude information. After time domain feature compression, spatial data filling, spatial domain feature compression and spatiotemporal attention feature fusion, it enters the Encoder-Decoder model. Through spatiotemporal feature fusion attention, the encoder learns the spatiotemporal feature vector representation of the input population thermal data information.

5. According to claim 4, a method for predicting population thermal data based on spatiotemporal feature attention mechanism is characterized in that: The step S05 comprises: The time feature vector and the space feature vector are respectively fused with the population thermal feature time and space vector in sequence according to feature weights, and the fused time and space features are the output values ​​of the convolutional neural network.

6. According to claim 4, a method for predicting population thermal data based on spatiotemporal feature attention mechanism is characterized in that: The spatiotemporal feature fusion attention is reintegrated with the backbone network to capture the spatiotemporal correlation between population heat tensors at different times, thereby generating new features associated with spatiotemporal context.

7. According to claim 4, a method for predicting population thermal data based on spatiotemporal feature attention mechanism is characterized in that: The starting point features and end point features of each node are calculated to obtain the final representation reflecting the spatiotemporal contextual information features. The output of the spatiotemporal attention transfer mechanism model layer is regressed through the fully connected layer and the tanh activation function to obtain the start / end flow value at the next moment, so as to complete the spatiotemporal attention fusion population thermal prediction.

8. According to claim 1, a method for predicting population thermal data based on spatiotemporal feature attention mechanism is characterized in that: Zero padding is performed on nodes that cannot generate data. The zero padding logic formula is as follows: Among them, M is the spatial data, and m is the true value of M.

9. According to claim 4, a method for predicting population thermal data based on spatiotemporal feature attention mechanism is characterized in that: The time domain feature compression includes: A time compression framework is established, taking the population density tensor matrix in the window of a given time lt as input, and predicting the population density tensor matrix at the output time t+k. The formula is as follows: Among them, K is the number of time output windows, Input frame for spacetime, Output framework for space-time; x t1 =F sq (x)=Avg[x(i,j)],i∈H,j∈W x t2 =F sq (x)=Max[x(i,j)],i∈H,j∈W x tn =F sq (x)=Sf[x(i,j)],i∈H,j∈W Among them, the above formulas correspond to time domain mean compression, time domain maximum value compression, and other time domain statistical compression from top to bottom, respectively. sq is the compression function.

10. The method for predicting population thermal data based on spatiotemporal feature attention mechanism according to claim 4, characterized in that: The spatial domain feature compression includes spatial domain mean compression, spatial domain maximum compression and other spatial domain statistical compression on the T dimension. The corresponding formula is as follows: x s1 =F sq (x)=Avg[x(k)],k∈T x s2 =F sq (x)=Max[x(k)],k∈T x sn =F sq (x)=Sf[x(k)],k∈T Among them, the above formulas correspond to spatial domain mean compression, spatial domain maximum compression, and other statistical compressions in spatial domain from top to bottom. sq is the compression function.

Images