Population space-time distribution prediction method and system considering behavior space-time heterogeneity

Abstract

The invention discloses a population spatio-temporal distribution prediction method and system considering behavior spatio-temporal heterogeneity. The method comprises the following steps: S1, constructing an explicit function relationship of spatio-temporal dependence between a social media digital footprint GDF and a population based on a TFDF (Tracking Frequency per capita) of the digital footprint; s2, based on a Transform encoder, fusing the static geographic features of the space unit, the dynamic environment features and a historical social media digital footprint GDF sequence, and generating space-time associated context features; s3, constructing a TFDF prediction model to predict a future period TFDF and constructing a population prediction model to predict a future period population; and training a TFDF prediction model and a population prediction model by taking the explicit function relationship of the space-time dependence between the social media digital footprint and the population as training constraints. According to the method, the spatial-temporal heterogeneity of the social media digital footprint and the population relationship is modeled, and the spatial heterogeneity of the population time sequence mode is considered based on the feature fusion of the multi-source geographic spatial data, so that the generalization ability of population spatial-temporal distribution prediction is improved by fully utilizing massive GDF data.

Description

Technical Field

[0001] The present invention relates to the field of urban geographic intelligent computing, and in particular to a method and system for predicting the spatiotemporal distribution of a population taking into account the spatiotemporal heterogeneity of behavior. Background Art

[0002] High-resolution population distribution data can reveal the instantaneous distribution and changing trends of urban populations, making them crucial for urban applications such as disaster emergency management, infectious disease prevention and control, dynamic resource allocation, and crowd monitoring. Regional population data typically comes from large-scale censuses conducted every five to ten years. Numerous studies have conducted fine-scale population distribution estimation based on these census data. However, individuals are constantly on the move, and censuses and their derived data cannot capture the temporal dynamics of population distribution. While daily behaviors exhibit specific spatiotemporal patterns, such as commuting, work, leisure, and staying at home, individual movements are also random, particularly during emergencies. Therefore, real-time prediction of the spatiotemporal distribution of urban populations while protecting individual privacy is both highly practical and challenging.

[0003] With the ubiquity of mobile devices and social media, humans are constantly generating massive amounts of geolocated digital footprints (GDFs), which record their activities and locations. These footprints are highly timely and can reveal relative activity hotspots across time and space, providing a data foundation for high-resolution population distribution modeling. However, each type of GDF is generated by specific online behaviors and, in most cases, can only represent relative population density. Furthermore, from a behavioral geography perspective, the activity of location-based behaviors on social media is influenced by both social rhythms and environmental factors. Therefore, the functional relationship between the number of GDFs and the actual population is not constant across time and space, but rather exhibits spatiotemporal heterogeneity.

[0004] Existing high-resolution population distribution forecasting methods typically employ a time series forecasting framework, employing autoregressive models to map historical and future population sequences across spatial units. Mainstream autoregressive models include ARIMA, LSTM, Transformer, and their variants. These methods neglect the integration of geographic spatial and temporal features, making it difficult to model the spatial heterogeneity of population time series patterns. Furthermore, when population training data is scarce, the spatial and temporal generalization of population forecasts cannot be guaranteed.

[0005] Therefore, this paper proposes a high-spatiotemporal-resolution population distribution autoregressive prediction method (HSTP) that accounts for spatiotemporal behavioral heterogeneity. First, to model the spatial heterogeneity of population temporal patterns, a Transformer-based encoder is designed to fuse static geographic features of spatial units, dynamic environmental factors, and historical GDF sequences to form spatiotemporal contextual features. Then, to reduce HSTP's reliance on population training data and fully utilize GDF data, the per capita trigger frequency of digital footprints (TFDF) is proposed to quantify the spatiotemporal activity of social media location-based behavior. Furthermore, an explicit functional relationship between the spatiotemporal dependency of GDF and population is constructed (GDF = population × TFDF). Finally, population and TFDF, as intermediate output variables of HSTP, are predicted by a Transformer decoder that uses spatiotemporal contextual features as input. This functional relationship serves as a training constraint, allowing the population prediction module to be continuously optimized based on GDF data even when population data is missing, ensuring its spatiotemporal generalization. Summary of the Invention

[0006] In response to the problems in existing population spatiotemporal distribution prediction methods, such as difficulty in modeling the spatial heterogeneity of population time series patterns and over-reliance on large amounts of population data, the present invention proposes a population spatiotemporal distribution prediction method that takes into account the spatiotemporal heterogeneity of behavioral patterns.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions: A method for predicting the spatiotemporal distribution of a population taking into account the spatiotemporal heterogeneity of behavior comprises the following steps: S1. Based on the per capita trigger frequency TFDF of digital footprints, we construct an explicit functional relationship between the temporal and spatial dependence of social media digital footprints (GDF) and population; S2, the Transformer-based encoder fuses the static geographic features of spatial units, dynamic environmental features, and the GDF sequence of historical social media digital footprints to generate spatiotemporal contextual features; S3. Based on the spatiotemporal contextual features, a TFDF prediction model is constructed through a masked multi-head self-attention mechanism to predict the future TFDF; based on the spatiotemporal contextual features and historical GDF data, a population prediction model is constructed to predict the future population; and the explicit functional relationship of the spatiotemporal dependence between the social media digital footprint GDF and the population is used as a training constraint to train the TFDF prediction model and the population prediction model.

[0008] Furthermore, the explicit functional relationship between the temporal and spatial dependence of the social media digital footprint GDF and population in step S1 is: GDF = population × TFDF Among them, GDF is the social media digital footprint, and TFDF is the per capita triggering frequency of the digital footprint.

[0009] Furthermore, step S2 specifically includes the following sub-steps: S21. Processing multi-source geospatial data into feature vectors, including static geographic features and dynamic environmental features; S22. Construct a Transformer-based spatiotemporal context feature encoder CF-Encoder, which uses a multi-head attention mechanism to fuse the static geographic features, dynamic environmental factors, and historical GDF sequences of spatial units to generate spatiotemporal context features of spatial units. .

[0010] Furthermore, the static geographic features include the area ratio of land cover types and point of interest features calculated based on buffer zone attenuation; the dynamic environmental features include weather forecast and holiday information encoded through a semantic model.

[0011] Furthermore, the calculation formula of the point of interest feature based on the buffer attenuation calculation is:

[0012]

[0013] in, For grid cells POI features, Grid cells Inside POI-like POI features; For building blocks The radius is The buffer zone, Then calculate the buffer The number of class POIs, is the attenuation coefficient of POI influence.

[0014] Furthermore, the spatiotemporal context feature encoder CF-Encoder includes an embedding layer, a position encoding layer and two Transformer encoding layers, denoted as TEL0 and TEL1 respectively; The embedding layer converts the original input Convert to the dimensions required for TEL0 Generate feature vectors , the position encoding layer converts the temporal information Incorporate into the feature vector In the example above, we get the feature vector with time series information ,in is the number of input samples, is the number of historical moments, is the input feature vector dimension; The feature vector with timing information As the input of TEL0, the multi-head attention module is first used to perform feature fusion in the time dimension; then the residual connection and layer normalization LN are used to improve the stability of the input; then, the feedforward propagation network FFN performs independent nonlinear transformation on the features at each time step; finally, the residual connection and layer normalization LN are used again to improve the stability of the output, and the output of TEL0 is obtained. ; The output of the TEL0 As the input of TEL1, the same calculation process of TEL0 is used to generate spatiotemporal context features. .

[0015] Furthermore, step S3 specifically includes the following sub-steps: S31. Construct TFDF prediction model; use attention mechanism to model the temporal association between spatial unit features and TFDF, and dynamically and adaptively extract context features from spatiotemporal associations. Extract the required information and then predict the TFDF of the spatial unit at the subsequent moment; the TFDF prediction model includes an embedding layer, a position encoding layer, two Transformer decoding layers and a linear mapping layer, and the two Transformer decoding layers are represented as TDL0 and TDL1 respectively; S32. Build a population prediction model, use the attention mechanism to capture the temporal correlation between spatial unit features, historical GDF and future population, and adaptively extract contextual features from spatiotemporal correlations. Extract the required information from the data and then predict the population of the spatial unit at subsequent moments; S33. Construct a joint loss function, and based on the joint loss function, use the explicit functional relationship of the spatiotemporal dependency between the social media digital footprint GDF and the population as a training constraint to train the TFDF prediction model and the population prediction model.

[0016] Furthermore, the joint loss function for:

[0017] in, and are the weights of population prediction loss and GDF prediction loss, is the population prediction loss function, is the GDF prediction loss function.

[0018] Furthermore, the population prediction loss function and GDF prediction loss function are:

[0019]

[0020] in, and Space units At the moment The population prediction value and the actual value; and Grid At the moment The GDF predicted value and the true value, represents the mean square error loss function.

[0021] In another aspect, the present invention provides a population spatiotemporal distribution prediction system that takes into account spatiotemporal heterogeneity of behavior, comprising: The data preprocessing module is used to construct an explicit functional relationship between the temporal and spatial dependence of social media digital footprints (GDF) and population based on the per capita trigger frequency (TFDF) of digital footprints; The spatiotemporal context feature encoding module is used to fuse the static geographic features of spatial units, dynamic environmental features, and historical social media digital footprint GDF sequences based on the Transformer encoder to generate spatiotemporal context features; A dual-decoder prediction module is configured to construct a TFDF prediction model based on the spatiotemporal contextual features using a masked multi-head self-attention mechanism to predict the future TFDF; construct a population prediction model based on the spatiotemporal contextual features and historical GDF data to predict the future population; and train the TFDF prediction model and the population prediction model using the explicit functional relationship of the spatiotemporal dependency between the social media digital footprint GDF and the population as a training constraint. Compared with the prior art, the present invention has the following beneficial effects: The high-spatiotemporal-resolution population distribution prediction method proposed in this paper, which takes into account the spatiotemporal heterogeneity of behavioral behavior, explicitly models the spatiotemporal heterogeneity of the relationship between social media digital footprints (GDF) and population, and considers the spatial heterogeneity of population time series patterns based on the feature fusion of multi-source geospatial data, thereby making full use of massive GDF data to improve the generalization ability of population spatiotemporal distribution prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0023] Figure 1 The present invention is a flowchart of the method. DETAILED DESCRIPTION

[0024] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0025] Example 1 like Figure 1 As shown, the present invention provides a method for predicting the spatiotemporal distribution of a population taking into account the spatiotemporal heterogeneity of behavior, comprising the following steps: S1. Based on the per capita trigger frequency of digital footprints (TFDF), we construct an explicit functional relationship between the temporal and spatial dependence of social media digital footprints (GDF) and population: GDF = population × TFDF; Furthermore, the specific implementation of the per capita trigger frequency (TFDF) of digital footprints proposed in step S1 to quantify the spatiotemporal activity of social media location-based behavior is as follows: Given a spatial unit ,and At the moment The population and GDF numbers are and , then the space unit at time The calculation formula of TFDF is as follows: (1) TFDF expresses the per capita frequency of people generating location data (GDF) on social media in a given spatial area and a given time period, and thus an explicit functional relationship between GDF and population can be constructed on this basis.

[0026] S2: The Transformer-based encoder fuses the static geographic features of the spatial unit, the dynamic environmental features, and the GDF sequence of the historical social media digital footprint to generate spatiotemporal contextual features. Step S2 specifically includes the following sub-steps: S21. Multi-source geospatial data preprocessing Multi-source geospatial data is processed into feature vectors. First, the static geographic features of the spatial unit are calculated using land cover type and point of interest (POI) data. The feature value of the land cover type data is the area ratio of each cover type within the spatial unit. For POI feature calculation, this embodiment uses AutoNavi POI data, which has 23 first-level classifications, covering catering services, shopping services, life services, sports and leisure services, medical care services, accommodation services, and scenic spots. In order to alleviate the feature sparsity problem and considering that the service range of POI is not limited to a single spatial grid and its influence decays with distance, formula (2) is used to calculate a certain grid unit. POI features: (2) (3) in, For grid cells POI features, Grid cells Inside POI-like POI features; For building blocks The radius is The buffer zone, Then calculate the buffer The number of class POIs, is the attenuation coefficient of POI influence.

[0027] In this embodiment, considering that a large buffer radius may cause feature homogeneity, the buffer radius list is set to meters, and .

[0028] Next, weather forecasts and holiday information are used as dynamic environmental features. Weather and holiday information includes categorical information in multiple dimensions (such as sunny, rainy, and cloudy weather types, and weekdays and weekends holiday types) and continuous values ​​(such as temperature). To avoid complex feature engineering when processing this type of data, this example uses the semantic model stella-base-zh-v2 to output the first 32-dimensional vectors as dynamic semantic features corresponding to weather and holiday information. Taking March 3, 2024, as an example, the weather and holiday description text input to the semantic model is "This day is {weekend}, the weather is {sunny / cloudy}, and the temperature is {2-7°C}."

[0029] S22, Transformer-based spatiotemporal context feature encoder (CF-Encoder) The spatiotemporal context feature encoding module (CF-Encoder) uses a multi-head attention mechanism to fuse the static geographic features, dynamic environmental factors, and historical GDF sequences of spatial units to generate spatiotemporal context features of spatial units, providing the necessary foundation for modeling the spatial heterogeneity of population temporal patterns. Specifically, the CF-Encoder consists of an embedding layer, a position encoding layer, and two Transformer encoding layers (represented as TEL0 and TEL1, respectively). As shown in formulas (4)-(5), the embedding layer converts the original input Convert to the dimensions required by TEL0 (i.e. ), the position encoding layer then converts the temporal information Incorporated into the transformed feature vector, we get ,in is the batch size of the current input, is the number of historical moments, is the input feature vector dimension. and Represent the weight matrix and bias respectively, which need to be optimized during training.

[0030] (4) (5) The input of TEL0 is shown in formulas (6)-(9). First, a multi-head attention module is used to fuse the temporal features, see formula (6); then, residual connection and layer normalization (LN) are used to improve the stability of the input and thus accelerate the convergence speed during model training, as shown in formula (7); then, the feed-forward network (FFN) performs an independent nonlinear transformation on the features at each time step, thereby enhancing the local feature extraction capability, see formula (8), where ReLU is the activation function; finally, residual connection and LN are used again to improve the stability of the output, and the output of TEL0 is obtained. .

[0031] (6) (7) (8) (9) Single-head attention in formula (6) The calculation formulas are shown in (10)-(12), where is the number of attention heads.

[0032] (10) (11) (12) TEL1 input This is the output of TEL0 , and then use the same calculation process of formula (6)-(9) to obtain the final spatiotemporal context features , which will serve as the input to TFDF-Decoder and POP-Decoder.

[0033] S3. Based on the spatiotemporal contextual features, a TFDF prediction model is constructed through a masked multi-head self-attention mechanism to predict the future TFDF; based on the spatiotemporal contextual features and historical GDF data, a population prediction model is constructed to predict the future population; and the explicit functional relationship of the spatiotemporal dependence between the social media digital footprint GDF and the population is used as a training constraint to train the TFDF prediction model and the population prediction model.

[0034] In this embodiment, step S3 specifically includes the following sub-steps: S31, TFDF-Decoder The TFDF prediction module (TFDF-Decoder) uses the attention mechanism to model the temporal association between spatial unit features and TFDF, and dynamically and adaptively extracts temporal and spatial context features from the spatial and temporal context features. The TFDF-Decoder is composed of an embedding layer, a position encoding layer, two Transformer decoding layers (denoted as TDL0 and TDL1), and a linear mapping layer.

[0035] The current time is , when predicting the current batch grid at time When the per capita trigger frequency is ,and to External dynamic characteristics of the moment Concatenate in the feature vector dimension as the original input of TFDF-Decoder In order to avoid label leakage, the GDF value is not used as the input of TFDF-Decoder. The feature dimension is The embedding layer and position encoding layer of TFDF-Decoder are calculated in the same way as in CF-Encoder, as shown in formulas (4)-(5). Convert to .

[0036] As the input of TDL0, the calculation process of TDL0 is shown in formulas (13)-(15). First, a masked multi-head self-attention module is used to make the features of each time step only integrate the temporal information before it, and residual connection and LN are used to improve the training stability, as shown in formula (13); then, a multi-head attention module is used to make the features of each time step can be output from the CF-Encoder. The required information is extracted from the network, and then passes through the residual connection and LN, as shown in formula (14); finally, the feed-forward network (FFN) performs an independent nonlinear transformation on the features at each time step, and then passes through the residual connection and LN again, as shown in formula (15).

[0037] (13) (14) (15) Input of TDL1 This is the output of TDL0 , and then use the same calculation process of formula (16)-(18) to get the output of TDL1 Finally, a linear layer is applied to Mapped into TFDF prediction value, considering that TFDF is always non-negative, the activation function ReLU is used to constrain the TFDF prediction value, see formula (16).

[0038] (16) Since the average trigger frequency per person in a grid at a certain moment is a real number, the dimension of the last dimension output by TFDF-Decoder is 1, that is, ,in The number of grids (samples) for the current input. For the current batch of grids at time to The predicted value of per capita trigger frequency.

[0039] S32, Population Decoder (POP-Decoder) The population prediction module (POP-Decoder) uses the attention mechanism to capture the temporal correlation between spatial unit features, historical GDF and future population, and adaptively extracts the temporal and spatial context features from the population prediction module. The required information is extracted from the GDF, and then the population prediction accuracy is improved by modeling the implicit relationship between GDF and population.

[0040] The structure of POP-Decoder is the same as TFDF-Decoder, and it is co-optimized through the GDF-population function relationship during training. The original input of POP-Decoder contains the historical GDF value, so the original input of POP-Decoder can be recorded as , and through the calculation process consistent with TFDF-Decoder (see formulas (4)-(5) and (13)-(16)), the current batch grid at time to Population forecast ,in is the batch size of the current input, is the number of future moments. Based on the autoregressive framework of HSTP and the prediction , we can continuously predict the population and TFDF at subsequent times.

[0041] Model training under the constraints of S33, GDF and population function relationship The training loss of HSTP is the weighted sum of population prediction loss and GDF prediction loss, as shown in formula (19). and Space units At the moment The population prediction value and the actual value; and Grid At the moment The GDF predicted value and true value; and are the weights of population prediction loss and GDF prediction loss, respectively, and their sum is equal to 1.

[0042] (17) (18) (19) When the population data for a certain period is missing, let , at this time HSTP can still predict the loss based on GDF Conduct training and optimization to improve the population prediction generalization ability of HSTP.

[0043] Specifically, the model was constructed and trained using the PyTorch deep learning framework, with stochastic gradient descent as the optimizer. To determine the optimal values ​​of the hyperparameters in HSTP, a random search method was used, with a search limit of 50 times. The search ranges for each hyperparameter and the optimal values ​​obtained are shown in Table 1. To stabilize the model training process, both the population and the true GDF values ​​were square-root transformed during training, and the predicted values ​​were square-transformed during prediction.

[0044] Table 1 Hyperparameter search range and its optimal value

[0045] The proposed model explicitly models the spatiotemporal heterogeneity of the relationship between social media digital footprints (GDF) and population. It also considers the spatial heterogeneity of population temporal patterns by integrating features from multi-source geospatial data. This model leverages massive amounts of GDF data to improve the generalization of population spatiotemporal distribution predictions. This method can support refined urban management, including disaster emergency management, infectious disease prevention and control, dynamic resource allocation, and population monitoring.

[0046] Example 2 This embodiment provides a population spatiotemporal distribution prediction system that takes into account spatiotemporal heterogeneity of behavior, including: The data preprocessing module is used to construct an explicit functional relationship between the temporal and spatial dependence of social media digital footprints (GDF) and population based on the per capita trigger frequency (TFDF) of digital footprints; The spatiotemporal context feature encoding module is used to fuse the static geographic features of spatial units, dynamic environmental features, and historical social media digital footprint GDF sequences based on the Transformer encoder to generate spatiotemporal context features; A dual-decoder prediction module is used to construct a TFDF prediction model based on the spatiotemporal contextual features through a masked multi-head self-attention mechanism to predict the future TFDF; to construct a population prediction model based on the spatiotemporal contextual features and historical GDF data to predict the future population; and to train the TFDF prediction model and the population prediction model using the explicit functional relationship of the spatiotemporal dependence between the social media digital footprint GDF and the population as a training constraint.

[0047] It should be understood that parts not elaborated in detail in this specification belong to the prior art.

[0048] It should be understood that the above description of the preferred embodiments is relatively detailed and cannot be considered as limiting the scope of protection of the present invention. It is not necessary and impossible to list all embodiments here. Under the guidance of the present invention, ordinary technicians in this field can also make substitutions or modifications without departing from the scope of protection of the claims of the present invention, which fall within the scope of protection of the present invention. The scope of protection of the present invention shall be based on the attached claims.

Claims

1. A method for predicting the spatiotemporal distribution of a population taking into account the spatiotemporal heterogeneity of behavior, characterized by: The steps include: S1. Based on the per capita trigger frequency TFDF of digital footprints, we construct an explicit functional relationship between the temporal and spatial dependence of social media digital footprints (GDF) and population; S2, the Transformer-based encoder fuses the static geographic features of spatial units, dynamic environmental features, and the GDF sequence of historical social media digital footprints to generate spatiotemporal contextual features; S3. Based on the spatiotemporal contextual features, a TFDF prediction model is constructed through a masked multi-head self-attention mechanism to predict the future TFDF; based on the spatiotemporal contextual features and historical GDF data, a population prediction model is constructed to predict the future population; and the explicit functional relationship of the spatiotemporal dependence between the social media digital footprint GDF and the population is used as a training constraint to train the TFDF prediction model and the population prediction model.

2. A method for predicting the spatiotemporal distribution of a population taking into account the spatiotemporal heterogeneity of behavior according to claim 1, characterized in that: The explicit functional relationship between the temporal and spatial dependence of social media digital footprint GDF and population in step S1 is: GDF = population × TFDF Among them, GDF is the social media digital footprint, and TFDF is the per capita triggering frequency of the digital footprint.

3. A method for predicting population spatiotemporal distribution taking into account spatiotemporal heterogeneity of behavior according to claim 1, characterized in that: Step S2 specifically includes the following sub-steps: S21. Processing multi-source geospatial data into feature vectors, including static geographic features and dynamic environmental features; S22. Construct a Transformer-based spatiotemporal context feature encoder CF-Encoder, which uses a multi-head attention mechanism to fuse the static geographic features, dynamic environmental factors, and historical GDF sequences of spatial units to generate spatiotemporal context features of spatial units. .

4. A method for predicting population spatiotemporal distribution taking into account spatiotemporal heterogeneity of behavior according to claim 3, characterized in that: The static geographic features include the area ratio of land cover types and the features of points of interest calculated based on buffer zone attenuation; The dynamic environment features include weather forecast and holiday information encoded by a semantic model.

5. A method for predicting population spatiotemporal distribution taking into account spatiotemporal heterogeneity of behavior according to claim 4, characterized in that: The calculation formula of interest point features based on buffer attenuation calculation is: in, For grid cells POI features, Grid cells Inside POI-like POI features; For building blocks The radius is The buffer zone, Then calculate the buffer The number of class POIs, is the attenuation coefficient of POI influence.

6. The method for predicting population spatiotemporal distribution taking into account spatiotemporal heterogeneity of behavior according to claim 3, characterized in that: The spatiotemporal context feature encoder CF-Encoder includes an embedding layer, a position encoding layer and two Transformer encoding layers, denoted as TEL0 and TEL1 respectively; The embedding layer converts the original input Convert to the dimensions required for TEL0 Generate feature vectors , the position encoding layer converts the temporal information Incorporate into the feature vector In the example above, we get the feature vector with time series information ,in is the number of input samples, is the number of historical moments, is the input feature vector dimension; The feature vector with timing information As the input of TEL0, a multi-head attention module is first used to fuse features in the time dimension; then residual connections and layer normalization (LN) are used to improve the stability of the input; then, a feedforward propagation network (FFN) performs independent nonlinear transformations on the features at each time step; Finally, we use residual connection and layer normalization LN again to improve the stability of the output and get the output of TEL0 ; The output of TEL0 As the input of TEL1, the same calculation process of TEL0 is used to generate spatiotemporal context features. .

7. The method for predicting population spatiotemporal distribution taking into account spatiotemporal heterogeneity of behavior according to claim 1, characterized in that: Step S3 specifically includes the following sub-steps: S31. Construct TFDF prediction model; use attention mechanism to model the temporal association between spatial unit features and TFDF, and dynamically and adaptively extract context features from spatiotemporal associations. Extract the required information and then predict the TFDF of the spatial unit at the subsequent moment; the TFDF prediction model includes an embedding layer, a position encoding layer, two Transformer decoding layers and a linear mapping layer, and the two Transformer decoding layers are represented as TDL0 and TDL1 respectively; S32. Build a population prediction model, use the attention mechanism to capture the temporal correlation between spatial unit features, historical GDF and future population, and adaptively extract contextual features from spatiotemporal correlations. Extract the required information from the data and then predict the population of the spatial unit at subsequent moments; S33. Construct a joint loss function, and based on the joint loss function, use the explicit functional relationship of the spatiotemporal dependency between the social media digital footprint GDF and the population as a training constraint to train the TFDF prediction model and the population prediction model.

8. A method for predicting population spatiotemporal distribution taking into account spatiotemporal heterogeneity of behavior according to claim 7, characterized in that: The joint loss function for: in, and are the weights of population prediction loss and GDF prediction loss, is the population prediction loss function, is the GDF prediction loss function.

9. A method for predicting population spatiotemporal distribution taking into account spatiotemporal heterogeneity of behavior according to claim 8, characterized in that: The population prediction loss function and GDF prediction loss function are: in, and Space units At the moment The population prediction value and the actual value; and Grid At the moment The GDF predicted value and the true value, represents the mean square error loss function.

10. A population spatiotemporal distribution prediction system that takes into account spatiotemporal heterogeneity of behavior, characterized by: include: The data preprocessing module is used to construct an explicit functional relationship between the temporal and spatial dependence of social media digital footprints (GDF) and population based on the per capita trigger frequency (TFDF) of digital footprints; The spatiotemporal context feature encoding module is used to fuse the static geographic features of spatial units, dynamic environmental features, and historical social media digital footprint GDF sequences based on the Transformer encoder to generate spatiotemporal context features; A dual-decoder prediction module is configured to construct a TFDF prediction model based on the spatiotemporal contextual features using a masked multi-head self-attention mechanism to predict the future TFDF; construct a population prediction model based on the spatiotemporal contextual features and historical GDF data to predict the future population; and train the TFDF prediction model and the population prediction model using the explicit functional relationship of the spatiotemporal dependency between the social media digital footprint GDF and the population as a training constraint. The high spatiotemporal resolution population distribution prediction system taking into account spatiotemporal heterogeneity of behaviors is used to execute the steps in the high spatiotemporal resolution population distribution prediction method taking into account spatiotemporal heterogeneity of behaviors as described in any one of claims 1-9.

Images