WO2023207411A1 - 一种基于时空数据的流量确定方法、装置、设备和介质 - Google Patents

一种基于时空数据的流量确定方法、装置、设备和介质 Download PDF

Info

Publication number
WO2023207411A1
WO2023207411A1 PCT/CN2023/082204 CN2023082204W WO2023207411A1 WO 2023207411 A1 WO2023207411 A1 WO 2023207411A1 CN 2023082204 W CN2023082204 W CN 2023082204W WO 2023207411 A1 WO2023207411 A1 WO 2023207411A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
historical traffic
spatio
spatiotemporal
traffic data
Prior art date
Application number
PCT/CN2023/082204
Other languages
English (en)
French (fr)
Inventor
宋礼
张钧波
郑宇�
Original Assignee
京东城市(北京)数字科技有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 京东城市(北京)数字科技有限公司 filed Critical 京东城市(北京)数字科技有限公司
Publication of WO2023207411A1 publication Critical patent/WO2023207411A1/zh

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent

Definitions

  • the present application relates to the field of computer technology, and in particular to a flow determination method, device, equipment and medium based on spatiotemporal data.
  • embodiments of the present invention provide a flow determination method, device, equipment and medium based on spatiotemporal data.
  • embodiments of the present invention provide a traffic determination method based on spatiotemporal data, which includes: obtaining the first historical traffic data in the target area; inputting the historical traffic data into a preset spatiotemporal network model as input data, Obtain the target traffic data of the target area within the target time period; wherein, the preset spatiotemporal network model includes a spatiotemporal feature layer, and the spatiotemporal feature layer obtains the historical traffic data through convolution kernels of multiple scales.
  • a variety of convolution features and determine the spatio-temporal characteristics of the historical traffic data from the multiple convolution features through a self-attention mechanism, so that the preset spatio-temporal network model determines the spatio-temporal characteristics based on the spatio-temporal characteristics Target traffic data for the target area during the target time period.
  • the method further includes: when the preset spatiotemporal network model includes multiple adjacent spatiotemporal feature layers, converting the previous one of the multiple adjacent spatiotemporal feature layers into The output data of the spatio-temporal feature layer is input into the subsequent spatio-temporal feature layer to determine the spatio-temporal features of the output data.
  • the method further includes: when the preset spatiotemporal network model also includes a first convolutional layer, input the historical traffic data into the first convolutional layer to obtain The high-dimensional features of the historical traffic data are input into the spatio-temporal feature layer.
  • the method further includes: when the preset spatiotemporal network model also includes a second convolutional layer, inputting the spatiotemporal features into the second convolutional layer, and Perform a down-sampling operation on the spatio-temporal features to obtain the target traffic data of the target area within the target time period.
  • the method further includes training to obtain the preset spatio-temporal network model according to the following steps: dividing the training area into grids and determining a plurality of grid areas; for each grid area , obtain the second historical traffic data of the grid area in multiple time periods; splice the second historical traffic data of the multiple grid areas in the same time period to obtain the second historical traffic data of the same time period.
  • two traffic characteristic data obtain a historical traffic sequence according to the second traffic characteristic data in the multiple time periods, and use the historical traffic sequence as a training data set; train the training data set to obtain the Describe the preset spatiotemporal network model.
  • training the training data set to obtain the preset spatiotemporal network model includes: determining the size of a sliding window; using the sliding window to extract input samples from the historical traffic sequence , the data volume of the input sample is consistent with the size of the sliding window; according to the extracted input sample, an output sample is extracted from the historical traffic sequence, and the output sample is used as the label corresponding to the input sample; based on the The input samples and labels are used for training to obtain the preset spatiotemporal network model.
  • determining the size of the sliding window includes: determining the size of the sliding window to be N;
  • Using the sliding window to extract input samples from the historical traffic sequence includes: determining a starting time; extracting historical traffic data corresponding to the starting time from the historical traffic sequence, and extracting data before the starting time.
  • the historical traffic data corresponding to (N-1) times, the historical traffic data corresponding to the starting time and the historical traffic data corresponding to (N-1) times before the starting time are used as input samples ;
  • extracting an output sample from the historical traffic sequence includes: extracting historical traffic data corresponding to the next time after the starting time from the historical traffic sequence as an output sample.
  • obtaining the first historical traffic data in the target area includes: rasterizing the target area and determining a plurality of grid areas; for each of the grid areas, obtaining the The first historical traffic data of the grid area; determining the first traffic characteristic data of the target area according to the first historical traffic data of each grid area; inputting the historical traffic data as input data into a preset
  • the spatiotemporal network model includes: inputting the first traffic characteristic data of the target area as input data into a preset spatiotemporal network model.
  • embodiments of the present invention provide a traffic determination device based on spatiotemporal data, including: an acquisition module, used to acquire the first historical traffic data in the target area; and a determination module, used to use the historical traffic data as Input data into a preset spatiotemporal network model to obtain target traffic data of the target area within the target time period; wherein the preset spatiotemporal network model includes at least one spatiotemporal feature layer, and the spatiotemporal feature layer passes through a variety of The scale convolution kernel obtains a variety of convolution features of the historical traffic data, and uses the self-attention machine to Determine the spatiotemporal characteristics of the historical traffic data from the multiple convolution features, so that the preset spatiotemporal network model determines the target traffic data of the target area within the target time period based on the spatiotemporal characteristics. .
  • the determination module is further configured to: when the preset spatiotemporal network model includes multiple adjacent spatiotemporal feature layers, combine the multiple adjacent spatiotemporal feature layers into The output data of the previous spatio-temporal feature layer is input into the subsequent spatio-temporal feature layer to determine the spatio-temporal features of the output data.
  • the determination module is further configured to: when the preset spatiotemporal network model also includes a first convolutional layer, input the historical traffic data into the first convolutional layer. , obtain the high-dimensional features of the historical traffic data, and input the high-dimensional features into the spatio-temporal feature layer.
  • the determination module is further configured to: when the preset spatiotemporal network model also includes a second convolutional layer, input the spatiotemporal features into the second convolutional layer, A downsampling operation is performed on the spatio-temporal features to obtain target traffic data of the target area within a target time period.
  • the device further includes a training module, configured to: divide the training area into grids and determine multiple grid areas; and for each grid area, obtain the location of the grid area.
  • second historical traffic data in multiple time periods splicing the second historical traffic data of the multiple grid areas in the same time period to obtain second traffic characteristic data in the same time period; according to the The second traffic characteristic data in multiple time periods is used to obtain a historical traffic sequence, and the historical traffic sequence is used as a training data set; the training data set is trained to obtain the preset spatiotemporal network model.
  • the training module is also used to: determine the size of the sliding window; use the sliding window to extract input samples from the historical traffic sequence, the data amount of the input sample is the same as the sliding window.
  • the training module is also used to determine the size of the sliding window to be N; determine a starting time; extract historical traffic data corresponding to the starting time from the historical traffic sequence, and Extract the historical traffic data corresponding to (N-1) moments before the starting time, and combine the historical traffic data corresponding to the starting time and the (N-1) moments before the starting time.
  • the historical traffic data is used as an input sample; the historical traffic data corresponding to the next moment after the starting time is extracted from the historical traffic sequence as an output sample.
  • the acquisition module is further configured to: rasterize the target area and determine a plurality of grid areas; for each grid area, obtain the information of the grid area.
  • first historical traffic data determining first traffic characteristic data of the target area based on the first historical traffic data of each grid area; Entering the historical traffic data into a preset spatiotemporal network model as input data includes: inputting the first traffic characteristic data of the target area into the preset spatiotemporal network model as input data.
  • embodiments of the present invention provide an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory complete communication with each other through the communication bus; the memory is used to Store at least one executable instruction.
  • the executable instruction causes the processor to perform the following steps: obtain the first historical traffic data in the target area; input the historical traffic data as input data into a preset spatiotemporal network model to obtain The target traffic data of the target area within the target time period; wherein the preset spatiotemporal network model includes at least one spatiotemporal feature layer, and the spatiotemporal feature layer obtains the historical traffic data through convolution kernels of multiple scales A variety of convolution features, and determine the spatiotemporal features of the historical traffic data from the multiple convolution features through a self-attention mechanism, so that the preset spatiotemporal network model determines the spatiotemporal features based on the spatiotemporal features. Describe the target traffic data of the target area within the target time period.
  • embodiments of the present invention provide a computer-readable storage medium on which a computer program is stored.
  • the computer program is executed by a processor, the flow determination method based on spatiotemporal data of the embodiment of the present invention is implemented.
  • the traffic determination method based on spatiotemporal data in the embodiment of the present invention can monitor, predict and control city-level traffic, provide more accurate passenger flow predictions, thereby achieving more sophisticated management and control solutions and more sophisticated urban planning; this method It can be used in scenarios such as discovery and advertising in densely populated areas in cities, and dynamic advertising space pricing.
  • Figure 1 schematically shows a schematic diagram of the spatio-temporal feature layer of the spatio-temporal network model according to an embodiment of the present invention
  • Figure 2 schematically shows the structural diagram of the spatio-temporal network model according to the embodiment of the present invention
  • Figure 3 schematically shows the main steps of the method for training a spatiotemporal network model according to an embodiment of the present invention
  • Figure 4 schematically shows a flow chart of the sub-steps of the method for training a spatiotemporal network model according to an embodiment of the present invention
  • Figure 5 schematically shows a flow chart of a traffic determination method based on spatiotemporal data according to an embodiment of the present invention
  • Figure 6 schematically shows a structural diagram of a flow determination device based on spatiotemporal data according to an embodiment of the present invention
  • Figure 7 schematically shows the system architecture of a traffic determination method based on spatiotemporal data suitable for an embodiment of the present invention
  • Figure 8 schematically shows a structural block diagram of an electronic device provided by an embodiment of the present invention.
  • neural networks can currently be used for spatiotemporal data modeling.
  • residual neural networks can be introduced into ST-ResNet to extract spatiotemporal features through residual neural networks.
  • spatiotemporal characteristics include characteristics in two dimensions: temporal dependence and spatial dependence.
  • Time dependence refers to the relationship between traffic changes in each area and historical traffic
  • spatial dependence refers to the relationship between traffic flows between different areas.
  • fixed-size receptive fields are used, such as 3*3, 5*5, etc., and the parameters of spatial positions are shared, which has spatial invariance.
  • application scenarios do not satisfy this spatial invariance relationship.
  • embodiments of the present invention provide a traffic determination method based on spatiotemporal data, which analyzes the historical traffic data in the target area through a preset spatiotemporal network model. Determine the target traffic data of the target area within the target time period, where the preset
  • the spatiotemporal network model includes a spatiotemporal feature layer, which includes convolution kernels of multiple scales to provide receptive fields of different sizes for different areas, and uses an attention mechanism to dynamically select receptive fields for each area, which can Accurately determine the relationship between traffic changes in each area and historical traffic, and determine the relationship between traffic flows between different areas, thereby accurately predicting the traffic in the target area.
  • the spatiotemporal network model in the embodiment of the present invention includes a spatiotemporal feature layer, which is mainly used to solve the problem of inconsistent receptive fields at different locations.
  • the spatio-temporal feature layer first uses different convolution kernels to extract spatio-temporal features, and then uses an attention mechanism to automatically select effective features for the extracted feature results in this area, thereby extracting the spatio-temporal features of the spatio-temporal data.
  • This enables the use of different convolution kernels for different areas, that is, the use of different receptive fields for different areas to extract spatiotemporal features.
  • spatiotemporal data is data with both time and space dimensions, such as people flow data and vehicle flow data in cities.
  • Figure 1 schematically shows the structural diagram of the spatio-temporal feature layer of the spatio-temporal network model according to the embodiment of the present invention.
  • H and W represent the number of grids in the longitude and latitude directions after a certain area is rasterized.
  • c, c’ represents the number of flow channels in the area, such as inflow, outflow, residence, etc.
  • K1 and K2 respectively represent the sizes of different convolution kernels, such as 3*3, 5*5, 7*7, etc.
  • this embodiment uses two convolution kernels of different sizes to extract spatiotemporal features for two different areas within the original range, and the size of the convolution kernel is dynamically selected through a self-attention mechanism. of.
  • Conv 3 (X); f 5
  • Conv 5 (X); f 7
  • Conv 7 (X); f 9
  • Conv 9 (X); f 11
  • Conv 11 (X).
  • the spatiotemporal feature layer uses a self-attention mechanism to select effective features from the various convolutional features.
  • the specific process is as follows:
  • W k is a learnable parameter matrix, which is used to increase the capacity of the spatiotemporal network model and can model complex relationships.
  • k 1 and k 2 are the sizes of the convolution kernel. is the mapped feature, It calculates the correlation of the output features of two different convolution kernels.
  • softmax() is a normalization function. is the normalized weight, s i, j is the feature fused through different convolution kernels, that is, the spatial and temporal features.
  • the spatio-temporal network model of the embodiment of the present invention is aimed at the same area.
  • the spatio-temporal feature layer first uses different convolution kernels to extract spatio-temporal features, and then uses an attention mechanism to automatically select the effective feature results extracted from this area.
  • Features thereby extracting the spatio-temporal features of spatio-temporal data, so as to use different convolution kernels for different areas, that is, to use different receptive fields for different areas to extract spatio-temporal features, which reflects the differences in different areas.
  • the characteristics of spatiotemporal data are fully considered.
  • the spatio-temporal network model of the embodiment of the present invention may include one spatio-temporal feature layer, or may include multiple adjacent spatio-temporal feature layers. Among the multiple adjacent spatio-temporal feature layers, the previous spatio-temporal feature layer The output data of the layer is the input data of the next spatiotemporal feature layer. In this embodiment, in order to fully consider the relationship between distant sub-regions in the original region, multiple adjacent spatio-temporal feature layers may be used.
  • the spatiotemporal network model of the embodiment of the present invention may also include a first convolutional layer and a second convolutional layer.
  • the spatiotemporal network model includes a first convolutional layer, two spatiotemporal feature layers and a second convolutional layer.
  • the first convolution layer is used to extract high-dimensional features, that is, to extract high-dimensional features of the original input data, and then the spatio-temporal feature layer extracts spatio-temporal features.
  • the second convolution layer performs dimensionality reduction operations on the spatio-temporal features to obtain The outcome to be predicted.
  • This embodiment uses the first convolution layer and the second convolution layer to improve the accuracy of the model while reducing the amount of calculation and improving efficiency.
  • training samples need to be obtained for training.
  • the training process of the spatiotemporal network model is shown in Figure 3. The process includes:
  • Step S301 Rasterize the training area and determine multiple grid areas
  • Step S302 For each grid area, obtain the second historical traffic data of the grid area in multiple time periods;
  • Step S303 Splice the second historical traffic data of the plurality of grid areas in the same time period to obtain the second traffic characteristic data of the same time period;
  • Step S304 Obtain a historical traffic sequence based on the second traffic characteristic data in the multiple time periods, and use the historical traffic sequence as a training data set;
  • Step S305 Train the training data set to obtain the preset spatiotemporal network model.
  • the training area is gridded to obtain multiple grid areas.
  • the training area can be a city or a certain area in the city.
  • the training area may be the target area, may include the target area, or may be included in the target area (that is, the training area is within the target area).
  • the trained spatiotemporal network model has the most accurate results when predicting the traffic in the target area.
  • Traffic characteristics include, for example, resident traffic, inflow traffic, and outflow traffic in the grid area.
  • the traffic characteristics of the grid area can be collected with a time step of 30 minutes.
  • the inflow flow of time segment t (within half an hour) is counted: time t-1 is not in the grid area D ij , and time t is in the grid area D ij .
  • time t-1 is in the grid area D ij
  • time t is not in the grid area D ij .
  • time t-1 is in the grid area D ij
  • time t is also in the grid area D ij .
  • the second historical traffic data of multiple grid areas in the same time period are spliced to obtain the second traffic characteristic data of the same time period (the traffic characteristic data
  • the data includes time dependence and spatial dependence).
  • the second historical traffic data of multiple grid areas at time t is spliced to obtain X t
  • the second historical traffic data at time t+1 is spliced to obtain X t+1
  • the historical traffic sequence ⁇ X 1 , X 2 ,..., X t , X t+1 ⁇ .
  • the historical traffic sequence is used as a training data set for training to obtain the spatiotemporal network model.
  • the historical traffic sequence is used as a training data set for training, and the process of obtaining the spatiotemporal network model includes:
  • Step S401 Determine the size of the sliding window
  • Step S402 Use the sliding window to extract input samples from the historical traffic sequence, and the data volume of the input samples is consistent with the size of the sliding window;
  • Step S403 Extract output samples from the historical traffic sequence according to the extracted input samples, and use the output samples as labels corresponding to the input samples;
  • Step S404 Perform training based on the input sample and the label to obtain the preset spatiotemporal network model.
  • this embodiment uses a sliding window method to construct training samples. More specifically, the process includes: setting the size of the sliding window to N, that is, the size of the sliding window is N; determining the starting time; and starting from the historical traffic Extract the historical traffic data corresponding to the starting time from the sequence, and extract the historical traffic data corresponding to (N-1) moments before the starting time, and combine the historical traffic data corresponding to the starting time and all the historical traffic data corresponding to the starting time.
  • the historical traffic data corresponding to (N-1) moments before the starting time is used as an input sample; the historical traffic data corresponding to the next time after the starting time is extracted from the historical traffic sequence as an output sample.
  • N the historical traffic data corresponding to (N-1) moments before the starting time
  • X t ,X t-1 ,...,X t-N+1 represents the input sample
  • X t+1 represents the output sample, that is, the label.
  • the BP algorithm After constructing the input samples and output samples, the BP algorithm is used for training to learn the parameters in the network.
  • the BP algorithm Error Back Propagation
  • the BP algorithm consists of two processes: forward propagation of signals and back propagation of errors. It is expected to adaptively adjust the connection weights between neurons during the training process of the neural network. value to find the best mapping function between input and output, so that the objective function or loss function can be minimized to complete tasks such as classification and regression.
  • the process of using the spatiotemporal network model to predict the traffic in the target area includes:
  • Step S501 Obtain the first historical traffic data in the target area.
  • the traffic data includes resident traffic, incoming traffic and outgoing traffic.
  • Step S502 Enter the historical traffic data into a preset spatiotemporal network model as input data to obtain the target traffic data of the target area within the target time period; wherein the preset spatiotemporal network model includes a spatiotemporal feature layer,
  • the spatiotemporal feature layer obtains multiple convolution features of the historical traffic data through convolution kernels of multiple scales, and determines the spatiotemporal features of the historical traffic data from the multiple convolution features through a self-attention mechanism. , so that the preset spatio-temporal network model determines the target traffic data of the target area within the target time period based on the spatio-temporal characteristics.
  • the process of obtaining the first historical traffic data in the target area may include performing Grid division to determine multiple grid areas; for each grid area, obtain the first historical traffic data of the grid area; determine based on the first historical traffic data of each grid area
  • the first traffic characteristic data of the target area is input into the preset spatiotemporal network model as input data to obtain the prediction result, that is, the target traffic data of the target area within the target time period.
  • the predicted traffic data is spatio-temporal data, which has particularity.
  • the relationship between traffic changes in each area and historical traffic, as well as the relationship between traffic flow between different areas, is different. Therefore, If the spatio-temporal network model is trained using the historical traffic data of the target area as training data, the spatio-temporal network model is better at predicting traffic in the target area than in predicting traffic in other areas. If you need to predict traffic in other areas, you need the spatiotemporal network model for transfer learning.
  • FIG 6 schematically shows a schematic diagram of the main modules of the spatiotemporal data-based traffic determination device 600 according to the embodiment of the present invention.
  • the spatiotemporal data-based traffic determination device 600 includes:
  • the acquisition module 601 is used to acquire the first historical traffic data in the target area
  • the determination module 602 is used to input the historical traffic data as input data into a preset spatiotemporal network model to obtain the target traffic data of the target area within the target time period; wherein the preset spatiotemporal network model includes at least A spatio-temporal feature layer that obtains multiple convolution features of the historical traffic data through convolution kernels of multiple scales, and determines the history from the multiple convolution features through a self-attention mechanism
  • the spatio-temporal characteristics of the traffic data are used so that the preset spatio-temporal network model determines the target traffic data of the target area within the target time period based on the spatio-temporal characteristics.
  • the traffic determination device based on spatiotemporal data in the embodiment of the present invention analyzes the historical traffic data in the target area through the preset spatiotemporal network model, and determines the target traffic data of the target area within the target time period, wherein the preset
  • the spatiotemporal network model includes a spatiotemporal feature layer, which includes convolution kernels of multiple scales to provide receptive fields of different sizes for different areas, and uses an attention mechanism to dynamically select receptive fields for each area, which can accurately Determine the relationship between traffic changes in each area and historical traffic, and determine the relationship between traffic flows between different areas, so as to accurately predict the traffic in the target area.
  • the determination module is further configured to: when the preset spatio-temporal network model includes multiple adjacent spatio-temporal feature layers, convert the previous one of the multiple adjacent spatio-temporal feature layers into The output data of the spatio-temporal feature layer is input into the subsequent spatio-temporal feature layer to determine the spatio-temporal features of the output data.
  • the preset spatiotemporal network model further includes a first convolution layer; the determination module is further configured to: input the historical traffic data into the first convolution layer, obtain the High-dimensional features of historical traffic data, and input the high-dimensional features into the spatio-temporal feature layer.
  • the preset spatiotemporal network model also includes a second convolution layer; the determination module It is also used to: input the spatio-temporal features into the second convolution layer, and perform a downsampling operation on the spatio-temporal features to obtain target flow data of the target area within a target time period.
  • the device further includes a training module, configured to: divide the training area into grids and determine multiple grid areas; and for each grid area, obtain the location of the grid area.
  • second historical traffic data in multiple time periods splicing the second historical traffic data of the multiple grid areas in the same time period to obtain second traffic characteristic data in the same time period; according to the The second traffic characteristic data in multiple time periods is used to obtain a historical traffic sequence, and the historical traffic sequence is used as a training data set; the training data set is trained to obtain the preset spatiotemporal network model.
  • the training module is also used to: determine the size of the sliding window; use the sliding window to extract input samples from the historical traffic sequence, the data amount of the input sample is the same as the sliding window.
  • the training module is also used to determine the size of the sliding window to be N; determine a starting time; extract historical traffic data corresponding to the starting time from the historical traffic sequence, and Extract the historical traffic data corresponding to (N-1) moments before the starting time, and combine the historical traffic data corresponding to the starting time and the (N-1) moments before the starting time.
  • the historical traffic data is used as an input sample; the historical traffic data corresponding to the next moment after the starting time is extracted from the historical traffic sequence as an output sample.
  • the acquisition module is further configured to: rasterize the target area and determine a plurality of grid areas; for each grid area, obtain the information of the grid area.
  • first historical traffic data determining the first traffic characteristic data of the target area according to the first historical traffic data of each grid area; inputting the historical traffic data as input data into a preset spatiotemporal network model including : Enter the first traffic characteristic data of the target area as input data into the preset spatiotemporal network model.
  • Figure 7 schematically shows the system architecture of a traffic determination method based on spatiotemporal data suitable for an embodiment of the present invention.
  • the system architecture 800 suitable for the traffic determination method based on spatiotemporal data includes: terminal devices 801, 802, 803, a network 804 and a server 805.
  • Network 804 is a medium used to provide communication links between terminal devices 801, 802, 803 and server 805.
  • Network 804 may include various connection types, such as wired, wireless communication links, fiber optic cables, etc.
  • Terminal devices 801, 802, 803 interact with the server 805 through the network 804 to receive or send messages, etc.
  • Various communication client applications can be installed on the terminal devices 801, 802, and 803.
  • Server 805 may be a server that provides various services.
  • the server can analyze and process the received requests or messages, and feed back the results obtained after data processing to the terminal device.
  • the traffic determination method based on spatiotemporal data provided by the embodiment of the present invention can generally be executed by the server 805.
  • the traffic determination method based on spatiotemporal data provided by the embodiment of the present invention can also be executed by a server or server cluster that is different from the server 805 and can communicate with the terminal devices 801, 802, 803 and/or the server 805.
  • Figure 8 schematically shows a schematic diagram of an electronic device according to an embodiment of the present invention.
  • the electronic device 800 provided by the embodiment of the present invention includes a processor 801, a communication interface 802, a memory 803, and a communication bus 804.
  • the processor 801, the communication interface 802, and the memory 803 complete interactions with each other through the communication bus 804. communication;
  • the memory 803 is used to store at least one executable instruction;
  • the processor 801 is used to implement the above-mentioned flow determination method based on spatiotemporal data when executing the executable instruction stored in the memory.
  • the above executable instructions when implementing the above model iteration method, the above executable instructions cause the above processor to perform the following steps:
  • the historical traffic data is input into a preset spatiotemporal network model as input data to obtain the target traffic data of the target area within the target time period.
  • the preset spatiotemporal network model includes a spatiotemporal feature layer.
  • the spatiotemporal feature layer Multiple convolution features of the historical traffic data are obtained through convolution kernels of multiple scales, and the spatiotemporal features of the historical traffic data are determined from the multiple convolution features through a self-attention mechanism.
  • the above-mentioned memory 803 may be an electronic memory such as flash memory, EEPROM (electrically erasable programmable read-only memory), EPROM, hard disk or ROM.
  • the memory 803 has storage space for program codes for executing any method steps in the above methods.
  • the storage space for the program code may include individual program codes respectively used to implement each step in the above method.
  • These program codes can be read from or written into one or more computer program products.
  • These computer program products include program code carriers such as hard disks, compact disks (CDs), memory cards or floppy disks. Such computer program products are usually portable or fixed storage units.
  • the storage unit may have storage segments or storage spaces arranged similarly to the memory 803 in the above-mentioned electronic device.
  • the program code may, for example, be compressed in a suitable form.
  • the storage unit includes a program for performing the steps of the method according to an embodiment of the invention, ie code that can be read by, for example, a processor such as 801, which code, when run by the electronic device, results in The electronic device performs the various steps in the method described above.
  • An embodiment of the present invention also provides a computer-readable storage medium.
  • a computer program is stored on the computer-readable storage medium.
  • the computer program is executed by a processor, the above-mentioned flow determination method based on spatiotemporal data is implemented.
  • the computer-readable storage medium may be included in the equipment/device described in the above embodiments; it may also exist independently without being assembled into the equipment/device.
  • the above computer-readable storage medium carries one or more programs. When the above one or more programs are executed, the method according to the embodiment of the present invention is implemented.
  • the computer-readable storage medium may be a non-volatile computer-readable storage medium, which may include, but is not limited to, portable computer disks, hard disks, random access memory (RAM), and read-only memory (ROM). , erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the above.
  • a computer-readable storage medium may be any tangible medium that contains or stores a program for use by or in conjunction with an instruction execution system, apparatus, or device.
  • Embodiments of the invention may be implemented in all or part of the steps in hardware, or in software modules running on one or more processors, or in a combination thereof.
  • a microprocessor or a digital signal processor (DSP) may be used in practice to implement some or all functions of some or all components in an electronic device according to embodiments of the present invention.
  • Embodiments of the invention may also be implemented as apparatus or apparatus programs (eg, computer programs and computer program products) for performing part or all of the methods described herein.
  • Programs implementing embodiments of the invention may be stored on a computer-readable medium, or may be in the form of one or more signals. Such signals may be downloaded from an Internet website, or provided on a carrier signal, or in any other form.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Business, Economics & Management (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Economics (AREA)
  • Strategic Management (AREA)
  • Human Resources & Organizations (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Development Economics (AREA)
  • Evolutionary Biology (AREA)
  • Game Theory and Decision Science (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Probability & Statistics with Applications (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Marketing (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Traffic Control Systems (AREA)

Abstract

本发明实施例公开了一种基于时空数据的流量确定方法、装置、电子设备和计算机可读存储介质，涉及计算机技术领域。该实施例包括：获取目标区域内的第一历史流量数据；将所述历史流量数据作为输入数据输入预设的时空网络模型，得到所述目标区域在目标时间段内的目标流量数据，所述预设的时空网络模型包括时空特征层，所述时空特征层通过多种尺度的卷积核获得所述历史流量数据的多种卷积特征，并通过自注意力机制从所述多种卷积特征中确定所述历史流量数据的时空特征。该实施例能够准确确定各个区域的流量变化和历史流量之间的关系以及确定不同区域之间的流量流转之间的关系，从而准确预测目标区域的流量。

Description

一种基于时空数据的流量确定方法、装置、设备和介质
相关申请的交叉引用
本公开要求于2022年4月28日向中华人民共和国国家知识产权局提交的申请号为202210470739.3、名称为“一种基于时空数据的流量确定方法、装置、设备和介质”的发明专利申请的全部权益，并通过引用的方式将其全部内容并入本文。
技术领域
本申请涉及计算机技术领域，尤其涉及一种基于时空数据的流量确定方法、装置、设备和介质。
背景技术
近年来，深度学习在自然语言处理、计算机视觉方面取得了很大成功。这些成功一方面依赖于大数据和硬件算力的发展，另一方面也依赖于神经网络架构的发展。在具体的应用场景中，如目标检测等，很多具有针对性、定制化的特征提取网络模型被研发。但是，在时空领域具有针对性的网络模型很少，而现有的网络模型是借鉴其他领域的模型进行特征提取，这些模型往往忽略了时空数据的特性，因此效果差强人意。
发明内容
为了解决上述技术问题或者至少部分地解决上述技术问题，本发明实施例提供一种基于时空数据的流量确定方法、装置、设备和介质。
第一方面，本发明实施例提供了一种基于时空数据的流量确定方法，包括：获取目标区域内的第一历史流量数据；将所述历史流量数据作为输入数据输入预设的时空网络模型，得到所述目标区域在目标时间段内的目标流量数据；其中，所述预设的时空网络模型包括时空特征层，所述时空特征层通过多种尺度的卷积核获得所述历史流量数据的多种卷积特征，并通过自注意力机制从所述多种卷积特征中确定所述历史流量数据的时空特征，以使所述预设的时空网络模型根据所述时空特征，确定所述目标区域在目标时间段内的目标流量数据。
在可选地实施例中，所述方法还包括：在所述预设的时空网络模型包括多个相邻的时空特征层的情况下，将所述多个相邻的时空特征层中前一个时空特征层的输出数据输入后一个时空特征层，以确定所述输出数据的时空特征。
在可选地实施例中，所述方法还包括：在所述预设的时空网络模型还包括第一卷积层的情况下，将所述历史流量数据输入所述第一卷积层，获取所述历史流量数据的高维特征，并将所述高维特征输入所述时空特征层。
在可选地实施例中，所述方法还包括：在所述预设的时空网络模型还包括第二卷积层的情况下，将所述时空特征输入所述第二卷积层，对所述时空特征进行降采样操作，以获得所述目标区域在目标时间段内的目标流量数据。
在可选地实施例中，所述方法还包括根据如下步骤训练得到所述预设的时空网络模型：对训练区域进行格化划分，确定多个网格区域；针对每一所述网格区域，获取所述网格区域在多个时间段内的第二历史流量数据；将所述多个网格区域在同一时间段内的第二历史流量数据进行拼接，得到所述同一时间段的第二流量特征数据；根据所述多个时间段内的所述第二流量特征数据，得到历史流量序列，将所述历史流量序列作为训练数据集；对所述训练数据集进行训练，以获得所述预设的时空网络模型。
在可选地实施例中，对所述训练数据集进行训练，以获得所述预设的时空网络模型包括：确定滑动窗口的尺寸；利用所述滑动窗口从所述历史流量序列中抽取输入样本，所述输入样本的数据量与所述滑动窗口的尺寸一致；根据所抽取的输入样本，从所述历史流量序列中抽取输出样本，所述输出样本作为所述输入样本对应的标签；基于所述输入样本和所述标签进行训练，以获得所述预设的时空网络模型。
在可选地实施例中，确定滑动窗口的尺寸包括：确定所述滑动窗口的尺寸为N；
利用所述滑动窗口从所述历史流量序列中抽取输入样本包括：确定起始时刻；从所述历史流量序列中抽取所述起始时刻对应的历史流量数据，以及抽取所述起始时刻之前的(N-1)个时刻所对应的历史流量数据，将所述起始时刻对应的历史流量数据以及所述起始时刻之前的(N-1)个时刻所对应的历史流量数据，作为输入样本；
根据所抽取的输入样本，从所述历史流量序列中抽取输出样本包括：从所述历史流量序列中抽取所述起始时刻后的下一时刻对应的历史流量数据作为输出样本。
在可选地实施例中，获取目标区域内的第一历史流量数据包括：对所述目标区域进行栅格化划分，确定多个网格区域；针对每一所述网格区域，获取所述网格区域的第一历史流量数据；根据每一所述网格区域的第一历史流量数据，确定所述目标区域的第一流量特征数据；将所述历史流量数据作为输入数据输入预设的时空网络模型包括：将所述目标区域的第一流量特征数据作为输入数据输入预设的时空网络模型。
第二方面，本发明实施例提供了一种基于时空数据的流量确定装置，包括：获取模块，用于获取目标区域内的第一历史流量数据；确定模块，用于将所述历史流量数据作为输入数据输入预设的时空网络模型，得到所述目标区域在目标时间段内的目标流量数据；其中，所述预设的时空网络模型包括至少一个时空特征层，所述时空特征层通过多种尺度的卷积核获得所述历史流量数据的多种卷积特征，并通过自注意力机 制从所述多种卷积特征中确定所述历史流量数据的时空特征，以使所述预设的时空网络模型根据所述时空特征，确定所述目标区域在目标时间段内的目标流量数据。
在可选地实施例中，所述确定模块还用于：在所述预设的时空网络模型包括多个相邻的时空特征层的情况下，将所述多个相邻的时空特征层中前一个时空特征层的输出数据输入后一个时空特征层，以确定所述输出数据的时空特征。
在可选地实施例中，所述确定模块还用于：在所述预设的时空网络模型还包括第一卷积层的情况下，将所述历史流量数据输入所述第一卷积层，获取所述历史流量数据的高维特征，并将所述高维特征输入所述时空特征层。
在可选地实施例中，所述确定模块还用于：在所述预设的时空网络模型还包括第二卷积层的情况下，将所述时空特征输入所述第二卷积层，对所述时空特征进行降采样操作，以获得所述目标区域在目标时间段内的目标流量数据。
在可选地实施例中，所述装置还包括训练模块，用于：对训练区域进行格化划分，确定多个网格区域；针对每一所述网格区域，获取所述网格区域在多个时间段内的第二历史流量数据；将所述多个网格区域在同一时间段内的第二历史流量数据进行拼接，得到所述同一时间段的第二流量特征数据；根据所述多个时间段内的所述第二流量特征数据，得到历史流量序列，将所述历史流量序列作为训练数据集；对所述训练数据集进行训练，以获得所述预设的时空网络模型。
在可选地实施例中，所述训练模块还用于：确定滑动窗口的尺寸；利用所述滑动窗口从所述历史流量序列中抽取输入样本，所述输入样本的数据量与所述滑动窗口的尺寸一致；根据所抽取的输入样本，从所述历史流量序列中抽取输出样本，所述输出样本作为所述输入样本对应的标签；基于所述输入样本和所述标签进行训练，以获得所述预设的时空网络模型。
在可选地实施例中，所述训练模块还用于确定所述滑动窗口的尺寸为N；确定起始时刻；从所述历史流量序列中抽取所述起始时刻对应的历史流量数据，以及抽取所述起始时刻之前的(N-1)个时刻所对应的历史流量数据，将所述起始时刻对应的历史流量数据以及所述起始时刻之前的(N-1)个时刻所对应的历史流量数据，作为输入样本；从所述历史流量序列中抽取所述起始时刻后的下一时刻对应的历史流量数据作为输出样本。
在可选地实施例中，所述获取模块还用于：对所述目标区域进行栅格化划分，确定多个网格区域；针对每一所述网格区域，获取所述网格区域的第一历史流量数据；根据每一所述网格区域的第一历史流量数据，确定所述目标区域的第一流量特征数据； 将所述历史流量数据作为输入数据输入预设的时空网络模型包括：将所述目标区域的第一流量特征数据作为输入数据输入预设的时空网络模型。
第三方面，本发明实施例提供了一种电子设备，包括处理器、通信接口、存储器和通信总线，其中，处理器、通信接口和存储器通过通信总线完成相互间的通信；所述存储器用于存放至少一可执行指令，所述可执行指令使得所述处理器执行以下步骤：获取目标区域内的第一历史流量数据；将所述历史流量数据作为输入数据输入预设的时空网络模型，得到所述目标区域在目标时间段内的目标流量数据；其中，所述预设的时空网络模型包括至少一个时空特征层，所述时空特征层通过多种尺度的卷积核获得所述历史流量数据的多种卷积特征，并通过自注意力机制从所述多种卷积特征中确定所述历史流量数据的时空特征，以使所述预设的时空网络模型根据所述时空特征，确定所述目标区域在目标时间段内的目标流量数据。
第四方面，本发明实施例提供了一种计算机可读存储介质，其上存储有计算机程序，所述计算机程序被处理器执行时实现本发明实施例的基于时空数据的流量确定方法。
上述实施例中的一个或多个技术方案至少具有如下优点的部分或全部：
通过预设的时空网络模型对目标区域内的历史流量数据进行分析，确定该目标区域在目标时间段内的目标流量数据，其中该预设的时空网络模型包括时空特征层，该时空特征层包括多种尺度的卷积核，以此为不同的区域提供大小不同的感受野，并使用注意力机制动态为每个区域选择感受野，能够准确确定各个区域的流量变化和历史流量之间的关系以及确定不同区域之间的流量流转之间的关系，从而准确预测目标区域的流量。本发明实施例的基于时空数据的流量确定方法，可以对城市级的流量进行监控预测和管控，提供更加精准的客流量预测，从而实现更加精细的管控方案，实现更精细的城市规划；该方法可以应用于城市中密集人流量区域的发现和广告投放、动态广告位定价等场景中。
附图说明
此处的附图被并入说明书中并构成本说明书的一部分，示出了符合本发明的实施例，并与说明书一起用于解释本发明的原理。
为了更清楚地说明本发明实施例或现有技术中的技术方案，下面将对实施例或相关技术描述中所需要使用的附图作简单地介绍，显而易见地，对于本领域普通技术人员而言，在不付出创造性劳动性的前提下，还可以根据这些附图获得其他的附图。
图1示意性地示出了本发明实施例的时空网络模型的时空特征层的示意图；
图2示意性地示出了本发明实施例时空网络模型的结构示意图；
图3示意性地示出了本发明实施例的训练时空网络模型的方法的主要步骤的示意图；
图4示意性地示出了本发明实施例训练时空网络模型的方法的子步骤的流程图；
图5示意性地示出了本发明实施例的基于时空数据的流量确定方法的流程图；
图6示意性地示出了本发明实施例的基于时空数据的流量确定装置的结构示意图；
图7示意性地示出了适用于本发明实施例的基于时空数据的流量确定方法的系统架构；
图8示意性示出了本发明实施例提供的电子设备的结构框图。
具体实施方式
为使本发明实施例的目的、技术方案和优点更加清楚，下面将结合本发明实施例中的附图，对本发明实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例是本发明的一部分实施例，而不是全部的实施例。基于本发明中的实施例，本领域普通技术人员在没有做出创造性劳动的前提下所获得的所有其他实施例，都属于本发明保护的范围。
在时空预测领域，目前可以使用神经网络进行时空数据建模，例如可以在ST-ResNet引入残差神经网络，通过残差神经网络进行时空特征提取。其中，时空特征包括了时序依赖和空间依赖两个维度的特征，时间依赖是指每个区域的流量变化和历史流量之间的关系，空间依赖指不同区域之间的流量流转之间的关系。在残差神经网络中，使用固定大小的感受野，如3*3，5*5等，并且进行了空间位置的参数共享，具有空间不变性。然而，在时空预测领域，应用场景并不满足这种空间不变性的关系。例如，在预测流量的场景中，流量的变化和流量流转的变化会受到居民活动的影响，如受到居民日常出行的影响，从而不同的城市区域的流量和周围区域的关系是不同的，不具备空间不变性。另外，使用固定大小的感受野也没有体现不同区域的差异性。如图1所示，对某市的流量进行分析，选取一段时间计算不同区域流量的皮尔逊相关系数，并选择该市的两个区域和其他区域进行对比，结果如图1所示。由该图可知，在城市范围内不同位置的流量具有较大的差异性，不同位置的区域和周围的区域之间的作用关系不同。因此，仅使用卷积神经网络并不能很好地建模时空数据。
为解决上述技术问题或者至少部分地解决上述技术问题，本发明实施例提供了一种基于时空数据的流量确定方法，该方法通过预设的时空网络模型对目标区域内的历史流量数据进行分析，确定该目标区域在目标时间段内的目标流量数据，其中该预设 的时空网络模型包括时空特征层，该时空特征层包括多种尺度的卷积核，以此为不同的区域提供大小不同的感受野，并使用注意力机制动态为每个区域选择感受野，能够准确确定各个区域的流量变化和历史流量之间的关系以及确定不同区域之间的流量流转之间的关系，从而准确预测目标区域的流量。
本发明实施例的时空网络模型包括时空特征层，该时空特征层主要为了解决不同位置的感受野不一致的问题。针对同一个区域，该时空特征层首先使用不同的卷积核进行时空特征的提取，然后对于这个区域提取的特征结果，使用一个注意力机制自动选择有效的特征，从而提取时空数据的时空特征，以此实现针对不同的区域使用不同的卷积核，即实现针对不同的区域使用不同的感受野进行时空特征的提取。其中，时空数据是同时具有时间和空间维度的数据，例如城市中的人流量数据、车流量数据等。为方便理解本发明实施例中的时空网络模型，下面以图1为例，对本发明的时空网络模型进行说明。
图1示意性地示出了本发明实施例的时空网络模型的时空特征层的结构示意图。在图1中，H，W表示某区域进行栅格化划分后经纬度方向的网格的数量。c，c’表示该区域的流量的通道数，如流入量、流出量、驻留量等。K1，K2分别表示不同的卷积核的大小，例如3*3，5*5，7*7等。如图1所示，本实施例针对原始范围内的两个不同区域，使用两个不同大小的卷积核进行时空特征的提取，而卷积核的大小是通过一个自注意力机制来动态选择的。
结合图1，本发明实施例的时空网络模型的时空特征层的实现过程如下：
首先，本实施例设置了一个多种尺度的卷积核集合K＝{3*3，5*5，7*7，9*9，11*11}，然后，对于原始区域的输入特征图X，在每个位置(xi，yj)，分别使用不同的卷积核进行特征图计算：
f3|＝Conv3(X)；
f5|＝Conv5(X)；
f7|＝Conv7(X)；
f9|＝Conv9(X)；
f11|＝Conv11(X)。
因此，对于给定位置(xi，yj)，获得了不同尺度的卷积核作用下的卷积特征
在获得多种卷积特征之后，该时空特征层使用自注意力机制从该多种卷积特征中选择有效的特征，具体过程如下：



其中，Wk是一个可学习的参数矩阵，用于增加时空网络模型的容量，可以建模复杂的关系，k1,k2是卷积核的大小。是通过映射后的特征，是计算了两个不同的卷积核输出特征的相关性，softmax()是归一化函数，是归一化后的权重，si,j是通过不同卷积核融合的特征，即时空特征。
综上，本发明实施例的时空网络模型针对同一个区域，该时空特征层首先使用不同的卷积核进行时空特征的提取，然后对于这个区域提取的特征结果，使用一个注意力机制自动选择有效的特征，从而提取时空数据的时空特征，以此实现针对不同的区域使用不同的卷积核，即实现针对不同的区域使用不同的感受野进行时空特征的提取，体现了不同区域的差异性，充分考虑了时空数据的特性。
在可选地实施例中，本发明实施例的时空网络模型可以包括一个时空特征层，也可以包括多个相邻的时空特征层，该多个相邻的时空特征层中，前一个时空特征层的输出数据为下一个时空特征层的输入数据。在本实施例中，为了充分考虑到原始区域中相距较远的子区域之间的关系，可以使用多个相邻的时空特征层。
在可选地实施例中，本发明实施例的时空网络模型除了时空特征层，还可以包括第一卷积层和第二卷积层。如图2所示，该时空网络模型包括第一卷积层、两个时空特征层以及第二卷积层。其中，第一卷积层用于进行高维特征的提取，即提取原始的输入数据的高维特征，然后时空特征层提取时空特征，第二卷积层对该时空特征进行降维操作，获得要预测的结果。本实施例通过第一卷积层和第二卷积层，在提高模型准确性的同时降低了计算量，提升了效率。
为了得到上述时空网络模型，需要获取训练样本进行训练。作为示例，该时空网络模型的训练过程如图3所示，该过程包括：
步骤S301：对训练区域进行栅格化划分，确定多个网格区域；
步骤S302：针对每一所述网格区域，获取所述网格区域在多个时间段内的第二历史流量数据；
步骤S303：将所述多个网格区域在同一时间段内的第二历史流量数据进行拼接，得到所述同一时间段的第二流量特征数据；
步骤S304：根据所述多个时间段内的所述第二流量特征数据，得到历史流量序列，将所述历史流量序列作为训练数据集；
步骤S305：对所述训练数据集进行训练，以获得所述预设的时空网络模型。
在训练时空网络模型的过程中，首先需要获取训练样本数据。在本实施例中，对训练区域进行栅格化划分，获得多个网格区域。其中，训练区域可以是城市，也可以是城市中的某个区域。更进一步的，该训练区域可以是目标区域，也可以包括目标区域，还可以包含在目标区域内(即训练区域在目标区域内)。在训练区域与目标区域的范围相同时，训练得到的时空网络模型在预测目标区域的流量时结果最准确。在获得多个网格区域之后，统计每个网格区域的第二历史流量数据x∈Rn×m×f，m和n分别表示网格区域的长度和宽度，f表示每个网格区域的流量特征，例如包括网格区域的驻留流量、流入流量和流出流量。在本实施例中，可以将以30分钟为时间步长统计网格区域的流量特征。对于给定的网格区域Dij，统计时间片段t(半个小时内)的流入流量：t-1时刻不在网格区域Dij中，t时刻在网格区域Dij中。统计时间片段t(半个小时内)的流出流量：t-1时刻在网格区域Dij中，t时刻不在网格区域Dij中。统计时间片段t(半个小时内)的驻留流量：t-1时刻在网格区域Dij中，t时刻也在网格区域Dij中。在确定各个网格区域的第二历史流量数据之后，将多个网格区域在同一时间段内的第二历史流量数据进行拼接，得到所述同一时间段的第二流量特征数据(该流量特征数据包含时间依赖和空间依赖)，例如将多个网格区域在t时刻的第二历史流量数据拼接，得到Xt，将t+1时刻的第二历史流量数据拼接得到Xt+1，从而得到历史流量序列{X1，X2，…，Xt，Xt+1}。然后，根据该历史流量序列作为训练数据集进行训练，得到该时空网络模型。
具体的，如图4所示，根据该历史流量序列作为训练数据集进行训练，得到该时空网络模型的过程包括：
步骤S401：确定滑动窗口的尺寸；
步骤S402：利用所述滑动窗口从所述历史流量序列中抽取输入样本，所述输入样本的数据量与所述滑动窗口的尺寸一致；
步骤S403：根据所抽取的输入样本，从所述历史流量序列中抽取输出样本，所述输出样本作为所述输入样本对应的标签；
步骤S404：基于所述输入样本和所述标签进行训练，以获得所述预设的时空网络模型。
在获得历史流量序列之后，基于该历史流量序列构造时空网络模型的输入样本和 输出样本。具体地，本实施例采用滑动窗口的方式构造训练样本，更具体地，该过程包括：设定滑动窗口的尺寸为N，即滑动窗口的大小为N；确定起始时刻；从所述历史流量序列中抽取所述起始时刻对应的历史流量数据，以及抽取所述起始时刻之前的(N-1)个时刻所对应的历史流量数据，将所述起始时刻对应的历史流量数据以及所述起始时刻之前的(N-1)个时刻所对应的历史流量数据，作为输入样本；从所述历史流量序列中抽取所述起始时刻后的下一时刻对应的历史流量数据作为输出样本。例如，在时刻t抽取历史长度为N的数据作为训练数据，并进行未来一步的预测，于是有：
其中，Xt,Xt-1,…,Xt-N+1表示输入样本，Xt+1表示输出样本，即标签。
在构造输入样本和输出样本之后，通过BP算法进行训练，学习网络中的参数。BP算法(Error Back Propagation，误差后向传播算法)由信号的正向传播和误差的反向传播两个过程组成，期望通过在神经网络的训练过程中自适应的调整各神经元间的连接权值，以寻求最佳的输入输出间的映射函数，使得目标函数或损失函数达到最小，完成分类、回归等任务。
在获得上述时空网络模型之后，可以利用该时空网络模型预测目标区域在目标时间段内的目标流量数据。例如使用滑动窗口，提出最后一段时间的历史流量特征：Input＝(XL,XL-1,…,XL-N+1)，然后利用训练得到的时空网络模型进行预测，获得时空网络模型预测目标区域在目标时间段内的目标流量数据。在可选地实施例中，在对目标区域的流量数据进行预测时，输入时空网络模型的输入数据也可以不是从第二历史流量数据中获取的。
具体的，如图5所示，利用该时空网络模型对目标区域的流量进行预测的过程包括：
步骤S501：获取目标区域内的第一历史流量数据。其中，流量数据包括驻留流量、流入流量和流出流量。
步骤S502：将所述历史流量数据作为输入数据输入预设的时空网络模型，得到所述目标区域在目标时间段内的目标流量数据；其中，所述预设的时空网络模型包括时空特征层，所述时空特征层通过多种尺度的卷积核获得所述历史流量数据的多种卷积特征，并通过自注意力机制从所述多种卷积特征中确定所述历史流量数据的时空特征，以使所述预设的时空网络模型根据所述时空特征，确定所述目标区域在目标时间段内的目标流量数据。
其中，获取目标区域内的第一历史流量数据的过程可以包括对所述目标区域进行 栅格化划分，确定多个网格区域；针对每一所述网格区域，获取所述网格区域的第一历史流量数据；根据每一所述网格区域的第一历史流量数据，确定所述目标区域的第一流量特征数据，将所述目标区域的第一流量特征数据作为输入数据输入预设的时空网络模型，获得预测结果，即目标区域在目标时间段内的目标流量数据。
在本实施例的应用场景中，所预测的流量数据为时空数据，具有特殊性，每个区域的流量变化和历史流量之间的关系，以及不同区域之间的流量流转的关系不同，因此，若时空网络模型是以目标区域的历史流量数据为训练数据训练得到的，则该时空网络模型预测目标区域的流量的效果优于预测其他区域的流量。若需要预测其他区域的流量则需要该时空网络模型进行迁移学习。
图6示意性地示出了本发明实施例的基于时空数据的流量确定装置600的主要模块的示意图，如图6所示，该基于时空数据的流量确定装置600包括：
获取模块601，用于获取目标区域内的第一历史流量数据；
确定模块602，用于将所述历史流量数据作为输入数据输入预设的时空网络模型，得到所述目标区域在目标时间段内的目标流量数据；其中，所述预设的时空网络模型包括至少一个时空特征层，所述时空特征层通过多种尺度的卷积核获得所述历史流量数据的多种卷积特征，并通过自注意力机制从所述多种卷积特征中确定所述历史流量数据的时空特征，以使所述预设的时空网络模型根据所述时空特征，确定所述目标区域在目标时间段内的目标流量数据。
本发明实施例的基于时空数据的流量确定装置，通过预设的时空网络模型对目标区域内的历史流量数据进行分析，确定该目标区域在目标时间段内的目标流量数据，其中该预设的时空网络模型包括时空特征层，该时空特征层包括多种尺度的卷积核，以此为不同的区域提供大小不同的感受野，并使用注意力机制动态为每个区域选择感受野，能够准确确定各个区域的流量变化和历史流量之间的关系以及确定不同区域之间的流量流转之间的关系，从而准确预测目标区域的流量。
在可选地实施例中，所述确定模块还用于：当所述预设的时空网络模型包括多个相邻的时空特征层时，将所述多个相邻的时空特征层中前一个时空特征层的输出数据输入后一个时空特征层，以确定所述输出数据的时空特征。
在可选地实施例中，所述预设的时空网络模型还包括第一卷积层；所述确定模块还用于：将所述历史流量数据输入所述第一卷积层，获取所述历史流量数据的高维特征，并将所述高维特征输入所述时空特征层。
在可选地实施例中，所述预设的时空网络模型还包括第二卷积层；所述确定模块 还用于：将所述时空特征输入所述第二卷积层，对所述时空特征进行降采样操作，以获得所述目标区域在目标时间段内的目标流量数据。
在可选地实施例中，所述装置还包括训练模块，用于：对训练区域进行格化划分，确定多个网格区域；针对每一所述网格区域，获取所述网格区域在多个时间段内的第二历史流量数据；将所述多个网格区域在同一时间段内的第二历史流量数据进行拼接，得到所述同一时间段的第二流量特征数据；根据所述多个时间段内的所述第二流量特征数据，得到历史流量序列，将所述历史流量序列作为训练数据集；对所述训练数据集进行训练，以获得所述预设的时空网络模型。
在可选地实施例中，所述训练模块还用于：确定滑动窗口的尺寸；利用所述滑动窗口从所述历史流量序列中抽取输入样本，所述输入样本的数据量与所述滑动窗口的尺寸一致；根据所抽取的输入样本，从所述历史流量序列中抽取输出样本，所述输出样本作为所述输入样本对应的标签；基于所述输入样本和所述标签进行训练，以获得所述预设的时空网络模型。
在可选地实施例中，所述训练模块还用于确定所述滑动窗口的尺寸为N；确定起始时刻；从所述历史流量序列中抽取所述起始时刻对应的历史流量数据，以及抽取所述起始时刻之前的(N-1)个时刻所对应的历史流量数据，将所述起始时刻对应的历史流量数据以及所述起始时刻之前的(N-1)个时刻所对应的历史流量数据，作为输入样本；从所述历史流量序列中抽取所述起始时刻后的下一时刻对应的历史流量数据作为输出样本。
在可选地实施例中，所述获取模块还用于：对所述目标区域进行栅格化划分，确定多个网格区域；针对每一所述网格区域，获取所述网格区域的第一历史流量数据；根据每一所述网格区域的第一历史流量数据，确定所述目标区域的第一流量特征数据；将所述历史流量数据作为输入数据输入预设的时空网络模型包括：将所述目标区域的第一流量特征数据作为输入数据输入预设的时空网络模型。
图7示意性地示出了适用于本发明实施例的基于时空数据的流量确定方法的系统架构。
如图7所示，适用于本发明实施例的基于时空数据的流量确定方法的系统架构800包括：终端设备801、802、803，网络804和服务器805。网络804用以在终端设备801、802、803和服务器805之间提供通信链路的介质。网络804可以包括各种连接类型，例如有线、无线通信链路或者光纤电缆等。
终端设备801、802、803通过网络804与服务器805交互，以接收或发送消息等。 终端设备801、802、803上可以安装有各种通讯客户端应用。
服务器805可以是提供各种服务的服务器。服务器可以对接收到的请求或消息进行分析和处理，并将数据处理后得到的结果反馈给终端设备。
需要说明的是，本发明实施例所提供的基于时空数据的流量确定方法一般可以由服务器805执行。本发明实施例所提供的基于时空数据的流量确定方法也可以由不同于服务器805且能够与终端设备801、802、803和/或服务器805通信的服务器或服务器集群执行。
应该理解的是，图7中的终端设备、网络和服务器的数目仅仅是示意性的。根据实现需要，可以具有任意数目的终端设备、网络和服务器。
图8示意性示出了本发明一实施例的电子设备的示意图。如图8所示，本发明实施例提供的电子设备800包括处理器801、通信接口802、存储器803和通信总线804，其中，处理器801、通信接口802和存储器803通过通信总线804完成相互间的通信；存储器803，用于存放至少一可执行指令；处理器801，用于执行存储器上所存放的可执行指令时，实现如上所述的基于时空数据的流量确定方法。
具体而言，当实现上述模型迭代方法时，上述可执行指令使得上述处理器执行以下步骤：
获取目标区域内的第一历史流量数据；
将所述历史流量数据作为输入数据输入预设的时空网络模型，得到所述目标区域在目标时间段内的目标流量数据，所述预设的时空网络模型包括时空特征层，所述时空特征层通过多种尺度的卷积核获得所述历史流量数据的多种卷积特征，并通过自注意力机制从所述多种卷积特征中确定所述历史流量数据的时空特征。
上述存储器803可以是诸如闪存、EEPROM(电可擦除可编程只读存储器)、EPROM、硬盘或者ROM之类的电子存储器。存储器803具有用于执行上述方法中的任何方法步骤的程序代码的存储空间。例如，用于程序代码的存储空间可以包括分别用于实现上面的方法中的各个步骤的各个程序代码。这些程序代码可以从一个或者多个计算机程序产品中读出或者写入到这一个或者多个计算机程序产品中。这些计算机程序产品包括诸如硬盘，光盘(CD)、存储卡或者软盘之类的程序代码载体。这样的计算机程序产品通常为便携式或者固定存储单元。该存储单元可以具有与上述电子设备中的存储器803类似布置的存储段或者存储空间等。程序代码可以例如以适当形式进行压缩。通常，存储单元包括用于执行根据本发明的实施例的方法步骤的程序，即可以由例如诸如801之类的处理器读取的代码，这些代码当由电子设备运行时，导致 该电子设备执行上面所描述的方法中的各个步骤。
本发明实施例还提供了一种计算机可读存储介质。上述计算机可读存储介质上存储有计算机程序，上述计算机程序被处理器执行时实现如上所述的基于时空数据的流量确定方法。
该计算机可读存储介质可以是上述实施例中描述的设备/装置中所包含的；也可以是单独存在，而未装配入该设备/装置中。上述计算机可读存储介质承载有一个或者多个程序，当上述一个或者多个程序被执行时，实现根据本发明实施例的方法。
根据本发明的实施例，计算机可读存储介质可以是非易失性的计算机可读存储介质，例如可以包括但不限于：便携式计算机磁盘、硬盘、随机访问存储器(RAM)、只读存储器(ROM)、可擦式可编程只读存储器(EPROM或闪存)、便携式紧凑磁盘只读存储器(CD-ROM)、光存储器件、磁存储器件、或者上述的任意合适的组合。在本发明中，计算机可读存储介质可以是任何包含或存储程序的有形介质，该程序可以被指令执行系统、装置或者器件使用或者与其结合使用。
本发明的实施例提供的上述各个技术方案可以全部或部分步骤以硬件实现，或者以在一个或者多个处理器上运行的软件模块实现，或者以它们的组合实现。本领域的技术人员应当理解，可以在实践中使用微处理器或者数字信号处理器(DSP)来实现根据本发明的实施例的电子设备中的一些或者全部部件的一些或者全部功能。本发明的实施例还可以实现为用于执行这里所描述的方法的一部分或者全部的设备或者装置程序(例如，计算机程序和计算机程序产品)。实现本发明的实施例的程序可以存储在计算机可读介质上，或者可以具有一个或者多个信号的形式。这样的信号可以从因特网网站上下载得到，或者在载体信号上提供，或者以任何其他形式提供。
需要说明的是，在本文中，诸如“第一”和“第二”等之类的关系术语仅仅用来将一个实体或者步骤与另一个实体或步骤区分开来，而不一定要求或者暗示这些实体或步骤之间存在任何这种实际的关系或者顺序。而且，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、物品或者设备不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、物品或者设备所固有的要素。在没有更多限制的情况下，由语句“包括一……”限定的要素，并不排除在包括所述要素的过程、方法、物品或者设备中还存在另外的相同要素。
以上所述仅是本发明的具体实施方式，使本领域技术人员能够理解或实现本发明。对这些实施例的多种修改对本领域的技术人员来说将是显而易见的，本文中所定义的 一般原理可以在不脱离本发明的精神或范围的情况下，在其它实施例中实现。因此，本发明将不会被限制于本文所示的这些实施例，而是要符合与本文所申请的原理和新颖特点相一致的最宽的范围。

Claims (11)

  1. 一种基于时空数据的流量确定方法，其特征在于，包括：
    获取目标区域内的第一历史流量数据；
    将所述历史流量数据作为输入数据输入预设的时空网络模型，得到所述目标区域在目标时间段内的目标流量数据；其中，所述预设的时空网络模型包括时空特征层，所述时空特征层通过多种尺度的卷积核获得所述历史流量数据的多种卷积特征，并通过自注意力机制从所述多种卷积特征中确定所述历史流量数据的时空特征，以使所述预设的时空网络模型根据所述时空特征，确定所述目标区域在目标时间段内的目标流量数据。
  2. 根据权利要求1所述的方法，其特征在于，所述方法还包括：在所述预设的时空网络模型包括多个相邻的时空特征层的情况下，将所述多个相邻的时空特征层中前一个时空特征层的输出数据输入后一个时空特征层，以确定所述输出数据的时空特征。
  3. 根据权利要求1所述的方法，其特征在于，所述方法还包括：在所述预设的时空网络模型还包括第一卷积层的情况下，将所述历史流量数据输入所述第一卷积层，获取所述历史流量数据的高维特征，并将所述高维特征输入所述时空特征层。
  4. 根据权利要求3所述的方法，其特征在于，所述方法还包括：在所述预设的时空网络模型还包括第二卷积层的情况下，将所述时空特征输入所述第二卷积层，对所述时空特征进行降采样操作，以获得所述目标区域在目标时间段内的目标流量数据。
  5. 根据权利要求1所述的方法，其特征在于，所述方法还包括根据如下步骤训练得到所述预设的时空网络模型：
    对训练区域进行格化划分，确定多个网格区域；
    针对每一所述网格区域，获取所述网格区域在多个时间段内的第二历史流量数据；
    将所述多个网格区域在同一时间段内的第二历史流量数据进行拼接，得到所述同一时间段的第二流量特征数据；
    根据所述多个时间段内的所述第二流量特征数据，得到历史流量序列，将所述历史流量序列作为训练数据集；
    对所述训练数据集进行训练，以获得所述预设的时空网络模型。
  6. 根据权利要求5所述的方法，其特征在于，对所述训练数据集进行训练，以获得所述预设的时空网络模型包括：
    确定滑动窗口的尺寸；
    利用所述滑动窗口从所述历史流量序列中抽取输入样本，所述输入样本的数据量与所述滑动窗口的尺寸一致；
    根据所抽取的输入样本，从所述历史流量序列中抽取输出样本，所述输出样本作为所述输入样本对应的标签；
    基于所述输入样本和所述标签进行训练，以获得所述预设的时空网络模型。
  7. 根据权利要求6所述的方法，其特征在于，确定滑动窗口的尺寸包括：确定所述滑动窗口的尺寸为N；
    利用所述滑动窗口从所述历史流量序列中抽取输入样本包括：确定起始时刻；从所述历史流量序列中抽取所述起始时刻对应的历史流量数据，以及抽取所述起始时刻之前的(N-1)个时刻所对应的历史流量数据，将所述起始时刻对应的历史流量数据以及所述起始时刻之前的(N-1)个时刻所对应的历史流量数据，作为输入样本；
    根据所抽取的输入样本，从所述历史流量序列中抽取输出样本包括：从所述历史流量序列中抽取所述起始时刻后的下一时刻对应的历史流量数据作为输出样本。
  8. 根据权利要求7所述的方法，其特征在于，获取目标区域内的第一历史流量数据包括：
    对所述目标区域进行栅格化划分，确定多个网格区域；
    针对每一所述网格区域，获取所述网格区域的第一历史流量数据；
    根据每一所述网格区域的第一历史流量数据，确定所述目标区域的第一流量特征数据；
    将所述历史流量数据作为输入数据输入预设的时空网络模型包括：将所述目标区域的第一流量特征数据作为输入数据输入预设的时空网络模型。
  9. 一种基于时空数据的流量确定装置，其特征在于，包括：
    获取模块，用于获取目标区域内的第一历史流量数据；
    确定模块，用于将所述历史流量数据作为输入数据输入预设的时空网络模型，得到所述目标区域在目标时间段内的目标流量数据；其中，所述预设的时空网络模型包括至少一个时空特征层，所述时空特征层通过多种尺度的卷积核获得所述历史流量数据的多种卷积特征，并通过自注意力机制从所述多种卷积特征中确定所述历史流量数据的时空特征，以使所述预设的时空网络模型根据所述时空特征，确定所述目标区域在目标时间段内的目标流量数据。
  10. 一种电子设备，其特征在于，包括处理器、通信接口、存储器和通信总线，其中，处理器、通信接口和存储器通过通信总线完成相互间的通信；
    所述存储器用于存放至少一可执行指令，所述可执行指令使得所述处理器执行以下步骤：
    获取目标区域内的第一历史流量数据；
    将所述历史流量数据作为输入数据输入预设的时空网络模型，得到所述目标区域在目标时间段内的目标流量数据；其中，所述预设的时空网络模型包括至少一个时空特征层，所述时空特征层通过多种尺度的卷积核获得所述历史流量数据的多种卷积特征，并通过自注意力机制从所述多种卷积特征中确定所述历史流量数据的时空特征，以使所述预设的时空网络模型根据所述时空特征，确定所述目标区域在目标时间段内的目标流量数据。
  11. 一种计算机可读存储介质，其上存储有计算机程序，其特征在于，所述计算机程序被处理器执行时实现权利要求1-8中任一项所述的方法。
PCT/CN2023/082204 2022-04-28 2023-03-17 一种基于时空数据的流量确定方法、装置、设备和介质 WO2023207411A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210470739.3 2022-04-28
CN202210470739.3A CN114862010A (zh) 2022-04-28 2022-04-28 一种基于时空数据的流量确定方法、装置、设备和介质

Publications (1)

Publication Number Publication Date
WO2023207411A1 true WO2023207411A1 (zh) 2023-11-02

Family

ID=82634920

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/082204 WO2023207411A1 (zh) 2022-04-28 2023-03-17 一种基于时空数据的流量确定方法、装置、设备和介质

Country Status (2)

Country Link
CN (1) CN114862010A (zh)
WO (1) WO2023207411A1 (zh)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114862010A (zh) * 2022-04-28 2022-08-05 京东城市(北京)数字科技有限公司 一种基于时空数据的流量确定方法、装置、设备和介质
CN115515184B (zh) * 2022-11-04 2023-04-07 北京众森信和科技有限公司 一种业务数据的发送方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112910711A (zh) * 2021-02-03 2021-06-04 山东大学 一种基于自注意力卷积网络的无线业务流量预测方法、设备及介质
US20210209938A1 (en) * 2020-09-25 2021-07-08 Beijing Baidu Netcom Science And Technology Co., Ltd. Method, apparatus, system, and computer-readable medium for traffic pattern prediction
CN113743505A (zh) * 2021-09-06 2021-12-03 辽宁工程技术大学 基于自注意力和特征融合的改进ssd目标检测方法
CN114039871A (zh) * 2021-10-25 2022-02-11 中山大学 一种蜂窝流量预测的方法、系统、装置及介质
CN114066052A (zh) * 2021-11-15 2022-02-18 国网江苏省电力有限公司营销服务中心 基于改进tcn修正累积误差的可调节负荷功率多步预测方法
CN114862010A (zh) * 2022-04-28 2022-08-05 京东城市(北京)数字科技有限公司 一种基于时空数据的流量确定方法、装置、设备和介质

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20210209938A1 (en) * 2020-09-25 2021-07-08 Beijing Baidu Netcom Science And Technology Co., Ltd. Method, apparatus, system, and computer-readable medium for traffic pattern prediction
CN112910711A (zh) * 2021-02-03 2021-06-04 山东大学 一种基于自注意力卷积网络的无线业务流量预测方法、设备及介质
CN113743505A (zh) * 2021-09-06 2021-12-03 辽宁工程技术大学 基于自注意力和特征融合的改进ssd目标检测方法
CN114039871A (zh) * 2021-10-25 2022-02-11 中山大学 一种蜂窝流量预测的方法、系统、装置及介质
CN114066052A (zh) * 2021-11-15 2022-02-18 国网江苏省电力有限公司营销服务中心 基于改进tcn修正累积误差的可调节负荷功率多步预测方法
CN114862010A (zh) * 2022-04-28 2022-08-05 京东城市(北京)数字科技有限公司 一种基于时空数据的流量确定方法、装置、设备和介质

Also Published As

Publication number Publication date
CN114862010A (zh) 2022-08-05

Similar Documents

Publication Publication Date Title
US11586880B2 (en) System and method for multi-horizon time series forecasting with dynamic temporal context learning
WO2023207411A1 (zh) 一种基于时空数据的流量确定方法、装置、设备和介质
CN110728317A (zh) 决策树模型的训练方法、系统、存储介质及预测方法
WO2018156942A1 (en) Optimizing neural network architectures
CN111523640B (zh) 神经网络模型的训练方法和装置
WO2021094920A1 (en) Fusing multimodal data using recurrent neural networks
CN110674636B (zh) 一种用电行为分析方法
CN112163601A (zh) 图像分类方法、系统、计算机设备及存储介质
CN116684330A (zh) 基于人工智能的流量预测方法、装置、设备及存储介质
CN111191825A (zh) 用户违约预测方法、装置及电子设备
CN112055038B (zh) 生成点击率预估模型的方法及预测点击概率的方法
CN114417174A (zh) 内容推荐方法、装置、设备及计算机存储介质
US11496775B2 (en) Neural network model compression with selective structured weight unification
US20220292315A1 (en) Accelerated k-fold cross-validation
US11935271B2 (en) Neural network model compression with selective structured weight unification
CN113128773A (zh) 地址预测模型的训练方法、地址预测方法及装置
WO2023197857A1 (zh) 一种模型切分方法及其相关设备
CN113723712B (zh) 风电功率预测方法、系统、设备及介质
CN114241411B (zh) 基于目标检测的计数模型处理方法、装置及计算机设备
US20210312265A1 (en) Response Generation using Memory Augmented Deep Neural Networks
US20210201157A1 (en) Neural network model compression with quantizability regularization
JP2023535202A (ja) コンテキスト・アウェア・アノマリ検出
CN112231546A (zh) 异构文档的排序方法、异构文档排序模型训练方法及装置
CN115037655B (zh) 压测方法和系统
WO2024114659A1 (zh) 一种摘要生成方法及其相关设备

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 23794846

Country of ref document: EP

Kind code of ref document: A1