Space-time data prediction, weather data prediction and energy data prediction method

Abstract

The embodiment of the present specification provides a spatio-temporal data prediction method, a weather data prediction method and an energy data prediction method, wherein the spatio-temporal data prediction method comprises: obtaining spatio-temporal sampling data of a target task, wherein the spatio-temporal sampling data is spatially distributed time series data; performing frequency domain mapping on the spatio-temporal sampling data to obtain frequency domain data of the spatio-temporal sampling data; and using a pre-trained spatio-temporal prediction model to predict target spatio-temporal prediction data of the target task according to the frequency domain data, wherein the target spatio-temporal prediction data is spatially distributed time series data. By performing frequency domain mapping on the spatio-temporal sampling data of the target task to obtain frequency domain data representing global information, and then using the pre-trained spatio-temporal prediction model to predict the target spatio-temporal prediction data of the target task according to the frequency domain data representing global information, the accuracy of the predicted target spatio-temporal prediction data is improved.

Description

Technical Field

[0001] The embodiments in this specification relate to the field of data prediction technology, and in particular to a spatiotemporal data method. Background Technology

[0002] With the development of computer technology, in multiple task domains, spatiotemporal prediction data for a future period of time is predicted based on the spatiotemporal sampling data of the target task. This data is then used to formulate task plans, make task decisions, and set task plans. Accurate spatiotemporal prediction data can ensure the effectiveness and efficiency of the above-mentioned task processing and reduce task costs.

[0003] Currently, due to the increase in data volume and the increasing requirements for prediction accuracy, the traditional method of manually plotting points for prediction has been replaced by prediction methods based on neural network models, such as CNN (Convolutional Neural Networks), LSTM (Long Short-Term Memory), and Transformer. However, while these neural network-based prediction methods can predict detailed information, they suffer from insufficient ability to extract global information, resulting in inaccurate spatiotemporal prediction data. Therefore, a more accurate spatiotemporal data prediction method is urgently needed. Summary of the Invention

[0004] In view of the above, this specification provides a spatiotemporal data prediction method. One or more embodiments of this specification also relate to a weather data prediction method, an energy data prediction method, a spatiotemporal data prediction data processing method, a spatiotemporal data prediction data processing system, a spatiotemporal data prediction device, a weather data prediction device, an energy data prediction device, a spatiotemporal data prediction data processing device, a computing device, a computer-readable storage medium, and a computer program, to address the technical deficiencies existing in the prior art.

[0005] According to a first aspect of the embodiments of this specification, a spatiotemporal data prediction method is provided, comprising:

[0006] Acquire spatiotemporal sampling data for the target task, where the spatiotemporal sampling data is spatially distributed time-series data;

[0007] Frequency domain mapping is performed on the spatiotemporal sampling data to obtain the frequency domain data of the spatiotemporal sampling data;

[0008] Based on frequency domain data, a pre-trained spatiotemporal prediction model is used to predict the target spatiotemporal prediction data of the target task. The target spatiotemporal prediction data is spatially distributed time series data.

[0009] According to a second aspect of the embodiments of this specification, a weather data forecasting method is provided, comprising:

[0010] Acquire weather sampling data for weather forecasting tasks, where the weather sampling data is spatially distributed time-series weather data;

[0011] Frequency domain mapping is performed on the weather sampling data to obtain the frequency domain data of the weather sampling data;

[0012] Based on the frequency domain data of the weather sampling data, the weather forecast data for the weather forecasting task is obtained by using a pre-trained spatiotemporal prediction model. The weather forecast data is spatially distributed time-series weather data.

[0013] According to a third aspect of the embodiments of this specification, an energy data prediction method is provided, comprising:

[0014] Acquire energy generation data for energy generation tasks, where the energy generation data is spatially distributed time-series energy data;

[0015] Frequency domain mapping is performed on the energy generation data to obtain the frequency domain data of the energy generation data;

[0016] Based on the frequency domain data of energy generation data, a pre-trained spatiotemporal prediction model is used to predict the energy generation prediction data for the energy generation task. The energy generation prediction data is spatially distributed energy time series data.

[0017] According to a fourth aspect of the embodiments of this specification, a data processing method for spatiotemporal data prediction is provided, applied to a cloud-side device, comprising:

[0018] Obtain a sample set, which includes multiple sample groups, and each sample group includes frequency domain data of sample spatiotemporal data and label spatiotemporal data;

[0019] Extract the first sample group from the sample set, wherein the first sample group is any sample group, and the first sample group includes the frequency domain data of the first sample spatiotemporal data and the first label spatiotemporal data;

[0020] Based on the frequency domain data of the first sample spatiotemporal data, the spatiotemporal prediction data of the sample is predicted using the spatiotemporal prediction model.

[0021] Calculate the loss value based on the sample spatiotemporal prediction data and the label spatiotemporal data;

[0022] Based on the loss value, adjust the model parameters of the spatiotemporal prediction model, return to the step of extracting the first sample group from the sample set, and obtain the spatiotemporal prediction model after training is completed if the preset training termination conditions are met.

[0023] The trained spatiotemporal prediction model parameters are sent to the edge device.

[0024] According to a fifth aspect of the embodiments of this specification, a data processing system for spatiotemporal data prediction is provided, comprising:

[0025] End-side devices are used to send training requests for spatiotemporal prediction models to cloud-side devices;

[0026] The cloud-side device, upon receiving a training request, acquires a sample set, which includes multiple sample groups. Each sample group includes frequency domain data of sample spatiotemporal data and label spatiotemporal data. It then extracts a first sample group from the sample set, which is any sample group including frequency domain data of first sample spatiotemporal data and first label spatiotemporal data. Based on the frequency domain data of the first sample spatiotemporal data, it uses a spatiotemporal prediction model to predict the sample spatiotemporal prediction data. Based on the sample spatiotemporal prediction data and label spatiotemporal data, it calculates a loss value. Based on the loss value, it adjusts the model parameters of the spatiotemporal prediction model and returns to the step of extracting the first sample group from the sample set. If a preset training termination condition is met, the trained spatiotemporal prediction model is obtained. Finally, it sends the model parameters of the trained spatiotemporal prediction model to the edge device.

[0027] The end-side device is also used to receive model parameters of the spatiotemporal prediction model.

[0028] According to a sixth aspect of the embodiments of this specification, a spatiotemporal data prediction apparatus is provided, comprising:

[0029] The first acquisition module is configured to acquire spatiotemporal sampling data of the target task, wherein the spatiotemporal sampling data is spatially distributed time-series data;

[0030] The first mapping module is configured to perform frequency domain mapping on the spatiotemporal sampling data to obtain the frequency domain data of the spatiotemporal sampling data.

[0031] The first prediction module is configured to predict the target spatiotemporal prediction data of the target task based on the frequency domain data and using a pre-trained spatiotemporal prediction model, wherein the target spatiotemporal prediction data is spatially distributed time-series data.

[0032] According to a seventh aspect of the embodiments of this specification, a weather data forecasting apparatus is provided, comprising:

[0033] The second acquisition module is configured to acquire weather sampling data for the weather forecasting task, wherein the weather sampling data is spatially distributed time-series weather data;

[0034] The second mapping module is configured to perform frequency domain mapping on the weather sampling data to obtain the frequency domain data of the weather sampling data.

[0035] The second prediction module is configured to predict weather prediction data for the weather prediction task based on the frequency domain data of the weather sampling data and using a pre-trained spatiotemporal prediction model, wherein the weather prediction data is spatially distributed weather time series data.

[0036] According to an eighth aspect of the embodiments of this specification, an energy data prediction apparatus is provided, comprising:

[0037] The third acquisition module is configured to acquire energy generation data of the energy generation task, wherein the energy generation data is spatially distributed energy time-series data;

[0038] The third mapping module is configured to perform frequency domain mapping on the energy generation data to obtain the frequency domain data of the energy generation data.

[0039] The third prediction module is configured to predict energy generation prediction data for the energy generation task based on the frequency domain data of the energy generation data and using a pre-trained spatiotemporal prediction model, wherein the energy generation prediction data is spatially distributed energy time-series data.

[0040] According to a ninth aspect of the embodiments of this specification, a data processing apparatus for spatiotemporal data prediction is provided, applied to cloud-side equipment, comprising:

[0041] The fourth acquisition module is configured to acquire a sample set, wherein the sample set includes multiple sample groups, and any sample group includes frequency domain data of sample spatiotemporal data and label spatiotemporal data;

[0042] The extraction module is configured to extract a first sample group from the sample set, wherein the first sample group is any sample group, and the first sample group includes the frequency domain data of the first sample spatiotemporal data and the first label spatiotemporal data;

[0043] The fourth prediction module is configured to predict the spatiotemporal prediction data of the sample based on the frequency domain data of the first sample spatiotemporal data using a spatiotemporal prediction model.

[0044] The calculation module is configured to calculate the loss value based on the sample spatiotemporal prediction data and the label spatiotemporal data;

[0045] The training module is configured to adjust the model parameters of the spatiotemporal prediction model according to the loss value, return to execute the step of extracting the first sample group from the sample set, and obtain the trained spatiotemporal prediction model when the preset training termination condition is met.

[0046] The sending module is configured to send the model parameters of the trained spatiotemporal prediction model to the edge device.

[0047] According to a tenth aspect of the embodiments of this specification, a computing device is provided, comprising:

[0048] Memory and processor;

[0049] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the above-mentioned spatiotemporal data prediction method, weather data prediction method, energy data prediction method, or spatiotemporal data prediction data processing method.

[0050] According to an eleventh aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores computer-executable instructions, which, when executed by a processor, implement the steps of the above-described spatiotemporal data prediction method, weather data prediction method, energy data prediction method, or spatiotemporal data prediction data processing method.

[0051] According to a twelfth aspect of the embodiments of this specification, a computer program is provided, wherein when the computer program is executed in a computer, it causes the computer to perform the steps of the above-described spatiotemporal data prediction method, weather data prediction method, energy data prediction method, or spatiotemporal data prediction data processing method.

[0052] In one or more embodiments of this specification, spatiotemporal sampling data of a target task is acquired. This spatiotemporal sampling data is spatially distributed time-series data. Frequency domain mapping is performed on the spatiotemporal sampling data to obtain frequency domain data. Based on this frequency domain data, a pre-trained spatiotemporal prediction model is used to predict the target spatiotemporal prediction data of the target task. This target spatiotemporal prediction data is spatially distributed time-series data. By performing frequency domain mapping on the spatiotemporal sampling data of the target task, frequency domain data representing global information is obtained. Then, based on this frequency domain data, a pre-trained spatiotemporal prediction model is used to predict the target spatiotemporal prediction data of the target task, thereby improving the accuracy of the predicted target spatiotemporal prediction data. Attached Figure Description

[0053] Figure 1 This is a flowchart of a spatiotemporal data prediction method provided in one embodiment of this specification;

[0054] Figure 2 This is a flowchart illustrating the spatiotemporal prediction model in a spatiotemporal data prediction method provided in one embodiment of this specification.

[0055] Figure 3This is a flowchart illustrating a weather data forecasting method provided in one embodiment of this specification;

[0056] Figure 4 This is a flowchart illustrating an energy data prediction method provided in one embodiment of this specification;

[0057] Figure 5 This is a flowchart of a data processing method for spatiotemporal data prediction provided in one embodiment of this specification;

[0058] Figure 6 This is a flowchart illustrating the processing procedure of an energy data prediction method applied to electrical load, provided in one embodiment of this specification.

[0059] Figure 7 This is a schematic diagram of the structure of a data processing method system for spatiotemporal data prediction provided in one embodiment of this specification;

[0060] Figure 8 This is a schematic diagram of the structure of a spatiotemporal data prediction device provided in one embodiment of this specification;

[0061] Figure 9 This is a schematic diagram of the structure of a weather data forecasting device provided in one embodiment of this specification;

[0062] Figure 10 This is a schematic diagram of the structure of an energy data prediction device provided in one embodiment of this specification;

[0063] Figure 11 This is a schematic diagram of the structure of a data processing device for spatiotemporal data prediction provided in one embodiment of this specification;

[0064] Figure 12 This is a structural block diagram of a computing device provided in one embodiment of this specification. Detailed Implementation

[0065] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0066] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0067] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0068] First, the terms and concepts used in one or more embodiments of this specification will be explained.

[0069] CNN (Convolutional Neural Networks) model: A multi-layer neural network model with forward and backward propagation, featuring convolutional kernels (filters) that process feature data.

[0070] LSTM (Long Short Term Memory) model: A neural network model with the ability to remember information in both short and long term, and has convolutional kernels (filters) to process feature data.

[0071] Transformer model: A neural network model based on the attention mechanism, which extracts and analyzes the features of data through the attention mechanism.

[0072] Time-series data: Data that changes over time and is represented in a coordinate system. The horizontal axis t represents time, and the vertical axis represents the task data of the target task. For example, the temperature change of a certain region every hour of the day, with the horizontal axis t representing the hour and the vertical axis representing the temperature.

[0073] Frequency domain data: The distribution of different task data on frequencies is a linear combination form, represented in a coordinate system. The horizontal axis w represents the frequency, and the vertical axis represents the energy intensity of the task data. For example, for multiple broadcast signals at different frequencies, the horizontal axis w represents the frequency, and the vertical axis represents the number of advertising signals at each frequency.

[0074] Frequency domain mapping: A coordinate dimension mapping method for converting time-series data into frequency domain data based on Fourier transform.

[0075] Fourier Transform (FT): A method that uses the complex form of trigonometric functions to transform time-series or spatial data into frequency-domain data represented by a linear combination of frequencies. The Fourier transform formula for time-series data f(t) is shown in Equation 1:

[0076]

[0077] Accordingly, the formula for the inverse Fourier transform is shown in Formula 2:

[0078]

[0079] For discrete time series data The formula for the Discrete Fourier Transform (DFT) is shown in Equation 3:

[0080]

[0081] Accordingly, the formula for the inverse discrete Fourier transform is shown in Formula 4:

[0082]

[0083] Furthermore, spatial data can also be converted into frequency domain data through Fourier transform. For spatial data f(x, y), x∈[0, M-1]&&y∈[0, N-1], the discrete Fourier transform formula is shown in Formula 5:

[0084]

[0085] Accordingly, the formula for the discrete Fourier transform is shown in Formula 6:

[0086]

[0087] Fourier Network Operator (FNO): A convolutional kernel for neural network models based on Fourier transform. The Fourier Network Operator performs feature processing on frequency domain data representing global information, and then obtains the output data through inverse frequency domain mapping.

[0088] U-Net: A neural network model with a symmetric structure. U-Net is a neural network model with an encoder-decoder structure, in which the encoder downsamples the input data, and the decoder upsamples the downsampled features to decode the output data.

[0089] This specification provides a spatiotemporal prediction method, a weather data prediction method, an energy data prediction method, a spatiotemporal data prediction data processing method, a spatiotemporal data prediction data processing system, a spatiotemporal data prediction device, a weather data prediction device, an energy data prediction device, a spatiotemporal data prediction data processing device, a computing device, a computer-readable storage medium, and a computer program, which will be described in detail in the following embodiments.

[0090] Figure 1 A flowchart of a spatiotemporal data prediction method according to an embodiment of this specification is shown, including the following specific steps:

[0091] Step 102: Obtain the spatiotemporal sampling data of the target task, wherein the spatiotemporal sampling data is spatially distributed time-series data.

[0092] This manual provides real-time examples applicable to applications, web page clients, or servers with spatiotemporal data prediction capabilities.

[0093] The target task is a data prediction task with multiple spatially distributed task points, such as weather forecasting, energy generation forecasting, traffic flow forecasting, scientific research data forecasting, modeling and simulation, and retail sales forecasting. Each task point samples and generates corresponding time-series data. Combined with the spatial distribution of these task points, spatiotemporal sampling data is constructed. For example, the target task could be a traffic flow forecasting task, where each task point is a distributed traffic flow node. Each traffic node P samples and generates corresponding traffic data. Combining the spatial distribution of each traffic node X k and Y l Spatial-Temporal Data is constructed as follows: For ease of explanation in the following description, the spatiotemporal sampling data will be referred to as "ST_Data" in the embodiments of this specification. Here, the spatial dimension is an abstract spatial dimension. For example, a pixel on a display screen can be regarded as a spatially distributed task point, and the chromaticity change value can be regarded as the temporal data sampled and generated from that pixel.

[0094] The spatiotemporal sampling data for the target task is a spatially distributed temporal sampling data of the target task, possessing both temporal and spatial dimensions. The spatial dimension can be two-dimensional or three-dimensional, without limitation here. The spatiotemporal sampling data can be understood as a matrix of data containing multiple spatially distributed elements. Each element represents a task point, and the value of each element is a variable corresponding to the temporal data sampled and generated from that task point.

[0095] It should be noted that the sampling method for each task point can be either continuous time sampling within a preset sampling time period or discrete time sampling. The corresponding generated time series data can be either continuous time series data or discrete time series data, and no limitation is made here.

[0096] The spatiotemporal sampling data of the target task can be obtained directly from the database of the target task, or it can be constructed by sampling at each sampling point and generating time series data. There is no limitation here.

[0097] For example, the spatiotemporal sampling data ST_Data of the traffic prediction task can be obtained directly from the database of the traffic prediction task.

[0098] Acquiring spatiotemporal sampling data of the target task lays the data foundation for subsequent frequency domain conversion to obtain the frequency domain data of the spatiotemporal sampling data, thus providing a spatially distributed time-series data basis.

[0099] Step 104: Perform frequency domain mapping on the spatiotemporal sampling data to obtain the frequency domain data of the spatiotemporal sampling data.

[0100] The frequency domain data of spatiotemporal sampling data is the frequency distribution data obtained by frequency domain mapping of the time and / or spatial dimensions of the spatiotemporal sampling data, and has a frequency space. The component data of spatiotemporal sampling data in the time and spatial dimensions are converted into frequency domain data in the frequency space through frequency domain mapping.

[0101] Frequency domain mapping is based on Fourier transform. As shown in Equations 1-6 above, Fourier transform is achieved by performing integration or discrete summation operations on temporal and spatial data, yielding a type of global information. Since frequency domain data is inherently global information, performing subsequent 106-step predictions based on this global information offers a higher degree of certainty regarding the accuracy of global information compared to methods like CNN models and LSTM convolutional kernels or Transformer-related attention mechanisms, which require training to determine global information extraction capabilities.

[0102] Frequency domain mapping is performed on the spatiotemporal sampled data to obtain its frequency domain data. Specifically, frequency domain mapping is performed on the spatiotemporal sampled data along the time and / or spatial dimensions to obtain its frequency domain data. The specific tool for implementing frequency domain mapping is a frequency domain mapping unit that includes a mapping matrix. It should be noted that step 104 can be implemented using a frequency domain mapping unit embedded in the spatiotemporal prediction model, or it can be implemented using a frequency domain mapping unit independent of the spatiotemporal prediction model; this is not limited here.

[0103] For example, the time-domain mapping of the spatiotemporal sampled data ST_Data is performed to obtain the frequency domain data Freq_ST_Data of the spatiotemporal sampled data ST_Data.

[0104] By performing frequency domain mapping on the spatiotemporal sampling data, frequency domain data of the spatiotemporal sampling data is obtained, providing frequency domain data representing global information for subsequent spatiotemporal data prediction using spatiotemporal prediction models.

[0105] Step 106: Based on the frequency domain data, use the pre-trained spatiotemporal prediction model to predict the target spatiotemporal prediction data of the target task, wherein the target spatiotemporal prediction data is spatially distributed time series data.

[0106] The target spatiotemporal prediction data for the objective task is the spatially distributed temporal prediction data of the objective task, possessing both temporal and spatial dimensions. The spatial dimension is consistent with the spatial dimension of the spatiotemporal sampling data and can be either two-dimensional or three-dimensional; no limitation is made here. The target spatiotemporal prediction data can also be understood as a matrix of data, containing multiple spatially distributed elements. Each element represents a task point, and the value of each element is a variable corresponding to the predicted temporal data for that task point. The target spatiotemporal prediction data refers to the prediction data under the target spatiotemporal context. The target spatiotemporal context is the target time and target space corresponding to the target task. The spatial distribution of the target space and the spatiotemporal sampling data can be consistent or inconsistent. For example, for a traffic prediction task, the target time is the next 3 days, and the target space is the target spatial distribution of each traffic node. Based on the frequency domain data Freq_ST_Data of the spatiotemporal sampling data, the time series data (predicted traffic) of the target time (next 3 days) and the spatial data (spatial distribution) of the target space of the traffic prediction task are predicted using a pre-trained spatiotemporal prediction model. This data can be used to adjust the spatial distribution of traffic nodes and to formulate traffic strategies for the next 3 days, such as flow limiting strategies and traffic diversion strategies.

[0107] The spatiotemporal prediction model is a neural network model with spatiotemporal data prediction capabilities. It is pre-trained based on frequency domain data and labeled spatiotemporal data of the sample spatiotemporal data. The model includes at least one spatiotemporal prediction unit. When the number of spatiotemporal prediction units is greater than or equal to two, the units can be sequentially connected (series), parallelly connected (parallel), or a combination of both; no limitation is imposed here.

[0108] Based on frequency domain data, a pre-trained spatiotemporal prediction model is used to predict the target spatiotemporal prediction data of the target task. Specifically, based on the frequency domain data, at least one spatiotemporal prediction unit in the pre-trained spatiotemporal prediction model is used for prediction to obtain the target spatiotemporal prediction data of the target task. Furthermore, the input of each spatiotemporal prediction unit can be spatiotemporal sampling data (when the frequency domain mapping unit is embedded in each spatiotemporal prediction unit), or frequency domain data of spatiotemporal sampling data (when the frequency domain mapping unit is independent of the spatiotemporal prediction model and the spatiotemporal prediction units are connected in parallel), or the output of the previously connected spatiotemporal prediction unit (when the spatiotemporal prediction units are connected in sequence). The above are some examples and are not limited.

[0109] For example, based on the frequency domain data Freq_ST_Data, prediction is performed using n spatiotemporal prediction units in the pre-trained spatiotemporal prediction model Model to obtain the target spatiotemporal prediction data Pred_ST_Data for the traffic prediction task.

[0110] In the embodiments of this specification, spatiotemporal sampling data of the target task is acquired. This spatiotemporal sampling data is spatially distributed time-series data. Frequency domain mapping is performed on the spatiotemporal sampling data to obtain frequency domain data. Based on this frequency domain data, a pre-trained spatiotemporal prediction model is used to predict the target spatiotemporal prediction data of the target task. This target spatiotemporal prediction data is spatially distributed time-series data. By performing frequency domain mapping on the spatiotemporal sampling data of the target task, frequency domain data representing global information is obtained. Then, based on this frequency domain data, the pre-trained spatiotemporal prediction model is used to predict the target spatiotemporal prediction data of the target task, thereby improving the accuracy of the predicted target spatiotemporal prediction data.

[0111] Optionally, the pre-trained spatiotemporal prediction model includes n sequentially connected spatiotemporal prediction units, where n is a positive integer greater than or equal to 2;

[0112] Accordingly, step 106 includes the following specific steps:

[0113] Based on the frequency domain data of the spatiotemporal sampling data, the first spatiotemporal prediction data is predicted using the first spatiotemporal prediction unit.

[0114] Based on the spatiotemporal sampling data and the (i-1)th spatiotemporal prediction data, the i-th spatiotemporal prediction data is predicted using the i-th spatiotemporal prediction unit, where 2≤i≤n, and the (i-1)th spatiotemporal prediction data is the output of the (i-1)th spatiotemporal prediction unit.

[0115] The i-th spatiotemporal prediction data is determined as the target spatiotemporal prediction data.

[0116] Generally, neural network models have multiple neural network units, such as hidden layers, convolutional layers, and normalization layers. Without fully guaranteeing the performance of the spatiotemporal prediction model, the layer-by-layer processing of data can lead to a large deviation (gradient vanishing) between the predicted target spatiotemporal data and the spatiotemporal sampled data. Therefore, in the embodiments of this specification, the sequentially connected spatiotemporal prediction units not only make predictions based on the output of the previous sequentially connected spatiotemporal prediction unit, but also use the spatiotemporal sampled data as a reference for prediction.

[0117] The spatiotemporal prediction model structure consists of an encoding unit, n sequentially connected spatiotemporal prediction units, and a decoding unit. Frequency domain data is sequentially input into these units to obtain the predicted target spatiotemporal prediction data. The spatiotemporal prediction model in the embodiments of this specification is a neural network model with a U-Net structure.

[0118] The spatiotemporal prediction unit is a neural network unit with spatiotemporal data prediction function, and each spatiotemporal prediction unit includes a Fourier network operator.

[0119] Based on the spatiotemporal sampling data and the (i-1)th spatiotemporal prediction data, the i-th spatiotemporal prediction data is predicted using the i-th spatiotemporal prediction unit. Specifically, based on the frequency domain data of the spatiotemporal sampling data and the (i-1)th frequency domain data corresponding to the (i-1)th spatiotemporal prediction data, the i-th spatiotemporal prediction data is predicted using the i-th spatiotemporal prediction unit. Here, the (i-1)th spatiotemporal prediction data is the output of the (i-1)th spatiotemporal prediction unit, which is spatially distributed time-series data. It needs to be frequency-domain mapped before prediction. The frequency domain mapping can be implemented through a frequency domain mapping unit embedded in the i-th spatiotemporal prediction unit, or through a frequency domain mapping unit independent of the i-th spatiotemporal prediction unit; this is not limited here.

[0120] For example, based on the frequency domain data Freq_ST_Data of the spatiotemporal sampling data, the first spatiotemporal prediction data Pred_ST_Data 1 is predicted using the first spatiotemporal prediction unit Unit_1. Based on the frequency domain data Freq_ST_Data of the spatiotemporal sampling data and the (i-1)th frequency domain data Freq_Pred_ST_Data1 corresponding to the (i-1)th spatiotemporal prediction data, the i-th spatiotemporal prediction data Pred_ST_Data i is predicted using the i-th spatiotemporal prediction unit Unit_i. The i-th spatiotemporal prediction data Pred_ST_Data i is then determined as the target spatiotemporal prediction data TargetPred_ST_Data.

[0121] Based on the frequency domain data of the spatiotemporal sampling data, the first spatiotemporal prediction data is predicted using the first spatiotemporal prediction unit. Based on the spatiotemporal sampling data and the (i-1)th spatiotemporal prediction data, the ith spatiotemporal prediction data is predicted using the ith spatiotemporal prediction unit, where 2 ≤ i ≤ n, and the (i-1)th spatiotemporal prediction data is the output of the (i-1)th spatiotemporal prediction unit. The ith spatiotemporal prediction data is determined as the target spatiotemporal prediction data. By sequentially connecting the spatiotemporal prediction units, the prediction accuracy is improved. Predicting the ith spatiotemporal prediction data using the ith spatiotemporal prediction unit, based on the spatiotemporal sampling data and the (i-1)th spatiotemporal prediction data, avoids significant deviations between the predicted target spatiotemporal prediction data and the spatiotemporal sampling data, further improving the accuracy of the target spatiotemporal prediction data.

[0122] Optionally, step 104 includes the following specific steps:

[0123] The spatiotemporal sampled data is mapped to the frequency domain in both spatial and temporal dimensions to obtain the spatiotemporal frequency domain data of the spatiotemporal sampled data;

[0124] By performing a spatial-dimensional frequency domain mapping on the spatiotemporal sampled data, we obtain the spatial frequency domain data of the spatiotemporal sampled data.

[0125] Since there is no linear relationship between the temporal data in the time dimension and the spatial data in the spatial dimension of the spatiotemporal sampling data (i.e., the data can be converted between different dimensions through the Ax+B form), and the neural network model processes the data in different dimensions at a higher level of abstraction, it is inevitable that the data between the non-linearly related dimensions will interfere with each other. Therefore, it is necessary to split the data into dimensions and use it as a reference for prediction to avoid interference between the time and spatial dimensions when using each spatiotemporal prediction unit for prediction, which would reduce the accuracy of the first spatiotemporal prediction data.

[0126] The spatiotemporal frequency domain data of the spatiotemporal sampling data is the frequency distribution data obtained by frequency domain mapping of the time and space dimensions of the spatiotemporal sampling data. The spatial frequency domain data of the spatiotemporal sampling data is the frequency distribution data obtained by frequency domain mapping of the spatial dimension of the spatiotemporal sampling data. The first spatiotemporal first reference data is the temporal prediction data of the spatial distribution of the target task corresponding to the first spatiotemporal prediction unit. The first spatiotemporal second reference data is the spatial prediction data of the target task corresponding to the first spatiotemporal prediction unit.

[0127] Frequency domain mapping is specifically implemented through a mapping matrix. That is, the spatiotemporal sampled data is mapped to the frequency domain using the frequency domain mapping matrix to obtain the frequency domain data. The mapping matrix can be obtained by multiplying each spatial dimension by the time dimension, or by adding each spatial dimension by the time dimension.

[0128] For example, using the frequency domain mapping matrix Matrix, which is obtained by multiplying each spatial dimension and the time dimension, the spatiotemporal sampled data ST_Data is frequency-domain mapped in both spatial and time dimensions to obtain the spatiotemporal frequency domain data Freq_ST_Data1. Using the same frequency domain mapping matrix Matrix, the spatiotemporal sampled data ST_Data is frequency-domain mapped in both spatial and time dimensions to obtain the spatial frequency domain data Freq_ST_Data2.

[0129] By performing frequency domain mapping on the spatiotemporal sampling data in both spatial and temporal dimensions, the spatiotemporal frequency domain data of the spatiotemporal sampling data is obtained. Similarly, by performing frequency domain mapping on the spatiotemporal sampling data in both spatial and temporal dimensions, the spatial frequency domain data of the spatiotemporal sampling data is obtained. By splitting and mapping the spatiotemporal frequency domain, the accuracy of the first spatiotemporal prediction data is improved.

[0130] Optionally, the spatiotemporal sampling data is mapped to the frequency domain in both spatial and temporal dimensions to obtain the spatiotemporal frequency domain data of the spatiotemporal sampling data, including the following specific steps:

[0131] Using the first frequency domain mapping matrix, the spatiotemporal sampled data is frequency domain mapped in both spatial and temporal dimensions to obtain the spatiotemporal frequency domain data of the spatiotemporal sampled data. The first frequency domain mapping matrix is ​​the frequency domain mapping matrix obtained by adding the spatial and temporal dimensions.

[0132] Accordingly, the spatiotemporal sampling data is mapped to the frequency domain in the spatial dimension to obtain the spatial frequency domain data of the spatiotemporal sampling data, including:

[0133] Using the second frequency domain mapping matrix, the spatiotemporal sampling data is frequency domain mapped in spatial dimensions to obtain the spatial frequency domain data of the spatiotemporal sampling data. The second frequency domain mapping matrix is ​​the frequency domain mapping matrix obtained by adding the spatial dimensions.

[0134] Because spatially distributed time-series data is large in volume, it is processed by multiple sequentially connected spatiotemporal prediction units to predict the target spatiotemporal prediction data. Traditionally, frequency domain mapping is performed using a multiplicative frequency domain mapping matrix, i.e., a concatenated mapping matrix: fxfyfz(h)=fxy(h)+fyz(h)+fxz(h). This results in excessively large matrix parameters, which places excessive demands on hardware computing power in practical applications and carries the risk of overfitting, affecting the efficiency and accuracy of spatiotemporal prediction.

[0135] In the embodiments of this specification, a first mapping matrix is ​​used to perform frequency domain mapping of the spatiotemporal sampling data in both spatial and temporal dimensions, and to perform frequency domain mapping of the spatiotemporal sampling data in both spatial and temporal dimensions. Since the first mapping matrix is ​​a frequency domain mapping matrix obtained by adding each spatial dimension and the time dimension, and the second mapping matrix is ​​a frequency domain mapping matrix obtained by adding each spatial dimension, both are parallel mapping matrices: fxfyfz(h)=fxy(h)+fyz(h)+fxz(h), thus reducing the amount of data processing, improving prediction efficiency, reducing hardware costs, effectively avoiding overfitting problems, improving the accuracy of obtaining each spatiotemporal frequency domain data, improving the accuracy of the i-th spatiotemporal prediction data, and thus improving the accuracy of the target spatiotemporal prediction data.

[0136] Optionally, based on the frequency domain data of the spatiotemporal sampling data, the first spatiotemporal prediction data is predicted using the first spatiotemporal prediction unit, including the following specific steps:

[0137] Fourier network operators are used to process the frequency domain data of the spatiotemporal sampling data to obtain the first spatiotemporal prediction data.

[0138] The Fourier network operator is a convolutional kernel that performs feature prediction processing on frequency domain data representing global information within a spatiotemporal prediction unit. After completing the feature prediction processing of the frequency domain data and obtaining the predicted frequency domain data, the Fourier network operator needs to perform a frequency domain inverse mapping on the predicted frequency domain data to obtain the corresponding spatiotemporal prediction data.

[0139] The frequency domain data of the spatiotemporal sampling data is processed by Fourier network operators to obtain the first spatiotemporal prediction data. Specifically, the Fourier network operator of the first spatiotemporal prediction unit is used to perform feature prediction processing on the frequency domain data of the spatiotemporal sampling data to obtain the first frequency domain prediction data. The first frequency domain prediction data is then subjected to frequency domain inverse mapping to obtain the first spatiotemporal prediction data.

[0140] For example, the Fourier network operator Kernel1 of the first spatiotemporal prediction unit is used to perform feature prediction processing on the frequency domain data Freq_ST_Data of the spatiotemporal sampling data to obtain the first frequency domain prediction data Pred_Freq_ST_Data 1. The first frequency domain prediction data Pred_Freq_ST_Data 1 is then subjected to a time-dimensional inverse frequency domain mapping to obtain the first spatiotemporal prediction data Pred_ST_Data 1.

[0141] The frequency domain data of the spatiotemporal sampling data is processed by Fourier network operators to obtain the first spatiotemporal prediction data, which is more accurate. This ensures that the subsequent i-th spatiotemporal prediction data is more accurate, and thus the target spatiotemporal prediction data is more accurate.

[0142] Optionally, the frequency domain data of the spatiotemporal sampling data is processed by a Fourier network operator to obtain the first spatiotemporal prediction data, including the following specific steps:

[0143] Fourier network operators are used to process the spatiotemporal frequency domain data of the spatiotemporal sampling data to obtain the first spatiotemporal first reference data;

[0144] Fourier network operators are used to process the spatial frequency domain data of the spatiotemporal sampling data to obtain the first spatiotemporal second reference data;

[0145] Based on the first reference data and the second reference data of the first spacetime, the prediction data of the first spacetime is obtained.

[0146] The first spatiotemporal reference data is obtained by processing the spatiotemporal frequency domain data of the spatiotemporal sampling data with Fourier network operators. Specifically, the feature prediction processing of the spatiotemporal frequency domain data of the spatiotemporal sampling data is performed using the Fourier network operator of the first spatiotemporal prediction unit to obtain the first frequency domain reference data. The first frequency domain reference data is then subjected to frequency domain inverse mapping to obtain the first spatiotemporal reference data.

[0147] The spatial frequency domain data of the spatiotemporal sampling data is processed by Fourier network operators to obtain the first spatiotemporal second reference data. Specifically, the Fourier network operator of the first spatiotemporal prediction unit is used to perform feature prediction processing on the spatial frequency domain data of the spatiotemporal sampling data to obtain the first frequency domain second reference data. The first frequency domain second reference data is then subjected to frequency domain inverse mapping to obtain the first spatiotemporal second reference data.

[0148] Based on the first and second reference data of the first spatiotemporal space, the first spatiotemporal prediction data is obtained. Specifically, the first and second reference data of the first spatiotemporal space are weighted and calculated to obtain the first spatiotemporal prediction data. Furthermore, the spatiotemporal sampled data, the first and second reference data of the first spatiotemporal space are weighted and calculated to obtain the first spatiotemporal prediction data. By combining the spatiotemporal sampled data, frequency domain mapping can be better avoided, which could lead to a large deviation between the predicted target spatiotemporal prediction data and the spatiotemporal sampled data, thus achieving residual preservation.

[0149] For example, using the Fourier network operator R1 of the first spatiotemporal prediction unit Unit_1, feature prediction processing is performed on the spatiotemporal frequency domain data Freq_ST_Data1 of the spatiotemporal sampling data to obtain the first reference data Freq_Pred_ST_Data 1_1 in the first frequency domain. Frequency domain inverse mapping is then performed on the first reference data Freq_Pred_ST_Data 1_1 in the first frequency domain to obtain the first spatiotemporal reference data Pred_ST_Data 1_1. Using the Fourier network operator R1 of the first spatiotemporal prediction unit Unit_1, feature prediction processing is performed on the spatial frequency domain data Freq_ST_Data2 of the spatiotemporal sampling data to obtain the second reference data Freq_Pred_ST_Data 1_2 in the first frequency domain. Frequency domain inverse mapping is then performed on the second reference data Freq_Pred_ST_Data 1_2 in the first frequency domain to obtain the second reference data Pred_ST_Data in the first spatiotemporal domain. 1_2, weighted calculation is performed on the spatiotemporal sampled data ST_Data, the first reference data Pred_ST_Data 1_1 of the first spatiotemporal space, and the second reference data Pred_ST_Data 1_2 of the first spatiotemporal space to obtain the predicted data Pred_ST_Data 1 of the first spatiotemporal space.

[0150] Fourier network operators are applied to the spatiotemporal frequency domain data of the spatiotemporal sampling data to obtain the first spatiotemporal reference data. Fourier network operators are then applied to the spatial frequency domain data of the spatiotemporal sampling data to obtain the second spatiotemporal reference data. Based on the first and second spatiotemporal reference data, the first spatiotemporal prediction data is obtained. By using the spatiotemporal frequency domain data of the dimensionally split spatiotemporal sampling data for prediction, and combining it with the spatial frequency domain data of the spatiotemporal sampling data as a reference, the accuracy of the first spatiotemporal prediction data corresponding to the first spatiotemporal prediction unit is improved. This, in turn, improves the prediction accuracy of the subsequent i-th spatiotemporal prediction model, ultimately enhancing the accuracy of the obtained target spatiotemporal prediction data.

[0151] Optionally, based on the spatiotemporal sampling data and the (i-1)th spatiotemporal prediction data, the ith spatiotemporal prediction unit is used to predict the ith spatiotemporal prediction data, including the following specific steps:

[0152] The i-1th spatiotemporal prediction data is mapped to the frequency domain to obtain the corresponding i-1th frequency domain data. The i-1th frequency domain data is then processed by Fourier network operators to obtain the i-th spatiotemporal first reference data.

[0153] Fourier network operators are used to process the frequency domain data of the spatiotemporal sampling data to obtain the i-th spatiotemporal second reference data;

[0154] Based on the first reference data and the second reference data of the i-th spacetime, the predicted data of the i-th spacetime is obtained.

[0155] After the first i-1 spatiotemporal prediction units of the spatiotemporal prediction model process the data layer by layer, a large deviation (gradient vanishing) occurs between the predicted target spatiotemporal prediction data and the spatiotemporal sampling data. Therefore, in the embodiments of this specification, the i-th spatiotemporal prediction unit connected in sequence not only makes predictions based on the output of the i-1-th spatiotemporal prediction unit, but also uses the spatiotemporal sampling data as a reference for prediction.

[0156] The Fourier network operator is a convolutional kernel that performs feature prediction processing on frequency domain data representing global information within a spatiotemporal prediction unit. After completing the feature prediction processing of the frequency domain data and obtaining the predicted frequency domain data, the Fourier network operator needs to perform a frequency domain inverse mapping on the predicted frequency domain data to obtain the corresponding spatiotemporal prediction data.

[0157] The first reference data in the i-th spatiotemporal space is the temporal prediction data of the spatial distribution of the target task corresponding to the i-th spatiotemporal prediction unit, and the second reference data in the i-th spatiotemporal space is the spatial prediction data of the target task corresponding to the i-th spatiotemporal prediction unit.

[0158] Frequency domain mapping is performed on the (i-1)th spatiotemporal prediction data to obtain the corresponding (i-1)th frequency domain data. Specifically, frequency domain mapping is performed on the (i-1)th spatiotemporal prediction data in terms of time and / or spatial dimensions to obtain the frequency domain data of the (i-1)th spatiotemporal prediction data. The specific frequency domain mapping dimension is consistent with the frequency domain mapping dimension in step 104.

[0159] The i-1 frequency domain data is processed by Fourier network operators to obtain the i-th spatiotemporal first reference data. Specifically, the feature prediction processing of the i-1 frequency domain data is performed using the Fourier network operator of the i-th spatiotemporal prediction unit to obtain the i-th frequency domain first reference data. The i-th frequency domain first reference data is then subjected to frequency domain inverse mapping to obtain the i-th spatiotemporal first reference data.

[0160] The frequency domain data of the spatiotemporal sampling data is processed by Fourier network operators to obtain the i-th spatiotemporal second reference data. Specifically, the Fourier network operators of the i-th spatiotemporal prediction unit are used to perform feature prediction processing on the frequency domain data of the spatiotemporal sampling data to obtain the i-th frequency domain second reference data. The i-th frequency domain second reference data is then subjected to frequency domain inverse mapping to obtain the i-th spatiotemporal second reference data.

[0161] Based on the first and second reference data of the i-th spatiotemporal space, the predicted data of the i-th spatiotemporal space is obtained. Specifically, the first and second reference data of the i-th spatiotemporal space are weighted and calculated. Further, the spatiotemporal sampled data, the first and second reference data of the i-th spatiotemporal space are weighted and calculated to obtain the predicted data of the i-th spatiotemporal space. By combining the spatiotemporal sampled data, the large deviation between the predicted target spatiotemporal data and the spatiotemporal sampled data caused by the previous (i-1)-layer feature processing can be better avoided, thus achieving residual preservation.

[0162] For example, the (i-1)th spatiotemporal prediction data Pred_ST_Data i-1 is frequency-domain mapped to obtain the corresponding (i-1)th frequency-domain data Freq_Pred_ST_Data i-1. Then, using the Fourier network operator Ri of the i-th spatiotemporal prediction unit Unit_i, feature prediction processing is performed on the (i-1)th frequency-domain data Freq_Pred_ST_Data i-1 to obtain the i-th first frequency-domain reference data Freq_Pred_ST_Data i_1. The i-th first frequency-domain reference data Freq_Pred_ST_Data i_1 is then inversely frequency-domain mapped to obtain the i-th spatiotemporal first reference data Pred_ST_Data i_1. Finally, using the Fourier network operator Ri of the i-th spatiotemporal prediction unit Unit_i, feature prediction processing is performed on the frequency-domain data Freq_ST_Data of the spatiotemporal sampling data to obtain the i-th second frequency-domain reference data Freq_Pred_ST_Data i_2. Finally, the i-th second frequency-domain reference data Freq_Pred_ST_Data... i_2 is subjected to frequency domain inverse mapping to obtain the i-th spatiotemporal second reference data Pred_ST_Data i_2. The spatiotemporal sampled data ST_Data, the i-th spatiotemporal first reference data Pred_ST_Data i_1, and the i-th spatiotemporal second reference data Pred_ST_Data i_2 are weighted and calculated to obtain the i-th spatiotemporal prediction data Pred_ST_Data i.

[0163] Frequency domain mapping is performed on the (i-1)th spatiotemporal prediction data to obtain the corresponding (i-1)th frequency domain data. Fourier network operator processing is then applied to the (i-1)th frequency domain data to obtain the first spatiotemporal reference data for the i-th time. Similarly, Fourier network operator processing is applied to the frequency domain data of the spatiotemporal sampling data to obtain the second spatiotemporal reference data for the i-th time. Based on the first and second spatiotemporal reference data, the i-th spatiotemporal prediction data is obtained. By utilizing a bi-branching spatiotemporal prediction unit and combining it with the frequency domain data of the spatiotemporal sampling data as a reference, the accuracy of the i-th spatiotemporal prediction data is improved, thereby enhancing the accuracy of the target spatiotemporal prediction data. Furthermore, Fourier network operator processing of the frequency domain data yields even more accurate i-th spatiotemporal prediction data, further improving the accuracy of the target spatiotemporal prediction data.

[0164] Optionally, the (i-1)th spatiotemporal prediction data is frequency-domain mapped to obtain the corresponding (i-1)th frequency domain data, and the (i-1)th frequency domain data is processed by a Fourier network operator to obtain the first spatiotemporal reference data of the ith, including the following specific steps:

[0165] The spatial and temporal dimensions of the (i-1)th spatiotemporal prediction data are mapped to the frequency domain to obtain the corresponding (i-1)th spatiotemporal frequency domain data. The (i-1)th spatiotemporal frequency domain data is then processed by a Fourier network operator to obtain the first reference data of the i-th spatiotemporal space.

[0166] Accordingly, the frequency domain data of the spatiotemporal sampling data is processed by Fourier network operators to obtain the i-th spatiotemporal second reference data, including the following specific steps:

[0167] The spatial frequency domain data of the spatiotemporal sampling data is processed by Fourier network operators to obtain the i-th spatiotemporal second reference data.

[0168] Since there is no linear relationship between the temporal data in the time dimension and the spatial data in the spatial dimension of the spatiotemporal sampling data (i.e., the data can be converted between different dimensions through the Ax+B form), and the neural network model processes the data in different dimensions at a higher level of abstraction, it is inevitable that the data between the non-linearly related dimensions will interfere with each other. Therefore, it is necessary to split the data into dimensions and use it as a reference for prediction to avoid interference between the time and spatial dimensions when using each spatiotemporal prediction unit for prediction, which would reduce the accuracy of the i-th spatiotemporal prediction data.

[0169] The specific methods for processing the (i-1)th spatiotemporal frequency domain data with Fourier network operators to obtain the i-th spatiotemporal first reference data, and for processing the spatial frequency domain data of the spatiotemporal sampled data with Fourier network operators to obtain the i-th spatiotemporal second reference data, have been explained above and will not be repeated here.

[0170] For example, using the Fourier network operator Ri of the i-th spatiotemporal prediction unit Unit_i, feature prediction processing is performed on the (i-1)-th frequency domain data Freq_Pred_ST_Data i-1 to obtain the i-th first frequency domain reference data Freq_Pred_ST_Data i_1. Frequency domain inverse mapping is then performed on the i-th first frequency domain reference data Freq_Pred_ST_Data i_1 to obtain the i-th spatiotemporal first reference data Pred_ST_Data i_1. Using the Fourier network operator Ri of the i-th spatiotemporal prediction unit Unit_i, feature prediction processing is performed on the frequency domain data Freq_ST_Data of the spatiotemporal sampled data to obtain the i-th second frequency domain reference data Freq_Pred_ST_Data i_2. Frequency domain inverse mapping is then performed on the i-th second frequency domain reference data Freq_Pred_ST_Data i_2 to obtain the i-th spatiotemporal second reference data Pred_ST_Data i_2.

[0171] Fourier network operators are applied to the (i-1)th spatiotemporal frequency domain data to obtain the first reference data for the i-th spatiotemporal space. Fourier network operators are also applied to the spatial frequency domain data of the spatiotemporal sampled data to obtain the second reference data for the i-th spatiotemporal space. Prediction is performed using the spatiotemporal frequency domain data of the (i-1)th spatiotemporal prediction data after dimensionality splitting, combined with the spatial frequency domain data of the spatiotemporal sampled data as a reference. This improves the accuracy of the i-th spatiotemporal prediction data corresponding to the i-th spatiotemporal prediction unit, and ultimately improves the accuracy of the obtained target spatiotemporal prediction data.

[0172] Optionally, the (i-1)th spatiotemporal prediction data is mapped to the frequency domain in both spatial and temporal dimensions to obtain the corresponding (i-1)th spatiotemporal frequency domain data, including the following specific steps:

[0173] Using the first frequency domain mapping matrix, the spatial and temporal dimensions of the (i-1)th spatiotemporal prediction data are mapped in the frequency domain to obtain the corresponding (i-1)th spatiotemporal frequency domain data. The first frequency domain mapping matrix is ​​the frequency domain mapping matrix obtained by adding the spatial and temporal dimensions.

[0174] Because spatially distributed time-series data is large in volume, it is processed by multiple sequentially connected spatiotemporal prediction units to predict the target spatiotemporal prediction data. Traditionally, frequency domain mapping is performed using a multiplicative frequency domain mapping matrix, i.e., a concatenated mapping matrix: fxfyfz(h)=fxy(h)+fyz(h)+fxz(h). This results in excessively large matrix parameters, which places excessive demands on hardware computing power in practical applications and carries the risk of overfitting, affecting the efficiency and accuracy of spatiotemporal prediction.

[0175] In the embodiments of this specification, the first mapping matrix is ​​used to perform frequency domain mapping of the spatial and temporal dimensions of the (i-1)th spatiotemporal prediction data. Since the first mapping matrix is ​​the frequency domain mapping matrix obtained by adding each spatial and temporal dimension, that is, the parallel mapping matrix: fxfyfz(h)=fx(h)+fy(h)+fz(h), the amount of data processing is reduced, the prediction efficiency is improved, the hardware cost is reduced, and the overfitting problem is effectively avoided, thereby improving the accuracy of the obtained (i-1)th spatiotemporal frequency domain data, improving the accuracy of the i-th spatiotemporal prediction data, and thus improving the accuracy of the target spatiotemporal prediction data.

[0176] Optionally, prior to step 106, the following specific steps are also included:

[0177] Obtain a sample set, which includes multiple sample groups, and each sample group includes frequency domain data of sample spatiotemporal data and label spatiotemporal data;

[0178] Extract the first sample group from the sample set, wherein the first sample group is any sample group, and the first sample group includes the frequency domain data of the first sample spatiotemporal data and the first label spatiotemporal data;

[0179] Based on the frequency domain data of the first sample spatiotemporal data, the spatiotemporal prediction data of the sample is predicted using the spatiotemporal prediction model.

[0180] Calculate the loss value based on the sample spatiotemporal prediction data and the label spatiotemporal data;

[0181] Based on the loss value, adjust the model parameters of the spatiotemporal prediction model, return to the step of extracting the first sample group from the sample set, and obtain the spatiotemporal prediction model after training is completed if the preset training termination conditions are met.

[0182] The sample set is a collection of spatiotemporal data samples, comprising multiple sample groups. Each sample group includes frequency domain data of the sample spatiotemporal data and label spatiotemporal data. The frequency domain data of the sample spatiotemporal data is used as input to the spatiotemporal prediction model. After the sample spatiotemporal prediction data is obtained, it is compared with the label spatiotemporal data for supervised training. The frequency domain data of the sample spatiotemporal data is obtained after frequency domain transformation of the sample spatiotemporal data. The sample spatiotemporal data or its frequency domain data can be stored in the local database of the target task or obtained from a remote database. The remote database can be an open-source database, which is not limited here. The label spatiotemporal data is time-series data with a corresponding spatial distribution to the sample spatiotemporal data under the same target task. The time-series data of the sample spatiotemporal data precedes the time-series data of the label spatiotemporal data. The frequency domain data of the sample spatiotemporal data is stored correspondingly to the label spatiotemporal data.

[0183] The loss value can be cross-entropy loss, CTC loss, cosine similarity loss, L1 loss, etc., and is not limited here.

[0184] The preset training termination condition is a pre-defined condition for determining the end of training. It can be a preset threshold for the number of training iterations, a preset threshold for the loss value, or a preset condition related to the sample group. For example, all sample groups in the sample set have completed training, or any sample group in the sample set has participated in training for a preset number of times.

[0185] Based on the loss value, the network parameters of the summarization generation network are adjusted. Specifically, the network parameters of the summarization generation network are adjusted using gradient descent based on the loss value.

[0186] For example, a sample set Sample_Set is obtained, wherein the sample set Sample_Set includes 10,000 sample groups Sample_Group(Sample_Group_j, j∈[1,10000]), and any sample group Sample_Group_j includes the frequency domain data Freq_Sample_ST_Data_j of the sample spatiotemporal data Sample_ST_Data_j and the label spatiotemporal data Label_ST_Data_j. A first sample group Sample_Group_1 is extracted from the sample set Sample_Set, wherein the first sample group Sample_Group_1 includes the frequency domain data Freq_Sample_ST of the first sample spatiotemporal data. Given _Data_1 and the first label spatiotemporal data Label_ST_Data_1, and using the frequency domain data Freq_Sample_ST_Data_1 of the first sample spatiotemporal data, the spatiotemporal prediction model Model1 is used to predict the sample spatiotemporal prediction data Pred_ST_Data_1. Based on the sample spatiotemporal prediction data Pred_ST_Data_1 and the label spatiotemporal data Label_ST_Data_1, the cross-entropy loss value Loss is calculated. Based on the cross-entropy loss value Loss, the model parameters of the spatiotemporal prediction model Model1 are adjusted using the gradient descent method. The process then returns to the step of extracting the first sample group Sample_Group_1 from the sample set Sample_Set, until the preset cross-entropy loss value threshold is reached. Under these conditions, the spatiotemporal prediction model Model1, which has been trained, is obtained.

[0187] Obtain a sample set, which includes multiple sample groups. Each sample group includes frequency domain data of the sample spatiotemporal data and label spatiotemporal data. Extract a first sample group from the sample set. The first sample group is any sample group that includes frequency domain data of the first sample spatiotemporal data and first label spatiotemporal data. Based on the frequency domain data of the first sample spatiotemporal data, use a spatiotemporal prediction model to predict the sample spatiotemporal prediction data. Calculate the loss value based on the sample spatiotemporal prediction data and label spatiotemporal data. Adjust the model parameters of the spatiotemporal prediction model based on the loss value. Return to the step of extracting the first sample group from the sample set. If the preset training termination condition is met, the trained spatiotemporal prediction model is obtained. Based on the frequency domain data of the first sample spatiotemporal data, use the spatiotemporal prediction model to predict the sample spatiotemporal prediction data. Combine this with the label spatiotemporal data to calculate the loss value. Train the spatiotemporal prediction model based on the loss value to obtain a spatiotemporal prediction model that can accurately predict spatiotemporal prediction data with high accuracy based on the global information of the frequency domain data.

[0188] Optionally, after step 106, the following specific steps are also included:

[0189] Send the target spatiotemporal prediction data to the user;

[0190] Receive feedback information sent by the user, wherein the feedback information is generated based on the user's editing operations on the target spatiotemporal prediction data;

[0191] Based on the feedback, adjust the model parameters of the spatiotemporal prediction model.

[0192] The user is the client user who logs into the application or webpage that has spatiotemporal data prediction capabilities.

[0193] The feedback information is generated after the user edits the target spatiotemporal prediction data through the client's front end. Editing operations can include replacing, adding, deleting, compressing, etc.

[0194] Based on feedback, fine-tuning is an operation that adaptively adjusts the model parameters of a neural network model to meet the actual needs of the target task. Specifically, fine-tuning involves adjusting the parameters of the normalization layer (Softmax layer) in the encoding unit of the spatiotemporal prediction model based on feedback information.

[0195] The system sends target spatiotemporal prediction data to the user and receives feedback from the user. This feedback is generated based on the user's editing operations on the target spatiotemporal prediction data. Based on this feedback, the model parameters of the spatiotemporal prediction model are fine-tuned. By interacting with the user and using feedback to fine-tune the model parameters, the performance of the spatiotemporal prediction model is improved, thereby enhancing the accuracy of subsequent spatiotemporal predictions and the resulting data.

[0196] Figure 2 This document illustrates a flowchart of a spatiotemporal prediction model in a spatiotemporal data prediction method provided in one embodiment of this specification.

[0197] like Figure 2 As shown, the spatiotemporal sampling data a(x, t) is input into the spatiotemporal prediction model. First, it is encoded using an encoding unit, and then it is input into sequentially connected spatiotemporal prediction subunits (the first spatiotemporal prediction subunit...the ith spatiotemporal prediction subunit...the nth spatiotemporal prediction subunit) for spatiotemporal data prediction. The i-th spatiotemporal prediction subunit includes two branches. For the first branch, the (i-1)-th spatiotemporal prediction data v_i-1(x,t) output by the (i-1)-th spatiotemporal prediction subunit is frequency-domain mapped. The frequency-domain mapped (i-1)-th frequency domain data is then input into the Fourier network operator Ri. After feature prediction processing, the first reference data in the i-th frequency domain is obtained. This data is then inversely frequency-domain mapped to obtain the first reference data in the i-th spatiotemporal domain. For the second branch, the spatiotemporal sampling data a(x,t) is frequency-domain mapped. The frequency domain data of the frequency-domain mapped (i-1)-th spatiotemporal sampling data is then input into the Fourier network operator Ri. After feature prediction processing, the second reference data in the i-th frequency domain is obtained. This data is then inversely frequency-domain mapped to obtain the second reference data in the i-th spatiotemporal domain. The (i-1)-th spatiotemporal prediction data is multiplied by the first weight, the first reference data in the i-th spatiotemporal domain is multiplied by the second weight, and the second reference data in the i-th spatiotemporal domain is multiplied by the third weight. Finally, a weighted sum is performed to obtain the i-th spatiotemporal prediction data v_i(x,t). The nth spatiotemporal prediction data output by the nth spatiotemporal prediction unit is input into the decoding unit and decoded to obtain the target spatiotemporal prediction data u(x,t).

[0198] It should be noted that the processing procedure in the i-th spatiotemporal prediction unit is as shown in Equation 7:

[0199]

[0200] Formula 7

[0201] in,

[0202] in,

[0203] The above v i(x, t) represents the spatiotemporal prediction data of the i-th generation, v i-1 (x, t) represents the (i-1)th spatiotemporal prediction data, a(x, t) represents the spatiotemporal sampling data, W represents the weighting weights, and σ() represents an iterable architecture. Characterizing the spatial frequency domain mapping, Characterizing the inverse mapping of the frequency domain in space, Characterizing Fourier network operators, Characterized by nonlinear processing in the time and space dimensions.

[0204] Figure 3 A flowchart of a weather data forecasting method according to an embodiment of this specification is shown, including the following specific steps:

[0205] Step 202: Obtain weather sampling data for the weather forecasting task, wherein the weather sampling data is spatially distributed time-series weather data.

[0206] Step 204: Perform frequency domain mapping on the weather sampling data to obtain the frequency domain data of the weather sampling data.

[0207] Step 206: Based on the frequency domain data of the weather sampling data, use a pre-trained spatiotemporal prediction model to predict the weather prediction data for the weather prediction task. The weather prediction data is spatially distributed time-series weather data.

[0208] Weather forecasting is a task that forecasts weather time-series data, which includes, but is not limited to, temperature, humidity, light intensity, precipitation, and wind speed.

[0209] Weather sampling data is time-series weather data collected from multiple spatially distributed sampling points for weather forecasting tasks.

[0210] Weather forecast data consists of forecast data of time series weather data from multiple spatially distributed sampling points for weather forecasting tasks.

[0211] Steps 202-206 are the same as above. Figure 1 Steps 102-106 in the embodiments are based on the same inventive concept and are implemented in the same way, and will not be described again here.

[0212] Optionally, after step 206, the method further includes:

[0213] Display weather forecast data on the front end.

[0214] In the embodiments of this specification, weather sampling data for a weather forecasting task is acquired. This weather sampling data is spatially distributed time-series weather data. Frequency domain mapping is performed on the weather sampling data to obtain its frequency domain data. Based on this frequency domain data, a pre-trained spatiotemporal prediction model is used to predict the weather forecast data for the weather forecasting task. This method improves the accuracy of the predicted weather data by obtaining frequency domain data representing global information from the weather sampling data and then using the pre-trained spatiotemporal prediction model based on this global information representation.

[0215] Figure 4 A flowchart of an energy data prediction method according to an embodiment of this specification is shown, including the following specific steps:

[0216] Step 302: Obtain energy generation data for the energy generation task, wherein the energy generation data is spatially distributed time-series energy data.

[0217] Step 304: Perform frequency domain mapping on the energy generation data to obtain the frequency domain data of the energy generation data.

[0218] Step 306: Based on the frequency domain data of energy generation data, use a pre-trained spatiotemporal prediction model to predict the energy generation prediction data for the energy generation task, wherein the energy generation prediction data is spatially distributed energy time series data.

[0219] The energy generation task is to predict the time series data of energy generation, which includes, but is not limited to, thermal power data, electricity data, nuclear power data, hydropower data, photovoltaic data, etc.

[0220] Energy generation data refers to time-series energy data collected and generated from multiple spatially distributed energy generation points for energy generation tasks. These energy generation points include, but are not limited to, thermal power plants, hydropower stations, nuclear power plants, and photovoltaic clusters.

[0221] The energy generation prediction data consists of prediction data of the time series data of energy generation from multiple spatially distributed energy generation points for energy generation tasks.

[0222] Steps 302-306 are the same as above. Figure 1 Steps 102-106 in the embodiments are based on the same inventive concept and are implemented in the same way, and will not be described again here.

[0223] In the embodiments of this specification, weather sampling data for a weather forecasting task is acquired. This weather sampling data is spatially distributed time-series weather data. Frequency domain mapping is performed on the weather sampling data to obtain its frequency domain data. Based on this frequency domain data, a pre-trained spatiotemporal prediction model is used to predict the weather forecast data for the weather forecasting task. This method improves the accuracy of the predicted weather data by obtaining frequency domain data representing global information from the weather sampling data and then using the pre-trained spatiotemporal prediction model based on this global information representation.

[0224] In the embodiments of this specification, energy generation data for an energy generation task is acquired. This energy generation data is spatially distributed time-series energy data. Frequency domain mapping is performed on the energy generation data to obtain frequency domain data. Based on this frequency domain data, a pre-trained spatiotemporal prediction model is used to predict the energy generation forecast data for the energy generation task. This predicted energy generation data is also spatially distributed time-series energy data. Similarly, by performing frequency domain mapping on weather sampling data for a weather forecast task, frequency domain data representing global information is obtained. Then, based on this frequency domain data, a pre-trained spatiotemporal prediction model is used to predict weather forecast data, thereby improving the accuracy of the predicted weather forecast data.

[0225] Figure 5 This specification illustrates a flowchart of a data processing method for spatiotemporal data prediction according to an embodiment of the present invention. The method is applied to cloud-side equipment and includes the following specific steps:

[0226] Step 402: Obtain a sample set, wherein the sample set includes multiple sample groups, and any sample group includes frequency domain data of sample spatiotemporal data and label spatiotemporal data;

[0227] Step 404: Extract the first sample group from the sample set, wherein the first sample group is any sample group, and the first sample group includes the frequency domain data of the first sample spatiotemporal data and the first label spatiotemporal data;

[0228] Step 406: Based on the frequency domain data of the first sample spatiotemporal data, use the spatiotemporal prediction model to predict the sample spatiotemporal prediction data;

[0229] Step 408: Calculate the loss value based on the sample spatiotemporal prediction data and the label spatiotemporal data;

[0230] Step 410: Adjust the model parameters of the spatiotemporal prediction model according to the loss value, return to the step of extracting the first sample group from the sample set, and obtain the spatiotemporal prediction model after training is completed if the preset training termination condition is met.

[0231] Step 412: Send the model parameters of the trained spatiotemporal prediction model to the edge device.

[0232] Cloud-side devices are network devices that provide model training capabilities; they are virtual devices. Edge devices are physical devices located on the client-side of applications, web pages, mini-programs, etc., that provide spatiotemporal data prediction capabilities. Cloud-side devices and edge devices are connected via network transmission channels for data transmission.

[0233] Steps 402-410 have been described above. Figure 1 The embodiments are described in detail and will not be repeated here.

[0234] In this embodiment, a sample set is obtained, which includes multiple sample groups. Each sample group includes frequency domain data of sample spatiotemporal data and label spatiotemporal data. A first sample group is extracted from the sample set, which is any sample group and includes frequency domain data of first sample spatiotemporal data and first label spatiotemporal data. Based on the frequency domain data of the first sample spatiotemporal data, a spatiotemporal prediction model is used to predict sample spatiotemporal prediction data. Based on the sample spatiotemporal prediction data and label spatiotemporal data, a loss value is calculated. Based on the loss value, the model parameters of the spatiotemporal prediction model are adjusted. The step of extracting the first sample group from the sample set is returned. If the preset training termination condition is met, the trained spatiotemporal prediction model is obtained. The model parameters of the trained spatiotemporal prediction model are sent to the edge device. Based on the frequency domain data of the first sample spatiotemporal data, the spatiotemporal prediction model is used to predict the sample spatiotemporal prediction data. The loss value is calculated by combining the label spatiotemporal data, and the spatiotemporal prediction model is trained based on the loss value. This allows the spatiotemporal prediction model to accurately predict high-accuracy spatiotemporal prediction data based on the global information of the frequency domain data. The above training is completed through cloud-side devices, which saves training costs for edge devices and improves training efficiency.

[0235] The following is in conjunction with the appendix Figure 6 Taking the application of the energy data prediction method provided in this specification in electricity load prediction as an example, the energy generation method will be further explained. Among other things, Figure 6 The present specification illustrates a process flowchart of an energy data prediction method for electrical load according to an embodiment of this specification, which specifically includes the following steps.

[0236] Step 502: Receive the power load prediction request sent by the client, wherein the power load request carries spatiotemporal sampling data of the power load;

[0237] As the proportion of renewable energy continues to rise in the future, it will eventually become the energy form with the highest installed capacity. In particular, the large-scale integration of distributed renewable energy into the grid poses significant challenges to grid power generation planning and scheduling, affecting the safe and stable operation of the grid. Therefore, accurate load forecasting capabilities will become a crucial technological foundation. Load forecasting refers to determining load data at a specific future moment based on various factors such as system operating characteristics, capacity expansion decisions, natural conditions, and social impacts, while meeting certain accuracy requirements. It is a prerequisite for power system dispatching, real-time control, operation planning, and development planning, and is essential information for grid dispatching and planning departments. Accurate load forecasting allows for the economical and rational scheduling of generator start-up and shutdown within the grid, maintaining the safe and stable operation of the grid, reducing unnecessary spinning reserve capacity, rationally scheduling generator maintenance plans, ensuring normal social production and life, effectively reducing power generation costs, and improving economic and social benefits.

[0238] Currently, the method of manually plotting data points for electricity load forecasting by relevant personnel is becoming increasingly difficult and complex with the integration of new energy sources, especially distributed photovoltaic (PV) systems. Traditional manual methods require personnel to spend a significant amount of time analyzing and manually plotting data points in the face of various external inputs, posing a considerable challenge and burden to their work. Therefore, there is an urgent need for an efficient and accurate energy data forecasting method for electricity load.

[0239] Step 504: Using the first frequency domain mapping matrix, perform frequency domain mapping on the spatiotemporal sampled data in both spatial and temporal dimensions to obtain the spatiotemporal frequency domain data of the spatiotemporal sampled data;

[0240] Step 506: Using the second frequency domain mapping matrix, perform spatial dimension frequency domain mapping on the spatiotemporal sampled data to obtain the spatial frequency domain data of the spatiotemporal sampled data;

[0241] Step 508: Perform Fourier network operator processing on the spatiotemporal frequency domain data of the spatiotemporal sampling data to obtain the first reference data of the first spatiotemporal period, and perform Fourier network operator processing on the spatial frequency domain data of the spatiotemporal sampling data to obtain the second reference data of the first spatiotemporal period.

[0242] Step 510: Obtain the prediction data for the first spacetime based on the first reference data and the second reference data for the first spacetime;

[0243] Step 512: Using the first frequency domain mapping matrix, perform frequency domain mapping on the (i-1)th spatiotemporal prediction data in both spatial and temporal dimensions to obtain the corresponding (i-1)th spatiotemporal frequency domain data. Then, process the (i-1)th spatiotemporal frequency domain data using Fourier network operators to obtain the first reference data for the i-th spatiotemporal space, where 2≤i≤n.

[0244] Step 514: Perform Fourier network operator processing on the spatial frequency domain data of the spatiotemporal sampling data to obtain the i-th spatiotemporal second reference data;

[0245] Step 516: Obtain the predicted data for the i-th spatiotemporal space based on the first reference data and the second reference data for the i-th spatiotemporal space.

[0246] Step 518: Determine the i-th spatiotemporal prediction data as the spatiotemporal prediction data of electricity load;

[0247] Step 520: Send the spatiotemporal forecast data of electricity load to the client.

[0248] In the embodiments of this specification, frequency domain mapping is performed on the spatiotemporal sampling data of the power load to obtain frequency domain data representing global information. Then, based on the frequency domain data representing global information, a pre-trained spatiotemporal prediction model is used to predict the spatiotemporal prediction data of the power load. This improves the accuracy and efficiency of the predicted spatiotemporal prediction data of the power load, effectively reduces the task pressure of power grid power generation planning and mode arrangement, and improves the work efficiency of relevant personnel.

[0249] It should be noted that the spatiotemporal sampling data, weather sampling data, energy generation data, sample sets, spatiotemporal prediction models, and other information and data involved in the above method embodiments are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use, and processing of the relevant data must comply with the relevant laws, regulations, and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0250] Corresponding to the above method embodiments, this specification also provides a system embodiment of a data processing method for spatiotemporal data prediction. Figure 7 This diagram illustrates the structure of a data processing method system for spatiotemporal data prediction according to an embodiment of this specification. Figure 7 As shown, the device includes:

[0251] The edge device 602 is used to send a training request for the spatiotemporal prediction model to the cloud device 604;

[0252] Cloud-side device 604, upon receiving a training request, acquires a sample set, wherein the sample set includes multiple sample groups, each sample group including frequency domain data of sample spatiotemporal data and label spatiotemporal data; extracts a first sample group from the sample set, wherein the first sample group is any sample group, and the first sample group includes frequency domain data of first sample spatiotemporal data and first label spatiotemporal data; predicts sample spatiotemporal prediction data using a spatiotemporal prediction model based on the frequency domain data of the first sample spatiotemporal data; calculates a loss value based on the sample spatiotemporal prediction data and label spatiotemporal data; adjusts the model parameters of the spatiotemporal prediction model based on the loss value, returns to the step of extracting the first sample group from the sample set, and obtains a trained spatiotemporal prediction model if a preset training termination condition is met; and sends the model parameters of the trained spatiotemporal prediction model to the edge device 602.

[0253] The end-side device 602 is also used to receive model parameters of the spatiotemporal prediction model.

[0254] In this embodiment, a sample set is obtained, which includes multiple sample groups. Each sample group includes frequency domain data of sample spatiotemporal data and label spatiotemporal data. A first sample group is extracted from the sample set, which is any sample group and includes frequency domain data of first sample spatiotemporal data and first label spatiotemporal data. Based on the frequency domain data of the first sample spatiotemporal data, a spatiotemporal prediction model is used to predict sample spatiotemporal prediction data. Based on the sample spatiotemporal prediction data and label spatiotemporal data, a loss value is calculated. Based on the loss value, the model parameters of the spatiotemporal prediction model are adjusted. The step of extracting the first sample group from the sample set is returned. If the preset training termination condition is met, the trained spatiotemporal prediction model is obtained. The model parameters of the trained spatiotemporal prediction model are sent to the edge device. Based on the frequency domain data of the first sample spatiotemporal data, the spatiotemporal prediction model is used to predict the sample spatiotemporal prediction data. The loss value is calculated by combining the label spatiotemporal data, and the spatiotemporal prediction model is trained based on the loss value. This allows the spatiotemporal prediction model to accurately predict high-accuracy spatiotemporal prediction data based on the global information of the frequency domain data. The above training is completed through cloud-side devices, which saves training costs for edge devices and improves training efficiency.

[0255] The above is an illustrative scheme of a data processing device for spatiotemporal data prediction according to this embodiment. It should be noted that the technical solution of this data processing device for spatiotemporal data prediction belongs to the same concept as the technical solution of the data processing method for spatiotemporal data prediction described above. For details not described in detail in the technical solution of the data processing device for spatiotemporal data prediction, please refer to the description of the technical solution of the data processing method for spatiotemporal data prediction described above.

[0256] Corresponding to the above method embodiments, this specification also provides embodiments of spatiotemporal data prediction devices. Figure 8 A schematic diagram of a spatiotemporal data prediction device according to one embodiment of this specification is shown. Figure 8 As shown, the device includes:

[0257] The first acquisition module 702 is configured to acquire spatiotemporal sampling data of the target task, wherein the spatiotemporal sampling data is spatially distributed time-series data;

[0258] The first mapping module 704 is configured to perform frequency domain mapping on the spatiotemporal sampling data to obtain the frequency domain data of the spatiotemporal sampling data.

[0259] The first prediction module 706 is configured to predict the target spatiotemporal prediction data of the target task based on the frequency domain data and using a pre-trained spatiotemporal prediction model, wherein the target spatiotemporal prediction data is spatially distributed time series data.

[0260] Optionally, the pre-trained spatiotemporal prediction model includes n sequentially connected spatiotemporal prediction units, where n is a positive integer greater than or equal to 2;

[0261] Correspondingly, the first prediction module 706 is further configured as follows:

[0262] Based on the frequency domain data of the spatiotemporal sampling data, the first spatiotemporal prediction data is predicted using the first spatiotemporal prediction unit; based on the spatiotemporal sampling data and the (i-1)th spatiotemporal prediction data, the i-th spatiotemporal prediction data is predicted using the i-th spatiotemporal prediction unit, where 2≤i≤n, and the (i-1)th spatiotemporal prediction data is the output of the (i-1)th spatiotemporal prediction unit; the i-th spatiotemporal prediction data is determined as the target spatiotemporal prediction data.

[0263] Optionally, the first prediction module 706 is further configured as follows:

[0264] Fourier network operators are used to process the frequency domain data of the spatiotemporal sampling data to obtain the first spatiotemporal prediction data.

[0265] Optionally, the first mapping module 704 is further configured as follows:

[0266] The spatiotemporal sampled data is mapped to the frequency domain in both spatial and temporal dimensions to obtain the spatiotemporal frequency domain data, and the spatiotemporal sampled data is mapped to the frequency domain in both spatial and temporal dimensions to obtain the spatial frequency domain data.

[0267] Correspondingly, the first prediction module 706 is further configured as follows:

[0268] Fourier network operator processing is performed on the spatiotemporal frequency domain data of the spatiotemporal sampling data to obtain the first reference data of the first spatiotemporal region; Fourier network operator processing is performed on the spatial frequency domain data of the spatiotemporal sampling data to obtain the second reference data of the first spatiotemporal region; based on the first reference data and the second reference data of the first spatiotemporal region, the first spatiotemporal prediction data is obtained.

[0269] Optionally, the first prediction module 706 is further configured as follows:

[0270] The (i-1)th spatiotemporal prediction data is mapped to the frequency domain to obtain the corresponding (i-1)th frequency domain data. The (i-1)th frequency domain data is then processed by a Fourier network operator to obtain the first spatiotemporal reference data of the i-th time. The frequency domain data of the spatiotemporal sampling data is processed by a Fourier network operator to obtain the second spatiotemporal reference data of the i-th time. Based on the first spatiotemporal reference data and the second spatiotemporal reference data of the i-th time, the spatiotemporal prediction data of the i-th time is obtained.

[0271] Optionally, the first mapping module 704 is further configured as follows:

[0272] The spatiotemporal sampled data is mapped to the frequency domain in both spatial and temporal dimensions to obtain the spatiotemporal frequency domain data, and the spatiotemporal sampled data is mapped to the frequency domain in both spatial and temporal dimensions to obtain the spatial frequency domain data.

[0273] Correspondingly, the first prediction module 706 is further configured as follows:

[0274] The spatial and temporal dimensions of the (i-1)th spatiotemporal prediction data are mapped to the frequency domain to obtain the corresponding (i-1)th spatiotemporal frequency domain data. The (i-1)th spatiotemporal frequency domain data is then processed by a Fourier network operator to obtain the first reference data of the i-th spatiotemporal region. The spatial frequency domain data of the spatiotemporal sampling data is then processed by a Fourier network operator to obtain the second reference data of the i-th spatiotemporal region.

[0275] Optionally, the first prediction module 706 is further configured as follows:

[0276] Using the first frequency domain mapping matrix, the spatial and temporal dimensions of the (i-1)th spatiotemporal prediction data are mapped in the frequency domain to obtain the corresponding (i-1)th spatiotemporal frequency domain data. The first frequency domain mapping matrix is ​​the frequency domain mapping matrix obtained by adding the spatial and temporal dimensions.

[0277] Optionally, the first mapping module 704 is further configured as follows:

[0278] Using the first frequency domain mapping matrix, the spatiotemporal sampled data is frequency domain mapped in both spatial and temporal dimensions to obtain the spatiotemporal frequency domain data of the spatiotemporal sampled data. The first frequency domain mapping matrix is ​​the frequency domain mapping matrix obtained by adding the spatial and temporal dimensions.

[0279] Correspondingly, the first prediction module 706 is further configured as follows:

[0280] Using the second frequency domain mapping matrix, the spatiotemporal sampling data is frequency domain mapped in spatial dimensions to obtain the spatial frequency domain data of the spatiotemporal sampling data. The second frequency domain mapping matrix is ​​the frequency domain mapping matrix obtained by adding the spatial dimensions.

[0281] Optionally, the device further includes:

[0282] The pre-training module is configured to acquire a sample set, which includes multiple sample groups, each sample group including frequency domain data of sample spatiotemporal data and label spatiotemporal data; extract a first sample group from the sample set, wherein the first sample group is any sample group, and the first sample group includes frequency domain data of first sample spatiotemporal data and first label spatiotemporal data; predict sample spatiotemporal prediction data using a spatiotemporal prediction model based on the frequency domain data of the first sample spatiotemporal data; calculate a loss value based on the sample spatiotemporal prediction data and label spatiotemporal data; adjust the model parameters of the spatiotemporal prediction model based on the loss value, and return to execute the step of extracting the first sample group from the sample set. Under the condition of satisfying the preset training termination condition, the trained spatiotemporal prediction model is obtained.

[0283] Optionally, the device further includes:

[0284] The fine-tuning module is configured to send target spatiotemporal prediction data to the user; receive feedback information sent by the user, wherein the feedback information is generated based on the user's editing operations on the target spatiotemporal prediction data; and fine-tune the model parameters of the spatiotemporal prediction model based on the feedback information.

[0285] In the embodiments of this specification, spatiotemporal sampling data of the target task is acquired. This spatiotemporal sampling data is spatially distributed time-series data. Frequency domain mapping is performed on the spatiotemporal sampling data to obtain frequency domain data. Based on this frequency domain data, a pre-trained spatiotemporal prediction model is used to predict the target spatiotemporal prediction data of the target task. This target spatiotemporal prediction data is spatially distributed time-series data. By performing frequency domain mapping on the spatiotemporal sampling data of the target task, frequency domain data representing global information is obtained. Then, based on this frequency domain data, the pre-trained spatiotemporal prediction model is used to predict the target spatiotemporal prediction data of the target task, thereby improving the accuracy of the predicted target spatiotemporal prediction data.

[0286] The above is a schematic scheme of a spatiotemporal data prediction device according to this embodiment. It should be noted that the technical solution of this spatiotemporal data prediction device and the technical solution of the above-described spatiotemporal data prediction method belong to the same concept. For details not described in detail in the technical solution of the spatiotemporal data prediction device, please refer to the description of the technical solution of the above-described spatiotemporal data prediction method.

[0287] Corresponding to the above method embodiments, this specification also provides embodiments of weather data forecasting devices. Figure 9 A schematic diagram of a weather data forecasting device according to one embodiment of this specification is shown. Figure 9 As shown, the device includes:

[0288] The second acquisition module 802 is configured to acquire weather sampling data for the weather forecasting task, wherein the weather sampling data is spatially distributed weather time series data;

[0289] The second mapping module 804 is configured to perform frequency domain mapping on the weather sampling data to obtain the frequency domain data of the weather sampling data.

[0290] The second prediction module 806 is configured to predict weather prediction data for the weather prediction task based on the frequency domain data of the weather sampling data and using a pre-trained spatiotemporal prediction model. The weather prediction data is spatially distributed weather time series data.

[0291] In the embodiments of this specification, weather sampling data for a weather forecasting task is acquired. This weather sampling data is spatially distributed time-series weather data. Frequency domain mapping is performed on the weather sampling data to obtain its frequency domain data. Based on this frequency domain data, a pre-trained spatiotemporal prediction model is used to predict the weather forecast data for the weather forecasting task. This method improves the accuracy of the predicted weather data by obtaining frequency domain data representing global information from the weather sampling data and then using the pre-trained spatiotemporal prediction model based on this global information representation.

[0292] The above is an illustrative scheme of a weather data forecasting device according to this embodiment. It should be noted that the technical solution of this weather data forecasting device and the technical solution of the weather data forecasting method described above belong to the same concept. For details not described in detail in the technical solution of the weather data forecasting device, please refer to the description of the technical solution of the weather data forecasting method described above.

[0293] Corresponding to the above method embodiments, this specification also provides embodiments of an energy data prediction device. Figure 10 A schematic diagram of an energy data prediction device according to one embodiment of this specification is shown. Figure 10 As shown, the device includes:

[0294] The third acquisition module 902 is configured to acquire energy generation data of the energy generation task, wherein the energy generation data is spatially distributed energy time-series data;

[0295] The third mapping module 904 is configured to perform frequency domain mapping on the energy generation data to obtain the frequency domain data of the energy generation data.

[0296] The third prediction module 906 is configured to predict energy generation prediction data for the energy generation task based on the frequency domain data of the energy generation data and using a pre-trained spatiotemporal prediction model. The energy generation prediction data is spatially distributed energy time series data.

[0297] In the embodiments of this specification, energy generation data for an energy generation task is acquired. This energy generation data is spatially distributed time-series energy data. Frequency domain mapping is performed on the energy generation data to obtain frequency domain data. Based on this frequency domain data, a pre-trained spatiotemporal prediction model is used to predict the energy generation forecast data for the energy generation task. This predicted energy generation data is also spatially distributed time-series energy data. Similarly, by performing frequency domain mapping on weather sampling data for a weather forecast task, frequency domain data representing global information is obtained. Then, based on this frequency domain data, a pre-trained spatiotemporal prediction model is used to predict weather forecast data, thereby improving the accuracy of the predicted weather forecast data.

[0298] The above is an illustrative scheme of an energy data prediction device according to this embodiment. It should be noted that the technical solution of this energy data prediction device and the technical solution of the energy data prediction method described above belong to the same concept. For details not described in detail in the technical solution of the energy data prediction device, please refer to the description of the technical solution of the energy data prediction method described above.

[0299] Corresponding to the above method embodiments, this specification also provides embodiments of a data processing apparatus for spatiotemporal data prediction. Figure 11 This specification illustrates a schematic diagram of a data processing apparatus for spatiotemporal data prediction according to one embodiment of the present specification. This apparatus is applied to cloud-side devices, such as… Figure 11 As shown, the device includes:

[0300] The fourth acquisition module 1002 is configured to acquire a sample set, wherein the sample set includes multiple sample groups, and any sample group includes frequency domain data of sample spatiotemporal data and label spatiotemporal data;

[0301] Extraction module 1004 is configured to extract a first sample group from the sample set, wherein the first sample group is any sample group, and the first sample group includes frequency domain data of the first sample spatiotemporal data and spatiotemporal data of the first label;

[0302] The fourth prediction module 1006 is configured to predict the sample spatiotemporal prediction data based on the frequency domain data of the first sample spatiotemporal data using a spatiotemporal prediction model.

[0303] The calculation module 1008 is configured to calculate the loss value based on the sample spatiotemporal prediction data and the label spatiotemporal data;

[0304] The training module 1010 is configured to adjust the model parameters of the spatiotemporal prediction model according to the loss value, return to the step of extracting the first sample group from the sample set, and obtain the spatiotemporal prediction model after training is completed if the preset training termination condition is met.

[0305] The sending module 1012 is configured to send the model parameters of the trained spatiotemporal prediction model to the edge device.

[0306] In this embodiment, a sample set is obtained, which includes multiple sample groups. Each sample group includes frequency domain data of sample spatiotemporal data and label spatiotemporal data. A first sample group is extracted from the sample set, which is any sample group and includes frequency domain data of first sample spatiotemporal data and first label spatiotemporal data. Based on the frequency domain data of the first sample spatiotemporal data, a spatiotemporal prediction model is used to predict sample spatiotemporal prediction data. Based on the sample spatiotemporal prediction data and label spatiotemporal data, a loss value is calculated. Based on the loss value, the model parameters of the spatiotemporal prediction model are adjusted. The step of extracting the first sample group from the sample set is returned. If the preset training termination condition is met, the trained spatiotemporal prediction model is obtained. The model parameters of the trained spatiotemporal prediction model are sent to the edge device. Based on the frequency domain data of the first sample spatiotemporal data, the spatiotemporal prediction model is used to predict the sample spatiotemporal prediction data. The loss value is calculated by combining the label spatiotemporal data, and the spatiotemporal prediction model is trained based on the loss value. This allows the spatiotemporal prediction model to accurately predict high-accuracy spatiotemporal prediction data based on the global information of the frequency domain data. The above training is completed through cloud-side devices, which saves training costs for edge devices and improves training efficiency.

[0307] The above is an illustrative scheme of a data processing device for spatiotemporal data prediction according to this embodiment. It should be noted that the technical solution of this data processing device for spatiotemporal data prediction belongs to the same concept as the technical solution of the data processing method for spatiotemporal data prediction described above. For details not described in detail in the technical solution of the data processing device for spatiotemporal data prediction, please refer to the description of the technical solution of the data processing method for spatiotemporal data prediction described above.

[0308] Figure 12 A structural block diagram of a computing device according to one embodiment of this specification is shown. The components of the computing device 1100 include, but are not limited to, a memory 1110 and a processor 1120. The processor 1120 is connected to the memory 1110 via a bus 1130, and a database 1150 is used to store data.

[0309] The computing device 1100 also includes an access device 1140, which enables the computing device 1100 to communicate via one or more networks 1160. Examples of these networks include PSTN (Public Switched Telephone Network), LAN (Local Area Network), WAN (Wide Area Network), PAN (Personal Area Network), or combinations of communication networks such as the Internet. The access device 1140 may include one or more of any type of wired or wireless network interface (e.g., NIC (Network Interface Controller)), such as an IEEE 802.12 WLAN (Wireless Local Area Networks) wireless interface, a Wi-MAX (World Interoperability for Microwave Access) interface, an Ethernet interface, a USB (Universal Serial Bus) interface, a cellular network interface, a Bluetooth interface, an NFC (Near Field Communication) interface, and so on.

[0310] In one embodiment of this specification, the aforementioned components of the computing device 1100 and Figure 12 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 12 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0311] The computing device 1100 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or PCs (Personal Computers). The computing device 1100 can also be a mobile or stationary server.

[0312] The processor 1120 is configured to execute the following computer-executable instructions, which, when executed by the processor, implement the steps of the aforementioned spatiotemporal data prediction method, weather data prediction method, energy data prediction method, or spatiotemporal data prediction data processing method.

[0313] The above is an illustrative scheme of a computing device according to this embodiment. It should be noted that the technical solution of this computing device belongs to the same concept as the technical solutions of the above-mentioned spatiotemporal data prediction method, weather data prediction method, energy data prediction method, and spatiotemporal data prediction data processing method. For details not described in detail in the technical solution of the computing device, please refer to the descriptions of the technical solutions of the above-mentioned spatiotemporal data prediction method, weather data prediction method, energy data prediction method, or spatiotemporal data prediction data processing method.

[0314] An embodiment of this specification also provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the above-described spatiotemporal data prediction method, weather data prediction method, energy data prediction method, or spatiotemporal data prediction data processing method.

[0315] The above is an illustrative scheme of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium belongs to the same concept as the technical solutions of the aforementioned spatiotemporal data prediction method, weather data prediction method, energy data prediction method, and spatiotemporal data prediction data processing method. Details not described in detail in the technical solution of the storage medium can be found in the descriptions of the technical solutions of the aforementioned spatiotemporal data prediction method, weather data prediction method, energy data prediction method, or spatiotemporal data prediction data processing method.

[0316] An embodiment of this specification also provides a computer program, wherein when the computer program is executed in a computer, it causes the computer to perform the steps of the above-described spatiotemporal data prediction method, weather data prediction method, energy data prediction method, or spatiotemporal data prediction data processing method.

[0317] The above is an illustrative scheme of a computer program according to this embodiment. It should be noted that the technical solution of this computer program belongs to the same concept as the technical solutions of the above-mentioned spatiotemporal data prediction method, weather data prediction method, energy data prediction method, and spatiotemporal data prediction data processing method. For details not described in detail in the technical solution of the computer program, please refer to the descriptions of the technical solutions of the above-mentioned spatiotemporal data prediction method, weather data prediction method, energy data prediction method, or spatiotemporal data prediction data processing method.

[0318] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0319] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0320] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0321] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0322] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A spatiotemporal data prediction method, comprising: Acquire spatiotemporal sampling data for the target task, wherein the spatiotemporal sampling data is spatially distributed time-series data, the spatiotemporal sampling data has time and spatial dimensions, and the spatiotemporal sampling data includes weather sampling data or energy generation data; The spatiotemporal sampling data is frequency-domain mapped to obtain the frequency domain data of the spatiotemporal sampling data. The frequency domain data includes spatiotemporal frequency domain data and spatial frequency domain data. The spatiotemporal frequency domain data is obtained by frequency-domain mapping of the spatiotemporal sampling data based on the time dimension and the spatial dimension. The spatial frequency domain data is obtained by frequency-domain mapping of the spatiotemporal sampling data based on the spatial dimension. Based on the frequency domain data, a pre-trained spatiotemporal prediction model is used to predict the target spatiotemporal prediction data of the target task, wherein the target spatiotemporal prediction data is spatially distributed time-series data, and the spatial frequency domain data provides a reference in the prediction.

2. The method according to claim 1, wherein the pre-trained spatiotemporal prediction model comprises n sequentially connected spatiotemporal prediction units, wherein, n is a positive integer greater than or equal to 2; The step of predicting target spatiotemporal prediction data based on the frequency domain data using a pre-trained spatiotemporal prediction model includes: Based on the frequency domain data of the spatiotemporal sampling data, the first spatiotemporal prediction data is predicted using the first spatiotemporal prediction unit. Based on the spatiotemporal sampling data and the (i-1)th spatiotemporal prediction data, the i-th spatiotemporal prediction data is predicted using the i-th spatiotemporal prediction unit, where 2≤i≤n, and the (i-1)th spatiotemporal prediction data is the output of the (i-1)th spatiotemporal prediction unit. The i-th spatiotemporal prediction data is determined as the target spatiotemporal prediction data.

3. The method according to claim 2, wherein performing frequency domain mapping of the spatiotemporal sampled data in spatial and temporal dimensions to obtain the spatiotemporal frequency domain data of the spatiotemporal sampled data includes: Using the first frequency domain mapping matrix, the spatiotemporal sampled data is frequency domain mapped in both spatial and temporal dimensions to obtain the spatiotemporal frequency domain data of the spatiotemporal sampled data. The first frequency domain mapping matrix is ​​a frequency domain mapping matrix obtained by adding the spatial and temporal dimensions. Accordingly, the step of performing a spatial-dimensional frequency domain mapping on the spatiotemporal sampled data to obtain the spatial frequency domain data of the spatiotemporal sampled data includes: Using the second frequency domain mapping matrix, the spatiotemporal sampled data is frequency domain mapped in spatial dimensions to obtain the spatial frequency domain data of the spatiotemporal sampled data. The second frequency domain mapping matrix is ​​a frequency domain mapping matrix obtained by adding the spatial dimensions.

4. The method according to claim 2, wherein predicting the first spatiotemporal prediction data using the first spatiotemporal prediction unit based on the frequency domain data of the spatiotemporal sampling data comprises: The spatiotemporal frequency domain data and the spatial frequency domain data are processed by Fourier network operators to obtain the first spatiotemporal prediction data.

5. The method according to claim 4, wherein performing Fourier network operator processing on the spatiotemporal frequency domain data and the spatial frequency domain data to obtain the first spatiotemporal prediction data includes: The spatiotemporal frequency domain data of the spatiotemporal sampling data are processed by Fourier network operators to obtain the first spatiotemporal first reference data; The spatial frequency domain data of the spatiotemporal sampling data is processed by Fourier network operators to obtain the first spatiotemporal second reference data; Based on the first reference data and the second reference data of the first spacetime, the prediction data of the first spacetime is obtained.

6. The method according to any one of claims 2-5, wherein predicting the i-th spatiotemporal prediction data using the i-th spatiotemporal prediction unit based on the spatiotemporal sampling data and the (i-1)-th spatiotemporal prediction data comprises: The i-1th spatiotemporal prediction data is mapped in the frequency domain to obtain the corresponding i-1th frequency domain data, and the i-1th frequency domain data is processed by Fourier network operators to obtain the i-th spatiotemporal first reference data. The frequency domain data of the spatiotemporal sampling data is processed by Fourier network operators to obtain the i-th spatiotemporal second reference data; Based on the first reference data and the second reference data of the i-th spacetime, the predicted data of the i-th spacetime is obtained.

7. The method according to claim 6, wherein the step of performing frequency domain mapping on the (i-1)th spatiotemporal prediction data to obtain the corresponding (i-1)th frequency domain data, and performing Fourier network operator processing on the (i-1)th frequency domain data to obtain the i-th spatiotemporal first reference data, comprises: The (i-1)th spatiotemporal prediction data is mapped to the frequency domain in both spatial and temporal dimensions to obtain the corresponding (i-1)th spatiotemporal frequency domain data. The (i-1)th spatiotemporal frequency domain data is then processed by a Fourier network operator to obtain the first reference data in the i-th spatiotemporal space. Accordingly, the step of processing the frequency domain data of the spatiotemporal sampled data using Fourier network operators to obtain the i-th spatiotemporal second reference data includes: The spatial frequency domain data of the spatiotemporal sampling data is processed by Fourier network operators to obtain the i-th spatiotemporal second reference data.

8. The method according to claim 7, wherein performing frequency domain mapping of the (i-1)th spatiotemporal prediction data in spatial and temporal dimensions to obtain the corresponding (i-1)th spatiotemporal frequency domain data includes: Using the first frequency domain mapping matrix, the (i-1)th spatiotemporal prediction data is frequency domain mapped in both spatial and temporal dimensions to obtain the corresponding (i-1)th spatiotemporal frequency domain data. The first frequency domain mapping matrix is ​​a frequency domain mapping matrix obtained by adding the spatial and temporal dimensions.

9. The method according to claim 1, further comprising, before predicting the target spatiotemporal prediction data based on the frequency domain data using a pre-trained spatiotemporal prediction model: Obtain a sample set, wherein the sample set includes multiple sample groups, and any sample group includes frequency domain data of sample spatiotemporal data and label spatiotemporal data; Extract a first sample group from the sample set, wherein the first sample group is any sample group, and the first sample group includes the frequency domain data of the first sample spatiotemporal data and the first label spatiotemporal data; Based on the frequency domain data of the first sample spatiotemporal data, the spatiotemporal prediction data of the sample is predicted using a spatiotemporal prediction model. The loss value is calculated based on the sample spatiotemporal prediction data and the label spatiotemporal data; Based on the loss value, the model parameters of the spatiotemporal prediction model are adjusted, and the step of extracting the first sample group from the sample set is returned to be executed. Under the condition of satisfying the preset training termination condition, the spatiotemporal prediction model that has been trained is obtained.

10. The method according to claim 1, further comprising, after obtaining the target spatiotemporal prediction data through prediction: Send the target spatiotemporal prediction data to the user; Receive feedback information sent by the user, wherein the feedback information is generated based on the user's editing operation on the target spatiotemporal prediction data; Based on the feedback information, the model parameters of the spatiotemporal prediction model are adjusted.

11. A weather data forecasting method, comprising: Acquire weather sampling data for weather forecasting tasks, wherein the weather sampling data is spatially distributed time-series weather data, and the weather sampling data has time and spatial dimensions; The weather sampling data is frequency-domain mapped to obtain frequency domain data of the weather sampling data. The frequency domain data includes spatiotemporal frequency domain data and spatial frequency domain data. The spatiotemporal frequency domain data is obtained by frequency-domain mapping of the weather sampling data based on the time dimension and the spatial dimension. The spatial frequency domain data is obtained by frequency-domain mapping of the weather sampling data based on the spatial dimension. Based on the frequency domain data of the weather sampling data, a pre-trained spatiotemporal prediction model is used to predict the weather prediction data for the weather prediction task. The weather prediction data is spatially distributed time-series weather data, and the spatial frequency domain data provides a reference in the prediction.

12. An energy data forecasting method, comprising: Acquire energy generation data for energy generation tasks, wherein the energy generation data is spatially distributed time-series energy data, and the energy generation data has time and spatial dimensions; The energy generation data is frequency-domain mapped to obtain the frequency domain data of the energy generation data. The frequency domain data includes spatiotemporal frequency domain data and spatial frequency domain data. The spatiotemporal frequency domain data is obtained by frequency-domain mapping of the energy generation data based on the time dimension and the spatial dimension. The spatial frequency domain data is obtained by frequency-domain mapping of the energy generation data based on the spatial dimension. Based on the frequency domain data of the energy generation data, a pre-trained spatiotemporal prediction model is used to predict the energy generation prediction data for the energy generation task. The energy generation prediction data is spatially distributed energy time-series data, and the spatial frequency domain data provides a reference in the prediction.

13. A data processing method for spatiotemporal data prediction, applied to cloud-side equipment, comprising: A sample set is obtained, wherein the sample set includes multiple sample groups, each sample group includes frequency domain data of sample spatiotemporal data and label spatiotemporal data, the sample spatiotemporal data has a time dimension and a spatial dimension, the frequency domain data includes spatiotemporal frequency domain data and spatial frequency domain data, the spatiotemporal frequency domain data is obtained by frequency domain mapping of the sample spatiotemporal data based on the time dimension and the spatial dimension, the spatial frequency domain data is obtained by frequency domain mapping of the sample spatiotemporal data based on the spatial dimension, and the sample spatiotemporal data includes weather sampling data or energy generation data; Extract a first sample group from the sample set, wherein the first sample group is any sample group, and the first sample group includes the frequency domain data of the first sample spatiotemporal data and the first label spatiotemporal data; Based on the frequency domain data of the first sample spatiotemporal data, a spatiotemporal prediction model is used to predict the sample spatiotemporal prediction data, wherein the spatial frequency domain data provides a reference in the prediction. The loss value is calculated based on the sample spatiotemporal prediction data and the label spatiotemporal data; Based on the loss value, adjust the model parameters of the spatiotemporal prediction model, return to the step of extracting the first sample group from the sample set, and obtain the spatiotemporal prediction model after training is completed if the preset training termination condition is met. The model parameters of the trained spatiotemporal prediction model are sent to the edge device.

14. A data processing system for spatiotemporal data prediction, comprising: End-side devices are used to send training requests for spatiotemporal prediction models to cloud-side devices; The cloud-side device is configured to, upon receiving the training request, acquire a sample set, wherein the sample set includes multiple sample groups, each sample group including frequency domain data and label spatiotemporal data of sample spatiotemporal data, the sample spatiotemporal data having time and spatial dimensions, the frequency domain data including spatiotemporal frequency domain data and spatial frequency domain data, the spatiotemporal frequency domain data being obtained by frequency domain mapping of the sample spatiotemporal data based on the time and spatial dimensions, the spatial frequency domain data being obtained by frequency domain mapping of the sample spatiotemporal data based on the spatial dimension, and the sample spatiotemporal data including weather sampling data or energy generation data; and to extract a first sample group from the sample set, wherein the first sample group... This group is any sample group. The first sample group includes the frequency domain data of the first sample spatiotemporal data and the first label spatiotemporal data. Based on the frequency domain data of the first sample spatiotemporal data, a spatiotemporal prediction model is used to predict the sample spatiotemporal prediction data, wherein the spatial frequency domain data provides a reference in the prediction. Based on the sample spatiotemporal prediction data and the label spatiotemporal data, a loss value is calculated. Based on the loss value, the model parameters of the spatiotemporal prediction model are adjusted, and the step of extracting the first sample group from the sample set is returned to be executed. Under the condition of satisfying the preset training end condition, the trained spatiotemporal prediction model is obtained. The model parameters of the trained spatiotemporal prediction model are sent to the edge device. The end-side device is also used to receive the model parameters of the spatiotemporal prediction model.

15. A computing device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the spatiotemporal data prediction method according to any one of claims 1 to 10, the weather data prediction method according to claim 11, the energy data prediction method according to claim 12, or the data processing method for spatiotemporal data prediction according to claim 13.

16. A computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the spatiotemporal data prediction method of any one of claims 1 to 10, the weather data prediction method of claim 11, the energy data prediction method of claim 12, or the data processing method for spatiotemporal data prediction of claim 13.

Images