A deep learning prediction method and system for multi-source spatio-temporal grid data

Abstract

The application relates to the technical field of deep learning, and discloses a deep learning prediction method and system for multi-source space-time grid point data, which comprises the following steps: acquiring live grid point data, forecast grid point data and static grid point data, and preprocessing to obtain unified grid point feature data. Missing mask grid point data is generated, confidence grid point data is generated according to observation coverage information, interpolation distance information, space-time consistency test results and forecast time decay rules, and model input grid point data is formed. Training samples are constructed and forecast time coding data is generated. A deep learning prediction model is constructed and trained, and the model comprises a space coding network, a time coding network and a forecast time segmented prediction network. In the training stage, confidence weighted regression loss is adopted. In the reasoning stage, the prediction result is obtained. The application improves the learning stability of multi-source grid point data under the conditions of missing and filling, and improves the reliability of continuous value prediction and graded alarm probability under different forecast times.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, in particular to a deep learning prediction method and system for multi-source spatiotemporal grid data. BACKGROUND

[0002] With the development of Internet of Things perception, remote sensing inversion, digital twinning and business data convergence, a large amount of spatiotemporal grid data products are generated in scenarios such as road traffic, city operation, environmental monitoring, energy scheduling and near-earth risk warning. Such data usually contains real-time grid data obtained by sensing network or fusion analysis, forecast grid data generated by numerical simulation or statistical model, and static grid data such as terrain, surface, administrative division and road network. In order to output continuous value prediction and probability results required by hierarchical warning under the conditions of hourly update and high spatial resolution, the industry gradually adopts computer- implemented prediction methods based on deep learning to jointly model and correct multi-source spatiotemporal grid data.

[0003] In the prior art, for example, CN111639787A discloses a spatiotemporal data prediction method based on graph convolutional neural network, which outputs prediction results by constructing node relationships and extracting spatiotemporal features. Although this scheme can improve prediction accuracy using deep learning, it usually inputs the grid points after missing data filling as the same as the original effective grid points, lacks mask representation of missing data source traceability, observation coverage information, interpolation distance information, spatiotemporal consistency test results and forecast time decay information, and confidence modeling driven by joint driving. The training stage is prone to propagate filling errors; and it usually uses a single structure to uniformly fit across the forecast time, lacks a gated fusion mechanism segmented by the forecast time, and is prone to performance degradation in a specific forecast time segment; meanwhile, the hierarchical probability output lacks sequence consistency control and calibration by the forecast time, resulting in unstable hierarchical warning and insufficient probability reliability.

[0004] Therefore, it is necessary to design a deep learning prediction method and system for multi-source spatiotemporal grid data to solve the problems in the current technology. SUMMARY

[0005] In view of this, the present application provides a deep learning prediction method and system for multi-source spatiotemporal grid data, aiming to solve the problems in the prior art that the spatial coordinates and time reference of multi-source spatiotemporal grid data are not unified, missing data filling introduces uncertainty and is difficult to quantify, prediction performance is not balanced under different forecast times, and hierarchical warning probability is not reliable and prone to sequence conflicts.

[0006] The present application provides a deep learning prediction method for multi-source spatiotemporal grid data, comprising:

[0007] Obtaining live grid data, forecast grid data and static grid data, preprocessing the data to obtain uniform grid feature data, the preprocessing including unifying spatial coordinates and time reference;

[0008] Generating missing mask grid data according to the missing or filling mark, generating confidence grid data according to observation coverage information, interpolation distance information, spatio-temporal consistency test results and forecast time decay rules, and splicing the confidence grid data and the uniform grid feature data to obtain model input grid data;

[0009] Based on the sliding time window, a training sample is constructed, so that the training sample contains the model input grid data and the forecast time encoding data;

[0010] A deep learning prediction model is constructed and trained, the deep learning prediction model including a spatial encoding network, a time encoding network and a forecast time segmentation prediction network, the forecast time segmentation prediction network including a gating network and a plurality of expert prediction sub-networks, the gating network weighting and fusing the plurality of expert prediction sub-networks to output continuous value prediction results, and outputting hierarchical probability outputs based on a threshold set;

[0011] During training, the confidence weighted regression loss is performed on the continuous value prediction results, and the sequence relationship consistency constraint is applied to the hierarchical probability outputs;

[0012] During inference, the continuous value prediction results and the hierarchical probability outputs are output, and the hierarchical probability outputs are probability calibrated to obtain the published prediction results.

[0013] Further, when preprocessing the data to obtain uniform grid feature data, it includes:

[0014] The target regular latitude and longitude grid is set as the uniform spatial coordinate, and the hourly whole point is set as the uniform time reference;

[0015] The live grid data, forecast grid data and static grid data are converted to the uniform spatial coordinate; the numerical element is resampled to the target regular latitude and longitude grid by bilinear interpolation, and the categorical element is resampled to the target regular latitude and longitude grid by nearest neighbor interpolation; the forecast grid data is converted to the effective time according to the start time and the forecast time information, and is aligned to the uniform time reference; the live grid data is taken within thirty minutes before and after the uniform time reference, and if there are multiple values, the average value is taken; finally, the uniform grid feature data is output.

[0016] Further, when generating the missing mask grid data, it includes:

[0017] The missing or filling marks are read by element and grid point for the unified grid feature data; when the missing or filling marks indicate original valid values, a valid mask layer is generated; when the missing or filling marks indicate interpolation filling values, an interpolation mask layer is generated; when the missing or filling marks indicate backfilling filling values, a backfilling mask layer is generated; and the valid mask layer, the interpolation mask layer and the backfilling mask layer are combined according to the element correspondence to form the missing mask grid data.

[0018] Further, when the confidence grid data is generated and the model input grid data is obtained, the following steps are included:

[0019] The observation coverage information is determined based on the valid grid point proportion of the valid mask layer in a preset spatial neighborhood; the interpolation distance information is determined based on the minimum distance from an observation point to a target grid point; and the spatial neighborhood consistency test and the time adjacent consistency test are performed on the unified grid feature data to obtain the spatiotemporal consistency test result;

[0020] The timeliness decay information is determined according to the prediction timeliness information of the prediction grid data and according to the prediction timeliness decay rule; the confidence grid data is generated according to the observation coverage information, the interpolation distance information, the spatiotemporal consistency test result and the timeliness decay information according to the configured fusion rule, and the confidence grid data is punished according to the interpolation mask layer and the backfilling mask layer; and the confidence grid data after the punishment and the unified grid feature data are spliced according to the channel to obtain the model input grid data.

[0021] Further, when the training sample is constructed based on the sliding time window, the following steps are included:

[0022] Each target prediction time under the unified time reference is taken as an alignment point, and the model input grid data is selected forward according to the number of continuous whole point times corresponding to a preset window length to form a model input grid data sequence, and the sliding time window is rolled with a whole point time as a step;

[0023] For each training sample, the corresponding prediction timeliness level is determined according to the start time and the valid time of the prediction grid data, and the prediction timeliness level is converted into prediction timeliness code data and written into the training sample; the training samples are grouped according to the prediction timeliness level, and each group is equally sampled to obtain a training sample set.

[0024] Further, when the deep learning prediction model is constructed, the following steps are included:

[0025] The spatial coding network is configured to perform multi-scale down-sampling on the model input grid data of a single integral time point to extract spatial features, and to output spatial coding features by jumping and fusing with an up-sampling branch of a corresponding scale; the temporal coding network is configured to perform time series coding on the spatial coding features corresponding to the model input grid data sequence, and the prediction lead time coding data is taken as a gating input for time series coding to modulate the time series weight of the spatial coding features under different prediction lead time levels, and to output time series coding features.

[0026] Further, when constructing the prediction lead time segmented prediction network, the following steps are included:

[0027] At least three expert prediction sub-networks are set, respectively corresponding to short-time prediction lead time levels, medium-time prediction lead time levels and long-time prediction lead time levels; the gating network is configured to generate fusion weights of each expert prediction sub-network based on the time series coding features and the prediction lead time coding data, and to perform normalization processing on the fusion weights; the outputs of each expert prediction sub-network are weighted and fused according to the fusion weights to obtain continuous value prediction results; and based on the continuous value prediction results, a hierarchical probability output corresponding one-to-one to the threshold set is generated through a hierarchical output layer, and a monotonic constraint according to the threshold set order is applied to the hierarchical probability output.

[0028] Further, when training the deep learning prediction model, the following steps are included:

[0029] Regression weights are set for each grid corresponding to the continuous value prediction results according to the confidence grid data, and upper and lower limits are set for the regression weights to form truncated weights; a regression loss between the continuous value prediction results and the training labels is calculated based on the truncated weights; the hierarchical probability output is subjected to sequence relationship consistency verification according to the ordering relationship of the threshold set, and when the hierarchical probability outputs corresponding to adjacent thresholds do not satisfy the order of the threshold set, a sequence conflict penalty term is generated; the model parameters of the deep learning prediction model are updated based on the joint loss of the regression loss and the sequence conflict penalty term.

[0030] Further, when reasoning and generating the publishing prediction results, the following steps are included:

[0031] The model input grid data of the to-be-predicted period and the corresponding prediction lead time coding data are input into the trained deep learning prediction model to output continuous value prediction results and hierarchical probability outputs.

[0032] statistical matching error of the hierarchical probability output and the training label on the verification sample, determine the calibration parameter, and set the calibration parameter according to the forecast time limit level respectively; the hierarchical probability output is calibrated by using the calibration parameter corresponding to the forecast time limit level to obtain the calibrated hierarchical probability output; consistency verification is performed based on the calibrated hierarchical probability output and the continuous value prediction result, and the publishing prediction result is output when the consistency verification passes.

[0033] Compared with the prior art, the beneficial effects of the present application are that: the missing data source is traceable through the missing mask grid data, the observation coverage, the interpolation distance, the spatio-temporal consistency and the time limit decay are uniformly represented as learnable reliability information through the confidence grid data, and the confidence weighted suppression is used to inhibit the interference of filling error on the model parameters in the training stage; at the same time, the adaptive ability under different forecast time limit levels is improved based on the fusion mechanism of the expert prediction sub-network and the gate network of the forecast time limit segmentation, and the sequence relationship consistency constraint and the probability calibration according to the forecast time limit level are introduced at the hierarchical probability output end, so that the continuous value prediction and the hierarchical alarm probability are more stable and more reliable in the publishing link, thereby improving the identification sensitivity and business availability of key events such as low visibility.

[0034] On the other hand, the present application also provides a deep learning prediction system for multi-source spatio-temporal grid data, which is used for applying the deep learning prediction method for multi-source spatio-temporal grid data, comprising:

[0035] The acquisition unit is configured to obtain live grid data, forecast grid data and static grid data, and to obtain uniform grid feature data by preprocessing the data, wherein the preprocessing includes unifying spatial coordinates and time reference;

[0036] The processing unit is configured to generate missing mask grid data according to missing or filling marks, and to generate confidence grid data according to observation coverage information, interpolation distance information, spatio-temporal consistency test results and forecast time limit decay rules, and to splice the confidence grid data and the uniform grid feature data to obtain model input grid data;

[0037] The construction unit is configured to construct training samples based on a sliding time window, so that the training samples contain the model input grid data sequence and the forecast time limit encoding data;

[0038] The training unit is configured to build and train a deep learning prediction model, the deep learning prediction model comprising a spatial coding network, a temporal coding network and a forecast time limit segmentation prediction network, the forecast time limit segmentation prediction network comprising a gating network and a plurality of expert prediction sub-networks, the gating network weighting and fusing the plurality of expert prediction sub-networks to output a continuous value prediction result according to the forecast time limit coding data, and outputting a hierarchical probability output based on a threshold set; during training, the continuous value prediction result is subjected to a confidence weighted regression loss, and the hierarchical probability output is subjected to a sequence relationship consistency constraint.

[0039] The inference unit is configured to output the continuous value prediction result and the hierarchical probability output during inference, and to obtain a published prediction result by performing probability calibration on the hierarchical probability output.

[0040] It can be understood that the deep learning prediction method and system for multi-source spatio-temporal grid data have the same beneficial effects, which will not be repeated here. BRIEF DESCRIPTION OF DRAWINGS

[0041] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The accompanying drawings are included to provide a description of the preferred embodiments and are not intended to limit the scope of the present application. Moreover, the same reference numerals are used throughout the various drawings to designate identical parts. In the drawings:

[0042] Figure 1 A flowchart of the deep learning prediction method for multi-source spatio-temporal grid data provided by the embodiments of the present application;

[0043] Figure 2 A visibility prediction business technology roadmap and model construction flowchart provided by the embodiments of the present application;

[0044] Figure 3 A residual U-shaped structure diagram provided by the embodiments of the present application;

[0045] Figure 4 A visibility forecast RMSE score diagram provided by the embodiments of the present application;

[0046] Figure 5 A functional block diagram of the deep learning prediction system for multi-source spatio-temporal grid data provided by the embodiments of the present application. DETAILED DESCRIPTION

[0047] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it is understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure can be more thoroughly understood, and so that the scope of the present disclosure can be conveyed to those skilled in the art. It should be noted that the embodiments in the present disclosure and the features in the embodiments can be combined with each other without conflict. The present disclosure will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0048] In some embodiments of the present application, referring to Figure 1 The present application proposes a deep learning prediction method for multi-source spatio-temporal grid data, including:

[0049] S100: Obtain live grid data, forecast grid data and static grid data, and pre-process the data to obtain unified grid feature data, the pre-processing including unifying spatial coordinates and time reference.

[0050] S200: Generate missing mask grid data according to the missing or filling mark, generate confidence grid data according to the observation coverage information, interpolation distance information, spatio-temporal consistency test results and forecast time decay rules, and splice the confidence grid data and the unified grid feature data to obtain model input grid data.

[0051] S300: Construct training samples based on sliding time windows, so that the training samples contain model input grid data sequences and forecast time encoding data.

[0052] S400: Construct and train a deep learning prediction model, the deep learning prediction model including a spatial encoding network, a temporal encoding network and a forecast time segmentation prediction network, the forecast time segmentation prediction network including a gating network and a plurality of expert prediction sub-networks, the gating network weighting and fusing the plurality of expert prediction sub-networks to output continuous value prediction results, and outputting hierarchical probability outputs based on a threshold set.

[0053] S500: During training, perform confidence weighted regression loss on the continuous value prediction results, and impose sequential relationship consistency constraints on the hierarchical probability outputs.

[0054] S600: During inference, output continuous value prediction results and hierarchical probability outputs, and perform probability calibration on the hierarchical probability outputs to obtain published prediction results.

[0055] Specifically, referring to Figure 2As shown, the method of the present application can be organized into an end-to-end technical route in the "fog visibility forecast" business scenario: first, real-time grid data, forecast grid data and static grid data are obtained, and unified spatial coordinates and time reference preprocessing is performed to form unified grid feature data; thereafter, missing mask grid data is generated according to missing or filling marks, and confidence grid data is generated in combination with observation coverage information, interpolation distance information, spatio-temporal consistency test results and forecast time decay rules, the confidence grid data and the unified grid feature data are spliced according to the channel to obtain model input grid data; then, based on a sliding time window, a training sample is constructed, so that the training sample contains model input grid data sequence and forecast time coding data, and training and verification are organized according to forecast time interval; in the model training stage, a deep learning prediction model including spatial coding network, time coding network and forecast time segmented prediction network is constructed, a plurality of expert prediction sub-networks are outputted by a gating network according to time weighting fusion to output continuous value prediction results, and based on a threshold set, a hierarchical probability output is outputted in parallel, at the same time, confidence weighted regression loss is implemented on the continuous value prediction results and sequence relationship consistency constraint is applied to the hierarchical probability output; in the inference and release stage, the continuous value prediction result and the hierarchical probability output are obtained by inputting the model to the prediction period to be predicted, the hierarchical probability output is probability calibrated and consistency checked according to the forecast time interval, and the release prediction result which can be released externally is formed, so that Figure 2 The business link shown and Figure 1 The step flow shown corresponds one by one in the data, model and release control level.

[0056] Specifically, the embodiment faces the nowcasting and short-term prediction of near-surface elements such as visibility, and adopts a computer-implemented deep learning method to uniformly organize and model multi-source spatio-temporal grid data: first, real-time grid data, forecast grid data and static grid data are obtained, and spatio-temporal alignment is completed through unified spatial coordinates and time reference to form uniform grid feature data. Secondly, the missing and filling sources are traceablely characterized at the data level, and missing mask grid data is generated according to the missing or filling marks, and further confidence grid data is generated by synthesizing observation coverage information, interpolation distance information, spatio-temporal consistency test results and forecast time decay rules, so that the model can distinguish different reliability sources such as "original effective", "interpolation filling", "backfill filling", etc. Subsequently, the confidence grid data and the uniform grid feature data are spliced according to the channel to obtain the model input grid data, and the training samples are constructed with a sliding time window, so that each training sample contains the model input grid data sequence and the forecast time coding data. On the model side, the deep learning prediction model is composed of a spatial coding network, a time coding network and a forecast time segmentation prediction network, wherein the forecast time segmentation prediction network outputs continuous value prediction results by weighted fusion of multiple expert prediction sub-networks through a gating network, and outputs hierarchical probability output based on a threshold set. In the training stage, the confidence weighted regression loss is applied to the continuous value prediction result, and the sequential relationship consistency constraint is applied to the hierarchical probability output. In the inference stage, the continuous value prediction result and the hierarchical probability output are output, and the hierarchical probability output is probability calibrated to obtain the published prediction result, so as to improve the stability and reliability of the continuous value and alarm classification in the scene where missing filling is common and different forecast time error characteristics are significantly different.

[0057] Specifically, Guizhou region is selected as an example region, and hourly multi-source fusion ground truth data is used as ground truth grid data, which contains visibility, precipitation, ten-meter wind direction, ten-meter wind speed, two-meter humidity, and two-meter temperature. The spatial resolution is processed to 3 km x 3 km, the spatial range is 100.105 E to 112.104 E, 22.1 N to 31.72 N, and the time sequence is from 00:00 on January 1, 2023 to 23:00 on June 30, 2024. The forecast grid data uses numerical model forecast data with the same range and resolution, and the forecast lead time is from 0 to 72 hours per hour, with the starting time covering from 00:00 on January 1, 2023 to 23:00 on June 30, 2024. Static grid data can be selected from terrain elevation, underlying surface type, etc. (if there is no static field, at least keep the latitude and longitude grid index as a static element to meet the consistency definition of static grid data input). In the data preprocessing stage, set the target regular latitude and longitude grid and hourly whole point as the unified spatial coordinate and unified time reference, and resample and time align the ground truth grid data, forecast grid data and static grid data. For grid elements with missing data, fill them in according to the idea of nearest point interpolation and climate value supplement, and generate missing or filling markers simultaneously when filling in: original effective value is marked as original effective, nearest point interpolation result is marked as interpolation filling, and climate value supplement result is marked as backfill filling. At the same time, normalize each element to improve training effect. On this basis, generate missing mask grid data (effective mask layer, interpolation mask layer, backfill mask layer), and generate confidence grid data: take the effective grid point ratio in the preset spatial neighborhood of the effective mask layer as the observation coverage information. Take the minimum distance from the observation point participating in interpolation to the target grid point as the interpolation distance information. Perform spatial neighborhood consistency test and time adjacent consistency test on the unified grid feature data to get the spatiotemporal consistency test result. According to the forecast lead time information of the forecast grid data and combined with the preset forecast lead time decay rule, the lead time decay information is obtained. Generate confidence grid data according to the configured fusion rule, and apply filling penalty to the corresponding grid points of interpolation mask layer and backfill mask layer to reduce the influence of filling error on training. Then, concatenate the confidence grid data and the unified grid feature data by channel to get the model input grid data. In the sample construction stage, select the number of continuous whole point times corresponding to the window length according to the unified time reference to form the model input grid data sequence, and roll the time window with a step of one hour. At the same time, determine the forecast lead time level according to the starting time and effective time of the forecast grid data and write the forecast lead time encoding data, and group them according to the forecast lead time level and then sample equally to form the training sample set. The training and validation division can be: the ratio of training set to validation set is two to one, the training time sequence is from 00:00 on January 1, 2023 to 23:00 on December 31, 2023, and the validation set is from 00:00 on January 1, 2024 to 23:00 on June 30, 2024.In the model construction, the spatial coding network can be implemented by a structure of multi-scale down-sampling and up-sampling jump connection fusion, for example, a multi-scale feature extraction network composed of multiple layers of encoding and decoding and residual embedding structure as an implementation manner. The time coding network can be implemented by bidirectional time sequence coding and introducing an attention mechanism to weight and converge the sequence features. The prediction time segmentation prediction network sets at least three expert prediction sub-networks corresponding to short, medium and long prediction time intervals, respectively, and generates fusion weights based on the time sequence coding features and the prediction time coding data by the gating network and normalizes them, and then weights and fuses the outputs of each expert to obtain a continuous value prediction result. Taking a set of visibility threshold values as an example, the hierarchical probability output is generated according to the visibility less than or equal to 200 meters, less than or equal to 500 meters, less than or equal to 1,000 meters, and less than or equal to 3,000 meters. In the training stage, the regression weight is set according to the confidence grid data for the continuous value prediction result, and the upper and lower limits of the regression weight are set to form a truncated weight to enhance the robustness. The hierarchical probability output is subjected to sequence consistency verification, and when the hierarchical probability output corresponding to adjacent threshold values does not satisfy the order that does not decrease as the threshold value is relaxed, a sequence conflict penalty term is generated, and the model parameters are updated by the joint loss of the regression loss and the sequence conflict penalty term. In the inference stage, the model input grid data and the corresponding prediction time coding data of the to-be-predicted period are input into the model to obtain the continuous value prediction result and the hierarchical probability output. The matching error between the hierarchical probability output and the training label is calculated according to the prediction time interval on the verification sample to determine the calibration parameter, and the hierarchical probability output is subjected to probability calibration by using the calibration parameter corresponding to the interval. Finally, the consistency verification is performed based on the calibrated hierarchical probability output and the continuous value prediction result, and the prediction result is output if the verification is passed, otherwise the consistency correction is triggered or the release state is marked as needing review, so as to ensure that the released result has explainable consistency between the continuous value and the hierarchical alarm.

[0058] In some embodiments of the present application, referring to Figure 3 As shown in the figure, the spatial coding network can use a U-shaped feature extraction unit containing residual embedding as an implementation manner, and the residual U-shaped structure is used to extract local detail features and multi-scale semantic features simultaneously in the same network level. Specifically, after the input feature map is mapped to shallow features by convolution, it is down-sampled along the encoding path to form feature representations of different spatial scales, and each level of encoded features is transmitted to the corresponding decoding path through jump connection. In the deepest encoding layer, dilated convolution or hybrid dilated convolution can be used to expand the receptive field and enhance the representation ability of sparse events and small-scale structures, and then the features are up-sampled along the decoding path and fused with the features from the same scale encoding path to obtain an output feature map consistent with the input resolution. To ensure training stability and suppress deep degradation, the output of the U-shaped structure is added to its input in a residual manner to obtain the final output, thereby improving the multi-scale extraction ability and detail preservation ability of the spatial coding network for multi-source spatio-temporal grid input.

[0059] It can be understood that, on the basis of multi-source grid point live and model prediction fusion modeling (low visibility sample proportion is high in the verification set, suitable for testing low visibility prediction ability), further through "missing mask representation and confidence modeling", the filling source and reliability are explicitly injected into the model input, and the filling error propagation is suppressed in the training loss with a confidence weight. At the same time, the error characteristics of different prediction time are modeled by segmenting the prediction time and using the gated expert fusion, so as to reduce the performance collapse and instability problem of a single structure fitting across time. Combined with the consistency constraint of hierarchical probability sequence relationship and the probability calibration according to time, the hierarchical alarm probability is more stable and reliable in the threshold sequence and time dimension, thereby improving the identification sensitivity and business release availability of key events such as visibility.

[0060] In some embodiments of the present application, when the data is preprocessed to obtain uniform grid point feature data, the target regular latitude and longitude grid is set as the uniform spatial coordinate, and the hourly whole point is set as the uniform time reference. The live grid point data, the prediction grid point data and the static grid point data are converted to the uniform spatial coordinate. The numerical type elements are resampled to the target regular latitude and longitude grid by bilinear interpolation, and the category type elements are resampled to the target regular latitude and longitude grid by nearest neighbor resampling. The prediction grid point data is converted to the effective time according to the reporting time and the prediction time information and aligned to the uniform time reference. The live grid point data is taken at the nearest time within thirty minutes before and after the uniform time reference, and if there are multiple values, the average value is taken. Finally, the uniform grid point feature data is output.

[0061] Specifically, the present embodiment aims at the sample alignment deviation problem caused by different sources of multi-source spatio-temporal grid point data, inconsistent grid systems and non-uniform time markers. A preprocessing process of "uniform spatial coordinate and uniform time reference" is provided: first, determine the target regular latitude and longitude grid as the uniform spatial coordinate, and take the hourly whole point as the uniform time reference. Then, the live grid point data, the prediction grid point data and the static grid point data are uniformly converted to the uniform spatial coordinate, and the bilinear interpolation and the nearest neighbor resampling are used respectively according to the element type to realize the spatial scale consistency. Subsequently, the prediction grid point data is converted to the effective time according to the reporting time and the prediction time, and is aligned to the uniform time reference. The live grid point data is taken at the nearest time within thirty minutes before and after the uniform time reference, and if there are multiple values, the average value is taken. Finally, the uniform grid point feature data which can be directly used for subsequent model training and reasoning is output, so as to ensure the strict correspondence of multi-source elements in space and time at the same whole point time.

[0062] Specifically, the hourly multi-source fused grid live data is selected as the live grid data, the numerical model prediction data is selected as the prediction grid data, and the spatial range and spatial resolution are consistent with the live grid data. The static grid data selects the terrain elevation and underlying surface type and other static elements with the same grid of the unified spatial coordinates, among which the terrain elevation is a numerical element and the underlying surface type is a category element. During preprocessing, first, the above-mentioned spatial range is divided into a target regular latitude and longitude grid as a unified spatial coordinate, and each hour is taken as a unified time reference. Secondly, the live grid data, the prediction grid data and the static grid data are mapped to the unified spatial coordinate: for numerical elements such as visibility, precipitation, ten-meter wind speed, two-meter humidity, two-meter temperature and terrain elevation, bilinear interpolation resampling is adopted to the target regular latitude and longitude grid. For category elements such as underlying surface type, nearest neighbor resampling is adopted to the target regular latitude and longitude grid. Thirdly, for the prediction grid data, the effective time is calculated according to the start time and the prediction validity of each record, and the effective time is aligned to the unified time reference. If the effective time is in the middle position of the adjacent two whole points, the previous whole point is taken as the alignment result to ensure the uniqueness of the alignment rule. For the live grid data, the available live records are searched within a thirty-minute window before and after the unified time reference, and the record with the smallest time difference from the whole point time is taken as the live value of the whole point. If there are multiple records with the same time difference, they are averaged to obtain the live value of the whole point. Finally, the elements after spatial alignment and time alignment at the same whole point time are combined in the order of the established channel sequence to form the unified grid feature data, and further normalization processing can be performed after the missing values are interpolated and supplemented by the climate values to improve the sample availability and training stability.

[0063] It can be understood that by unifying the multi-source live grid, numerical model prediction grid and static grid to the same target regular latitude and longitude grid and hourly whole point time reference, and adopting differentiated resampling strategies for numerical and category elements, and determining the effective time of the prediction data by "start time plus prediction validity" and determining the whole point value of the live data by "taking the nearest value before and after the whole point and averaging", each whole point sample has determinacy and consistency in spatial position and time direction, significantly reducing label noise and sample alignment error caused by multi-source misalignment. At the same time, the alignment rule can directly reproduce the fusion process of live and prediction data in the unified space-time framework, providing stable basic input for subsequent missing mask and confidence modeling, thereby improving the repeatability and engineering landing maintainability of the overall model training.

[0064] In some embodiments of the present application, when generating the missing mask grid data, the missing or filling marks are read from the unified grid feature data by element and grid point. When the missing or filling mark indicates an original valid value, an effective mask layer is generated. When the missing or filling mark indicates an interpolation filling value, an interpolation mask layer is generated. When the missing or filling mark indicates a backfilling filling value, a backfilling mask layer is generated. The effective mask layer, the interpolation mask layer and the backfilling mask layer are combined into the missing mask grid data according to the corresponding relationship of the elements.

[0065] In some embodiments of the present application, when generating the confidence grid data and obtaining the model input grid data, the observation coverage information is determined based on the proportion of effective grid points of the effective mask layer in a preset spatial neighborhood. The interpolation distance information is determined based on the minimum distance from the observation point to the target grid point. The spatial neighborhood consistency test and the time adjacent consistency test are performed on the unified grid feature data to obtain the spatiotemporal consistency test result. The timeliness decay information is determined according to the timeliness information of the forecast grid data and according to the timeliness decay rule. The confidence grid data is generated according to the configured fusion rule of the observation coverage information, the interpolation distance information, the spatiotemporal consistency test result and the timeliness decay information, and the confidence grid data is punished according to the interpolation mask layer and the backfilling mask layer. The confidence grid data after the punishment and the unified grid feature data are spliced according to the channel to obtain the model input grid data.

[0066] Specifically, the present embodiment aims at the missing and filling error propagation problems that inevitably occur in the quality control, transmission link and spatiotemporal fusion process of multi-source spatiotemporal grid data, and provides a missing traceable mask representation and a multi-factor joint driven confidence modeling implementation manner. The missing or filling marks are read from the unified grid feature data by element and grid point. The data source state of the same grid point under the same element is clearly distinguished into an original valid value, an interpolation filling value and a backfilling filling value, and the effective mask layer, the interpolation mask layer and the backfilling mask layer are generated respectively, and then the missing mask grid data is composed according to the corresponding relationship of the elements. On this basis, the confidence grid data is constructed from four dimensions of observation coverage, interpolation distance, spatiotemporal consistency and forecast timeliness decay, and the interpolation mask layer and the backfilling mask layer are punished, so that the confidence can describe the confidence degree of the grid point, the element and the time point. Finally, the confidence grid data and the unified grid feature data are spliced according to the channel to form the model input grid data, so as to provide controllable quality weight and error suppression entrance for deep learning model training.

[0067] Specifically, in the sample construction phase, missing value processing is first performed on each element point by point, and a missing or filling mark is generated synchronously: when the grid point value is missing, the nearest point interpolation is used to obtain the filling value in the preset spatial neighborhood, and the missing or filling mark of the grid point at that time of the element is recorded as the interpolation filling value. When the nearest point interpolation still cannot produce the filling value, the climate value is used to generate the backfill filling value, and the missing or filling mark is recorded as the backfill filling value. When the grid point value is the original value passed by quality control, the missing or filling mark is recorded as the original effective value. Subsequently, the missing or filling mark is read for each element point by point: the grid point marked as the original effective value is set as valid in the effective mask layer, the grid point marked as the interpolation filling value is set as valid in the interpolation mask layer, the grid point marked as the backfill filling value is set as valid in the backfill mask layer, and the three types of mask layers are combined in the order of elements to obtain the missing mask grid point data. Further, the confidence calculation uses reproducible configuration rules: first, the observation coverage information is obtained from the effective grid point ratio in the preset spatial neighborhood in the effective mask layer, which reflects the density of original effective values in the neighborhood. Second, the interpolation distance information is calculated for the grid point that has been interpolated and filled, which reflects the degree of interpolation extrapolation. Third, the spatiotemporal consistency test result is obtained from the spatial neighborhood consistency test and the time adjacent consistency test, the spatial neighborhood consistency test is used to determine whether the change of the target grid point and the neighborhood original effective grid point is abrupt, and the time adjacent consistency test is used to determine whether the change of the target grid point at adjacent integral points is abnormal. Fourth, the time decay information is determined according to the forecast time interval of the forecast grid point data and according to the preset decay rule, which reflects the rule that the credibility decreases with the growth of the forecast time interval. Finally, the observation coverage information, the interpolation distance information, the spatiotemporal consistency test result and the time decay information are generated into confidence grid point data according to the configured fusion rule, and the confidence grid point data is subjected to filling penalty according to the interpolation mask layer and the backfill mask layer, so that the confidence corresponding to the backfill filling is lower than that of the interpolation filling, and the confidence corresponding to the interpolation filling is lower than that of the original effective value. The confidence grid point data after filling penalty and the unified grid point feature data are spliced according to the channel to obtain the model input grid point data, which is used for subsequent training and reasoning.

[0068] It can be understood that by explicitly solidifying the missing source state as missing mask grid data, and jointly mapping the observation coverage, interpolation distance, spatio-temporal consistency and time decay into confidence grid data, the model can automatically reduce the weight of low confidence samples during the training stage, suppress the filling error propagation caused by the interpolation of nearby points and the supplement of climate values from the source, and avoid the parameter bias caused by treating the filling values and the original effective values as the same. At the same time, the confidence can provide a consistent quality reference for continuous value prediction and hierarchical probability output during the inference stage, making the output more stable and more interpretable, and the marking generation, mask construction, confidence fusion and punishment rules can be reproduced according to the data and configuration, meeting the requirements of engineering landing and review verifiability.

[0069] In some embodiments of the present application, when constructing training samples based on a sliding time window, the following steps are included: taking each target prediction time under the unified time reference as an alignment point, selecting model input grid data corresponding to a number of consecutive whole point times of a preset window length, forming a model input grid data sequence, and rolling the sliding time window with a whole point time as a step. For each training sample, determine the corresponding prediction time validity level according to the start time and the effective time of the prediction grid data, and write the prediction time validity level converted into prediction time validity encoding data into the training sample. Group the training samples according to the prediction time validity level, and perform equal sampling on each group to obtain a training sample set.

[0070] Specifically, after completing the alignment of the unified spatial coordinates and the unified time reference, and obtaining the model input grid data corresponding to each whole point time, a sliding time window is constructed with each target prediction time under the unified time reference as an alignment point: selecting model input grid data corresponding to a number of consecutive whole point times of a preset window length, arranging the model input grid data in time sequence to form a model input grid data sequence, and rolling the sliding time window with a whole point time as a step. For alignment points that are in the initial stage of the training period and cannot meet the complete window length, no training sample is constructed. At the same time, for each training sample, determine the prediction time validity level based on the start time and the effective time of the prediction grid data, and write the prediction time validity level converted into prediction time validity encoding data into the training sample, wherein the prediction time validity encoding data is represented by encoding bits corresponding one-to-one to the prediction time validity level, and the encoding bit corresponding to the target prediction time validity level is valid, and other encoding bits are invalid. Further, group the training samples according to the prediction time validity level, and take the minimum of the number of samples in each group as the unified sampling quantity, and perform equal sampling on each group to obtain a training sample set, thereby reducing the training bias caused by the uneven distribution of different prediction time validity samples.

[0071] Specifically, when the preset window length is set as six consecutive whole point times and the step is one whole point time, for each target prediction time in the training set, six whole point time corresponding model input grid data are selected to form a model input grid data sequence. For each sample, the starting time is obtained by matching the prediction grid data record with the effective time equal to the target prediction time, and the prediction validity level is determined accordingly. The prediction validity level is further divided into three categories: short-term, medium-term and long-term, and prediction validity encoding data is generated. Finally, the training sample set is formed by grouping according to the three categories of prediction validity levels and randomly sampling the remaining groups in equal amounts according to the minimum group sample size. In the verification stage, the predictions starting at 8:00 and 20:00 can be compared and evaluated to verify the consistency of the sample construction and training effect under different prediction validity conditions.

[0072] It can be understood that by using the target prediction time as the alignment point of the sliding time window, the near-time sequence evolution of multi-source grid elements under a unified time reference is directly included in the training sample, enabling the model to learn the time sequence triggering and continuous process of low-visibility events. By encoding the prediction validity in levels and grouping and sampling in equal amounts, the number bias of the training sample under different validity is explicitly weakened, avoiding performance collapse of the model in a specific prediction validity segment, thereby improving the stability and consistency of cross-validity prediction.

[0073] In some embodiments of the present application, when constructing a deep learning prediction model, the spatial encoding network is configured to perform multi-scale down-sampling on the model input grid data of a single whole point time to extract spatial features, and the spatial encoding features are output by jumping and fusing with the up-sampling branch of the corresponding scale. The time encoding network is configured to perform time sequence encoding on the spatial encoding features corresponding to the model input grid data sequence, and the prediction validity encoding data is used as the gate input of the time sequence encoding to modulate the time sequence weight of the spatial encoding features under different prediction validity levels, and the time sequence encoding features are output.

[0074] In some embodiments of the present application, when constructing a prediction validity segmented prediction network, at least three expert prediction sub-networks are set up, corresponding to short-term, medium-term and long-term prediction validity levels respectively. The gating network is configured to generate the fusion weight of each expert prediction sub-network based on the time sequence encoding features and the prediction validity encoding data, and the fusion weight is normalized. The output of each expert prediction sub-network is weighted and fused according to the fusion weight to obtain a continuous value prediction result. Based on the continuous value prediction result, a hierarchical probability output corresponding to a threshold set is generated through a hierarchical output layer, and a monotonic constraint according to the order of the threshold set is applied to the hierarchical probability output to avoid order conflict of the hierarchical probability output.

[0075] Specifically, the embodiment first performs spatial representation on the model input grid data of the same whole hour moment, then performs time series modeling on the spatial representation sequence formed by the continuous whole hour moments, and finally performs gated fusion output according to the forecast time limit segmentation under the overall framework of the “spatial coding network-time coding network-forecast time limit segmentation prediction network”. The spatial coding network adopts a multi-scale structure with symmetric encoding and decoding: the model input grid data of a single whole hour moment is down-sampled to extract multi-scale spatial features, and the spatial details are recovered by up-sampling at the corresponding decoding layer, while the spatial coding features are obtained by jumping and fusing with the same scale encoding branch. The time coding network performs time series coding on the spatial coding features corresponding to the model input grid data sequence, and introduces the forecast time limit coding data as the gated input: specifically, the forecast time limit coding data is mapped to the time limit gating weight through gating, and the spatial coding features or their time series hidden states of each whole hour moment are modulated by the time limit gating weight, so that the time series weight distribution is different under different forecast time limit gears, thereby maintaining the stability of feature extraction and memory update when the time limit changes. To enhance the extraction ability of the sequence context, the time coding network of the embodiment can adopt a bidirectional time series structure with an attention mechanism, model the forward and backward information together, and weight the time series output through the attention weight to form more robust time series coding features. In the forecast time limit segmentation prediction network, at least three expert prediction sub-networks are set to correspond to the short-time forecast time limit gear, the medium-time forecast time limit gear and the long-time forecast time limit gear, respectively. The gating network generates the fusion weight of each expert prediction sub-network based on the time series coding features and the forecast time limit coding data, and performs normalization processing on the fusion weight. Each expert prediction sub-network outputs a continuous value prediction sub-result and a hierarchical probability sub-result, and the gating network weights and fuses each sub-result according to the fusion weight to obtain a continuous value prediction result and a hierarchical probability output. The hierarchical probability output corresponds to a threshold set generated level by level, and a monotonic constraint is applied to the hierarchical probability output, so that the probability corresponding to a stricter threshold is not greater than the probability corresponding to a more lenient threshold, to avoid order conflict of hierarchical alarm.

[0076] Specifically, after the sample is constructed, a model input grid data sequence is formed for each target prediction time at consecutive integral time points, and the spatial coding network independently extracts spatial coding features for the model input grid data at each integral time point in the sequence. The spatial coding network can be implemented in a multi-scale structure of encoding layers and decoding layers, and the spatial coding features are output by jumping and fusing the encoding results of corresponding scales in the decoding stage. The time coding network inputs the spatial coding features in the above sequence into a bidirectional time sequence structure in time order and obtains time sequence coding features by combining an attention mechanism. The attention mechanism is realized by converting time sequence output into attention scores and performing weighted aggregation. On this basis, the prediction time coding data is used as a gate input to participate in the time sequence coding process, which is used to modulate the time sequence weight in different gears of short time, medium time and long time. In the prediction time segmentation prediction network, the short-time expert prediction sub-network, the medium-time expert prediction sub-network and the long-time expert prediction sub-network respectively map the time sequence coding features to output corresponding continuous value prediction sub-results and hierarchical probability sub-results, and the gating network generates fusion weights according to the time sequence coding features and the prediction time coding data, and then performs weighted fusion after normalization. The hierarchical output layer generates hierarchical probability output according to the threshold set corresponding to the visibility hierarchical interval, and applies monotonic constraint according to the threshold set order to ensure the consistency of the hierarchical probability output. The following evaluation settings can be used: the predictions made at 8 o'clock and 20 o'clock every day are comprehensively compared and evaluated, and the test is performed according to the hierarchical interval and the time range to verify the stability of different prediction time segments.

[0077] It can be understood that the spatial coding network enhances the representation ability of the large fog spatial structure and local details by multi-scale downsampling and upsampling combined with jumping and fusion. The time coding network introduces prediction time gating modulation based on bidirectional time sequence modeling and attention aggregation, so that the model can adaptively allocate time sequence weights for different prediction time gears, and reduce the feature mismatch caused by the change of time. In combination with the expert prediction sub-network and the gating network, the phenomenon of low value area, insufficient stability and the like of part of the model in a specific prediction time segment can be directed inhibited, and the monotonic constraint of the hierarchical probability output can reduce the order conflict of the hierarchical alarm, thereby improving the stability of the continuous value prediction and the reliability of the hierarchical alarm across time.

[0078] In some embodiments of the present application, when training a deep learning prediction model, the following steps are included: setting regression weights for each grid corresponding to the continuous value prediction result according to the confidence grid data, and setting upper and lower limits for the regression weights to form truncated weights. The regression loss between the continuous value prediction result and the training label is calculated based on the truncated weights. The order relationship consistency of the hierarchical probability output is verified according to the order relationship of the threshold set. When the hierarchical probability outputs corresponding to adjacent thresholds do not satisfy the order of the threshold set, an order conflict penalty term is generated. The model parameters of the deep learning prediction model are updated based on the joint loss of the regression loss and the order conflict penalty term.

[0079] Specifically, after the aforementioned data preprocessing under unified spatial coordinates and unified time reference, the construction of missing mask grid data and confidence grid data, and the construction of sliding time window training samples are completed, the embodiment further gives the training details of the deep learning prediction model. In the training stage, for the continuous value prediction result, a "confidence weighted regression loss" is introduced, that is, based on the confidence grid data, a regression weight is generated for each sample, each element channel and each grid position, and the grid with higher confidence is given greater regression weight. At the same time, in order to avoid the unreasonable dominance of a small number of extremely high or extremely low confidence grids on the training, the regression weight is set with an upper limit and a lower limit to form a truncated weight, and then the regression loss between the continuous value prediction result and the training label is calculated based on the truncated weight. For the graded probability output, the order consistency check is performed according to the ordering relationship of the threshold set: when the graded probability output corresponding to adjacent thresholds appears "order conflict" (for example, the probability corresponding to the small threshold is higher than the probability corresponding to the large threshold, resulting in self-contradiction of the alarm level logic), an order conflict penalty term is generated. Finally, the model parameters are updated with the joint loss of the regression loss and the order conflict penalty term, so as to balance the continuous value accuracy and the explainable consistency of the graded alarm.

[0080] Specifically, taking the "fog visibility forecast" as an example, the hourly CLDADS multi-source fusion point real-time data is selected as the main training label source and part of the input elements, including at least visibility, precipitation, 10-meter wind direction, 10-meter wind speed, 2-meter humidity, and 2-meter temperature. The spatial resolution is processed to 3 km x 3 km regional grid data. At the same time, the CMA-GD model visibility forecast is introduced as the forecast grid data, covering 0-72 hour forecast time and used to construct sample and time coding data of different forecast time intervals. To enhance the training sample size, the missing values in the data set are first filled in by using the interpolation of adjacent points and the supplement of climate values, and the input elements and training labels are normalized to [0, 1] interval to improve the stability and convergence of the deep learning model training. In terms of data set division, the training set is selected from 00:00 on January 1, 2023 to 23:00 on December 31, 2023, and the number of training samples is 8784. The validation set is selected from 00:00 on January 1, 2024 to 23:00 on June 30, 2024, and can be compared and evaluated according to the 0-72 hour forecast time every 3 hours. In this embodiment, the threshold set can use the commonly used four-level threshold of visibility classification (visibility ≤200m, ≤500m, ≤1000m, ≤3000m), and the corresponding classification probability output is generated at the output end of the model for each threshold. During training, for each training sample: first read its confidence grid data and map it to the regression weight (for example, the confidence value is converted to the weight according to the preset monotone mapping, and the lower the confidence, the smaller the weight), and then the regression weight is clipped to the preset lower limit and upper limit to obtain the truncated weight. Then, the error between the continuous value prediction result and the visibility training label is weighted and summed to obtain the regression loss. At the same time, for the classification probability sequence output at the same grid position, the consistency is verified according to the threshold set from small to large in pairs: if the probability of adjacent thresholds appears in conflict with the order, then the order conflict penalty term is accumulated according to the conflict degree. Finally, the joint loss is used to iteratively update the model parameters (which can be realized by using the conventional gradient descent type optimizer and small batch training method), until the convergence and generalization requirements are met on the validation set.

[0081] Specifically, the regression loss is weighted in combination with the confidence grid data, and extreme weights are suppressed through upper / lower limit truncation. In the presence of interpolation filling, backfilling, and other scenarios, the pollution of low-confidence grid errors on model gradients can be effectively reduced, the learning of high-confidence observations and high-coverage areas by the model can be strengthened, and the benefits of missing data supplement and normalization are further amplified, which helps to improve the quality of sample data and model training. By applying sequence relationship consistency checking and order conflict penalty terms to the hierarchical probability output, the logical self-consistency of the hierarchical alarm can be constrained from the training stage, making the probability output of different threshold levels more stable and more consistent with the monotonic rule that "the wider the threshold, the lower the trigger probability" in business. This is especially suitable for heavy fog forecasting scenarios with a high proportion of low-visibility samples and the need for hierarchical verification and evaluation, thereby improving the credibility and availability of the release-end hierarchical alarm.

[0082] In some embodiments of the present application, when reasoning and generating the release prediction result, the model input grid data and corresponding prediction validity coding data of the to-be-predicted period are input into the trained deep learning prediction model, and continuous value prediction results and hierarchical probability outputs are output. The matching error between the hierarchical probability output and the training label is statistically analyzed on the verification sample to determine the calibration parameter, and the calibration parameter is set according to the prediction validity level. The hierarchical probability output is calibrated using the calibration parameter corresponding to the prediction validity level to obtain the calibrated hierarchical probability output. Consistency checking is performed based on the calibrated hierarchical probability output and the continuous value prediction result, and when the consistency checking is passed, the release prediction result is output.

[0083] Specifically, the inference and generation of the release prediction result are used to deploy the trained deep learning prediction model into the business operation link, output continuous value prediction results and hierarchical probability outputs for a specified prediction period, and improve the credibility and availability of the release result through probability calibration and consistency check. Specifically, the model input grid data of the to-be-predicted period and the corresponding prediction validity coding data are input into the trained deep learning prediction model to obtain the continuous value prediction result and the hierarchical probability output corresponding to the threshold set. The hierarchical probability output is used to represent the occurrence probability of the visibility falling within each threshold constraint. Subsequently, based on the matching error of the verification sample statistical hierarchical probability output and the training label, the calibration parameters are determined according to the prediction validity level, and the hierarchical probability output is calibrated using the calibration parameters corresponding to the prediction validity level to obtain the calibrated hierarchical probability output. Finally, the calibrated hierarchical probability output and the continuous value prediction result are subjected to consistency check: the continuous value prediction result is mapped to the corresponding hierarchical interval according to the threshold set, and the hierarchical probability output corresponding to the adjacent threshold is used to generate the probability of each hierarchical interval, which is used to verify whether the hierarchical interval to which the continuous value prediction result belongs and the dominant hierarchical interval of the interval probability are consistent. When the consistency check passes, the release prediction result containing the continuous value prediction result and the calibrated hierarchical probability output is output, and when the consistency check fails, the exception is recorded and the release output of the grid point is terminated to avoid business misjudgment caused by release alarm sequence conflict or probability distortion.

[0084] Specifically, after the training is completed, January 1, 2024, 00:00 to June 30, 2024, 23:00 are selected as the verification sample, the forecast at 8:00 and 20:00 every day is subjected to hierarchical inspection and evaluation, and the visibility is classified according to 200 meters, 500 meters, 1,000 meters and 3,000 meters to cover the alarm requirements of heavy fog and low visibility scenarios. In the inference stage, for a to-be-predicted period corresponding to a daily reporting time, the model input grid data of the period is constructed and the prediction validity coding data is written, and the deep learning prediction model is input to obtain the continuous value prediction result and the hierarchical probability output. In the calibration stage, the verification sample is binned according to the prediction validity level to form the calibration parameters, which at least include the bin boundary and the probability mapping relationship corresponding to each bin, so that the hierarchical probability output in the same bin matches the actual occurrence frequency of the training label. In the release stage, the calibration parameters consistent with the current prediction validity level are used to calibrate the hierarchical probability output, and the consistency check is performed. After the check passes, the release prediction result is output, and if the check fails, the grid point is not released and the reason is recorded to support subsequent quality analysis and threshold optimization.

[0085] In some embodiments of the present application, reference is made to Figure 4As shown, the numerical error of the continuous value prediction result under different forecast time interval levels is visualized and verified, the root mean square error RMSE is counted according to the forecast time interval on the verification sample, and the error score curve with the change of the forecast time interval is formed under two test standards of grid point prediction and station point prediction. Among them, the grid point prediction test can calculate the RMSE of the continuous value prediction result and the corresponding training label (or the label after real-time alignment) in the unified spatial coordinates for each forecast time interval in the verification sample, and then the error score of the time interval is obtained by summarizing. The station point prediction test can interpolate or map the grid point prediction result to the station point according to the station point position and calculate the RMSE score; as an example, the RMSE score change of different model implementations / contrast schemes within 0-72 hour forecast time interval is given in the figure, it can be observed that there are stage differences between different implementation schemes in some forecast time interval, and there may also be abnormal large value fluctuations in individual time interval, while the implementation scheme which emphasizes more on time sequence stable expression performs more stably and has smaller overall error in most time intervals, thereby illustrating that the overall link of organizing samples and output according to the forecast time interval and performing calibration on the publishing side in the present application is beneficial to improve the stability and business availability of cross-time interval prediction.

[0086] It can be understood that through the cooperation of the above reasoning, probability calibration and consistency verification, the embodiment can suppress the systematic deviation of the graded probability output for different forecast time interval levels, make the graded probability output more consistent with the real event frequency, and thus improve the stability and interpretability of the graded alarm. At the same time, by taking the threshold set as the link to constrain the consistency of the continuous value prediction result and the graded probability output, the risk of "continuous value and alarm probability contradicting each other" under low visibility events is reduced, and the business disturbance caused by false alarm and missed alarm is reduced.

[0087] It can be understood that, in some embodiments of the present application, in order to make the missing mask, confidence, training / inference link reproducible, the missing or filling marks can be generated according to the data source quality control results: for each grid point of each element, if the live / forecast / static original value meets the value range and quality identification requirements, it is marked as an original valid value. If the original value is missing and can be interpolated and filled by adjacent valid grid points using bilinear or nearest neighbor method under unified spatial coordinates, it is marked as an interpolated and filled value. If interpolation still fails, climate statistical value or historical same hour average value is used to complete backfilling and filling and marked as a backfilled and filled value. On this basis, the preset spatial neighborhood can take a k×k neighborhood centered on the target grid point (for example, k=3 or 5), the observation coverage information takes the proportion of grid points with valid mask layer of 1 in the neighborhood, the interpolation distance information takes the minimum spherical distance / grid distance from the target grid point to the nearest observation source grid point or station, and the spatio-temporal consistency check can be performed on the spatial neighborhood (for example, the deviation from the median / mean of the neighborhood) and the time adjacent (for example, the change rate from the previous integral point value) respectively to obtain consistent / inconsistent marks. The forecast time decay information can segment the decay coefficient or exponential decay coefficient according to the forecast time interval setting. The confidence grid point data can normalize and weight fuse the observation coverage, interpolation distance, consistency test and time decay according to the preset weight to obtain the basic confidence, and multiply the interpolation mask layer and the backfill mask layer corresponding grid points by a penalty coefficient less than 1 to reflect the filling uncertainty. The regression weight in the training stage is obtained by mapping the confidence and is truncated to avoid extreme samples dominating the gradient, and the consistency constraint of the order relationship of the hierarchical probability can generate a penalty term for the monotonicity violation of the output corresponding to adjacent thresholds and weighted sum with the regression loss. In the inference stage, the probability calibration can fit temperature scaling / equivalent regression calibration parameters respectively according to the forecast time interval on the validation sample to minimize the matching error, and based on the hierarchical relationship derived from the continuous value prediction result and the threshold set, the consistency of the calibrated hierarchical probability output is verified (for example, requiring the hierarchical probability matching the threshold interval where the continuous value is located to reach a preset confidence threshold or meet the monotonic sequence constraint), and when the consistency verification is passed, the prediction result is output.

[0088] In an optional release control mode, when the calibrated hierarchical probability output and the continuous value prediction result do not pass the consistency check, consistency correction can be performed: first, map the continuous value prediction result to the corresponding event interval according to the threshold set, and check whether the hierarchical probability output satisfies the monotonicity constraint that does not decrease with the threshold relaxation; if there is a probability item that violates the monotonicity, the hierarchical probability output is modified and corrected in a minimum way using monotonic regression or adjacent item merging to make it satisfy the monotonicity and keep overall proximity to the calibration result; if the continuous value prediction result is in the threshold critical region and the probability output is in significant conflict with it, the grid point is marked as a recheck release state and the corresponding mark bit is output, so that the business side selects a conservative release strategy or triggers secondary inference.

[0089] In some embodiments of the present application, the training label is used to represent the true state at the target prediction time, which can be generated by the corresponding grid element value of the live grid data after alignment with the target prediction time; for continuous value prediction tasks, the training label is the continuous measured / fused value of the target prediction element; for hierarchical probability output tasks, the same training label is thresholded based on a threshold set to generate hierarchical labels corresponding to the threshold set, where each hierarchical label is used to indicate whether the training label meets the corresponding threshold condition, so that the continuous value regression training and the hierarchical probability training share consistent data sources and spatio-temporal alignment benchmarks.

[0090] In some embodiments of the present application, to evaluate the prediction stability under different prediction time ranges, the following evaluation method can be used: select multiple prediction samples at different starting times and cover the preset prediction time range, for each prediction time range, respectively, count the matching error between the hierarchical probability output and the hierarchical label, and the error index between the continuous value prediction result and the continuous value training label, and compare them by grouping according to the hierarchical interval and the time range, so as to verify the stability and consistency of the model output under different time ranges.

[0091] In summary, the present application realizes traceability of missing data sources through missing mask grid data, unifies observation coverage, interpolation distance, spatio-temporal consistency and time decay into reliable degree information that can be learned through confidence grid data, and suppresses the interference of filling errors on model parameters in the training stage with confidence weighting. At the same time, based on the fusion mechanism of the expert prediction sub-network and the gating network under the prediction time segmentation, the adaptive ability under different prediction time ranges is improved, and the sequence relationship consistency constraint and the probability calibration according to the prediction time range are introduced at the hierarchical probability output end, so that the continuous value prediction and the hierarchical alarm probability are more stable and reliable in the publishing link, thereby improving the identification sensitivity and business availability of key events such as low visibility.

[0092] Based on another preferred way of the above embodiments, referring to Figure 5 The present embodiment provides a deep learning prediction system for multi-source spatio-temporal grid data, which is used to apply the deep learning prediction method of the multi-source spatio-temporal grid data as described above, comprising:

[0093] The acquisition unit is configured to obtain live grid data, prediction grid data and static grid data, and to obtain uniform grid feature data by preprocessing the data, the preprocessing including unifying spatial coordinates and time benchmarks.

[0094] The processing unit is configured to generate missing mask grid data according to the missing or filling mark, generate confidence grid data according to the observation coverage information, the interpolation distance information, the spatio-temporal consistency test result and the prediction time decay rule, and splice the confidence grid data with the unified grid feature data to obtain model input grid data.

[0095] The constructing unit is configured to construct a training sample based on a sliding time window, so that the training sample contains a model input grid data sequence and prediction time encoding data.

[0096] The training unit is configured to construct and train a deep learning prediction model, the deep learning prediction model including a spatial encoding network, a temporal encoding network and a prediction time segmentation prediction network, the prediction time segmentation prediction network including a gating network and a plurality of expert prediction sub-networks, the gating network weighting and fusing the plurality of expert prediction sub-networks to output a continuous value prediction result, and outputting a hierarchical probability output based on a threshold set. The continuous value prediction result is subjected to confidence weighted regression loss, and the hierarchical probability output is subjected to sequence relationship consistency constraint.

[0097] The inference unit is configured to output the continuous value prediction result and the hierarchical probability output when inferring, and to obtain a published prediction result by probability calibration of the hierarchical probability output.

[0098] Specifically, step one, data preprocessing. The acquisition unit obtains live grid data, prediction grid data and static grid data from a business data source. In a reproducible embodiment, the live grid data can be selected from grid products fused by multi-site observations, the prediction grid data can be selected from grid prediction fields generated by numerical simulation models or statistical prediction models, and the static grid data can include terrain elevation, underlying surface category, etc. A target regular latitude-longitude grid is set as a unified spatial coordinate, and an hourly whole point is set as a unified time reference. The live grid data, the prediction grid data and the static grid data are converted to the unified spatial coordinate. Numerical elements are resampled to the target regular latitude-longitude grid by bilinear interpolation, and categorical elements are resampled to the target regular latitude-longitude grid by nearest neighbor interpolation. The prediction grid data is converted to an effective time according to the reporting time and the prediction time information and aligned to the unified time reference. The live grid data is taken as the nearest time value within thirty minutes before and after the unified time reference, and if there are multiple values, the average value is taken, and finally the unified grid feature data is output.

[0099] Step two, missing data mask and confidence modeling. The processing unit reads the missing data or filling marks by elements and grid points. When the mark indicates an original valid value, a valid mask layer is generated. When the mark indicates an interpolated filling value, an interpolation mask layer is generated. When the mark indicates a backfill filling value, a backfill mask layer is generated. The missing data mask grid point data is composed according to the corresponding relationship of the elements. On this basis, the processing unit calculates the effective grid point proportion of the valid mask layer in the preset spatial neighborhood to obtain the observation coverage information. Based on the minimum distance from the observation point to the target grid point in the target regular latitude-longitude grid for generating the interpolated filling value, the interpolation distance information is obtained. The spatial neighborhood consistency test and the time adjacent consistency test are performed on the unified grid point feature data to obtain the spatiotemporal consistency test result. The time decay information is generated according to the prediction time of the prediction grid point data and according to the prediction time decay rule. The processing unit generates the confidence grid point data according to the configured fusion rule based on the above information. The fusion rule can be that each item of information is normalized to the same value interval and then weighted and synthesized according to the preset weight. The confidence grid point data is filled and punished according to the interpolation mask layer and the backfill mask layer, wherein the confidence corresponding to the interpolation filling value and the backfill filling value is reduced. Finally, the confidence grid point data after filling and punishment and the unified grid point feature data are spliced according to the channel to obtain the model input grid point data.

[0100] Step three, training sample construction. The construction unit takes the target prediction time under the unified time reference as the alignment point, selects the model input grid point data forward according to the number of continuous whole point times corresponding to the preset window length, forms the model input grid point data sequence, and rolls the time window with a whole point time as a step. For each training sample, the construction unit determines the prediction time interval according to the start time and the valid time of the prediction grid point data, and converts the prediction time interval into prediction time coding data and writes it into the training sample. The training samples are grouped according to the prediction time interval and equally sampled to avoid training bias caused by the difference in the number of samples of each prediction time interval.

[0101] Step four, model structure and training. The training unit constructs a deep learning prediction model. The spatial encoding network performs multi-scale down-sampling on the model input grid data of a single integral time to extract spatial features, and outputs spatial encoding features by jumping and fusing with the up-sampling branch of the corresponding scale. The time encoding network performs time series encoding on the spatial encoding features corresponding to the model input grid data sequence, and takes the prediction lead time encoding data as the gate input to modulate the time series weight of the spatial encoding features under different prediction lead time levels, and outputs the time series encoding features. The prediction lead time segmented prediction network sets at least three expert prediction sub-networks corresponding to short, medium and long prediction lead time levels. The gating network generates the fusion weight of each expert prediction sub-network based on the time series encoding features and the prediction lead time encoding data, normalizes the fusion weight, and then weights and fuses the output of each expert prediction sub-network to obtain a continuous value prediction result. The hierarchical output layer generates a hierarchical probability output corresponding to the threshold set based on the continuous value prediction result, and applies a monotonic constraint on the hierarchical probability output according to the order of the threshold set to avoid order conflict. In the training stage, the training unit sets regression weights for each grid according to the confidence grid data, sets an upper limit and a lower limit to form a truncated weight, and calculates the confidence weighted regression loss between the continuous value prediction result and the training label. At the same time, the hierarchical probability output is subjected to sequence relationship consistency check, and when the hierarchical probability outputs corresponding to adjacent thresholds violate the threshold set order, an order conflict penalty term is generated, and the model parameters are updated based on the joint loss of the regression loss and the order conflict penalty term.

[0102] Step five, reasoning and publishing. The reasoning unit inputs the model input grid data and the corresponding prediction lead time encoding data of the to-be-predicted period into the trained deep learning prediction model to output the continuous value prediction result and the hierarchical probability output. On the validation sample, the reasoning unit calculates the matching error between the hierarchical probability output and the training label, determines the calibration parameter, and sets the calibration parameter according to the prediction lead time level. Then, the hierarchical probability output is probability calibrated using the calibration parameter corresponding to the prediction lead time level to obtain the calibrated hierarchical probability output. The consistency check can use the following rules: mapping the continuous value prediction result to the hierarchical interval corresponding to the threshold set, checking whether the maximum probability level of the calibrated hierarchical probability output is consistent with the hierarchical interval, and the calibrated hierarchical probability output satisfies the threshold set order. When the consistency check passes, the prediction result is output and published.

[0103] Specific reproducible embodiment: Taking visibility prediction in heavy fog as an example, a unified spatial coordinate is set as a target regular latitude-longitude grid covering a target area, and a unified time reference is set as every hour. Live grid point data can use fused live visibility grid point products, forecast grid point data can use visibility grid point prediction fields generated by a numerical simulation model, and static grid point data can include terrain elevation and underlying surface categories. After the observation and prior prediction data are unified to the unified spatial coordinate and the unified time reference, unified grid point feature data are formed for subsequent missing mask generation, confidence modeling, and model training. The forecast validity information is determined by the start time and the effective time of the forecast grid point data.

[0104] The embodiment can suppress the propagation of filling errors, improve the prediction stability of different forecast validity levels, and make the hierarchical alarm probability and the continuous value prediction result consistent on the release side, thereby improving the identification sensitivity and business availability of key events such as low visibility, through explicit modeling of data reliability of missing mask grid point data + confidence grid point data, release control of forecast validity segmented prediction network, sequence consistency constraint, and time validity probability calibration.

[0105] It can be understood that by explicitly passing the missing or filling state as missing mask grid point data on the input side, and fusing the observation coverage, interpolation distance, spatio-temporal consistency test results, and forecast validity decay rules into confidence grid point data, the model can distinguish the reliable differences between original valid information and filling information, suppress the amplification of missing filling errors in spatio-temporal propagation from the source, and improve the identification and positioning ability of rare events such as low visibility. By introducing forecast validity encoding on the model side and performing time validity weighted fusion of multiple expert prediction sub-networks through a gating network, different forecast validity intervals obtain targeted expression capability, alleviate the performance collapse of a single model in a specific validity interval caused by cross-validity unified fitting, and improve the stability of long-validity prediction. At the same time, on the output and release side, continuous value prediction and hierarchical probability output are generated in parallel, sequence consistency constraints are imposed on hierarchical probability output in the training stage to avoid level order conflicts, and hierarchical probability output is calibrated according to time validity probability in the reasoning stage to form release prediction results, so that the hierarchical alarm meets the monotonic consistency and probability confidence of the business threshold, reduces false positives and false negatives, and enhances the result interpretability, reproducibility, and reliability of engineering deployment.

[0106] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit it, although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that the specific embodiments of the present application can be modified or replaced equivalently without departing from the spirit and scope of the present application, any modification or equivalent replacement without departing from the spirit and scope of the present application should be covered in the protection scope of the claims of the present application.

Claims

1. A deep learning prediction method for multi-source spatiotemporal point data, characterized in that, include: Acquire real-time grid data, forecast grid data, and static grid data, and preprocess the data to obtain unified grid feature data. The preprocessing includes unifying spatial coordinates and time references. Missing measurement mask grid data is generated based on missing measurement or filling marks. Confidence grid data is generated based on observation coverage information, interpolation distance information, spatiotemporal consistency test results and forecast lead time decay rules. The confidence grid data is then concatenated with the unified grid feature data to obtain model input grid data. Training samples are constructed based on a sliding time window, such that the training samples contain the model input grid data and the forecast lead time coding data; A deep learning prediction model is constructed and trained. The deep learning prediction model includes a spatial coding network, a temporal coding network, and a prediction network for forecast timeliness. The prediction network for forecast timeliness includes a gating network and multiple expert prediction subnetworks. The gating network weights and fuses the multiple expert prediction subnetworks based on the forecast timeliness coding data to output continuous value prediction results, and outputs hierarchical probability output based on a threshold set. During training, a confidence-weighted regression loss is applied to the continuous value prediction results, and an order relation consistency constraint is imposed on the hierarchical probability output. During inference, the continuous value prediction result and the hierarchical probability output are output, and the hierarchical probability output is probabilistically calibrated to obtain the published prediction result; Generating the missing test mask grid data includes: The missing or infill markers are read point by point according to the features of the unified grid feature data; when the missing or infill marker indicates an original valid value, a valid mask layer is generated; when the missing or infill marker indicates an interpolated infill value, an interpolation mask layer is generated; when the missing or infill marker indicates a backfill value, a backfill mask layer is generated; the valid mask layer, the interpolation mask layer, and the backfill mask layer are combined according to the feature correspondence to form the missing mask grid data; Generating the confidence grid data and obtaining the model input grid data includes: The observation coverage information is determined based on the proportion of effective grid points in the preset spatial neighborhood of the effective mask layer; the interpolation distance information is determined based on the minimum distance from the observation point to the target grid point; the spatial neighborhood consistency test and the temporal adjacency consistency test are performed on the unified grid point feature data to obtain the spatiotemporal consistency test result. Based on the forecast lead time information of the forecast grid data and the lead time decay information determined according to the forecast lead time decay rule, the observation coverage information, the interpolation distance information, the spatiotemporal consistency test result, and the lead time decay information are combined with the lead time decay information to generate confidence grid data according to the configured fusion rule, and the confidence grid data is penalized by the interpolation mask layer and the backfill mask layer; the confidence grid data after the penalization is combined with the unified grid feature data by channel to obtain the model input grid data.

2. The deep learning prediction method for multi-source spatiotemporal point data according to claim 1, characterized in that, Preprocessing data to obtain uniform grid feature data includes: Set the target rule latitude and longitude grid as a unified spatial coordinate system, and set the hourly time as a unified time reference; The real-time grid data, forecast grid data, and static grid data are converted to the unified spatial coordinates; numerical features are resampled to the target regular latitude and longitude grid using bilinear interpolation, and categorical features are resampled to the target regular latitude and longitude grid using nearest neighbor interpolation; the forecast grid data is converted to valid times based on the reporting time and forecast lead time information and aligned to the unified time reference; the real-time grid data is taken from the nearest time value within 30 minutes before and after the unified time reference, and if multiple values ​​exist, the average value is taken; finally, the unified grid feature data is output.

3. The deep learning prediction method for multi-source spatiotemporal point data according to claim 2, characterized in that, When constructing training samples based on a sliding time window, the following is included: Using the predicted times of each target under the unified time reference as alignment points, the model input grid data is selected forward according to the number of consecutive whole-hour times corresponding to the preset window length to form a model input grid data sequence, and the sliding time window is rolled with a step size of one whole-hour time. For each training sample, the corresponding forecast lead time level is determined according to the start time and effective time of the forecast grid data, and the forecast lead time level is converted into forecast lead time code data and written into the training sample; the training samples are grouped according to the forecast lead time level, and equal sampling is performed on each group to obtain the training sample set.

4. The deep learning prediction method for multi-source spatiotemporal point data according to claim 3, characterized in that, Constructing the deep learning prediction model includes: The spatial coding network is configured to perform multi-scale downsampling on the model input grid data at a single hour to extract spatial features, and then output spatial coding features by merging with the upsampling branch at the corresponding scale. The temporal coding network is configured to perform temporal coding on the spatial coding features corresponding to the model input grid data sequence, and to use the forecast lead time coding data as the gating input for temporal coding to modulate the temporal weights of the spatial coding features under different forecast lead time levels, and output temporal coding features.

5. The deep learning prediction method for multi-source spatiotemporal point data according to claim 4, characterized in that, Constructing the forecast time-lead time segmentation prediction network includes: At least three expert prediction subnetworks are set up, corresponding to short-term, medium-term, and long-term forecast lead times, respectively. The gating network is configured to generate fusion weights for each expert prediction subnetwork based on the time-series coding features and the forecast lead time coding data, and the fusion weights are normalized. The outputs of each expert prediction subnetwork are weighted and fused according to the fusion weights to obtain continuous value prediction results. Based on the continuous value prediction results, a hierarchical probability output corresponding one-to-one with the threshold set is generated through a hierarchical output layer, and a monotonic constraint is applied to the hierarchical probability output according to the order of the threshold set.

6. The deep learning prediction method for multi-source spatiotemporal point data according to claim 5, characterized in that, Training the deep learning prediction model includes: Regression weights are assigned to each grid point corresponding to the continuous value prediction result based on the confidence grid data, and upper and lower limits are set for the regression weights to form truncation weights; the regression loss between the continuous value prediction result and the training label is calculated based on the truncation weights; the order relationship consistency of the hierarchical probability output is checked according to the sorting relationship of the threshold set, and when the hierarchical probability outputs corresponding to adjacent thresholds do not satisfy the order of the threshold set, an order conflict penalty term is generated; the model parameters of the deep learning prediction model are updated based on the joint loss of the regression loss and the order conflict penalty term.

7. The deep learning prediction method for multi-source spatiotemporal point data according to claim 6, characterized in that, When reasoning and generating the published prediction results, the process includes: The model input grid data and corresponding forecast lead time coding data for the period to be predicted are input into the trained deep learning prediction model, and the continuous value prediction results and hierarchical probability outputs are output. The matching error between the hierarchical probability output and the training label is statistically analyzed on the validation sample to determine the calibration parameters, and the calibration parameters are set according to the forecast lead time level. The hierarchical probability output is probabilistically calibrated using the calibration parameters corresponding to the forecast lead time level to obtain the calibrated hierarchical probability output. The consistency verification is performed between the calibrated hierarchical probability output and the continuous value prediction result. When the consistency verification passes, the release prediction result is output.

8. A deep learning prediction system for multi-source spatiotemporal point data, used to apply the deep learning prediction method for multi-source spatiotemporal point data as described in any one of claims 1-7, characterized in that, include: The acquisition unit is configured to acquire real-time grid data, forecast grid data and static grid data, and to preprocess the data to obtain unified grid feature data. The preprocessing includes unified spatial coordinates and time reference. The processing unit is configured to generate missing measurement mask grid data based on missing measurement or filling marks, generate confidence grid data based on observation coverage information, interpolation distance information, spatiotemporal consistency test results and forecast lead time decay rules, and concatenate the confidence grid data with the unified grid feature data to obtain model input grid data. The building unit is configured to build training samples based on a sliding time window, such that the training samples include the model input grid data and the forecast timeliness coding data; The training unit is configured to build and train a deep learning prediction model, which includes a spatial encoding network, a temporal encoding network, and a forecast time-leading segmented prediction network. The forecast time-leading segmented prediction network includes a gating network and multiple expert prediction subnetworks. The gating network weights and fuses the multiple expert prediction subnetworks based on the forecast time-leading encoded data to output continuous value prediction results, and outputs hierarchical probability outputs based on a threshold set. During training, a confidence-weighted regression loss is applied to the continuous value prediction results, and an order relation consistency constraint is imposed on the hierarchical probability outputs. The inference unit is configured to output the continuous value prediction result and the hierarchical probability output during inference, and to perform probability calibration on the hierarchical probability output to obtain the published prediction result.

Images