A method for ultra-short-term irradiance prediction based on FY-4B satellite

Abstract

The application belongs to the field of photovoltaic prediction, and discloses a method for super-short-term irradiance prediction based on Fengyun 4B satellite, comprising the following steps: acquiring cloud phase and irradiance data of the Fengyun 4B satellite, constructing a spatial grid based on the longitude and latitude of a target position and extracting grid point data, and respectively constructing data matrices and then preprocessing; converting the preprocessed data into time sequence feature vectors, combining discrete time characteristics to construct model input features; constructing a deep learning model containing a multi-source time sequence modeling module, a time characteristic embedding module and a multi-feature fusion prediction module, and training the model, inputting the feature vectors to infer and output super-short-term irradiance results. The application improves prediction accuracy and model robustness, is suitable for fields such as photovoltaic power generation and weather prediction, and has high engineering application value.

Description

Technical Field

[0001] This invention belongs to the field of photovoltaic forecasting, and specifically relates to a method for ultra-short-term irradiance forecasting based on the Fengyun-4B satellite. Background Technology

[0002] Accurate ultra-short-term solar irradiance forecasts are of great significance to solar power generation, weather forecasting, and agricultural production. In solar power generation, accurate irradiance forecasts can improve the accuracy of photovoltaic power plant power prediction, optimize grid dispatch, and reduce energy losses. In weather forecasting, irradiance is a key parameter affecting temperature and cloud distribution, which can help improve short-term weather forecasting capabilities. In agricultural production, it can assist in planning agricultural activities to increase crop yields, and also provide data support for ecological monitoring and climate change research. Therefore, efficient ultra-short-term irradiance forecasting technology is urgently needed.

[0003] Current mainstream irradiance forecasting technologies have significant shortcomings. Numerical weather prediction models are computationally expensive, and their short-term forecast accuracy is easily affected by initial condition errors; statistical regression methods struggle to capture dynamic changes in cloud cover, resulting in limited forecast accuracy; empirical models based on satellite data rely on empirical formulas and have poor adaptability; even deep learning methods, which have been applied in recent years, suffer from insufficient diversity of input data and inadequate spatiotemporal feature modeling, failing to fully meet the high-precision requirements of ultra-short-term forecasts.

[0004] The Fengyun-4B satellite provides a high-quality data foundation for ultra-short-term irradiance forecasting. Its advanced visible-infrared imager acquires multispectral remote sensing data to extract cloud phase data (characterizing the physical state of clouds), and the ground product service system provides irradiance data (characterizing the shortwave solar radiation energy received by the Earth's surface per unit time), all with high spatiotemporal resolution. However, current technologies have not yet designed suitable preprocessing procedures, feature fusion strategies, and deep learning models for the multimodal data from the Fengyun-4B satellite, thus failing to fully leverage the advantages of this satellite data. How to improve the accuracy of ultra-short-term irradiance forecasting based on this data has become an urgent problem to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an ultra-short-term irradiance forecasting method based on the Fengyun-4B satellite. Based on multi-modal meteorological data from the Fengyun-4B satellite, this method innovatively integrates cloud phase data and irradiance data. By extracting the spatiotemporal dynamic features of the multi-modal data and combining a parallel time-series modeling network and gating mechanisms, it captures the dynamic change patterns and long-term dependencies between data. This effectively solves the problems of insufficient input data diversity and inadequate spatiotemporal feature modeling in traditional irradiance forecasting methods, optimizes the accuracy of ultra-short-term irradiance forecasts, and better meets the needs of high-precision forecasting.

[0006] The technical solution adopted by this invention to solve its technical problem is: to provide a method for ultra-short-term irradiance forecasting based on the Fengyun-4B satellite, comprising:

[0007] Acquire historical multimodal meteorological data from the Fengyun-4B satellite, including cloud phase data and irradiance data;

[0008] Based on the latitude and longitude coordinates of the target predicted location, a spatial grid is constructed with the location as the center, a preset spatial range and a preset resolution. Multimodal meteorological data corresponding to each grid point in the spatial grid is extracted from the multimodal meteorological data of the historical Fengyun 4B satellite.

[0009] Based on the extracted grid point data, cloud phase data matrix and irradiance data matrix are constructed respectively;

[0010] The cloud phase data matrix and irradiance data matrix are preprocessed;

[0011] The preprocessed cloud phase data matrix and irradiance data matrix are converted into feature vectors;

[0012] The model input feature vector is constructed based on the cloud phase temporal feature vector, the irradiance temporal feature vector, and discrete time class features.

[0013] Build a deep learning model and train it.

[0014] The feature vector is input into a pre-built and trained deep learning model for inference, and the irradiance prediction result is output with a preset time step.

[0015] Furthermore, acquiring multimodal meteorological data from the historical Fengyun-4B satellite includes:

[0016] Multispectral remote sensing data were acquired using the Advanced Visible and Infrared Imager carried by the Fengyun-4B satellite.

[0017] Cloud phase data are extracted from multispectral remote sensing observation data based on remote sensing inversion algorithms to represent the physical state of cloud bodies;

[0018] Irradiance data is obtained based on the inversion results provided by the Fengyun-4B satellite ground product service system; the irradiance data is used to characterize the solar shortwave radiation energy received by the Earth's surface per unit time.

[0019] Furthermore, constructing the cloud phase data matrix and irradiance data matrix includes:

[0020] Using the latitude and longitude coordinates of the predicted target location as the center, delineate the target area within a preset latitude and longitude range;

[0021] The target area is divided into grids according to a preset spatial resolution to obtain several grid points;

[0022] Obtain cloud phase data and irradiance data corresponding to each grid point, and construct a cloud phase data matrix and an irradiance data matrix with the spatial distribution of the grid points as the dimension.

[0023] Furthermore, preprocessing includes handling missing values, specifically including:

[0024] Calculate the missing rate of data for each grid point, preset a missing rate threshold, interpolate and supplement data with missing rates lower than the missing rate threshold, and remove data for the corresponding time period with missing rates higher than the threshold.

[0025] Furthermore, preprocessing includes daytime period filtering, specifically including:

[0026] By calculating the solar altitude angle of the target area, daytime and nighttime periods are distinguished, and data from the daytime period is selected for subsequent processing.

[0027] Furthermore, the preprocessed cloud phase data matrix and irradiance data matrix are converted into feature vectors, including:

[0028] The preprocessed cloud phase data matrix is ​​expanded according to the spatial grid dimension and the time dimension. The cloud phase temporal features of each grid point within the preset lag time step are extracted. The discrete cloud phase category values ​​are converted into dense real-valued vectors through encoding to obtain the cloud phase temporal feature vector.

[0029] The preprocessed irradiance data matrix is ​​expanded according to the spatial grid dimension and the time dimension. The temporal features of the irradiance values ​​of each grid point within the preset lag time step are extracted and combined with the normalized irradiance data to form an irradiance temporal feature vector.

[0030] The cloud phase temporal feature vector and the irradiance temporal feature vector are dimensionally aligned to ensure that they are consistent in terms of the number of spatial grids and the length of time steps, and are used as the base vectors for subsequent model input features.

[0031] Furthermore, the model input feature vector is constructed based on the cloud phase temporal feature vector, the irradiance temporal feature vector, and discrete-time class features, including:

[0032] Extract discrete time-related features corresponding to meteorological data within a preset lag time step. The discrete time-related features include the month, hour, and frame sequence ID of the time period. The frame sequence ID is the frame number corresponding to each meteorological data after dividing a day into several frames based on a preset time granularity.

[0033] The discrete-time features are converted into time feature vectors using a preset encoding method, and the batch dimension of the vectors is kept consistent with the cloud phase time-series feature vectors and the irradiance time-series feature vectors.

[0034] Following the splicing rules of space-time-semantic dimensions, the cloud phase temporal feature vector, the irradiance temporal feature vector, and the time feature vector are horizontally spliced ​​together to integrate them into a model input feature vector that contains spatiotemporal dynamic information and temporal context information of multimodal meteorological data.

[0035] Furthermore, building deep learning models includes:

[0036] A multi-source temporal feature modeling module is constructed to process the cloud phase temporal features and irradiance temporal features in the model input feature vector. The module is configured with a parallel temporal modeling network and captures the dynamic change patterns and long-term dependencies of multimodal temporal data through a gating mechanism.

[0037] A discrete-time feature embedding module is constructed to further optimize the expression of discrete-time features, mapping the time feature vectors into semantic vectors that are more suitable for the irradiance prediction task, and maintaining dimensionality adaptation with the output features of the multi-source time-series feature modeling module.

[0038] A multi-feature fusion prediction module is constructed, which adopts a multi-layer fully connected network structure to fuse the temporal feature vector output by the multi-source temporal feature modeling module with the semantic vector output by the discrete-time feature embedding module. The nonlinear expression and generalization capabilities of the model are enhanced by activation functions and regularization operations, and finally the irradiance prediction value dimension corresponding to the preset time step is output.

[0039] Furthermore, the deep learning model is trained, including:

[0040] The constructed model input feature vector and the corresponding measured irradiance label (the actual irradiance value of the period to be predicted) are divided into training set and validation set according to a preset ratio;

[0041] A preset loss function is used to measure the deviation between the model's predicted value and the measured irradiance label. The preset loss function includes the mean squared error loss function.

[0042] Configure a preset optimizer for model parameter updates. The preset optimizer includes the Adam optimizer, and set the initial learning rate and weight decay coefficient.

[0043] Set a learning rate scheduling strategy: if the validation set loss does not decrease for a preset number of consecutive rounds, reduce the learning rate by a preset percentage and limit the minimum learning rate; at the same time, set an early stopping strategy: if the validation set loss does not improve for a preset number of consecutive rounds, terminate model training to avoid model overfitting.

[0044] Furthermore, training deep learning models also includes:

[0045] The model input feature vectors of the training set are input into the constructed deep learning model in batches, and the irradiance prediction value corresponding to each batch of data is calculated through forward propagation.

[0046] The loss value between the predicted value and the measured irradiance label in the training set is calculated based on the preset loss function, and the parameters of each module of the model are updated through the backpropagation algorithm.

[0047] After each round of training, the model input feature vectors of the validation set are input into the model for inference, the model's predictive performance on the validation set is evaluated, and the model parameters of the current round are recorded.

[0048] Repeat the above steps until the early stop strategy is triggered or the preset maximum number of training rounds is reached. Select the parameters with the best performance on the validation set from the recorded model parameters as the final model parameters.

[0049] A cross-validation strategy with a preset number of folds is used to train a corresponding number of independent deep learning models, and the optimal parameters of each model are saved for result integration in the subsequent inference stage.

[0050] The present invention has the following beneficial effects:

[0051] (1) Based on the multimodal meteorological data of Fengyun-4B satellite, this invention innovatively integrates cloud phase data and irradiance data. By extracting the spatiotemporal dynamic features of the multimodal data, and combining the parallel time series modeling network and gating mechanism to capture the dynamic change patterns and long-term dependencies between data, this invention effectively solves the problems of insufficient input data diversity and insufficient spatiotemporal feature modeling in traditional irradiance forecasting methods, optimizes the accuracy of ultra-short-term irradiance forecasting, and better meets the needs of high-precision forecasting.

[0052] (2) This invention significantly improves the quality of the original data through targeted data preprocessing strategies, including interpolation supplementation and invalid data removal based on missing rate, daytime period screening based on solar altitude angle, and unified temporal resolution of multimodal data. During the model training stage, an appropriate loss function and optimizer are adopted, along with dynamic scheduling of learning rate and early stopping strategy. Cross-validation and multi-model integration are combined to avoid overfitting. At the same time, the time-type feature expression is optimized through discrete time feature embedding module, so that the model can better adapt to different time dimensions and scenarios and maintain stable performance under diverse application conditions.

[0053] (3) This invention focuses on the actual needs of fields such as solar power generation, weather forecasting, and agricultural production. The ultra-short-term forecast results can directly provide data support for photovoltaic power plant power prediction and grid dispatch optimization, short-term weather forecast accuracy improvement, and agricultural activity planning. The model reasoning process can be adapted to real-time satellite data processing, and can quickly convert real-time data into feature vectors that meet the input requirements, output standardized time-series irradiance forecast results, which is convenient for engineering application and provides reliable technical support for decision-making in related fields. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of this specification, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 A flowchart illustrating an ultra-short-term irradiance forecasting method based on the Fengyun-4B satellite, provided as an embodiment of this specification;

[0056] Figure 2 A flowchart illustrating an ultra-short-term irradiance forecasting method based on the Fengyun-4B satellite, provided as an embodiment of this specification;

[0057] Figure 3 A flowchart illustrating an ultra-short-term irradiance forecasting method based on the Fengyun-4B satellite, provided as an embodiment of this specification;

[0058] Figure 4 A prediction error analysis diagram of an ultra-short-term irradiance forecasting method based on the Fengyun-4B satellite, provided for an embodiment of this specification;

[0059] Figure 5 A 4-hour prediction effect diagram of an ultra-short-term irradiance forecasting method based on the Fengyun-4B satellite, provided as an embodiment of this specification;

[0060] Figure 6 The image shows the 6-hour prediction effect of an ultra-short-term irradiance forecasting method based on the Fengyun-4B satellite, which is provided as an embodiment of this specification. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] The terms "first," "second," "third," etc., in the description, claims, and accompanying drawings are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such processes, methods, products, or apparatus.

[0063] In the following description, terms such as “inner,” “outer,” “upper,” “lower,” “left,” and “right” are used only to facilitate the description of the embodiments and to simplify the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this specification.

[0064] All data involved in this application is information and data authorized by the user or fully authorized by all parties, and the collection of relevant data complies with the relevant laws, regulations and standards of the relevant countries and regions.

[0065] Example 1:

[0066] A method for ultra-short-term irradiance forecasting based on the Fengyun-4B satellite, such as Figure 1 As shown, it includes:

[0067] Acquire historical multimodal meteorological data from the Fengyun-4B satellite, including cloud phase data and irradiance data;

[0068] Based on the latitude and longitude coordinates of the target predicted location, a spatial grid is constructed with the location as the center, a preset spatial range and a preset resolution. Multimodal meteorological data corresponding to each grid point in the spatial grid is extracted from the multimodal meteorological data of the historical Fengyun 4B satellite.

[0069] Based on the extracted grid point data, cloud phase data matrix and irradiance data matrix are constructed respectively;

[0070] The cloud phase data matrix and irradiance data matrix are preprocessed;

[0071] The preprocessed cloud phase data matrix and irradiance data matrix are converted into feature vectors;

[0072] The model input feature vector is constructed based on the "cloud phase temporal feature vector, irradiance temporal feature vector, and discrete time-type features";

[0073] Build a deep learning model and train it.

[0074] The model input feature vector is fed into a pre-built and trained deep learning model for inference, and the output is the irradiance prediction result with a preset time step.

[0075] Example 2: Satellite multimodal data acquisition.

[0076] The original data required for this invention all come from the FY-4B satellite and its ground service system. The specific acquisition methods and data parameters are as follows:

[0077] Multispectral remote sensing data acquisition: The Advanced Visible Infrared Imager (AGRI) carried by the FY-4B satellite collects multispectral remote sensing data of the target area. This data contains information in multiple bands such as visible light and infrared, which is the basis for extracting cloud phase data.

[0078] Cloud phase data extraction: Based on a remote sensing inversion algorithm, specifically a thresholding method using AGRI multi-band reflectance and brightness temperature combined with a machine learning classifier, cloud phase product data is extracted from multispectral remote sensing data to characterize the physical states of clouds, including categories such as water clouds, ice clouds, and mixed clouds. The spatial resolution of this cloud phase data is 0.04° × 0.04°, and the temporal resolution is 60 minutes.

[0079] Irradiance data acquisition: The surface solar irradiance (SSI) inversion results provided by the FY-4B satellite ground product service system are directly called. This data is used to characterize the solar shortwave radiation energy received by the surface per unit time, with a spatial resolution of 0.1°×0.1° and a temporal resolution of 15 minutes.

[0080] Example 3:

[0081] Based on the latitude and longitude coordinates of the target predicted location, a spatial grid is constructed with the location as the center, a preset spatial range and a preset resolution. Multimodal meteorological data corresponding to each grid point in the spatial grid is extracted from the multimodal meteorological data of the historical Fengyun 4B satellite.

[0082] Based on the extracted grid point data, cloud phase data matrix and irradiance data matrix are constructed respectively.

[0083] Determine the target area range: Based on the latitude and longitude coordinates of the target predicted location input by the user (denoted as (φ0,λ0), where φ0 is latitude and λ0 is longitude), take this point as the center and expand ±1° in both the longitude and latitude directions to form a rectangular target area with "latitude range [φ0-1°, φ0+1°] and longitude range [λ0-1°, λ0+1°]".

[0084] Spatial grid division: The target area is divided into grids with a spatial resolution of 0.1°. Since the range of latitude / longitude is 2° (from "-1°" to "+1°"), the resulting spatial grid has a dimension of "21×21" (calculation logic: 2°÷0.1°+1=21), containing a total of 441 spatial grid points;

[0085] Grid point data extraction and matrix construction: Traverse each node of the 21×21 grid and record the latitude and longitude coordinates (φ) of each node. ij ,λ ij ), where i=1,2,...,21 (latitude direction index), j=1,2,...,21 (longitude direction index);

[0086] Based on the historical satellite data obtained in Example 2, the cloud phase category data and irradiance value corresponding to each grid point (φᵢⱼ, λᵢⱼ) are extracted;

[0087] Using the spatial distribution of grid points as the dimension, the cloud phase data of 441 points are organized into a 21×21 two-dimensional matrix (denoted as C, where C(i,j) represents the cloud phase category of the grid point in the i-th row and j-th column), and the irradiance data of 441 points are organized into a 21×21 two-dimensional matrix (denoted as S, where S(i,j) represents the irradiance value of the grid point in the i-th row and j-th column), thus completing the construction of the spatial matrix of the target area.

[0088] Example 4: Preprocess the cloud phase data matrix and irradiance data matrix.

[0089] Because the cloud phase data and irradiance data of FY-4B satellite have differences in spatiotemporal resolution, and the raw data may have problems such as missing data and invalid data at night, data preprocessing is required to improve data quality.

[0090] (1) Downsampling: Because the temporal resolution of cloud phase data (60 minutes) is inconsistent with that of irradiance data (15 minutes), the cloud phase data needs to be downsampled: linear interpolation is used to resample the temporal resolution of cloud phase data from 60 minutes to 15 minutes to ensure that the two types of data are aligned in the time dimension. For example, if the cloud phase category in a certain hour is "water cloud", then the hour is divided into 4 15-minute time steps, and the cloud phase category of each time step is inherited as "water cloud"; if the cloud phase category changes in adjacent hours (e.g., "water cloud" at t=0 and "ice cloud" at t=60 minutes), then the cloud phase category at t=15 minutes, t=30 minutes, and t=45 minutes is determined by linear interpolation.

[0091] Preprocessing includes missing value handling, specifically including: calculating the missing rate of data for each grid point, setting a missing rate threshold, interpolating and supplementing data with a missing rate lower than the missing rate threshold, and removing data for the corresponding time period with a missing rate higher than the threshold.

[0092] (2) Handling missing values: The original satellite data may be missing due to observation obstruction, transmission errors, etc., and should be handled according to the following strategies:

[0093] Missing rate calculation: For the time series data of each grid point, cloud phase and irradiance are calculated separately, and the missing rate is calculated using the following formula:

[0094]

[0095] Missing value imputation and data removal:

[0096] The preset missing rate threshold is 50%: If the missing rate of a certain grid point is less than 50%, the missing value is supplemented using a rolling time window linear interpolation method—the rolling window size is set to 2N (N is 3, that is, the window contains 3 valid data points before and after the missing point). If the number of missing values ​​in the 2N+1=7 data points within the window is ≤3 (interpolation threshold L=3), then the linear interpolation formula is used:

[0097]

[0098] Where, x t-k x t+k Here, x represents the nearest valid data point before and after the missing point, k is the time step interval between the valid data point and the missing point, and d is the distance between the missing point and x. t-k Time step distance;

[0099] If the missing rate of a certain grid point is higher than 50%, all data for the corresponding time period of that grid point will be directly removed to avoid too many missing values ​​affecting model training.

[0100] (3) Preprocessing includes daytime period filtering, specifically: distinguishing between daytime and nighttime periods by calculating the solar altitude angle of the target area, and filtering out the daytime period data for subsequent processing.

[0101] Nighttime irradiance is always 0, which can interfere with model training. Therefore, daytime data needs to be filtered by calculating the solar altitude angle.

[0102] Formula for calculating solar altitude angle:

[0103]

[0104] in:

[0105] h is the solar altitude angle (unit: °);

[0106] δ is the solar declination (unit: °, calculated according to the date, such as summer solstice δ≈23.5°, winter solstice δ≈-23.5°);

[0107] φ represents the latitude of the predicted target location (unit: °, positive for North latitude and negative for South latitude).

[0108] ω is the local time angle of the predicted target location (unit: °, ω=0° at noon, changes by 15° per hour, negative in the morning and positive in the afternoon);

[0109] Day / night time criteria: When h > 0°, it is determined to be a daytime period, and the cloud phase and irradiance data for that period are retained; when h ≤ 0°, it is determined to be a nighttime period, and the data for that period are discarded.

[0110] (4) Data normalization: To eliminate the impact of differences in data units on model training, the irradiance data is normalized using Max-min, as shown in the following formula:

[0111] First, calculate the maximum value (x) of the irradiance data for all grid points in the training set. max ) and minimum value (x) min ):

[0112] x max =max{S(i,j,t)|i=1…21,j=1…21,t=1…T};

[0113] x min =min{S(i,j,t)|i=1…21,j=1…21,t=1…T}.

[0114] Where T is the total number of time steps in the training set;

[0115] Normalize the irradiance value S(i,j,t) at time step t for each grid point (i,j):

[0116]

[0117] Normalized data S norm (i,j,t) is mapped to the interval [0,1] for subsequent feature engineering and model input.

[0118] Example 5: Convert the preprocessed cloud phase data matrix and irradiance data matrix into feature vectors; construct the model input feature vector based on the "cloud phase time-series feature vector, irradiance time-series feature vector, and discrete-time class features". After data preprocessing, feature engineering is needed to transform the raw meteorological data into structured features adapted to the deep learning model. The core includes time feature construction, meteorological lag feature construction, classification feature encoding, and feature dimension alignment to ensure that the model can fully capture the spatiotemporal dynamics and semantic relationships of multimodal data.

[0119] (1) Construction of time characteristics.

[0120] Temporal features are a key factor influencing the periodic changes in irradiance. They need to be constructed from three aspects: basic field extraction, periodic feature transformation, and frame sequence ID generation, as detailed below:

[0121] Extracting the basic time field.

[0122] Extract the five basic time fields (year, month, day, hour, and minute) from the timestamps of the preprocessed data:

[0123] "Month" (denoted as month): Used to characterize seasonal differences. For example, irradiance is high in summer (June-August) and low in winter (December-February);

[0124] "Hour": Used to depict intraday periodic variations. For example, peak irradiance between noon and 2 PM, and lower irradiance in the early morning / evening.

[0125] "Day" (denoted as "day"): helps identify special dates, such as periods with a high concentration of rainy days.

[0126] The year and minute fields are used as auxiliary features to eliminate interference from year-end trends and refine the time positioning of the 15-minute time step.

[0127] (2) Construction of periodic time features.

[0128] For time fields with cyclical attributes, such as "hour" and "month," trigonometric function transformation is used to map them into continuous periodic signals, solving the problem that "the values ​​of 23:00 and 1:00 are far apart but their time sequences are adjacent." The specific formula is as follows:

[0129] Periodic processing of hourly fields:

[0130] ;

[0131] The value of hour ranges from 0 to 23. It is converted into a continuous value in the interval [-1, 1] by sine and cosine functions, thus preserving the 24-hour cycle within the day.

[0132] Periodic processing of the month field:

[0133]

[0134] The value of month ranges from 1 to 12. It is converted into a continuous value in the interval [-1, 1] using trigonometric functions to depict the seasonal cycle of the 12 months of the year.

[0135] (3) Frame sequence ID generation and encoding.

[0136] Based on the 15-minute time resolution of the data, a day (24 hours) is divided into 96 frames (calculation logic: 24 × 60 ÷ 15 = 96), and a unique frame sequence ID (denoted as frame) is assigned to each data sample. id ):

[0137] frame id The value range is 0-95, corresponding to 00:00 (frame) of a day. id =0), 00:15 (frame id =1), ..., 23:45 (frame id =95);

[0138] For frame id Embedding is used for processing, through a preset embedding layer nn.Embedding(cardinality=96, embedding) dim =16) is mapped to a 16-dimensional dense real-valued vector, enabling the model to perceive high-frequency temporal change trends at the 15-minute granularity within a day.

[0139] (4) Meteorological characteristics processing.

[0140] To capture the temporal dependence of irradiance on cloud phase, a lag meteorological feature is introduced, which uses historical time step data to predict future irradiance, including:

[0141] Lag step size determination: Combining domain experience and experimental optimization, select the first 1-4 time steps (corresponding to the past 1 hour, each step is 15 minutes) as the lag window, that is, for each target time step t, extract meteorological data of four historical time steps t-1, t-2, t-3, t-4;

[0142] Lag Feature Assembly:

[0143] For irradiance data: For each grid point (i,j), the normalized irradiance value S at the current time step t is... norm (i,j,t) and the S of the historical time step norm (i,j,t-1),S norm (i,j,t-2),S norm (i,j,t-3),S norm (i,j,t-4) are concatenated to form the irradiance time series S of this grid point. seq (i,j)=[S norm (i,j,t-4),S norm (i,j,t-3),S norm (i,j,t-2),S norm (i,j,t-1),Snorm (i,j,t)];

[0144] For cloud phase data: For each grid point (i,j), the cloud phase category C(i,j,t) at the current time step t is concatenated with the cloud phase categories C(i,j,t-1), C(i,j,t-2), C(i,j,t-3), and C(i,j,t-4) at historical time steps to form the cloud phase time series sequence C for that grid point. seq (i,j)=[C(i,j,t-4),C(i,j,t-3),C(i,j,t-2),C(i,j,t-1),C(i,j,t)];

[0145] Sequence Dimension Unfolding: Unfold the irradiance time series S of all points in a 21×21 grid. seq (i,j) is expanded according to "spatial grid dimension (21×21) + time step dimension (5)" to form an irradiance temporal feature matrix of shape [5,441]; similarly, the cloud phase temporal sequence C is expanded according to "spatial grid dimension (21×21) + time step dimension (5)". seq (i,j) is expanded into a cloud phase temporal feature matrix of shape [5,441].

[0146] (5) Classification feature coding (cloud phase and time category).

[0147] For discrete classification features (cloud phase, month, hour, frame) id This needs to be converted into a dense real-valued vector that the model can recognize. The specific encoding method is as follows:

[0148] Cloud phase coding: Cloud phases include four discrete categories: "water cloud, ice cloud, mixed cloud, and cloudless". One-hot encoding is used for processing.

[0149] Assign a unique binary vector to each category: water cloud → [1,0,0,0], ice cloud → [0,1,0,0], mixed cloud → [0,0,1,0], cloudless → [0,0,0,1];

[0150] The encoded cloud phase vector is combined with the corresponding time sequence to expand the original cloud phase time sequence feature matrix [5,441] into a three-dimensional feature tensor [5,441,4].

[0151] Time-based classification feature encoding: for month, hour, frame id The three time-related classification features are processed using embedding encoding:

[0152] Month (12 categories): mapped to a 16-dimensional vector using nn.Embedding(12,16);

[0153] Hour (24 categories): mapped to a 16-dimensional vector using nn.Embedding(24,16);

[0154] Frame Sequence ID (frame id (96 categories): mapped to a 16-dimensional vector using nn.Embedding(96,16);

[0155] The three embedding vectors are concatenated horizontally to form a time feature vector with dimensions [16+16+16=48], denoted as Time. Emb .

[0156] (6) Feature dimension alignment.

[0157] To ensure that multimodal features can be adapted to the input of subsequent deep learning models, the cloud phase temporal feature vector and the irradiance temporal feature vector need to be dimensionally aligned:

[0158] Spatial grid dimension alignment: Both are built based on 21×21 (441) grid points, and their spatial dimensions are naturally consistent (both are 441).

[0159] Time step dimension alignment: Both use "the previous 4 historical steps + the current step" for a total of 5 time steps, with the same time dimension (5 for both).

[0160] Feature dimension integration:

[0161] The cloud phase state three-dimensional feature tensor [5,441,4] is flattened into a one-dimensional vector according to "time step × spatial grid × encoding dimension", i.e., Cloud Feature =reshape([5,441,4]), the dimension is 5×441×4=8820;

[0162] Flatten the time-series feature matrix of irradiance [5,441] into a one-dimensional vector, i.e., Irradiance Feature =reshape([5,441]), the dimension is 5×441=2205;

[0163] Ultimately, a basic feature set of "cloud phase features + irradiance features + temporal embedding features" is formed, providing data support for the subsequent construction of model input feature vectors.

[0164] Example 6: Constructing and training a deep learning model; such as Figure 2 and Figure 3As shown, after feature engineering is completed, deep learning models are used to achieve deep fusion of multimodal features and ultra-short-term irradiance prediction. The core process includes three main stages: model structure design, model training and optimization, and model inference. All parameters and procedures are designed based on the characteristics of Fengyun-4B satellite multimodal data and forecast mission requirements to ensure the accuracy and engineering feasibility of the model. The specific implementation steps are as follows:

[0165] This invention designs a three-level serial model architecture of "temporal modeling-feature embedding-fusion prediction", which is adapted to multimodal inputs of cloud phase, irradiance and time features. Specifically, it includes a multi-source temporal feature modeling module, a discrete time feature embedding module and a multi-feature fusion prediction module. The functions and structural details of each module are as follows.

[0166] (1) Multi-source temporal feature modeling module (Multi-GRU).

[0167] This module is used to capture the temporal dynamic changes and long-term dependencies between cloud phase and irradiance. It adopts a parallel GRU network structure (3 independent GRU channels, respectively processing irradiance, cloud phase, and irradiance-cloud phase cross features). Specific design details are as follows:

[0168] Input feature specifications: The input is a temporal feature tensor generated during the feature engineering stage, with a shape of [batch_size, num_gru_layers, seq_len, input_size], where:

[0169] batch_size: Training batch size, set to 32 based on hardware resources (adjustable);

[0170] num_gru_layers: Number of parallel GRU channels, set to 3 (corresponding to 3 types of time-series features);

[0171] seq_len: Time step length, set to 4 (corresponding to the past 4 15-minute time steps, i.e. 1 hour of historical data);

[0172] input_size: The feature dimension of each time step, which is equal to the number of spatial grid points (21×21=441).

[0173] GRU network parameters: Each GRU channel adopts a 2-layer stacked structure, the number of hidden units is set to 128, the activation function is tanh, and a gating mechanism (reset gate and update gate) is introduced to capture long-term dependencies. The formula is as follows (taking a single GRU unit as an example):

[0174] Reset Door: ;

[0175] Update Gate: ;

[0176] Candidate hidden state: ;

[0177] Hidden state output: ;

[0178] Where, x t Given the input features at time t, h t-1 W is the hidden state at time t-1. r W z W h Let b be the weight matrix. r ,b z ,b h Here, δ is the bias term, δ is the Sigmoid activation function, and ʘ is the element-wise product;

[0179] Output feature specifications: Each GRU channel outputs the hidden state of the last time step (shape [batch_size, 128]). The three channels are concatenated to form a temporal feature vector of [batch_size, 384], denoted as TimeSeq. Feature .

[0180] (2) Discrete-time feature embedding module.

[0181] This module is used to optimize time-related features (month, hour, frame). id The semantic expression of ) is adapted to the irradiance prediction task, and the specific design is as follows:

[0182] Input feature specifications: The input is a discrete-time feature vector generated during the feature engineering stage, containing 3 categories of features (month: 12 categories, hour: 24 categories, frame ... frame: 12 categories, frame: 12 categories, frame: 12 categories, frame: 12 categories, frame: 12 categories, frame: id : 96 categories), with a shape of [batch] size ,3];

[0183] Embedding layer parameters: Each category feature is mapped to a dense real-valued vector through an independent nn.Embedding layer, with the parameters set as follows:

[0184] Month embedding: nn.Embedding(cardinality=12, embedding) dim =16), input [batch_size,1], output [batch_size,16];

[0185] Hourly embedding: nn.Embedding(cardinality=24, embedding_dim=16), input [batch_size,1], output [batch_size,16];

[0186] Frame sequence ID embedding: nn.Embedding(cardinality=96,embedding_dim=16), input [batch_size,1], output [batch_size,16];

[0187] Feature projection and fusion: The three embedding vectors are concatenated horizontally (shape [batch_size, 48]), and semantic projection is performed through a fully connected layer nn.Linear(48, 32) and the GELU activation function. The final output is a time semantic vector of [batch_size, 32], denoted as TimeSemantic_Feature, which ensures that it is adapted to the output dimension of the multi-source time series feature module.

[0188] (3) Multi-feature fusion prediction module.

[0189] This module fuses temporal features and temporal semantic features using a multilayer perceptron (MLP) to output a predicted value for future irradiance. Specific design details are as follows:

[0190] Input feature fusion: The TimeSeq_Feature (384-dimensional) and TimeSemantic_Feature (32-dimensional) are concatenated horizontally to form a high-dimensional fused feature vector of [batch_size, 416].

[0191] MLP network structure: It adopts a "2 hidden layers + 1 output layer" structure, with specific parameters as follows:

[0192] First hidden layer: nn.Linear(416,256) + GELU activation + nn.BatchNorm1d(256) (batch normalization) + nn.Dropout(0.2) (regularization to prevent overfitting);

[0193] Second hidden layer: nn.Linear(256,128) + GELU activation + nn.BatchNorm1d(128) + nn.Dropout(0.2);

[0194] Output layer: nn.Linear(128,24), with an output dimension of 24 (corresponding to 24 future 15-minute time steps, i.e., 6-hour ultra-short-term forecast).

[0195] Output constraints: The output value is constrained by the ReLU activation function nn.ReLU() to ensure that the irradiance prediction value is non-negative (in accordance with physical meaning, irradiance ≥ 0).

[0196] (4) Model training process.

[0197] The model was trained using supervised learning, with measured irradiance as the label, and the model parameters were optimized using a loss function.

[0198] Dataset partitioning and label definition.

[0199] Data partitioning: The model input feature vector generated by feature engineering and the corresponding measured irradiance label (the actual irradiance value of the forecast period) are divided into training set and validation set in a 7:3 ratio to ensure consistent data distribution and partitioning in time series order to avoid data leakage;

[0200] Input features: The fused high-dimensional feature vector of [batch_size, 416];

[0201] Measured labels: The actual irradiance values ​​for the next 24 time steps, in shape [batch_size,24]. The labels must be aligned with the time steps of the input features (e.g., if the input is the features from time t-4 to time t, the labels are the measured values ​​from time t+1 to time t+24).

[0202] Data normalization: All numerical features (irradiance time series features) were normalized using Max-Min before training (same as the formula in Section 2.4), with the normalization parameter (x) max ,x min Statistics are only collected from the training set to avoid leaking validation set information.

[0203] Training parameters and optimization strategies.

[0204] Loss function: The mean squared error loss function (MSELoss) is used to measure the deviation between the predicted value and the measured label. The formula is as follows:

[0205] ;

[0206] Where N is the sample size, H=24 is the prediction time step, and y i,k Let i be the measured value of the i-th sample at time k. This corresponds to the predicted value;

[0207] Optimizer: Adam optimizer selected, parameter settings:

[0208] Initial learning rate: lr=10 -3 ;

[0209] Weight decay factor: weight decay}=10 -4 (Suppress overfitting);

[0210] Momentum parameters: β1=0.9, β2=0.999;

[0211] Learning rate scheduling: The "ReduceLROnPlateau" strategy is adopted. If the validation set loss does not decrease for 5 consecutive epochs, the learning rate is multiplied by 0.5, and the minimum learning rate is set to 10. -5 To avoid subsequent fluctuations;

[0212] Early stopping strategy: Set the early stopping epoch to 20. If the validation set loss does not improve for 20 consecutive epochs, then terminate the training, save the model parameters with the best performance on the validation set before the current epoch, and prevent overfitting.

[0213] Training process and parameter saving.

[0214] Forward propagation: Input the training set features into the model in batches (batch_size=32), and calculate the irradiance prediction value for each batch through each module of the model;

[0215] Loss calculation and backpropagation: The loss value between the predicted value and the training set label is calculated based on MSELoss, and the weight parameters of each module of the model (GRU weights, embedding layer weights, MLP weights) are updated through the backpropagation algorithm (automatic differentiation).

[0216] Validation and parameter recording: After each epoch of training, input the validation set features into the model for inference, calculate the validation set loss and prediction accuracy (refer to the accuracy formula in the handover document, constructed based on RMSE), and record the model parameters for the current epoch.

[0217] Cross-validation and model ensemble: A 5-fold cross-validation strategy is adopted, dividing the training set into 5 independent subsets. Four subsets are used as the training set and one subset as the validation set to train 5 independent models. The optimal parameters of each model are saved for result ensemble in the inference stage.

[0218] Example 7: Input the model input feature vector into a pre-built and trained deep learning model for inference, and output the irradiance prediction result for a preset time step.

[0219] Model reasoning process.

[0220] After the model training is completed, the saved optimal parameters are loaded to perform ultra-short-term irradiance forecasting. The inference process strictly follows the logic of "feature input - module calculation - output processing", and the specific steps are as follows:

[0221] Inference data preparation.

[0222] Real-time data acquisition: Acquire the latest hourly Fengyun-4B satellite data (4 15-minute time steps) of the target's predicted location, including cloud phase data (21×21 grid) and irradiance data (21×21 grid), and extract the corresponding time features (month, hour, frame). id );

[0223] Data preprocessing and feature generation: Repeat the processes in Sections 2.1-2.4 (data downsampling, missing value interpolation, daytime filtering, normalization) and Sections 3.1-3.4 (temporal feature construction, lag feature construction, classification feature encoding, dimension alignment) for real-time data to generate feature vectors that meet the model input specifications (temporal features [1,3,4,441], time features [1,3], batch_size=1, single-sample inference).

[0224] Multi-step reasoning calculation.

[0225] Step 1: Category Embedding and Projection. Input the time features (month, hour, frame_id) into the discrete time feature embedding module:

[0226] The three categorical features are mapped to 16-dimensional vectors by nn.Embedding(12,16), nn.Embedding(24,16), and nn.Embedding(96,16), respectively, and then concatenated to form [1,48].

[0227] The time semantic vector TimeSemantic_Feature is output by nn.Linear(48,32) and GELU activation, which is in the range of [1,32].

[0228] Step 2: Multi-channel GRU sequence encoding inputs temporal features into the multi-source temporal feature modeling module:

[0229] Three parallel GRU channels process different temporal features respectively. Each GRU inputs [1,4,441] and outputs a hidden state of [1,128].

[0230] The three GRU outputs are concatenated horizontally to obtain the time-series feature vector TimeSeq_Feature in the range of [1,384].

[0231] Step 3: Multimodal Feature Fusion and MLP Prediction. This involves fusing temporal and temporal semantic features and outputting predicted values ​​via MLP.

[0232] The TimeSeq_Feature (384-dimensional) and TimeSemantic_Feature (32-dimensional) are concatenated to form a [1,416] fused feature;

[0233] The original predicted value of [1,24] is obtained by passing the MLP through the first hidden layer (416→256, GELU+BN+Dropout), the second hidden layer (256→128, GELU+BN+Dropout), and the output layer (128→24).

[0234] Step 4: Output Layer Processing

[0235] Apply the ReLU activation function to the original predicted values ​​to ensure that all predicted values ​​are ≥0;

[0236] The normalized predicted values ​​are then converted back to the actual irradiance values ​​using the "inverse normalization formula," which is: ;

[0237] Output the final result: the predicted irradiance values ​​for the next 24 15-minute time steps (T+0h15m to T+6h00m).

[0238] Verification of reasoning results: After the reasoning is completed, if... Figure 4 , Figure 5 and Figure 6 As shown, after completing the inference, the model performance needs to be verified by combining real-time measured data from ground-based irradiance observation stations at the target location. Figure 4 The mean absolute error analysis chart for the predicted time clearly shows the error change trend from T+0h00m to T+5h45m. The first significant increase in error occurs at T+0h00m. The mean absolute error and error range for different prediction times are also clearly presented. Figure 5 A scatter plot of the 4-hour forecast results shows an accuracy of 92.85%, with the scatter points closely distributed near the ideal forecast line, intuitively reflecting the high degree of fit between the predicted and actual values. If the prediction error exceeds a preset threshold during the validation process, the parameters of other independent deep learning models saved in the cross-validation stage are reloaded for secondary inference, and the optimal result is selected as the final forecast value. This process significantly improves the reliability of the ultra-short-term irradiance forecasting method in engineering applications, and can better meet the actual needs for forecast accuracy in fields such as photovoltaic power generation and weather forecasting.

[0239] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for ultra-short-term irradiance forecasting based on the Fengyun-4B satellite, characterized in that, include: Acquire historical multimodal meteorological data from the Fengyun-4B satellite, including cloud phase data and irradiance data; Based on the latitude and longitude coordinates of the target's predicted location, a spatial grid is constructed with the target's predicted location as the center, a preset spatial range, and a preset resolution. Multimodal meteorological data corresponding to each grid point within the spatial grid is extracted from the multimodal meteorological data of the historical Fengyun 4B satellite. Based on the extracted grid point data, cloud phase data matrix and irradiance data matrix are constructed respectively; The cloud phase data matrix and irradiance data matrix are preprocessed; The preprocessed cloud phase data matrix and irradiance data matrix are converted into feature vectors; The model input feature vector is constructed based on the cloud phase temporal feature vector, the irradiance temporal feature vector, and discrete time class features. Build a deep learning model and train it. The model input feature vector is fed into a pre-built and trained deep learning model for inference, and the irradiance prediction result is output for a preset time step. The construction of deep learning models includes: A multi-source temporal feature modeling module is constructed to process the cloud phase temporal features and irradiance temporal features in the model input feature vector. The multi-source temporal feature modeling module is configured with a parallel temporal modeling network, which captures the dynamic change patterns and long-term dependencies of multimodal temporal data through a gating mechanism. A discrete-time feature embedding module is constructed to further optimize the expression of discrete-time features, mapping the time feature vectors into semantic vectors that are more suitable for the irradiance prediction task, and maintaining dimensionality adaptation with the output features of the multi-source time-series feature modeling module. A multi-feature fusion prediction module is constructed, which adopts a multi-layer fully connected network structure to fuse the temporal feature vector output by the multi-source temporal feature modeling module with the semantic vector output by the discrete-time feature embedding module. The nonlinear expression and generalization capabilities of the model are enhanced by activation functions and regularization operations, and finally the irradiance prediction value dimension corresponding to the preset time step is output.

2. The ultra-short-term irradiance forecasting method based on the Fengyun-4B satellite according to claim 1, characterized in that, The acquisition of multimodal meteorological data from the historical Fengyun-4B satellite includes: Multispectral remote sensing data were acquired using the Advanced Visible and Infrared Imager carried by the Fengyun-4B satellite. Cloud phase data are extracted from multispectral remote sensing observation data based on remote sensing inversion algorithms to represent the physical state of cloud bodies; Irradiance data is obtained based on the inversion results provided by the Fengyun-4B satellite ground product service system; the irradiance data is used to characterize the solar shortwave radiation energy received by the Earth's surface per unit time.

3. The ultra-short-term irradiance forecasting method based on the Fengyun-4B satellite according to claim 1, characterized in that, Constructing the cloud phase data matrix and irradiance data matrix includes: Using the latitude and longitude coordinates of the predicted target location as the center, delineate the target area within a preset latitude and longitude range; The target area is divided into grids according to a preset spatial resolution to obtain several grid points; Obtain cloud phase data and irradiance data corresponding to each grid point, and construct a cloud phase data matrix and an irradiance data matrix with the spatial distribution of the grid points as the dimension.

4. The ultra-short-term irradiance forecasting method based on the Fengyun-4B satellite according to claim 1, characterized in that, Preprocessing includes handling missing values, specifically including: Calculate the missing rate of data for each grid point, preset a missing rate threshold, interpolate and supplement data with missing rates lower than the missing rate threshold, and remove data for the corresponding time period with missing rates higher than the threshold.

5. The ultra-short-term irradiance forecasting method based on the Fengyun-4B satellite according to claim 1, characterized in that, Preprocessing includes daytime period filtering, specifically including: By calculating the solar altitude angle of the target area, daytime and nighttime periods are distinguished, and data from the daytime period is selected for subsequent processing.

6. The ultra-short-term irradiance forecasting method based on the Fengyun-4B satellite according to claim 1, characterized in that, The preprocessed cloud phase data matrix and irradiance data matrix are converted into feature vectors, including: The preprocessed cloud phase data matrix is ​​expanded according to the spatial grid dimension and the time dimension. The cloud phase temporal features of each grid point within the preset lag time step are extracted. The discrete cloud phase category values ​​are converted into dense real-valued vectors through encoding to obtain the cloud phase temporal feature vector. The preprocessed irradiance data matrix is ​​expanded according to the spatial grid dimension and the time dimension. The temporal features of the irradiance values ​​of each grid point within the preset lag time step are extracted and combined with the normalized irradiance data to form an irradiance temporal feature vector. The cloud phase temporal feature vector and the irradiance temporal feature vector are dimensionally aligned to ensure that they are consistent in terms of the number of spatial grids and the length of time steps, and are used as the base vectors for subsequent model input features.

7. The ultra-short-term irradiance forecasting method based on the Fengyun-4B satellite according to claim 1, characterized in that, The model input feature vector is constructed based on the cloud phase temporal feature vector, the irradiance temporal feature vector, and discrete-time class features, including: Extract discrete time-related features corresponding to meteorological data within a preset lag time step. The discrete time-related features include the month, hour, and frame sequence ID corresponding to the preset lag time step. The frame sequence ID is the frame number corresponding to each meteorological data after dividing a day into several frames based on a preset time granularity. The discrete-time features are converted into time feature vectors using a preset encoding method, and the batch dimension of the time feature vectors is kept consistent with the cloud phase time-series feature vectors and the irradiance time-series feature vectors. Following the splicing rules of space-time-semantic dimensions, the cloud phase temporal feature vector, the irradiance temporal feature vector, and the time feature vector are horizontally spliced ​​together to integrate them into a model input feature vector that contains spatiotemporal dynamic information and temporal context information of multimodal meteorological data.

8. The ultra-short-term irradiance forecasting method based on the Fengyun-4B satellite according to claim 1, characterized in that, Training a deep learning model includes: The constructed model input feature vector and the corresponding measured irradiance label are divided into training set and validation set according to a preset ratio; A preset loss function is used to measure the deviation between the model's predicted value and the measured irradiance label. The preset loss function includes the mean squared error loss function. Configure a preset optimizer for model parameter updates. The preset optimizer includes the Adam optimizer, and set the initial learning rate and weight decay coefficient. Set a learning rate scheduling strategy: if the validation set loss does not decrease for a preset number of consecutive rounds, reduce the learning rate by a preset percentage and limit the minimum learning rate; at the same time, set an early stopping strategy: if the validation set loss does not improve for a preset number of consecutive rounds, terminate model training to avoid model overfitting.

9. The ultra-short-term irradiance forecasting method based on the Fengyun-4B satellite according to claim 8, characterized in that, Training a deep learning model also includes: S1. Input the model input feature vector of the training set into the constructed deep learning model in batches, and calculate the irradiance prediction value corresponding to each batch of data through forward propagation; S2. Calculate the loss value between the predicted value and the measured irradiance label of the training set based on the preset loss function, and update the parameters of each module of the model through the backpropagation algorithm; S3. After each round of training, input the model input feature vector of the validation set into the model for inference, evaluate the model's prediction performance on the validation set, and record the model parameters for the current round. S4. Repeat steps S1, S2 and S3 until the early stop strategy is triggered or the preset maximum number of training rounds is reached. Select the parameters with the best performance on the validation set from the recorded model parameters as the final model parameters. S5. Employ a cross-validation strategy with a preset number of folds to train a corresponding number of independent deep learning models, and save the optimal parameters of each model for result integration in the subsequent inference stage.

Images