An ultra-short-term solar radiation prediction method, device, equipment and storage medium

Abstract

The application discloses a kind of ultra-short-term solar radiation prediction methods, comprising: collecting the meteorological observation data of target radiation area, the meteorological observation data includes satellite cloud picture data, radar data and ground meteorological data;Using the pre-constructed preprocessing multichannel and feature extraction multichannel, respectively, the meteorological observation data is preprocessed and feature extraction is carried out, and the feature data extracted is fused, and meteorological fusion feature data is obtained;Using the pre-constructed target domain solar radiation predictor, the meteorological fusion feature data is predicted and analyzed, and the ultra-short-term solar radiation prediction result is obtained;The application can more accurately analyze the relationship between meteorological fusion feature data and solar radiation, significantly improve the timeliness and accuracy of ultra-short-term solar radiation prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of solar radiation, and in particular to a super-short-term solar radiation prediction method, device, equipment and storage medium. BACKGROUND

[0002] In the field of new energy power generation, accurate prediction of solar radiation is crucial for efficient operation of photovoltaic power stations and grid scheduling. Existing super-short-term solar radiation prediction methods rely on a single meteorological data source, which is difficult to fully capture the impact of weather changes on solar radiation. Moreover, the model has poor adaptability to different regions, often resulting in insufficient timeliness and low accuracy of prediction, which cannot meet the high requirements of super-short-term prediction in practical applications. SUMMARY

[0003] The embodiments of the present application provide a super-short-term solar radiation prediction method, which can more accurately analyze the relationship between meteorological fusion feature data and solar radiation, significantly improving the timeliness and accuracy of super-short-term solar radiation prediction.

[0004] In a first aspect, the embodiments of the present application provide a super-short-term solar radiation prediction method, comprising:

[0005] Collecting meteorological observation data of a target radiation area, the meteorological observation data including satellite cloud image data, radar data and ground meteorological data;

[0006] Using a pre-constructed preprocessing multi-channel and a feature extraction multi-channel, respectively preprocessing and feature extracting the meteorological observation data, and fusing the extracted feature data to obtain meteorological fusion feature data;

[0007] Using a pre-constructed target domain solar radiation predictor, performing prediction analysis on the meteorological fusion feature data to obtain a super-short-term solar radiation prediction result.

[0008] Further, the preprocessing and feature extraction of the meteorological observation data respectively include:

[0009] The meteorological observation data is preprocessed respectively, and the preprocessing includes noise removal of the satellite cloud image data, clutter suppression of the radar data, and missing value filling of the ground meteorological data;

[0010] Obtaining associated parameters related to solar radiation of the preprocessed meteorological observation data, the associated parameters including cloud top temperature and cloud cover ratio in the satellite cloud image data, echo top height and vertical cumulative liquid water content in the radar data, and sunshine duration and visibility in the ground meteorological data;

[0011] Feature extraction is performed on the associated parameters to obtain multi-dimensional associated feature data.

[0012] Further, the extracted feature data is fused to obtain meteorological fusion feature data, including:

[0013] A spatial reference coordinate is set, and the extracted multi-dimensional correlation feature data is mapped into the spatial reference coordinate to obtain spatial feature data;

[0014] A timestamp identifier is added to each spatial feature point in the spatial feature data;

[0015] According to a preset time granularity, a time window is divided, and the spatial feature data is classified into a corresponding time window to obtain spatiotemporal feature data;

[0016] The spatiotemporal feature data is weighted and fused to obtain meteorological fusion feature data.

[0017] Further, the construction process of the target domain solar radiation predictor includes:

[0018] The source domain dataset and the target domain dataset are determined, and the source domain dataset and the target domain dataset are associated analyzed to determine a transfer learning strategy;

[0019] The source domain solar radiation predictor is generated by using the source domain dataset to perform radiation prediction training on a pre-constructed LSTM model;

[0020] Based on the prediction target of the ultra-short-term solar radiation and the target domain dataset, the source domain solar radiation predictor is transferred and trained using the transfer learning strategy to generate a target domain solar radiation predictor.

[0021] Further, the associated analysis of the source domain dataset and the target domain dataset to determine a transfer learning strategy includes:

[0022] The cosine similarity and the KL divergence of the core meteorological features of the source domain dataset and the target domain dataset are calculated;

[0023] Based on the combination result of the cosine similarity and the KL divergence, the correlation degree level of the source domain dataset and the target domain dataset is determined;

[0024] According to the correlation degree level, a corresponding transfer learning strategy is selected from a plurality of preset candidate transfer learning strategies.

[0025] Further, the source domain dataset includes source domain historical meteorological data and source domain historical solar radiation data of different radiation regions, and the source domain solar radiation predictor is generated by using the source domain dataset to perform radiation prediction training on a pre-constructed LSTM model, including:

[0026] Construct an LSTM neural network model that includes an input layer, hidden layers, and an output layer;

[0027] The LSTM neural network model is trained by taking the historical meteorological data of the source region as input and the historical solar radiation data of the source region as output.

[0028] When the model reaches the preset accuracy requirement, training stops, and a trained source region solar radiation predictor is obtained.

[0029] Furthermore, the target domain dataset includes historical meteorological data and historical solar radiation data of the target domain. Then, the prediction target based on ultra-short-term solar radiation and the target domain dataset are used to perform transfer learning training on the source domain solar radiation predictor to generate a target domain solar radiation predictor, including:

[0030] Using the feature vectors of the aforementioned meteorological fusion feature data as the retrieval benchmark, a search is performed in the meteorological database of the target radiation area to obtain historical meteorological data of the target domain.

[0031] Obtain the historical solar radiation data of the target domain corresponding to the historical meteorological data of the target domain;

[0032] Based on the prediction target of ultra-short-term solar radiation, the historical meteorological data of the target domain is used as input and the historical solar radiation data of the target domain is used as output. The transfer learning strategy is used to transfer train the source domain solar radiation predictor to generate the target domain solar radiation predictor.

[0033] Secondly, embodiments of the present invention provide an ultra-short-term solar radiation prediction device, comprising:

[0034] The data acquisition module is used to collect meteorological observation data of the target radiation area, including satellite cloud image data, radar data and ground meteorological data;

[0035] The data processing module is used to preprocess and extract features from the meteorological observation data using pre-built preprocessing multi-channel and feature extraction multi-channel, respectively, and to fuse the extracted feature data to obtain meteorological fusion feature data.

[0036] The radiation prediction module is used to perform predictive analysis on the meteorological fusion feature data using a pre-built target domain solar radiation predictor to obtain ultra-short-term solar radiation prediction results.

[0037] Thirdly, embodiments of the present invention provide an electronic device, comprising:

[0038] Memory, used to store computer programs;

[0039] A processor for executing the computer program;

[0040] Wherein, when the processor executes the computer program, it implements the ultra-short-term solar radiation prediction method described in any of the first aspects above.

[0041] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program that, when executed, implements the ultra-short-term solar radiation prediction method described in any of the first aspects above.

[0042] This invention provides an ultra-short-term solar radiation prediction method. The method involves collecting meteorological observation data of a target radiation area, including satellite cloud imagery, radar data, and ground meteorological data. The meteorological observation data is preprocessed and feature extracted, and the extracted feature data is fused to obtain meteorological fusion feature data. A pre-constructed target domain solar radiation predictor is then used to predict and analyze the meteorological fusion feature data to obtain ultra-short-term solar radiation prediction results. Compared with existing technologies, this invention can more accurately analyze the relationship between meteorological fusion feature data and solar radiation, significantly improving the timeliness and accuracy of ultra-short-term solar radiation prediction. Attached Figure Description

[0043] To more clearly illustrate the technical features of the embodiments of the present invention, the drawings used in the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a flowchart illustrating an ultra-short-term solar radiation prediction method provided in an embodiment of the present invention;

[0045] Figure 2 This is a schematic diagram of the structure of an ultra-short-term solar radiation prediction device provided in an embodiment of the present invention;

[0046] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0047] 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.

[0048] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0049] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein is for the purpose of describing embodiments of the invention only and is not intended to limit the invention.

[0050] In a first aspect, embodiments of the present invention provide a method for predicting ultra-short-term solar radiation, see [link to previous document]. Figure 1 This is a flowchart illustrating an embodiment of an ultra-short-term solar radiation prediction method provided by the present invention.

[0051] like Figure 1 As shown, the method includes the following steps:

[0052] S1: Collect meteorological observation data of the target radiation area, including satellite cloud image data, radar data and ground meteorological data;

[0053] S2: Using pre-built preprocessing multi-channel and feature extraction multi-channel, the meteorological observation data is preprocessed and features are extracted respectively, and the extracted feature data is fused to obtain meteorological fusion feature data;

[0054] S3: Using a pre-built target domain solar radiation predictor, predictive analysis is performed on the meteorological fusion feature data to obtain ultra-short-term solar radiation prediction results.

[0055] In practice, meteorological observation data of the target radiation area is collected through meteorological observation equipment and data receiving systems. This includes satellite cloud image data, radar data, and surface meteorological data. Satellite cloud image data covers image information such as the distribution, thickness, type, and movement trajectory of clouds over the target area. Radar data contains monitoring data reflecting the structure and evolution of precipitation systems, such as the echo intensity, radial velocity, and spectral width of precipitation particles in the atmosphere of the area. Surface meteorological data involves real-time and short-time series data of near-surface meteorological elements such as temperature, humidity, air pressure, wind speed, wind direction, and sunshine duration recorded by ground observation stations in the target area. This forms a multi-source, multi-dimensional short-term meteorological correlation data set, providing basic data support for subsequent solar radiation prediction and processing.

[0056] A multi-channel preprocessing and feature extraction system for meteorological observation data is pre-built. Satellite cloud image data, radar data, and ground meteorological data are preprocessed and feature extracted separately. The extracted feature data are then fused to obtain meteorological fusion feature data. A pre-built target domain solar radiation predictor is used to predict and analyze the meteorological fusion feature data to obtain ultra-short-term solar radiation prediction results.

[0057] In summary, this invention avoids the limitations of a single data source by simultaneously acquiring multi-source data from satellite cloud images, radar, and ground meteorological data, providing a more comprehensive and accurate information foundation for forecasting. Furthermore, by utilizing pre-built preprocessing and feature extraction multi-channels to process and fuse the data, key features are further extracted, enhancing the data's expressiveness and usability. Finally, by employing a pre-built target domain solar radiation predictor for forecasting analysis, the relationship between meteorological fusion feature data and solar radiation can be analyzed more accurately, significantly improving the timeliness and accuracy of ultra-short-term solar radiation forecasts. This provides more reliable technical support for the rational utilization of solar energy, meteorological disaster early warning, and other related fields.

[0058] In one optional implementation, the preprocessing and feature extraction of the meteorological observation data respectively include:

[0059] The meteorological observation data are preprocessed respectively, including noise removal for the satellite cloud image data, clutter suppression for the radar data, and missing value imputation for the ground meteorological data;

[0060] The preprocessed meteorological observation data is used to obtain correlation parameters related to solar radiation. These correlation parameters include cloud top temperature and cloud cover percentage from the satellite cloud image data, echo top height and vertical cumulative liquid water content from the radar data, and sunshine duration and visibility from the ground meteorological data.

[0061] Feature extraction is performed on the correlation parameters to obtain multidimensional correlation feature data.

[0062] Specifically, based on the source dimensional characteristics of satellite cloud images, radar, and ground meteorological data, dimensional differentiated processing is implemented to construct preprocessing multi-channels and feature extraction multi-channels, ultimately forming a complete short-term meteorological processing multi-channel. The short-term meteorological processing multi-channel is used to preprocess and extract features from the meteorological observation data to obtain multidimensional correlated feature data. The specific construction process of the preprocessing multi-channel and feature extraction multi-channel is as follows:

[0063] First, considering the differences in dimensional characteristics among satellite cloud imagery, radar, and ground meteorological data, differentiated anomaly identification rules are designed to accurately mark single-dimensional anomalies. For example, the anomaly identification rule for satellite cloud imagery data is pixels where the cloud top temperature exceeds a reasonable range; the anomaly identification rule for radar data is abnormal jumps in echo intensity; and the anomaly identification rule for ground meteorological data is observations of a certain element value deviating from three times the standard deviation. Then, based on a spatiotemporal coordinate matching association algorithm, correlation analysis is performed on the above-mentioned single-dimensional anomaly data. Multi-dimensional anomaly data with significant correlations are integrated to finally determine the short-term meteorological multi-dimensional anomaly data.

[0064] To address the anomalous characteristics of the three types of data in the aforementioned multidimensional anomaly data, a multidimensional independent preprocessing workflow is constructed based on a pre-defined preprocessing rule base, forming a multi-channel data preprocessing system. Specifically, satellite cloud image anomaly data mainly exhibits edge noise and spatial distortion, with corresponding preprocessing steps of image denoising and geometric correction. Radar anomaly data mainly exhibits ground clutter and echo attenuation, with corresponding preprocessing steps of clutter suppression and echo calibration. Ground meteorological anomaly data mainly exhibits missing data and outliers, with corresponding preprocessing steps of missing value imputation and outlier replacement. The above three preprocessing workflows are encapsulated according to the source dimension to form a multi-channel data preprocessing system, including a satellite sub-channel preprocessing module, a radar sub-channel preprocessing module, and a ground sub-channel preprocessing module, which can independently preprocess the anomaly data of the corresponding dimension.

[0065] Furthermore, based on the source dimension characteristics of satellite cloud images, radar, and ground meteorological data, the intrinsic relationship between these three types of data and solar radiation is correlated to obtain multidimensional correlated radiation parameters, including satellite cloud image correlated radiation parameters (including cloud top temperature and cloud cover ratio), radar correlated radiation parameters (including echo top height and vertical cumulative liquid water content), and ground correlated radiation parameters (including sunshine duration and visibility). For these correlated radiation parameters, combined with the characteristic features of the three types of data, corresponding extraction algorithms are adopted to obtain multidimensional correlated radiation features. For example, image segmentation algorithms can be used for satellite cloud image data, signal filtering algorithms can be used for radar data, and time series analysis algorithms can be used for ground meteorological data. These feature extraction algorithms are categorized according to source dimension, forming a satellite cloud image feature extraction algorithm set, a radar data feature extraction algorithm set, and a ground data feature extraction algorithm set. These three algorithm sets are encapsulated as independent modules and integrated to form a feature extraction multi-channel, corresponding one-to-one with the sub-channels of the data preprocessing multi-channel. The sub-channels of the data preprocessing multi-channel and the corresponding algorithm sets of the feature extraction multi-channel are then cascaded and merged to construct a complete short-term meteorological processing multi-channel.

[0066] In one optional implementation, the step of fusing the extracted feature data to obtain meteorological fused feature data includes:

[0067] Set spatial reference coordinates, and map the extracted multidimensional correlation feature data to the spatial reference coordinates to obtain spatial feature data;

[0068] Add a timestamp identifier to each spatial feature point in the spatial feature data;

[0069] The spatial feature data is divided into time windows according to a preset time granularity, and then classified into the corresponding time windows to obtain spatiotemporal feature data.

[0070] The spatiotemporal feature data are weighted and fused to obtain meteorological fusion feature data.

[0071] Specifically, firstly, the central latitude and longitude of the target radiation area is selected as the origin to establish a unified spatial reference coordinate system, such as using the UTM projection coordinate system, and the coordinate units and projection parameters are specified. The three types of multidimensional associated feature data extracted are then mapped to the same spatial dimension. For example, a projection transformation algorithm is used to convert the pixels corresponding to features such as "cloud cover percentage" and "cloud top temperature" extracted from satellite cloud images into UTM coordinates to achieve spatial positioning of satellite feature data. The polar coordinate to rectangular coordinate conversion formula is used to convert feature data such as "echo intensity" and "echo top height" extracted from radar data into UTM rectangular coordinate data. For feature data such as "sunshine hours" and "humidity" extracted from ground meteorological data, they are directly mapped to the spatial reference coordinates based on the latitude and longitude coordinates of the ground observation station. Through the above mapping, the three types of feature data are unified into the spatial reference coordinate system to generate spatial feature data containing "spatial coordinates + feature values".

[0072] Based on the generated spatial feature data, according to the original acquisition time of the three types of feature data (e.g., the imaging time of satellite cloud images, the scanning time of radar, and the observation time of ground meteorological stations), a corresponding second-level precision timestamp is added to each spatial feature point. According to the time granularity requirements of the ultra-short-term solar radiation prediction target, such as every 30 minutes in the next 6 hours as a prediction unit, time windows are divided, and spatial feature data with different timestamps are assigned to the corresponding windows in chronological order. Through interpolation or matching, it is ensured that each time window contains spatial feature data from three sources: satellite, radar, and ground, achieving alignment in the time dimension. Finally, spatiotemporal feature data of "spatial coordinates + time window + feature value" is generated.

[0073] Furthermore, for the three dimensions of meteorological data in the spatiotemporal feature data—satellite cloud imagery, radar, and ground meteorology—historical measured solar radiation data from the same season and time period are selected as the benchmark dataset. Deviation indices between the three types of feature data and the benchmark dataset are calculated to obtain the accuracy coefficients for the short-term meteorological dimension data. For example, for satellite cloud imagery spatiotemporal feature data, the Pearson correlation coefficients between "cloud cover percentage" and "cloud top temperature" and historical measured solar radiation data are calculated, and the average value is taken as the satellite dimension correlation coefficient. For radar spatiotemporal feature data, the Spearman rank correlation coefficients between "echo intensity" and "echo top height" and historical measured solar radiation data are calculated, and the average value is taken as the radar dimension correlation coefficient. For ground meteorological feature data, the average absolute error between "sunshine hours" and "humidity" and historical measured solar radiation data is calculated, converted into error coefficients, and the average value is taken as the ground dimension error coefficient. The above evaluation indicators are normalized and mapped to the 0-1 range. Then, a weighted average is applied (e.g., the weight of correlation coefficient indicators is 0.6, and the weight of error indicators is 0.4) to obtain the comprehensive accuracy score for each dimension, i.e., the accuracy coefficient.

[0074] Furthermore, key indicators of short-term meteorological spatiotemporal characteristic data (such as cloud cover, precipitation intensity, visibility, etc.) are analyzed. Based on preset scene classification rules, the current short-term meteorological scene type is determined. A dynamic weight allocation model is established based on the accuracy coefficient and scene type. For example, in a clear weather scene, ground meteorological data is assigned 40% weight due to its high accuracy coefficient, satellite cloud image data is assigned 35% weight, and radar data is assigned 25% weight. In a precipitation scene, the weight of radar data is increased to 50%, satellite cloud image data accounts for 30%, and ground meteorological data accounts for 20%. Through this model, the weights of each dimension of data are calculated and adjusted, and finally, the decision weights of the short-term meteorological dimension data reflecting the importance of each dimension of data in the current scene are determined.

[0075] Based on the established decision weights, the feature information of satellite cloud images, radar, and surface meteorology in the short-term meteorological spatiotemporal feature data is weighted and fused. For feature data of the same point in space and the same window in time, the feature value of satellite cloud image is multiplied by its decision weight, the feature value of radar is multiplied by its decision weight, and the feature value of surface meteorology is multiplied by its decision weight. Then, these three weighted feature values ​​are summed. The above weighted fusion operation is performed on feature data of all time windows and all spatial points to form a unified feature set containing "spatiotemporal coordinates + fused feature value", that is, meteorological fused feature data.

[0076] In one alternative implementation, the process of constructing the target domain solar radiation predictor includes:

[0077] Determine the source domain dataset and the target domain dataset, perform correlation analysis on the source domain dataset and the target domain dataset, and determine the transfer learning strategy;

[0078] The source domain dataset is used to train a pre-built LSTM model for radiation prediction, generating a source domain solar radiation predictor.

[0079] Based on the prediction target of ultra-short-term solar radiation and the target domain dataset, the source domain solar radiation predictor is transferred and trained using the transfer learning strategy to generate the target domain solar radiation predictor.

[0080] Specifically, a target domain solar radiation predictor is constructed using a transfer learning method. First, the source domain dataset and the target domain dataset are determined. By performing correlation analysis on the source domain dataset and the target domain dataset, the transfer learning strategy is determined, an LSTM neural network model is constructed, and the LSTM model is trained for radiation prediction using the source domain dataset to generate the source domain solar radiation predictor. Based on the prediction target of ultra-short-term solar radiation and the target domain dataset, the transfer learning strategy is used to transfer train the source domain solar radiation predictor to generate the target domain solar radiation predictor.

[0081] In one optional implementation, the step of performing correlation analysis on the source domain dataset and the target domain dataset to determine the transfer learning strategy includes:

[0082] Calculate the cosine similarity and KL divergence of the core meteorological features between the source domain dataset and the target domain dataset;

[0083] Based on the combined results of the cosine similarity and the KL divergence, the correlation level between the source domain dataset and the target domain dataset is determined;

[0084] Based on the correlation level, a corresponding transfer learning strategy is selected from a set of preset candidate transfer learning strategies.

[0085] Specifically, core meteorological features are extracted from the source and target domain datasets, such as cloud cover features from satellite cloud images, radar echo features, and surface temperature and humidity features. Each feature is normalized to eliminate the influence of dimensions. The cosine similarity and KL divergence of the two datasets on each feature distribution are calculated to quantitatively evaluate the inter-domain differences between the two datasets. Based on the combined results of cosine similarity and KL divergence, the correlation level between the source and target domain datasets is determined. According to the correlation level, a corresponding transfer learning strategy is selected from a set of pre-defined candidate transfer learning strategies. For example, if the cosine similarity... If the similarity is higher than 0.7 and the KL divergence is lower than 0.3, the correlation between the two is considered high, and a transfer learning strategy with parameter fine-tuning is adopted, adjusting only the top-level network parameters of the source domain predictor. If the cosine similarity is between 0.4 and 0.7 and the KL divergence is between 0.3 and 0.6, the correlation is considered moderate, and a feature transfer strategy is adopted, using the feature extraction layer of the source domain predictor combined with the target domain data to train a new output layer. If the cosine similarity is lower than 0.4 and the KL divergence is higher than 0.6, the correlation is considered low, and a domain adaptation strategy is adopted, introducing an adversarial network to reduce inter-domain differences and improve model adaptability.

[0086] In one optional implementation, the source domain dataset includes historical meteorological data and historical solar radiation data for different radiation regions. Then, the step of using the source domain dataset to train a pre-built LSTM model for radiation prediction to generate a source domain solar radiation predictor includes:

[0087] Construct an LSTM neural network model that includes an input layer, hidden layers, and an output layer;

[0088] The LSTM neural network model is trained by taking the historical meteorological data of the source region as input and the historical solar radiation data of the source region as output.

[0089] When the model reaches the preset accuracy requirement, training stops, and a trained source region solar radiation predictor is obtained.

[0090] Specifically, based on the prediction target of ultra-short-term solar radiation (e.g., predicting solar radiation values ​​for the next 1 hour, 3 hours, or 6 hours), historical meteorological data covering different radiation zones such as tropical, temperate, and frigid zones are screened from meteorological databases. Short-term meteorological characteristic data of these zones are extracted (including cloud cover and cloud top temperature from satellite cloud images, echo intensity and echo top height from radar, and surface meteorological data such as temperature, humidity, and air pressure). These data are then matched with historical measured solar radiation data for the corresponding time period (e.g., irradiance data recorded by radiation observation stations). The screened data are cleaned to remove outliers and missing values, ensuring that the sample size for each radiation zone is balanced and covers various meteorological conditions such as sunny, cloudy, overcast, and precipitation. Finally, the data is integrated into a source domain dataset.

[0091] A neural network model consisting of an input layer, hidden layers (using LSTM units), and an output layer is constructed. Historical meteorological data from the source domain dataset is used as input, and corresponding historical solar radiation data is used as output. Network parameters are set (e.g., 64 hidden layer units, learning rate of 0.001, and 500 iterations). The LSTM network is trained using the Adam optimizer and mean squared error loss function. During training, the predicted values ​​are calculated through forward propagation, and the network weights are adjusted through backpropagation to minimize the loss function. An early stopping strategy is adopted to prevent overfitting. Training is stopped when the model's loss on the validation set tends to stabilize and reaches the preset accuracy requirement, thus generating a source domain solar radiation predictor that can predict solar radiation based on short-term meteorological feature data.

[0092] In one optional implementation, the target domain dataset includes historical meteorological data and historical solar radiation data of the target domain. Then, the target domain solar radiation predictor, based on ultra-short-term solar radiation prediction and the target domain dataset, is trained using the transfer learning strategy on the source domain solar radiation predictor to generate a target domain solar radiation predictor, including:

[0093] Using the feature vectors of the aforementioned meteorological fusion feature data as the retrieval benchmark, a search is performed in the meteorological database of the target radiation area to obtain historical meteorological data of the target domain.

[0094] Obtain the historical solar radiation data of the target domain corresponding to the historical meteorological data of the target domain;

[0095] Based on the prediction target of ultra-short-term solar radiation, the historical meteorological data of the target domain is used as input and the historical solar radiation data of the target domain is used as output. The transfer learning strategy is used to transfer train the source domain solar radiation predictor to generate the target domain solar radiation predictor.

[0096] Specifically, using the feature vector of meteorological fusion feature data (including spatial coordinates, timestamps, and meteorological feature weights of each dimension) as the retrieval benchmark, the K-Nearest Neighbors (KNN) algorithm is used to search the meteorological database of the target radiation area. The top 100 historical records with the highest similarity to the current meteorological fusion feature vector are used as historical meteorological data of the target domain. The solar radiation measurement data corresponding to these historical meteorological data of the target domain are extracted. The extracted data is cleaned to remove records with missing or abnormal solar radiation data. Finally, the data is organized in chronological order to form a target domain dataset containing historical meteorological data and historical solar radiation data of the target domain.

[0097] Based on the prediction target of ultra-short-term solar radiation, historical meteorological data of the target domain is used as input and historical solar radiation data of the target domain is used as output. These are then compiled into training samples. A model parameter transfer approach is adopted, using the LSTM network structure and trained parameters of the source domain solar radiation predictor as a basis to construct a target domain solar radiation predictor. Specifically, if a parameter fine-tuning strategy is adopted, 80% of the bottom LSTM network parameters of the source domain solar radiation predictor are frozen, and only the top 20% of the parameters are updated. Iterative training is then performed using target domain data with a small learning rate (e.g., 0.0001). If a feature transfer strategy is adopted, the feature extraction layer parameters of the source domain model are fixed, and a new fully connected output layer is trained based on the target domain feature data extracted by it. The output layer weights are optimized to adapt to the radiation patterns of the target region. If a domain adaptation strategy is adopted, a domain classifier is introduced on the basis of the source domain model. Through adversarial training, the feature extraction layer learns domain-invariant features while minimizing the solar radiation prediction loss. During the training process, the prediction accuracy of the validation set is used as an indicator. Training is stopped when the accuracy does not improve for several consecutive rounds. Finally, a target domain solar radiation predictor adapted to the target region is obtained.

[0098] In one optional implementation, the radiation prediction time granularity, i.e. the time interval between each prediction, is determined according to the prediction target of ultra-short-term solar radiation. For example, if the prediction target is the change in solar radiation in the next hour, the time granularity can be set to 10 minutes to achieve a more refined prediction. If the prediction target is the next 6 hours, the time granularity can be set to 30 minutes to balance prediction efficiency and accuracy. The determination of the time granularity must ensure that the dynamic changes of solar radiation in the prediction period can be accurately captured, and at the same time, it should match the update frequency of short-term meteorological data to provide a unified time reference for subsequent dataset partitioning and model training.

[0099] Based on the determined radiation prediction time granularity, the target domain dataset is partitioned using a sliding window method, cutting continuous time series data into several equal-length sequence segments. Each segment contains historical meteorological data and historical solar radiation data for the target domain at that time granularity. Simultaneously, a timestamp (such as start time and end time) and a sequence number are added to each sequence segment to clarify its position in the historical time series, ultimately forming a structured target domain dataset. Each sample can fully reflect the correlation between meteorological characteristics and solar radiation within that time granularity. The structured target domain dataset is then divided into... The samples are sorted in descending order of timestamp, with the most recent sample at the beginning of the sequence. As time progresses, the samples are sorted sequentially backward. Weights are then assigned based on the order of the samples, using a gradient-increasing weight distribution rule. For example, the weight of the first sample is set to 1.0, the weight of the second sample is set to 0.95, and so on, decreasing by 0.05 for each subsequent position (with a minimum weight of 0.3). This ensures that the most recent samples at the beginning of the sequence receive a larger training weight. In this way, the sample weight factor reflecting the temporal correlation of the samples is calculated.

[0100] The target domain solar radiation predictor is dynamically adjusted and optimized based on the weighting factors of sample data. Specifically, the weighting factors are integrated into the loss function calculation during model training, so that the prediction error of recent high-weight samples accounts for a larger proportion of the loss value. During backpropagation, the weights and biases of the LSTM network are adjusted according to the weighted loss value, so that the model focuses more on learning the solar radiation change patterns contained in recent samples. Through multiple rounds of iterative training, the network parameters are continuously optimized, so that the target domain solar radiation predictor can better adapt to the recent meteorological characteristics and solar radiation correlation patterns of the target area, thereby improving the accuracy of ultra-short-term solar radiation prediction.

[0101] Secondly, embodiments of the present invention provide an ultra-short-term solar radiation prediction device, see [link to previous document]. Figure 2 This is a schematic diagram of an embodiment of an ultra-short-term solar radiation prediction device provided by the present invention.

[0102] like Figure 2 As shown, the device includes:

[0103] Data acquisition module 21 is used to collect meteorological observation data of the target radiation area, including satellite cloud image data, radar data and ground meteorological data;

[0104] The data processing module 22 is used to preprocess and extract features from the meteorological observation data using a pre-built preprocessing multi-channel and feature extraction multi-channel, and to fuse the extracted feature data to obtain meteorological fusion feature data.

[0105] The radiation prediction module 23 is used to perform predictive analysis on the meteorological fusion feature data using a pre-built target domain solar radiation predictor to obtain ultra-short-term solar radiation prediction results.

[0106] In an optional implementation, the data processing module 22 is further configured to:

[0107] The meteorological observation data are preprocessed respectively, including noise removal for the satellite cloud image data, clutter suppression for the radar data, and missing value imputation for the ground meteorological data;

[0108] The preprocessed meteorological observation data is used to obtain correlation parameters related to solar radiation. These correlation parameters include cloud top temperature and cloud cover percentage from the satellite cloud image data, echo top height and vertical cumulative liquid water content from the radar data, and sunshine duration and visibility from the ground meteorological data.

[0109] Feature extraction is performed on the correlation parameters to obtain multidimensional correlation feature data.

[0110] In an optional implementation, the data processing module 22 is further configured to:

[0111] Set spatial reference coordinates, and map the extracted multidimensional correlation feature data to the spatial reference coordinates to obtain spatial feature data;

[0112] Add a timestamp identifier to each spatial feature point in the spatial feature data;

[0113] The spatial feature data is divided into time windows according to a preset time granularity, and then classified into the corresponding time windows to obtain spatiotemporal feature data.

[0114] The spatiotemporal feature data are weighted and fused to obtain meteorological fusion feature data.

[0115] In one alternative implementation, the process of constructing the target domain solar radiation predictor includes:

[0116] Determine the source domain dataset and the target domain dataset, perform correlation analysis on the source domain dataset and the target domain dataset, and determine the transfer learning strategy;

[0117] The source domain dataset is used to train a pre-built LSTM model for radiation prediction, generating a source domain solar radiation predictor.

[0118] Based on the prediction target of ultra-short-term solar radiation and the target domain dataset, the source domain solar radiation predictor is transferred and trained using the transfer learning strategy to generate the target domain solar radiation predictor.

[0119] In one optional implementation, the step of performing correlation analysis on the source domain dataset and the target domain dataset to determine the transfer learning strategy includes:

[0120] Calculate the cosine similarity and KL divergence of the core meteorological features between the source domain dataset and the target domain dataset;

[0121] Based on the combined results of the cosine similarity and the KL divergence, the correlation level between the source domain dataset and the target domain dataset is determined;

[0122] Based on the correlation level, a corresponding transfer learning strategy is selected from a set of preset candidate transfer learning strategies.

[0123] In one optional implementation, the source domain dataset includes historical meteorological data and historical solar radiation data for different radiation regions. Then, the step of using the source domain dataset to train a pre-built LSTM model for radiation prediction to generate a source domain solar radiation predictor includes:

[0124] Construct an LSTM neural network model that includes an input layer, hidden layers, and an output layer;

[0125] The LSTM neural network model is trained by taking the historical meteorological data of the source region as input and the historical solar radiation data of the source region as output.

[0126] When the model reaches the preset accuracy requirement, training stops, and a trained source region solar radiation predictor is obtained.

[0127] In one optional implementation, the target domain dataset includes historical meteorological data and historical solar radiation data of the target domain. Then, the target domain solar radiation predictor, based on ultra-short-term solar radiation prediction and the target domain dataset, is trained using the transfer learning strategy on the source domain solar radiation predictor to generate a target domain solar radiation predictor, including:

[0128] Using the feature vectors of the aforementioned meteorological fusion feature data as the retrieval benchmark, a search is performed in the meteorological database of the target radiation area to obtain historical meteorological data of the target domain.

[0129] Obtain the historical solar radiation data of the target domain corresponding to the historical meteorological data of the target domain;

[0130] Based on the prediction target of ultra-short-term solar radiation, the historical meteorological data of the target domain is used as input and the historical solar radiation data of the target domain is used as output. The transfer learning strategy is used to transfer train the source domain solar radiation predictor to generate the target domain solar radiation predictor.

[0131] Thirdly, embodiments of the present invention provide an electronic device, see [link to previous document]. Figure 3 The diagram shown is a structural schematic of an electronic device provided in an embodiment of the present invention.

[0132] like Figure 3 As shown, the device includes:

[0133] Memory 31 is used to store computer programs;

[0134] Processor 32 is used to execute the computer program;

[0135] When the processor 32 executes the computer program, it implements the ultra-short-term solar radiation prediction method as described in any of the above embodiments.

[0136] For example, the computer program may be divided into one or more modules / units, which are stored in the memory 31 and executed by the processor 32 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0137] The processor 32 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0138] The memory 31 can be used to store the computer programs and / or modules. The processor 32 implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory 31 and calling the data stored in the memory 31. The memory 31 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory 31 may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital card (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0139] It should be noted that the aforementioned electronic devices include, but are not limited to, processors and memory, as will be understood by those skilled in the art. Figure 3 The structural diagram is merely an example of the electronic device described above and does not constitute a limitation on the electronic device. It may include more components than shown in the diagram, or combine certain components, or use different components.

[0140] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing a computer program that, when executed, implements the ultra-short-term solar radiation prediction method described in any of the above embodiments.

[0141] It should be understood that the present invention can implement all or part of the processes in the above-described ultra-short-term solar radiation prediction method by instructing related hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the above-described ultra-short-term solar radiation prediction method. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0142] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. It should be noted that, for those skilled in the art, several equivalent obvious modifications and / or equivalent substitutions can be made without departing from the technical principles of the present invention, and these obvious modifications and / or equivalent substitutions should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting ultra-short-term solar radiation, characterized in that, include: Meteorological observation data of the target radiation area are collected, including satellite cloud image data, radar data, and ground meteorological data; The meteorological observation data are preprocessed and feature extracted using pre-built preprocessing multi-channel and feature extraction multi-channel, respectively. The extracted feature data are then fused to obtain meteorological fusion feature data. Using a pre-built target domain solar radiation predictor, the meteorological fusion feature data is used to perform predictive analysis to obtain ultra-short-term solar radiation prediction results. The construction process of the target domain solar radiation predictor includes: Determine the source domain dataset and the target domain dataset, perform correlation analysis on the source domain dataset and the target domain dataset, and determine the transfer learning strategy; The source domain dataset is used to train a pre-built LSTM model for radiation prediction, generating a source domain solar radiation predictor. Based on the prediction target of ultra-short-term solar radiation and the target domain dataset, the source domain solar radiation predictor is transferred and trained using the transfer learning strategy to generate the target domain solar radiation predictor. The step of performing correlation analysis on the source domain dataset and the target domain dataset to determine the transfer learning strategy includes: Calculate the cosine similarity and KL divergence of the core meteorological features between the source domain dataset and the target domain dataset; Based on the combined results of the cosine similarity and the KL divergence, the correlation level between the source domain dataset and the target domain dataset is determined; Based on the correlation level, a corresponding transfer learning strategy is selected from a plurality of preset candidate transfer learning strategies; When the correlation level is high, a transfer learning strategy with fine-tuned parameters is adopted. When the correlation level is medium, a feature transfer strategy is adopted. When the relevance level is low, a domain-adaptive strategy is adopted.

2. The ultra-short-term solar radiation prediction method as described in claim 1, characterized in that, The preprocessing and feature extraction of the meteorological observation data include: The meteorological observation data are preprocessed respectively, including noise removal for the satellite cloud image data, clutter suppression for the radar data, and missing value imputation for the ground meteorological data; The preprocessed meteorological observation data is used to obtain correlation parameters related to solar radiation. These correlation parameters include cloud top temperature and cloud cover percentage from the satellite cloud image data, echo top height and vertical cumulative liquid water content from the radar data, and sunshine duration and visibility from the ground meteorological data. Feature extraction is performed on the correlation parameters to obtain multidimensional correlation feature data.

3. The ultra-short-term solar radiation prediction method as described in claim 1, characterized in that, The process of fusing the extracted feature data to obtain meteorological fusion feature data includes: Set spatial reference coordinates, and map the extracted multidimensional correlation feature data to the spatial reference coordinates to obtain spatial feature data; Add a timestamp identifier to each spatial feature point in the spatial feature data; The spatial feature data is divided into time windows according to a preset time granularity, and then classified into the corresponding time windows to obtain spatiotemporal feature data. The spatiotemporal feature data are weighted and fused to obtain meteorological fusion feature data.

4. The ultra-short-term solar radiation prediction method as described in claim 1, characterized in that, The source domain dataset includes historical meteorological data and historical solar radiation data for different radiation regions. The step of using the source domain dataset to train a pre-built LSTM model for radiation prediction to generate a source domain solar radiation predictor includes: Construct an LSTM neural network model that includes an input layer, hidden layers, and an output layer; The LSTM neural network model is trained by taking the historical meteorological data of the source region as input and the historical solar radiation data of the source region as output. When the model reaches the preset accuracy requirement, training stops, and a trained source region solar radiation predictor is obtained.

5. The ultra-short-term solar radiation prediction method as described in claim 1, characterized in that, The target domain dataset includes historical meteorological data and historical solar radiation data for the target domain. Then, the target domain solar radiation predictor, based on ultra-short-term solar radiation prediction and the target domain dataset, is trained using the transfer learning strategy on the source domain solar radiation predictor to generate a target domain solar radiation predictor, including: Using the feature vectors of the aforementioned meteorological fusion feature data as the retrieval benchmark, a search is performed in the meteorological database of the target radiation area to obtain historical meteorological data of the target domain. Obtain the historical solar radiation data of the target domain corresponding to the historical meteorological data of the target domain; Based on the prediction target of ultra-short-term solar radiation, the historical meteorological data of the target domain is used as input and the historical solar radiation data of the target domain is used as output. The transfer learning strategy is used to transfer train the source domain solar radiation predictor to generate the target domain solar radiation predictor.

6. A short-term solar radiation prediction device, characterized in that, include: The data acquisition module is used to collect meteorological observation data of the target radiation area, including satellite cloud image data, radar data and ground meteorological data; The data processing module is used to preprocess and extract features from the meteorological observation data using pre-built preprocessing multi-channel and feature extraction multi-channel, and to fuse the extracted feature data to obtain meteorological fusion feature data. The radiation prediction module is used to perform predictive analysis on the meteorological fusion feature data using a pre-built target domain solar radiation predictor to obtain ultra-short-term solar radiation prediction results. The construction process of the target domain solar radiation predictor includes: Determine the source domain dataset and the target domain dataset, perform correlation analysis on the source domain dataset and the target domain dataset, and determine the transfer learning strategy; The source domain dataset is used to train a pre-built LSTM model for radiation prediction, generating a source domain solar radiation predictor. Based on the prediction target of ultra-short-term solar radiation and the target domain dataset, the source domain solar radiation predictor is transferred and trained using the transfer learning strategy to generate the target domain solar radiation predictor. The step of performing correlation analysis on the source domain dataset and the target domain dataset to determine the transfer learning strategy includes: Calculate the cosine similarity and KL divergence of the core meteorological features between the source domain dataset and the target domain dataset; Based on the combined results of the cosine similarity and the KL divergence, the correlation level between the source domain dataset and the target domain dataset is determined; Based on the correlation level, a corresponding transfer learning strategy is selected from a plurality of preset candidate transfer learning strategies; When the correlation level is high, a transfer learning strategy with fine-tuned parameters is adopted. When the correlation level is medium, a feature transfer strategy is adopted. When the relevance level is low, a domain-adaptive strategy is adopted.

7. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program; The processor executes the computer program to implement the ultra-short-term solar radiation prediction method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed, implements the ultra-short-term solar radiation prediction method as described in any one of claims 1 to 5.

Images