Multi-modal and multi-scale solar radiation prediction method and device
Through the multimodal multi-scale solar radiation prediction method, combining meteorological variables and all-sky image data, the characteristics are extracted and fused, and the problem of inaccurate solar radiation prediction in the existing technology is solved, achieving higher prediction accuracy and reliability .
Patent Information
- Application Number
- CN202411209180.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-30
- Publication Date
- 2024-12-06
AI Technical Summary
The prior art cannot effectively ensure the accuracy of solar radiation prediction. Direct modeling relies on single data and simple feature extraction, indirect modeling is complex and empirically set thresholds are unreliable.
By using multimodal multi-scale solar radiation prediction method, we can obtain meteorological variables and all-sky image data at multiple time points, and use a pre-trained radiation prediction model to extract meteorological variables and all-sky image features, and perform feature fusion to predict the future Target radiation parameters at the time point.
The accuracy and reliability of solar radiation prediction are improved, and the ability of prediction models is enhanced by combining multi-purpose meteorological parameters and all-sky image data to extract spatial and temporal features.
Smart Images

Figure CN119089384A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of radiation information analysis, and in particular to a multi-modal and multi-scale solar radiation prediction method and device. Background Art
[0002] Due to the abundant solar energy resources in the world and the huge potential for photovoltaic power generation, photovoltaic power generation has received more and more attention, and photovoltaic installed capacity is also increasing rapidly. Photovoltaic power generation is closely related to the solar radiation. Therefore, accurate prediction of future solar radiation can provide future power generation information for the operation of the power grid, so as to make the correct operation strategy.
[0003] The current methods for predicting solar radiation information are mainly divided into direct modeling and indirect modeling. The direct modeling method mainly uses a single meteorological variable data or image data as the input of the black box model, so as to obtain the predicted value of solar radiation through the black box model. The indirect modeling method mainly targets image data, applies image recognition technology to sky images, and performs cloud momentum modeling by identifying the cloud distribution in the image, so as to predict the future cloud distribution, and then map the future solar radiation according to the future cloud distribution.
[0004] However, the direct modeling method only uses a single data and the feature extraction method is simple, so the information that can be extracted is limited, which limits the accuracy of the model. For the indirect modeling method, the process is more complicated, and this method needs to rely on experience to set thresholds in the image recognition stage to distinguish between clouds and sky, which has certain unreliability, so neither method can effectively guarantee the accurate prediction of solar radiation. Summary of the invention
[0005] Based on the above-mentioned deficiencies of the prior art, the present application provides a multi-modal and multi-scale solar radiation prediction method and device to solve the problem that the prior art cannot effectively guarantee accurate prediction of solar radiation.
[0006] In order to achieve the above objectives, this application provides the following technical solutions:
[0007] The first aspect of the present application provides a multi-modal and multi-scale solar radiation prediction method, comprising:
[0008] Acquire data of various meteorological variables and full sky images at multiple time points within the current specified time period; wherein the meteorological variables include multiple target meteorological parameters and multiple radiation parameters;
[0009] Inputting a meteorological variable sequence composed of data of each of the meteorological variables at each time point, and a full sky image sequence composed of each of the full sky images, into a pre-trained radiation prediction model;
[0010] Extracting features of the meteorological variable sequence through a meteorological variable feature extraction module in the radiation prediction model to obtain meteorological variable features, and extracting features of the full sky image sequence through a full sky image feature extraction module in the radiation prediction model to obtain full sky image features;
[0011] The meteorological variable features are fused with the all-sky image features through the multimodal feature fusion module in the radiation prediction model to obtain a fused feature vector, and based on the fused feature vector, predicted values of the target radiation parameters at multiple future time points are predicted.
[0012] Optionally, in the above multi-modal and multi-scale solar radiation prediction method, it also includes:
[0013] Obtain data on various meteorological parameters at multiple time points within a preset time period;
[0014] Analyze and remove abnormal values in the data of each meteorological parameter;
[0015] Filling in missing values in the data of each of the meteorological parameters;
[0016] Based on the data of each meteorological parameter, calculating the radiation correlation coefficient of each meteorological parameter;
[0017] According to the order of the radiation correlation coefficient of each meteorological parameter from large to small, multiple meteorological parameters are selected to be determined as the target meteorological parameters.
[0018] Optionally, in the above-mentioned multi-modal and multi-scale solar radiation prediction method, the step of analyzing and removing abnormal values in the data of each of the meteorological parameters includes:
[0019] For each of the meteorological parameters, the distribution of the data of the meteorological parameter is represented in the form of a box plot;
[0020] Determine each data of the meteorological parameter exceeding the upper limit line in the box plot as a preliminary abnormal value;
[0021] Based on the data of the adjacent moments of each of the preliminary abnormal values, each of the preliminary abnormal values is divided into a first abnormal value and a second abnormal value; wherein the first abnormal value is an abnormal value caused by a collection problem; and the second abnormal value is an abnormal value caused by extreme weather;
[0022] Each of the first outliers is determined as a final outlier in the data of the meteorological parameter and is removed.
[0023] Optionally, in the above-mentioned multi-modal multi-scale solar radiation prediction method, before inputting the meteorological variable sequence composed of the data of each of the meteorological variables at each time point and the full sky image sequence composed of each of the full sky images into the pre-trained radiation prediction model, it also includes:
[0024] Cleaning each of the full sky images and aligning them with the time tags of the data of each of the meteorological variables;
[0025] According to a preset cropping radius, each of the full sky images after time tag alignment is cropped using a binary mask method; wherein pixels within the preset cropping radius in the cropped full sky image are valid pixels, and pixels outside the preset cropping radius are invalid pixels;
[0026] The resolution of each of the cropped all-sky images is reduced and normalized.
[0027] Optionally, in the above-mentioned multi-modal and multi-scale solar radiation prediction method, extracting features of the meteorological variable sequence by a meteorological variable feature extraction module in the radiation prediction model to obtain meteorological variable features includes:
[0028] The meteorological variable sequence is sequentially passed through each convolution network in the meteorological variable feature extraction module to extract spatial features; wherein one of the convolution networks is composed of a one-dimensional convolution layer, a maximum pooling layer and a Dropout layer connected in sequence;
[0029] The feature vector output by the last convolutional network is input into the bidirectional long short-term memory network for temporal feature extraction;
[0030] Inputting the feature vector output by the bidirectional long short-term memory network into the attention mechanism layer for feature weight allocation;
[0031] The feature vector output by the attention mechanism layer is input into the flattening layer for flattening to obtain the final feature of the meteorological variable.
[0032] Optionally, in the above-mentioned multi-modal multi-scale solar radiation prediction method, extracting features from the full sky image sequence by the full sky image feature extraction module in the radiation prediction model to obtain full sky image features includes:
[0033] Inputting each of the full sky images in the full sky image sequence into a branch network of the full sky image feature extraction module; wherein the structures of the branch networks are consistent;
[0034] The input full sky image is subjected to spatial feature extraction through the single-scale two-dimensional convolution module in the branch network; wherein the single-scale two-dimensional convolution module is composed of a two-dimensional convolution layer, a batch normalization layer, an activation function layer, a maximum pooling layer, and a Dropout layer connected in sequence;
[0035] The feature vector output by the single-scale two-dimensional convolution module is input into the multi-scale two-dimensional convolution module in the branch network for secondary spatial feature extraction; wherein the multi-scale two-dimensional convolution module is composed of a two-dimensional convolution layer, a batch normalization layer, an activation function layer, a spatial pyramid pooling layer, and a Dropout layer connected in sequence;
[0036] The feature vector output by the multi-scale two-dimensional convolution module is input into the bidirectional long short-term memory network in the branch network to extract time features, so as to obtain the full sky image features corresponding to the input full sky image.
[0037] Optionally, in the above-mentioned multi-modal and multi-scale solar radiation prediction method, the step of predicting the predicted values of the target radiation parameters at multiple future time points based on the fused feature vector includes:
[0038] The fused feature vector is processed by a fully connected layer in the multimodal feature fusion module to map predicted values of clear sky factors at multiple future time points;
[0039] According to the conversion relationship between the clear sky factor and the horizontal total radiation, the predicted values of the clear sky factor at multiple future time points are used to calculate the predicted values of the horizontal total radiation at multiple future time points.
[0040] The second aspect of the present application provides a multi-modal and multi-scale solar radiation prediction device, comprising:
[0041] A data acquisition unit, used to acquire data of various meteorological variables and a full sky image at multiple time points within a currently specified time period; wherein the meteorological variables include multiple target meteorological parameters and multiple radiation parameters;
[0042] A prediction unit is used to input a meteorological variable sequence composed of data of each of the meteorological variables at each time point and a full-sky image sequence composed of each of the full-sky images into a pre-trained radiation prediction model, perform feature extraction on the meteorological variable sequence through a meteorological variable feature extraction module in the radiation prediction model to obtain meteorological variable features, and perform feature extraction on the full-sky image sequence through a full-sky image feature extraction module in the radiation prediction model to obtain full-sky image features, then fuse the meteorological variable features with the full-sky image features through a multimodal feature fusion module in the radiation prediction model to obtain a fused feature vector, and predict the predicted values of the target radiation parameters at multiple future time points based on the fused feature vector.
[0043] Optionally, in the above-mentioned multi-modal and multi-scale solar radiation prediction device, it also includes:
[0044] A data acquisition unit, used to obtain data of various meteorological parameters at multiple time points within a preset time period;
[0045] A filtering unit, used to analyze and remove abnormal values in the data of each of the meteorological parameters;
[0046] A filling unit, used to fill in missing values in the data of each of the meteorological parameters;
[0047] A coefficient calculation unit, used for calculating the radiation correlation coefficient of each meteorological parameter based on the data of each meteorological parameter;
[0048] The selection unit is used to select multiple meteorological parameters as the target meteorological parameters in descending order of the radiation correlation coefficients of the meteorological parameters.
[0049] Optionally, in the above-mentioned multi-modal and multi-scale solar radiation prediction device, the filtering unit comprises:
[0050] A representation unit, for representing the distribution of the data of each meteorological parameter in the form of a box plot;
[0051] A first determining unit is used to determine each data of the meteorological parameter exceeding the upper limit line in the box plot as a preliminary abnormal value;
[0052] A division unit, configured to divide each of the preliminary abnormal values into a first abnormal value and a second abnormal value based on data of the adjacent moments of each of the preliminary abnormal values; wherein the first abnormal value is an abnormal value caused by a collection problem; and the second abnormal value is an abnormal value caused by extreme weather;
[0053] The second determination unit is used to determine each of the first abnormal values as a final abnormal value in the data of the meteorological parameter and remove it.
[0054] Optionally, in the above-mentioned multi-modal and multi-scale solar radiation prediction device, it also includes:
[0055] An alignment unit, used for cleaning each of the full sky images and aligning them with the time tags of the data of each of the meteorological variables;
[0056] A cropping unit, used to crop the full sky images after time tag alignment according to a preset cropping radius using a binary mask method; wherein the pixels within the preset cropping radius in the cropped full sky image are valid pixels, and the pixels outside the preset cropping radius are invalid pixels;
[0057] The normalization processing unit is used to reduce the resolution of each of the cropped full-sky images and perform normalization processing.
[0058] Optionally, in the above-mentioned multi-modal multi-scale solar radiation prediction device, when the prediction unit performs the feature extraction of the meteorological variable sequence through the meteorological variable feature extraction module in the radiation prediction model to obtain the meteorological variable features, it is used to:
[0059] The meteorological variable sequence is sequentially passed through each convolution network in the meteorological variable feature extraction module to extract spatial features; wherein one of the convolution networks is composed of a one-dimensional convolution layer, a maximum pooling layer and a Dropout layer connected in sequence;
[0060] The feature vector output by the last convolutional network is input into the bidirectional long short-term memory network for temporal feature extraction;
[0061] Inputting the feature vector output by the bidirectional long short-term memory network into the attention mechanism layer for feature weight allocation;
[0062] The feature vector output by the attention mechanism layer is input into the flattening layer for flattening to obtain the final feature of the meteorological variable.
[0063] Optionally, in the above-mentioned multi-modal multi-scale solar radiation prediction device, when the prediction unit performs the feature extraction of the full sky image sequence through the full sky image feature extraction module in the radiation prediction model to obtain the full sky image features, it is used to:
[0064] Inputting each of the full sky images in the full sky image sequence into a branch network of the full sky image feature extraction module; wherein the structures of the branch networks are consistent;
[0065] The input full sky image is subjected to spatial feature extraction through the single-scale two-dimensional convolution module in the branch network; wherein the single-scale two-dimensional convolution module is composed of a two-dimensional convolution layer, a batch normalization layer, an activation function layer, a maximum pooling layer, and a Dropout layer connected in sequence;
[0066] The feature vector output by the single-scale two-dimensional convolution module is input into the multi-scale two-dimensional convolution module in the branch network for secondary spatial feature extraction; wherein the multi-scale two-dimensional convolution module is composed of a two-dimensional convolution layer, a batch normalization layer, an activation function layer, a spatial pyramid pooling layer, and a Dropout layer connected in sequence;
[0067] The feature vector output by the multi-scale two-dimensional convolution module is input into the bidirectional long short-term memory network in the branch network to extract time features, so as to obtain the full sky image features corresponding to the input full sky image.
[0068] Optionally, in the above-mentioned multi-modal multi-scale solar radiation prediction device, when the prediction unit performs the step of predicting the target radiation parameter at multiple future time points based on the fused feature vector, it is used to:
[0069] The fused feature vector is processed by a fully connected layer in the multimodal feature fusion module to map predicted values of clear sky factors at multiple future time points;
[0070] According to the conversion relationship between the clear sky factor and the horizontal total radiation, the predicted values of the clear sky factor at multiple future time points are used to calculate the predicted values of the horizontal total radiation at multiple future time points.
[0071] The present application provides a multimodal and multiscale solar radiation prediction method, which obtains data of various meteorological variables and full-sky images at multiple time points within the current specified time period. Among them, meteorological variables include data of multiple target meteorological parameters and multiple radiation parameters, so in the embodiment of the present application, multiple target meteorological parameters are combined with full-sky images to ensure that there is sufficient radiation-related data. In addition, multiple radiation parameters are introduced to further increase the data used for prediction, thereby improving the accuracy of the prediction. The meteorological variable sequence composed of the data of various meteorological variables at each time point and the full-sky image sequence composed of each full-sky image are input into a pre-trained radiation prediction model, and then input into the model in a sequence manner to extract spatial and temporal features to ensure the accuracy of the prediction. Therefore, the meteorological variable sequence is feature extracted by the meteorological variable feature extraction module in the radiation prediction model to obtain meteorological variable features, and the full-sky image sequence is feature extracted by the full-sky image feature extraction module in the radiation prediction model to obtain full-sky image features. Finally, the meteorological variable features are fused with the full-sky image features through the multimodal feature fusion module in the radiation prediction model to obtain the fused feature vector. Based on the fused feature vector, the predicted values of the target radiation parameters at multiple future time points are predicted, thus realizing a multimodal and multi-scale solar radiation prediction method, which effectively ensures the accuracy of the prediction results. BRIEF DESCRIPTION OF THE DRAWINGS
[0072] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.
[0073] Figure 1 A flowchart of a multi-modal and multi-scale solar radiation prediction method provided in an embodiment of the present application;
[0074] Figure 2 A flow chart of a method for determining target meteorological parameters provided in an embodiment of the present application;
[0075] Figure 3 A flowchart of a method for analyzing and removing abnormal values in data of various meteorological parameters provided in an embodiment of the present application;
[0076] Figure 4 A flowchart of a method for preprocessing a full sky image provided in an embodiment of the present application;
[0077] Figure 5 A schematic diagram of the architecture of a radiation prediction model provided in an embodiment of the present application;
[0078] Figure 6 A flowchart of a method for extracting features from a meteorological variable sequence through a meteorological variable feature extraction module in a radiation prediction model provided in an embodiment of the present application;
[0079] Figure 7 A flowchart of a method for extracting features from a full sky image sequence by using a full sky image feature extraction module in a radiation prediction model provided in an embodiment of the present application;
[0080] Figure 8 A schematic diagram of the architecture of a spatial pyramid pooling layer provided in an embodiment of the present application;
[0081] Fig. 9 A result curve diagram of a one-step-ahead prediction example provided in an embodiment of the present application;
[0082] Fig.10 A result curve diagram of a prediction example 3 steps ahead provided in an embodiment of the present application;
[0083] Fig.11 A result curve diagram of a 5-step-ahead prediction example provided in an embodiment of the present application;
[0084] Fig.12 A schematic diagram of the architecture of a multi-modal and multi-scale solar radiation prediction device provided in an embodiment of the present application. DETAILED DESCRIPTION
[0085] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0086] In this application, relational terms such as first and second, etc. are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the statement "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element.
[0087] The present application embodiment provides a multi-modal and multi-scale solar radiation prediction method, such as Figure 1 As shown, the following steps are included:
[0088] S101, obtaining data of various meteorological variables and full sky images at multiple time points within a currently specified time period.
[0089] Among them, meteorological variables include multiple target meteorological parameters and multiple radiation parameters. The target meteorological parameters are designated meteorological parameters with a greater correlation with radiation. Meteorological parameters refer to meteorological related parameters collected by meteorological collection instruments and displayed in numerical form, such as ambient temperature, humidity, wind direction, etc. Optionally, in the embodiment of the present application, the target meteorological parameters may specifically include horizontal total radiation, temperature and humidity.
[0090] Radiation parameters are parameters directly related to solar radiation, such as solar zenith angle, hour angle, declination, etc. Therefore, it can be seen that in the embodiment of the present application, not only meteorological parameters and image data are predicted, but also multiple radiation parameters are introduced as auxiliary variables, so that more features can be extracted from more information, thereby effectively ensuring the accuracy of the prediction results.
[0091] The total sky image is an image of the sky collected by a total sky imager. The instrument uses a fisheye lens with a 180-degree viewing angle to take pictures of the sky from the ground. It can obtain sky status information for a larger area above the collection point, thereby obtaining characteristic information of the sky over a larger range.
[0092] Optionally, specific target meteorological parameters and full sky images can be directly collected by corresponding equipment at regular intervals, while radiation parameters need to be collected until the parameters required for calculation are calculated and then calculated.
[0093] Optionally, in the embodiment of the present application, the various radiation parameters introduced specifically include the solar zenith angle, clear sky solar radiation and clear sky factor. Among them, the solar zenith angle is the angle between a beam of light from the sun to a point on the ground and the straight line perpendicular to the ground at this point. The clear sky solar radiation is the total solar radiation value obtained by the outer space of the earth. The clear sky factor is the total solar radiation incident on the horizontal plane of the earth, that is, the horizontal total solar radiation, and the ratio of the clear sky solar radiation, and the value is between 0 and 1.
[0094] Specifically, the solar zenith angle can be calculated by the following method:
[0095] cosθ z =cosφ*cosδ*cosω+sinφ*sinδ
[0096]
[0097] Among them, θ z is the solar zenith angle; φ is the latitude, ranging from -90 degrees to 90 degrees; δ is the declination; ω is the hour angle, one hour is 15 degrees, negative in the morning and positive in the afternoon; n is the day number of the day in the year.
[0098] Clear sky solar radiation can be calculated as follows:
[0099]
[0100] Among them, I O is the clear sky solar radiation; G SC is the solar constant, which is 1367W / m 2 ;ω 2 is the hour angle at the end time; ω 1 is the hour angle of the starting time.
[0101] The clear sky factor can be expressed as:
[0102]
[0103] Among them, k T is the clear sky factor; I is the total horizontal solar radiation received by the strategy at the current moment.
[0104] Optionally, in another embodiment of the present application, a method for determining a target meteorological parameter is provided, such as Figure 2 As shown, the following steps are included:
[0105] S201, obtaining data of various meteorological parameters at multiple time points within a preset time period.
[0106] Optionally, the data of all meteorological parameters that may affect solar radiation may be collected at intervals within a preset time period, so as to obtain the data of various meteorological parameters at multiple time points within the preset time period.
[0107] S202: Analyze and remove abnormal values in the data of various meteorological parameters.
[0108] It should be noted that, considering the possibility of failure of the data collection equipment or improper operation of the equipment, abnormal values that cannot truly reflect the meteorological conditions of the environment are collected, thereby affecting the accuracy of the determination of the target meteorological parameters and further affecting the accuracy of subsequent predictions. Therefore, it is necessary to first analyze these abnormal data and eliminate them.
[0109] Optionally, in another embodiment of the present application, a specific implementation of step S202 is as follows: Figure 3 As shown, the following steps are included:
[0110] S301. For each meteorological parameter, the distribution of the meteorological parameter data is represented in the form of a box plot.
[0111] In order to easily identify abnormal data that exceeds other data, in the embodiment of the present application, the distribution of the meteorological parameter data is represented in the form of a box plot.
[0112] S302, determining each data of the meteorological parameters exceeding the upper limit line in the box plot as preliminary abnormal values.
[0113] It should be noted that the abnormal data may be caused by extreme weather, improper operation, or instrument failure. Although the abnormal data caused by extreme weather appears abnormal relative to other data, it is a parameter that truly reflects the environmental meteorology and cannot be eliminated. Therefore, we can only first determine the data of the meteorological parameters that exceed the upper limit as preliminary abnormal values, and then further execute step S303.
[0114] S303 . Based on the data of the adjacent moments of the preliminary outliers, divide the preliminary outliers into first outliers and second outliers.
[0115] The first abnormal value is an abnormal value caused by a collection problem. The second abnormal value is an abnormal value caused by extreme weather. Collection problems can be problems that occur during the collection process, such as improper operation or instrument failure.
[0116] Optionally, although the abnormal data caused by extreme weather has a large deviation from most of the data, its deviation from the data at the nearby time is relatively small. The abnormal value caused by the acquisition problem is sudden, so it has a large deviation from the data at the nearby time. Therefore, the first abnormal value and the second abnormal value can be specifically divided based on the change rate of the preliminary abnormal value and the data at the nearby time.
[0117] S304: Determine each first abnormal value as a final abnormal value in the meteorological parameter data and remove it.
[0118] S203, filling in the missing values in the data of various meteorological parameters.
[0119] After an outlier is removed, the data of the meteorological parameter at that time point will be missing, and there may also be missing data when collecting data, so it is necessary to fill in the missing values in the data of each meteorological parameter.
[0120] Optionally, missing values may be filled using linear interpolation.
[0121] S204. Calculate the radiation correlation coefficient of each meteorological parameter based on the data of each meteorological parameter.
[0122] In order to screen out the meteorological parameters that have a greater impact on radiation, it is necessary to calculate the correlation between each meteorological parameter and solar radiation. Optionally, the radiation correlation coefficient of each meteorological parameter may specifically be a Pearson correlation coefficient.
[0123] S205. Select multiple meteorological parameters as target meteorological parameters in descending order of the radiation correlation coefficients of the meteorological parameters.
[0124] Specifically, N meteorological parameters with the largest radiation correlation coefficients are selected as target meteorological parameters as needed.
[0125] S102, inputting a meteorological variable sequence composed of data of various meteorological variables at various time points, and a full sky image sequence composed of various full sky images, into a pre-trained radiation prediction model.
[0126] In order to not only extract spatial features directly from the data, but also extract temporal features, so as to more accurately predict future solar radiation information, in the embodiment of the present application, the data of various meteorological variables at various time points and the full sky images at various time points are respectively composed of meteorological variable sequences and full sky image sequences according to the collected time points. Among them, the data of each point in the meteorological variable sequence includes the data of various meteorological variables collected at a time point, for example, clear sky factor, clear sky radiation, horizontal total radiation, temperature, humidity and solar zenith angle at a certain time.
[0127] It should be noted that since the meteorological variable sequence and the full sky image sequence are two different types of data, in order to extract more accurate features, in the embodiment of the present application, a meteorological variable feature extraction module and a full sky image feature extraction module are respectively designed in the radiation prediction model to extract features of the two types of data respectively, so as to be able to perform feature extraction in a targeted manner. Therefore, when inputting data into the radiation prediction model, the meteorological variable sequence is specifically input into the meteorological variable feature extraction module of the radiation prediction model, and the full sky image sequence is input into the full sky image feature extraction module of the radiation prediction model.
[0128] Optionally, in order to better extract features from the full sky image, in another embodiment of the present application, before executing step S102, each full sky image may be further preprocessed. Figure 4 As shown, a method for preprocessing a full sky image provided by an embodiment of the present application includes:
[0129] S401, cleaning each full sky image and aligning the image with the time tags of the data of each meteorological variable.
[0130] The time label of the data of each meteorological variable may be its collection time. Therefore, the full sky image with the same collection time is aligned with the data of the meteorological variable, so that the solar radiation can be predicted based on the full sky image and the data of the meteorological variable at the same time point.
[0131] S402 , cropping each of the full sky images after the time tag alignment using a binary mask method according to a preset cropping radius.
[0132] Among them, the pixels within the preset cropping radius in the cropped full sky image are valid pixels, and the pixels outside the preset cropping radius are invalid pixels.
[0133] It should be noted that in order to remove the occlusions and other interfering information at the edge of the image, the binary mask method is used to crop each full sky image. Specifically, according to the preset cropping radius, a circle is formed with the center of the image as the circle. Among them, the image within the circular range is the required image, that is, the valid image, which can be retained. For the image outside the circular range, it needs to be cropped, so it can be processed as black.
[0134] S403: reducing the resolution of each cropped full-sky image and performing normalization processing.
[0135] In order to reduce redundant information in the image and reduce the training burden of the model, in this application, the resolution of the full sky image is also reduced and normalized.
[0136] S103, extracting features of the meteorological variable sequence through the meteorological variable feature extraction module in the radiation prediction model to obtain meteorological variable features, and extracting features of the full sky image sequence through the full sky image feature extraction module in the radiation prediction model to obtain full sky image features.
[0137] Optionally, the meteorological variable feature extraction module can be used to extract spatial and temporal features of the entire meteorological variable sequence to obtain meteorological variable features. Similarly, the full sky image feature extraction module in the radiation prediction model can be used to extract spatial and temporal features of each full sky image in the full sky image sequence to obtain full sky image features, so that more comprehensive features can be obtained to predict solar radiation and ensure the accuracy of the prediction results.
[0138] Optionally, in order to extract the temporal and spatial features of the meteorological variable sequence and focus more on important features, in the embodiment of the present application, Figure 5As shown, the meteorological variable feature extraction module mainly uses the "CNN-BiLSTM-Attention" structure for feature extraction, that is, the meteorological variable feature extraction module is mainly composed of a convolutional network, a bidirectional long short-term memory network, and an attention mechanism layer. Accordingly, in the embodiment of the present application, a specific implementation method of extracting features from a meteorological variable sequence by using the meteorological variable feature extraction module in the radiation prediction model is as follows: Figure 6 As shown, the following steps are included:
[0139] S601, extracting spatial features from the meteorological variable sequence through each convolutional network in the meteorological variable feature extraction module.
[0140] Among them, Figure 5 As shown, a convolutional network is composed of a one-dimensional convolutional layer, a maximum pooling layer, and a Dropout layer connected in sequence. In an embodiment of the present application, convolution is performed in sequence through multiple convolutional networks, so that feature extraction omissions can be avoided as much as possible. Optionally, the meteorological variable feature extraction module can specifically include two convolutional networks.
[0141] Specifically, the meteorological variable sequence is first convolved through a 1D convolutional layer to extract various features.
[0142] Among them, the specific calculation process of the 1D convolutional layer is as follows:
[0143]
[0144] in, is the i-th output of the k-th layer; is the jth output of the k-1th layer; is the convolution kernel weight matrix corresponding to the kth layer; is the bias coefficient of the kth layer; f is the activation function; n is the number of neurons in this layer.
[0145] After extracting features through the one-dimensional convolution layer, the extracted features are then filtered through the maximum pooling layer to retain valid features. The calculation process of the maximum pooling layer is as follows:
[0146]
[0147] Among them, P i k (j) is the value of the corresponding position passed to the kth layer by the maximum pooling process; is the activation value of the lth neuron in the ith feature vector of the previous layer; W is the width of the maximum pooling.
[0148] A Dropout layer is connected after the maximum pooling layer to discard neurons with a certain probability to prevent the model from overfitting.
[0149] S602: Input the feature vector output by the last convolutional network into a bidirectional long short-term memory network for temporal feature extraction.
[0150] Specifically, the bidirectional long short-term memory network BiLSTM updates, protects or deletes the cell state information through the forget gate, input gate and output gate. The calculation process of these three pieces of information is as follows:
[0151]
[0152] Among them, h t-1 is the hidden state of the previous moment; x t is the input vector at that moment; ω f ,ω i and ω o are the weight matrices corresponding to the three gates; b f 、b i and b o are the bias coefficients corresponding to the three gates respectively; σ is the sigmoid activation function, which generates a value between 0 and 1.
[0153] LSTM internal cell state vector at time t The calculation process is as follows:
[0154]
[0155] Among them, ω c and b c is the weight matrix and bias coefficient; tanh is the activation function used to update the internal cell state.
[0156] The final cell state at time t is c t The forget gate interacts with the final cell state at the previous moment to determine the information that needs to be forgotten. The input gate interacts with the internal cell state at the current moment to determine the information that needs to be retained. The specific calculation process is shown in the following formula:
[0157]
[0158] Among them, the hidden state h at the current moment t Calculated by the following formula:
[0159] h t =o t *tanh(c t )
[0160] BiLSTM consists of a forward LSTM and a backward LSTM. The forward LSTM calculates the output in chronological order, and the backward LSTM calculates the output in reverse chronological order. The hidden states of the LSTMs in both directions are concatenated into the output of the BiLSTM. Through the above-mentioned calculation method of bidirectional information flow, BiLSTM can better capture the temporal dependency between the previous and next steps.
[0161] S603: Input the feature vector output by the bidirectional long short-term memory network into the attention mechanism layer for feature weight allocation.
[0162] It should be noted that the attention mechanism obtains the importance of different features based on the information of the input features, and assigns greater weights to features with greater importance, so that it can focus on features that are more useful for prediction. Therefore, after all the features are extracted, the feature vector is input into the attention mechanism layer for feature weight assignment, thereby improving the accuracy of the prediction results.
[0163] Specifically, the calculation process of the attention mechanism layer is as follows:
[0164] e i =tanh(ω h *h i +b h )
[0165]
[0166] h i ′=α i *h i
[0167] Among them, e i is the importance of the input feature, h i is the output vector of the BiLSTM hidden layer, ω h and ω h are the corresponding weight matrices and bias coefficients, α i is the feature weight coefficient, h i ′ is the feature after the attention mechanism pays attention to it.
[0168] Therefore, it can be seen that in the embodiment of the present application, the spatial characteristics of the meteorological variable data are first extracted through the convolutional network, and then the temporal characteristics of the meteorological variable data are extracted through the BiLSTM layer, and finally the attention mechanism is used to focus on the features that are more favorable to the prediction. In this way, not only comprehensive and sufficient features can be extracted, but also more attention can be paid to more important features, thereby effectively improving the accuracy of the prediction.
[0169] S604: Input the feature vector output by the attention mechanism layer into the flattening layer for flattening to obtain the final meteorological variable feature.
[0170] It should be noted that since the extracted feature vector needs to be input into the subsequent multimodal feature fusion module for fusion, in order to process the extracted feature vector into a format that conforms to the input of the multimodal feature fusion module, the feature vector is finally input into the flattening layer for flattening to obtain the final meteorological variable feature.
[0171] Optionally, since an image in this application is equivalent to including a large amount of data, that is, a collection of data, feature extraction can be performed separately. However, it is also necessary to ensure the sequence relationship between images, so in another embodiment of this application, such as Figure 5 As shown in the figure, the full sky image feature extraction module adopts a time-distributed architecture, that is, a parallel network is used to process the full sky image sequence. Specifically, the image of each time step is used as the input of a branch network in the parallel network, so that feature extraction is performed in the corresponding branch network. Each branch network consists of a single-scale two-dimensional convolution module, a multi-scale two-dimensional convolution module, and a bidirectional long short-term memory network.
[0172] Therefore, in the embodiment of the present application, an implementation method of extracting features from the full sky image sequence by using the full sky image feature extraction module in the radiation prediction model is as follows: Figure 7 As shown, the following steps are included:
[0173] S701, input each full sky image in the full sky image sequence into a branch network of the full sky image feature extraction module.
[0174] Among them, the structures of each branch network are consistent.
[0175] S702: extract spatial features of the input full sky image through the single-scale two-dimensional convolution module in the branch network.
[0176] Among them, the single-scale two-dimensional convolution module is composed of a two-dimensional convolution layer, a batch normalization layer, an activation function layer, a maximum pooling layer, and a Dropout layer connected in sequence.
[0177] It should be noted that, since the image is feature extracted, a two-dimensional convolution layer is used in the embodiment of the present application. Specifically, the two-dimensional convolution layer is used to extract image spatial features, the batch normalization layer and the Dropout layer are used to prevent the model from overfitting, the activation function layer is used to enhance the expression ability of the network, and the maximum pooling layer is used to filter features.
[0178] S703: Input the feature vector output by the single-scale two-dimensional convolution module into the multi-scale two-dimensional convolution module in the branch network to perform secondary space feature extraction.
[0179] In order to ensure that sufficient effective features can be extracted to improve the accuracy of prediction, in an embodiment of the present application, after feature extraction is performed through a single-scale two-dimensional convolution module, it is further performed through a multi-scale two-dimensional convolution module.
[0180] Among them, the multi-scale two-dimensional convolution module is composed of a two-dimensional convolution layer, a batch normalization layer, an activation function layer, a spatial pyramid pooling layer, and a Dropout layer connected in sequence. It can be seen that the main difference between the multi-scale two-dimensional convolution module and the single-scale two-dimensional convolution module is that the maximum pooling layer is replaced by the spatial pyramid pooling layer. By introducing the spatial pyramid pooling layer, maximum pooling can be performed from multiple scales to capture features of different scales. Therefore, the specific processing process is almost the same as that of the single-scale two-dimensional convolution module, and the only difference is the processing process of the spatial pyramid pooling layer.
[0181] Specifically, Figure 8 As shown in the figure, the spatial pyramid pooling layer divides the feature map into different numbers of grids, representing different scales, performs maximum pooling on these grids, and finally flattens and concatenates the pooling results of each grid.
[0182] S704: Input the feature vector output by the multi-scale two-dimensional convolution module into the bidirectional long short-term memory network in the branch network to extract temporal features, so as to obtain the full sky image features corresponding to the input full sky image.
[0183] It should be noted that, in order to extract the temporal features of the full sky image sequence, the temporal features will be further extracted through the bidirectional long short-term memory network. Figure 5 As shown in FIG. 1 , the feature vector output by the multi-scale two-dimensional convolution module is input into a flattening layer for flattening, and then input into a bidirectional long short-term memory network for temporal feature extraction. The extraction process is consistent with the extraction process in the meteorological variable feature extraction model, and will not be repeated here.
[0184] S104. The meteorological variable features are fused with the full sky image features through the multimodal feature fusion module in the radiation prediction model to obtain a fused feature vector, and based on the fused feature vector, the predicted values of the target radiation parameters at multiple future time points are predicted.
[0185] Since the meteorological variable features and the all-sky image features are extracted separately, and both features are needed in the prediction to obtain accurate prediction results, the meteorological variable features and a large number of features in the all-sky image features are spliced and fused through the multimodal feature fusion module to obtain a fused feature vector. Finally, the predicted values of the target radiation parameters at multiple time points in the future can be obtained based on the fused feature vector.
[0186] The predicted value of the target radiation parameter is a specified parameter that can reflect the radiation situation.
[0187] Optionally, in another embodiment of the present application, a specific implementation method of predicting predicted values of target radiation parameters at multiple future time points based on the fused feature vector includes:
[0188] The fused feature vector is processed by the fully connected layer in the multimodal feature fusion module to map the predicted values of the clear sky factor at multiple future time points. Based on the conversion relationship between the clear sky factor and the horizontal total radiation, the predicted values of the horizontal total radiation at multiple future time points are calculated using the predicted values of the clear sky factor at multiple future time points.
[0189] That is to say, in the embodiment of the present application, the target radiation parameter is the horizontal total radiation. Based on the fused feature vector, it is mapped to the predicted value of the clear sky factor, and then further based on the conversion relationship between the clear sky factor and the horizontal total radiation, the predicted value of the horizontal total radiation is calculated, so as to better meet the current prediction process and demand for solar radiation information.
[0190] Therefore, through the multi-modal multi-scale model provided in the embodiment of the present application, the predicted value of horizontal total radiation can be accurately predicted. Fig. 9 , Fig.10 , Fig.11 As shown in the figure, it is divided into 1 step ahead, 3 steps ahead and 5 steps ahead, that is, the prediction results of the multimodal multiscale model at the 1st, 3rd and 5th time points in the future, the prediction results of the existing image unimodal model, the prediction results of the text unimodal model and the true value are compared. From the figure, it can be seen that the overall prediction accuracy of the multimodal multiscale model at each prediction step is better than that of the meteorological variable unimodal model, that is, better than the text unimodal model and the image unimodal model, which can achieve more effective and accurate predictions and is closer to the true value.
[0191] A multimodal and multiscale solar radiation prediction method provided in an embodiment of the present application obtains data of various meteorological variables and full-sky images at multiple time points within the current specified time period. Among them, meteorological variables include data of multiple target meteorological parameters and multiple radiation parameters, so in an embodiment of the present application, multiple target meteorological parameters are combined with full-sky images to ensure that there is sufficient radiation-related data. In addition, multiple radiation parameters are introduced to further increase the data used for prediction, thereby improving the accuracy of the prediction. The meteorological variable sequence composed of the data of various meteorological variables at each time point and the full-sky image sequence composed of each full-sky image are input into a pre-trained radiation prediction model, and then input into the model in a sequence manner to extract spatial and temporal features to ensure the accuracy of the prediction. Therefore, the meteorological variable sequence is feature extracted by the meteorological variable feature extraction module in the radiation prediction model to obtain meteorological variable features, and the full-sky image sequence is feature extracted by the full-sky image feature extraction module in the radiation prediction model to obtain full-sky image features. Finally, the meteorological variable features are fused with the full-sky image features through the multimodal feature fusion module in the radiation prediction model to obtain the fused feature vector. Based on the fused feature vector, the predicted values of the target radiation parameters at multiple future time points are predicted, thus realizing a multimodal and multi-scale solar radiation prediction method, which effectively ensures the accuracy of the prediction results.
[0192] Another embodiment of the present application provides a multi-modal and multi-scale solar radiation prediction device, such as Fig.12 As shown, including:
[0193] The data acquisition unit 1201 is used to acquire various meteorological variables and full sky images at multiple time points within the current specified time period, wherein the meteorological variables include data of multiple target meteorological parameters and multiple radiation parameters.
[0194] The prediction unit 1202 is used to input the meteorological variable sequence composed of the data of various meteorological variables at various time points and the all-sky image sequence composed of various all-sky images into a pre-trained radiation prediction model, perform feature extraction on the meteorological variable sequence through the meteorological variable feature extraction module in the radiation prediction model to obtain meteorological variable features, and perform feature extraction on the all-sky image sequence through the all-sky image feature extraction module in the radiation prediction model to obtain all-sky image features, then fuse the meteorological variable features with the all-sky image features through the multimodal feature fusion module in the radiation prediction model to obtain a fused feature vector, and based on the fused feature vector, predict the predicted values of the target radiation parameters at multiple future time points.
[0195] Optionally, in the multi-modal multi-scale solar radiation prediction device provided in another embodiment of the present application, it also includes:
[0196] The data acquisition unit is used to obtain data of various meteorological parameters at multiple time points within a preset time period.
[0197] The filtering unit is used to analyze and remove abnormal values in the data of various meteorological parameters.
[0198] Filling unit is used to fill the missing values in the data of various meteorological parameters.
[0199] The coefficient calculation unit is used to calculate the radiation correlation coefficient of each meteorological parameter based on the data of each meteorological parameter.
[0200] The selection unit is used to select multiple meteorological parameters as target meteorological parameters in the order of radiation correlation coefficients of various meteorological parameters from large to small.
[0201] Optionally, in the multi-modal multi-scale solar radiation prediction device provided in another embodiment of the present application, the filtering unit includes:
[0202] The representation unit is used to represent the distribution of the data of the meteorological parameter in the form of a box plot for each meteorological parameter.
[0203] The first determination unit is used to determine each data of the meteorological parameter exceeding the upper limit line in the box plot as a preliminary abnormal value.
[0204] The division unit is used to divide each preliminary abnormal value into a first abnormal value and a second abnormal value based on the data of the adjacent time of each preliminary abnormal value, wherein the first abnormal value is an abnormal value caused by a collection problem, and the second abnormal value is an abnormal value caused by extreme weather.
[0205] The second determination unit is used to determine each first abnormal value as a final abnormal value in the data of the meteorological parameter and remove it.
[0206] Optionally, in the multi-modal multi-scale solar radiation prediction device provided in another embodiment of the present application, it also includes:
[0207] The alignment unit is used to clean each full sky image and align it with the time tags of the data of each meteorological variable.
[0208] The cropping unit is used to crop the full sky images after time tag alignment according to a preset cropping radius using a binary mask method, wherein the pixels within the preset cropping radius in the cropped full sky image are valid pixels, and the pixels outside the preset cropping radius are invalid pixels.
[0209] The normalization processing unit is used to reduce the resolution of each cropped full-sky image and perform normalization processing.
[0210] Optionally, in a multi-modal multi-scale solar radiation prediction device provided in another embodiment of the present application, the prediction unit performs feature extraction on a meteorological variable sequence through a meteorological variable feature extraction module in a radiation prediction model to obtain meteorological variable features, which is used to:
[0211] The meteorological variable sequence is sequentially passed through each convolutional network in the meteorological variable feature extraction module for spatial feature extraction. Among them, a convolutional network is composed of a one-dimensional convolution layer, a maximum pooling layer and a Dropout layer connected in sequence.
[0212] The feature vector output by the last convolutional network is input into the bidirectional long short-term memory network for temporal feature extraction.
[0213] The feature vector output by the bidirectional long short-term memory network is input into the attention mechanism layer for feature weight assignment.
[0214] The feature vector output by the attention mechanism layer is input into the flattening layer for flattening to obtain the final meteorological variable features.
[0215] Optionally, in a multimodal multiscale solar radiation prediction device provided in another embodiment of the present application, when the prediction unit performs feature extraction on the full sky image sequence through the full sky image feature extraction module in the radiation prediction model to obtain the full sky image features, it is used to:
[0216] Each full sky image in the full sky image sequence is input into a branch network of the full sky image feature extraction module, wherein the structures of each branch network are consistent.
[0217] The spatial features of the input full sky image are extracted through the single-scale two-dimensional convolution module in the branch network. The single-scale two-dimensional convolution module is composed of a two-dimensional convolution layer, a batch normalization layer, an activation function layer, a maximum pooling layer, and a Dropout layer connected in sequence.
[0218] The feature vector output by the single-scale two-dimensional convolution module is input into the multi-scale two-dimensional convolution module in the branch network for secondary spatial feature extraction. The multi-scale two-dimensional convolution module is composed of a two-dimensional convolution layer, a batch normalization layer, an activation function layer, a spatial pyramid pooling layer, and a Dropout layer connected in sequence.
[0219] The feature vector output by the multi-scale two-dimensional convolution module is input into the bidirectional long short-term memory network in the branch network for temporal feature extraction to obtain the full sky image features corresponding to the input full sky image.
[0220] Optionally, in the multi-modal multi-scale solar radiation prediction device provided in another embodiment of the present application, when the prediction unit performs prediction based on the fused feature vector to obtain predicted values of target radiation parameters at multiple future time points, it is used to:
[0221] The fused feature vector is processed through the fully connected layer in the multimodal feature fusion module to map the predicted values of the clear sky factor at multiple future time points.
[0222] According to the conversion relationship between the clear sky factor and the horizontal total radiation, the predicted values of the clear sky factor at multiple future time points are used to calculate the predicted values of the horizontal total radiation at multiple future time points.
[0223] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.
[0224] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A multi-modal and multi-scale solar radiation prediction method, characterized in that: include: Acquire data of various meteorological variables and full sky images at multiple time points within the current specified time period; wherein the meteorological variables include multiple target meteorological parameters and multiple radiation parameters; Inputting a meteorological variable sequence composed of data of each of the meteorological variables at each time point, and a full sky image sequence composed of each of the full sky images, into a pre-trained radiation prediction model; Extracting features of the meteorological variable sequence through a meteorological variable feature extraction module in the radiation prediction model to obtain meteorological variable features, and extracting features of the full sky image sequence through a full sky image feature extraction module in the radiation prediction model to obtain full sky image features; The meteorological variable features are fused with the all-sky image features through the multimodal feature fusion module in the radiation prediction model to obtain a fused feature vector, and based on the fused feature vector, predicted values of the target radiation parameters at multiple future time points are predicted.
2. The method according to claim 1, characterized in that Also includes: Obtain data on various meteorological parameters at multiple time points within a preset time period; Analyze and remove abnormal values in the data of each meteorological parameter; Filling in missing values in the data of each of the meteorological parameters; Based on the data of each meteorological parameter, calculating the radiation correlation coefficient of each meteorological parameter; According to the order of the radiation correlation coefficient of each meteorological parameter from large to small, multiple meteorological parameters are selected to be determined as the target meteorological parameters.
3. The method according to claim 2, characterized in that The analyzing and removing abnormal values in the data of each meteorological parameter includes: For each of the meteorological parameters, the distribution of the data of the meteorological parameter is represented in the form of a box plot; Determine each data of the meteorological parameter exceeding the upper limit line in the box plot as a preliminary abnormal value; Based on the data of the adjacent moments of each of the preliminary abnormal values, each of the preliminary abnormal values is divided into a first abnormal value and a second abnormal value; wherein the first abnormal value is an abnormal value caused by a collection problem; and the second abnormal value is an abnormal value caused by extreme weather; Each of the first outliers is determined as a final outlier in the data of the meteorological parameter and is removed.
4. The method according to claim 1, characterized in that: Before inputting the meteorological variable sequence composed of the data of each of the meteorological variables at each time point and the full sky image sequence composed of each of the full sky images into the pre-trained radiation prediction model, the method further includes: Cleaning each of the full sky images and aligning them with the time tags of the data of each of the meteorological variables; According to a preset cropping radius, each of the full sky images after time tag alignment is cropped using a binary mask method; wherein pixels within the preset cropping radius in the cropped full sky image are valid pixels, and pixels outside the preset cropping radius are invalid pixels; The resolution of each of the cropped all-sky images is reduced and normalized.
5. The method according to claim 1, characterized in that The step of extracting features of the meteorological variable sequence through a meteorological variable feature extraction module in the radiation prediction model to obtain meteorological variable features includes: The meteorological variable sequence is sequentially passed through each convolution network in the meteorological variable feature extraction module to extract spatial features; wherein one of the convolution networks is composed of a one-dimensional convolution layer, a maximum pooling layer and a Dropout layer connected in sequence; The feature vector output by the last convolutional network is input into the bidirectional long short-term memory network for temporal feature extraction; Inputting the feature vector output by the bidirectional long short-term memory network into the attention mechanism layer for feature weight allocation; The feature vector output by the attention mechanism layer is input into the flattening layer for flattening to obtain the final feature of the meteorological variable.
6. The method according to claim 1, characterized in that The step of extracting features from the full sky image sequence by using the full sky image feature extraction module in the radiation prediction model to obtain full sky image features includes: Inputting each of the full sky images in the full sky image sequence into a branch network of the full sky image feature extraction module; wherein the structures of the branch networks are consistent; The input full sky image is subjected to spatial feature extraction through the single-scale two-dimensional convolution module in the branch network; wherein the single-scale two-dimensional convolution module is composed of a two-dimensional convolution layer, a batch normalization layer, an activation function layer, a maximum pooling layer, and a Dropout layer connected in sequence; The feature vector output by the single-scale two-dimensional convolution module is input into the multi-scale two-dimensional convolution module in the branch network for secondary spatial feature extraction; wherein the multi-scale two-dimensional convolution module is composed of a two-dimensional convolution layer, a batch normalization layer, an activation function layer, a spatial pyramid pooling layer, and a Dropout layer connected in sequence; The feature vector output by the multi-scale two-dimensional convolution module is input into the bidirectional long short-term memory network in the branch network to extract time features, so as to obtain the full sky image features corresponding to the input full sky image.
7. The method according to claim 1, characterized in that The step of predicting the target radiation parameters at multiple future time points based on the fused feature vector includes: The fused feature vector is processed by a fully connected layer in the multimodal feature fusion module to map predicted values of clear sky factors at multiple future time points; According to the conversion relationship between the clear sky factor and the horizontal total radiation, the predicted values of the clear sky factor at multiple future time points are used to calculate the predicted values of the horizontal total radiation at multiple future time points.
8. A multi-modal and multi-scale solar radiation prediction device, characterized in that: include: A data acquisition unit, used to acquire data of various meteorological variables and a full sky image at multiple time points within a currently specified time period; wherein the meteorological variables include multiple target meteorological parameters and multiple radiation parameters; A prediction unit is used to input a meteorological variable sequence composed of data of each of the meteorological variables at each time point and a full-sky image sequence composed of each of the full-sky images into a pre-trained radiation prediction model, perform feature extraction on the meteorological variable sequence through a meteorological variable feature extraction module in the radiation prediction model to obtain meteorological variable features, and perform feature extraction on the full-sky image sequence through a full-sky image feature extraction module in the radiation prediction model to obtain full-sky image features, then fuse the meteorological variable features with the full-sky image features through a multimodal feature fusion module in the radiation prediction model to obtain a fused feature vector, and predict the predicted values of the target radiation parameters at multiple future time points based on the fused feature vector.
9. The device according to claim 8, characterized in that Also includes: A data acquisition unit, used to obtain data of various meteorological parameters at multiple time points within a preset time period; A filtering unit, used to analyze and remove abnormal values in the data of each of the meteorological parameters; A filling unit, used to fill in missing values in the data of each of the meteorological parameters; A coefficient calculation unit, used for calculating the radiation correlation coefficient of each meteorological parameter based on the data of each meteorological parameter; The selection unit is used to select multiple meteorological parameters as the target meteorological parameters in descending order of the radiation correlation coefficients of the meteorological parameters.
10. The device according to claim 9, characterized in that The filtering unit comprises: A representation unit, for representing the distribution of the data of each meteorological parameter in the form of a box plot; A first determining unit is used to determine each data of the meteorological parameter exceeding the upper limit line in the box plot as a preliminary abnormal value; A division unit, configured to divide each of the preliminary abnormal values into a first abnormal value and a second abnormal value based on data of the adjacent moments of each of the preliminary abnormal values; wherein the first abnormal value is an abnormal value caused by a collection problem; and the second abnormal value is an abnormal value caused by extreme weather; The second determination unit is used to determine each of the first abnormal values as a final abnormal value in the data of the meteorological parameter and remove it.
Citation Information
Patent Citations
Solar radiation prediction method and system based on double-branch feature extraction
CN115099461A
Ultra-short-term solar radiation prediction method and system based on sparse spatial-temporal feature descriptor
CN115457389A
Ultra-short-term solar irradiance prediction method based on multi-modal feature fusion
CN118470486A
Short-term photovoltaic generating capacity prediction method and device
CN118554427A
System and method for predicting photovoltaics power generation
KR102577207B1
Cited By
Multi-modal irradiance prediction method and system of deep fusion physical model
CN122241619A
