A ground-based cloud image based solar irradiance ultra-short-term prediction method

Abstract

The application discloses a kind of solar irradiance ultra-short term prediction method based on ground-based cloud image, it is related to photovoltaic power prediction technical field, the method includes: obtaining the cloud image sequence and irradiance numerical sequence of photovoltaic power station historical time;Using binary mask matrix conversion technology to extract the target sky area image in cloud image sequence not affected by building, target sky area image is sequentially carried out contrast enhancement and standardization processing, obtain the cloud feature matrix of each time;Irradiance numerical value and cloud feature matrix are integrated to construct optimal training sample, using optimal training sample to construct and train the ultra-short term prediction model based on comprehensive similarity evaluation index, the ultra-short term prediction result of future time solar irradiance is output by ultra-short term prediction model.The application realizes the ultra-short term prediction of future time solar irradiance considering cloud feature, so as to effectively enhance the application effect of cloud information in solar irradiance ultra-short term prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of photovoltaic power prediction, in particular to a solar irradiance ultra-short-term prediction method based on ground-based cloud images. BACKGROUND

[0002] The ultra-short-term (0-4 hours) power prediction of photovoltaic power stations is crucial for the efficient operation and maintenance, power dispatching and energy market transactions of photovoltaic power stations. As the core parameter determining the photovoltaic power generation, the prediction accuracy of irradiance directly affects the power generation income, power grid stability and power supply and demand balance of the power station.

[0003] At present, the common methods for ultra-short-term irradiance prediction mainly include physical model method, statistical model method and machine learning method. The physical model method is based on the principle of atmospheric radiation transmission, and simulates the change of irradiance by inputting meteorological parameters, but its adaptability to complex atmospheric conditions is poor. The statistical model method relies on historical data to establish an empirical relationship, and it is difficult to capture the nonlinear dynamic changes of irradiance. Although the machine learning method can improve the prediction ability through data training, it requires a large amount of computing resources to complete the training, and in the case of rapid movement and mutation of clouds, the computing efficiency and tracking ability will decrease significantly.

[0004] With the development of remote sensing technology, ground-based cloud images provide a new breakthrough direction for irradiance prediction due to their high spatial and temporal resolution cloud observation capability. How to deeply explore the value of cloud image data and build an irradiance ultra-short-term prediction model that takes into account accuracy and efficiency is a technical problem that needs to be solved.

[0005] In view of the problems in the related art, no effective solutions have been proposed so far. SUMMARY

[0006] In view of the problems in the related art, the present application proposes a solar irradiance ultra-short-term prediction method based on ground-based cloud images to overcome the above technical problems existing in the prior art.

[0007] To this end, the specific technical solutions adopted by the present application are as follows:

[0008] A solar irradiance ultra-short-term prediction method based on ground-based cloud images, the method comprising:

[0009] S1, obtaining a cloud image sequence and an irradiance value sequence of a photovoltaic power station at a historical time;

[0010] S2, using a binary mask matrix conversion technology to extract the target sky area image in the cloud image sequence that is not affected by buildings, and sequentially performing contrast enhancement and standardization processing on the target sky area image to obtain a cloud feature matrix at each time;

[0011] S3, the optimal training sample is constructed by the integrated irradiance value and the cloud image feature matrix, the super short-term prediction model based on the integrated similar evaluation index is constructed and trained by using the optimal training sample, and the super short-term prediction result of the solar irradiance at the future time is output by the super short-term prediction model.

[0012] Preferably, the cloud image sequence at the historical time of the photovoltaic power station is a three-channel matrix sequence; and the irradiance value sequence is the irradiance data of the same position of the cloud monitoring station.

[0013] Preferably, the target sky area image in the cloud image sequence which is not affected by the building is extracted by using the binary mask matrix conversion technology, and the target sky area image is sequentially subjected to contrast enhancement and standardization processing to obtain the cloud image feature matrix at each time, which includes:

[0014] S21, the width and height of the cloud image in the cloud image sequence are obtained, and the cloud center coordinates are calculated; the cloud image is converted to the HSV color space to generate an HSV cloud image;

[0015] S22, the threshold range is set according to the color range of the sky area, the pixel-level judgment of the HSV cloud image is performed, and the binary mask matrix is constructed based on the pixel-level judgment result of the HSV cloud image;

[0016] S23, the maximum inscribed circle radius of the sky area in the binary mask matrix is obtained with the cloud center coordinates as the center, and the target sky area image in the cloud image which is not affected by the building is cut off in combination with the center and the maximum inscribed circle radius;

[0017] S24, the target sky area image is subjected to contrast enhancement processing to obtain the cloud image pixel matrix after cloud feature enhancement, and the cloud image pixel matrix and the irradiance value sequence are respectively subjected to standardization processing to obtain the cloud image feature matrix at each time and the normalized irradiance value.

[0018] Preferably, the threshold range is set according to the color range of the sky area, the pixel-level judgment of the HSV cloud image is performed, and the binary mask matrix is constructed based on the pixel-level judgment result of the HSV cloud image, which includes:

[0019] S221, the threshold range is set based on the color range of the sky area; the hue, brightness and saturation of the undetermined pixel in the HSV cloud image are extracted and compared with the threshold range in sequence;

[0020] S222, if the hue, brightness and saturation of the current pixel are all within the threshold range, the current pixel is determined as the sky area which is not affected by the building, and the corresponding position in the mask matrix is marked as 1;

[0021] S223, if any of the hue, brightness and saturation of the current pixel exceeds the threshold range, the current pixel is determined as a non-sky region, and the corresponding position in the mask matrix is marked as 0;

[0022] S224, repeating steps S222-S223 until all pixels of the HSV cloud chart are traversed to stop, forming a binary mask matrix.

[0023] Preferably, the target sky region image is subjected to contrast enhancement processing to obtain a cloud feature enhanced cloud image pixel matrix, which comprises:

[0024] Separate the brightness channel from the HSV cloud chart, and use the adaptive histogram equalization method to enhance the brightness contrast of the brightness channel;

[0025] After merging the contrast-enhanced brightness channel with the remaining channels, return to the original cloud chart image to obtain the cloud feature-enhanced cloud image pixel matrix.

[0026] Preferably, the comprehensive irradiance value and cloud feature matrix are used to construct an optimal training sample, and the optimal training sample is used to construct and train a super-short-term prediction model based on a comprehensive similarity evaluation index, and the super-short-term prediction result of the future time of the solar irradiance output by the super-short-term prediction model comprises:

[0027] S31, extracting the cloud feature matrix and the normalized irradiance value at the same historical time, respectively constructing the irradiance view training sample and the cloud view training sample;

[0028] S32, taking the irradiance value as the first feature and the cloud feature matrix as the second feature, constructing a comprehensive similarity calculation model, and iteratively training the comprehensive similarity calculation model through the irradiance view training sample and the cloud view training sample;

[0029] S33, outputting the comprehensive similarity score of the irradiance view training sample through the trained comprehensive similarity calculation model, and calculating the super-short-term prediction result of the future time of the solar irradiance based on the comprehensive similarity score.

[0030] Preferably, the comprehensive similarity calculation model is constructed by taking the irradiance value as the first feature and the cloud feature matrix as the second feature, and the comprehensive similarity calculation model is iteratively trained through the irradiance view training sample and the cloud view training sample, which comprises:

[0031] S321, constructing an irradiance similarity matrix based on the normalized irradiance value, and constructing a cloud similarity matrix based on the structural index of the cloud feature matrix;

[0032] S322, set a clustering cluster number threshold, and perform spectral clustering on the irradiance view and the cloud view based on the clustering cluster number threshold respectively to obtain an irradiance clustering result and a cloud clustering result;

[0033] S323, based on the irradiance clustering result, adjust the cloud similarity matrix according to a preset constraint rule, and simultaneously adjust the irradiance similarity matrix based on the cloud clustering result, re-cluster the adjusted cloud similarity matrix and the adjusted irradiance similarity matrix, and terminate iteration when the clustering result no longer changes or meets a maximum iteration number, to obtain a final irradiance clustering result and a final cloud clustering result;

[0034] S324, calculate a contour coefficient between the final irradiance clustering result and the final cloud clustering result, assign weights to the final irradiance clustering result and the final cloud clustering result based on the contour coefficient respectively, and obtain a comprehensive similarity calculation model.

[0035] Preferably, the expression of the preset constraint rule for adjusting the cloud similarity matrix according to the preset constraint rule is:

[0036] ;

[0037] In the formula, indicates the cloud similarity matrix of the i-th day and the j-th day at t time; and α and β respectively indicate an enhancement coefficient and a weakening coefficient.

[0038] Preferably, the comprehensive similarity calculation model outputted by the training includes:

[0039] S331, calculate the comprehensive similarity score of the irradiance view training sample by the trained comprehensive similarity calculation model, arrange the comprehensive similarity scores in descending order, and select the first M groups of irradiance view training samples that meet a preset training sample amount threshold;

[0040] S332, based on the comprehensive similarity scores of the first M groups of irradiance view training samples, construct a weight matrix by using a comprehensive similarity numerical weighting method;

[0041] S333, extract irradiance values of future 1st to Tth time after the current time of the first M groups of irradiance view training samples, obtain a new irradiance value sequence by weighted calculation, and take the new irradiance value sequence as the super short-term prediction value of the solar irradiance from the current time to the future time.

[0042] Preferably, the expression of the comprehensive similarity calculation model is:

[0043] ;

[0044] In the formula, denotes the final comprehensive similarity matrix; denotes the weight of the irradiance similarity matrix; denotes the final irradiance similarity matrix; denotes the final cloud image similarity matrix.

[0045] The beneficial effects of the present application are:

[0046] 1. The present application realizes key pixel feature enhancement extraction of ground-based cloud images by recognizing the effective sky area of the ground-based cloud image and enhancing the contrast of cloud features, realizes cloud feature enhancement through cutting and strengthening processing of cloud image features, thereby providing more intuitive and accurate input feature quantities for subsequent prediction modeling, reducing the influence of redundant features in the model training process, and further improving the accuracy of model training.

[0047] 2. The present application innovatively proposes a solar irradiance ultra-short-term prediction method based on irradiance and cloud feature matrix collaborative training, dynamically fuses cloud image time sequence features and irradiance time sequence features, constructs a comprehensive similarity evaluation index, makes the historical similar sample selection closer to the current real-time state, improves the accuracy of solar irradiance ultra-short-term prediction, at the same time reasonably reduces the sample search range, improves the modeling efficiency of ultra-short-term prediction, improves the traditional photovoltaic ultra-short-term training method, and further can accurately extract the optimal training sample.

[0048] 3. The present application comprehensively considers the correlation characteristics of cloud image features and irradiance, constructs a comprehensive evaluation index, realizes the ultra-short-term prediction of future time solar irradiance considering cloud image features, thereby effectively enhancing the application effect of cloud image information in solar irradiance ultra-short-term prediction, and through constructing an irradiance ultra-short-term prediction model considering recognition accuracy and calculation efficiency, provides more accurate meteorological change information for the ultra-short-term prediction of photovoltaic power stations. BRIEF DESCRIPTION OF DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0050] Figure 1 is a flowchart of a solar irradiance ultra-short-term prediction method based on ground-based cloud images according to an embodiment of the present application. DETAILED DESCRIPTION

[0051] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0052] According to an embodiment of the present invention, a method for ultra-short-term prediction of solar irradiance based on ground-based cloud maps is provided.

[0053] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, according to an embodiment of the present invention, a method for ultra-short-term solar irradiance prediction based on ground-based cloud maps includes:

[0054] S1. Obtain the cloud image sequence and irradiance value sequence of the photovoltaic power station at historical moments.

[0055] The cloud image sequence of the photovoltaic power station at a historical moment is a three-channel matrix sequence; the irradiance value sequence is the irradiance data of the same location of the cloud image monitoring station.

[0056] S2. Use binary mask matrix transformation technology to extract the target sky region image that is not affected by buildings in the cloud image sequence. Perform contrast enhancement and standardization processing on the target sky region image in sequence to obtain the cloud image feature matrix at each time step.

[0057] The step of extracting the target sky region image unaffected by buildings from the cloud image sequence using binary mask matrix transformation technology, and sequentially performing contrast enhancement and standardization processing on the target sky region image to obtain the cloud image feature matrix at each time step includes:

[0058] S21. Obtain the width and height of the cloud image in the cloud image sequence, calculate the center coordinates of the cloud image; convert the cloud image to the HSV color space to generate an HSV cloud image;

[0059] S22. Set a threshold range based on the color range of the sky region to perform pixel-level determination on the HSV cloud map, and construct a binary mask matrix based on the pixel-level determination result of the HSV cloud map.

[0060] The step of setting a threshold range based on the color range of the sky region to perform pixel-level determination of the HSV cloud image, and constructing a binary mask matrix based on the pixel-level determination result of the HSV cloud image includes:

[0061] S221. Set a threshold range based on the color range of the sky region; extract the hue, brightness, and saturation of undetermined pixels in the HSV cloud image and compare them sequentially with the threshold range;

[0062] S222, if the hue, brightness and saturation of the current pixel are all within the threshold range, the current pixel is determined as a sky region not affected by buildings, and the corresponding position in the mask matrix is marked as 1;

[0063] S223, if any of the hue, brightness and saturation of the current pixel is out of the threshold range, the current pixel is determined as a non-sky region, and the corresponding position in the mask matrix is marked as 0;

[0064] S224, repeating steps S222-S223 until all pixels of the HSV cloud chart are traversed to stop, forming a binary mask matrix.

[0065] S23, taking the center coordinates of the cloud chart as the center of a circle, obtaining the maximum inscribed circle radius of the sky region in the binary mask matrix, and combining the center and the maximum inscribed circle radius to cut the target sky region image in the cloud chart image not affected by buildings;

[0066] S24, performing contrast enhancement processing on the target sky region image to obtain a cloud feature enhanced cloud image pixel matrix, and performing standardization processing on the cloud image pixel matrix and the irradiance value sequence respectively to obtain a cloud feature matrix and a normalized irradiance value at each time.

[0067] Wherein, the contrast enhancement processing on the target sky region image to obtain a cloud feature enhanced cloud image pixel matrix includes:

[0068] Separate the brightness channel from the HSV cloud chart, and use the adaptive histogram equalization method to enhance the brightness contrast of the brightness channel;

[0069] After merging the contrast-enhanced brightness channel with the remaining channels, return to the original cloud chart image to obtain a cloud feature-enhanced cloud image pixel matrix.

[0070] S3, constructing an optimal training sample by integrating the irradiance value and the cloud feature matrix, constructing and training an ultra-short-term prediction model based on a comprehensive similarity evaluation index using the optimal training sample, and outputting an ultra-short-term prediction result of the solar irradiance at the future time through the ultra-short-term prediction model.

[0071] Wherein, the contrast enhancement processing on the target sky region image to obtain a cloud feature enhanced cloud image pixel matrix includes:

[0072] S31, extracting the cloud feature matrix and the normalized irradiance value at the same historical time, and constructing an irradiance view training sample and a cloud view training sample respectively;

[0073] S32, taking the irradiance value as the first feature and the cloud image feature matrix as the second feature, constructing a comprehensive similarity calculation model, and performing collaborative iterative training on the comprehensive similarity calculation model through the irradiance view training sample and the cloud image view training sample.

[0074] The comprehensive similarity calculation model is constructed by taking the irradiance value as the first feature and the cloud image feature matrix as the second feature, and the comprehensive similarity calculation model is collaboratively iteratively trained by the irradiance view training sample and the cloud image view training sample.

[0075] S321, based on the normalized irradiance value, an irradiance similarity matrix is constructed, and based on the structural index of the cloud image feature matrix, a cloud image similarity matrix is constructed;

[0076] S322, a clustering cluster number threshold is set, and spectral clustering is performed on the irradiance view and the cloud image view based on the clustering cluster number threshold, to obtain an irradiance clustering result and a cloud image clustering result;

[0077] S323, based on the irradiance clustering result, the cloud image similarity matrix is adjusted according to a preset constraint rule, and based on the cloud image clustering result, the irradiance similarity matrix is adjusted, and the adjusted cloud image similarity matrix and the irradiance similarity matrix are re-clustered until the clustering result no longer changes or the maximum iteration number is met, to obtain the final irradiance clustering result and the cloud image clustering result;

[0078] S324, the contour coefficient between the final irradiance clustering result and the cloud image clustering result is calculated, the final irradiance clustering result and the cloud image clustering result are respectively assigned weights based on the contour coefficient, and a comprehensive similarity calculation model is obtained.

[0079] S33, the comprehensive similarity calculation model outputting the comprehensive similarity score of the irradiance view training sample is trained, and the super-short-term prediction result of the solar irradiance at the future time is calculated based on the comprehensive similarity score.

[0080] The comprehensive similarity calculation model outputting the comprehensive similarity score of the irradiance view training sample is trained, and the super-short-term prediction result of the solar irradiance at the future time is calculated based on the comprehensive similarity score.

[0081] S331, the comprehensive similarity score of the irradiance view training sample is calculated by the trained comprehensive similarity calculation model, the comprehensive similarity score is arranged in descending order, and the first M groups of irradiance view training samples satisfying the preset training sample quantity threshold are selected;

[0082] S332, based on the comprehensive similarity score of the first M group of irradiance view training samples, a weight matrix is constructed by using a comprehensive similarity value weighting method;

[0083] S333, extract the irradiance values of the future 1st to Tth time after the current time of the first M group of irradiance view training samples, obtain a new irradiance value sequence through weighted calculation, as the ultra-short-term prediction value of the solar irradiance from the current time to the future time.

[0084] The solar irradiance ultra-short-term prediction method based on ground-based cloud images provided by the present application will be further described below in conjunction with specific embodiments.

[0085] The present application introduces an irradiance and cloud feature collaborative training clustering method by effective area recognition and cloud feature enhancement of ground-based cloud images, comprehensively considers the information of irradiance values and cloud feature matrix, constructs an ultra-short-term prediction model based on comprehensive similarity evaluation index, and obtains the ultra-short-term prediction result of solar irradiance in the future 4 hours, which specifically includes:

[0086] Step one, obtain a cloud image sequence and a corresponding irradiance value sequence of a photovoltaic power station in a period of time.

[0087] Among them, the cloud image is a three-channel matrix (RGB) sequence after transformation , the irradiance value sequence is the irradiance information of the same position of the cloud monitoring station , wherein N represents the length of the data sequence, and n represents the nth value of the data sequence, .

[0088] Step two, considering the strong correlation between cloud changes and photovoltaic power station power generation, the present application takes the cloud image of the photovoltaic power station as the center, considers the difference of the cloud image construction site of different power stations, extracts and determines the effective sky area radius R which is not affected by surrounding buildings through binary mask matrix transformation, and reduces the influence of buildings on cloud feature extraction.

[0089] On the basis of the traditional technical scheme, the present application adds a regional sky-cloud contrast enhancement module, strengthens the cloud features in the cloud image pixel matrix, and on this basis, constructs a per-time effective sky area cloud feature matrix, and respectively standardizes the irradiance sequence and the pixel matrix.

[0090] Specifically, taking the cloud image as the center, considering the difference of the cloud image construction site of different power stations, extracting and determining the effective sky area radius R which is not affected by surrounding buildings through binary mask matrix transformation, including:

[0091] Obtain the width and height of the cloud image and calculate the center coordinates, convert the three-channel RGB cloud image (the cloud image is a three-channel RGB cloud image) to the HSV color space and obtain the HSV cloud image, set the threshold according to the color range of the sky area, realize the differentiation of the sky and the building, construct a binary mask matrix, the inside of the mask is 1, and the outside (the building area) is 0, take the center coordinates as the center, calculate the maximum inscribed circle radius R of the sky area, and realize the effective sky area image (equivalent to the target sky area image) interception.

[0092] In view of the demand of the photovoltaic power station for cloud feature extraction, a regional sky-cloud contrast enhancement module is added to strengthen the cloud features in the cloud pixel matrix, and on this basis, a per-time cloud feature matrix is constructed, which specifically includes:

[0093] The effective sky area image is subjected to contrast enhancement, the brightness channel V is separated from the HSV color space image, the adaptive histogram equalization method is applied to the brightness channel V for brightness contrast enhancement, the enhanced brightness channel is combined with other channels, the three-channel RGB cloud image is converted back, and the cloud feature enhanced cloud pixel matrix is obtained.

[0094] The irradiance sequence and the cloud pixel matrix are respectively subjected to standardization processing, the obtained cloud pixel matrix is subjected to standardization processing (0-1 numerical range), and a new cloud feature matrix is constructed The irradiance sequence is subjected to min-max normalization processing.

[0095] Among them, the HSV color space image is a color model visualization tool for representing three attributes of color: hue (H) represents the type of color, which is represented by an angle (0°-360°); saturation (S) represents the vividness of the color, which changes from the center to the edge; and brightness (V) represents the lightness of the color, which changes from black to white.

[0096] Step three, in order to improve the training effect of the ultra-short-term power prediction model, the present application proposes a per-time optimal training sample subset construction module considering the irradiance value and the cloud feature matrix. First, the historical sequence is twice split according to the resolution of the prediction time sequence, the irradiance value and the pixel matrix at the same time in the historical sequence are extracted, and the irradiance view and the cloud view training sample subset are constructed; then, based on the correlation characteristics of irradiance and cloud and photovoltaic power generation, through double-view collaborative iterative training, taking the irradiance value as the first feature, the cloud feature enhanced and standardized cloud feature matrix as the second feature, a comprehensive similarity calculation model is constructed.

[0097] Specifically, the comprehensive similarity calculation model is constructed by comprehensively considering the irradiance value and the hourly optimal training sample subset of the cloud image feature matrix, the irradiance value and the pixel matrix at the same time in the historical sequence are extracted, the time point is taken as the modeling unit, the accuracy of modeling is improved, the training subset is constructed by screening according to the time label of the sample, and the data sample at the same time as the current time is screened out, for example, the current time is 8 o'clock, and the cloud image feature matrix at 8 o'clock of each day in the history is screened.

[0098] Wherein, the double-view collaborative iterative training is adopted, the irradiance value is taken as the first feature, the cloud image pixel matrix is taken as the second feature, and the comprehensive similarity calculation model is constructed, which specifically includes:

[0099] (1) Constructing a similarity matrix, the initial irradiance similarity matrix The calculation is performed by using the standardized exponential similarity, wherein the calculation formula of the irradiance similarity matrix is:

[0100] ;

[0101] In the formula, denotes the irradiance similarity matrix of the i-th day and the j-th day, and denote the irradiance value of the i-th day and the j-th day respectively, denotes the irradiance standard deviation.

[0102] Cloud image similarity matrix It is characterized by calculating the structural similarity index (SSIM) of the enhanced image matrix.

[0103] (2) Initialization clustering, setting the clustering cluster number threshold K, performing spectral clustering on the irradiance view and the cloud image view respectively, and obtaining the initial clustering result, wherein the data labels of the irradiance view and the cloud image view are and The two view weights and are initialized, each of which is 0.5.

[0104] (3) Fusion iterative training, first, based on the current irradiance clustering result, adjust the cloud image similarity matrix, enhance the cloud image similarity of samples in the same cluster, and weaken the cloud image similarity of samples between different clusters;

[0105] The adjustment rule is as follows:

[0106] ;

[0107] In the formula, denotes the cloud image similarity matrix of the i-th day and the j-th day at t time; and β denote the enhancement coefficient and the weakening coefficient respectively, and the new cloud image similarity clustering result and its label are obtained by adjusting and.

[0108] Symmetrically, the irradiance similarity matrix is adjusted according to the cloud map clustering result, and re-clustering is performed based on the adjusted irradiance similarity. When the clustering result no longer changes significantly or the maximum number of iterations is reached, the iteration is terminated.

[0109] (4) According to the final clustering result, the contour coefficients of the two view clustering results are calculated, the respective weights are dynamically determined, and the comprehensive similarity calculation formula is obtained:

[0110] ;

[0111] In the formula, indicates the final comprehensive similarity matrix; indicates the weight of the irradiance similarity matrix; indicates the final irradiance similarity matrix, i.e. the final irradiance clustering result; indicates the final cloud map similarity matrix, i.e. the final cloud map clustering result.

[0112] Through the double-view collaborative iterative training mode, the irradiance monitoring and cloud map are innovatively integrated, the training set construction method in the traditional technical solution is improved, and the accuracy of modeling is further improved.

[0113] Step four, in order to improve the calculation efficiency of the prediction link, considering the continuity law of cloud change, on the basis of steps two to three, the training samples are arranged in descending order according to the comprehensive similarity score, a training sample amount threshold is set, the irradiance data of the first M group of samples with higher comprehensive similarity score is extracted, and a sample weight matrix is constructed according to the similarity score of the M group of samples, which specifically includes:

[0114] M groups of irradiance values are selected , according to the length requirement T of the ultra-short-term prediction at the moment, T irradiance values after are selected. .

[0115] According to the comprehensive similarity between the M group of irradiance values and the current moment, a weight matrix is constructed. The weight matrix is obtained by weighting the comprehensive similarity value. Here, the traditional similarity prediction method is improved, and a weight calculation method based on comprehensive similarity is specially introduced to realize more reasonable weight distribution.

[0116] Step five, the irradiance values of the first to T moments in the future after the current moment of the M group of samples are extracted, and a new irradiance value sequence is obtained by weighted calculation, which is used as the ultra-short-term prediction value of the future irradiance at the current moment.

[0117] The new irradiance value sequence is obtained by weighted calculation, as the ultra-short-term prediction value of future irradiance at the current time, and specifically includes:

[0118] The calculation method of the irradiance prediction value of the future t time is as follows:

[0119]

[0120] The irradiance prediction values of the future 1 to T times are calculated, and the ultra-short-term prediction sequence of future irradiance at the current time is obtained .

[0121] In summary, by means of the above technical solutions of the present application, the present application realizes key pixel feature enhancement extraction of ground-based cloud images through ground-based cloud image sky area recognition and cloud feature contrast enhancement; through cloud feature segmentation and strengthening processing, cloud feature enhancement is realized, thereby providing more intuitive and accurate input feature quantity for subsequent prediction modeling, reducing the influence of redundant features in the model training process, and further improving the accuracy of model training; the present application innovatively proposes a solar irradiance ultra-short-term prediction method based on irradiance and cloud feature matrix collaborative training, dynamically fuses cloud image time sequence features and irradiance time sequence features, constructs a comprehensive similarity evaluation index, makes the historical similar sample selection more close to the current real-time state, improves the accuracy of solar irradiance ultra-short-term prediction, and at the same time reasonably reduces the sample search range, improves the modeling efficiency of ultra-short-term prediction, improves the traditional photovoltaic ultra-short-term training method, and further can accurately extract the optimal training sample; the present application comprehensively considers the correlation characteristics of cloud features and irradiance, constructs a comprehensive evaluation index, realizes the ultra-short-term prediction of future solar irradiance considering cloud features, thereby effectively enhancing the application effect of cloud information in solar irradiance ultra-short-term prediction, and through the construction of an irradiance ultra-short-term prediction model considering recognition accuracy and calculation efficiency, more accurate meteorological change information is provided for photovoltaic power station ultra-short-term prediction.

[0122] The above only describes the preferred embodiments of the present application and is not intended to limit the present application, and any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.​​

Claims

1. A method for ultra-short-term prediction of solar irradiance based on ground-based cloud maps, characterized in that, The method includes: S1. Obtain the cloud image sequence and irradiance value sequence of the photovoltaic power station at historical moments; S2. Use binary mask matrix transformation technology to extract the target sky region image that is not affected by buildings in the cloud image sequence. Perform contrast enhancement and standardization processing on the target sky region image in sequence to obtain the cloud image feature matrix at each time step. S3. Extract the cloud map feature matrix and the normalized irradiance value at the same historical moment, and construct irradiance view training samples and cloud map view training samples respectively. An irradiance similarity matrix is ​​constructed based on the normalized irradiance values, and a cloud map similarity matrix is ​​constructed based on the structural index of the cloud map feature matrix. Set a threshold for the number of clusters, and perform spectral clustering on the irradiance view and the cloud map view based on the threshold to obtain the irradiance clustering results and the cloud map clustering results; Based on the irradiance clustering results, the cloud map similarity matrix is ​​adjusted according to the preset constraint rules. At the same time, the irradiance similarity matrix is ​​adjusted based on the cloud map clustering results. The adjusted cloud map similarity matrix and irradiance similarity matrix are re-clustered until the clustering results no longer change or the maximum number of iterations is met, at which point the iteration is terminated, and the final irradiance clustering result and cloud map clustering result are obtained. The silhouette coefficient between the final irradiance clustering result and the cloud map clustering result is calculated. Based on the silhouette coefficient, weights are assigned to the final irradiance clustering result and the cloud map clustering result respectively to obtain the comprehensive similarity calculation model. The trained comprehensive similarity calculation model outputs the comprehensive similarity score of the irradiance view training samples, and the ultra-short-term prediction result of the solar irradiance at future times is calculated based on the comprehensive similarity score.

2. The method for ultra-short-term solar irradiance prediction based on ground-based cloud maps according to claim 1, characterized in that, The cloud image sequence of the photovoltaic power station at a historical moment is a three-channel matrix sequence; the irradiance value sequence is the irradiance data of the same location of the cloud image monitoring station.

3. The method for ultra-short-term solar irradiance prediction based on ground-based cloud maps according to claim 1, characterized in that, The method of extracting the target sky region image unaffected by buildings from the cloud image sequence using binary mask matrix transformation technology, and sequentially performing contrast enhancement and standardization processing on the target sky region image to obtain the cloud image feature matrix at each time step includes: S21. Obtain the width and height of the cloud image in the cloud image sequence, calculate the center coordinates of the cloud image; convert the cloud image to the HSV color space to generate an HSV cloud image; S22. Set a threshold range based on the color range of the sky region to perform pixel-level determination on the HSV cloud image, and construct a binary mask matrix based on the pixel-level determination result of the HSV cloud image. S23. Using the center coordinates of the cloud image as the center, obtain the maximum inscribed circle radius of the sky region in the binary mask matrix, and combine the center and the maximum inscribed circle radius to extract the target sky region image that is not affected by buildings in the cloud image. S24. Perform contrast enhancement processing on the target sky region image to obtain the cloud image pixel matrix after cloud feature enhancement, and perform standardization processing on the cloud image pixel matrix and irradiance value sequence respectively to obtain the cloud image feature matrix and normalized irradiance value at each time step.

4. The method for ultra-short-term solar irradiance prediction based on ground-based cloud maps according to claim 3, characterized in that, The step of setting a threshold range based on the color range of the sky region to perform pixel-level determination of the HSV cloud image, and constructing a binary mask matrix based on the pixel-level determination result of the HSV cloud image includes: S221. Set a threshold range based on the color range of the sky region; extract the hue, brightness, and saturation of undetermined pixels in the HSV cloud image and compare them sequentially with the threshold range; S222. If the hue, brightness, and saturation of the current pixel are all within the threshold range, then the current pixel is determined to be a sky area unaffected by buildings, and marked as 1 at the corresponding position in the mask matrix. S223. If any of the hue, brightness, or saturation of the current pixel exceeds the threshold range, the current pixel is determined to be a non-sky region and marked as 0 at the corresponding position in the mask matrix. S224. Repeat steps S222-S223 until all pixels of the HSV cloud map have been traversed, forming a binary mask matrix.

5. The method for ultra-short-term solar irradiance prediction based on ground-based cloud maps according to claim 4, characterized in that, The contrast enhancement processing of the target sky region image to obtain the cloud image pixel matrix with enhanced cloud features includes: The luminance channel is separated from the HSV cloud map, and the luminance contrast is enhanced by the adaptive histogram equalization method. After merging the contrast-enhanced luminance channel with the other channels, the image is converted back to the original cloud image to obtain a cloud image pixel matrix with enhanced cloud features.

6. The method for ultra-short-term solar irradiance prediction based on ground-based cloud maps according to claim 1, characterized in that, The expression for adjusting the preset constraint rules in the cloud map similarity matrix according to the preset constraint rules is as follows: ； In the formula, Let represent the similarity matrix of cloud maps on day i and day j at time t; α and β represent the enhancement coefficient and the reduction coefficient, respectively.

7. The method for ultra-short-term solar irradiance prediction based on ground-based cloud maps according to claim 6, characterized in that, The comprehensive similarity calculation model, after training, outputs a comprehensive similarity score of the irradiance view training samples. Based on the comprehensive similarity score, the ultra-short-term prediction results of solar irradiance at future times are calculated, including: The comprehensive similarity score of the irradiance view training samples is calculated by the trained comprehensive similarity calculation model. The comprehensive similarity scores are sorted in descending order, and the top M groups of irradiance view training samples that meet the preset training sample size threshold are selected. Based on the comprehensive similarity scores of the training samples of the first M groups of irradiance views, a weight matrix is ​​constructed using a comprehensive similarity numerical weighting method; Extract the irradiance values ​​from time 1 to time T after the current time from the training samples of the first M groups of irradiance views, and obtain a new irradiance value sequence through weighted calculation, which serves as the ultra-short-term prediction value of solar irradiance at the current time for future times.

8. The method for ultra-short-term solar irradiance prediction based on ground-based cloud maps according to claim 7, characterized in that, The expression for the comprehensive similarity calculation model is: ； In the formula, This represents the final comprehensive similarity matrix; The weights represent the irradiance similarity matrix; This represents the final irradiance similarity matrix; This represents the final similarity matrix of the cloud map.

Images