Device and method for determining at least one component of hemispherical irradiance of solar radiation in an arbitrary plane

Abstract

The invention relates to a device and a method for determining at least one component of a hemispherical irradiance (120) of solar radiation in an arbitrary plane, wherein the component comprises a diffuse irradiance (122) and / or a direct irradiance (124), comprising: determining measurement data of the hemispherical irradiance (120) with a radiation sensor (12) in a field of view (26) above a plane (46) of the radiation sensor (12); capturing an image (110) of the sky (50) with a camera (14) in a field of view (32) above a plane (48) of the camera (14); extracting features from the image (110) of the sky (50) using a first machine learning model and generating a result data set; combining the measurement data of the radiation sensor (12) and the result data set of the camera (14) into a common data set;Determining at least one component of the hemispheric irradiance (120) in the arbitrary plane, in particular the diffuse irradiance (122) and / or the direct irradiance (124), from the data set using a second machine learning model.

Description

State of the art

[0001] The invention relates to a method and a device for determining at least one component of a hemispheric irradiance of solar radiation in any plane. Furthermore, the invention relates to a method for training a first machine learning model for extracting features, in particular structures, from a camera image of the sky, as well as a computer program for determining at least one component of a hemispheric irradiance of solar radiation in any plane, a data processing device, a trained first machine learning model for extracting features, in particular structures, from an image of the sky, and a trained second machine learning model for determining at least one component of a hemispheric irradiance of solar radiation in any plane.

[0002] Large-scale solar energy systems, such as solar power plants and photovoltaic systems, which are installed or planned at different latitudes around the world, are becoming increasingly important for energy supply. The economic viability of operating the solar energy system is becoming a key consideration in the decision for or against its installation. Due to varying solar radiation levels at different latitudes, the location and efficiency of the solar energy system are of great importance for its economical operation.

[0003] To plan the location for the installation and economical operation of the photovoltaic system, for example, the irradiance of the solar radiation at the respective location must be known or it must be possible to measure it. The irradiance of solar radiation, also referred to as solar irradiance, is the intensity and thus the light output of the solar radiation that hits one square meter of a solar panel of the photovoltaic system to be installed. In order to optimize the solar radiation on the solar panels and the efficiency as best as possible, the solar panels are often tracked in order to change the azimuth and the angle of inclination. In bifacial photovoltaic systems, the solar radiation on the front of the solar panel as well as the back of the solar panel is used.Since solar radiation has individual components, such as diffuse irradiance in the horizontal plane (DHI), ground-reflected radiation, and direct normal irradiance (DNI), all of which contribute differently to the overall performance of a solar system, it is necessary to have access to the individual components of solar irradiance in addition to the hemispheric solar irradiance. It is of interest to measure and know the hemispheric irradiance and its components at any desired plane.

[0004] It is known that an irradiance sensor, typically a pyranometer, is used to measure solar irradiance. Measurements are always taken within a field of view of a half-space above the installed plane of the irradiance sensor, typically the pyranometer. The half-space describes a hemisphere, corresponding to a solid angle of at least approximately 2π steradians. It is also known to measure the degree of sky coverage using a surveillance camera or cloud camera. However, determining diffuse horizontal radiation and direct normal radiation requires a significant amount of measurement.

[0005] For example, a sun-tracking pyrheliometer and a sun-tracking pyranometer can be used. The measured data are analyzed using transposition and decomposition models, which, however, are subject to significant errors.

[0006] WO 2021 / 219570 A1 discloses a measuring device comprising an irradiance sensor, for example, a pyranometer, and a camera. Using the measuring device, a method can be used to determine the hemispheric solar irradiance at any plane by analyzing the measured data. Furthermore, the measuring device and method can also be used to determine the radiance distribution of the sky and the resulting proportions of diffuse radiation. Disclosure of the invention

[0007] The object of the invention is to provide an improved method and an improved device for determining at least one component of a hemispheric irradiance of solar radiation in any plane.

[0008] Furthermore, it is the object of the invention to provide an improved method for training a first machine learning model for extracting features, in particular structures, from the image of the sky.

[0009] Furthermore, it is the object of the invention to provide an improved computer program for determining at least one component of a hemispheric irradiance of solar radiation in any plane.

[0010] Furthermore, it is the object of the invention to provide a data processing device for a device for determining at least one component of a hemispherical irradiance of solar radiation in any plane.

[0011] Furthermore, it is the object of the invention to provide an improved trained first machine learning model for extracting features, in particular structures, from the image of the sky and an improved trained second machine learning model for determining at least one component of a hemispheric irradiance of solar radiation in any plane.

[0012] The objects are achieved by the features of the independent claims. Advantageous embodiments and advantages of the invention emerge from the further claims, the description, and the drawings.

[0013] The features listed individually in the patent claims can be combined with one another in a technologically meaningful manner and can be supplemented by explanatory facts from the description and by details from the figures, whereby further embodiments of the invention are shown.

[0014] According to a first aspect of the invention, a method is proposed for determining at least one component of a hemispheric irradiance of solar radiation in any plane, wherein the at least one component comprises a diffuse irradiance and / or a direct irradiance. The method comprises the following steps: (i) obtaining measurement data of the hemispherical irradiance with a radiation sensor in a field of view above a plane of the radiation sensor; (ii) capturing an image of the sky with a camera in a field of view above a plane of the camera; (iii) extracting features, in particular structures, from the image of the sky using a first machine learning model and generating a result data set; (iv) merging the measurement data from the radiation sensor and the result data set from the camera into a common data set; (v) determining the at least one component of the hemispheric irradiance in the arbitrary plane, in particular the diffuse irradiance and / or the direct irradiance, from the data set by means of a second machine learning model.

[0015] Any plane shall be understood to be a horizontal plane or an inclined plane.

[0016] The irradiance sensor and the camera form a measuring unit, which forms the basis of the proposed method. The measuring unit can thus provide a first measurement data set from the irradiance sensor, which comprises the hemispherical irradiance measured in a field of view of a half-space above the irradiance sensor. The half-space can be a hemisphere and / or hemisphere corresponding to a solid angle of at least approximately 2π steradians. The measuring device can provide a second measurement data set, which comprises measurement data measured by the camera. The measurement data set from the camera comprises an image of the sky, which depicts the sky in a field of view of a half-space above the camera. Thus, the method is based on two measurement data sets from two different instruments: the irradiance sensor and the camera. The irradiance sensor can be a pyranometer, for example.The camera can be a cloud camera or a surveillance camera. The advantage of this is that it allows for improved accuracy in determining one of the components of hemispheric irradiance.

[0017] In particular, measurement inaccuracies and / or errors of the irradiance sensor or camera can be compensated for. The reason for this is that the camera generally cannot measure irradiance as accurately as the irradiance sensor. Thus, accuracy is improved compared to the state of the art, which typically uses only one camera.

[0018] The image of the sky typically contains the sun and structures around the sun, such as cloud cover and arrangement, shading of the sun, blurring by cirrus clouds, and / or image artifacts caused by the influence of direct radiation on the camera image. Thus, the image of the sky contains structures and textures. These structures and textures provide the image features extracted by the first machine learning model. In particular, the image features around the sun can be used. The first machine learning model can extract the image features from the image of the sky. This allows effects such as lens refraction and image saturation to be detected. A type of spectral correction can be applied. These image features can be used in the subsequent analysis of the combined dataset using the second machine learning model.

[0019] The image of the sky can contain the structures and textures, each as angle-resolved information, due to the information stored pixel by pixel, particularly the measured intensities. The structures and textures can be used as image features for analyzing the measurement data by the first machine learning model and form the resulting dataset. The first machine learning model can, for example, be a so-called convolutional neural network (CNN).

[0020] The CNN algorithm is a widely used algorithm in a wide range of applications and, in particular, provides reliable iteration results.

[0021] In contrast to the analysis of the sky image with the first machine learning model, the proposed method uses a physical model based on physical assumptions and corrections. The first machine learning model, on the other hand, can directly analyze the image features of the structures and textures present in the sky image. This can have the advantage of preventing any information loss.

[0022] The second machine-learning model can correct the measurement errors of a camera-based radiation measurement, as an independent radiation measurement is performed using the irradiance sensor and additional image features are utilized. The second machine-learning model determines the contribution of diffuse radiation to the hemispheric irradiance based on the diffuse irradiance in the horizontal plane (DHI) and / or the direct irradiance (DNI). Thus, the second machine-learning model in the present method assumes the role of a decomposition model of the prior art method.

[0023] Diffuse irradiance can be determined in the horizontal plane and / or an inclined plane. Direct irradiance can be determined in the horizontal plane and / or an inclined plane. The corrections applied in the horizontal plane can be used preferentially for the inclined plane.

[0024] According to a favorable embodiment of the method, the second machine learning model can be different from the first machine learning model, wherein the first machine learning model is trained with first training data sets, and wherein the second machine learning model is trained with second training data sets. The first training data sets can preferably contain reference data. The second machine learning model typically comprises a different algorithm than the first machine learning model. The second machine learning model can have a so-called multilayer perception algorithm (MLP: multilayer perceptron). The data set underlying the second machine learning model includes both the measured value of the hemispheric irradiance of the irradiation sensor and the result data set of the image of the sky analyzed with the first machine learning model. The data set is also referred to as a data vector or vector data set.

[0025] The second machine-learning model estimates the diffuse irradiance in the horizontal plane (DHI) and / or the direct irradiance (DNI) from the shared dataset, in particular the data vector, depending on the second training data used. Typically, the second machine-learning model can be trained for the diffuse irradiance in the horizontal plane (DHI). The hemispheric horizontal irradiance (GHI) and the diffuse irradiance in the horizontal plane (DHI) and the direct irradiance (DNI) are related by a known equation: GHI - DHI = DNI x sin (sun elevation angle). Here, the sun elevation angle is known at any time of day. Thus, the DNI, for example, can be determined from the DHI extracted from the sky image using the second machine-learning model and the measured GHI.

[0026] With the second machine learning model, in addition to the measurement of the hemispheric irradiance with the irradiance sensor and the image features from the structure and texture of the sky, the following features can also be used to estimate the diffuse irradiance in the horizontal plane (DHI) and / or the direct irradiance (DNI): solar elevation, number of saturated pixels in the image captured by the camera, DHI measurement from the classical analysis using a physical model, DNI measurement from the classical analysis using a physical model, irradiance in the horizontal plane from an angular range of, for example, 25° around the sun, intermediate results from the classical analysis using a physical model, in particular a DHI measurement per color channel of the camera and ratios of these DHI measurements per color channel.The advantage here is that measurement errors caused, for example, by lens errors, overexposure and / or lens refraction effects can be detected.

[0027] The first training data sets for training the first machine learning model can advantageously be extended with the additional features listed above.

[0028] According to a further advantageous embodiment of the method, features can be extracted in an image analysis using the camera image. In particular, features can optionally also be extracted from the measurement data of the radiation sensor unit. The image analysis can use information contained in the image pixels. For example, features can be calculated from the intensities of the image pixels. These can be, for example, a number of saturated image pixels, weighted sums of the intensities of the image pixels across the entire image and / or across the image area around the sun, and ratios of these sums.

[0029] According to a further advantageous embodiment of the method, features can be derived from the intensities of the image pixels in the image analysis using the camera image. In particular, a physical camera model can be applied. Alternatively or additionally, a number of saturated pixels and / or at least one sum and / or weighted sum of the intensities of the image pixels and / or at least one ratio of sums and / or weighted sums of the intensities of the image pixels can be included for feature extraction.In this way, advantageous features can be obtained, particularly for the second machine learning model: solar elevation, number of saturated pixels in the image captured by the camera, DHI measurement from the classical analysis using a physical model, DNI measurement from the classical analysis using a physical model, irradiance in the horizontal plane from an angular range of, for example, 25° around the sun, intermediate results from the classical analysis using a physical model, in particular a DHI measurement per color channel of the camera and ratios of these DHI measurements per color channel.

[0030] According to a favorable embodiment of the method, the features extracted from the camera image and / or from the radiation sensor measurement data can be passed to a third machine learning model and processed using the third machine learning model, and a radiation sensor dataset can be generated. The radiation sensor dataset and the camera result dataset can be merged before the second machine learning model is applied. The second machine learning model can be applied to a joint dataset consisting of features extracted from the irradiation sensor measurement data using the third machine learning model and the camera measurement data analyzed using the first machine learning model. The joint dataset can also be referred to as a data vector or vector dataset.

[0031] In a favorable design of the method, an image transformation of the sky image can be performed before applying the first machine-learning model. This has the advantage that only image regions with high information content regarding direct radiation are transmitted to the first machine-learning model. This is important when resources are limited.

[0032] A raw image of the sky captured by the camera is subjected to an image transformation. The raw image contains the camera's measurement data prior to any analysis. This allows a transformed image of the sky to be generated. This has the advantage that only image regions with high information content, particularly regarding direct radiation, are transmitted to the first machine learning model. This is important when resources are limited. In addition, errors in diffuse radiation measurements can also be passed on to the first machine learning model. One or more of the following steps can be performed as an image transformation: (i) Applying image rectification and image rotation and image distortion according to a camera model such that the sun's disk is always located in the image center. This image transformation can simulate camera tracking.(ii) Crop the image to a square area around the sun, encompassing angular intervals of approximately 25° around the sun. (iii) Downscale the resolution of the camera image. For example, a size of 224x224 pixels can be used. This is a common size and has the advantage of allowing the use of known pre-trained initial machine learning models from other applications.

[0033] According to a favorable embodiment of the method, the first machine learning model and / or the second machine learning model can be trained with reference data of at least one of the components of the hemispheric irradiance, in particular the diffuse irradiance in the horizontal plane (DHI), in particular from solar trackers and / or the direct irradiance (DNI), in particular from a pyrheliometer tracking the sun, and / or with regular input data, in particular with images of the camera, data of the hemispheric irradiance and / or intermediate results.

[0034] Intermediate results can be analysis results of measurement data recorded with the measuring device consisting of the irradiation sensor and the camera, whereby the measurement data are then evaluated using a physical model as described, for example, in WO 2021 / 219570 A1.

[0035] According to a favorable embodiment of the method, the first machine learning model can be pre-trained using at least one of the following steps: (i) Use of publicly available weights; (ii) training using at least one unsupervised or self-supervised training approach, for example using a DeepCluster v2 method according to CARON, Mathilde et al., in particular with a number of k=30 clusters / groups; (iii) augmenting data with Gaussian blurs and / or distorting image colors and / or mirroring images; (iv) Training with images of the sky taken by the camera.

[0036] The number k=30 is intended as an example. Fewer clusters / groups can also be used, for example, k=25. More clusters / groups can also be used, for example, k=35 or k=40. The number of clusters / groups should be understood as indicating the order of magnitude of the clusters / groups.

[0037] The advantage, especially of step (iii), may be the acquisition of additional training data.

[0038] Pre-training the first machine learning model typically means training the first machine learning model with known datasets. The known datasets preferably contain structures and textures that are similar or identical to those in the sky images measured with the camera.

[0039] According to a favorable embodiment of the method, a so-called convolutional neural algorithm, in particular a convolutional neural network (CNN), can be used in the first machine learning model. Preferably, image features that occur around the sun's surface can be extracted. For example, artifacts such as rays and / or circles around the sun's disk in the image of the sky can indicate the occurrence of lens refraction effects.

[0040] According to a favorable embodiment of the method, at least one algorithm from a multilayer perceptron (MLP), random forest algorithm, recurrent neural network (RNN), LSTM long-short-term memory (LSTM) algorithm, transformer model algorithm, k-nearest neighbor (k-NN) algorithm, and / or support vector machine (SVM) can be used in the second machine learning model.

[0041] With a favorable design of the method, the first training data sets and / or the second training data sets can be filtered so that different atmospheric conditions are represented to a similar extent. The atmospheric conditions can be, for example, cloud cover, cirrus clouds, or partial cloud cover.

[0042] One-minute averages can advantageously be used as the time resolution for processing measurement data according to the method.

[0043] According to a favorable embodiment of the method, the first machine learning model, the second machine learning model, and / or the third machine learning model can be trained in a supervised manner, using input data and / or correct reference data from at least one of the components of the hemispheric irradiance. The at least one component can be diffuse irradiance in the horizontal plane (DHI) or direct irradiance (DNI).

[0044] According to a favorable design of the method, the first machine learning model, the second machine learning model, and the third machine learning model can also be trained jointly, particularly under supervision. This is referred to as fine-tuning. Joint training can take place, in particular, after the individual training of the first machine learning model, the second machine learning model, and the third machine learning model.

[0045] It may also be planned to train the first machine learning model unsupervised. This can typically be done using existing images from a cloud camera. For example, unsupervised training can use only input data and no reference data.

[0046] According to a favorable embodiment of the method, image features from the sky image can be used as additional input data for the third machine learning model or the merging process using estimated values ​​of a red channel, a blue channel, and / or a green channel from the measurement data of the sky image. Advantageously, intensity values ​​of the red channel, the blue channel, and / or the green channel can be used.

[0047] According to a further aspect of the invention, a device is proposed for carrying out a method for determining at least one component of a hemispheric irradiance of solar radiation in any plane, wherein the component comprises a diffuse irradiance and / or a direct irradiance. The device comprises at least one radiation sensor unit, a camera, and an evaluation unit which is provided for evaluating measurement data from the radiation sensor unit and the camera. The radiation sensor unit determines the irradiance of solar radiation in a field of view of 180° above a plane, wherein the plane is a horizontal plane or an inclined plane. The camera captures measurement data comprising an image of the sky in a field of view of 180° above a plane, wherein the plane is a horizontal plane or an inclined plane.The evaluation unit has at least a first machine learning model and a second machine learning model. The evaluation unit is configured to perform at least the following steps: (i) extracting features, in particular structures, from the image of the sky using a first machine learning model and generating a result data set; (ii) merging the measurement data from the radiation sensor and the result data set from the camera into a common data set; (iii) determining the at least one component of the hemispheric irradiance in the arbitrary plane, in particular the diffuse irradiance and / or the direct irradiance, from the data set by means of a second machine learning model.

[0048] The image of the sky includes an image of the sky within a field of view of up to 180° above the camera. The camera can be, in particular, a surveillance camera or a cloud camera. The image of the sky can also be referred to as a sky image.

[0049] Based on the analysis with the first machine learning model, one or more components of the hemispheric irradiance can be determined from the hemispheric irradiance measured with the irradiance sensor and the sky image from the camera. In particular, a proportion of the respective components can be determined.

[0050] The at least one component can be determined in any plane. The advantage is that the determined components each exhibit a high degree of accuracy. This is particularly true compared to prior art methods in which the hemispheric irradiance used is determined solely with the camera.

[0051] According to a favorable design of the device, the camera's measurement data can contain angle-resolved radiation information. The structures and textures in the sky image can each be available as angle-resolved information. The structures and textures can be used as image features for analyzing the measurement data by the camera's first machine learning model, forming the resulting dataset.

[0052] According to a further aspect of the invention, a method for training a first machine learning model for extracting features, in particular structures, from an image of the sky is proposed, comprising the following steps: (i) Determining or using data sets of images of the sky taken with a camera, in particular a cloud camera (ii) training the first machine learning model with the data sets of images of the sky or (iii) Training the first machine learning model with known images of the sky taken at other locations.

[0053] In this way, the first machine learning model can be efficiently trained and used to extract features, especially structures, from an image of the sky.

[0054] According to a favorable design of the method, pre-training with at least one of the following steps can be used to train the first machine learning model: (i) Use of publicly available weights; (ii) training using at least one unsupervised or self-supervised training approach; (iii) augmenting data with Gaussian blurs and / or distorting image colors and / or mirroring the images; (iv) Training with images of the sky taken by the camera.

[0055] When training using at least one unsupervised or self-supervised training approach, DeepClusterv2, a method according to CARON, Mathilde et al., can be used, in particular with a number of k=30 clusters / groups.

[0056] In this way, the first machine learning model can be efficiently trained and used to extract features, especially structures, from one of the images of the sky.

[0057] According to a further aspect of the invention, a computer program for determining at least one component of a hemispheric irradiance of solar radiation in any plane is proposed, which comprises instructions which, when the program is executed by a computer, cause the computer to carry out the steps of a method for determining at least one component of a hemispheric irradiance of solar radiation in any plane.

[0058] The computer program can advantageously be used in a computer of a data processing device of a device for determining at least one component of a hemispheric irradiance of solar radiation in any plane.

[0059] According to a further aspect of the invention, a data processing device for a device for determining at least one component of a hemispheric irradiance of solar radiation is proposed, which comprises at least one measurement data acquisition unit, an evaluation unit and a computer.

[0060] The data processing device can advantageously be used in a device for determining at least one component of a hemispherical irradiance of solar radiation in any plane.

[0061] According to a further aspect of the invention, a trained first machine learning model is proposed for a method for determining at least one component of a hemispheric irradiance of solar radiation in any plane. The trained first machine learning model serves to extract features, in particular structures, from an image of the sky and is trained according to a method for training the first machine learning model. The trained machine learning model can advantageously have been trained with first training data sets that contain image features relevant to an image of the sky.

[0062] The trained first machine learning model can advantageously be used in a device for determining at least one component of a hemispheric irradiance of solar radiation in any plane.

[0063] According to a further aspect of the invention, a trained second machine learning model is proposed for a method for determining at least one component of a hemispheric irradiance of solar radiation in any plane. The trained second machine learning model serves to determine at least one component of a hemispheric irradiance of solar radiation in any plane. The trained second machine learning model can be trained in a supervised manner, wherein, in particular, input data and / or reference data can be used. The trained second machine learning model in particular comprises a multilayer perceptron algorithm. The trained second machine learning model can advantageously have been trained using second training data sets. The second training data sets can comprise image features.

[0064] The trained second machine learning model can advantageously be used in a device for determining at least one component of a hemispheric irradiance of solar radiation in any plane. drawing

[0065] Further advantages will become apparent from the following description of the drawings. The figures illustrate exemplary embodiments of the invention. The figures, the description, and the claims contain numerous features in combination. Those skilled in the art will also expediently consider the features individually and combine them into useful further combinations.

[0066] Examples include: Fig. 1 shows a schematic structure of a device for determining at least one component of a hemispherical irradiance of solar radiation in any plane in a plan view; Fig. 2 the structure of the device according to Fig. 1 according to an embodiment of the invention in a side view; Fig. 3 the structure of the device according to Fig. 1 in an alternative setup in a side view, wherein the planes of the radiation sensor unit and the camera are inclined relative to the horizontal plane; Fig. 4 a schematic representation of an image of the sky taken with a camera; Fig. 5 is a flowchart of an embodiment of the method for determining at least one component of a hemispheric irradiance of solar radiation; Fig. 6 is a flowchart of a further embodiment of the method for determining at least one component of a hemispheric irradiance of solar radiation; Fig. 7 is a flowchart of a further embodiment of the method for determining at least one component of a hemispheric irradiance of solar radiation; Fig. 8 is a flowchart of a further embodiment of the method for determining at least one component of a hemispheric irradiance of solar radiation; Fig. 9 is a flowchart of another embodiment of the method for determining at least one component of a hemispheric irradiance of solar radiation; Fig. 10 a flow chart of the procedure according to Fig. 9 in a further embodiment; Fig. 11 parameters for training the first machine learning model. Embodiments of the invention

[0067] In the figures, components of the same type or function similarly are designated by the same reference numerals. The figures are merely examples and are not to be construed as limiting.

[0068] Before describing the invention in detail, it should be noted that it is not limited to the specific components of the device and the specific method steps, as these components and methods may vary. The terms used herein are intended only to describe particular embodiments and are not intended to be limiting. Furthermore, when the singular or indefinite articles are used in the description or claims, this also refers to the plural of these elements, unless the overall context clearly indicates otherwise.

[0069] The directional terminology used below, including terms such as “left”, “right”, “top”, “bottom”, “before”, “behind”, “after” and the like, is intended solely to improve understanding of the figures and is in no way intended to limit the scope of the illustrations.

[0070] The components and elements shown, their design and use may vary according to the considerations of a specialist and may be adapted to the respective applications.

[0071] Fig. Figure 1 shows a schematic structure of the device 5 for determining a hemispherical irradiance of solar radiation in any plane and its components: direct irradiance in a horizontal plane (GHI) or in an inclined plane (GTI), diffuse irradiance in the horizontal plane (DHI), direct irradiance (DNI), and irradiance reflected from the ground, in a plan view. The device 5 has a measuring unit 10, which has a radiation sensor unit 12 and a camera 14. Furthermore, a data processing device 20 with a measurement data acquisition unit 16 and an evaluation unit 18 is provided. The evaluation unit 18 is provided for evaluating measurement data from the radiation sensor unit 12 and / or the camera 14.The data processing device 20 with a control and regulation unit, a measurement data acquisition unit 16, and an evaluation unit 18 can, for example, realize synchronous measurement data acquisition of the radiation sensor unit 12 and the camera 14. The measurement data acquisition unit 16 is configured to acquire the measurement data of the radiation sensor unit 12 and the camera 14. The evaluation unit 18 is configured to evaluate the measurement data of the radiation sensor unit 12 and the camera 14. The evaluation unit 18 can have at least one first machine learning model, preferably a second and third machine learning model. The data processing device 20 or the evaluation unit 18 has a computer 19. The at least one first machine learning model is configured to analyze the measurement data of the camera 14.

[0072] Typically, the first trained machine learning model is implemented in the evaluation unit 18. It may be provided to train the first machine learning model in or with the evaluation unit 18. The particular configuration depends on the size and structure of the data processing device 20. The data processing device 20 may be a computer with a small storage capacity and low computing power. In this case, the training of the first machine learning model can take place on a server with a large storage capacity and high computing power, and the trained first machine learning model can only be stored on the data processing device 20. If necessary, the evaluation of the raw data during measurement operation can also take place on a server in order to use even fewer resources on the computer.

[0073] The radiation sensor unit 12 and the camera 14 can advantageously be arranged in close proximity to each other, horizontally leveled at the same height. In addition, the installation location can be expediently selected so that further obstacles in the fields of view 26, 32 (the fields of view 26, 32 are in the Fig. 2 and Fig. 3) of the radiation sensor unit 12 and the camera 14 can be avoided.

[0074] The irradiation sensor unit 12 and the camera 14 can be arranged in the horizontal plane 36 according to an embodiment of the invention, which in Fig. 2. In another embodiment, which is shown in Fig. 3, the radiation sensor unit 12 and / or the camera 14 are each arranged in a plane 46 and 48 inclined to the horizontal plane.

[0075] The radiation sensor unit 12 comprises a housing 22 and a sensor 24. The sensor 24 detects radiation in an upper half-space with a field of view 26 of 180°.

[0076] The radiation sensor unit 12 measures the irradiance of solar radiation in an upper half-space in a field of view 24 of 180° above the plane 38 in which it is arranged.

[0077] The half-space refers to a hemisphere, also called a hemisphere, which covers a solid angle of at least approximately 2π steradians.

[0078] The radiation sensor unit 12 can be configured to detect solar radiation in a wavelength range from 0.3 µm to 3 µm. In particular, the sensor 24 can detect light in the wavelength range from 0.3 µm to 3 µm.

[0079] The radiation sensor unit 12 is designed, for example, as a pyranometer, in particular a thermopile pyranometer. The radiation sensor unit 12 can alternatively be designed as a photodiode or photovoltaic reference cell if lower requirements are placed on the accuracy of determining the hemispheric irradiance.

[0080] The camera 14 is designed, for example, as a fisheye surveillance camera, in particular, for example, as a Mobotix Q25 surveillance camera or as a cloud camera. The camera 14 comprises a housing 28, above which a sensor 30 is arranged to capture the upper half-space with a field of view 32 of 180°. The camera 14 is preferably designed to record the sky in the field of view 32. The camera 14 generates an image 110 of the sky 50 (this is shown in Fig. 4). For example, image 110 of sky 50 contains information about clouds, the solar disk, and areas around the solar disk.

[0081] The camera 14 is advantageously designed so that it can capture the entire field of view 32 in a single shot without the need for mechanically moving parts.

[0082] The radiation sensor unit 12 and the camera 14 are arranged at a distance 34, each measured from a center point of the sensors 24 and 30. The distance 34 is preferably less than 3 m. The minimum distance 34 is determined by the diameters of the respective housings 22 and 28.

[0083] Fig. Figure 2 shows the device 5 in a schematic representation in a side view. The horizontal plane on which the radiation sensor unit 12 and the camera 14 are arranged is designated by the reference numeral 36. The sensor 24 of the radiation sensor unit 12 is arranged in or above a plane 38. The sensor 30 of the camera 14 is arranged in or above a plane 40. The planes 38 and 40 extend in the plane shown in Fig. 2 illustrated embodiment parallel to the horizontal plane 36.

[0084] In Fig. 3 is the structure of the device 5 according to Fig. 1 is shown in an alternative configuration, in which the plane of the radiation sensor unit 12 and the plane of the camera 14 are inclined relative to the horizontal plane 36. The camera 14 is arranged in plane 48. The radiation sensor 12 is arranged in plane 46.

[0085] The device 5 comprises the radiation sensor unit 12, the camera 14, as well as the data processing device 20 with the measurement data acquisition unit 16 and the evaluation unit 18 (in Fig. 1), which is provided for evaluating measurement data from the radiation sensor unit 12 and / or the camera 14, and the computer 19. The radiation sensor unit 12 is provided for determining the irradiance of solar radiation in a field of view 16 of 180°, i.e. the half-space above the plane 38, 46, and the camera 14 is also provided for detecting a field of view 26 of 180°, i.e. the half-space above the plane 40, 48.

[0086] At the Fig. In the setup of the device 5 shown in Figure 1, the two planes 38, 40 are aligned in the horizontal plane 36 and coincide with the horizontal plane 36.

[0087] At the Fig. In the setup of the device 5 shown in Figure 3, the planes 46, 48 of the radiation sensor unit 12 and the camera 14 are each inclined at an angle 42 and an angle 44 relative to the horizontal plane 36. The angle 44 can be adjusted between 0° and 90°.

[0088] The radiation sensor unit 12 and the camera 14 are arranged in this embodiment, for example, on a north-south axis, with the radiation sensor unit 12 being located in the northern hemisphere north of the camera 14, as in Fig. 2. In the southern hemisphere, the radiation sensor unit 12 would be located south of the camera 14.

[0089] In the advantageous embodiment shown here, the sensor 24 of the radiation sensor unit 12 and the sensor 30 of the camera 14 are each arranged in the horizontal plane 36 such that the field of view 26, 32 of the two sensors 24, 30 is each above the horizontal plane 36 and ends with the horizontal plane 36.

[0090] The distance 34 between the radiation sensor unit 12 and the camera 14 can be adjusted such that the sensor 24 of the radiation sensor unit 12 is visible in the field of view 32 of the camera 14 with an elevation 43 of at most 10°.

[0091] The radiation sensor unit 12 and the camera 14 are coupled in such a way that the measurement data are recorded by the radiation sensor unit 12 and the camera 14 in a synchronized manner. The measurement data can be conveniently evaluated in the evaluation unit 18.

[0092] The measurement data recording of the radiation sensor unit 12 can advantageously be carried out with a high temporal resolution, in particular with a temporal resolution in the second range, in particular of less than 10 s, preferably less than 5 s, particularly preferably less than or equal to 1 s.

[0093] The camera 14 can be expediently configured such that a single image is captured at a fixed time interval, in particular every half and full minute. Advantageously, the at least one sensor 30 of the camera 14 can have a constant color temperature. The camera 14 can advantageously have a constant exposure time for each individual image. A predetermined minimum value of an average image brightness can be set for the exposure control of the camera 14, with the exposure time remaining unchanged at higher image brightness.

[0094] In particular, the predetermined minimum value of an average image brightness can preferably be at most 10%, particularly preferably at most 8%, very particularly preferably at least 5%.

[0095] The image from camera 14 provides in real time the image 110 of the sky 50, which is Fig. 4. Image 110 of sky 50 contains raw sky data.

[0096] Shown is a typical camera image, specifically a fisheye lens with a 180° field of view in the half-space above the horizontal. Near the zenith, the bright solar disk 54 can be seen, while clouds 52 can be seen at the edge of the image, i.e., closer to the horizon.

[0097] The image of the sky can advantageously contain the following information: - Clouds and cloud cover, - Thin cirrus clouds, - solar disk, shaded solar disk, - lens refraction effects around the solar disk, - Shape of the solar disk, cloudiness of the solar disk, - Clarity of the edge of the solar disk, - Decrease in brightness in the vicinity of the solar disk.

[0098] This information is referred to as image features. If a first machine learning model is applied to image 110 of sky 50, these image features can be recognized and evaluated.

[0099] The image 110 of the sky 50 can be recorded as a digital image. The digital image 110, also referred to as the sky image 110, has a matrix of pixels, referred to as a pixel matrix, wherein the information is contained as the intensity of each pixel, referred to as the pixel intensity. Each pixel has a red channel, a green channel, and a blue channel. Each of the channels thus has its own pixel intensity. Thus, the information of the image is contained in the pixel intensities of the red channel, the green channel, and the blue channel and in the pixel position in the pixel matrix. The pixel position contains the information about the location or position in the sky assigned to the respective pixel. Thus, the pixel intensity of the pixel matrix contains radiation information. The image of the sky contains angle-resolved radiation information via the pixel position and the pixel intensity of the pixel in the pixel matrix.

[0100] In the Fig. 5 to 9 describe methods S100, S200, S300, S400, S500 that use machine learning models to analyze images 110 from camera 14 and measurement data from radiation sensor 12. Reference numeral 122 denotes the diffuse irradiance in the horizontal plane extracted using one of methods S100, S200, S300, S400, S500.

[0101] Reference numeral 124 denotes the direct irradiance extracted using one of the methods S100, S200, S300, S400, S500. The diffuse irradiance in the horizontal plane and the direct irradiance, which were determined or measured by other means, are used without reference numerals.

[0102] In Fig. Figure 5 shows an embodiment of a method S100 for analyzing the information recorded by the device 5. The method steps S130 to S150 are listed below: In method step S130, an image 110 captured with the camera 14 is analyzed with a first machine learning model by extracting image features, in particular structures and textures, from the image 110.

[0103] In method step S140, the image features and a value of the hemispherical irradiance 120 recorded with the radiation sensor unit 12 are concatenated and a common data set, in particular a data vector, which is also referred to as a vector data set, is formed.

[0104] In process step S150, the joint data set is analyzed with a second machine learning model and values ​​for the diffuse irradiance 122 in the horizontal plane (DHI) and the direct irradiance (DNI) 124 are extracted.

[0105] The first machine learning model is preferably a pre-trained machine learning model. The first machine learning model can, for example, be a so-called convolutional neural network (CNN). The CNN is a so-called deep learning algorithm that is primarily used for analyzing images and videos. It extracts visual features—image features—from images.

[0106] The image features are present in the images as structures and / or textures that form a pattern. The CNN is typically trained with a large amount of data, especially images with structures and textures, and then applied to new images to analyze them, especially automatically.

[0107] The term neural network refers to a logical structure inspired by the human brain and can form the basis for deep learning algorithms. For example, the term "deep learning" refers to a multi-layered neural network that learns from large data sets and from image features contained in the data sets. The term "convolutional" refers to the application of convolutional layers with filters to the input data to learn the image features.

[0108] Typically, there are multiple convolutional layers connected by pooling layers. The first convolutional layers extract general or low-level features such as lines and edges, while the later layers learn finer details or high-level features such as the sun's disk. Pooling layers are used to reduce the size of the convolutional features and thus reduce computation time and cost.

[0109] Fully connected layers learn global patterns based on the high-level features output by the convolutional and pooling layers, generating global patterns, such as those for the solar disk or clouds. After the input data passes through the fully connected layer, the final layer activates the machine learning model, allowing the neural network to make its predictions.

[0110] The first machine learning model can be trained differently than the second machine learning model. For example, a multilayer perceptron (MLP) model can be used as the second machine learning model. Different training means that different data sets are used and the algorithm is different.

[0111] In Fig. 6 is an extended embodiment of the Fig. 5, which is referred to as method S200.

[0112] The method S200 comprises the following steps: In method step S230, the first machine learning model is applied to the camera image 110 and image features are extracted from the image 110 of the sky 50.

[0113] In method step S235, an image analysis is performed using the image 110 from the camera 14 and / or other information. Optionally, measured values ​​of the hemispheric irradiance 120 can also be used to extract features.

[0114] Advantageously, in the image analysis in step S235, using the image 110 from the camera 14, features can be derived from the intensities of the image pixels. In particular, a physical camera model can be applied. Alternatively or additionally, a number of saturated pixels and / or sums and / or weighted sums of the intensities of the image pixels and / or ratios of sums and / or weighted sums of the intensities of the image pixels can be included for feature extraction.

[0115] In this way, advantageous features can be obtained, particularly for the second machine learning model: solar elevation, number of saturated pixels in the image captured by the camera, DHI measurement from the classical analysis using a physical model, DNI measurement from the classical analysis using a physical model, irradiance in the horizontal plane from an angular range of, for example, 25° around the sun, intermediate results from the classical analysis using a physical model, in particular a DHI measurement per color channel of the camera and ratios of these DHI measurements per color channel.

[0116] When analyzing the camera image 110 in the image analysis in method step S235, the following features can advantageously be obtained: - solar elevation, - Number of saturated pixels in the image 110, - Measurement of diffuse irradiance in the horizontal plane using physical models or analyzed values ​​of diffuse irradiance from the analysis of the camera image using the physical model, - Determining the direct normal radiation using physical models or analyzed values ​​of the direct radiation from the analysis of the camera image using the physical model, - Measurement of irradiance in the horizontal plane from an angular range of 25° around the sun, using only direct radiation - Estimated illuminance (radiance) from the red channel, the green channel and the blue channel of image 110, - Ratio of illuminance estimates based on each channel.

[0117] The direct normal radiation as well as the diffuse irradiance is determined using physical models or analyzed values ​​from the analysis of the camera image using the physical model.

[0118] The features listed above are particularly suitable for determining the least one component of a hemispheric irradiance of solar radiation using the second machine learning model, as this allows for the determination of errors inherent in evaluations using physical models. This is because there is a close relationship between the features listed above and such a measurement of diffuse irradiance in the horizontal plane. For example, the irradiance in the horizontal plane from an angular range of 25° around the sun shows measurement errors that may have been caused by overexposure and / or lens refraction effects. Irradiance estimates based on the individual color channels (red channel, green channel, blue channel) and their ratios provide information on the spectral composition of the diffuse irradiance.The spectral composition has an influence on the measured values ​​of the diffuse irradiance 122, in particular the accuracy of the measurement of the intensity of the diffuse irradiance 122. If the spectral composition of the diffuse irradiance 122 is known, the measured values ​​of the diffuse irradiance 122 can be corrected accordingly.

[0119] In process step S240, the extracted data from process step S230 and process step S235, as well as the hemispheric irradiance 120, are concatenated. Features of the resulting data set, also referred to as a vector data set, contain data from S230, S235, and the hemispheric irradiance 120.

[0120] In method step S250, the second machine learning model is applied to the vector data set to extract the diffuse irradiance 122 in the horizontal plane and / or the direct irradiance 124.

[0121] Fig. 7 shows a further extended process S300 with the following process steps: In method step S315, an image transformation is carried out using the measured values ​​of the image 110 taken with the camera 14.

[0122] In method step S315, image distortion correction and / or image rotation are / is performed. The image distortion correction and / or image rotation is applied according to the camera model. As a result, the sun disk 54 is always positioned in the center of the image in camera image 110. This simulates a camera 14 tracking the sun 54. Furthermore, in method step S315, the image is cropped to a quasi-square area around the sun. This includes angular distances of up to 25° around the sun. Furthermore, the resolution of the camera image is scaled down, for example, to a format of 224x224 pixels. This makes it possible to use pre-trained machine learning models from other applications to analyze the image data from camera 14. This saves the time and costs required for pre-training the first machine learning model. Furthermore, larger amounts of data are available for use as training datasets.

[0123] In method step S330, the first machine learning model is applied to the camera image 110 to extract image features from the image 110 of the sky 50.

[0124] In method step S335, the camera image 110 is analyzed as described in method step S235 and optionally using the hemispheric irradiance 120 and / or other information.

[0125] In process step S340, the extracted data from process step S330 and process step S335, as well as the hemispheric irradiance 120, are concatenated, and a vector data set is generated. Features of the data set contain data from S330, S335, and the hemispheric irradiance 120.

[0126] In method step 350, the second machine learning model is applied to the vector data set to extract the diffuse irradiance 122 in the horizontal plane and / or the direct irradiance 124.

[0127] Fig. 8 shows a further embodiment of a method S400 for determining the diffuse irradiance 122 in the horizontal plane and / or the direct irradiance 124. The method S400 is similar to that in Fig. 7. However, no image analysis is performed as described in step S335. Image features are extracted solely using the first machine learning model in method step S430.

[0128] The S400 process comprises the following process steps: In method step S415, an image transformation is carried out which corresponds to the image transformation described in method step S315. In method step S430, the first machine learning model is applied to the camera image 110 after method step S315 to extract image features from the image 110 of the sky 50.

[0129] In process step S440, the extracted data from process step S430 and the hemispheric irradiance 120 are concatenated. Features of the data set contain data from S430 as well as values ​​of the hemispheric irradiance 120.

[0130] In method step S450, the second machine learning model is applied to the vector data set to determine the diffuse irradiance 122 in the horizontal plane and / or the direct irradiance 124.

[0131] Fig. Figure 9 shows a further embodiment of a method S500, which includes the following steps. In step S538, the method S300 is expanded by a third machine learning model.

[0132] In method step S515, an image transformation of the image 110 of the sun 54 is carried out, which corresponds to the image transformation described in method step S315.

[0133] In method step S530, the first machine learning model is applied to the camera image 110 after method step S515 to extract image features from the image 110 of the sky.

[0134] In method step S535, an image analysis of the data of the image 110 and optionally the hemispherical irradiance 120 recorded with the radiation sensor 12 is carried out similarly to the method step S235 described above.

[0135] In method step S538, the third machine learning model is used to extract features from the hemispheric irradiance 120 and from the features extracted in method step S535.

[0136] In method step S538, the third machine learning model is applied to the measurement data of the hemispheric irradiance 120, which was analyzed together with the information from the camera image 110 in method step S535. This results in the vector data set being less complex than the data set passed to the second machine learning model in methods S100, S200, S300, and S400. This reduces the training effort for the second machine learning model. At the same time, potential overfitting can be at least partially avoided.

[0137] In process step S540, the extracted data from process step S530 and process step S538 of the hemispheric irradiance 120 are concatenated, and a vector data set is generated. Features of the data set contain data from S530 and S538.

[0138] In method step S550, the second machine learning model is applied to the vector data set to determine the diffuse irradiance 122 in the horizontal plane and / or the direct irradiance 124.

[0139] It is also possible to combine the S100, S200, S300, S400, and S500 methods differently than shown in the figures. Furthermore, the components of each machine learning model can be weighted. This involves applying a special training procedure for the machine learning models.

[0140] The machine learning models can be trained differently, meaning different training strategies and / or different training datasets are used. Different algorithms are also used for the three machine learning models.

[0141] The training procedure of the first machine learning model can be supervised or unsupervised.

[0142] If the first machine learning model is trained supervised, it may include the steps and strategies listed below and use data that uses one or more datasets from the list below: - Data on diffuse irradiance in the horizontal plane, for example from solar trackers, - data from a shaded radiation sensor 12, in particular a pyranometer, - direct normal radiation of a sun-tracking pyrheliometer, - regular input data, for example images from a cloud camera, - regular input data, for example data of hemispheric irradiance, intermediate results of an evaluation with physical models, data from an irradiance sensor, in particular pyranometer.

[0143] When the first machine learning model is trained unsupervised, images 110 from camera 14, in particular from a cloud camera, are used. For example, existing images 110 from cloud camera 14 or existing images from any cloud camera can be used. In particular, image features in the region of the sun 54 are extracted.

[0144] Training can be carried out using at least one unsupervised or self-supervised training approach, for example using a DeepCluster v2 method according to CARON, Mathilde et al., in particular with a number of k=30 clusters / groups.

[0145] From the image features around the sun 54, or from the area of ​​the sun 54, effects such as lens refraction and image saturation can be detected and then taken into account in the analysis of the currently recorded images 110. Thus, a spectral correction of the recorded images 110 can be performed.

[0146] The second machine learning model can be trained in a supervised manner. Depending on the implementation of the method, the second machine learning model can, in particular, correct measurement errors from a cloud camera-based radiation measurement using provided image features and intermediate results from the cloud-based radiation measurement and the hemispheric irradiance 120. This utilizes data from both a hemispheric radiation measurement from the camera 14 and the radiation sensor 12. This leads to better and increased accuracy of the data for the diffuse irradiance 122 in the horizontal plane and the direct irradiance 124, which are obtained from the hemispheric irradiance 120 and the camera image 110.In method S100, the second machine learning model uses the image features and the hemispheric irradiance 120 to estimate the diffuse irradiance 122 in the horizontal plane and the direct irradiance 124. In method S100, the second machine learning model thus performs the analysis of the contribution of the diffuse irradiance 122 in the horizontal plane and the direct irradiance 124 to the hemispheric irradiance 120. Thus, the second machine learning model has taken over the task of the decomposition model or the split model used in the prior art.

[0147] The second machine learning model, for example, is a machine learning model known as a multilayer perceptron. It can also be a so-called RNN (recurrent neural network), a so-called LSTM (long-short-term memory), or so-called transformer models.

[0148] The training of the first machine learning model therefore aims to extract image features that support the tasks of the second machine learning model.

[0149] For the third machine learning model, which is applied in method S500 in method step S538, the same training data as for the second machine learning model can, in principle, be used. The third machine learning model can, for example, be a multilayer perceptron.

[0150] For example, reference data from more than a year and from multiple locations under very different atmospheric conditions can be used as training data for the first machine learning model and the second machine learning model. Atmospheric conditions include: sky cloud cover, opacity and optical thickness of the atmosphere, and the position of the sun. The training data is filtered to ensure a similar representation of all conditions. For example, averages over one minute (one-minute averages) can be used as a time resolution.

[0151] High-quality measurements of diffuse irradiance in the horizontal plane and direct irradiance obtained from the measurements and measures listed below can be used as reference data: - A two-axis solar tracker accurately tracks instruments and shadow balls or shadow shields to the sun. The solar tracker is precisely leveled horizontally to obtain accurate measurements at all positions of the sun. - A pyrheliometer tracking the sun 54 measures the direct radiation in a plane perpendicular to the sun 54 and obtains data on the direct-normal irradiance. - A sun sensor is attached to the solar tracker and connected to the solar tracker's controller. The sun sensor helps reduce small errors in the solar tracker's tracking. - A shaded radiation sensor, particularly a pyranometer, measures diffuse irradiance in the horizontal plane. The radiation sensor, particularly a pyranometer, complies with the ISO 9060:2018 standard with accuracy class A (spectrally flat). A ventilation and heating unit is used to reduce fogging caused by dew, pollution, and temperature influences. During the measurement, the radiation sensor is mounted on the solar tracker and leveled horizontally. The shadow ball or shade shield shades the radiation sensor, particularly a pyranometer. The shadow ball or shade shield is adjusted so that an aperture angle of 2.5° around the sun 54 with an inclination angle of 1° is shaded. - An unshaded horizontally leveled radiation sensor, in particular an unshaded pyranometer, is used to measure the hemispheric irradiance 120 in the horizontal plane (GHI) in order to detect measurement errors of the pyrheliometer, which measures the direct normal irradiance 124, and the shaded radiation sensor, in particular the pyranometer, which measures the diffuse irradiance 122 in the horizontal plane. A consistency check is performed on the measured values ​​according to the formula: DNI*sin(sun elevation) + DHI = GHI. The unshaded radiation sensor, in particular the unshaded pyranometer, fulfills the same requirements as the shaded radiation sensor, in particular the shaded pyranometer. - During the workweek, daily cleaning and inspection of the measurement setup and on-site inspection are typically performed, allowing measurement errors to be detected promptly. Contamination of the sensors of radiation sensor 12 and camera 14, and especially of the pyrheliometer and the shaded pyranometer, is detected and remedied in a timely manner. - On-site check of the leveling of all sensors (radiation sensor 12 and camera 14), tracking of the shadow ball and the shadow aperture and the pyrheliometer. - On-site inspection of the wiring and the glass dome of the radiation sensor 12 and the window of the pyrheliometer and all other instruments for damage and / or wear. - Regular calibration of all instruments and a data logger used. - Semi-automatic test for data control, as described for example in the reference: Geuder, N. et al., Energy Procedia, 2015, Volume 69, pp. 1989-1998.

[0152] The hemispheric irradiance 120 is used because it is very often available in solar applications without additional measurement technology. The diffuse irradiance in the horizontal plane and the direct irradiance are components of the hemispheric irradiance 120.

[0153] Thus, the hemispherical irradiance 120 provides an indication of the values ​​of diffuse irradiance in the horizontal plane and direct irradiance in the respective case.

[0154] The pyrheliometer used above complies with ISO 9060:2018 with accuracy class A. The pyrheliometer is mounted on the solar tracker and tracks the sun accurately.

[0155] Images 110 from camera 14, in particular a cloud camera, for example, a surveillance camera with a fisheye lens, are used for training. For this purpose, cameras are selected whose images exhibit minimal lens refraction effects, which can, if possible, do not require a protective dome, whose images have the highest possible dynamic range, and whose exposure settings are consistent with the image data. Furthermore, image correction may be omitted under certain circumstances.

[0156] The dynamic range of a surveillance camera's image sensor is typically 10 bits per color channel. Alternatively, it may be possible to capture exposure time series within a few seconds. This high dynamic range is advantageous because it allows the brightest and darkest areas of the sky to be captured, ensuring neither underexposure nor overexposure.

[0157] Data automatically stored in the image data typically includes: exposure time, color temperature, analog gain, digital gain, and gamma correction. This data, which is part of the captured image data, is referred to as metadata.

[0158] All measuring devices are synchronized, for example, via a so-called NTP protocol, which maintains a time offset of a few seconds. NTP stands for Network Time Protocol.

[0159] If transformed camera images from the area of ​​the sun 54 are used as input to the first machine learning model, such as in method S300, only image areas with high information content are passed to the first machine learning model.

[0160] If image features from the camera images 110 and the hemispheric irradiance 120 are used for training, an underestimation of radiation due to too many saturated pixels in the digital image 110 can be avoided. Knowledge of the solar elevation provides a general influence on the solar radiation measurement, especially on the irradiance.

[0161] Fig. 10 shows a detailed implementation of method S500 according to another embodiment. First, images 110 and values ​​of the hemispheric irradiance 120 are input as input parameters of the camera 14 and the radiation sensor 12. The input parameters 120 may also contain tabular values ​​from previous measurements. The input parameters are subjected to various method steps S538 and S530 for modification and analysis. S538 may, for example, be implemented as a multilayer perceptron with a single hidden layer. The analyzed data sets are combined, in particular concatenated, into a common data set in method step S540. The common data set represents a data vector or vector data set.

[0162] In the example shown in Fig. 9 shown optional method step S535 for image analysis of the hemispherical irradiance 120 and / or the image 110 of the camera 14 is carried out in the Fig. 10 is omitted. The image transformation step S515 can be applied optionally.

[0163] The data vector after method step S540 is then transferred to the second machine learning model S550 for analysis. Using the second machine learning model, the desired components of the hemispheric radiation 120 are extracted. The extracted components are the diffuse irradiance 122 in the horizontal plane and / or the direct irradiance 124. In method step S550, the hidden layers of the machine learning model are summarized. Two hidden layers 555 and 556 are shown as examples in the illustration in Fig. 10. However, it may also be intended to use more or fewer hidden layers in process step S550.

[0164] Method steps S530 and S516 summarize various substeps of the first machine learning model. Method steps S530 and / or S516, for example, relate to convolution steps of the data used in the CNN algorithm. Method steps S530 and / or S516 may also include method steps that involve pooling the mean values, taking maximum values ​​(maximum pooling) and / or minimum values ​​(minimal pooling) into account. Following these method steps, a multilayer perceptron, particularly one with a single hidden layer, may be used, for example, in S538, to reduce the number of extracted image features.

[0165] The architecture is standard. For example, an architecture known as ResNet-18 is used. ResNet stands for Residual Neural Network. A residual neural network is a deep learning model in which the weight layers learn residual functions with respect to the layer inputs. A residual neural network is a network with jump connections that perform identity mappings and are merged with the layer outputs by summing them. The number "18" denotes the number of layers.

[0166] Fig. Figure 11 shows parameters used in training the first machine learning model. Fig. Figure 11 is a representation created using data from fastai's OneCycleScheduler class, described at https: / / fastai1.fast.ai / callbacks.onecycle.html#OneCycleScheduler. The training strategy is referred to as a one-cycle policy. The one-cycle policy is a learning strategy applied over one learning cycle and enables faster training of a model. After the first cycle, the learning rate can be applied repeatedly to obtain better values.

[0167] In the left part of Fig. 11 shows a curve 600. The curve 600 describes a progression of a learning rate 630 plotted as a function of the number of iterations 620. At the beginning of the process for training the first machine learning model, the learning rate 630 increases at a low iteration rate, and at the end of the process for training the first machine learning model, the learning rate 630 decreases again.

[0168] In the right part of Fig. 11 shows a curve 640 of an impulse 650 as a function of the number of iterations 620.

[0169] The momentum is a parameter that is used to update the weights. The momentum accumulates the gradients from the last iteration steps, enabling faster convergence and overcoming saddle points and local minima. The momentum 640 drops to a minimum after a relatively high value at the beginning of the process and rises again at the end of the process for training the first machine learning model.

[0170] In addition to the structures and textures of the image 110 of the sky 50, intermediate results of a method for determining the irradiance of solar radiation and / or its components in any plane, in particular a plane inclined to a horizontal plane, can also be used as features. The method uses physical models that take physical conditions and assumptions into account. The intermediate results can be one of the following features: - values ​​of diffuse irradiance in the horizontal plane, - direct radiation values, - values ​​of irradiance in any plane, where the plane is different from the plane in which the radiation sensor and / or the camera are arranged, - Estimates of the irradiance in the red channel, the green channel and / or the blue channel.

[0171] The method is described in detail in WO 2021 / 219570 A1. The method comprises the following method steps, with one or more of the method steps being applied to obtain the intermediate results. The raw data used can be measurement data from the camera 14 and the radiation sensor 12 or measurement data recorded at other locations and with other cameras and / or radiation sensors and stored, for example, in a database or table.

[0172] The red-green-blue (RGB) color channels of the sky image are weighted together. This weighting of the channels ensures the camera's sensitivity is as uniform as possible in the visible wavelength range. This gray value is multiplied by a broadband correction to account for radiation at wavelengths outside the camera's measurement range.

[0173] Using a standard geometric internal and external calibration for cloud cameras, a sky region (azimuth and zenith angle) is specified for each pixel of the camera image, yielding an estimate of the sky's radiance distribution.

[0174] Analogous to the radiance distribution, a luminance distribution is calculated. For this purpose, the RGB color channels are weighted according to the sensitivity of the human eye before summation. Integration of the luminance distribution across all angular ranges yields a measured value of the illuminance. The illuminance output by the camera and the illuminance calculated from the camera image are compared. The radiance distribution is scaled according to the ratio of the two values ​​to compensate for any influence of the camera control on the camera's sensitivity.

[0175] The area of ​​the solar disk is masked. For an evaluated sensor plane, each sky region in the radiance image is weighted according to a projection into the plane. Integrating the radiance distribution across all sky regions within the field of view of the inclined plane yields the diffuse irradiance from the sky for the respective plane.

[0176] The horizontal diffuse irradiance in the plane of the radiation sensor unit is calculated accordingly. The direct normal irradiance (DNI) is calculated by comparing it with the horizontal hemispheric irradiance measured by a pyranometer and taking into account the current position of the sun.

[0177] To correct for refraction effects in the camera lens, the initial estimate of diffuse irradiance in the plane of the irradiance sensor unit, as well as all other calculated diffuse irradiances, are reduced by a portion of the DNI (lens refraction correction). The correction is then added accordingly to the direct irradiance in the plane of the irradiance sensor unit. The DNI is then recalculated.

[0178] The GTI in an evaluated plane is ultimately determined from a direct component, a diffuse component from the sky, and a component reflected from the ground. The DNI is projected into the evaluated plane and thus yields the direct component. The diffuse irradiance is calculated from the camera image for this plane as described above. The reflected component is determined as the GHI multiplied by the albedo of the subsurface and the term (1 - cos(angle of inclination of the inclined plane to the horizontal)) / 2.According to the specified method, in order to convert the hemispherical irradiance of the solar radiation determined with the radiation sensor unit in the plane of the radiation sensor unit into the irradiance and / or their respective components, direct irradiance, diffuse irradiance, irradiance reflected on the ground, in the horizontal plane and / or in the plane inclined to the horizontal plane, at least one of the quantities of irradiance reflected on the ground, and / or diffuse irradiance, and / or the position of the sun during the radiation measurement, and / or a sensor-specific correction factor, which in particular includes lens parameters of the camera, can be used.

[0179] Factors such as the position of the sun that influence the radiation measurement can be taken into account in the conversion.

[0180] Furthermore, according to the specified method for converting camera measurement values, at least one of the variables from a ratio of broadband radiation to the portion of radiation recorded by the camera, and / or a spectral intensity of RGB channels of the camera, and / or an internal and / or external calibration of the camera, and / or an inclination and orientation of the camera sensor, and / or the position of the sun during the radiation measurement, and / or a camera sensitivity determined from an illuminance of the camera and / or the spectral sensitivity of the RGB channels and / or recording settings and / or the RGB camera image and / or the internal and / or external calibration of the camera can be used. Factors influencing the radiation measurement, such as the position of the sun, can be taken into account during the conversion.

[0181] The individual steps for determining the hemispheric irradiance of solar radiation and / or its respective components, direct irradiance, diffuse irradiance, ground-reflected irradiance, in the horizontal plane and in the plane inclined to the horizontal plane are described.

[0182] The radiation reflected from the ground is determined using albedo, inclination and orientation of the inclined plane, as well as a measured value of the hemispheric irradiance in the plane of the radiation sensor unit.

[0183] The direct irradiance in the plane of the radiation sensor unit is determined by subtracting the measured value of the diffuse irradiance, evaluated for the plane of the radiation sensor unit, from the hemispherical irradiance in the plane of the radiation sensor unit. The diffuse irradiance in the plane of the radiation sensor unit is determined as described in more detail below.

[0184] The direct irradiance is then determined by inverting the projection into the plane of the radiation sensor unit using the position of the sun calculated from the location and time.

[0185] The direct normal radiation is multiplied by a correction factor, which includes, in particular, the lens parameters of the camera used. This results in a lens refraction correction.

[0186] The direct irradiance in the plane of the radiation sensor unit and the lens refraction correction are added, and the projection into the plane of the radiation sensor unit is inverted, taking into account the calculated position of the sun. This direct normal irradiance is then projected into the horizontal and / or inclined plane, taking into account the inclination and orientation of the horizontal and / or inclined plane, to obtain the corrected measured value of the direct irradiance in this plane.

[0187] The lens refraction correction is subtracted from the diffuse irradiance evaluated for the inclined or horizontal plane as described above. This results in a corrected measured value of the diffuse irradiance in the respective inclined or horizontal plane.

[0188] The hemispheric irradiance in the horizontal and / or inclined plane can then be determined by summing the radiation reflected from the ground, the direct radiation as a component of the hemispheric irradiance in the horizontal and / or inclined plane, and the diffuse irradiance as a component of the hemispheric irradiance, evaluated for the horizontal and / or inclined plane, and corrected. The determination of the diffuse irradiance originating from the sky in the horizontal or inclined plane, particularly in the plane of the radiation sensor unit, is summarized below.

[0189] First, a broadband correction factor is determined from the ratio of broadband radiation to the portion registered by the camera using the daylight spectrum and the spectral sensitivity of RGB channels of the camera.

[0190] For this purpose, weights of the RGB channels are determined according to the inverse sensitivity using the camera's recording settings.

[0191] This allows the weighted RGB channels of the camera image to be summed.

[0192] The summed RGB channels are then multiplied by the determined broadband correction factor.

[0193] This allows angular areas of the sky to be assigned to camera image pixels using the camera's internal and / or external calibration values.

[0194] These image areas are then weighted according to the projection into the horizontal and / or inclined plane.

[0195] In parallel, the angular range of the field of view of the horizontal and / or inclined plane can be determined from the inclination and orientation of this plane, while the angular range of the solar disk is determined from the location and time of day.

[0196] This allows the angular range of the solar disk to be excluded from the angular range of the field of view of the horizontal and / or inclined plane.

[0197] The image areas can then be integrated across the field of view of the horizontal or inclined plane.

[0198] The diffuse irradiance in the horizontal or inclined plane, in particular in the plane of the radiation sensor unit, can then be calculated by multiplying it by the camera sensitivity correction factor, which was previously determined.

[0199] The calculation of the camera sensitivity correction factor based on a comparison of the illuminance output by the camera and the calculated illuminance can be done as described below.

[0200] First, weights corresponding to the sensitivity of each RGB channel of the camera used are determined using the spectral sensitivity of the RGB channels and the camera's recording settings.

[0201] Weights are then determined based on human perception.

[0202] Using these weights, RGB channels from the RGB camera image can then be summed together in a weighted manner.

[0203] In parallel, angular areas of the sky are assigned to image pixels of the camera using the camera's internal and / or external calibration values.

[0204] The weighted RGB channels are then integrated over the hemisphere of the sky.

[0205] This allows the correction factor of the camera sensitivity to be determined by calculating the ratio of the camera's illuminance to the integrated weighted RGB camera image. Reference symbol 5 Device 10 measuring units 12 Radiation sensor unit 14 Camera 16 Measurement data acquisition unit 18 Evaluation unit 19 computers 20 Data processing device 22 housings 24 sensors 26 Field of view of the radiation sensor 28 Camera housing 30 Camera sensor 32 Camera field of view 34 distance 36 horizontal plane 38 Level of the sensor 24 40 Level of the sensor 30 42 Elevation, Inclination 43 Elevation 44 angles, angle of inclination 46 Level of the radiation sensor 48 Camera level 50 Heavens 52 Cloud 54 Sun 100 procedures 110 image taken with the camera 120 hemispheric irradiance 122 diffuse irradiance 124 direct irradiance 530 hidden layer 538 hidden layer 555 hidden layer 556 hidden layer 600 Curve of the learning rate as a function of iterations 620 620 iterations 630 learning rate 640 Curve of the impulse plotted as a function of iterations 620 650 pulse S100 procedure S130 Extracting features using the first machine learning model S140 Concatenation, merging into a common data set S150 Determining the diffuse irradiance in the horizontal plane and / or the direct irradiance using the second machine learning model S200 procedure S230 Extracting features using the first machine learning model S235 Image analysis S240 Concatenation, merging into a common data set S250 Determining the diffuse irradiance in the horizontal plane and / or the direct irradiance using the second machine learning model S300 procedure S315 Image Transformation S330 Extracting features using the first machine learning model S335 Image analysis S340 Concatenation, merging into a common data set S350 Determining the diffuse irradiance in the horizontal plane and / or the direct irradiance using the second machine learning model S400 procedure S415 Image transformation S430 Extracting features using the first machine learning model S440 Concatenation, merging into a common data set S450 Determining diffuse irradiance in the horizontal plane and / or direct irradiance using the second machine learning model S500 procedure S515 Image Transformation S516 Process steps, for example folding steps S530 Extracting features using the first machine learning model S535 Image Analysis S538 Processing features from image analysis using the third machine learning model S540 Concatenation, merging into a common data set S550 Determining the diffuse irradiance in the horizontal plane and / or the direct irradiance using the second machine learning model QUOTES CONTAINED IN THE DESCRIPTION

[0000] This list of documents submitted by the applicant was generated automatically and is included solely for the convenience of the reader. This list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions. Cited patent literature

[0000] WO 2021 / 219570 A1 [0006, 0034, 0171] Cited non-patent literature

[0000] Geuder, N. et al., Energy Procedia, 2015, Volume 69, pp. 1989-1998

[0151]

Claims

[1] A method (S100, S200, S300, S400, S500) for determining at least one component of a hemispheric irradiance (120) of solar radiation in any plane, wherein the at least one component comprises a diffuse irradiance (122) and / or a direct irradiance (124), comprising the following steps: (i) determining measurement data of the hemispherical irradiance (120) with a radiation sensor (12) in a field of view (26) above a plane (46) of the radiation sensor (12); (ii) capturing an image (110) of the sky (50) with a camera (14) in a field of view (32) above a plane (48) of the camera (14); (iii) extracting (S130, S230, S330, S430, S530) features, in particular structures, from the image (110) of the sky (50) by means of a first machine learning model and generating a result data set; (iv) merging (S140, S240, S340, S440, S540) the measurement data of the radiation sensor (12) and the result data set of the camera (14) into a common data set; (v) determining (S150, S250, S350, S450, S550) the at least one component of the hemispherical irradiance (120) in the arbitrary plane, in particular the diffuse irradiance (122) and / or the direct irradiance (124), from the data set by means of a second machine learning model. [2] The method of claim 1, wherein the second machine learning model is different from the first machine learning model, wherein the first machine learning model is trained with first training data sets, wherein the second machine learning model is trained with second training data sets. [3] Method according to claim 1 or 2, wherein features are extracted in an image analysis (S235, S335, S535) using the image (110) of the camera (14), in particular wherein features are extracted from the measurement data of the radiation sensor unit (12). [4] Method according to claim 3, wherein in the image analysis (S235, S335, S535) using the image (110) of the camera (14) features are derived from the intensities of the image pixels, in particular wherein a physical camera model is applied and / or a number of saturated pixels and / or at least one sum and / or weighted sum of the intensities of the image pixels and / or at least one ratio of sums and / or weighted sums of the intensities of the image pixels are included for the extraction of features. [5] Method according to claim 3 or 4, wherein the features extracted from the image (110) of the camera (14) and / or from the measurement data of the radiation sensor (12) are transferred to a third machine learning model and processed by means of the third machine learning model (S538) and a radiation sensor data set is generated, wherein the radiation sensor data set and the result data set of the camera (14) are merged (S540) before the second machine learning model is applied (S550). [6] Method according to one of the preceding claims, wherein an image transformation (S315, S415, S515) of the image (110) of the sky (50) takes place before the application of the first machine learning model. [7] Method according to one of the preceding claims, wherein the first machine learning model and / or the second machine learning model is trained with reference data of at least one of the components of the hemispheric irradiance (120), in particular the diffuse irradiance (122) in the horizontal and / or inclined plane, in particular from solar trackers, and / or the direct irradiance (124), in particular from a pyrheliometer tracking the sun, and / or with regular input data, in particular with images (110) of the camera (14), data of the hemispheric irradiance (120) and / or intermediate results, which were obtained in particular from the application of a physical model. [8] A method according to any one of the preceding claims, wherein the first machine learning model is pre-trained using at least one of the following steps: (i) Use of publicly available weights, (ii) training using at least one unsupervised or self-supervised training approach, (iii) extending data with Gaussian blurs and / or distorting image colors and / or mirroring the images (110), (iv) Training with images (110) of the sky (50) taken by the camera (14). [9] Method according to one of the preceding claims, wherein a convolutional neural algorithm, in particular a convolutional neural network, is used in the first machine learning model. [10] Method according to one of the preceding claims, wherein in the second machine learning model at least one algorithm from a multilayer perceptron (MLP), random forest algorithm, recurrent neural network, long short term memory algorithm, transformer model algorithm, k-nearest neighbor (k-NN) algorithm, and / or support vector machine algorithm is used. [11] Method according to one of the preceding claims, wherein the first training data sets and / or the second training data sets are filtered such that different atmospheric conditions are represented to a similar extent. [12] Method according to one of the preceding claims, wherein the first machine learning model, the second machine learning model and / or the third machine learning model are trained in a supervised manner using input data and / or correct reference data of at least one of the components of the hemispheric irradiance (120). [13] Method according to one of the preceding claims, wherein the first machine learning model, the second machine learning model and the third machine learning model are trained together, in particular supervised, in particular after a respective individual training. [14] Method according to one of the preceding claims, wherein image features from the image (110) of the sky (50) using estimated values of a red channel, and / or a blue channel, and / or a green channel from the measurement data of the image (110) of the sky (50) are used as additional input data of the third machine learning model or the merging (S140, S240, S340, S440, S540). [15] Device (5) for carrying out a method (S100, S200, S300, S400, S500) for determining at least one component of a hemispherical irradiance (120) of solar radiation in any plane, according to one of the preceding claims, wherein the at least one component comprises a diffuse irradiance (122) and / or a direct irradiance (124), comprising at least one radiation sensor unit (12), a camera (14), and an evaluation unit (18) which is provided for evaluating measurement data from the radiation sensor unit (12) and the camera (14), wherein the radiation sensor unit (12) determines the irradiance of solar radiation (120) in a field of view (26) of 180° above a plane (38, 46), wherein the plane (38, 46) is a horizontal plane or an inclined plane, wherein the camera (14) acquires measurement data comprising an image (110) of the sky (50) in a field of view (32) of 180° above a plane (40, 48), wherein the plane (40, 48) is a horizontal plane or an inclined plane, wherein the evaluation unit (18) has at least a first machine learning model and a second machine learning model, wherein the evaluation unit (18) is at least designed to carry out the following steps: (i) extracting (S130) features, in particular structures, from the image (110) of the sky (50) by means of a first machine learning model and generating a result data set; (ii) merging (S140) the measurement data of the radiation sensor (12) and the result data set of the camera (14) into a common data set; (iii) determining (S150) the at least one component of the hemispherical irradiance (120) in the arbitrary plane, in particular the diffuse irradiance (122) and / or the direct irradiance (124), from the data vector by means of a second machine learning model. [16] Device according to claim 15, wherein the measurement data of the camera (14) contain angle-resolved radiation information. [17] Method for training a first machine learning model for extracting features, in particular structures, from an image (110) of the sky (50), comprising at least one of the following steps: (i) determining or using data sets of images (110) of the sky (50) taken with a camera, in particular a cloud camera; (ii) training the machine learning model with the data sets of images (110) of the sky (50); (iii) Training the first machine learning model with known images (110) of the sky (50) taken at other locations. [18] The method of claim 17, wherein pretraining is used with at least one of the following steps: (i) Use of publicly available weights; (ii) training using at least one unsupervised or self-supervised training approach; (iii) augmenting data with Gaussian blurs and / or distorting image colors and / or mirroring the images; (iv) Training with images (110) of the sky (50) taken by the camera (14). [19] Computer program for determining at least one component of a hemispheric irradiance (120) of solar radiation in any plane, comprising instructions which, when the program is executed by a computer (19), cause the computer (19) to carry out the steps of a method (S100, S200, S300, S400, S500) for determining at least one component of a hemispheric irradiance of solar radiation according to one of claims 1 to 14. [20] Data processing device (20) for a device (5) for determining at least one component of a hemispherical irradiance (120) of solar radiation in any plane, according to claim 15 or 16, at least comprising a measurement data acquisition unit (16), an evaluation unit (18) and a computer (19). [21] A trained first machine learning model for a method (S100, S200, S300, S400, S500) for determining at least one component of a hemispheric irradiance (120) of solar radiation in any plane according to one of claims 1 to 14, for extracting features, in particular structures, from an image (110) of the sky (50) which is trained according to a method according to claim 17 or 18. [22] A trained second machine learning model for a method (S100, S200, S300, S400, S500) for determining at least one component of a hemispheric irradiance (120) of solar radiation in any plane according to one of claims 1 to 14, for determining at least one component of a hemispheric irradiance (120) of solar radiation in any plane, which has been trained in a supervised manner, in particular using input data and reference data, in particular comprising a multilayer perceptron algorithm.

Images