A photovoltaic power prediction method and medium

Abstract

A photovoltaic power prediction method and medium, relating to the photovoltaic technical field, comprising the following steps: S1, acquiring original observation data; S2, acquiring a clear sky reference power, and acquiring an image-power alignment sample; S3, acquiring image-derived physical diagnostic features from the image-power alignment sample; S4, grouping according to the image-derived physical diagnostic features to acquire grouped image-derived physical quantities; S5, constituting a historical window sample; S6, respectively encoding the grouped image-derived physical quantities to obtain each physical mechanism group representation; fusing the each physical mechanism group representation with an image sequence representation and a power history sequence representation to obtain a fusion representation object; and S7, inputting the fusion representation object into a prediction model to output photovoltaic power prediction results for multiple future prediction time intervals. According to the present application, image features are divided into four groups according to physical characteristics, and each group is respectively time-sequentially encoded, thereby improving the accuracy of prediction.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic technology, and more specifically, to a photovoltaic power prediction method and medium. Background Technology

[0002] As the proportion of photovoltaic (PV) power generation in the power system continues to increase, the randomness, intermittency, and volatility of PV power place higher demands on grid dispatch, energy storage coordinated control, and renewable energy consumption. Especially in short-time and ultra-short-time forecasting scenarios, PV power is significantly affected by factors such as local cloud obstruction, cloud movement, changes in solar altitude angle, and changes in sky conditions near the solar disk, leading to rapid power fluctuations on a minute-level timescale. Therefore, improving the accuracy, stability, and interpretability of short-term PV power forecasting has become an important research issue in the field of renewable energy forecasting.

[0003] Photovoltaic power generation is easily affected by factors such as cloud cover, cloud edge movement, changes in solar position, and local sky brightness variations within a short timescale, resulting in frequent fluctuations in power output, including sudden drops and rises, and unstable recovery processes. For applications such as photovoltaic dispatching, energy storage coordination, grid connection control, and short-term early warning, it is often necessary to provide power forecasts for multiple forecast intervals in advance. Therefore, how to more accurately describe and predict photovoltaic power fluctuations caused by changes in sky conditions within a short period has become a crucial problem that existing technologies need to address.

[0004] In existing technologies, one type of method mainly relies on historical photovoltaic power series for prediction. This type of method can utilize the inertia and recent trend information of power plant output, but it is difficult to detect the impending shading before the clouds actually affect the power plant's power. Therefore, for sudden power changes caused by rapid cloud movement, there are often problems of response lag and large prediction errors.

[0005] Another type of approach incorporates visual information, such as sky images, in addition to historical power data, enhancing predictive capabilities through a multimodal approach. While this type of method can reflect cloud distribution and sky conditions to some extent, many existing schemes still primarily rely on black-box learning of raw image features by the model, lacking explicit characterization of the formation, development, and resolution of solar shading. In other words, although these methods "see the image," they may not be able to clearly distinguish which image changes are truly directly related to short-term fluctuations in photovoltaic power.

[0006] Another approach attempts to further extract image-derived physical quantities such as cloud cover, brightness, and solar visibility from sky images and use them as additional inputs for prediction. This approach is closer to the actual physical processes than the aforementioned methods. However, a common problem with existing practices is that they directly flatten physical diagnostic features from different sources, with different mechanisms, and at different stages of action—such as solar geometry, cloud movement, the relative position of clouds and the sun, and changes in the appearance of the solar disk—and input them into the same encoding path. This causes information from different physical mechanisms to become mixed before entering the model, making it difficult for the model to distinguish between different levels of questions such as "where is the sun?", "how are the clouds moving?", "when might the sun be obscured?", and "is the sun currently obscured?". This not only affects the stable learning of the prediction model but also hinders the subsequent interpretation of the basis for the prediction results.

[0007] Therefore, the main shortcoming of existing technologies is not merely the lack of a single type of input data, but rather the lack of a technical solution capable of systematically organizing, hierarchically representing, and structurally utilizing physical diagnostic information derived from sky images related to the solar shading process. Especially in short-term, multi-step photovoltaic power prediction scenarios, how to effectively combine sky images, historical photovoltaic power data, and image-derived physical diagnostic features, while avoiding information interference caused by the direct mixing of different mechanistic features as input, remains a key technical problem that has not yet been clearly resolved. Summary of the Invention

[0008] To address the aforementioned technical problems in related technologies, in a first aspect, the present invention proposes a photovoltaic power prediction method, which includes the following steps:

[0009] Step S1: Obtain raw observation data; the raw observation data includes: time-stamped sky image data, time-stamped photovoltaic power data, and basic parameters related to the power station;

[0010] Step S2: Calculate the clear-sky reference power under unobstructed conditions based on the original observation data. ,in, The current time is indicated; using the sky image timestamp as the anchor point, each image is matched with the photovoltaic power record of the corresponding time or the most recent time, and the measured power and clear sky reference power are written into the same sample record to form an image-power alignment sample.

[0011] Step S3: Obtain image-derived physical diagnostic features from the image-power aligned samples; the image-derived physical diagnostic features include: solar geometric features, cloud cover and cloud motion features, cloud-sun relative occlusion geometric features, and solar disk appearance features;

[0012] Step S4: Based on the physical mechanism of the image-derived physical diagnostic features, group the images to obtain the grouped image-derived physical quantities; the grouping includes at least: solar geometry group, cloud motion group, solar occlusion geometry group, and solar disk appearance group.

[0013] Step S5, using the current predicted time Using the anchor point as an example, sky images, measured power, image-derived physical diagnostic features, and corresponding clear-sky reference power are extracted from a continuous historical period preceding it to form a historical window sample. ;

[0014] Step S6: Encode the derived physical quantities of the grouped images to obtain the representation of each physical mechanism group; fuse the representations of each physical mechanism group with the image sequence representation and the power history sequence representation to obtain the fused representation object. ;

[0015] Step S7, the fused representation object Input the prediction model to output photovoltaic power prediction results for multiple future prediction time intervals.

[0016] Specifically, step S3 includes:

[0017] Step S31: Obtain the current sky image, timestamp, and measured photovoltaic power at the corresponding time from the image-power alignment sample. and clear sky reference power ;

[0018] Step S32: Calculate the solar position and solar geometric fundamentals based on the image timestamp, power station latitude and longitude, altitude, and time zone;

[0019] Step S33: Obtain a clear sky reference image under similar solar altitude conditions;

[0020] Step S34: Generate an intermediate mask for the cloud region based on the current image and the clear sky reference image;

[0021] Step S35: Calculate the cloud coverage intensity and the cloud coverage amount in the near-sun region using the intermediate mask of the cloud region;

[0022] Step S36: Estimate cloud cluster movement characteristics based on cloud region location at consecutive time points;

[0023] Step S37: Calculate the relative occlusion geometry between the cloud region and the sun's position;

[0024] Step S38: Construct the estimated occlusion / removal time and time-division occlusion physical characteristics based on the relative motion;

[0025] Step S39: Calculate the image-derived physical features of the solar disk's brightness, visibility, and appearance.

[0026] Specifically, the prediction model is a multi-step prediction head used to output power prediction values ​​at multiple prediction time intervals.

[0027] Specifically, the calculation of the clear sky reference power is as follows: the Ineichen clear sky model is used, the open_rack_glass_glass temperature model is used, the ambient temperature is 20°C, the wind speed is 1m / s, and the chunk_size parameter is 20000.

[0028] Specifically, the duration of a continuous historical segment in step S5 is 45 minutes.

[0029] Specifically, step S32 involves: calculating the solar altitude angle, solar azimuth angle, and apparent zenith angle based on the timestamp and the power station's geographical location; and calculating the rate of change of the solar altitude angle, the rate of change of the solar azimuth angle, and the amount of motion of the solar center in the image coordinates based on the changes in the solar position at adjacent times.

[0030] Step S33 specifically involves: calculating the clear sky index based on the measured power and the clear sky reference power.

[0031] ,

[0032] in, Indicates the first Measured photovoltaic power at any given time This represents the reference power at that moment under unobstructed, clear sky conditions. To prevent extremely small constants with a denominator of zero, Used to measure how close the current measured power is to the clear-sky reference power;

[0033] The samples were grouped according to the numerical range of the solar altitude angle. Within each altitude angle range, the samples with higher clear sky indices were selected. The samples are used as approximate clear sky samples, and the median of the sky images of these samples is taken pixel by pixel to obtain a clear sky reference image under the corresponding solar altitude conditions. .

[0034] Specifically, step S34 is as follows:

[0035] For the coordinates in the current sky image, The pixel, whose red channel value is recorded. The value of the blue channel is Calculate the normalized red-blue difference:

[0036] ,

[0037] Meanwhile, for clear sky reference images Calculation of pixels with the same coordinates And calculate the difference between the current image and the clear sky reference image:

[0038] ,

[0039] Then, based on the above two quantities, a cloud region intermediate mask is generated. :

[0040] ,

[0041] in, This indicates that the value is 1 when the condition is true and 0 when the condition is false. Represents pixels Pixels identified as cloud-related This indicates that the pixel was not identified as a cloud-related pixel.

[0042] Specifically, step S35 is as follows:

[0043] Calculate the original cloud cover intensity in the intermediate mask image of the cloud region:

[0044] ,

[0045] in, and These represent the image height and width, respectively.

[0046] Specifically, step S36 is as follows:

[0047] First according to Calculate the centroid of the cloud region using the pixel coordinates with a value of 1. Then, based on the changes in the centroid coordinates at consecutive time intervals, the motion velocity of the cloud region in the image coordinates is estimated; the time interval between two adjacent images is obtained. ,but:

[0048] ,

[0049] ,

[0050] in, and These represent the motion components of the cloud region's centroid in the horizontal and vertical directions of the image, respectively. , Let x and y represent the x and y coordinates of the cloud region's centroid in the image coordinate system at time t, respectively. , Let x and y represent the x and y coordinates of the cloud region centroid in the image coordinate system at time t-1, respectively. This indicates the actual observation time corresponding to the previous image;

[0051] Further calculations were made regarding the speed and direction of cloud movement:

[0052] ,

[0053] ,

[0054] in, Indicates the speed of cloud movement. Indicates the direction of cloud movement.

[0055] In a second aspect, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a photovoltaic power prediction method as described in any one of the first aspects above.

[0056] The photovoltaic power prediction method of the present invention improves the completeness of the information used in short-term prediction by introducing joint modeling of sky image sequences and historical photovoltaic power sequences.

[0057] This invention introduces sky image sequences as synchronous observation information in addition to historical photovoltaic power sequences, enabling the prediction process to utilize both existing power variation information from the power plant and changes in external sky conditions. In this way, when making short-term predictions, the model can retain the temporal continuity of historical power while also being aware of prior cues from cloud distribution, solar position, and local brightness changes, thus providing a more complete input basis for subsequent predictions.

[0058] 2. By constructing image-derived physical diagnostic features, the solar occlusion process can participate in prediction as an explicit intermediate object, thus improving the interpretability of the scheme.

[0059] 3. By grouping and encoding the image-derived physical diagnostic features according to their physical mechanisms, interference caused by heterogeneous information mixing is reduced.

[0060] The most significant technical effect of this invention lies in the organization and utilization of image-derived physical diagnostic features. Existing methods commonly involve directly tiling physical features from different sources and mechanisms and inputting them into the same encoding path. In this approach, information such as the sun's position, cloud movement, occlusion formation conditions, and the current appearance of the solar disk are mixed together during the input stage. The model needs to manually distinguish the hierarchical relationships between these pieces of information, which can easily lead to mutual interference.

[0061] This invention first divides these features into a solar geometry group, a cloud motion group, a solar occlusion geometry group, and a solar disk appearance group, and then performs temporal encoding on each group. Through this process, information from different physical mechanisms first forms a representation along their respective pathways, and then is fused at a higher level. This helps maintain the consistency of the mechanisms within each type of information and also helps reduce mutual interference between heterogeneous features. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 This is a schematic diagram of the photovoltaic power prediction structure provided in an embodiment of the present invention;

[0064] Figure 2 This is a flowchart of the photovoltaic power prediction method provided in the embodiments of the present invention;

[0065] Figure 3 This is a schematic diagram of the sky image provided in an embodiment of the present invention. Figure 1 ;

[0066] Figure 4 This is a schematic diagram of the sky image provided in an embodiment of the present invention. Figure 2 ;

[0067] Figure 5 This is a schematic diagram of the sky image provided in an embodiment of the present invention. Figure 3 ;

[0068] Figure 6 This is a comparison chart of prediction results for different schemes under the multi-cloud occlusion case provided in the embodiments of the present invention;

[0069] Figure 7 This is a schematic diagram of a photovoltaic power prediction device provided in an embodiment of the present invention. Detailed Implementation

[0070] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0071] Example 1

[0072] refer to Figure 1-2This embodiment discloses a photovoltaic power prediction method, which includes the following steps:

[0073] Step S1: Obtain raw observation data; the raw observation data includes: time-stamped sky image data, time-stamped photovoltaic power data, and basic parameters related to the photovoltaic power station;

[0074] The process involves acquiring timestamped sky imagery and timestamped photovoltaic (PV) power data (or historical power data), along with fundamental parameters related to the PV power plant, including geographical location, time information, and PV system parameters. Sky images reflect cloud distribution, sun position, and sky brightness. PV power data reflects the actual output power of the PV power plant at historical moments, serving as a historical basis for future power predictions. PV power plant parameters are used to subsequently calculate the sun's position and clear-sky reference power (referring to the reference power calculated based on time, geographical location, and system parameters under a clear-sky assumption).

[0075] Step S2: Calculate the clear-sky reference power under unobstructed conditions based on the original observation data. ,in, The current time is indicated; using the sky image timestamp as the anchor point, each image is matched with the photovoltaic power record of the corresponding time or the most recent time, and the measured power and clear sky reference power are written into the same sample record to form an image-power alignment sample.

[0076] After obtaining the raw observation data, the clear-sky reference power under unobstructed conditions is calculated based on the observation time, geographical location, and power station parameters, and denoted as . .in, Indicates the current moment. This represents the theoretically achievable reference power of the power station under clear sky conditions at that moment. The purpose of introducing this reference value is to provide a physical reference for subsequent judgments on whether there is any obstruction in the current sky, the degree of obstruction, and how it may evolve in the future.

[0077] Specifically, clear-sky reference power The parameters were calculated using the pvlib tool based on the image timestamp, power station latitude and longitude, altitude, module tilt angle, module azimuth angle, rated DC power, and temperature.

[0078] Subsequently, using the sky image timestamp as anchor, each image is matched with the photovoltaic power record of the corresponding or most recent time, and the measured power and clear-sky reference power are written into the same sample record to form an image-power aligned sample (aligning the time of the image and the power to obtain the power corresponding to each image). After this step, the original observation object is transformed into a basic sample object that can be directly used in subsequent analysis and modeling.

[0079] The output of this step is an aligned sample object. This object is used in the next step to generate image-derived physical diagnostic features.

[0080] Step S3: Obtain image-derived physical diagnostic features from the image-power aligned samples; the image-derived physical diagnostic features include: solar geometric features, cloud cover and cloud motion features, cloud-sun relative occlusion geometric features, and solar disk appearance features;

[0081] This step converts the aligned sky image into image-derived physical diagnostic features. Constructed around the solar occlusion process, it primarily describes the sun's position, cloud cover and movement, the relative occlusion relationship between cloud regions and the sun, and the appearance of the solar disk. This object serves as input for the next step, and is subsequently divided into solar geometry, cloud movement, solar occlusion geometry, and solar disk appearance groups according to physical mechanisms, each entering its corresponding encoding path.

[0082] Step S31: Obtain the current sky image, timestamp, and measured photovoltaic power at the corresponding time from the image-power alignment sample. and clear sky reference power ;

[0083] The input for this step is the aforementioned aligned sample object. For the first... A sample of time points, including sky images. Image timestamp, corresponding measured photovoltaic power at the time. and clear sky reference power In this embodiment, the sky image is uniformly processed as follows: The pixels are processed, and the RGB pixel values ​​are normalized to a numerical image. This ensures that subsequent pixel-level calculations are performed on different image samples under the same spatial size and numerical scale.

[0084] The output of this step is a standardized sky image and its corresponding time information, which is used for subsequent calculation of the sun's position, matching of clear sky reference images, determination of cloud regions, and diagnosis of the appearance of the solar disk.

[0085] Step S32: Calculate the solar position and solar geometric fundamentals based on the image timestamp, power station latitude and longitude, altitude, and time zone;

[0086] This step is used to determine the first... The system displays the position and changing trend of the sun under current observation conditions. Inputs include image timestamps, power station latitude and longitude, altitude, and time zone. During processing, solar position parameters such as solar altitude angle, solar azimuth angle, and apparent zenith angle are calculated based on the timestamps and the power station's geographical location. In continuous time samples, the system can also calculate the rate of change of solar altitude angle, the rate of change of solar azimuth angle, and the movement of the sun's center within the image coordinates based on the changes in the sun's position at adjacent times.

[0087] The output of this step includes the solar altitude angle, solar azimuth angle, apparent zenith angle, rate of change of solar angle, and the sun's position in image coordinates. These results constitute the main source of solar geometry, used to characterize "where the sun is and how it changes".

[0088] Step S33: Obtain a clear sky reference image under similar solar altitude conditions;

[0089] This step is used to obtain an unobstructed reference for the current sky image under conditions of similar solar positions. The inputs are sky images from historical samples, solar altitude angle, and measured photovoltaic power. and clear sky reference power .

[0090] During processing, first calculate the clear sky index based on the measured power and the clear sky reference power:

[0091] ,

[0092] in, Indicates the first Measured photovoltaic power at any given time This represents the reference power at that moment under unobstructed, clear sky conditions. To prevent extremely small constants with a denominator of zero. Used to measure how close the current measured power is to the clear-sky reference power.

[0093] In this embodiment, the samples are grouped according to the numerical range of the solar altitude angle, for example, every [number] [time range]. Divide the altitude angle range, and keep the solar altitude angle no lower than 100°C. The sample. Within each altitude angle interval, the sample with the highest clear sky index was selected. The samples are used as approximate clear sky samples, and the median of the sky images of these samples is taken pixel by pixel to obtain a clear sky reference image under the corresponding solar altitude conditions. .

[0094] The output of this step is a clear sky reference image. This is used to subsequently determine the degree of deviation of the current sky image from a clear sky state.

[0095] Step S34: Generate an intermediate mask for the cloud region based on the current image and the clear sky reference image;

[0096] This step is used to obtain pixel-level cloud region determination results. The input is the current sky image. Reference image of clear sky under similar solar altitude conditions. .

[0097] For the coordinates in the image are The pixel, whose red channel value is recorded. The value of the blue channel is Calculate the normalized red-blue difference:

[0098] ,

[0099] Meanwhile, for clear sky reference images Calculation of pixels with the same coordinates And calculate the difference between the current image and the clear sky reference image:

[0100] ,

[0101] Then, based on the above two quantities, a cloud region intermediate mask is generated. :

[0102] ,

[0103] in, This indicates that the value is 1 when the condition is true and 0 when the condition is false. Represents pixels Pixels identified as cloud-related This indicates that the pixel was not identified as a cloud-related pixel. In this embodiment, 0.05 is acceptable. 0.175 is acceptable.

[0104] The output of this step is the intermediate mask for the cloud region. This mask is used for subsequent calculations of cloud cover intensity, cloud cover amount in the near-solar region, cloud centroid, and cloud cluster motion characteristics.

[0105] Step S35: Calculate the cloud coverage intensity and the cloud coverage amount in the near-sun region using the intermediate mask of the cloud region;

[0106] This step converts the pixel-level cloud region intermediate mask into cloud-overlaid image derived physical features. The input is the cloud region intermediate mask. And the position of the sun.

[0107] First, calculate the original cloud cover intensity in the entire image:

[0108] ,

[0109] in, and These represent the image height and width, respectively. This formula indicates the proportion of pixels identified as cloud-related to the total number of pixels in the image, thus reflecting the original cloud cover intensity in the entire image.

[0110] In this embodiment, cloud cover status quantities can also be generated based on preset thresholds. For example, when the original cloud cover intensity is not higher than 0.045, it is recorded as an approximately cloudless state; when the original cloud cover intensity is not lower than 0.35, it is recorded as a high cloud cover state. At the same time, the cloud cover amount in the near-solar region is calculated based on the local area near the sun's position to describe the cloud cover situation around the solar disk.

[0111] The output of this step includes the original cloud cover intensity, cloud cover state quantity, and cloud cover amount in the near-solar region. Among them, the continuous cloud cover intensity and cloud cover amount in the near-solar region can be regarded as part of the cloud cover and cloud motion characteristics.

[0112] Step S36: Estimate cloud cluster movement characteristics based on cloud region location at consecutive time points;

[0113] This step is used to extract the motion trend of cloud regions from continuous sky images. The input is the intermediate mask of the cloud region at different times. Cloud region location and timestamp.

[0114] During processing, first according to Calculate the centroid of the cloud region using the pixel coordinates with a value of 1. If cloud-related pixels exist in a local area near the sun, the cloud centroid in the near-sun region is also calculated. Then, based on the changes in the centroid coordinates over consecutive time intervals, the motion velocity of the cloud region in the image coordinates is estimated. The time interval between two adjacent images is obtained. ,but:

[0115] ,

[0116] ,

[0117] in, and These represent the motion components of the cloud region's centroid in the horizontal and vertical directions of the image, respectively. , Let x and y represent the x and y coordinates of the cloud region's centroid in the image coordinate system at time t, respectively. , Let x and y represent the x and y coordinates of the cloud region centroid in the image coordinate system at time t-1, respectively. This indicates the actual observation time corresponding to the previous image.

[0118] Further calculations were made regarding the speed and direction of cloud movement:

[0119] ,

[0120] ,

[0121] in, Indicates the speed of cloud movement. This indicates the direction of cloud movement. For the near-solar region, the local cloud movement components can be calculated in the same way. Combining the consistency between the overall cloud movement and the cloud movement in the near-solar region, the local cloud coverage, and the number of local cloud pixels, a motion reliability index can also be formed.

[0122] The output of this step includes cloud motion velocity, cloud motion direction, cloud motion components in the near-solar region, and motion confidence. These quantities are used to form cloud cover and cloud motion characteristics, characterizing how clouds move.

[0123] Step S37: Calculate the relative occlusion geometry between the cloud region and the sun's position;

[0124] This step describes the relative position of the cloud region and the sun. The input is the centroid of the cloud region. The position of the center of the sun The speed of cloud movement and the apparent speed of the sun's movement in the image coordinates.

[0125] First, calculate the relative displacement and relative distance between the cloud region and the sun's position:

[0126] ,

[0127] ,

[0128] ,

[0129] in, and This represents the lateral and longitudinal relative displacement of the cloud region from the sun's position in the image coordinate system. This represents the image coordinate distance between the two. Further, let the speed of the sun's movement in the image coordinates be denoted as... Calculate the approach velocity of the cloud region relative to the direction of the sun:

[0130] ,

[0131] in, This represents the relative motion component of the cloud region along the direction from the cloud to the sun.

[0132] When the value is positive, it indicates that the cloud region has a tendency to move closer to the sun's position.

[0133] When the value is negative, it indicates that the cloud region tends to move away from the sun.

[0134] The output of this step includes the relative displacement, relative distance, and approach velocity between the cloud and the sun. These quantities are used to form the cloud-sun relative occlusion geometry.

[0135] Step S38: Construct the estimated occlusion / removal time and time-division occlusion physical characteristics based on the relative motion;

[0136] This step converts the cloud-sun relative geometry into image-derived occlusion physical features related to future predicted time distances. The input is the distance from the cloud region to the sun's position. and approximation speed .

[0137] when When the value exceeds a preset positive threshold, calculate the estimated occlusion time:

[0138] ,

[0139] when When the time is less than the preset negative threshold, calculate the estimated occlusion release time:

[0140] ,

[0141] in, This represents the estimated timeframe for occlusion to occur in the context of image coordinates. These represent the estimated timeframe for occlusion removal in the image coordinates sense. Both quantities are derived from the cloud-sun relative motion estimate in the image coordinates and are used as occlusion geometric and physical features.

[0142] For future prediction time interval Furthermore, the expected occlusion time is mapped to bounded occlusion physical characteristics:

[0143] ,

[0144] in, This represents the Sigmoid function. This represents the scaling factor. In this embodiment, 0.3 is acceptable. Take the future Minutes, 15 minutes, or 30 minutes. This quantity describes the relative strength of occlusion risk within a given prediction time interval and is a physical characteristic of occlusion.

[0145] The output of this step includes the estimated occlusion duration, the estimated occlusion release time, and the time-spacing occlusion physical characteristics. These quantities are used to form the cloud-sun relative occlusion geometry.

[0146] Step S39: Calculate the image-derived physical features of the solar disk's brightness, visibility, and appearance.

[0147] This step describes the visibility and appearance of the solar disk in the current image. The input is the current sky image. Image regions near the center of the sun.

[0148] During processing, the maximum value of the RGB three channels of each pixel is first taken as the luminance channel:

[0149] ,

[0150] Then, the highest quantile of the luminance channel is taken as the luminance threshold. In this embodiment, the quantile can be 0.98. This satisfies... The pixels with the best coordinates are selected as candidate pixels for the sun, and the average coordinates of these candidate pixels are used as the center position of the sun. A solar neighborhood is then constructed with the sun's center as the center. The radius of the solar neighborhood can be determined as 0.08 times the length of the shorter side of the image.

[0151] Calculate the solar visibility index based on the solar neighborhood:

[0152] ,

[0153] in, Represents the solar neighborhood. This represents the number of pixels in that neighborhood. This formula is used to describe the proportion of bright pixels in the solar neighborhood.

[0154] Furthermore, the solar neighborhood is divided into the solar core region and the annular region, and the average brightness of the solar core region is calculated separately. and the average brightness of the solar ring region And construct an index for the relationship between the brightness of the solar disk:

[0155] ,

[0156] in, This quantity describes the brightness relationship between the Sun's core region and its surrounding ring region. It is one of the image-derived physical features of the solar disk's appearance, and its meaning is limited to describing brightness relationships.

[0157] The outputs of this step include the solar visibility index, solar visibility trend, solar core brightness, solar zonal brightness, solar disk brightness relationship index, solar core saturation ratio, and solar central motion. These quantities are used to form the appearance characteristics of the solar disk.

[0158] Step S310: Obtain image-derived physical diagnostic features; the image-derived physical diagnostic features include: solar geometric features, cloud coverage and cloud movement features, cloud-sun relative occlusion geometric features, and solar disk appearance features.

[0159] After the above processing, the first The time-series sky images and their adjacent time series are converted into image-derived physical diagnostic features. This object includes the following four categories of information:

[0160] 1. Solar geometric features, used to describe the sun's position and its changing trends;

[0161] 2. Cloud coverage and cloud motion characteristics, used to describe overall cloud coverage, cloud coverage in the near-solar region, cloud movement direction, cloud movement speed, and motion reliability;

[0162] 3. Relative cloud-sun occlusion geometry, used to describe the relative displacement, relative distance, approach velocity, expected occlusion duration, and expected occlusion release time between the cloud region and the sun;

[0163] 4. Solar disk appearance characteristics, used to describe the relationship between solar visibility, solar core brightness, solar ring brightness, solar disk brightness, and solar disk appearance variation trends.

[0164] Taking an engineering sample as an example, for a sky image at a certain moment, the system first processes it into... A pixel image is generated, and the sun's position is calculated based on the timestamp. This image is then compared with a clear-sky reference image under similar sun altitude conditions to obtain a mask for the cloud region. Then by The cloud cover intensity, cloud centroid, and cloud motion are calculated. Simultaneously, the solar visibility index and solar disk brightness relationship are calculated within the solar neighborhood. Finally, by combining the relative displacement and relative motion between the cloud region and the sun's position, the expected duration of occlusion and future... The occlusion physical characteristics are calculated for minutes, 15 minutes, and 30 minutes. In this way, the same frame of sky image forms a set of image-derived physical diagnostic features organized around the mechanism of solar occlusion.

[0165] Step S4: Based on the physical mechanism of the image-derived physical diagnostic features, group the images to obtain the grouped image-derived physical quantities; the grouping includes at least: solar geometry group, cloud motion group, solar occlusion geometry group, and solar disk appearance group.

[0166] This step is one of the key steps in this invention. A common practice in the prior art is to directly concatenate multiple image-derived physical quantities into a single long vector and then input it into the same encoding path—essentially, direct tiling encoding. While this also "uses physical features," information from different physical mechanisms is already mixed together before entering the model. The model struggles to distinguish which variables describe the sun's position, which describe cloud movement, which describe impending occlusion, and which describe changes in the appearance of the solar disk after it has already been occluded.

[0167] To address the aforementioned problems, this invention groups image-derived physical diagnostic features according to physical mechanism rather than field source or storage order. Preferably, it includes at least the following four groups:

[0168] 1. Solar geometry group;

[0169] 2. Cloud Sports Group;

[0170] 3. Solar shading geometry;

[0171] 4. Solar disk appearance assembly.

[0172] The above grouping method corresponds to a complete shading mechanism chain: the solar geometry group answers "Where is the sun?", the cloud motion group answers "How do the clouds move?", the solar shading geometry group answers "When might the clouds shade or leave the sun?", and the solar disk appearance group answers "What state is the solar disk currently in which it is shaded?". Therefore, this invention adopts a method of first organizing these features according to the physical action chain, and then proceeding to the subsequent encoding and fusion steps.

[0173] The output of this step is a mechanism grouping object. This object is used in the next step to construct and encode time window samples.

[0174] Step S5, using the current prediction time Using a point as an anchor, sky images, measured power, image-derived physical diagnostic features, and corresponding clear-sky reference power are extracted from a continuous historical period preceding it to form a historical window sample. ;

[0175] After completing sample alignment and physical grouping, the samples at discrete time points need to be organized into historical time windows that can be used to predict future power. The purpose of this is to allow the model to see not only the static state at the current time point, but also the continuous changes in image, power, and physical diagnostic information over a period of time prior to the current time point.

[0176] Based on the current prediction time Using a point as an anchor, sky images, measured power, image-derived physical diagnostic features, and corresponding clear-sky reference power are extracted from a continuous historical period preceding it to form a historical window sample. This historical window sample can be represented as:

[0177] .

[0178] in, Indicates time The corresponding complete history window sample; This indicates the number of sampling times contained within the history window; Indicates from time At the time A sequence of sky images; This represents the historical power sequence of photovoltaic power within the same time range; This represents a sequence of image-derived physical diagnostic features within the same time frame; This represents a clear-sky reference power sequence within the same time range. The formula indicates that, for the same prediction task, image, power, and physical diagnostic information are uniformly organized into a time-continuous window object for use by different subsequent branches.

[0179] For example, in the current preferred embodiment, the historical window length can be 45 minutes, and the sampling step size can be 1 minute. If the current prediction time... If it is 12:00, then This can be understood as a "historical sample window constructed around the prediction task of 12:00," which includes the sky image sequence from 11:16 to 12:00, the measured power sequence from 11:16 to 12:00, the physical diagnostic feature sequence within the same time range, and the corresponding clear-sky reference power sequence. This window object will be fed into the image branch, power branch, and physical mechanism branch in the next step. (The sky image sequence is input with image encoding, the photovoltaic historical power sequence is input with power encoding, and the image-derived physical diagnostic feature sequence is grouped according to physical mechanism and then input with the corresponding physical feature encoding).

[0180] The output of this step is a timing window object. .

[0181] Step S6: Encode the derived physical quantities of the grouped images to obtain the representation of each physical mechanism group; fuse the representations of each physical mechanism group with the image sequence representation and the power history sequence representation to obtain the fused representation object. ;

[0182] After obtaining the time-series window object, this invention extracts three types of information: one is visual time-series information in the sky image sequence, another is power change information in the photovoltaic historical power sequence, and the third is image physical diagnosis information grouped according to physical mechanisms.

[0183] In this invention, image sequences are encoded using both image coding and temporal coding to obtain image branch representations, while historical photovoltaic power sequences are encoded using temporal coding to obtain power branch representations. For physical diagnostic features, this invention employs grouping by physical mechanism, establishing a separate coding path for each mechanism group. It does not use a method of "all features sharing a single encoder."

[0184] Set time The set of physical diagnostic features is as follows:

[0185] .

[0186] in, Indicates time The complete set of image-derived physical diagnostic features; Indicates the number of mechanism groups; Indicates the first A feature subset of each mechanism group In the current preferred embodiment, These correspond to the solar geometry group, cloud motion group, solar occlusion geometry group, and solar disk appearance group, respectively. The purpose of this formula is to explicitly decompose "a whole package of physical features at the same moment" into "several mechanism groups with clear physical meanings under the same prediction task".

[0187] Subsequently, each mechanism group is time-coded separately, and the process can be represented as follows:

[0188] ,

[0189] in, Indicates the first The encoders corresponding to each mechanism group; Indicates the first The temporal characteristic sequence of each mechanism group within the historical window range; Indicates the first The temporal representation results obtained after encoding each mechanism group. This formula indicates that each mechanism group independently learns temporal patterns along its own encoding path, rather than allowing all physical features to interfere with each other within the same encoding path. The obtained representation results will be subsequently fused with image branch representations and power branch representations.

[0190] For example, if the current predicted time is still 12:00, and the historical window covers 11:16 to 12:00, then:

[0191] This can be understood as a sequence of solar geometric features within these 45 moments, such as solar altitude angle, solar azimuth angle, apparent zenith angle and their rate of change;

[0192] This can be understood as a sequence of cloud motion characteristics within these 45 moments, such as cloud coverage ratio, cloud speed, cloud direction, and cloud motion status in the near-sun region.

[0193] It can be understood as a sequence of geometric features of cloud-sun occlusion within these 45 moments, such as the relative displacement, relative distance, approach speed and expected occlusion time of the cloud to the sun;

[0194] This can be understood as a sequence of solar disk appearance characteristics over these 45 moments, such as the solar visibility index, solar core brightness, solar ring brightness, and the relationship between solar disk brightness and other indicators.

[0195] After encoding separately, we can obtain , , and These four symbols are not the final output, but rather "four intermediate representations of the same prediction task under four different physical mechanisms".

[0196] In obtaining image branch representation Power branch characterization and the characterization of each mechanism group Then, a unified fusion process is performed to obtain the final fused representation. This fusion process can be represented as:

[0197] .

[0198] in, This represents the unified representation after fusion; Indicates the fusion function; Represents image sequence characterization; Represents the power history sequence; Indicates the first The encoding results of the mechanism group. This formula means that information from different sources is not directly mixed at the original input layer, but rather fused in a higher-level semantic space after each source has completed its own representation. This is more conducive to preserving the structural features of image information, power history information, and occlusion physical mechanism information.

[0199] For example, for the prediction task of 12:00, This can be understood as a "compressed representation of the changes in the sky image between 11:16 and 12:00". This can be understood as a "compressed representation of the power change trend from 11:16 to 12:00", while to These correspond to compressed representations of "changes in solar position," "changes in cloud movement," "changes in occlusion geometry," and "changes in solar disk appearance," respectively. The result after fusing these representations is... Essentially, this is the comprehensive judgment basis formed by the model for the prediction task of 12:00. This fused representation is used in the next step to output the power prediction results for multiple future prediction time intervals.

[0200] The output of this step is the fusion representation object. .

[0201] Step S7, the fused representation object Input the prediction model to output photovoltaic power prediction results for multiple future prediction time intervals;

[0202] The prediction model in this embodiment is a pre-trained prediction model, and its training method is existing technology, which will not be described in detail in this embodiment.

[0203] The prediction model in this embodiment includes multiple prediction heads, which can output multiple prediction results simultaneously, namely, prediction results for the next 5 minutes, 15 minutes, and 30 minutes.

[0204] After obtaining the fused representation, it is fed into the prediction model. This is equivalent to mapping the comprehensive information contained in the fused representation, such as weather conditions, shading conditions, and historical power changes, into a power prediction value for a future target time interval. This invention outputs photovoltaic power prediction results for multiple future prediction time intervals based on this representation, rather than outputting a prediction value for a single future moment.

[0205] Let the predicted time interval set be... The final predicted output can then be expressed as:

[0206] ,

[0207] in, Indicates time The corresponding multi-step prediction result vector; Indicates the number of predicted time intervals; Indicates the first One predicted time interval; Indicates future time The predicted value of photovoltaic power, This formula means: for the same current moment... This invention can simultaneously provide multiple future time-distance prediction results to meet the needs of different scheduling and control scenarios for prediction lead time.

[0208] In the current preferred embodiment, the prediction time interval can be 5 minutes, 15 minutes, or 30 minutes in the future, that is:

[0209] ,

[0210] For example, if the current prediction time At 12:00, the output of a certain prediction was: This means that, based on sky images, photovoltaic power, and physical diagnostic information from the 45 minutes prior to 12:00, the model predicts that the power output of the power station at 12:05, 12:15, and 12:30 will be approximately 428.6 kW, 401.2 kW, and 365.7 kW, respectively. This multi-step prediction result can be directly used for short-term dispatching, early warning analysis, energy storage collaborative control, or grid-connected operation management.

[0211] The output of this step is the final predicted output object. This object is the final output result in the main solution of this invention.

[0212] The photovoltaic power prediction method in this embodiment improves the completeness of information used in short-term prediction by introducing joint modeling of sky image sequences and historical photovoltaic power sequences.

[0213] This invention introduces sky image sequences as synchronous observation information in addition to historical photovoltaic power sequences, enabling the prediction process to utilize both existing power variation information from the power plant and changes in external sky conditions. In this way, when making short-term predictions, the model can retain the temporal continuity of historical power while also being aware of prior cues from cloud distribution, solar position, and local brightness changes, thus providing a more complete input basis for subsequent predictions.

[0214] From the current implementation, this joint modeling framework not only supports multimodal inputs but also supports unified output across multiple future prediction time intervals. Existing training and testing metric files have recorded the average absolute error (MAE) between predicted and measured power, the root mean square error (RMSE) of the predicted error, and other metrics for 5-minute, 15-minute, and 30-minute prediction time intervals, respectively. These metrics reflect the overall magnitude of the prediction error. The normalized error metric (NRMSE) is used to compare prediction errors at different power scales. The mean percentage error (MAPE) of the prediction error relative to the measured power demonstrates that the multi-step prediction results of this invention are independently quantifiable, separately evaluable, and continuously trackable. This effect is beneficial for subsequent engineering deployment and operational evaluation for different scheduling lead times.

[0215] For example, in the current comparative implementation, compared with the scheme that only uses historical power sequences, the RMSE of the 30-minute prediction decreased from 1.5414 to 1.4274 after introducing sky image sequences, indicating that visual observation information has a supplementary role for longer prediction intervals.

[0216] 2. By constructing image-derived physical diagnostic features, the solar occlusion process can participate in prediction as an explicit intermediate object, thus improving the interpretability of the scheme.

[0217] This invention goes beyond simply "using raw sky images"; it further constructs solar geometric features, cloud movement features, cloud-sun relative occlusion geometric features, and solar disk appearance features from the sky images. These intermediate objects explicitly express the occlusion mechanisms in the images, enabling the model not only to "see image changes" but also to utilize information closer to actual physical processes, such as "where the sun is, how the clouds move, when occlusion might occur, and the current extent to which the solar disk is occluded."

[0218] Current experimental results show that adding image-derived physical diagnostic features to the joint input of image and photovoltaic history has a positive effect on 15-minute and 30-minute predictions, while the 5-minute prediction shows a slight degradation. This indicates that such physical diagnostic features are more helpful in describing the occlusion evolution process at slightly longer lead times. Although this effect is not entirely consistent across different prediction time intervals, it demonstrates that explicitly extracting occlusion-related physical information from images can provide valuable supplementary information for mid-range predictions, while improving the physical interpretability of the entire scheme.

[0219] 3. By grouping and encoding the image-derived physical diagnostic features according to their physical mechanisms, the interference caused by heterogeneous information mixing is reduced, which is the most critical technical effect of this invention.

[0220] The most significant technical effect of this invention lies in the organization and utilization of image-derived physical diagnostic features. Existing methods commonly involve directly tiling physical features from different sources and mechanisms and inputting them into the same encoding path. In this approach, information such as the sun's position, cloud movement, occlusion formation conditions, and the current appearance of the solar disk are mixed together during the input stage. The model needs to manually distinguish the hierarchical relationships between these pieces of information, which can easily lead to mutual interference.

[0221] This invention first divides these features into a solar geometry group, a cloud motion group, a solar occlusion geometry group, and a solar disk appearance group, and then performs temporal encoding on each group. Through this process, information from different physical mechanisms first forms a representation along their respective pathways, and then is fused at a higher level. This helps maintain the consistency of the mechanisms within each type of information and also helps reduce mutual interference between heterogeneous features.

[0222] This point has been clearly supported by current controlled comparative experiments. Under the condition of a fixed feature set, a comparison of the direct tiling coding scheme and the physical mechanism-based grouping coding scheme using multiple random seeds shows that:

[0223] -NRMSE_rated decreased from 4.250142±0.448794% to 3.848608±0.083263%;

[0224] -RMSE decreased from 1.279293±0.135087 to 1.158431±0.025062;

[0225] - The 5-minute forecast RMSE decreased from 1.033192 to 0.921401;

[0226] The 15-minute forecast RMSE decreased from 1.262402 to 1.092354;

[0227] The 30-minute forecast RMSE decreased from 1.499262 to 1.407997.

[0228] The results above demonstrate that, provided the input feature set remains consistent, the improvement primarily stems from the feature organization method itself, rather than from simply changing feature fields. Simultaneously, the fluctuation range under multiple random seeds is significantly narrowed, indicating that this structured organization method has a positive effect on the stability of the results.

[0229] Table 1. Key results of the direct tiling coding scheme for image-derived physical diagnostic features and the scheme of grouping and coding physical diagnostic features by physical mechanism under controlled comparison conditions.

[0230] NRMSE_rated(%) 4.250142±0.448794 3.848608±0.083263 RMSE 1.279293±0.135087 1.158431±0.025062 RMSE_5min 1.033192 0.921401 RMSE_15min 1.262402 1.092354 RMSE_30min 1.499262 1.407997

[0231] refer to Figure 6 , Figure 6 The diagram illustrates the differences in target power, absolute error, and diagnostic metrics between the direct tiling encoding scheme and the scheme that groups and encodes based on physical mechanisms under typical cloud cover scenarios. As shown in the figure, the direct tiling encoding scheme significantly overestimates future power during this period, while the scheme that groups and encodes based on physical mechanisms is closer to the actual power. The corresponding 15-minute absolute error decreases from 7.4616kW to 2.5593kW, an improvement of 4.9023kW. Combined with the diagnostic metric curve below, it can be seen that there is significant cloud cover disturbance during this period. Therefore, this case demonstrates that the proposed solution can more effectively utilize image-derived physical diagnostic features in specific cover events, thereby improving prediction accuracy. This case can be used to illustrate the improvement effect of the proposed solution in specific cover events.

[0232] 4. The improvement effect of the present invention can still be achieved when the number of parameters is controlled, indicating that the benefits do not mainly come from the increase in model capacity.

[0233] When comparing the image-derived physical diagnostic feature direct tiling encoding scheme with the physical diagnostic feature grouping and encoding scheme based on physical mechanisms, the current implementation also explicitly checks the difference in model parameter count. The current summary results show that the difference in parameter count between the two schemes is approximately 0.007857630551686588, or about 0.79%, which is within a pre-controlled, small range and does not result in significant model capacity expansion.

[0234] Therefore, the aforementioned reduction in error and improvement in stability cannot be simply attributed to "the model becoming larger." This further illustrates that the effects brought about by this invention are mainly related to the organizational method of "grouping and encoding image-derived physical diagnostic features according to physical mechanisms," rather than the superficial benefits directly obtained by increasing the number of parameters.

[0235] 5. In weather scenarios closely related to the shielding process, the advantages of this invention can be applied to specific cases, demonstrating good engineering relevance.

[0236] Based on the current comparison results of weather slices, the structured grouping coding scheme of this invention performs better than the direct tiling coding scheme in partially cloudy and overcast weather scenarios. For clear weather scenarios, there is currently a lack of effective sample support, so no absolute extrapolation is made for them.

[0237] Furthermore, in the three case studies selected in the partially cloudy scenario, the improvement in absolute prediction error over 15 minutes reached 4.9023kW, 2.9352kW, and 2.6732kW, respectively. This demonstrates that the improvements of this invention are not only reflected in overall statistical indicators but also applicable to specific multi-cloudy events. For short-term prediction tasks where cloud obstruction disturbances are the primary source of problems, this kind of improvement, which can be applied to specific events, specific time periods, and specific prediction targets, better reflects the engineering practical value of this invention.

[0238] Example 2

[0239] This embodiment uses a short-term multi-step power prediction task for a single photovoltaic power station as an example. The sky images are historical sky observation images with timestamps, and the photovoltaic power records are the historical measured power data of the corresponding power station. For ease of explanation, this embodiment uses the following power station parameters: latitude 37.424107, longitude -122.174199, altitude 30m, module tilt angle 22.5°, module azimuth angle 195°, DC rated power 30100W, power temperature coefficient -0.004 / °C, and time zone US / Pacific.

[0240] The software environment uses Python 3.9.23, PyTorch 2.5.1, and torchvision 0.20.1, combined with libraries such as pandas, numpy, and pvlib to complete data processing, model training, and evaluation. The computing device uses a CUDA-enabled computing platform; when the GPU is unavailable, execution can fall back to the CPU. The specific model of the sky imaging device is not a necessary technical feature for the establishment of this invention, and therefore is not limited in this embodiment.

[0241] The common data processing flow in this embodiment is as follows.

[0242] First, read the timestamped sky images and photovoltaic power records, and standardize the time base. The root directory for the images is datas / , and the image file format is ** / *.jpg. The power records use single-site historical data files, with the timestamp column as Date and the power column as Huang_E4102_kW.

[0243] Sky images such as Figure 3-5 The above, Figure 3-5 These are sky images taken at the same location but at different times using the same camera.

[0244] Second, clear-sky reference power was pre-calculated. The Ineichen clear-sky model was used for the calculation, with the open_rack_glass_glass temperature model employed. The ambient temperature was set to 20°C, the wind speed to 1 m / s, and the chunk_size parameter to 20000. The calculation results were output as a time-by-time clear-sky reference power table, serving as the basis for subsequent sample alignment and physical diagnostics.

[0245] Third, construct image-power aligned samples. Using image timestamps as anchors, each image is matched with the power record of the corresponding or most recent time, and the measured power and clear-sky reference power are written into a unified sample table to form aligned samples. Subsequently, the samples are divided into training set, validation set and test set in chronological order, with a division ratio of 0.70:0.15:0.15.

[0246] Fourth, construct image-derived physical diagnostic features. After uniformly processing the images to 224×224 pixels, further calculate image-derived physical features related to the solar occlusion process. To improve feature stability, first construct a clear-sky reference image library based on the training set. The main parameters used in constructing the image physical features include: csl_bin_width_deg=10.0, csl_min_elevation_deg=5.0, csl_top_percent=0.05, nrbr_threshold=0.05, dnrbr_threshold=0.175, cloud_cover_gate_low=0.045, cloud_cover_gate_high=0.35, occ_sigmoid_scale=0.3, svi_quantile=0.98, svi_roi_radius_frac=0.08, and svi_trend_window_minutes=10. Here, csl_bin_width_deg=10.0 indicates that the clear-sky reference image library is divided into intervals of 10° based on the solar altitude angle; csl_min_elevation_deg=5.0 indicates that only images with a solar altitude angle of at least 5° are selected. The samples are used to construct the clear sky reference image to avoid interference from samples taken at night or with low sun altitude. `csl_top_percent=0.05` indicates that the top 5% of samples with higher clear sky indexes within each sun altitude angle interval are selected as approximate clear sky samples. `nrbr_threshold=0.05` and `dnrbr_threshold=0.175` are the cloud region determination thresholds for the normalized red-blue difference and its difference relative to the clear sky reference image, respectively, used to generate intermediate masks for cloud regions. `cloud_cover_gate_low=0.045` and `cloud_cover_gate_high=0.35` are used to distinguish between low and high cloud coverage states, respectively. `occ_sigmoid_scale=0.3` is the Sigmoid scale coefficient when the expected occlusion time is mapped to the time-separated occlusion physical features. `svi_quantile=0.98` indicates that the 98th quantile of the luminance channel is used to determine the threshold for candidate bright pixels of the sun; `svi_roi_radius_frac=0.08` indicates that the radius of the solar neighborhood is determined as 8% of the shorter side length of the image; `svi_trend_window_minutes=10` indicates that a 10-minute time window is used when calculating the trend of solar visibility changes. After processing, intermediate results such as cloud cover, cloud movement, solar visibility, solar disk brightness relationship indicators, cloud-sun relative geometric relationship, and physical characteristics of occlusion in the next 5, 15, and 30 minutes can be obtained.

[0247] Fifth, construct a multi-step prediction window. Around the current prediction time, extract historical image sequences, historical power sequences, and historical physical diagnostic feature sequences from the preceding 45 minutes, with a sampling step size of 1 minute. The prediction output step sizes are set to 5 minutes, 15 minutes, and 30 minutes.

[0248] The following embodiments are all based on the above common data processing flow, and differ only in the input modality, physical feature organization method and model branch structure.

[0249] Example 1: Multi-step prediction example using only historical photovoltaic power series

[0250] This embodiment illustrates the basic implementation method of short-term multi-step prediction using only historical photovoltaic power sequences.

[0251] (1) Input data

[0252] This embodiment uses only historical photovoltaic power sequences as model input, without using the sky image branch or the image-derived physical diagnostic feature branch. The historical time window length is 45 minutes, the sampling step size is 1 minute, and the prediction time intervals are 5 minutes, 15 minutes, and 30 minutes in the future.

[0253] (2) Model parameters

[0254] In this embodiment, the training rounds are 100, the batch size is 10, and the learning rate is... The weight decay coefficient is The early stopping strategy has a patience round count of 10 and a minimum training round count of 10. Power history encoding uses a gated recurrent unit temporal encoding structure with a hidden dimension of 128.

[0255] (3) Operation process

[0256] First, read historical power window samples from the training set, validation set, and test set;

[0257] Secondly, the 45 historical power points within each time window are input into the power timing encoder to obtain the power history representation;

[0258] Finally, based on this characterization, the predicted power values ​​for the next 5 minutes, 15 minutes, and 30 minutes are output simultaneously, and the mean absolute error, root mean square error, normalized root mean square error, and mean absolute percentage error are calculated on the test set.

[0259] (4) Results

[0260] The results of this embodiment in the current test log are as follows:

[0261] 1. The overall RMSE is 1.1780;

[0262] 2. The overall NRMSE is 0.0839;

[0263] The 3.5-minute forecast RMSE is 0.8088;

[0264] The RMSE for the 4.15-minute forecast is 1.0644.

[0265] The RMSE for the 5.30-minute forecast is 1.5414.

[0266] This embodiment can complete multi-step photovoltaic power prediction, but since it only relies on historical power sequences, it lacks direct observation of external disturbances caused by changes in sky conditions in future periods. Therefore, there is room for further improvement over longer prediction time intervals.

[0267] Example 2: Multimodal multistep prediction using sky image sequences and historical photovoltaic power sequences

[0268] This embodiment adds a sky image branch to the existing embodiment 1 to illustrate the implementation of short-term multi-step prediction under multimodal input.

[0269] (1) Input data

[0270] This embodiment uses both sky image sequences and historical photovoltaic power sequences as input. The image window length is 45 minutes, with one frame per minute; the historical power window length is also 45 minutes. The prediction time intervals are 5 minutes, 15 minutes, and 30 minutes in the future.

[0271] (2) Model parameters

[0272] The number of training epochs, batch size, learning rate, weight decay, and early stopping strategy remain consistent with Example 1. The image branch uses the EfficientNet-B0 convolutional visual coding network to extract frame-by-frame image features, which are then encoded using an image temporal encoder. The image temporal encoding has a hidden dimension of 256; the power temporal encoding has a hidden dimension of 128. The image branch and the power branch are fused using a gated fusion module.

[0273] (3) Operation process

[0274] First, 45 frames of sky images and 45 historical power points are read from each time window;

[0275] Secondly, each frame of sky image is preprocessed and features are extracted, and image branch representation is formed by an image temporal encoder;

[0276] Next, the historical power sequence is time-series encoded to form a power branch representation;

[0277] Subsequently, the image branch representation and the power branch representation are fused.

[0278] Finally, based on the unified representation after fusion, the prediction results for the next 5 minutes, 15 minutes, and 30 minutes are output, and various evaluation indicators are calculated on the test set.

[0279] (4) Results

[0280] The results of this embodiment in the current test log are as follows:

[0281] 1. The overall RMSE is 1.1990;

[0282] 2. The overall NRMSE is 0.0854;

[0283] 3. The 5-minute forecast RMSE is 0.9635;

[0284] 4. The 15-minute forecast RMSE is 1.1605;

[0285] 5. The 30-minute forecast RMSE is 1.4274.

[0286] Compared to Example 1, the error in this example at a 30-minute prediction time interval decreased from 1.5414 to 1.4274, indicating that sky image information has a supplementary effect on longer prediction time intervals. However, considering the overall indicators and performance at shorter time intervals, the gains still depend on the prediction time interval when only the original sky image branch is introduced.

[0287] Example 3: Multimodal, multi-step prediction example that introduces image-derived physical diagnostic features and groups and encodes them according to physical mechanisms.

[0288] This embodiment is the main embodiment of the present invention.

[0289] (1) Input data

[0290] This embodiment uses sky image sequences, historical photovoltaic power sequences, and image-derived physical diagnostic feature sequences as inputs. The image-derived physical diagnostic features are divided into four groups according to their physical mechanisms:

[0291] 1. Solar geometry group;

[0292] 2. Cloud Sports Group;

[0293] 3. Solar shading geometry;

[0294] 4. Solar disk appearance assembly.

[0295] In this embodiment, the physical input dimension is 31-dimensional, the historical time window length is 45 minutes, the sampling step size is 1 minute, and the prediction time interval is 5 minutes, 15 minutes, and 30 minutes in the future.

[0296] (2) Model parameters

[0297] In this embodiment, the training rounds are 100, the batch size is 10, and the learning rate is... The weight decay coefficient is The image temporal coding hidden dimension is 256, the power temporal coding hidden dimension is 128, the physical mechanism coding hidden dimension is 128, the fusion dimension is 256, and the prediction head hidden dimension is 256. Automatic mixed precision is used during training, and the data type is float16.

[0298] (3) Operation process

[0299] First, the clear-sky reference power pre-calculation, sample alignment, and image-derived physical diagnostic features were completed according to the common data processing flow.

[0300] This set of sky images illustrates typical cloudy scenarios at different times. These images allow for the extraction of physical diagnostic information related to power fluctuations from variations in solar neighborhood brightness, local cloud location, and sky cover.

[0301] Secondly, the image-derived physical diagnostic features are divided into solar geometry group, cloud motion group, solar occlusion geometry group, and solar disk appearance group.

[0302] Next, historical image sequences, historical power sequences, and physical feature sequences of each mechanism group are extracted from 45 minutes prior to each prediction time.

[0303] Subsequently, the image sequence is encoded frame by frame and temporally to obtain the image branch representation; the historical power sequence is encoded temporally to obtain the power branch representation; and the four physical mechanism groups are encoded temporally to obtain the four mechanism group representations.

[0304] Then, the image branch representation and the power branch representation are fused together, and the physical mechanism branch representation is incorporated to obtain a unified fused representation.

[0305] Finally, based on the unified fusion representation, the predicted power values ​​for the next 5, 15, and 30 minutes are output.

[0306] (4) Results

[0307] In the current specific implementation record, the results of this embodiment are as follows:

[0308] 1. The overall RMSE is 1.1636;

[0309] 2. The overall NRMSE is 0.0829;

[0310] The 3.5-minute forecast RMSE is 0.8912;

[0311] The RMSE for the 4.15-minute forecast is 1.0956;

[0312] The 5.30-minute predicted RMSE is 1.4378.

[0313] Table 2. Comparison of key prediction results for Examples 1, 2, and 3

[0314] RMSE 1.1780 1.1990 1.1636 NRMSE 0.0839 0.0854 0.0829 RMSE_5min 0.8088 0.9635 0.8912 RMSE_15min 1.0644 1.1605 1.0956 RMSE_30min 1.5414 1.4274 1.4378

[0315] This embodiment illustrates that, based on multimodal input, organizing image-derived physical diagnostic features according to physical mechanisms, encoding them separately, and then participating in fusion can result in a clearer image.

[0316] A clear mechanism characterization path was obtained, and relatively stable prediction results were achieved.

[0317] Comparative Example 1: Comparative Example of Direct Tile Encoding of Image-Derived Physical Diagnostic Features

[0318] This comparative example illustrates the impact of "feature organization method" on prediction results.

[0319] (1) Input data

[0320] This comparative example uses the same sky image sequence, historical photovoltaic power sequence, and the same set of image-derived physical diagnostic features as Example 3. To ensure a fair comparison, the input feature set remains consistent, and the prediction time intervals are also 5 minutes, 15 minutes, and 30 minutes in the future.

[0321] (2) Model parameters

[0322] Except for the organization of physical features, the main training parameters in this comparative embodiment are consistent with those in Embodiment 3. The physical input dimension is also 31, the number of training rounds is also 100, the batch size is also 10, and the learning rate is also the same. The physical branch hiding dimension is also 128. The number of model parameters differs from that of Example 3 by approximately 0.79%, and there is no significant expansion of model capacity.

[0323] (3) Operation process

[0324] First, the clear-sky reference power construction, sample alignment, and image-derived physical diagnostic feature extraction were completed following the same procedure as in Example 3.

[0325] Secondly, instead of grouping according to physical mechanisms, all image-derived physical diagnostic features are directly spliced ​​into a flat feature sequence;

[0326] Next, the entire flat physical feature sequence is fed into the same physical temporal encoder;

[0327] Subsequently, the results are fused with the image branch representation and the power branch representation, and the prediction results for the next 5 minutes, 15 minutes and 30 minutes are output.

[0328] (4) Results

[0329] The results of this comparative embodiment are as follows in the corresponding implementation record:

[0330] 1. The overall RMSE is 1.2563;

[0331] 2. The overall NRMSE is 0.0895;

[0332] The 3.5-minute forecast RMSE is 1.0158;

[0333] The 4.15-minute RMSE forecast is 1.2291;

[0334] The RMSE for the 5.30-minute forecast is 1.4806.

[0335] Compared to Example 3, under the condition that the input feature set remains unchanged and the difference in the number of parameters is small, the overall RMSE of this comparative example is higher, and the RMSE at the three prediction time intervals is also higher than that of Example 3. This result shows that the improvement mainly comes from the technical arrangement of grouping physical features according to physical mechanisms and encoding them separately, rather than from simply increasing the number of features or increasing the model size.

[0336] Comparative Example 2: Controlled Comparative Example under Multiple Random Seed Conditions

[0337] To further eliminate the randomness of single training results, this comparative embodiment compares the "image-derived physical diagnostic feature direct tiling encoding scheme" and the "image-derived physical diagnostic feature grouping and encoding scheme according to physical mechanism" using multiple random seeds, while keeping the set of occlusion-related physical diagnostic features unchanged.

[0338] (1) Comparison settings

[0339] This comparative example uses three random seeds: 11, 22, and 33. Both schemes use the same sky image sequence, historical photovoltaic power sequence, and the same set of image-derived physical diagnostic features, with prediction time intervals of 5 minutes, 15 minutes, and 30 minutes in advance.

[0340] (2) Results

[0341] The results of the three random seeds are summarized below.

[0342] 1. The overall RMSE of the direct tiling encoding scheme is 1.2793, and the overall NRMSE is 0.0911;

[0343] 2. The overall RMSE of the mechanism block coding scheme is 1.1584, and the overall NRMSE is 0.0825;

[0344] 3. The RMSE of the direct tiling coding scheme at 5 minutes, 15 minutes, and 30 minutes are 1.0332, 1.2624, and 1.4993, respectively;

[0345] 4. The RMSE predictions for the 5-minute, 15-minute, and 30-minute timeframes of the mechanism-based block coding scheme are 0.9214, 1.0924, and 1.4080, respectively.

[0346] (3) Conclusion

[0347] The above comparison results of multiple random seeds show that, under the premise of keeping the input feature set consistent, the scheme of grouping and encoding according to physical mechanism is better than the direct tiling encoding scheme in terms of overall error and each prediction time interval error, indicating that the structured organization method has a more stable technical effect.

[0348] Comparative Example 3: Case-level Comparative Example under Multi-Cloud Obscuration Scenarios

[0349] To further illustrate the engineering effect of Example 3 in specific occlusion events, three 15-minute prediction cases were selected for comparison in some multi-cloud scenarios.

[0350] (1) Case selection rules

[0351] The case selection criteria are as follows: in the partial cloudy weather subset, the cases are sorted by the difference between the 15-minute average absolute error of the direct flat coding scheme and the 15-minute average absolute error of the mechanism group coding scheme, and the time interval between adjacent cases is controlled to be no less than 30 minutes.

[0352] (2) Case moment

[0353] The three case times obtained from the screening are as follows:

[0354] 1.2019-10-16T21:49:10+00:00;

[0355] 2.2019-10-16T20:58:10+00:00;

[0356] 3.2019-10-16T22:22:10+00:00.

[0357] (3) Results

[0358] In the three cases above, the error improvement of Example 3 compared to Comparative Example 1 in predicting the target at 15 minutes is as follows:

[0359] 1.49023kW;

[0360] 2.2.9352kW;

[0361] 3.2.6732kW.

[0362] (4) Conclusion

[0363] The results demonstrate that the advantages of Example 3 are not only reflected in the overall statistical indicators, but also can be applied to the prediction process of specific multi-cloud occlusion events, thus providing more intuitive support for the engineering application value of the present invention.

[0364] Example 3

[0365] refer to Figure 7 , Figure 7 This is a schematic diagram of the structure of a photovoltaic power prediction device according to this embodiment. The photovoltaic power prediction device 20 of this embodiment includes a processor 21, a memory 22, and a computer program stored in the memory 22 and executable on the processor 21. When the processor 21 executes the computer program, it implements the steps in the above method embodiments. Alternatively, when the processor 21 executes the computer program, it implements the functions of each module / unit in the above device embodiments.

[0366] For example, the computer program can be divided into one or more modules / units, which are stored in the memory 22 and executed by the processor 21 to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the photovoltaic power prediction device 20. For example, the computer program can be divided into the modules shown in Embodiment 2. The specific functions of each module are described in the working process of the device described in the above embodiments, and will not be repeated here.

[0367] The photovoltaic power prediction device 20 may include, but is not limited to, a processor 21 and a memory 22. Those skilled in the art will understand that the schematic diagram is merely an example of the photovoltaic power prediction device 20 and does not constitute a limitation on the device. It may include more or fewer components than illustrated, or combine certain components, or use different components. For example, the photovoltaic power prediction device 20 may also include input / output devices, network access devices, buses, etc.

[0368] The processor 21 can be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, off-the-shelf programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. The processor 21 is the control center of the photovoltaic power prediction device 20, connecting all parts of the photovoltaic power prediction device 20 via various interfaces and lines.

[0369] The memory 22 can be used to store the computer programs and / or modules. The processor 21 implements various functions of the photovoltaic power prediction device 20 by running or executing the computer programs and / or modules stored in the memory 22 and calling the data stored in the memory 22. The memory 22 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created based on the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory 22 may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart memory card, secure digital card, flash memory card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0370] If the integrated modules / units of the photovoltaic power prediction device 20 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by the processor 21, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory, random access memory, electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0371] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0372] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A photovoltaic power prediction method, characterized in that, It includes the following steps: Step S1: Obtain raw observation data; the raw observation data includes: time-stamped sky image data, time-stamped photovoltaic power data, and basic parameters related to the power station; Step S2: Calculate the clear-sky reference power under unobstructed conditions based on the original observation data. ,in, The current time is indicated; using the sky image timestamp as the anchor point, each image is matched with the photovoltaic power record of the corresponding time or the most recent time, and the measured power and clear sky reference power are written into the same sample record to form an image-power alignment sample. Step S3: Obtain image-derived physical diagnostic features from the image-power aligned samples; the image-derived physical diagnostic features include: solar geometric features, cloud cover and cloud motion features, cloud-sun relative occlusion geometric features, and solar disk appearance features; Step S4: Based on the physical mechanism of the image-derived physical diagnostic features, group the images to obtain the grouped image-derived physical quantities; the grouping includes at least: solar geometry group, cloud motion group, solar occlusion geometry group, and solar disk appearance group. Step S5, using the current predicted time Using the anchor point as an example, sky images, measured power, image-derived physical diagnostic features, and corresponding clear-sky reference power are extracted from a continuous historical period preceding it to form a historical window sample. ; Step S6: Encode the derived physical quantities of the grouped images to obtain the representation of each physical mechanism group; fuse the representations of each physical mechanism group with the image sequence representation and the power history sequence representation to obtain the fused representation object. ; Step S7, the fused representation object Input the prediction model to output photovoltaic power prediction results for multiple future prediction time intervals.

2. The method according to claim 1, characterized in that, Step S3 includes: Step S31: Obtain the current sky image, timestamp, and measured photovoltaic power at the corresponding time from the image-power alignment sample. and clear sky reference power ; Step S32: Calculate the solar position and solar geometric fundamentals based on the image timestamp, power station latitude and longitude, altitude, and time zone; Step S33: Obtain a clear sky reference image under similar solar altitude conditions; Step S34: Generate an intermediate mask for the cloud region based on the current image and the clear sky reference image; Step S35: Calculate the cloud coverage intensity and the cloud coverage amount in the near-sun region using the intermediate mask of the cloud region; Step S36: Estimate cloud cluster movement characteristics based on cloud region location at consecutive time points; Step S37: Calculate the relative occlusion geometry between the cloud region and the sun's position; Step S38: Construct the estimated occlusion / removal time and time-division occlusion physical characteristics based on the relative motion; Step S39: Calculate the image-derived physical features of the solar disk's brightness, visibility, and appearance.

3. The method according to claim 1, characterized in that, The prediction model is a multi-step prediction head used to output power prediction values ​​at multiple prediction time intervals.

4. The method according to claim 1, characterized in that, The calculation of the clear-sky reference power is as follows: the Ineichen clear-sky model is used, the open_rack_glass_glass temperature model is used, the ambient temperature is taken as 20°C, the wind speed is taken as 1m / s, and the chunk_size parameter is taken as 20000.

5. The method according to claim 2, characterized in that, Specifically, S32 involves: calculating the solar altitude angle, solar azimuth angle, and apparent zenith angle based on the timestamp and the power station's geographical location; and calculating the rate of change of the solar altitude angle, the rate of change of the solar azimuth angle, and the amount of motion of the solar center in the image coordinates based on the changes in the solar position at adjacent times.

6. The method according to claim 2, characterized in that, Step S33 specifically involves: calculating the clear sky index based on the measured power and the clear sky reference power. ， in, Indicates the first Measured photovoltaic power at any given time This represents the reference power at that moment under unobstructed, clear sky conditions. To prevent extremely small constants with a denominator of zero, Used to measure how close the current measured power is to the clear-sky reference power; The samples were grouped according to the numerical range of the solar altitude angle. Within each altitude angle range, the samples with higher clear sky indices were selected. The samples are used as approximate clear sky samples, and the median of the sky images of these samples is taken pixel by pixel to obtain a clear sky reference image under the corresponding solar altitude conditions. .

7. The method according to claim 2, characterized in that, Step S34 specifically involves: For the coordinates in the current sky image, The pixel, whose red channel value is recorded. The value of the blue channel is Calculate the normalized red-blue difference. : ， in, To prevent extremely small constants with a denominator of zero; Reference image for clear sky Calculation of pixels with the same coordinates It also calculates the difference between the current image and a clear-sky reference image. : ， Then, based on the above normalized red-blue difference... Differences Generate cloud region intermediate mask : ， in, This indicates that the value is 1 when the condition is true and 0 when the condition is false. Represents pixels Pixels identified as cloud-related This indicates that the pixel was not identified as a cloud-related pixel.

8. The method according to claim 7, characterized in that, Step S35 specifically involves: Calculate the original cloud cover intensity in the intermediate mask image of the cloud region. : ， in, and These represent the image height and width, respectively, and t represents time.

9. The method according to claim 8, characterized in that, Step S36 specifically involves: First according to Calculate the centroid of the cloud region using the pixel coordinates with a value of 1. Then, based on the changes in the centroid coordinates at consecutive time intervals, the motion velocity of the cloud region in the image coordinates is estimated; the time interval between two adjacent images is obtained. ,but: ， ， in, and These represent the motion components of the cloud region's centroid in the horizontal and vertical directions of the image, respectively. , Let x and y represent the x and y coordinates of the cloud region's centroid in the image coordinate system at time t, respectively. , Let x and y represent the x and y coordinates of the cloud region centroid in the image coordinate system at time t-1, respectively. This indicates the actual observation time corresponding to the previous image; Further calculations were made regarding the speed and direction of cloud movement: ， ， in, Indicates the speed of cloud movement. Indicates the direction of cloud movement.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When executed by a processor, the program implements a photovoltaic power prediction method as described in any one of claims 1-9.

Images