Photovoltaic power station power prediction method and device, storage medium and program product

Abstract

The invention relates to the technical field of photovoltaic power stations, and discloses a photovoltaic power station power prediction method and device, a storage medium and a program product, a plurality of photovoltaic power fluctuation modes are determined by fusing a plurality of real-time target space-based cloud pictures and a plurality of real-time target foundation cloud pictures, and the fineness of fluctuation mode division is improved. Furthermore, ultra-short-term time sequence dependence can be captured by using a long-short-term memory neural network, and meanwhile, short-term prediction is performed in combination with a feedback neural network, so that the adaptability of prediction is improved. Furthermore, a multi-element conjugate nonlinear optimization method is combined to solve a mode-divided combination weight, so that a combination prediction result is closer to actually measured power, the problem of error amplification of a traditional single weight during mode change is avoided, and dynamic correction of prediction results in different fluctuation modes can be realized. Furthermore, prediction is performed on the basis of the multi-modal image sequence in combination with a Transform model, so that the prediction result is closer to the actually measured power, and the accuracy of the prediction result is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power stations, and in particular to a photovoltaic power station power prediction method, device, storage medium and program product. Background Art

[0002] Current photovoltaic power generation prediction methods generally rely on historical power time series for extrapolation modeling. This single data-driven prediction logic has significant flaws: when the fluctuation pattern of the historical power series input to the model differs significantly from the fluctuation pattern of the actual prediction period, such as in power surges caused by rapid changes in cloud cover on rainy days, the prediction results will have significant deviations. Summary of the Invention

[0003] In view of this, the present invention provides a photovoltaic power station power prediction method, device, storage medium and program product to solve the problem that when the fluctuation pattern of the historical power sequence differs greatly from the fluctuation pattern of the actual prediction period, the traditional photovoltaic power generation power prediction results will have significant deviations.

[0004] In a first aspect, the present invention provides a photovoltaic power station power prediction method, the method comprising:

[0005] Multiple real-time target space-based cloud images, multiple real-time target ground-based cloud images and multimodal image sequences of the photovoltaic power station to be predicted are obtained; based on the multiple real-time target space-based cloud images and multiple real-time target ground-based cloud images, the movement trajectory of the cloud group is estimated and multiple photovoltaic power fluctuation patterns of the photovoltaic power station to be predicted are determined; based on the multiple photovoltaic power fluctuation patterns, long short-term memory neural networks and feedback neural networks are used to perform ultra-short-term power prediction and short-term power prediction for the photovoltaic power station to be predicted, respectively, to obtain ultra-short-term predicted power values ​​and short-term predicted power values; based on the ultra-short-term predicted power values ​​and short-term predicted power values, a multivariate conjugate nonlinear optimization method is used to calculate multiple combination weight values ​​of ultra-short-term power prediction and short-term power prediction corresponding to the multiple photovoltaic power fluctuation patterns; based on the multimodal image sequence, the ultra-short-term predicted power values, the short-term predicted power values ​​and the multiple combination weight values, after processing by a Transformer model, the target power prediction value of the photovoltaic power station to be predicted is determined.

[0006] The photovoltaic power station power prediction method provided by the present invention obtains multiple real-time target space-based cloud images reflecting macroscopic long-term motion trends and multiple real-time target ground-based cloud images reflecting microscopic short-term deformation characteristics, and fuses these two types of data to determine multiple photovoltaic power fluctuation patterns of the photovoltaic power station to be predicted. This method solves the pattern recognition bias caused by traditional methods relying solely on historical power series and improves the precision of fluctuation pattern classification. Furthermore, the use of long-short-term memory neural networks can capture ultra-short-term temporal dependencies. At the same time, combined with feedback neural networks for short-term prediction, it improves the adaptability of prediction. Furthermore, the multivariate conjugate nonlinear optimization method is combined to solve the combined weights of the sub-modes, making the combined prediction results closer to the measured power, avoiding the error amplification problem of traditional single weights when the mode changes, and thus enabling dynamic correction of prediction results under different fluctuation patterns. Furthermore, prediction based on multimodal image sequences and combined with the Transformer model makes the prediction results closer to the measured power, improving the accuracy of the prediction results. Therefore, by implementing the present invention, dynamic correction of prediction results under different fluctuation patterns is achieved, improving the accuracy of photovoltaic power prediction results.

[0007] In an optional embodiment, obtaining multiple real-time target space-based cloud images of the photovoltaic power station to be predicted includes:

[0008] Acquire multiple real-time initial space-based cloud images of the photovoltaic power station to be predicted; use a super-resolution convolutional neural network to process the resolution of the multiple real-time initial space-based cloud images to obtain multiple real-time target space-based cloud images.

[0009] The photovoltaic power station power prediction method provided by the present invention processes multiple real-time initial space-based cloud images through a super-resolution convolutional neural network, which can improve the resolution of the space-based cloud images and provide data support for subsequent mode division.

[0010] In an optional embodiment, a super-resolution convolutional neural network is used to process the resolution of multiple real-time initial space-based cloud images to obtain multiple real-time target space-based cloud images, including:

[0011] A plurality of real-time initial space-based cloud images are magnified using a bicubic interpolation method to obtain a plurality of first real-time space-based cloud images; a convolutional neural network is used to reconstruct the plurality of first real-time space-based cloud images to obtain a plurality of second real-time space-based cloud images; a peak signal-to-noise ratio and structural similarity are used to evaluate the plurality of second real-time space-based cloud images to obtain a plurality of evaluation results; and a plurality of real-time target space-based cloud images are determined from the plurality of second real-time space-based cloud images based on the plurality of evaluation results.

[0012] The photovoltaic power station power prediction method provided by the present invention uses a bicubic interpolation method to amplify the resolution of multiple real-time initial space-based cloud images, which can quickly enlarge the size of low-resolution space-based cloud images. Furthermore, the feature extraction and reconstruction capabilities of convolutional neural networks are used to reconstruct multiple first real-time space-based cloud images, solving the problem of low original spatial resolution of the space-based cloud images and clarifying the macroscopic outlines of the clouds. At the same time, it can restore the high-frequency details missing in the image and improve the structural similarity and visual clarity of the cloud images. Furthermore, by using peak signal-to-noise ratio and structural similarity for evaluation, the quality of the second real-time space-based cloud image can be quantitatively assessed, and multiple real-time target space-based cloud images that are close to the original high-resolution image can be screened out, significantly improving the macroscopic feature recognition of the space-based cloud image and providing accurate macroscopic data support for subsequent pattern division.

[0013] In an optional embodiment, obtaining a multimodal image sequence of a photovoltaic power station to be predicted includes:

[0014] A first historical space-based image sequence and a first historical ground-based cloud image sequence of the photovoltaic power station to be predicted are obtained; the first historical space-based image sequence and the first historical ground-based cloud image sequence are respectively selected and enhanced to obtain a second historical space-based image sequence and a second historical ground-based cloud image sequence; the second historical space-based image sequence is preprocessed to obtain a third historical space-based image sequence; and a multimodal image sequence is determined based on the third historical space-based image sequence and the second historical ground-based cloud image sequence.

[0015] The photovoltaic power station power prediction method provided by the present invention selects and enhances a first historical space-based image sequence and a first historical ground-based cloud image sequence of the photovoltaic power station to be predicted, and preprocesses the second historical space-based image sequence. This reduces noise interference, highlights cloud features, ensures that the image sequence has clear cloud features, and avoids prediction deviations caused by data quality issues in the subsequent Transformer model.

[0016] In an optional embodiment, based on multiple real-time target space-based cloud images and multiple real-time target ground-based cloud images, the cloud cluster's trajectory is estimated and multiple photovoltaic power fluctuation patterns of the photovoltaic power station to be predicted are determined, including:

[0017] Based on multiple real-time target space-based cloud images, the motion trajectory of the cloud cluster is estimated to obtain a space-based cloud cluster characteristic parameter set; based on multiple real-time target ground-based cloud images, the motion trajectory of the cloud cluster is estimated to obtain a ground-based cloud image characteristic parameter set; based on the space-based cloud cluster characteristic parameter set and the ground-based cloud image characteristic parameter set, a comprehensive fluctuation index is determined; according to the comprehensive fluctuation index, the fluctuation patterns of the photovoltaic power generation power of the photovoltaic power station to be predicted are divided to obtain multiple photovoltaic power fluctuation patterns.

[0018] The photovoltaic power station power prediction method provided by the present invention uses multiple real-time target space-based cloud images and multiple real-time target ground-based cloud images to estimate the movement trajectory of the cloud group, which can obtain different characteristic parameters reflecting the movement trend of the cloud group, and then integrates the space-based and ground-based characteristic parameters, and quantitatively divides the power fluctuation pattern through the comprehensive fluctuation index, which solves the defect of traditional methods that cannot take into account both macro and micro characteristics, and improves the precision of the fluctuation pattern division.

[0019] In an optional embodiment, based on multiple photovoltaic power fluctuation patterns, a long short-term memory neural network and a feedback neural network are used to perform ultra-short-term power prediction and short-term power prediction for the photovoltaic power station to be predicted, respectively, to obtain ultra-short-term predicted power values ​​and short-term predicted power values, including:

[0020] Acquire multiple historical power data sets of the photovoltaic power station to be predicted under multiple photovoltaic power fluctuation modes and irradiance data sets of the predicted time period; based on the multiple historical power data sets, use the long short-term memory neural network to perform ultra-short-term power prediction on the photovoltaic power station to be predicted, and obtain the ultra-short-term predicted power value; based on the irradiance data set, use the feedback neural network to perform short-term power prediction on the photovoltaic power station to be predicted, and obtain the short-term predicted power value.

[0021] The photovoltaic power station power prediction method provided by the present invention utilizes a long short-term memory neural network to capture ultra-short-term timing dependencies. At the same time, it combines a feedback neural network for short-term prediction, thereby improving the adaptability of the prediction.

[0022] In an optional embodiment, based on the multimodal image sequence, the ultra-short-term predicted power value, the short-term predicted power value, and multiple combined weight values, a target power prediction value of the photovoltaic power station to be predicted is determined through Transformer model processing, including:

[0023] The multimodal image sequence is processed by the Transformer model to obtain the photovoltaic power prediction fluctuation pattern; according to the photovoltaic power prediction fluctuation pattern, the target combination weight value is determined among multiple combination weight values; based on the ultra-short-term prediction power value, the short-term prediction power value and the target combination weight value, the target power prediction value of the photovoltaic power station to be predicted is determined.

[0024] The photovoltaic power station power prediction method proposed in this paper feeds a multimodal image sequence into a Transformer model, capturing the long-range spatiotemporal dependencies of cloud image sequences and accurately predicting future fluctuation patterns. Furthermore, based on the predicted fluctuation pattern, the target weight corresponding to the pattern is selected from the combined weights obtained through multivariate conjugate nonlinear optimization. This achieves dynamic matching of pattern and weight, avoiding the limited adaptability of traditional single weights. Finally, by fusing the two prediction results through dynamic weighting, errors caused by pattern changes are corrected, thereby improving the accuracy of photovoltaic power prediction results.

[0025] In a second aspect, the present invention provides a photovoltaic power station power prediction device, the device comprising:

[0026] The acquisition module is used to acquire multiple real-time target space-based cloud images, multiple real-time target ground-based cloud images and multimodal image sequences of the photovoltaic power station to be predicted; the estimation and determination module is used to estimate the movement trajectory of the cloud group and determine multiple photovoltaic power fluctuation patterns of the photovoltaic power station to be predicted based on the multiple real-time target space-based cloud images and multiple real-time target ground-based cloud images; the prediction module is used to perform ultra-short-term power prediction and short-term power prediction on the photovoltaic power station to be predicted based on the multiple photovoltaic power fluctuation patterns using a long short-term memory neural network and a feedback neural network, respectively, to obtain ultra-short-term predicted power values ​​and short-term predicted power values; the calculation module is used to calculate multiple combined weight values ​​of ultra-short-term power prediction and short-term power prediction corresponding to the multiple photovoltaic power fluctuation patterns using a multivariate conjugate nonlinear optimization method; the processing and determination module is used to determine the target power prediction value of the photovoltaic power station to be predicted based on the multimodal image sequence, the ultra-short-term predicted power value, the short-term predicted power value and the multiple combined weight values ​​after processing by a Transformer model.

[0027] In a third aspect, the present invention provides a computer-readable storage medium having computer instructions stored thereon, the computer instructions being used to enable a computer to execute the photovoltaic power station power prediction method of the first aspect or any corresponding embodiment thereof.

[0028] In a fourth aspect, the present invention provides a computer program product comprising computer instructions for causing a computer to execute the photovoltaic power station power prediction method according to the first aspect or any corresponding embodiment thereof. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0030] Figure 1 is a schematic flow chart of a photovoltaic power station power prediction method according to an embodiment of the present invention;

[0031] Figure 2 is a flow chart of another photovoltaic power station power prediction method according to an embodiment of the present invention;

[0032] Figure 3 is a flow chart of another photovoltaic power station power prediction method according to an embodiment of the present invention;

[0033] Figure 4 1. It is a flow chart of a method for classifying and predicting photovoltaic power generation fluctuation patterns by integrating space-based and ground-based cloud images according to an embodiment of the present invention;

[0034] Figure 5 4-hour forecast results before and after correction according to an embodiment of the present invention;

[0035] Figure 6 is a structural block diagram of a photovoltaic power station power prediction device according to an embodiment of the present invention;

[0036] Figure 7 Schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0037] To make the purpose, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative efforts shall fall within the scope of protection of the present invention.

[0038] An embodiment of the present invention provides a photovoltaic power plant power forecasting method. By integrating multiple real-time target space-based cloud images reflecting macroscopic, long-term motion trends with multiple real-time target ground-based cloud images reflecting microscopic, short-term deformation characteristics, this method achieves a more refined and effective classification of photovoltaic output fluctuation patterns. Furthermore, based on a multivariate conjugate nonlinear optimization method and combined with short-term forecasting, ultra-short-term correction vectors are fitted for different fluctuation patterns, achieving optimal correction and improvement of ultra-short-term forecasting under different fluctuation patterns.

[0039] According to an embodiment of the present invention, an embodiment of a photovoltaic power station power prediction method is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0040] In this embodiment, a photovoltaic power station power prediction method is provided, which can be used for electronic devices such as computers, mobile phones, tablet computers, etc. Figure 1 FIG. 1 is a flow chart of a photovoltaic power station power prediction method according to an embodiment of the present invention. Figure 1 As shown, the process includes the following steps:

[0041] Step S101 : obtaining a plurality of real-time target space-based cloud images, a plurality of real-time target ground-based cloud images, and a multimodal image sequence of a photovoltaic power station to be predicted.

[0042] Among them, the real-time target space-based cloud map refers to the cloud map of the area where the photovoltaic power station to be predicted is located, which is obtained in real time by satellite. After super-resolution processing, the image data meets the prediction requirements and can reflect the overall movement trend of the cloud group, such as moving speed and direction.

[0043] Furthermore, the real-time target ground-based cloud map represents the cloud map collected in real time by the local ground-based equipment of the photovoltaic power station, focusing on the local area around the power station. It has high temporal and spatial resolution and can capture microscopic details such as cloud thickness, density changes, and edge deformation.

[0044] Furthermore, the multimodal image sequence represents a multidimensional data sequence that integrates the historical space-based image sequence and the ground-based cloud image sequence, including the dynamic characteristics of the cloud layer in the time dimension.

[0045] Step S102 : Based on a plurality of real-time target space-based cloud images and a plurality of real-time target ground-based cloud images, the movement trajectory of the cloud cluster is estimated and a plurality of photovoltaic power fluctuation patterns of the photovoltaic power station to be predicted are determined.

[0046] Among them, multiple photovoltaic power fluctuation modes represent three types of photovoltaic power fluctuation states divided according to the correlation between space-based and ground-based cloud image characteristics and power fluctuations, which can include:

[0047] (1) Smooth mode: The cloud cover is relatively stable, and the power output of the photovoltaic power station is basically stable with minimal fluctuations;

[0048] (2) Slight fluctuation mode: clouds are sparsely distributed or move slowly, and the photovoltaic power fluctuates slightly, which does not significantly affect the stability of power generation;

[0049] (3) Severe fluctuation mode: Clouds are dense and move rapidly, and the shielding of solar radiation changes dramatically, resulting in significant fluctuations in photovoltaic power.

[0050] Specifically, by integrating multiple real-time target space-based cloud images reflecting macroscopic long-scale motion trends and multiple real-time target ground-based cloud images reflecting microscopic short-term deformation characteristics, the fluctuation state of the photovoltaic power generation power of the photovoltaic power station to be predicted is divided and multiple photovoltaic power fluctuation patterns of the photovoltaic power station to be predicted are determined. This solves the pattern identification bias problem caused by the traditional method relying only on historical power sequences, and improves the precision of the fluctuation pattern division.

[0051] Step S103 , based on multiple photovoltaic power fluctuation patterns, using a long short-term memory neural network and a feedback neural network, respectively perform ultra-short-term power prediction and short-term power prediction on the photovoltaic power station to be predicted, to obtain an ultra-short-term predicted power value and a short-term predicted power value.

[0052] Among them, the Long Short-Term Memory (LSTM) neural network represents a special recurrent neural network (RNN). It solves the long-term dependency problem of traditional RNN through a gating mechanism (input gate, forget gate, output gate) and is suitable for processing long-term dependency features in time series data.

[0053] Furthermore, a feedback neural network (BP) refers to a neural network based on the back propagation algorithm, which adjusts weights by back propagating errors to achieve nonlinear mapping from input to output, and can be used for pattern recognition and function approximation.

[0054] Specifically, for each photovoltaic power fluctuation pattern, a long short-term memory neural network and a feedback neural network are used to perform ultra-short-term power forecasting and short-term power forecasting for the photovoltaic power station being predicted. Furthermore, the long short-term memory neural network can capture ultra-short-term temporal dependencies during the forecast process, while the feedback neural network is combined with short-term forecasting to improve the adaptability of the forecast.

[0055] Step S104 , based on the ultra-short-term predicted power value and the short-term predicted power value, a multivariate conjugate nonlinear optimization method is used to calculate multiple combined weight values ​​of ultra-short-term power prediction and short-term power prediction corresponding to multiple photovoltaic power fluctuation patterns.

[0056] Among them, the multivariate conjugate nonlinear optimization method represents an iterative algorithm for solving nonlinear optimization problems. The core idea is to gradually approach the optimal solution in the multidimensional parameter space by constructing conjugate direction vectors. It is suitable for processing optimization scenarios with multivariable and nonlinear objective functions.

[0057] Specifically, the combined prediction value is the weighted sum of the ultra-short-term and short-term predictions, and the weight vector w is a 16-dimensional vector to be optimized (one weight for each step), as shown in the following relationship (1):

[0058]

[0059] In the formula: represents the ultra-short-term predicted power value; represents the short-term predicted power value.

[0060] Further, the objective function is to minimize the combined harmonic error, as shown in the following relationship (2):

[0061]

[0062] In the formula: p represents the measured power; e(·) represents the harmonic error formula, as shown in the following relationship (3):

[0063]

[0064] In the formula: P N represents the installed capacity of the photovoltaic power station; i represents the prediction step; p i represents the true power value of the i-th step; represents the predicted power value of the i-th step.

[0065] Further, because the error characteristics are different for different fluctuation modes (smooth, slight, and severe), the weights need to be optimized separately for each mode.

[0066] First, a validation set corresponding to a fluctuation mode (such as a “severe fluctuation mode”) can be separated from historical data, containing the true power values and predicted power values under that mode. p.

[0067] Further, an initial weight w0 is set.

[0068] Second, the harmonic error under the current weight is calculated using a conjugate nonlinear algorithm (such as L-BFGS).

[0069] Finally, the weight w is adjusted along the conjugate direction to reduce the harmonic error e until the stopping condition (such as error convergence or iteration limit) is met, obtaining the optimal combined weight value under the corresponding fluctuation mode.

[0070] Further, through the above process, multiple optimal combined weight values corresponding to multiple photovoltaic power fluctuation modes, i.e., multiple combined weight values, can be obtained.

[0071] Step S105, based on the multi-modal image sequence, the ultra-short-term predicted power value, the short-term predicted power value, and the multiple combined weight values, the target power prediction value of the photovoltaic power station to be predicted is determined through the Transformer model processing.

[0072] Among them, the Transformer model represents a deep learning model based on the self-attention mechanism, which can efficiently capture long-distance dependencies in sequence data.

[0073] Specifically, based on the ultra-short-term predicted power value, the short-term predicted power value and multiple combined weight values, prediction is performed in combination with a multimodal image sequence, so that the predicted result is closer to the measured power, thereby improving the accuracy of the prediction result.

[0074] In some optional implementations, a Transformer model is used to fuse the cloud sequence + power prediction + weight to output a more accurate target power. This includes three steps: input preprocessing, coding layer processing, and decoding output:

[0075] (1) Input preprocessing: Convert the three types of input into sequence vectors that can be processed by Transformer and unify the dimensions.

[0076] (2) Encoding layer processing: The Transformer encoder learns the associations between multi-source data through self-attention and feedforward networks. Specifically, the self-attention mechanism calculates the “attention score” of each element in the input sequence with respect to other elements, capturing the dependencies between cloud features, power predictions, and weights. Furthermore, the feedforward network performs nonlinear transformations on the self-attention output and further extracts complex features.

[0077] (3) Decoding output: After the encoding layer outputs the fused features, the decoder generates the final target power prediction. Specifically, the decoder uses masked self-attention to avoid "seeing future time data" to ensure the prediction logic is reasonable. Furthermore, the decoded features are mapped to a power range (e.g., 0 to installed capacity) through a linear layer, and the target power is output.

[0078] The photovoltaic power station power prediction method provided in this embodiment obtains multiple real-time target space-based cloud images reflecting macroscopic long-term motion trends and multiple real-time target ground-based cloud images reflecting microscopic short-term deformation characteristics, and fuses these two types of data to determine multiple photovoltaic power fluctuation patterns of the photovoltaic power station to be predicted. This solves the pattern recognition bias caused by traditional methods relying solely on historical power series, and improves the precision of fluctuation pattern classification. Furthermore, the use of long-short-term memory neural networks can capture ultra-short-term temporal dependencies. At the same time, combined with feedback neural networks for short-term prediction, it improves the adaptability of prediction. Furthermore, the multivariate conjugate nonlinear optimization method is combined to solve the combined weights of the sub-modes, making the combined prediction results closer to the measured power, avoiding the error amplification problem of traditional single weights when the mode changes, and thus enabling dynamic correction of prediction results under different fluctuation patterns. Furthermore, prediction based on multimodal image sequences and combined with the Transformer model makes the prediction results closer to the measured power, improving the accuracy of the prediction results. Therefore, through the implementation of the present invention, dynamic correction of prediction results under different fluctuation patterns is achieved, improving the accuracy of photovoltaic power prediction results.

[0079] In this embodiment, a photovoltaic power station power prediction method is provided, which can be used for electronic devices such as computers, mobile phones, tablet computers, etc. Figure 2 FIG. 1 is a flow chart of a photovoltaic power station power prediction method according to an embodiment of the present invention. Figure 2 As shown, the process includes the following steps:

[0080] Step S201 : Acquire multiple real-time target space-based cloud images, multiple real-time target ground-based cloud images, and multimodal image sequences of the photovoltaic power station to be predicted.

[0081] Among them, multiple real-time target space-based cloud images of the photovoltaic power station to be predicted can be obtained through the following steps:

[0082] Step a1: Acquire multiple real-time initial space-based cloud images of the photovoltaic power station to be predicted.

[0083] The real-time initial space-based cloud image refers to a cloud image of the area where the photovoltaic power station to be predicted is located, which is obtained in real time by satellite.

[0084] For example, a satellite remote sensing system (such as a meteorological satellite) can be used to collect and obtain original satellite cloud image data, i.e., multiple real-time initial space-based cloud images, at fixed time intervals (such as every 5 minutes) for the geographical area (longitude and latitude range) where the photovoltaic power station to be predicted is located.

[0085] In step a2, a super-resolution convolutional neural network is used to process the resolution of multiple real-time initial space-based cloud images to obtain multiple real-time target space-based cloud images.

[0086] Among them, Super-Resolution Convolutional Neural Network (SRCNN) represents a deep learning model for image super-resolution tasks. Its purpose is to convert low-resolution (LR) images into high-resolution (HR) images through algorithm processing to restore the detail information lost in the image due to insufficient resolution.

[0087] In some optional implementations, the above step a2 includes:

[0088] Step a21 : amplifying the multiple real-time initial space-based cloud images using a bicubic interpolation method to obtain multiple first real-time space-based cloud images.

[0089] Among them, the bicubic interpolation method represents a classic image scaling algorithm, which calculates the new pixel value by fitting a cubic polynomial of 16 neighborhood points around the pixel point, thereby achieving the enlargement of the low-resolution image to a high-resolution size.

[0090] Specifically, multiple real-time initial space-based cloud images with original low resolution are interpolated according to the target resolution to obtain corresponding multiple first real-time space-based cloud images.

[0091] By enlarging the image, you can quickly increase its size.

[0092] Step a22: reconstruct the multiple first real-time space-based cloud images using a convolutional neural network to obtain multiple second real-time space-based cloud images.

[0093] Specifically, a convolutional neural network (CNN) is used to learn the mapping relationship of "low-resolution image → high-resolution image" through the convolution layer, and multiple first real-time space-based cloud images are processed into multiple second real-time space-based cloud images with high resolution.

[0094] First, in the feature extraction layer of the convolutional neural network, the convolution kernel is used to extract the feature map of the low-resolution image.

[0095] Secondly, in the nonlinear mapping layer of the convolutional neural network, larger convolution kernels are used to map features to a high-resolution space.

[0096] Finally, the extracted features are reconstructed into high-resolution images through deconvolution or mapping layers to obtain the corresponding multiple second real-time space-based cloud images.

[0097] Furthermore, since no real new information is injected, the bicubic interpolation method in step a21 will cause blurred image edges and loss of details. Therefore, the above reconstruction process can supplement the details missing in the bicubic interpolation process.

[0098] Step a23, evaluate the plurality of second real-time space-based cloud images using the peak signal-to-noise ratio and the structural similarity to obtain a plurality of evaluation results.

[0099] wherein the peak signal-to-noise ratio (PSNR) is used to measure the pixel error between the super-resolution image and the ideal high-resolution image (if any), and the higher the value, the better the quality, as shown in the following relationship (4):

[0100]

[0101] MAX I denotes the maximum value of the image pixel value; MSE denotes the mean square error, which is an index for measuring the pixel error between the super-resolution image and the ideal high-resolution image (or reference image) after super-resolution processing of the image, and is used to calculate the average value of the square of the difference between the corresponding pixel values of the super-resolution image and the reference image. The smaller the value, the smaller the pixel error between the super-resolution image and the reference image, and the higher the super-resolution quality.

[0102] Further, the structural similarity index (SSIM) is used to measure the similarity of image structure, brightness and contrast, and the closer the value is to 1, the better the quality, as shown in the following relationship (5):

[0103]

[0104] μ x and μ y denote the mean values of x and y, respectively; σ y and σ y denote the standard deviations of x and y, respectively; σ xy denotes the covariance of x and y; c1 and c2 denote constants to prevent division by zero.

[0105] Specifically, if there is an "ideal high-resolution cloud image" (such as a higher precision satellite / ground truth), the PSNR and SSIM of the second real-time space-based cloud image after super-resolution can be directly calculated.

[0106] Further, if there is no truth value, the "subjective quality" (such as edge sharpness) of the image before and after super-resolution can be compared, or the cross-validation method (different SRCNN models are compared with each other) is used to finally output the PSNR value and SSIM value of each second real-time space-based cloud image as the evaluation result.

[0107] Step a24, determining a plurality of real-time target space-based cloud images from the plurality of second real-time space-based cloud images according to the plurality of evaluation results.

[0108] Specifically, the PSNR value and the SSIM value of each second real-time space-based cloud image calculated in step a23 can be compared with the corresponding threshold value respectively, and the second real-time space-based cloud image with up-sampling quality up to the standard can be screened out as the real-time target space-based cloud image.

[0109] Further, the plurality of real-time target space-based cloud images obtained finally not only retain the macroscopic cloud layer movement information of the space-based cloud image, but also improve the local detail recognition degree through up-sampling.

[0110] Through screening, the cloud images with poor up-sampling effect (such as serious edge blurring and many artifacts) can be removed, so as to ensure the quality of the input data for subsequent photovoltaic power fluctuation mode division.

[0111] In some optional embodiments, the multi-modal image sequence can be obtained through the following steps:

[0112] Step b1, obtaining a first historical space-based image sequence and a first historical ground-based cloud image sequence of a photovoltaic power station to be predicted.

[0113] The collection time period can be several hours to several days (such as 24 hours), which can be determined according to the model training requirement; the collection frequency of the space-based image sequence is related to the satellite transit period (such as once every 30 minutes to 2 hours), and the ground-based cloud image sequence is collected in real time at a minute level (such as 5 minutes / time).

[0114] Step b2, performing selection enhancement processing on the first historical space-based image sequence and the first historical ground-based cloud image sequence respectively, to obtain a second historical space-based image sequence and a second historical ground-based cloud image sequence.

[0115] Specifically, the first historical space-based image sequence and the first historical ground-based cloud image sequence can be processed by the following methods:

[0116] (1) Invalid frame screening: invalid frames (such as infrared bright temperature higher than the threshold value, indicating clear sky) or seriously noisy frames (such as data missing, strip interference) are removed by threshold method;

[0117] (2) Feature saliency screening: valid frames containing obvious cloud layer boundaries, dense cloud clusters or rapidly moving cloud systems are retained to ensure that the input sequence has analyzable cloud layer dynamic characteristics.

[0118] Further, the space-based image is enhanced. Specifically, the dynamic range of gray scale can be expanded by histogram stretching to highlight the contrast between the cloud layer and the background. At the same time, median filtering can be used to remove the salt and pepper noise in the satellite signal transmission.

[0119] Further, the ground-based image is enhanced. Specifically, the cloud layer edge definition can be enhanced by sharpening filter, and the local details can be improved by adaptive histogram equalization.

[0120] Through the above selection and enhancement process, the signal-to-noise ratio and feature identifiability of the data are improved, and the interference of invalid frames and noise on model training is avoided, which enables the subsequent Transformer to more accurately learn the mapping relationship between cloud dynamics and power fluctuations.

[0121] Step b3: preprocess the second historical space-based image sequence to obtain a third historical space-based image sequence.

[0122] The preprocessing may include target region clipping and histogram equalization.

[0123] Specifically, a rectangular area (e.g., 50 km × 50 km) centered on the photovoltaic power station can be delineated in the space-based image based on the latitude and longitude of the photovoltaic power station to be predicted, so as to exclude interference from surrounding irrelevant geographical areas.

[0124] Furthermore, the cropped area can be uniformly scaled to a fixed size (e.g., 256×256 pixels) to ensure that the image size input to the Transformer is consistent.

[0125] Furthermore, the captured space-based images can be histogram equalized to expand the grayscale value distribution to the entire dynamic range (0-255), making the cloud brightness temperature difference (reflecting thickness) more obvious.

[0126] Furthermore, multi-band space-based images (such as infrared 10.8μm and visible light 0.6μm) are equalized separately to retain the physical characteristics of each band.

[0127] Furthermore, through target area interception and histogram equalization processing, noise interference is reduced, cloud features are highlighted, and the image sequence is ensured to have clear cloud features, avoiding prediction deviations caused by data quality issues in the subsequent Transformer model.

[0128] Step b4: determining a multimodal image sequence based on the third historical space-based image sequence and the second historical ground-based cloud image sequence.

[0129] Specifically, the obtained third historical space-based image sequence and the second historical ground-based cloud image sequence are integrated to form a corresponding multimodal image sequence.

[0130] In some optional implementations, the space-based image sequence is first interpolated or resampled based on the acquisition time of the ground-based cloud image to ensure that the two correspond one-to-one in the time dimension.

[0131] Secondly, the preprocessed space-based image (e.g., 3 channels: infrared, visible light, and water vapor) and the ground-based image (e.g., 3 channels: RGB) are spliced ​​along the channel dimension to form a 6-channel multimodal image frame.

[0132] Then, the spliced multi-modal image frames are arranged in time sequence to form a time sequence containing macro features in space and micro features on the ground.

[0133] Finally, the pixel values of the multi-modal sequence are normalized (e.g., scaled to [0, 1]) to eliminate the magnitude difference of data from different sensors and form a final multi-modal image sequence.

[0134] Step S202, based on the plurality of real-time target space-based cloud images and the plurality of real-time target ground-based cloud images, estimating the running track of the cloud cluster and determining a plurality of photovoltaic power fluctuation patterns of the photovoltaic power station to be predicted.

[0135] Specifically, the above step S202 includes:

[0136] Step S2021, based on the plurality of real-time target space-based cloud images, estimating the running track of the cloud cluster to obtain a set of space-based cloud cluster feature parameters.

[0137] Among them, the plurality of real-time target space-based cloud images with improved resolution can provide large-scale cloud cluster position, shape and motion direction.

[0138] Specifically, the running track of the cloud cluster can be estimated by frame-by-frame image matching or optical flow method. For example, the optical flow method can be used to calculate the displacement of each cloud cluster between two frames to obtain the motion direction and speed; further, based on the infrared band brightness temperature and height information, the cloud layer thickness can be approximated, and the relationship between brightness temperature and thickness is converted to obtain the cloud thickness estimation by using the empirical formula.

[0139] Further, the corresponding set of space-based cloud cluster feature parameters can be formed by combining the obtained motion direction, speed, cloud layer thickness and other parameters.

[0140] Step S2022, based on the plurality of real-time target ground-based cloud images, estimating the running track of the cloud cluster to obtain a set of ground-based cloud image feature parameters.

[0141] Among them, the time resolution and spatial resolution of the plurality of real-time target ground-based cloud images are high, and the morphological changes and details of the local cloud layer, such as the thickness, density changes, edge shape, etc. of the cloud layer, can be clearly observed.

[0142] Specifically, the edge information can be obtained by using an edge detection algorithm (such as Canny Edge Detection), and the change rate of the edge information in a short time can be calculated to obtain the edge change rate, as shown in the following relationship (6):

[0143]

[0144] In the formula: e grepresents the edge change rate; ΔE represents the change in edge information obtained by the edge detection algorithm of the ground-based cloud image within the time interval Δt, that is, the quantitative value of the difference in edge features between the two frames of ground-based cloud images, such as the change in indicators such as edge length and number of edge pixels; Δt represents the time interval selected when calculating the edge change rate, which represents the time difference between the two frames of ground-based cloud images, and is used to measure the time scale of edge change.

[0145] Furthermore, the grayscale value of ground-based cloud images is positively correlated with the optical thickness of the cloud layer (the lower the grayscale, the thicker the cloud layer and the greater the density). Therefore, the mean or variance of the grayscale values ​​can be used to characterize density fluctuations. A larger mean value indicates a thinner cloud layer (more light transmittance, higher grayscale); a smaller mean value indicates greater local density variation (fragmented clouds with holes).

[0146] Furthermore, the corresponding ground-based cloud image feature parameter set is formed by combining parameters such as edge change rate and cloud density.

[0147] Step S2023: Determine a comprehensive fluctuation index based on the space-based cloud cluster characteristic parameter set and the ground-based cloud image characteristic parameter set.

[0148] Specifically, by integrating the space-based cloud cluster characteristic parameter set and the ground-based cloud image characteristic parameter set, the impact of clouds on photovoltaic power stations can be more accurately judged, as shown in the following relationship (7):

[0149] F t =w1·v s +w2·T C +w3·D C +w4·e g (7)

[0150] Where: F t represents the comprehensive fluctuation index; w1, w2, w3, and w4 are weight coefficients used to balance the influence of satellite and ground-based data; v s represents the speed of space-based cloud cluster movement; T C represents the space-based cloud thickness; D C Indicates the cloud density.

[0151] Step S2024 : dividing the fluctuation patterns of the photovoltaic power generation of the photovoltaic power station to be predicted according to the comprehensive fluctuation index to obtain a plurality of photovoltaic power fluctuation patterns.

[0152] Specifically, the fluctuation modes of photovoltaic power generation can be divided into three types according to the specific value of the calculated comprehensive fluctuation index. For details, please refer to the description in the above step S102.

[0153] Step S203: Based on multiple photovoltaic power fluctuation patterns, the long short-term memory neural network and the feedback neural network are used to perform ultra-short-term power prediction and short-term power prediction on the photovoltaic power station to be predicted, respectively, to obtain ultra-short-term predicted power values ​​and short-term predicted power values. Figure 1 Step S103 of the illustrated embodiment will not be described in detail here.

[0154] Step S204: Based on the ultra-short-term predicted power value and the short-term predicted power value, a multivariate conjugate nonlinear optimization method is used to calculate multiple combined weight values ​​of ultra-short-term power prediction and short-term power prediction corresponding to multiple photovoltaic power fluctuation patterns. Figure 1 Step S104 of the illustrated embodiment will not be described in detail here.

[0155] Step S205: Based on the multimodal image sequence, the ultra-short-term predicted power value, the short-term predicted power value and multiple combined weight values, the target power prediction value of the photovoltaic power station to be predicted is determined through Transformer model processing. Figure 1 Step S105 of the illustrated embodiment will not be described in detail here.

[0156] The photovoltaic power station power prediction method provided in this embodiment uses a bicubic interpolation method to upscale the resolution of multiple real-time initial space-based cloud images, enabling rapid scaling of low-resolution space-based cloud images. Furthermore, the method utilizes the feature extraction and reconstruction capabilities of convolutional neural networks to reconstruct multiple first real-time space-based cloud images. This addresses the low spatial resolution of the original space-based cloud images, clarifies the macroscopic outlines of clouds, and restores missing high-frequency details in the images, improving the structural similarity and visual clarity of the cloud images. Furthermore, using peak signal-to-noise ratio and structural similarity for evaluation, the quality of the second real-time space-based cloud images can be quantitatively assessed. This allows the selection of multiple real-time target space-based cloud images that closely resemble the original high-resolution images, significantly improving the macroscopic feature recognition of the space-based cloud images. Furthermore, the method selects and enhances the first historical space-based image sequence and the first historical ground-based cloud image sequence of the photovoltaic power station to be predicted, and preprocesses the second historical space-based image sequence to reduce noise interference, highlight cloud features, and ensure clear cloud features in the image sequence, thereby avoiding prediction bias caused by data quality issues in the subsequent Transformer model. Furthermore, by estimating the motion trajectory of the cloud cluster using multiple real-time target space-based cloud images and multiple real-time target ground-based cloud images respectively, different characteristic parameters reflecting the motion trend of the cloud cluster can be obtained. Then, the space-based and ground-based characteristic parameters are integrated, and the power fluctuation pattern is quantitatively divided through the comprehensive fluctuation index, which solves the defect of traditional methods that cannot take into account both macro and micro characteristics and improves the precision of fluctuation pattern division.

[0157] In this embodiment, a photovoltaic power station power prediction method is provided, which can be used for electronic devices such as computers, mobile phones, tablet computers, etc. Figure 3 FIG. 1 is a flow chart of a photovoltaic power station power prediction method according to an embodiment of the present invention. Figure 3 As shown, the process includes the following steps:

[0158] Step S301: Acquire multiple real-time target space-based cloud images, multiple real-time target ground-based cloud images, and multimodal image sequences of the photovoltaic power station to be predicted. Figure 2 Step S201 of the illustrated embodiment will not be described in detail here.

[0159] Step S302: Based on multiple real-time target space-based cloud images and multiple real-time target ground-based cloud images, estimate the cloud cluster's trajectory and determine multiple photovoltaic power fluctuation patterns of the photovoltaic power station to be predicted. Figure 2 Step S202 of the illustrated embodiment will not be described in detail here.

[0160] Step S303 , based on multiple photovoltaic power fluctuation patterns, using a long short-term memory neural network and a feedback neural network, respectively perform ultra-short-term power prediction and short-term power prediction on the photovoltaic power station to be predicted, to obtain an ultra-short-term predicted power value and a short-term predicted power value.

[0161] Specifically, the above step S303 includes:

[0162] Step S3031 : obtaining a plurality of historical power data sets of the photovoltaic power station to be predicted under a plurality of photovoltaic power fluctuation modes and an irradiance data set of the predicted time period.

[0163] Specifically, historical power data sets corresponding to multiple different photovoltaic power fluctuation modes may be obtained from the SCADA system or database of the photovoltaic power station.

[0164] Furthermore, historical power data sets can be filtered based on the three types of fluctuation patterns (smooth, slight fluctuations, and sharp fluctuations). For example, in the smooth mode, power data from periods with few clouds and stable power levels can be filtered; in the sharp fluctuation mode, power data from periods with dense clouds and sudden power fluctuations can be filtered.

[0165] Furthermore, the power data can be downsampled to a 15-minute resolution (matching the prediction step) to obtain a historical power dataset for each mode.

[0166] Furthermore, an irradiance dataset for the predicted time period may be obtained from a Numerical Weather Prediction (NWP) system.

[0167] Step S3032 : Based on multiple historical power data sets, a long short-term memory neural network is used to perform ultra-short-term power prediction on the photovoltaic power station to be predicted, and an ultra-short-term predicted power value is obtained.

[0168] Specifically, the long short-term memory neural network can be used to learn the "time series law of historical power in different modes" and predict the ultra-short-term power of the photovoltaic power station to be predicted.

[0169] In some optional implementations, the power in the period t-95 to t is used as input (using a power time resolution of 15 minutes) to predict the future power generation power in the period t+1 to t+16 to obtain a preliminary ultra-short-term power prediction result.

[0170] First, the historical power of the sub-fluctuation pattern is obtained as the input sequence of the long short-term memory neural network.

[0171] Secondly, the long short-term memory neural network is trained using the historical power data of different modes and the measured power of the corresponding time period, and the corresponding loss function is optimized until a trained long short-term memory neural network is obtained.

[0172] Among them, the initial long short-term memory neural network includes:

[0173] (1) Input layer: converts the power sequence into a tensor that the model can process;

[0174] (2) LSTM layer: learns the long-term dependency of historical power through the gating mechanism (input gate, forget gate, output gate), such as the stable trend in smooth mode and the mutation pattern in violent mode;

[0175] (3) Output layer: The fully connected layer maps the LSTM hidden state to the prediction power and outputs the ultra-short-term prediction sequence.

[0176] Finally, multiple historical power data sets (power in the period t-95 to t) are input into the trained long short-term memory neural network, which can output the ultra-short-term predicted power for the next 16 steps.

[0177] Step S3033 : Based on the irradiance data set, a feedback neural network is used to perform short-term power prediction on the photovoltaic power station to be predicted, and a short-term predicted power value is obtained.

[0178] Specifically, the BP neural network can be used to learn the mapping relationship of "meteorological irradiance → power" to predict the short-term power of the day.

[0179] In some optional implementations, NWP irradiance data for the forecast day, ie, the forecast time period, is obtained as an input sequence for the BP.

[0180] Furthermore, the historical meteorological irradiance data (input sequence) and the measured power of the corresponding period are used to train the BP network with a multi-layer perceptron (MLP) structure, and the corresponding loss function is optimized until a trained BP network is obtained.

[0181] Among them, the BP network of the multi-layer perceptron (MLP) structure includes:

[0182] (1) Input layer: converts the irradiance sequence into a feature vector;

[0183] (2) Hidden layer: learns the nonlinear mapping between irradiance and power through nonlinear activation functions (such as ReLU), such as how to convert irradiance fluctuations caused by cloud cover into power changes;

[0184] (3) Output layer: The fully connected layer outputs the short-term predicted power sequence.

[0185] Furthermore, the irradiance data set is input into the trained long short-term memory neural network, which can output the short-term predicted power of the corresponding predicted time period.

[0186] Step S304: Based on the ultra-short-term predicted power value and the short-term predicted power value, a multivariate conjugate nonlinear optimization method is used to calculate multiple combined weight values ​​of ultra-short-term power prediction and short-term power prediction corresponding to multiple photovoltaic power fluctuation patterns. Figure 1 Step S104 of the illustrated embodiment will not be described in detail here.

[0187] Step S305 : Based on the multimodal image sequence, the ultra-short-term predicted power value, the short-term predicted power value and the multiple combined weight values, the target power prediction value of the photovoltaic power station to be predicted is determined through Transformer model processing.

[0188] Specifically, the above step S305 includes:

[0189] Step S3051: Process the multimodal image sequence through the Transformer model to obtain a photovoltaic power prediction fluctuation pattern.

[0190] Specifically, a multimodal image sequence is input into the Transformer model.

[0191] Furthermore, the Transformer extracts the spatiotemporal features of the multimodal sequence through the encoder, and then outputs the corresponding photovoltaic power prediction fluctuation pattern through the decoder.

[0192] In some optional implementations, the multimodal image sequence is converted into a feature vector sequence, and linear transformation + position encoding is performed to allow the model to perceive the temporal order.

[0193] Furthermore, multiple self-attention layers are run in parallel to capture dependencies at different “spatiotemporal scales”.

[0194] Furthermore, the Feed-Forward Network (FFN) performs nonlinear transformation on the self-attention output and extracts more complex features.

[0195] Furthermore, the decoder inputs the feature sequence output by the encoder into the fully connected layer + Softmax to predict the probability distribution of three types of fluctuation patterns (smooth, slight fluctuation, and violent fluctuation).

[0196] Furthermore, the category with the highest probability is taken as the prediction result, i.e., the final photovoltaic power prediction fluctuation pattern.

[0197] Step S3052: determining a target combination weight value from a plurality of combination weight values ​​according to the photovoltaic power prediction fluctuation pattern.

[0198] Specifically, after the photovoltaic power prediction fluctuation pattern is determined, a target combination weight value corresponding to the photovoltaic power prediction fluctuation pattern is determined from the obtained multiple combination weight values.

[0199] Step S3053 : determining a target power prediction value of the photovoltaic power station to be predicted based on the ultra-short-term prediction power value, the short-term prediction power value, and the target combination weight value.

[0200] Specifically, the obtained ultra-short-term predicted power value, short-term predicted power value and target combination weight value are substituted into the above relationship (1), and the target power prediction value of the photovoltaic power station to be predicted can be calculated as follows:

[0201] The photovoltaic power station power prediction method provided in this embodiment uses a long short-term memory neural network to capture ultra-short-term temporal dependencies. At the same time, it combines a feedback neural network for short-term prediction, thereby improving the adaptability of the prediction. Furthermore, by inputting a multimodal image sequence into the Transformer model, it is possible to capture the long-distance spatiotemporal dependencies of the cloud sequence, thereby accurately predicting future fluctuation patterns. Furthermore, based on the predicted fluctuation pattern, the target weight of the corresponding pattern is selected from the combined weights obtained by multivariate conjugate nonlinear optimization, thereby achieving dynamic matching of the pattern and the weight, and avoiding the lack of adaptability of the traditional single weight. Finally, by fusing the two types of prediction results through dynamic weights, the errors caused by pattern changes are corrected, thereby improving the accuracy of the photovoltaic power prediction results.

[0202] In one example, Figure 4As shown in the figure, a photovoltaic power generation power fluctuation pattern classification and prediction correction method that integrates space-based and ground-based cloud images is provided. It can effectively utilize the macro-motion characteristics of space-based satellite cloud images and the local deformation characteristics of ground-based cloud images to realize fluctuation pattern identification, and then adopt targeted correction strategies to effectively solve the prediction errors caused by fluctuation pattern changes. Specifically, it includes:

[0203] S10: Photovoltaic power fluctuation pattern classification by integrating space-based and ground-based cloud images. Space-based cloud images can cover an area of ​​hundreds or even thousands of kilometers from Earth orbit, and can observe the dynamics of the atmosphere and clouds from a macroscopic level. They can effectively capture the overall movement and change trend of clouds. Through image sequences, the speed and direction of cloud movement can be estimated, which can provide photovoltaic power stations with cloud movement trajectories on a longer time scale, helping to judge the cloud dynamics in the next few hours. However, due to the low spatial resolution of space-based cloud images, an image super-resolution algorithm, the Super-Resolution Convolutional Neural Network (SRCNN), is used to improve its resolution. The low-resolution image is enlarged to the size of the high-resolution image through bicubic interpolation, and then reconstructed using a convolutional neural network. The evaluation indicators of the super-resolution image are the Peak Signal-to-Noise Ratio (PSNR) and the Structural Similarity Index (SSIM), as shown in equations (4) and (5), respectively.

[0204] Furthermore, satellite cloud images with improved resolution can provide large-scale cloud locations, shapes, and movement directions.

[0205] Furthermore, the motion trajectory of clouds can be estimated through frame-by-frame image matching or optical flow method. For example, the optical flow method can be used to calculate the displacement of each cloud between two frames to obtain the direction and speed of movement. The cloud thickness can be approximated based on the brightness temperature and height information of the infrared band, and the relationship between brightness temperature and thickness can be converted using empirical formulas to obtain an estimate of cloud thickness.

[0206] Furthermore, the temporal and spatial resolutions of ground-based cloud images are high, and the morphological changes and details of local clouds, such as the thickness, density changes, and edge shapes of clouds, can be clearly observed. Edge detection algorithms (such as Canny Edge Detection) can be used to obtain edge information and calculate its rate of change in a short period of time to obtain the edge change rate, as shown in the above equation (6).

[0207] Furthermore, the optical thickness of the ground-based image is used to approximate the density of the cloud cluster. The mean or variance of the image grayscale value is used to represent the change in grayscale value, which can reflect the density fluctuation of the cloud. CThe ground-based cloud image can focus on the rapid changes of local clouds, providing real-time dynamics at a micro scale. Using the ground-based cloud image of the photovoltaic power station, the rapid transformation and moving speed of the cloud cluster are monitored, which can help predict rapid power fluctuations.

[0208] Further, the combination of macro and micro can more accurately determine the impact of the cloud cluster on the photovoltaic power station. The comprehensive fluctuation index F t As shown in the above relationship (7).

[0209] Further, the fluctuation mode of photovoltaic power can be divided into the following 3 types according to the value of the comprehensive fluctuation index: smooth mode: less cloud, stable power output; slight fluctuation mode: sparse or slow-moving cloud cluster, small fluctuation of photovoltaic power; severe fluctuation mode: dense or fast-moving cloud cluster, significant fluctuation of photovoltaic power.

[0210] S20: Fitting of ultra-short-term prediction correction vector under different fluctuation modes. The photovoltaic power prediction is modeled according to the fluctuation mode, i.e. based on the historical power of the corresponding fluctuation mode, the ultra-short-term power prediction is performed. Based on the long short-term memory neural network (LSTM) model, the power in the period of t-95~t is taken as the input (the power time resolution is 15 min), the power in the period of t+1~t+16 is predicted, and the preliminary ultra-short-term prediction result of the power is obtained. The short-term power prediction result is based on the back propagation neural network (BP), and the predicted day NWP irradiance is taken as the input to predict the power of the predicted day. Then, based on the multivariate conjugate nonlinear optimization method, the combination weights of ultra-short-term and short-term under different fluctuation modes are calculated respectively. The specific implementation is to optimize the combination weights on the validation set respectively, and the objective function is to minimize the combined harmonic error, as shown in the above relationship (2).

[0211] S30: Prediction of photovoltaic power fluctuation mode based on space-based-ground-based cloud image data. The input of the model uses historical space-based image sequences and ground-based cloud image sequences, then the satellite images are preprocessed, including target region cutting and histogram equalization, at the same time, the original satellite images and ground-based cloud images are selected and enhanced, finally the preprocessed historical image sequences are input into the Transformer model for training, and the output is the power fluctuation mode in the next 4 hours.

[0212] S40: Fusion of short-term prediction and ultra-short-term prediction correction. According to the fluctuation mode predicted in the previous step, the short-term prediction result is used to correct the ultra-short-term prediction result according to the fluctuation mode type label and the selected combination weight of S30 corresponding to S20. The combination of short-term prediction result and ultra-short-term prediction result is used to realize the correction of ultra-short-term prediction result, as shown in the above relationship (1).

[0213] In some optional embodiments, as Figure 5As shown in FIG, after the correction strategy provided in this example is implemented, the predicted result is closer to the actual power value than before the correction.

[0214] Furthermore, from the comparison of forecast accuracy from January to June 2024, it can be seen that the forecast accuracy after correction is significantly higher than that before correction, as shown in Table 1 below.

[0215] Table 1. Comparison of prediction accuracy before and after correction

[0216]

[0217]

[0218] Therefore, the photovoltaic power generation power fluctuation mode classification and prediction correction method provided in this example, which integrates space-based and ground-based cloud images, has the following effects:

[0219] 1. Considering that space-based cloud images have a wide observation range but low spatiotemporal resolution, they can effectively explore long-term cloud motion trends. Ground-based cloud images have high spatiotemporal resolution but a small observation range, and can reflect rapid and short-term cloud deformation. By integrating the macroscopic motion characteristics of space-based cloud images with the local deformation characteristics of ground-based cloud images, a more refined and effective classification of photovoltaic output fluctuation patterns is achieved. The photovoltaic power is divided into three modes: smooth, slightly fluctuating, and violently fluctuating, with a basic unit of 4 hours.

[0220] 2. Based on the multivariate conjugate nonlinear optimization method and combined with short-term prediction, the ultra-short-term correction vectors under different fluctuation modes are fitted to achieve the optimal correction and improvement of the ultra-short-term prediction effect under different fluctuation modes.

[0221] This embodiment also provides a photovoltaic power plant power prediction device for implementing the aforementioned embodiments and preferred implementations. Details already described are omitted for clarity. As used below, the term "module" may refer to a combination of software and / or hardware that implements a predetermined function. While the devices described in the following embodiments are preferably implemented using software, implementation using hardware, or a combination of software and hardware, is also possible and contemplated.

[0222] This embodiment provides a photovoltaic power station power prediction device, such as Figure 6 As shown, the device includes:

[0223] The acquisition module 601 is used to acquire multiple real-time target space-based cloud images, multiple real-time target ground-based cloud images and multimodal image sequences of the photovoltaic power station to be predicted.

[0224] The estimation and determination module 602 is used to estimate the movement trajectory of the cloud cluster and determine multiple photovoltaic power fluctuation patterns of the photovoltaic power station to be predicted based on multiple real-time target space-based cloud images and multiple real-time target ground-based cloud images.

[0225] The prediction module 603 is used to perform ultra-short-term power prediction and short-term power prediction on the photovoltaic power station to be predicted based on multiple photovoltaic power fluctuation patterns using a long short-term memory neural network and a feedback neural network to obtain ultra-short-term predicted power values ​​and short-term predicted power values.

[0226] The calculation module 604 is used to calculate multiple combined weight values ​​of ultra-short-term power prediction and short-term power prediction corresponding to multiple photovoltaic power fluctuation patterns based on the ultra-short-term predicted power value and the short-term predicted power value using a multivariate conjugate nonlinear optimization method.

[0227] The processing and determination module 605 is used to determine the target power prediction value of the photovoltaic power station to be predicted based on the multimodal image sequence, the ultra-short-term predicted power value, the short-term predicted power value and multiple combination weight values ​​through Transformer model processing.

[0228] In some optional implementations, the acquisition module 601 includes:

[0229] The first acquisition submodule is used to obtain multiple real-time initial space-based cloud images of the photovoltaic power station to be predicted;

[0230] The first processing submodule is used to process the resolution of multiple real-time initial space-based cloud images using a super-resolution convolutional neural network to obtain multiple real-time target space-based cloud images.

[0231] In some optional implementations, the first processing submodule includes:

[0232] The processing unit is used to amplify the multiple real-time initial space-based cloud images using a bicubic interpolation method to obtain multiple first real-time space-based cloud images.

[0233] The reconstruction unit is used to reconstruct the multiple first real-time space-based cloud images using a convolutional neural network to obtain multiple second real-time space-based cloud images.

[0234] The evaluation unit is used to evaluate the plurality of second real-time space-based cloud images by using the peak signal-to-noise ratio and the structural similarity to obtain a plurality of evaluation results.

[0235] The determining unit is used to determine a plurality of real-time target space-based cloud images from a plurality of second real-time space-based cloud images according to a plurality of evaluation results.

[0236] In some optional implementations, the acquisition module 601 further includes:

[0237] The second acquisition submodule is used to acquire a first historical space-based image sequence and a first historical ground-based cloud image sequence of the photovoltaic power station to be predicted.

[0238] The second processing submodule is used to perform selection and enhancement processing on the first historical space-based image sequence and the first historical ground-based cloud image sequence respectively to obtain a second historical space-based image sequence and a second historical ground-based cloud image sequence.

[0239] The third processing submodule is used to preprocess the second historical space-based image sequence to obtain a third historical space-based image sequence.

[0240] The first determining submodule is configured to determine a multimodal image sequence according to the third historical space-based image sequence and the second historical ground-based cloud image sequence.

[0241] In some optional implementations, the estimation and determination module 602 includes:

[0242] The first estimation submodule is used to estimate the motion trajectory of the cloud cluster based on multiple real-time target space-based cloud images to obtain a space-based cloud cluster characteristic parameter set.

[0243] The second estimation submodule is used to estimate the motion trajectory of the cloud cluster based on multiple real-time target ground-based cloud images to obtain a ground-based cloud image feature parameter set.

[0244] The second determination submodule is used to determine the comprehensive fluctuation index based on the space-based cloud cluster characteristic parameter set and the ground-based cloud image characteristic parameter set.

[0245] The division submodule is used to divide the fluctuation mode of the photovoltaic power generation of the photovoltaic power station to be predicted according to the comprehensive fluctuation index to obtain multiple photovoltaic power fluctuation modes.

[0246] In some optional implementations, the prediction module 603 includes:

[0247] The third acquisition submodule is used to acquire multiple historical power data sets of the photovoltaic power station to be predicted under multiple photovoltaic power fluctuation modes and an irradiance data set of the predicted time period.

[0248] The first prediction submodule is used to perform ultra-short-term power prediction on the photovoltaic power station to be predicted based on multiple historical power data sets using a long short-term memory neural network to obtain an ultra-short-term predicted power value.

[0249] The second prediction submodule is used to perform short-term power prediction on the photovoltaic power station to be predicted based on the irradiance data set by using a feedback neural network to obtain a short-term predicted power value.

[0250] In some optional implementations, the processing determination module 605 includes:

[0251] The fourth processing submodule is used to process the multimodal image sequence through the Transformer model to obtain the photovoltaic power prediction fluctuation pattern.

[0252] The third determination submodule is configured to determine a target combination weight value from among a plurality of combination weight values ​​according to the photovoltaic power prediction fluctuation pattern.

[0253] The fourth determining submodule is configured to determine a target power prediction value of the photovoltaic power station to be predicted based on the ultra-short-term prediction power value, the short-term prediction power value, and the target combination weight value.

[0254] The further functional description of each of the above modules and units is the same as that of the above corresponding embodiments and will not be repeated here.

[0255] The photovoltaic power station power prediction device in this embodiment is presented in the form of a functional unit, where the unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that executes one or more software or fixed programs, and / or other devices that can provide the above functions.

[0256] The embodiment of the present invention also provides a computer device having the above Figure 6 The photovoltaic power station power prediction device shown.

[0257] See also Figure 7 , Figure 7 is a structural diagram of a computer device provided by an optional embodiment of the present invention, such as Figure 7 As shown, the computer device includes: one or more processors 10, a memory 20, and interfaces for connecting various components, including high-speed interfaces and low-speed interfaces. Various components utilize different buses to communicate with each other and can be installed on a common mainboard or installed in other ways as needed. The processor can process the instructions executed in the computer device, including instructions stored in or on the memory to display the graphical information of a GUI on an external input / output device (such as, a display device coupled to the interface). In some optional embodiments, if necessary, multiple processors and / or multiple buses can be used together with multiple memories. Equally, multiple computer devices can be connected, and each device provides part of the necessary operations (for example, as a server array, a group of blade servers, or a multi-processor system). Figure 7 A processor 10 is taken as an example.

[0258] The processor 10 may be a central processing unit, a network processor, or a combination thereof. The processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit, a programmable logic device, or a combination thereof. The programmable logic device may be a complex programmable logic device, a field programmable gate array, a general purpose array logic, or any combination thereof.

[0259] The memory 20 stores instructions that can be executed by at least one processor 10, so as to enable at least one processor 10 to execute the method shown in the above embodiment.

[0260] The memory 20 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and application programs required for at least one function; the data storage area may store data created based on the use of the computer device, etc. In addition, the memory 20 may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some optional embodiments, the memory 20 may optionally include a memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0261] The memory 20 may include a volatile memory, such as a random access memory; the memory may also include a non-volatile memory, such as a flash memory, a hard disk or a solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0262] The computer device further includes a communication interface 30 for the computer device to communicate with other devices or a communication network.

[0263] The embodiment of the present invention also provides a computer-readable storage medium. The above-mentioned method according to the embodiment of the present invention can be implemented in hardware, firmware, or implemented as a computer code that can be recorded in a storage medium, or implemented as a computer code that is originally stored in a remote storage medium or a non-temporary machine-readable storage medium and downloaded through a network and will be stored in a local storage medium, so that the method described herein can be stored in such software processing on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only storage memory, a random access memory, a flash memory, a hard disk or a solid-state drive, etc.; further, the storage medium can also include a combination of the above-mentioned types of memory. It can be understood that a computer, a processor, a microprocessor controller or programmable hardware includes a storage component that can store or receive software or computer code. When the software or computer code is accessed and executed by a computer, a processor or hardware, the method shown in the above embodiment is implemented.

[0264] A portion of the present invention may be applied as a computer program product, such as a computer program instruction, which, when executed by a computer, can call or provide the method and / or technical solution according to the present invention through the operation of the computer. Those skilled in the art should understand that the form in which the computer program instruction exists in a computer-readable medium includes, but is not limited to, a source file, an executable file, an installation package file, etc. Accordingly, the way in which the computer program instruction is executed by the computer includes, but is not limited to: the computer directly executes the instruction, or the computer compiles the instruction and then executes the corresponding compiled program, or the computer reads and executes the instruction, or the computer reads and installs the instruction and then executes the corresponding installed program. Here, the computer-readable medium may be any available computer-readable storage medium or communication medium that can be accessed by the computer.

[0265] Although the embodiments of the present invention have been described with reference to the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention. Such modifications and variations are all within the scope defined by the appended claims.

Claims

1. A photovoltaic power station power prediction method, characterized in that: The method comprises: Acquire multiple real-time target space-based cloud images, multiple real-time target ground-based cloud images, and multimodal image sequences of the photovoltaic power station to be predicted; Based on the multiple real-time target space-based cloud images and the multiple real-time target ground-based cloud images, estimating the movement trajectory of the cloud cluster and determining multiple photovoltaic power fluctuation patterns of the photovoltaic power station to be predicted; Based on the multiple photovoltaic power fluctuation patterns, using a long short-term memory neural network and a feedback neural network, respectively, performing ultra-short-term power prediction and short-term power prediction on the photovoltaic power station to be predicted, to obtain an ultra-short-term predicted power value and a short-term predicted power value; Based on the ultra-short-term predicted power value and the short-term predicted power value, a multivariate conjugate nonlinear optimization method is used to calculate a plurality of combined weight values ​​of the ultra-short-term power prediction and the short-term power prediction corresponding to the plurality of photovoltaic power fluctuation patterns; Based on the multimodal image sequence, the ultra-short-term predicted power value, the short-term predicted power value and the multiple combined weight values, a target power prediction value of the photovoltaic power station to be predicted is determined through Transformer model processing.

2. The method according to claim 1, characterized in that Obtain multiple real-time target space-based cloud images of the photovoltaic power station to be predicted, including: Acquiring a plurality of real-time initial space-based cloud images of the photovoltaic power station to be predicted; The resolution of the multiple real-time initial space-based cloud images is processed using a super-resolution convolutional neural network to obtain the multiple real-time target space-based cloud images.

3. The method according to claim 2, characterized in that The method comprises: processing the resolution of the plurality of real-time initial space-based cloud images using a super-resolution convolutional neural network to obtain the plurality of real-time target space-based cloud images; and Amplifying the multiple real-time initial space-based cloud images using a bicubic interpolation method to obtain multiple first real-time space-based cloud images; Reconstructing the plurality of first real-time space-based cloud images using a convolutional neural network to obtain a plurality of second real-time space-based cloud images; Evaluating the plurality of second real-time space-based cloud images using a peak signal-to-noise ratio and a structural similarity to obtain a plurality of evaluation results; The multiple real-time target space-based cloud images are determined from the multiple second real-time space-based cloud images according to the multiple evaluation results.

4. The method according to claim 1, wherein Obtain a multimodal image sequence of the photovoltaic power plant to be predicted, including: Acquire a first historical space-based image sequence and a first historical ground-based cloud image sequence of the photovoltaic power station to be predicted; performing selection and enhancement processing on the first historical space-based image sequence and the first historical ground-based cloud image sequence respectively to obtain a second historical space-based image sequence and a second historical ground-based cloud image sequence; preprocessing the second historical space-based image sequence to obtain a third historical space-based image sequence; The multimodal image sequence is determined according to the third historical space-based image sequence and the second historical ground-based cloud image sequence.

5. The method according to claim 1, wherein Based on the multiple real-time target space-based cloud images and the multiple real-time target ground-based cloud images, the cloud cluster's trajectory is estimated and multiple photovoltaic power fluctuation patterns of the photovoltaic power station to be predicted are determined, including: estimating the motion trajectory of the cloud cluster based on the multiple real-time target space-based cloud images to obtain a space-based cloud cluster characteristic parameter set; estimating the motion trajectory of the cloud cluster based on the multiple real-time target ground-based cloud images to obtain a ground-based cloud image feature parameter set; determining a comprehensive fluctuation index based on the space-based cloud cluster characteristic parameter set and the ground-based cloud image characteristic parameter set; According to the comprehensive fluctuation index, the fluctuation patterns of the photovoltaic power generation power of the photovoltaic power station to be predicted are divided to obtain the multiple photovoltaic power fluctuation patterns.

6. The method according to claim 1, characterized in that Based on the multiple photovoltaic power fluctuation patterns, using a long short-term memory neural network and a feedback neural network, ultra-short-term power prediction and short-term power prediction are performed on the photovoltaic power station to be predicted, respectively, to obtain an ultra-short-term predicted power value and a short-term predicted power value, including: Acquire a plurality of historical power data sets of the photovoltaic power station to be predicted under the plurality of photovoltaic power fluctuation modes and an irradiance data set of the predicted time period; Based on the multiple historical power data sets, using the long short-term memory neural network to perform ultra-short-term power prediction on the photovoltaic power station to be predicted, to obtain the ultra-short-term predicted power value; Based on the irradiance data set, the feedback neural network is used to perform short-term power prediction on the photovoltaic power station to be predicted to obtain the short-term predicted power value.

7. The method according to claim 1, characterized in that Determining a target power prediction value of the photovoltaic power station to be predicted based on the multimodal image sequence, the ultra-short-term predicted power value, the short-term predicted power value, and the multiple combined weight values ​​through Transformer model processing includes: Processing the multimodal image sequence through the Transformer model to obtain a photovoltaic power prediction fluctuation pattern; determining a target combination weight value from among the plurality of combination weight values ​​according to the photovoltaic power prediction fluctuation pattern; A target power prediction value of the photovoltaic power station to be predicted is determined based on the ultra-short-term prediction power value, the short-term prediction power value, and the target combination weight value.

8. A photovoltaic power station power prediction device, characterized in that: The device comprises: An acquisition module is used to acquire multiple real-time target space-based cloud images, multiple real-time target ground-based cloud images, and multimodal image sequences of the photovoltaic power station to be predicted; an estimation and determination module, configured to estimate the movement trajectory of the cloud cluster and determine multiple photovoltaic power fluctuation patterns of the photovoltaic power station to be predicted based on the multiple real-time target space-based cloud images and the multiple real-time target ground-based cloud images; A prediction module is configured to perform ultra-short-term power prediction and short-term power prediction on the photovoltaic power station to be predicted based on the multiple photovoltaic power fluctuation patterns using a long short-term memory neural network and a feedback neural network to obtain an ultra-short-term predicted power value and a short-term predicted power value; a calculation module, configured to calculate, based on the ultra-short-term predicted power value and the short-term predicted power value, a plurality of combined weight values ​​of the ultra-short-term power prediction and the short-term power prediction corresponding to the plurality of photovoltaic power fluctuation modes using a multivariate conjugate nonlinear optimization method; A processing and determination module is used to determine the target power prediction value of the photovoltaic power station to be predicted based on the multimodal image sequence, the ultra-short-term predicted power value, the short-term predicted power value and the multiple combined weight values ​​through Transformer model processing.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a computer to execute the photovoltaic power station power prediction method according to any one of claims 1 to 7.

10. A computer program product, characterized in that The method comprises computer instructions, wherein the computer instructions are used to enable a computer to execute the photovoltaic power station power prediction method according to any one of claims 1 to 7.

Images