Distributed photovoltaic ultra-short-term prediction method and system based on satellite cloud picture information

Abstract

The invention relates to a distributed photovoltaic ultra-short-term prediction method and system based on satellite cloud picture information, and belongs to the technical field of photovoltaic power prediction. Firstly, the correlation degree of meteorological variables and distributed photovoltaic power is measured through a Pearson correlation coefficient, and cloud layer distribution and movement characteristics thereof are analyzed and found to be key factors influencing photovoltaic output precision, so that a cloud cluster movement prediction framework under multiple time scales is constructed: a CNN-LSTM hybrid model is adopted to capture a nonlinear evolution rule of a cloud cluster; secondly, a feature cloud region dynamic selection algorithm is innovatively designed by means of a satellite cloud picture data source, an ultra-short-term distributed photovoltaic power prediction model based on the convolutional neural network is constructed by accurately quantifying a cloud layer shielding effect and combining the convolutional neural network, and the prediction precision is remarkably improved by effectively fusing cloud motion features.

Description

Technical Field

[0001] The present invention relates to a distributed photovoltaic ultra-short-term prediction method and system based on satellite cloud image information, belonging to the technical field of photovoltaic power prediction. Background Art

[0002] my country's energy structure is transitioning towards cleaner energy, and distributed photovoltaics has become an important development direction due to its flexible layout and environmental friendliness. However, its decentralized and small-scale characteristics make traditional centralized meteorological monitoring methods less applicable, and photovoltaic power forecasting faces significant challenges. Solar irradiance is the core factor affecting photovoltaic power generation, and dynamic changes in cloud cover can significantly change irradiance intensity, causing fluctuations in photovoltaic output. Distributed photovoltaics cannot easily deploy high-precision meteorological monitoring equipment at each site. If data from nearby centralized power stations is directly used, the dynamic shadow effect of clouds on irradiance over time will be ignored, resulting in increased prediction errors. Therefore, improving the accuracy of distributed photovoltaic power forecasting has become a key technical challenge in the energy field.

[0003] With the development of artificial intelligence technology, deep learning has shown great potential in the field of photovoltaic forecasting. Existing research shows that the obstruction effect of clouds is a key factor affecting the fluctuation of distributed photovoltaic output, and the extent of its influence mainly depends on characteristic parameters such as cloud thickness and type. Satellite cloud images can capture key characteristics such as cloud distribution, movement trajectory, and thickness in real time, providing a new path for quantifying the cloud obstruction effect. However, traditional methods have limitations in processing the nonlinear motion characteristics of clouds and dynamically locating characteristic cloud areas. It is difficult to accurately depict the real-time impact of cloud dynamics on photovoltaic power. There is an urgent need to combine satellite cloud image information with advanced deep learning models to build a more efficient ultra-short-term forecasting framework. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a distributed photovoltaic ultra-short-term prediction method and system based on satellite cloud image information. First, the correlation between meteorological variables and distributed photovoltaic power is measured by the Pearson correlation coefficient. Analysis shows that cloud distribution and its movement characteristics are key factors affecting the accuracy of photovoltaic output. For this purpose, a cloud movement prediction framework under multiple time scales is constructed: a CNN-LSTM hybrid model is used to capture the nonlinear evolution of clouds; secondly, with the help of satellite cloud image data sources, a characteristic cloud area dynamic selection algorithm is innovatively designed. By accurately quantifying the cloud shading effect and combining it with a convolutional neural network, an ultra-short-term distributed photovoltaic power prediction model based on a convolutional neural network is constructed. The present invention significantly improves the prediction accuracy by effectively integrating cloud motion characteristics.

[0005] The technical solutions of the present invention are as follows:

[0006] A distributed photovoltaic ultra-short-term prediction method based on satellite cloud image information, the steps are as follows:

[0007] (1) Process meteorological data and analyze the correlation between meteorological variables and distributed photovoltaic power;

[0008] (2) Explore the nonlinear evolution of clouds and construct a cloud movement prediction framework at multiple time scales;

[0009] (3) Construct a dynamic selection algorithm for characteristic cloud regions to accurately quantify the impact of cloud clusters on distributed photovoltaic power fluctuations;

[0010] (4) Combined with neural networks, an ultra-short-term distributed photovoltaic power prediction model is constructed to obtain prediction results.

[0011] Preferably, according to the present invention, in step (1), the meteorological variables include temperature, relative humidity, air pressure, wind speed, wind direction, precipitation and shortwave radiation, and the Pearson correlation coefficient is used to measure the correlation between the meteorological variables and the distributed photovoltaic power. The results show that shortwave radiation has the highest correlation with photovoltaic power.

[0012] Preferably, in step (2), a convolutional-long short-term memory neural network (CNN-LSTM) hybrid model is constructed to perform nonlinear cloud movement prediction;

[0013] Convolutional neural networks consist of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. To address the influencing factors of nonlinearity, convolutional neural networks introduce activation functions to map nonlinear features to a high-dimensional nonlinear region.

[0014] Among them, the convolution layer and pooling layer are the key layers, and the convolution operation is expressed as:

[0015]

[0016] in, is the weight of the g-th convolution filter in the l-th layer, is the bias of the lth layer, and the input area of ​​the lth layer position (i, j) is expressed as f conv represents the activation function;

[0017] At the same time, pooling layers are regularly introduced between convolutional layers to perform pooling in each depth dimension separately. The depth of the image remains unchanged, and the final output of the pooling layer is:

[0018]

[0019] in, is the result after the convolution operation, Represents the result after pooling, f poolRepresents the pooling operation. After the convolutional and pooling layers, each node in the fully connected layer is connected to all nodes in the previous layer, and the features extracted by the previous layer are integrated. The extracted feature map is converted into the final output of the network, resulting in multiple cloud image patches representing local cloud shape and thickness information;

[0020] The Long Short-Term Memory (LSTM) neural network is an improved neural network model based on the Recursive Neural Network (RNN). Compared with the RNN, the LSTM neural network dynamically controls the increase and decrease of information by introducing three gating units: the forget gate, the input gate, and the output gate, effectively solving the problem of long-term dependency. Its core memory unit can continuously store key information during training, significantly improving the model's prediction accuracy.

[0021] At time t, the status of each door is:

[0022] Forget Gate:

[0023] f t =σ(W f ·[h t-1 , x t ]+b f ) (3)

[0024] Input Gate:

[0025] i t =σ(W i ·[h t-1 , x t ]+b i ) (4)

[0026] Output gate:

[0027] o t =σ(W o ·[h t-1 , x t ]+b o ) (5)

[0028] Where t represents the current time step; t-1 represents the previous time step; W f 、W i 、W o represents the weight matrix of the gate; b f 、b i 、b o Indicates the gate bias; h t-1 is the hidden state output at the previous moment; x t is the current input; σ is the sigmoid activation function;

[0029] The memory cell normalizes the input by updating the function, which multiplies i t , the obtained gt is used to update the memory unit, the formula is:

[0030] g t =tanh(W g ·[h t-1 , x t ]+b g ) (6)

[0031] Where W g represents the weight matrix of the update function; b g is the bias; tanh is the activation function;

[0032] The update of memory units is the core part of the long short-term memory neural network. Its state is determined by the historical state and the current state. The memory unit will be updated as follows:

[0033] c t =f t c t-1 +i t ·g t (7)

[0034] Where C t-1 Indicates the memory unit at the previous moment;

[0035] The final output of the long short-term memory neural network is determined by the output gate and the unit state:

[0036] h t =o t *tanh(c t )(8).

[0037] According to the present invention, further preferably, in step (2):

[0038] The nonlinear cloud movement prediction method uses historical satellite cloud images at time T-2, T-1, and T as input to predict the cloud image distribution from time T+1 to T+3. To focus on the target area, the original cloud image is cropped into a 60×60 pixel rectangular area centered on the photovoltaic area.

[0039] A 6-layer convolutional-long short-term memory neural network cascade structure is used to extract the spatiotemporal features of the cloud image through multiple convolution kernels. Each layer of CNN-LSTM outputs a three-dimensional feature map (length × width × number of channels). Finally, the 3D features are converted into 2D prediction results through the Conv3D layer.

[0040] The specific steps of the nonlinear cloud movement prediction method are as follows:

[0041] Firstly, a 3×3 convolution kernel is used to extract the spatial distribution and thickness characteristics of cloud clusters from continuous satellite cloud images, and the cloud images are divided into multiple small cloud blocks, which represent the local cloud shape and thickness information respectively.

[0042] Secondly, a long short-term memory neural network is used to analyze the temporal evolution of cloud conditions in each region, learn local dynamic features and integrate them into overall motion trends. Since the change trends of each cloud block are different, a convolutional-long short-term memory neural network is used to learn the movement patterns of each region, realize nonlinear motion prediction of cloud clusters, and use the structural similarity metric (SSIM) as the evaluation indicator.

[0043] According to the preferred embodiment of the present invention, in step (3), a characteristic cloud area positioning algorithm based on the law of solar motion is proposed to identify the cloud area that blocks the incident light of the sun in real time. In the satellite cloud image I (x, y), it is assumed that a certain distributed photovoltaic O (x0, y0) is located at the center of the image, and the vertical projection of the intersection S' of the sunlight and the cloud group on the cloud image is S (x s ,y s ), through the solar altitude angle α s and the solar azimuth γ s , combined with the cloud height H and distance L, the solar incidence angle of the photovoltaic area at any moment is calculated. Finally, the cloud clusters that intersect with the propagation trajectory of the sun's rays are determined to be characteristic cloud areas with blocking effects;

[0044] The specific steps of the characteristic cloud area positioning algorithm are as follows:

[0045] First, calculate the sun's altitude:

[0046]

[0047] Where ψ represents latitude; ω represents hour angle; δ is the solar declination angle. The hour angle and solar declination angle are calculated using the following two equations:

[0048] ω=(t-12)×15° (10)

[0049]

[0050] Where t represents a certain moment; d represents the serial number of a certain day in a year; based on the above physical quantities, the solar azimuth angle is calculated by the following formula:

[0051]

[0052] According to the right triangle law, L is obtained from H:

[0053] L=H / tanα S (13)

[0054] Finally, through L and γs Obtain the position coordinates of point S, and then calculate the intersection of the sun's rays and the cloud cluster in the satellite cloud image. The specific process is as follows:

[0055] x S =x O +L×sinγ S (14)

[0056] y S =y O +L×cosγ S (15)

[0057] Where x s and y s They are the east-west and north-south distances between point S and point O in the satellite cloud image, respectively. Based on the above operations, the intersection position is mapped from three-dimensional space to two-dimensional satellite cloud image, and the intersection position of sunlight and cloud clusters in the cloud image can be accurately located.

[0058] In order to eliminate the S estimation bias caused by the idealized assumption of cloud height H, the present invention selects a 30×30 pixel rectangular area R with the intersection S as the center as the characteristic cloud area to more comprehensively reflect the real occlusion effect of the cloud.

[0059] According to the preferred embodiment of the present invention, in step (4), the ultra-short-term distributed photovoltaic power prediction model includes two stages: training and prediction;

[0060] During the training phase, cloud images at historical moments are first acquired through a dynamic selection algorithm for characteristic cloud regions. These images are then fed into a convolutional neural network. After convolution, normalization, and pooling, the cloud occlusion influencing factor is extracted and combined with other influencing factors. A mapping relationship between this factor and distributed photovoltaic power is established through a fully connected layer.

[0061] In the prediction stage, the satellite cloud image at the time to be predicted is first input into the convolution-pooling layer of the convolutional neural network. Each convolution-pooling layer consists of a convolution layer, a normalization layer, and a pooling layer arranged in sequence. In the pooling layer, a rank-ordered pooling method is used, with a kernel size of 2×2 and a step size of 2. After two layers of convolution-pooling layers, two fully connected layers are added, each with 1024 neurons. The variables that quantify the cloud occlusion information are obtained from the fully connected layer, which are set as eight cloud occlusion factors here. In addition, eight other influencing variables are selected for fusion modeling with the cloud occlusion factors, namely the temperature, humidity, irradiance, sine value of the solar altitude angle at the time to be predicted, and the historical power at times T-1, T-2, T-3, and T-4. The 16 variables are fused and input into a fully connected layer consisting of 256 neurons. After weighted operation, the distributed photovoltaic power prediction result is obtained.

[0062] A distributed photovoltaic ultra-short-term prediction system based on satellite cloud image information, comprising:

[0063] Data analysis module, used to process meteorological data and analyze the correlation between meteorological variables and distributed photovoltaic power;

[0064] The cloud prediction module is used to explore the nonlinear evolution of clouds and build a cloud movement prediction framework at multiple time scales;

[0065] The characteristic cloud region selection module is used to build a dynamic selection algorithm for characteristic cloud regions, accurately quantify the impact of cloud clusters on distributed photovoltaic power fluctuations, locate relevant cloud characteristic regions in real time, and obtain cloud images at historical moments;

[0066] The power prediction module is used to build an ultra-short-term distributed photovoltaic power prediction model and obtain prediction results.

[0067] The present invention provides a computer-readable storage medium, which adopts the following technical solution:

[0068] A computer-readable storage medium stores a program, which, when executed by a processor, implements the steps of the above-mentioned distributed photovoltaic ultra-short-term prediction method based on satellite cloud image information.

[0069] The present invention provides an electronic device, which adopts the following technical solution:

[0070] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps of the above-mentioned distributed photovoltaic ultra-short-term prediction method based on satellite cloud image information are implemented.

[0071] The beneficial effects of the present invention are:

[0072] This paper studies cloud motion prediction technology at multiple time scales and proposes a nonlinear cloud motion prediction method based on a convolutional-long short-term memory neural network (CNN-LSTM) for hourly cloud image prediction. Then, with the help of satellite cloud image data sources, a dynamic selection algorithm for characteristic cloud areas is further proposed. An ultra-short-term distributed photovoltaic power prediction method based on a convolutional neural network is established. The introduction of cloud feature information into the input variables further improves the prediction accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 This is a correlation coefficient diagram of distributed photovoltaic power at different times during the daytime in Example 1 of the present invention;

[0074] Figure 2 This is a flow chart of the nonlinear cloud movement prediction method based on CNN-LSTM in Example 1 of the present invention;

[0075] Figure 3 Schematic diagram of the intersection of sunlight and clouds in Example 1 of the present invention;

[0076] Figure 4 This is a diagram of the internal structure of the convolutional neural network used in Example 1 of the present invention;

[0077] Figure 5 This is a comparison chart of the small time cloud map prediction results in Example 1 of the present invention;

[0078] Figure 6 1 is a graph comparing the errors of the photovoltaic power prediction method in Example 1 of the present invention and other methods;

[0079] Figure 7 is a prediction curve diagram of the present invention under cloudy and sunny conditions in Example 1 of the present invention, wherein, Figure 7 (a) is the forecast curve under cloudy conditions. Figure 7 (b) is the forecast curve under sunny conditions; DETAILED DESCRIPTION

[0080] The present invention will be further described below with reference to embodiments and accompanying drawings, but is not limited thereto.

[0081] Example 1:

[0082] This embodiment discloses a distributed photovoltaic ultra-short-term prediction method based on satellite cloud image information, including:

[0083] (1) The Pearson correlation coefficient is used to measure the correlation between distributed photovoltaic power and meteorological variables, providing a data basis for studying cloud distribution and its mobility characteristics.

[0084] (2) A cloud movement prediction framework at multiple time scales was constructed, and a CNN-LSTM hybrid model was used to capture the nonlinear evolution of clouds. A dynamic selection algorithm for characteristic cloud areas was designed with the help of satellite cloud image data sources to achieve accurate prediction of cloud changes.

[0085] (3) An ultra-short-term distributed photovoltaic power prediction model based on convolutional neural networks was constructed to achieve accurate prediction of distributed photovoltaic power under different scenarios.

[0086] (4) The prediction errors of the prediction method proposed in the present invention are compared with those of other comparison methods using error evaluation indicators, and the prediction performance of the prediction method proposed in the present invention is analyzed through the comparison results.

[0087] Specifically, the detailed implementation process of this example is as follows:

[0088] The changing trend of distributed photovoltaic power roughly mirrors that of solar irradiance. Under clear skies, distributed photovoltaic power shows a pattern of first rising and then falling. Under cloudy conditions, irradiance fluctuates dramatically due to cloud movement, leading to increased volatility in distributed photovoltaic output. Under overcast skies, thicker and more widely distributed clouds result in consistently lower irradiance, resulting in lower and more gradual fluctuations in distributed photovoltaic output. Furthermore, meteorological factors such as temperature, relative humidity, air pressure, wind speed, and wind direction also exhibit a certain degree of correlation with distributed photovoltaic power.

[0089] The Pearson correlation coefficient was used to analyze the correlation between meteorological variables and distributed photovoltaic power. Table 1 shows the correlation results. Shortwave radiation has the highest correlation with photovoltaic power. Therefore, shortwave radiation, temperature and relative humidity should be selected as key input variables in the construction of the prediction model.

[0090] Table 1: Correlation coefficients between different meteorological variables and PV power

[0091]

[0092] The distributed photovoltaic power generation time series has obvious temporal correlation within a certain time scale. The Pearson correlation coefficient is used to calculate the correlation between different times during the daytime. The results are as follows: Figure 1 As shown in Figure 2, it can be seen that, on the one hand, the distributed photovoltaic power at adjacent times has a high correlation due to the existence of temporal inertia; on the other hand, due to the existence of the daily cycle characteristics, the symmetrical times with noon as the central axis also have a high correlation.

[0093] Cloud obstruction is a key factor in distributed photovoltaic power output fluctuations. Accurately quantifying the dynamic obstruction effect of clouds is crucial for improving power forecast accuracy. This paper proposes a nonlinear cloud motion prediction method using a convolutional-long short-term memory neural network (CNN-LSTM) for cloud image forecasting. Finally, leveraging satellite cloud image data sources, a dynamic selection algorithm for characteristic cloud regions is proposed to more accurately quantify the impact of clouds on distributed photovoltaic power fluctuations.

[0094] Convolutional Neural Networks (CNN) have demonstrated good learning ability and powerful feature extraction capabilities in the face of large amounts of data. They are also widely used in related fields such as image recognition, fault diagnosis, and time series prediction.

[0095] Convolutional neural networks consist of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. To address the influencing factors of nonlinearity, convolutional neural networks introduce activation functions to map nonlinear features to a high-dimensional nonlinear region.

[0096] Among them, the convolution layer and pooling layer are the key layers, and the convolution operation is expressed as:

[0097]

[0098] in, is the weight of the g-th convolution filter in the l-th layer, is the bias of the lth layer, and the input area of ​​the lth layer position (i, j) is expressed as f conv represents the activation function;

[0099] At the same time, pooling layers are regularly introduced between convolutional layers to perform pooling in each depth dimension separately. The depth of the image remains unchanged, and the final output of the pooling layer is:

[0100]

[0101] in, is the result after the convolution operation, Represents the result after pooling, f pool Represents the pooling operation. After the convolutional and pooling layers, each node in the fully connected layer is connected to all nodes in the previous layer, and the features extracted by the previous layer are integrated. The extracted feature map is converted into the final output of the network, resulting in multiple cloud image patches representing local cloud shape and thickness information;

[0102] The Long Short-Term Memory (LSTM) neural network is an improved neural network model based on the Recursive Neural Network (RNN). Compared with the RNN, the LSTM neural network dynamically controls the increase and decrease of information by introducing three gating units: the forget gate, the input gate, and the output gate, effectively solving the problem of long-term dependence. Its core memory unit can continuously store key information during the training process, significantly improving the model's prediction accuracy.

[0103] At time t, the status of each door is:

[0104] Forget Gate:

[0105] f t =σ(W f ·[h t-1 , x t ]+b f ) (3)

[0106] Input Gate:

[0107] i t =σ(W i ·[h t-1 , x t]+b i ) (4)

[0108] Output gate:

[0109] o t =σ(W o ·[h t-1 , x t ]+b o ) (5)

[0110] Where t represents the current time step; t-1 represents the previous time step; W f 、W i 、W o represents the weight matrix of the gate; b f 、b i 、b o Indicates the gate bias; h t-1 is the hidden state output at the previous moment; x t is the current input; σ is the sigmoid activation function;

[0111] The memory cell normalizes the input by updating the function, which multiplies i t , the obtained gt is used to update the memory unit, the formula is:

[0112] g t =tanh(W g ·[h t-1 , x t ]+b g ) (6)

[0113] Where W g represents the weight matrix of the update function; b g is the bias; tanh is the activation function;

[0114] The update of memory units is the core part of the long short-term memory neural network. Its state is determined by the historical state and the current state. The memory unit will be updated as follows:

[0115] c t =f t c t-1 +i t ·g t (7)

[0116] Where C t-1 Indicates the memory unit at the previous moment;

[0117] The final output of the long short-term memory neural network is determined by the output gate and the unit state:

[0118] h t =o t *tanh(ct )(8).

[0119] Figure 2 This article presents a flowchart for a nonlinear cloud motion prediction method based on a CNN-LSTM algorithm. This method uses historical satellite cloud images from times T-2, T-1, and T as input and predicts the cloud distribution from times T+1 to T+3. To focus on the target area, the original cloud image is cropped into a 60×60 pixel rectangle centered on the photovoltaic region.

[0120] The model uses a 6-layer CNN-LSTM cascade structure to extract spatiotemporal features of cloud images using multiple convolutional kernels. Each CNN-LSTM layer outputs a 3D feature map (length × width × number of channels), which is ultimately converted into a 2D prediction result through the Conv3D layer.

[0121] The specific steps of the nonlinear cloud movement prediction method are as follows:

[0122] Firstly, a 3×3 convolution kernel is used to extract the spatial distribution and thickness characteristics of cloud clusters from continuous satellite cloud images, and the cloud images are divided into multiple small cloud blocks, which represent the local cloud shape and thickness information respectively.

[0123] Secondly, an LSTM network is used to analyze the temporal evolution of cloud conditions in each region, learning local dynamic features and integrating them into overall motion trends. Because the changing trends of each cloud patch are different, a CNN-LSTM hybrid architecture is used to learn the movement patterns of each region, enabling nonlinear cloud motion prediction. The Structural Similarity metric (SSIM) is used as an evaluation metric.

[0124] This embodiment proposes a characteristic cloud region positioning algorithm based on the law of solar motion, which is used to identify cloud regions that block incident solar light in real time. The core of the algorithm is to determine the geometric intersection of solar rays and cloud clusters: Figure 3 As shown in the figure, in a satellite cloud image I(x,y), assume that a distributed photovoltaic system O(x0,y0) is located at the center of the image. The vertical projection of the intersection S' of the sunlight and the cloud cluster on the cloud image is S(xs,ys). Using the solar zenith angle αs and the solar azimuth angle γs, combined with the cloud height H and distance L, the solar incidence angle at the photovoltaic area at any moment can be accurately calculated. Ultimately, clouds that intersect the propagation path of sunlight are identified as characteristic cloud regions with an obstructing effect.

[0125] The specific steps of the characteristic cloud area positioning algorithm are as follows:

[0126] First, calculate the sun's altitude:

[0127]

[0128] Where ψ represents latitude; ω represents hour angle; δ is the solar declination angle. The hour angle and solar declination angle are calculated using the following two equations:

[0129] ω=(t-12)×15° (10)

[0130]

[0131] Where t represents a certain moment; d represents the serial number of a certain day in a year; based on the above physical quantities, the solar azimuth angle is calculated by the following formula:

[0132]

[0133] According to the right triangle law, L is obtained from H:

[0134] L=H / tanα S (13)

[0135] Finally, through L and γ s Obtain the position coordinates of point S, and then calculate the intersection of the sun's rays and the cloud cluster in the satellite cloud image. The specific process is as follows:

[0136] x S =x O +L×sinγ S (14)

[0137] y S =y O +L×cosγ S (15)

[0138] Where x s and y s They are the east-west and north-south distances between point S and point O in the satellite cloud image, respectively. Based on the above operations, the intersection position is mapped from three-dimensional space to two-dimensional satellite cloud image, and the intersection position of sunlight and cloud clusters in the cloud image can be accurately located.

[0139] In order to eliminate the S estimation bias caused by the idealized assumption of cloud height H, the present invention selects a 30×30 pixel rectangular area R with the intersection S as the center as the characteristic cloud area to more comprehensively reflect the real occlusion effect of the cloud.

[0140] The ultra-short-term distributed photovoltaic power prediction framework based on CNN is as follows: Figure 4 As shown, it includes two stages: training and prediction.

[0141] During the training phase, cloud images at historical moments are first obtained through a dynamic selection algorithm for characteristic cloud areas. Secondly, the cloud images are input into a convolutional neural network. After operations such as convolution, normalization, and pooling, the cloud occlusion influencing factor is extracted and combined with other influencing factors. A mapping relationship between the cloud image and the distributed photovoltaic power is established through a fully connected layer.

[0142] In the prediction phase, the satellite cloud image at the time of prediction is first fed into the convolution-pooling layer of the CNN. Each convolution-pooling layer consists of a convolutional layer, a normalization layer, and a pooling layer in sequence. In the pooling layer, this study uses a rank-ordered pooling scheme with a kernel size of 2×2 and a stride of 2. After two convolution-pooling layers, two fully connected layers are added, each with 1024 neurons. These fully connected layers then provide the variables that quantify cloud obscuration information, here defined as eight cloud obscuration factors. Furthermore, after correlation testing, this study selected eight additional influencing variables for fusion modeling with the cloud obscuration factors: temperature, humidity, irradiance, the sine of the solar altitude angle at the time of prediction, and historical power at times T-1, T-2, T-3, and T-4. These 16 variables are fused and fed into a fully connected layer consisting of 256 neurons. After a series of weighted operations, the distributed photovoltaic power forecast is obtained.

[0143] Finally, the prediction performance of the model proposed in the present invention is analyzed through actual examples. The data used in the examples in this embodiment are all from the distributed photovoltaic operation data and corresponding meteorological data of a certain region throughout the year.

[0144] Figure 5 Comparative cloud image prediction results are presented, including those of a CNN-LSTM model using both the root mean square error (MSE) and the structural similarity metric (SSIM) as loss functions. The figure shows that the CNN-LSTM model using MSE suffers from image blurring, while the CNN-LSTM model using SSIM achieves closer approximations to the real image in terms of cloud shape and edge clarity. By factoring in variations in cloud thickness and morphology, the CNN-LSTM model achieves accurate nonlinear motion prediction, demonstrating its superior performance in spatiotemporal sequence prediction tasks.

[0145] Because distributed photovoltaic output is low in the early morning and evening, the analysis selected valid data from 9:00 AM to 4:00 PM daily. To fully validate the model's predictive performance, data from the first three weeks of each month was used as the training set, and the last week as the test set. Model performance was quantitatively evaluated using the normalized mean absolute error (NMAE) and normalized root mean square error (NRMSE) to ensure the reliability of the evaluation results.

[0146] Figure 6This figure shows a comparison of the prediction errors of the proposed method and the comparison method. CNN (unlocalized) indicates that the feature cloud regions were not localized using the feature region localization algorithm. All other conditions are the same as those of the proposed method. VGG is a variant of CNN that is deeper and denser than standard CNNs.

[0147] As can be seen from the figure, VGG did not demonstrate good prediction performance in the test, indicating that an overly deep and dense network structure is not necessary to extract cloud features. On the other hand, the CNN (unlocalized) method exhibited a larger prediction error than the proposed method, demonstrating the positive role of the feature region localization algorithm in improving prediction accuracy.

[0148] To further validate the predictive performance of the proposed method, we selected support vector machines (SVMs) and LSTMs, two commonly used and highly effective methods in the photovoltaic forecasting field, as comparison models. Using distributed photovoltaic data from regions A, B, and C, we tested these three methods over a four-hour forecasting timeframe. The results are shown in Table 2.

[0149] Table 2: Prediction errors of different methods at a prediction scale of 4 hours

[0150]

[0151] Analysis shows that the errors of each method increase with the increase of prediction time, but the CNN method can still maintain a relatively small prediction error and achieve more accurate photovoltaic prediction.

[0152] In order to further explore the prediction performance of the proposed method under cloudy and sunny weather conditions, Figure 7 The four-hour-ahead forecast curve for area A under cloudy and sunny conditions on a certain day is shown.

[0153] Depend on Figure 7 It can be seen that the distributed photovoltaic power curve is relatively smooth under sunny conditions, while the distributed photovoltaic power fluctuation is more obvious under cloudy conditions. Whether it is sunny or cloudy weather, the predicted curve of this method is relatively close to the actual curve, indicating that the proposed method can achieve accurate prediction of distributed photovoltaic power in different scenarios.

[0154] Example 2

[0155] This embodiment discloses a distributed photovoltaic ultra-short-term prediction system based on satellite cloud image information, including:

[0156] Data analysis module, used to process meteorological data and analyze the correlation between meteorological variables and distributed photovoltaic power;

[0157] The cloud prediction module is used to explore the nonlinear evolution of clouds and build a cloud movement prediction framework at multiple time scales;

[0158] The characteristic cloud region selection module is used to build a dynamic selection algorithm for characteristic cloud regions, accurately quantify the impact of cloud clusters on distributed photovoltaic power fluctuations, locate relevant cloud characteristic regions in real time, and obtain cloud images at historical moments;

[0159] The power prediction module is used to build an ultra-short-term distributed photovoltaic power prediction model and obtain prediction results.

[0160] Example 3:

[0161] This embodiment provides a computer-readable storage medium:

[0162] A computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of a distributed photovoltaic ultra-short-term prediction method based on satellite cloud image information as described in Example 1.

[0163] The detailed steps are the same as those of the distributed photovoltaic ultra-short-term prediction method based on satellite cloud image information provided in Example 1, and will not be repeated here.

[0164] Example 4:

[0165] This embodiment provides an electronic device.

[0166] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps of the distributed photovoltaic ultra-short-term prediction method based on satellite cloud image information as described in Example 1 are implemented.

[0167] The detailed steps are the same as those of the distributed photovoltaic ultra-short-term prediction method based on satellite cloud image information provided in Example 1, and will not be repeated here.

[0168] The foregoing description is merely a preferred embodiment of the present disclosure and is not intended to limit the present disclosure. Those skilled in the art will readily appreciate that various modifications and variations are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present disclosure shall be included within the scope of protection of the present disclosure.

[0169] Although the above describes the specific implementation methods of the present disclosure in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present disclosure. Those skilled in the art should understand that on the basis of the technical solution of the present disclosure, various modifications or variations that can be made by those skilled in the art without creative work are still within the scope of protection of the present disclosure.

Claims

1. A distributed photovoltaic ultra-short-term prediction method based on satellite cloud image information, characterized in that: Here are the steps: (1) Process meteorological data and analyze the correlation between meteorological variables and distributed photovoltaic power; (2) Explore the nonlinear evolution of clouds and construct a cloud movement prediction framework at multiple time scales; (3) Construct a dynamic selection algorithm for characteristic cloud regions to accurately quantify the impact of cloud clusters on distributed photovoltaic power fluctuations; (4) Combined with neural networks, an ultra-short-term distributed photovoltaic power prediction model is constructed to obtain prediction results.

2. The distributed photovoltaic ultra-short-term prediction method based on satellite cloud image information according to claim 1 is characterized in that: In step (1), meteorological variables include temperature, relative humidity, air pressure, wind speed, wind direction, precipitation and shortwave radiation. The Pearson correlation coefficient is used to measure the correlation between meteorological variables and distributed photovoltaic power.

3. The distributed photovoltaic ultra-short-term prediction method based on satellite cloud image information according to claim 2 is characterized in that: In step (2), a convolutional-long short-term memory neural network hybrid model is constructed to perform nonlinear cloud movement prediction; Convolutional neural networks consist of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. To address the influencing factors of nonlinearity, convolutional neural networks introduce activation functions to map nonlinear features to a high-dimensional nonlinear region. Among them, the convolution layer and pooling layer are the key layers, and the convolution operation is expressed as: in, is the weight of the g-th convolution filter in the l-th layer, is the bias of the lth layer, and the input area of ​​the lth layer position (i, j) is expressed as f conv represents the activation function; At the same time, pooling layers are regularly introduced between convolutional layers to perform pooling in each depth dimension separately. The depth of the image remains unchanged, and the final output of the pooling layer is: in, is the result after the convolution operation, Represents the result after pooling, f pool Represents the pooling operation. After the convolutional and pooling layers, each node in the fully connected layer is connected to all nodes in the previous layer, and the features extracted by the previous layer are integrated. The extracted feature map is converted into the final output of the network, resulting in multiple cloud image patches representing local cloud shape and thickness information; The long short-term memory neural network is a neural network model improved based on the recurrent neural network. Compared with the recurrent neural network, the long short-term memory neural network dynamically controls the increase and decrease of information by introducing three gating units: the forget gate, the input gate, and the output gate. At time t, the status of each door is: Forget Gate: Input Gate: Output gate: Where t represents the current time step; t-1 represents the previous time step; W f 、W i 、W o represents the weight matrix of the gate; b f 、b i 、b o Indicates the gate bias; h t-1 is the hidden state output at the previous moment; x t is the current input; σ is the sigmoid activation function; The memory cell normalizes the input by updating the function, which multiplies i t , the obtained gt is used to update the memory unit, the formula is: Where W g represents the weight matrix of the update function; b g is the bias; tanh is the activation function; The update of memory units is the core part of the long short-term memory neural network. Its state is determined by the historical state and the current state. The memory unit will be updated as follows: Where C t-1 Indicates the memory unit at the previous moment; The final output of the long short-term memory neural network is determined by the output gate and the unit state: h t =o t *fishy(c) t )(8)。 4. The distributed photovoltaic ultra-short-term prediction method based on satellite cloud image information according to claim 3 is characterized in that: In step (2): The nonlinear cloud movement prediction method uses historical satellite cloud images at time T-2, T-1, and T as input to predict the cloud image distribution from time T+1 to T+3. To focus on the target area, the original cloud image is cropped into a 60×60 pixel rectangular area centered on the photovoltaic area. A 6-layer convolutional-long short-term memory neural network cascade structure is used to extract the spatiotemporal features of the cloud image through multiple convolution kernels. Each layer of CNN-LSTM outputs a three-dimensional feature map, and finally the 3D features are converted into 2D prediction results through the Conv3D layer. The specific steps of the nonlinear cloud movement prediction method are as follows: Firstly, a 3×3 convolution kernel is used to extract the spatial distribution and thickness characteristics of cloud clusters from continuous satellite cloud images, and the cloud images are divided into multiple small cloud blocks, which represent the local cloud shape and thickness information respectively. Secondly, the long short-term memory neural network is used to analyze the temporal evolution of cloud conditions in each region, learn local dynamic features and integrate them into the overall movement trend. Since the change trends of each cloud block are different, the convolution-long short-term memory neural network is used to learn the movement laws of each region, realize the nonlinear movement prediction of cloud clusters, and use structural similarity measurement as the evaluation indicator.

5. The distributed photovoltaic ultra-short-term prediction method based on satellite cloud image information according to claim 4 is characterized in that: In step (3), a characteristic cloud area positioning algorithm based on the law of solar motion is proposed to identify the cloud area that blocks the incident light of the sun in real time. In the satellite cloud image I(x,y), it is assumed that a distributed photovoltaic O(x0,y0) is located at the center of the image, and the vertical projection of the intersection point S' of the sunlight and the cloud group on the cloud image is S(x s ,y s ), through the solar altitude angle α s and the solar azimuth γ s , combined with the cloud height H and distance L, the solar incidence angle of the photovoltaic area at any moment is calculated. Finally, the cloud clusters that intersect with the propagation trajectory of sunlight are judged as characteristic cloud areas with blocking effects.

6. The distributed photovoltaic ultra-short-term prediction method based on satellite cloud image information according to claim 5, characterized in that: The specific steps of the characteristic cloud area positioning algorithm are as follows: First, calculate the sun's altitude: Where ψ represents latitude; ω represents hour angle; δ is the solar declination angle. The hour angle and solar declination angle are calculated using the following two equations: ω=(t-12)×15° (10) Where t represents a certain moment; d represents the serial number of a certain day in a year; based on the above physical quantities, the solar azimuth angle is calculated by the following formula: According to the right triangle law, L is obtained from H: L=H / tanα S (13) Finally, through L and γ s Obtain the position coordinates of point S, and then calculate the intersection of the sun's rays and the cloud cluster in the satellite cloud image. The specific process is as follows: x S =x O +L×sinγ S (14) and S =and O +L×cosγ S (15) Where x s and y s They are the east-west and north-south distances between point S and point O in the satellite cloud map respectively; based on the above operations, the intersection position is mapped from three-dimensional space to two-dimensional satellite cloud map.

7. The distributed photovoltaic ultra-short-term prediction method based on satellite cloud image information according to claim 6, characterized in that: In step (4), the ultra-short-term distributed photovoltaic power prediction model includes two stages: training and prediction; During the training phase, cloud images at historical moments are first acquired through a dynamic selection algorithm for characteristic cloud regions. These images are then fed into a convolutional neural network. After convolution, normalization, and pooling, the cloud occlusion influencing factor is extracted and combined with other influencing factors. A mapping relationship between this factor and distributed photovoltaic power is established through a fully connected layer. In the prediction stage, the satellite cloud image at the time to be predicted is first input into the convolution-pooling layer of the convolutional neural network. Each convolution-pooling layer consists of a convolution layer, a normalization layer, and a pooling layer arranged in sequence. In the pooling layer, a rank-ordered pooling method is used, with a kernel size of 2×2 and a step size of 2. After two layers of convolution-pooling layers, two fully connected layers are added, each with 1024 neurons. The variables that quantify the cloud occlusion information are obtained from the fully connected layer, which are set as eight cloud occlusion factors here. In addition, eight other influencing variables are selected for fusion modeling with the cloud occlusion factors, namely the temperature, humidity, irradiance, sine value of the solar altitude angle at the time to be predicted, and the historical power at times T-1, T-2, T-3, and T-4. The 16 variables are fused and input into a fully connected layer consisting of 256 neurons. After weighted operation, the distributed photovoltaic power prediction result is obtained.

8. A distributed photovoltaic ultra-short-term prediction system based on satellite cloud image information, characterized in that: include: Data analysis module, used to process meteorological data and analyze the correlation between meteorological variables and distributed photovoltaic power; The cloud prediction module is used to explore the nonlinear evolution of clouds and build a cloud movement prediction framework at multiple time scales; The characteristic cloud region selection module is used to build a dynamic selection algorithm for characteristic cloud regions, accurately quantify the impact of cloud clusters on distributed photovoltaic power fluctuations, locate relevant cloud characteristic regions in real time, and obtain cloud images at historical moments; The power prediction module is used to build an ultra-short-term distributed photovoltaic power prediction model and obtain prediction results.

9. A computer-readable storage medium, characterized in that A program is stored thereon, and when the program is executed by the processor, the steps of the distributed photovoltaic ultra-short-term prediction method based on satellite cloud image information as claimed in claim 1 are implemented.

10. An electronic device, characterized in that: The method comprises a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps of the distributed photovoltaic ultra-short-term prediction method based on satellite cloud image information as claimed in claim 1 are implemented.

Images