A photovoltaic output power prediction method, a terminal device and a storage medium

Abstract

The application relates to a photovoltaic output power prediction method, a terminal device and a storage medium, the method comprising the following steps: collecting satellite images and extracting cloud movement features and cloud position features; extracting global information and local information of two types of band images, namely visible and invisible band images, and extracting global feature vectors and local feature vectors; performing weighted summation on the four feature vectors based on an attention mechanism to obtain enhanced image features; predicting rainfall probabilities at each moment based on historical meteorological data through a decision tree model to obtain prediction values; and performing weighted summation on the cloud movement features, the cloud position features, the image features, the rainfall probability prediction values at the current moment and the photovoltaic output power values at the current moment based on the attention mechanism, and then obtaining prediction values of photovoltaic output power values at the next moment through an LSTM network. The application further improves the accuracy of a deep learning model in the photovoltaic prediction field.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of photovoltaic power generation, and particularly relates to a photovoltaic output power prediction method, a terminal device and a storage medium. BACKGROUND

[0002] At present, the main energy on the earth is still fossil energy, but the fossil energy will be exhausted one day, so countries are very active in exploring new energy, including photovoltaic energy, which is an important direction of new energy research. Photovoltaic energy can be obtained anywhere in the world and does not produce pollution during operation, and is a completely green energy. Therefore, how to improve the utilization rate of photovoltaic energy is a very worthy research topic. At present, due to the influence of weather, installation angle, position and other factors, and the influence of seasonal factors, the current photovoltaic prediction is difficult to fully consider all the influencing factors. SUMMARY

[0003] In order to solve the above problems, the present application provides a photovoltaic output power prediction method, a terminal device and a storage medium.

[0004] The specific scheme is as follows:

[0005] A photovoltaic output power prediction method, comprising the following steps:

[0006] S1: collecting satellite images, obtaining cloud motion features corresponding to each time satellite images through an optical flow method, and obtaining cloud position features corresponding to each time satellite images through an image feature extraction model;

[0007] S2: selecting visible band images and invisible band images from the satellite images, and extracting global information and local information of the two types of band images respectively, and then obtaining global images and local images corresponding to the two types of band images;

[0008] S3: processing the global images and the local images corresponding to the two types of band images through a convolutional neural network respectively, to obtain global feature vectors and local feature vectors corresponding to the two types of band images;

[0009] S4: based on the global feature vectors and the local feature vectors corresponding to the two types of band images, calculating the image features of the satellite images through the following formula

[0010]

[0011]

[0012] alpha pi = MLP (Wh p , WI i )

[0013] α p j=MLP(Wh p ,WI j )

[0014] in, represents the concatenation of four feature vectors, σ(·) represents the activation function, β pi Represents the weight coefficient of the i-th eigenvector, W p Represents the second parameter matrix, I k represents the kth eigenvector, I i represents the i-th eigenvector, I j Represents the jth eigenvector, i, k, j all represent the sequence number of the eigenvector, i∈[1,4], j∈[1,4], and j≠i, k∈[1,4], α pi represents the fusion feature of the i-th feature vector, α pj represents the fusion feature of the jth feature vector, MLP(·) represents the neural network layer, W represents the first parameter matrix, h p represents the photovoltaic output power eigenvector;

[0015] S5: Based on historical meteorological data, the rainfall probability at each moment is predicted using a decision tree model to obtain a predicted value;

[0016] S6: Based on the attention mechanism, the cloud motion features, cloud position features, image features of the current image, the rainfall probability prediction value at the current moment, and the photovoltaic output power value at the current moment are weighted and summed, and the photovoltaic output power value prediction value at the next moment is obtained through the LSTM network.

[0017] Furthermore, the image feature extraction model adopts the VGG network model.

[0018] Furthermore, the process of acquiring satellite images is as follows: based on satellite remote sensing data, aerosol optical depth data is integrated and then converted into satellite images.

[0019] Furthermore, global information is obtained through Fourier forward transform and inverse Fourier transform, and local information is obtained through high-pass filtering.

[0020] Furthermore, in step S4, the historical meteorological data includes features of multiple dimensions. By calculating the information gain of each feature, the feature with the largest information gain is selected as the root node of the decision tree. The number of branches of the tree is equal to the number of values ​​of the feature to which the root node belongs.

[0021] A photovoltaic output power prediction terminal device includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the method described above are implemented.

[0022] A computer readable storage medium stores a computer program, and the computer program, when executed by a processor, implements the steps of the method of the above-mentioned embodiment of the application.

[0023] The application adopts the above technical solution and has the beneficial effects:

[0024] (1) The influence of position perception in satellite images is innovatively explored, and a dynamic and static combination idea is adopted, not only optical flow analysis is used to learn motion features at different times, but also rich image features of satellite images under static conditions are fully considered.

[0025] (2) An attention mechanism is introduced to learn global information and local information of different band satellite cloud images.

[0026] (3) Better feature representation is achieved through the idea of image enhancement and denoising.

[0027] (4) The influence of meteorological factors is also considered to integrate meteorological factors into the photovoltaic power prediction task to better assist the time series prediction task. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 The satellite image sequence optical flow calculation process in the embodiment one of the application is shown.

[0029] Figure 2 The visible light band image and the invisible light band image in the embodiment one of the application are shown.

[0030] Figure 3 The decision tree model structure diagram in the embodiment one of the application is shown.

[0031] Figure 4 The flowchart of the method in the embodiment one of the application is shown.

[0032] Figure 5 The overall framework diagram of the model in the embodiment one of the application is shown. DETAILED DESCRIPTION

[0033] To further illustrate the embodiments, the application provides accompanying drawings. These drawings are part of the disclosure of the application, mainly used to illustrate the embodiments, and can be used to explain the operating principle of the embodiments in conjunction with the related description of the specification. Those skilled in the art should understand other possible implementation manners and advantages of the application by referring to these contents.

[0034] The application will be further described in conjunction with the accompanying drawings and specific embodiments.

[0035] Embodiment one:

[0036] The embodiment proposes a time series prediction model integrating time, space and weather conditions, named STMLSTM. Similar to the long short-term memory network structure, the cloud position, thickness and motion are considered in the continuous updating process, various weather indicators and image data of different wave bands in history are considered, and a self-attention mechanism is introduced to weigh the importance of different factors, and finally the future photovoltaic output is predicted. The modules of the model are introduced below.

[0037] (1) Cloud motion feature extraction module

[0038] The moving trend of the cloud is very important for photovoltaic prediction, and the cloud shielding will lead to the reduction of photovoltaic efficiency. By extracting the motion features of the cloud, the moving trend of the cloud can be predicted, thereby indirectly helping to predict the trend of photovoltaic change.

[0039] The cloud motion feature extraction module is committed to mining the cloud motion information on the satellite image sequence, and providing a prediction basis for subsequent photovoltaic prediction. Since the optical flow method can find the offset of the pixel points between different frames of the image sequence, and can also calculate the correlation between adjacent frames to obtain the motion information of the objects in the adjacent frames. Therefore, in the embodiment, the optical flow method is used to extract the motion features of the satellite image, and the satellite images are arranged in time sequence to form a satellite image sequence. The optical flow method uses the change of the pixels in the time domain and the correlation between adjacent frames of the satellite image sequence to find the corresponding relationship between the previous frame and the current frame, and calculate the motion information of the cloud between adjacent frames.

[0040] The optical flow method detects the motion of the cloud, that is, analyzes the vector field formed by the motion vector of the pixel stored in the satellite image. When the cloud in the satellite image does not move, the optical flow field in the image should be uniform; when the cloud starts to move, the optical flow vector formed by the moving cloud is inconsistent with the optical flow vector in the background of the satellite image, and the optical flow field is no longer continuous and uniform, so that the motion and position of the cloud can be detected. Therefore, the result of the optical flow calculation contains the structure of the cloud and the motion information of the cloud itself. The advantage of the optical flow method in this case is that it does not require prior knowledge and does not need to artificially construct the features of the image.

[0041] The core of the cloud motion feature extraction module in the embodiment is the Horn-Schunck optical flow method, which is a global method that estimates the dense optical flow field of the satellite image by calculating the optical flow of each pixel in the image. The algorithm converts the problem of finding the optical flow value of the optical flow field into the problem of finding the minimum value of the energy function of the image by constructing an energy function for the image. Assuming that a satellite image sequence P(x, y, t) is given, the optical flow field F(x, y) between the image sequences can be converted into finding the optical flow components m(x, y) and n(x, y) in the x and y directions, respectively. For the satellite image, we assume the following energy function

[0042]

[0043] where P x , P y , P t denote the derivatives of the image sequence in x, y, t dimensions, respectively, (P x m+P y n+P t ) 2 denote the image gray level variation factor, denote the smoothness constraint factor that limits the variation rate of the optical flow components m, n.

[0044] The cloud motion feature extraction module can perform optical flow calculation on the input 8-bit depth satellite image sequence, output a 24-bit depth optical flow image sequence, and convert it to an 8-bit depth optical flow sequence, which is convenient for subsequent feature extraction. This module uses a sliding window to select satellite images in a time sequence with a window size of 2. This is to calculate the optical flow between two consecutive time points. The calculation goal is to minimize the energy function E, which is a generalized polar coordinate problem, solved by Euler-Lagrange equations. After continuous iteration, when the error is less than a pre-set threshold, the final optical flow value is output.

[0045] As Figure 1 shown is an example of a satellite image sequence optical flow calculation process. T1 and T2 are two satellite images at T1 and T2 time points, respectively, and T1 and T2 are consecutive time periods. The optical flow image below is the result of optical flow calculation between T1 and T2 images, and we calculate the optical flow between every two time points of the satellite image sequence.

[0046] For a ground photovoltaic solar power station, the sunlight transmittance on it is an influencing factor affecting the photovoltaic power generation efficiency. Therefore, studying the specific position of the cloud and the thickness of the cloud in the geographical area where the power station is located will be conducive to the accurate prediction of photovoltaic. In order to study the position and thickness information of the cloud, the aerosol optical thickness image is synthesized by satellite remote sensing data in this embodiment.

[0047] The atmosphere of the earth can be considered as an aerosol. Aerosol optical thickness is one of the important optical properties of aerosol, which represents the light transmittance per unit cross section of the vertical atmospheric column in the atmosphere, and is also widely known as atmospheric transmittance. The greater the aerosol optical thickness, the lower the atmospheric transmittance, and the lower the ground light intensity. In this embodiment, aerosol optical thickness data is synthesized based on satellite remote sensing data, and the data is converted into satellite image data.

[0048] In order to obtain more image features while controlling the number of parameters and improve the response rate of the photovoltaic prediction system, the VGG network model is selected in this embodiment to extract location features from satellite images. The VGG network has more layers than the ordinary CNN, but the application of small convolution kernels and the role of the pooling layer reduce the computational complexity and the complexity of the model structure. In the VGG network used in this embodiment, the input satellite image sequence is a fixed size 2401x2401 RGB image sequence, and the 32-bit depth satellite image needs to be converted to an 8-bit depth image in advance. During training, the moving convolution kernel size is 3 and the moving step is 1 pixel. The number of channels in the convolution layer increases from 64 to 512, and a two-dimensional pooling layer is passed after each convolution layer. After a series of convolution layers, the data is compressed into one-dimensional data. Then add a fully connected layer to get the cloud location feature vector of the satellite image. In the entire convolution network, the nonlinear activation function of all hidden layers is set to ReLU to make the network training faster.

[0049] (3) Spectral learning module based on attention

[0050] The satellite image includes invisible bands and visible bands. In this embodiment, the invisible band is selected as the infrared image with a wavelength of 0.86 μm greater than 760 nm. The visible light band is selected as the image with a wavelength of 0.64 μm. As shown in FIG. 2, the left is the image of the visible light band, and the right is the image of the invisible light band. Figure 2

[0051] In this embodiment, the Fourier transform method is used to process the images of the two types of bands. This method decomposes the image into sine and cosine components. It converts the image from the spatial domain to the frequency domain. The formula for Fourier transform of the image is shown in equation 2.

[0052] F(u,v) = ∫∫f(x,y)e -i(ux+vy) dxdy (2)

[0053] where x, y represent the pixel coordinates in the spatial domain, and u, v represent the pixel coordinates in the frequency domain.

[0054] The pixel points of the image are discrete and finite. Therefore, the actual Fourier transform of the image is performed using the discrete Fourier transform. Because the integral in equation 2 becomes a summation. After Fourier forward transform and Fourier inverse transform, we can get the enhanced image information of different bands, that is, the global information.

[0055] ​For the two types of images, the importance of edge information should also be considered. Therefore, high-pass filtering is used in this embodiment to enhance the clear details in the image, obtaining the local information of the image. The high-pass filter attenuates low-frequency signals while allowing high-frequency signals to pass. The inverse Fourier transform moves the zero-frequency component to the center region of the image. Therefore, the diffusion frequency from the center region outward is from low to high. In this method, we shield the low-frequency signals of the part around the center and retain the high-frequency signals.

[0056] Global information and local information can better help us learn the features of the image. The images corresponding to the global information and local information are input into a convolutional neural network (such as VGG16) for feature learning, obtaining four feature vectors: the feature vector visg of the global visible image, the feature vector visl of the local visible image, the feature vector invisg of the global invisible image, and the feature vector invisl of the local invisible image.

[0057] For each satellite image at a time, four feature vectors are obtained. Here, an attention mechanism is introduced in this embodiment to focus on the importance between different feature vectors. By learning the attention mechanism, the importance of different band information and the importance of global and local information can be better revealed.

[0058] For convenience of notation, we use the set I = {I1, I2, I3, I4} to represent the four feature vectors obtained by the spectral method at time t, where I1 to I4 correspond to visg, visl, invisg, and invisl above, respectively. At the same time, the hourly photovoltaic power data is processed to obtain the photovoltaic power feature vector h p , which has a size of 1 × N, where N is the number of photovoltaic power collection times per hour. We use an attention mechanism for each image feature vector. The first step is to use a multi-layer perceptron to learn the nonlinear relationship between the photovoltaic power feature vector h p and a single feature vector I i . As shown in equation 3.

[0059] α pi = MLP (Wh p , WI i )

[0060] α pj = MLP(Wh p , WI j ) (3)

[0061] where α pi represents the fusion feature of the i-th feature vector, α pj represents the fusion feature of the j-th feature vector, MLP(·) represents a neural network layer, W represents a first parameter matrix determined by learning, and Ii denotes the i-th feature vector, i j denotes the j-th feature vector, i∈[1,4], j∈[1,4], and j≠i.

[0062] Therefore, for the photovoltaic power feature vector h p we can calculate the weight coefficient β pi as shown in equation 4.

[0063]

[0064]

[0065] where σ(·) denotes the activation function, β pi denotes the weight coefficient of the i-th feature vector, W p denotes the second parameter matrix, i k denotes the k-th feature vector, h t denotes the final representation of the image features of the satellite image obtained by the spectral method based on the attention mechanism. By learning the global and local information of different band images, it can well assist the photovoltaic power prediction task.

[0066] (4) Weather feature learning

[0067] The weather data of the photovoltaic power station is used to predict whether it will rain in the future, and the rainfall has a significant impact on photovoltaic power generation. However, historical weather data contains many irrelevant parameters, such as wind speed, wind direction, humidity, temperature, etc. In order to find the precipitation pattern from numerous irrelevant meteorological data, a decision tree model is used in this embodiment to extract the precipitation pattern from each historical meteorological parameter. The extracted precipitation pattern is used to predict the rainfall in the future. When some attributes of the data set are missing, the decision tree model can perform well. They can handle irrelevant features well and can be trained quickly without the need for special hardware acceleration, which is crucial when dealing with large data sets. The decision tree model is the best choice for this scenario. This is because meteorological historical data often has many missing and previously irrelevant features, and photovoltaic prediction has a requirement for fast response.

[0068] The embodiment proposes a rainfall probability prediction module which can collect various meteorological data at historical time and predict the rainfall probability of the next day. The core of the prediction algorithm is the ID3 algorithm based on decision tree. The rainfall prediction can be converted into a classification problem of future rainfall probability. Assuming that the sample space of the rainfall prediction system is (F, R), F is the number of samples, which contains m features. In this module, the number of features is 6, which are temperature, evaporation, wind direction, wind speed, humidity and air pressure. R is the number of categories classified by the classification system, which contains n categories. In this module, it represents the probability of raining on the day. For the rainfall probability classification system, the possible choices in the system are D1, D2, …, Dn, and the probability of each category appearing is P(D1), P(D2), …, P(Dn). The entropy of the rainfall probability classification system is shown in formula 6.

[0069]

[0070] In addition, in order to calculate the influence of a certain feature on the system of the classification system, the concept of conditional entropy is introduced. The conditional entropy in the classification system is defined as the information entropy of the sample when the one-dimensional feature F is fixed and represented. For example, H(D|F) is the conditional entropy of the rainfall probability classification system when the feature F=(f1, f2, …, f n ) is fixed.

[0071]

[0072] Among them, the sample feature F in the classification system has a probability P i f i , when the feature F takes a fixed value f i , H(D|F=f i ) is the conditional information entropy.

[0073] The entropy change of the rainfall probability classification prediction system is defined as the information gain specific to a certain feature. That is, the information gain is the difference between the information entropy of the system with the feature and the information entropy of the system without the feature. For example, the information gain of the feature F in the rainfall probability classification system is gain(F, R)=H(D)-H(D|F). The calculation process is to continuously calculate the information gain of each feature, and then select the feature with the maximum information gain, which is the basic principle of the rainfall probability classification system.

[0074] The module calculates the information gain of each feature first in the decision tree calculation based on information entropy, and selects the feature with the maximum information gain as the root node of the decision tree. The number of branches of the tree is the number of values of the characteristic to which the root node belongs. In the classification process, each branch generates a new data subset, which is iteratively recursive until all data subsets are classified into the same class. In this module, multiple meteorological factors are used as features of the classification system, a decision tree model with a depth of 3 is used, and historical weather data is used to predict rainfall. The tree model is as shown in Figure 3 .

[0075] (5) Time series prediction module

[0076] When dealing with problems related to the event timeline, such as photovoltaic power prediction problems, there is a certain correlation between the upper and lower in the time series data, which cannot be solved by traditional neural networks. Recurrent neural network (RNN) solves the problem related to time span by forming a connection matrix between neurons and recursively generating updated prediction data. However, when the interval between relevant information and the current prediction position becomes very large, RNN loses the ability to learn information connected at such a distance, and the root cause is the problem of gradient disappearance or gradient divergence. The photovoltaic output prediction time range in this embodiment is wide, and RNN cannot complete this task, but in the prediction module, the core iteration module of Long Short-term Memory (LSTM) is used, which overcomes the problem of insufficient accuracy in photovoltaic medium and long-term prediction.

[0077] Similar to RNN, LSTM has a repeated recursive structure, but due to the innovative introduction of gate structure in LSTM, it selectively extracts sequence information and reduces information redundancy and gradient disappearance. Therefore, LSTM has the ability to solve medium and long-term time series problems. The key to the gate structure is the dot product operation kernel sigmoid activation neural layer. In the LSTM core module, there are three gate structures for filtering and controlling information, namely the input gate, the forget gate, and the output gate.

[0078] Based on the above model, a photovoltaic output power prediction method is proposed as shown in Figure 4 , including the following steps:

[0079] S1: Collect satellite images, obtain cloud motion features corresponding to each time satellite image through optical flow method, and obtain cloud position features corresponding to each time satellite image through image feature extraction model;

[0080] S2: Select visible band images and invisible band images from the satellite images, and extract global information and local information of the two types of band images respectively, and then obtain global images and local images corresponding to the two types of band images;

[0081] S3: The global image and the local image corresponding to the two types of band images are respectively processed by a convolutional neural network to obtain a global feature vector and a local feature vector corresponding to the two types of band images;

[0082] S4: Based on the global feature vector and the local feature vector corresponding to the two types of band images, an attention mechanism is used to calculate the image features of the satellite image enhancement;

[0083] S5: Based on the historical meteorological data, a decision tree model is used to predict the rainfall probability at each time to obtain a prediction value;

[0084] S6: Based on the attention mechanism, the cloud motion feature, the cloud position feature, the image feature of the current time image, the rainfall probability prediction value of the current time, and the photovoltaic output power value of the current time are weighted and summed, and then the next time photovoltaic output power value prediction value is obtained through the LSTM network.

[0085] The specific method of the embodiment can simultaneously process historical photovoltaic output power data, satellite image data and historical meteorological data, thereby improving the accuracy of photovoltaic prediction.

[0086] The embodiment improves the traditional LSTM model by adding time, space and weather modules, which will be updated with the time sequence of the LSTM model. In addition, since the time granularity of various data is not consistent, a unified time granularity of 15 minutes is selected in this embodiment. For data with a time granularity greater than 15 minutes, the time granularity of various data is unified to 15 minutes using the padding method, which uses the average of the two most recent values to fill in.

[0087] Referring to the overall block diagram of the model as shown in Figure 5 In step S1, the satellite images at t-1 and t time are processed to obtain the optical flow image at t time, and then an independent image feature extraction model is used to extract the optical flow image to obtain the 512-dimensional feature of the cloud. Then, the features of each dimension are accumulated to obtain the cloud motion feature value Tt at t time. Similarly, another independent image feature extraction model processes the satellite image at t time into a 512-dimensional sparse feature, and adds them to obtain the cloud position feature value St at t time. In steps S2-S4, spectral learning is performed on the satellite image at time t to obtain the global information and local information of the two types of wavelength images at t time, and then the attention mechanism is combined with the photovoltaic power feature vector to learn the weight, thereby obtaining the image feature At at t time (At is h tThe input data in step S5 is all the meteorological data from time 0 to time t, and the output is the prediction value Mt of whether it will rain at time t. In step S6, the attention mechanism is applied to Tt, St, At, Mt, and the photovoltaic output power vector xt at time t is focused, and the weighted sum is input into the LSTM module to obtain the predicted photovoltaic output power value at time t. It is worth noting that when t is less than the current time t0, we input the historical photovoltaic power generation data, and when t is greater than t0, the input is the last moment of photovoltaic power generation prediction value. The implicit output ht of each layer of the model is taken as the implicit input of the next layer of the model, and the iteration is performed, and finally the photovoltaic output power prediction sequence of the next period is obtained.

[0088] The embodiment of the present application further improves the accuracy of the deep learning model in the field of photovoltaic prediction, and overcomes the deficiency of lacking response to uncertain disturbance in traditional photovoltaic prediction.

[0089] Embodiment two:

[0090] The present application also provides a photovoltaic output power prediction terminal device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the steps in the above method embodiments of the first embodiment of the present application when executing the computer program.

[0091] Further, as an executable scheme, the photovoltaic output power prediction terminal device can be a desktop computer, a notebook computer, a palm computer, and a cloud server, etc. The photovoltaic output power prediction terminal device can include, but is not limited to, a processor and a memory. Those skilled in the art can understand that the above-mentioned composition structure of the photovoltaic output power prediction terminal device is only an example of the photovoltaic output power prediction terminal device, and does not constitute a limitation on the photovoltaic output power prediction terminal device, and can include more or fewer components than the above, or combine certain components, or different components, for example, the photovoltaic output power prediction terminal device can also include an input and output device, a network access device, a bus, etc., and the present application does not limit this.

[0092] Further, as an executable solution, the processor can be a central processing unit (CPU), and can also be other general-purpose processors, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware component, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor or the like, which is a control center of the photovoltaic output power prediction terminal device and connects various parts of the photovoltaic output power prediction terminal device through various interfaces and lines.

[0093] The memory can be used to store the computer program and / or modules, and the processor realizes various functions of the photovoltaic output power prediction terminal device by running or executing the computer program and / or modules stored in the memory and calling data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system and at least one application required by a function; and the data storage area can store data created according to use of the mobile phone, etc. In addition, the memory can include a high-speed random access memory, and can also include a nonvolatile memory, for example, a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other volatile solid-state memory device.

[0094] The application further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to realize the steps of the method provided in the embodiments of the application.

[0095] The module / unit of the photovoltaic output power prediction terminal device integration, if realized in the form of a software function unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on such an understanding, all or part of the processes in the above-mentioned embodiment methods can also be completed by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, the steps of the above-mentioned various method embodiments can be implemented. The computer program includes computer program code, which can be in the form of source code, object code, an executable file, or some intermediate form. The computer-readable medium can include any entity or device capable of carrying the computer program code, a recording medium, a U disk, a mobile hard disk, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), and a software distribution medium, etc.

[0096] Although the present application has been specifically shown and described with respect to the preferred embodiments, it should be understood that the application is entitled to protection within the full scope of the appended claims, without departing from the spirit and scope of the application.

Claims

1. A photovoltaic output power prediction method, characterized by, The method comprises the following steps: S1: collecting satellite images, obtaining cloud movement features corresponding to the satellite images at each time point by an optical flow method, and obtaining cloud position features corresponding to the satellite images at each time point by an image feature extraction model; S2: selecting visible band images and invisible band images from the satellite images, and extracting global information and local information of the two types of band images respectively, and then obtaining global images and local images corresponding to the two types of band images; S3: processing the global images and the local images corresponding to the two types of band images by a convolutional neural network respectively, and obtaining global feature vectors and local feature vectors corresponding to the two types of band images; S4: Based on the global feature vectors and the local feature vectors corresponding to the two types of waveband images, the image features of the satellite image are calculated by the following formula a pi = MLP(Wi p ,WI i ) a pj = MLP(Wi p , W j i) wherein, denotes the concatenation of 4 feature vectors, σ(·) denotes an activation function, β pi denotes the weight coefficient of the i-th feature vector, W p denotes the second parameter matrix, I k denotes the k-th feature vector, I i denotes the i-th feature vector, I j denotes the j-th feature vector, i, k, j all represent the serial number of the feature vector, i∈[1, 4], j∈[1, 4], and j≠i, k∈[1, 4], α pi denotes the fusion feature of the i-th feature vector, α pj denotes the fusion feature of the j-th feature vector, MLP(·) denotes a neural network layer, W denotes the first parameter matrix, h p denotes a photovoltaic output power feature vector; S5: predicting the rainfall probability at each time point based on historical meteorological data by a decision tree model to obtain a prediction value; S6: based on an attention mechanism, weighting and summing the cloud movement features, the cloud position features, the image features of the current time image, the rainfall probability prediction value of the current time and the photovoltaic output power value of the current time, and then obtaining a prediction value of the photovoltaic output power value of the next time point by an LSTM network.

2. The photovoltaic output power prediction method of claim 1, wherein: The image feature extraction model adopts a VGG network model.

3. The photovoltaic output power prediction method of claim 1, wherein: The acquisition process of the satellite image is: on the basis of satellite remote sensing data, after comprehensively integrating aerosol optical depth data, the satellite image is converted.

4. The photovoltaic output power prediction method of claim 1, wherein: The global information is obtained by Fourier forward transform and Fourier inverse transform, and the local information is obtained by high-pass filtering.

5. The photovoltaic output power prediction method of claim 1, wherein: In step S4, the historical meteorological data includes multiple dimensions of features, the information gain of each feature is calculated, the feature with the largest information gain is selected as the root node of the decision tree, and the number of branches of the tree is equal to the number of values of the feature to which the root node belongs.

6. A photovoltaic output power prediction terminal device characterized by comprising: The computer program is executed by the processor to realize the steps of the method of any one of claims 1-5.

7. A computer readable storage medium storing a computer program, characterized in that: The computer program is executed by the processor to realize the steps of the method of any one of claims 1-5.

Images