Photovoltaic power prediction method and system based on deep learning and multi-source data fusion

Abstract

This invention discloses a photovoltaic power prediction method and system driven by deep learning and multi-source data fusion. The method includes the following steps: collecting photovoltaic power output, multi-source meteorological data, and multi-wavelength channel satellite cloud images of the target area to form a data sample library for the prediction model. This invention proposes a bidirectional dynamic residual mechanism applicable to both long short-term memory neural networks and self-attention convolutional long short-term memory neural networks, effectively improving the problems of gradient explosion and insufficient long-term temporal feature modeling. The improved self-attention convolutional long short-term memory neural network is used to predict satellite cloud images at the prediction time. Guided by future cloud images, the prediction model can fully capture the spatiotemporal dynamic information in historical cloud image sequences. This method can be used for practical distributed photovoltaic power prediction, improving the generalization and robustness of the prediction method and ensuring the safe and stable operation of the power system.

Description

Technical Field

[0001] This invention belongs to the field of photovoltaic power generation technology, specifically relating to a photovoltaic power prediction method and system driven by deep learning and multi-source data fusion. Background Art

[0002] The continued rapid development of new energy sources is an inevitable trend under the current energy transition and policy background. With the proposal of the "dual carbon" goals of "peaking carbon in 2030 and achieving carbon neutrality in 2060", the installed capacity of distributed photovoltaic power generation has grown rapidly in recent years, driven by both cost reduction and increased sustainability requirements.

[0003] The volatility of photovoltaic (PV) power generation manifests itself in the fact that even if the predicted output value at a certain moment is accurate, PV output can still fluctuate drastically over a smaller timescale due to cloud disturbances, affecting the stability of the power system and power quality. Therefore, the need for accurate power prediction is increasingly urgent. Current PV power prediction technologies suffer from problems such as limited data, insufficient information, low sample quality, and missing key characteristic information. Furthermore, due to the uncertainty of atmospheric motion, PV power sequences, especially PV power generation sequences, exhibit strong volatility and high noise, significantly reducing their measurability.

[0004] To address these issues, we propose a photovoltaic power prediction method and system based on deep learning and multi-source data fusion to solve the problems mentioned in the background section. Summary of the Invention

[0005] The purpose of this invention is to provide a photovoltaic power prediction method and system based on deep learning and multi-source data fusion to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a photovoltaic power prediction method driven by deep learning and multi-source data fusion, comprising the following steps:

[0007] S1. Collect photovoltaic power output, multi-source meteorological data, and multi-wavelength channel satellite cloud images of the target area to form a data sample library for the prediction model. The collected data sequence is as follows:

[0008] I = [i t-n i t-n+1 ,...,i t ]

[0009] P = [p] t-n ,p t-n+1 ,...,p t ]

[0010] W = [w t+1 ,w t+2,...,w t+m ];

[0011] S2. Perform data cleaning on photovoltaic power output and multi-source meteorological data, and use the pyramid algorithm to fuse multi-wavelength channel satellite cloud images. Use the multi-level pyramid algorithm to fuse satellite cloud images of visible light, near infrared light, infrared light, water vapor and other spectral bands to obtain a fused satellite cloud image that can fully characterize cloud information.

[0012] S3. Based on historical multi-step satellite cloud images, an improved self-attention convolutional long short-term memory neural network is used to predict satellite cloud images at future prediction times. Two-dimensional and three-dimensional convolutional neural network models are used to extract the spatial distribution characteristics of the predicted cloud images and the spatiotemporal dynamic characteristics of historical cloud images, respectively. The dimensions of the historical satellite cloud images are N, H, W, and C; N is the number of historical cloud image frames, H and W are the height, width, and number of channels of the cloud image, respectively, and C is the length of the historical satellite cloud image sequence used. Continuous historical satellite cloud images are used as input, and the improved self-attention convolutional long short-term memory neural network is used for image prediction to obtain the satellite cloud image at the prediction time. The improved self-attention convolutional long short-term memory neural network is defined as follows:

[0013]

[0014] S4. The photovoltaic power sequence is decomposed using the wavelet transform algorithm to remove noise and obtain the inherent characteristic components. The process can be represented as follows:

[0015]

[0016] The process of capturing the temporal characteristics of intrinsic feature components based on an improved long short-term memory neural network can be described as follows:

[0017]

[0018] Extracting deep features from multi-source weather forecast data using a multilayer perceptron:

[0019]

[0020] S5. Using the spatiotemporal characteristics of historical cloud images, the spatial characteristics of predicted cloud images, the temporal characteristics of historical power, and the deep meteorological characteristics as inputs to the multilayer sensor, the photovoltaic power prediction result O is obtained. The process can be described as follows:

[0021]

[0022] O = [o t+1 ,o t+2 ,...,o t+m ];

[0023] Where O represents the multi-step prediction result, [o t+1 ,o t+2 ,...,o t+m [This refers to the photovoltaic power prediction results for each time point from t+1 to t+m;]

[0024] S6. Based on real-time new data, use incremental update methods to optimize and correct the prediction model in real time.

[0025] Preferably, in the S1 formula: I, P, and W represent satellite cloud imagery, photovoltaic power, and weather forecast data, respectively, i t Let t be the set of satellite cloud images with multiple wavelength channels, n be the length of the historical sequence, and m be the length of the predicted sequence.

[0026] Preferably, in step S2, the photovoltaic power and multi-source weather forecast data are cleaned, including outlier screening and data imputation. Outliers in the original data are identified and deleted using the box plot method. Missing data are queried along the time series direction and filled using the K-nearest neighbor algorithm.

[0027] Preferably, in the S3 formula: where x t h t-1 Let be the input value at time t and the output value at time t-1; For x t Updated features obtained through the self-attention module; for h t-1 The updated values ​​obtained by sequentially passing through the dynamic residual module and the self-attention module; c t c t-1 Represents the cell state at time t and time t-1; BDR is the dynamic residual module; SA is the self-attention memory module; i t f t g t These are the input gate, forget gate, and output gate, respectively; W xi W hi b i W xf W hf b f W xc W hc b c W xg W hg b g Here are the learning parameters; tanh is the hyperbolic tangent function, and σ is the activation function. Represents the Hadamard product; h t Let t be the cell state at time t.

[0028] Preferably, a two-dimensional convolutional neural network model is used to extract features from the satellite cloud image at the predicted time to obtain the spatial distribution features of the predicted cloud image. The process can be described as follows:

[0029]

[0030] in, To predict the spatial distribution characteristics of cloud maps, f 2DCNN For a two-dimensional convolutional neural network model, [i t+1 i t+2 ,...,i t+m [This is for predicting satellite cloud image sequences.]

[0031] Preferably, feature extraction is performed on historical satellite cloud image sequences using a three-dimensional convolutional neural network model to obtain the spatiotemporal dynamic features of the cloud layer. The process can be described as follows:

[0032]

[0033] in, For the spatiotemporal dynamic characteristics of historical cloud map sequences, f 3DCNN For a three-dimensional convolutional neural network model, [i t-n i t-n+1 ,...,i t [This is a sequence of historical satellite cloud images.]

[0034] Preferably, in S4, p denoised The result is the denoised photovoltaic power sequence, where t is the time variable. W represents the summation over all scales s and all frequency bands k. k,s These are the wavelet coefficients after thresholding, where k represents the wavelet's frequency band, s represents the scaling parameter, and ψ... k,s It is a wavelet function corresponding to scale s and frequency band k;

[0035] in, For the time-series characteristics of historical photovoltaic power series, f IM-LSTM For the improved Long Short-Term Memory neural network model, This is the denoised historical photovoltaic power sequence;

[0036] For the deep meteorological characteristics of multi-source meteorological forecast data, [w t+1 ,w t+2 ,...,w t+m [f] represents the meteorological forecast data for each predicted time point from t+1 to t+m. MLP This is a multilayer perceptron model.

[0037] A photovoltaic power prediction system driven by deep learning and multi-source data fusion includes:

[0038] The data acquisition module collects photovoltaic power output, multi-source meteorological data, and multi-wavelength channel satellite cloud images of the target area to form a data sample library for the prediction model;

[0039] The data preprocessing module cleans the photovoltaic power output and multi-source meteorological data, and uses the pyramid algorithm to fuse multi-wavelength channel satellite cloud images;

[0040] The feature extraction module, based on historical multi-step satellite cloud images, uses an improved self-attention convolutional long short-term memory neural network to predict satellite cloud images at future prediction times. It extracts the spatial distribution characteristics of the predicted cloud images and the spatiotemporal dynamic characteristics of historical cloud images through two-dimensional convolutional neural network and three-dimensional convolutional neural network models, respectively. It decomposes the photovoltaic power sequence using wavelet transform algorithm to obtain inherent feature components. Based on the improved long short-term memory neural network, it captures the temporal characteristics of the inherent feature components. Finally, it uses a multilayer perceptron to extract deep features from multi-source weather forecast data.

[0041] The power prediction module takes the spatiotemporal characteristics of historical cloud images, the spatial characteristics of predicted cloud images, the temporal characteristics of historical power, and the deep meteorological characteristics as inputs to the multilayer sensor to obtain multi-step prediction results.

[0042] The incremental update module optimizes and corrects the prediction model in real time based on newly added data using the incremental update method.

[0043] A computer device, comprising:

[0044] one or more processors;

[0045] Memory;

[0046] And one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by one or more processors, wherein when the programs are executed by the processors, they implement the steps of the photovoltaic power prediction method driven by deep learning and multi-source data fusion as described.

[0047] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the photovoltaic power prediction method driven by deep learning and multi-source data fusion as described above.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] (1) This invention proposes a bidirectional dynamic residual mechanism applicable to long short-term memory neural networks and self-attention convolutional long short-term memory neural networks, which effectively improves the problems of gradient explosion and insufficient modeling of long-term temporal features.

[0050] (2) The present invention uses an improved self-attention convolutional long short-term memory neural network to predict satellite cloud images at the prediction time. Guided by future cloud images, the prediction model can fully capture the spatiotemporal dynamic information in the historical cloud image sequence.

[0051] (3) This invention can be used for actual distributed photovoltaic power prediction, improve the generalization and robustness of the prediction method, and ensure the safe and stable operation of the power system. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the photovoltaic power prediction method of the present invention;

[0053] Figure 2 This is a schematic diagram of the self-attention convolutional long short-term memory neural network used in this invention;

[0054] Figure 3 This is a schematic diagram of the bidirectional dynamic residual mechanism proposed in this invention;

[0055] Figure 4 This is a schematic diagram of the two-dimensional convolutional neural network used in this invention;

[0056] Figure 5 This is a schematic diagram of the prediction result fitting using the proposed method in an embodiment of the present invention. Detailed Implementation

[0057] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only 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 ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0058] Please see Figure 1-5 The present invention provides a technical solution:

[0059] A photovoltaic power prediction method driven by deep learning and multi-source data fusion includes the following steps:

[0060] S1. Collect photovoltaic power output, multi-source meteorological data, and multi-wavelength channel satellite cloud images of the target area to form a data sample library for the prediction model. The collected data sequence is as follows:

[0061] I = [i t-n i t-n+1 ,...,i t ]

[0062] P = [p] t-n ,p t-n+1 ,...,p t]

[0063] W = [w t+1 ,w t+2 ,...,w t+m ];

[0064] S2. Perform data cleaning on photovoltaic power output and multi-source meteorological data, and use the pyramid algorithm to fuse multi-wavelength channel satellite cloud images. Use the multi-level pyramid algorithm to fuse satellite cloud images of visible light, near infrared light, infrared light, water vapor and other spectral bands to obtain a fused satellite cloud image that can fully characterize cloud information.

[0065] S3. Based on historical multi-step satellite cloud images, an improved self-attention convolutional long short-term memory neural network is used to predict satellite cloud images at future prediction times. Two-dimensional and three-dimensional convolutional neural network models are used to extract the spatial distribution characteristics of the predicted cloud images and the spatiotemporal dynamic characteristics of historical cloud images, respectively. The dimensions of the historical satellite cloud images are N, H, and W; N is the number of historical cloud image frames, H and W are the height, width, and number of channels of the cloud image, respectively, and C is the length of the historical satellite cloud image sequence used. Continuous historical satellite cloud images are used as input, and the improved self-attention convolutional long short-term memory neural network is used for image prediction to obtain the satellite cloud image at the prediction time. The improved self-attention convolutional long short-term memory neural network is defined as follows:

[0066]

[0067] S4. The photovoltaic power sequence is decomposed using the wavelet transform algorithm to remove noise and obtain the inherent characteristic components. The process can be represented as follows:

[0068]

[0069] The process of capturing the temporal characteristics of intrinsic feature components based on an improved long short-term memory neural network can be described as follows:

[0070]

[0071] Extracting deep features from multi-source weather forecast data using a multilayer perceptron:

[0072]

[0073] S5. Using the spatiotemporal characteristics of historical cloud images, the spatial characteristics of predicted cloud images, the temporal characteristics of historical power, and the deep meteorological characteristics as inputs to the multilayer sensor, the photovoltaic power prediction result O is obtained. The process can be described as follows:

[0074]

[0075] O = [o t+1 ,ot+2 ,...,o t+m ];

[0076] Where O represents the multi-step prediction result, [o t+1 ,o t+2 ,...,o t+m [This refers to the photovoltaic power prediction results for each time point from t+1 to t+m;]

[0077] S6. Based on real-time new data, use incremental update methods to optimize and correct the prediction model in real time.

[0078] In S1, photovoltaic power output, multi-source meteorological data, and multi-wavelength channel satellite cloud images of the target area are collected to form a data sample library for the prediction model. The collected data sequence is as follows:

[0079] I = [i t-n i t-n+1 ,...,i t ]

[0080] P = [p] t-n ,p t-n+1 ,...,p t ]

[0081] W = [w t+1 ,w t+2 ,...,w t+m ]

[0082] Where I, P, and W represent satellite cloud imagery, photovoltaic power, and weather forecast data, respectively, i t is a set of satellite cloud images with multiple wavelength channels at time t, where n is the length of the historical sequence and m is the length of the predicted sequence;

[0083] In step S2 of this invention, data cleaning is performed on photovoltaic power and multi-source weather forecasts, including outlier screening and data imputation; outliers in the original data are identified and deleted using the box plot method; missing data are queried along the time series direction and filled using the K-nearest neighbor algorithm;

[0084] By using a multi-level pyramid algorithm, satellite cloud images in the visible light, near-infrared, infrared light, and water vapor spectral bands are fused to obtain a fused satellite cloud image that can fully characterize cloud information.

[0085] In S3 of this invention, the dimensions of the historical satellite cloud image are N, H, W, and C; N is the number of historical cloud image frames, H and W are the height, width, and number of channels of the cloud image, respectively, and C is the length of the historical satellite cloud image sequence used; for example... Figure 2As shown, continuous historical satellite cloud images are used as input, and an improved self-attention convolutional long short-term memory neural network is used for image prediction to obtain the satellite cloud image at the predicted time. The improved self-attention convolutional long short-term memory neural network is defined as follows:

[0086]

[0087] Where, x t h t-1 Let be the input value at time t and the output value at time t-1; For x t Updated features obtained through the self-attention module; for h t-1 The updated values ​​obtained by sequentially passing through the dynamic residual module and the self-attention module; c t c t-1 Represents the cell state at time t and time t-1; BDR is the dynamic residual module; SA is the self-attention memory module; i t f t g t These are the input gate, forget gate, and output gate, respectively; W xi W hi b i W xf W hf b f W xc W hc b c W xg W hg b g σ represents the learnable parameters; tanh is the hyperbolic tangent function, and σ is the activation function. Represents the Hadamard product; h t The cell state at time t; such as Figure 3 As shown, the specific implementation process of the two-way dynamic residual mechanism can be represented by the following formula:

[0088]

[0089] Where v represents the longitudinal residual information, z represents the lateral residual information, e represents the number of layers in the convolutional long short-term memory neural network, t represents the time series length, tanh represents the hyperbolic tangent function, and σ represents the activation function. Indicates the Hadamardi product, W h1 、W v1 b v1 、W h2 、W v2 b v2 、W h3 、W z1 bz1 、W h4 、W z2 b z2 These are learnable parameters;

[0090] The spatial distribution features of the predicted cloud image are obtained by extracting features from the satellite cloud image at the predicted time using a 2DCNN model. The process can be described as follows:

[0091]

[0092] in, To predict the spatial distribution characteristics of cloud maps, f 2DCNN For a 2DCNN model, [i t+1 i t+2 ,...,i t+m [This refers to] predicting satellite cloud image sequences; such as... Figure 4 As shown, the 2DCNN model can be represented by the following formula:

[0093]

[0094] in, It is the input of the l-th layer. It is the c-th convolutional kernel in the l-th layer, a (l-1) This is the 2D convolutional output feature map of layer l-1, where * indicates the convolution operation. It is the bias of the c-th convolutional kernel. is the activation value of the c-th feature map in the l-th layer at position (x, y), and f is the mapping function;

[0095] Feature extraction of historical satellite cloud image sequences using a 3DCNN model yields spatiotemporal dynamic features of clouds. The process can be described as follows:

[0096]

[0097] in, For the spatiotemporal dynamic characteristics of historical cloud map sequences, f 3DCNN For a 3DCNN model, [i t-n i t-n+1 ,...,i t [This refers to] a historical satellite cloud image sequence; the 3DCNN model can be represented by the following formula:

[0098]

[0099] in, It is the activation value of the c-th feature map in the l-th layer at position (x, y, n). It is the input of the l-th layer;

[0100] In step S4 of this invention, the photovoltaic power sequence is decomposed using a wavelet transform algorithm to remove noise and obtain inherent feature components. This process can be represented as follows:

[0101]

[0102] Where, p denoised The result is the denoised photovoltaic power sequence, where t is the time variable. W represents the summation over all scales s and all frequency bands k. k,s These are the wavelet coefficients after thresholding, where k represents the wavelet's frequency band, s represents the scaling parameter, and ψ... k,s It is a wavelet function corresponding to scale s and frequency band k; the selection process of wavelet coefficients is as follows:

[0103]

[0104] Among them, W original These are the coefficients obtained from the original wavelet transform; θ is the threshold used to determine which coefficients are considered noise.

[0105] The process of capturing the temporal characteristics of intrinsic feature components based on an improved long short-term memory neural network can be described as follows:

[0106]

[0107] in, For the time-series characteristics of historical photovoltaic power series, f IM-LSTM For the improved Long Short-Term Memory neural network model, This is the denoised historical photovoltaic power sequence;

[0108] Extracting deep features from multi-source weather forecast data using a multilayer perceptron:

[0109]

[0110] in, For the deep meteorological characteristics of multi-source meteorological forecast data, [w t+1 ,w t+2 ,...,w t+m [f] represents the meteorological forecast data for each predicted time point from t+1 to t+m. MLP It is a multilayer perceptron model;

[0111] In step S5 of this invention, the spatiotemporal dynamic features of historical cloud maps, the spatial features of predicted cloud maps, and the time-series features of historical power are used as inputs to a multilayer perceptron to obtain the photovoltaic power prediction result O. This process can be described as follows:

[0112]

[0113] O = [o t+1 ,o t+2 ,...,o t+m ]

[0114] In step S6 of this invention, based on real-time new data, an incremental update method is used to optimize and correct the prediction model in real time, ensuring that the prediction model can continuously improve over time and better adapt to new data and environmental changes.

[0115] To verify the performance of the method proposed in this invention, the following experiments were conducted in the embodiments. The dataset used was obtained from [database name missing]. Model performance was evaluated using quantitative error metrics. The satellite cloud images and photovoltaic power data used in this study came from a photovoltaic power station in China (2017-2018). 80% of this dataset was used for training and 20% for testing. Three error analysis metrics were used in this embodiment: Mean Error (MAE), Root Mean Square Error (RMSE), and Coefficient of Determination (R²). 2 ), defined as follows:

[0116]

[0117]

[0118] Where N represents the total number of test samples, y i and Let these represent the predicted value and the actual value of the i-th sample, respectively. This represents the mean of the actual values.

[0119] To visually reflect the prediction results, Figure 5 The photovoltaic power generation prediction curve obtained one hour in advance using the method of this invention is shown, demonstrating that the method can accurately predict photovoltaic power. To further evaluate the prediction performance of the prediction model, four comparison models were set up, denoted as S1, S2, S3, and S4. S1 uses Transformer to extract temporal features, S2 uses ConvLSTM to extract spatiotemporal features of cloud images, and S3 uses SA-ConvLSTM to extract feature information from satellite cloud images. Furthermore, Table 1 shows the error indices for multi-step prediction of each model. It can be seen that, compared to the comparison models, the model constructed based on this invention has smaller prediction errors and superior performance.

[0120] Table 1 Comparison of prediction results from different models

[0121]

[0122]

[0123] A photovoltaic power prediction system driven by deep learning and multi-source data fusion includes:

[0124] The data acquisition module collects photovoltaic power output, multi-source meteorological data, and multi-wavelength channel satellite cloud images of the target area to form a data sample library for the prediction model;

[0125] The data preprocessing module cleans the photovoltaic power output and multi-source meteorological data, and uses the pyramid algorithm to fuse multi-wavelength channel satellite cloud images;

[0126] The feature extraction module, based on historical multi-step satellite cloud images, uses an improved self-attention convolutional long short-term memory neural network to predict satellite cloud images at future prediction times; it extracts the spatial distribution features of the predicted cloud images and the spatiotemporal dynamic features of historical cloud images through 2DCNN and 3DCNN models, respectively; it decomposes the photovoltaic power sequence using a wavelet transform algorithm to obtain inherent feature components; it captures the temporal characteristics of the inherent feature components based on the improved long short-term memory neural network; and it extracts deep features from multi-source weather forecast data using a multilayer perceptron.

[0127] The power prediction module takes the spatiotemporal characteristics of historical cloud images, the spatial characteristics of predicted cloud images, the temporal characteristics of historical power, and the deep meteorological characteristics as inputs to the multilayer sensor to obtain multi-step prediction results.

[0128] The incremental update module optimizes and corrects the prediction model in real time based on the newly added data using the incremental update method.

[0129] It should be understood that the photovoltaic power prediction system based on deep learning and multi-source data fusion in the embodiments of the present invention can realize all the technical solutions in the above method embodiments. The functions of each functional module can be specifically implemented according to the methods in the above method embodiments. The specific implementation process can be referred to the relevant descriptions in the above embodiments, which will not be repeated here.

[0130] The present invention also provides a computer device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, wherein when the programs are executed by the processors, they implement the steps of the photovoltaic power prediction method driven by deep learning and multi-source data fusion as described above.

[0131] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the photovoltaic power prediction method driven by deep learning and multi-source data fusion as described above.

[0132] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus (systems), computer devices, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0133] This invention is described with reference to a flowchart of a method according to embodiments of the invention. It should be understood that each step in the flowchart and combinations thereof can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the process. Figure 1 A device for a function specified in one or more processes.

[0134] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 The function specified in one or more processes.

[0135] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 Steps of a specified function in one or more processes.

[0136] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A photovoltaic power prediction method based on deep learning and multi-source data fusion, characterized in that, Includes the following steps: S1. Collect photovoltaic power output, multi-source meteorological data, and multi-wavelength channel satellite cloud images of the target area to form a data sample library for the prediction model. The collected data sequence is as follows: I=[i t-n ,i t-n+1 ,...,i t ] P=[p t-n ,p t-n+1 ,...,p t ] In=[in t+1 ,In t+2 ,...,In t+m ]; S2. Perform data cleaning on photovoltaic power output and multi-source meteorological data, and use the pyramid algorithm to fuse multi-wavelength channel satellite cloud images. Use the multi-level pyramid algorithm to fuse satellite cloud images in visible light, near infrared, infrared light and water vapor spectral bands to obtain a fused satellite cloud image that can fully characterize cloud information. S3. Based on historical multi-step satellite cloud images, an improved self-attention convolutional long short-term memory neural network is used to predict satellite cloud images at future prediction times. Two-dimensional and three-dimensional convolutional neural network models are used to extract the spatial distribution characteristics of the predicted cloud images and the spatiotemporal dynamic characteristics of historical cloud images, respectively. The dimensions of the historical satellite cloud images are N, H, W, and C; N is the number of historical cloud image frames, H and W are the height, width, and number of channels of the cloud image, respectively, and C is the length of the historical satellite cloud image sequence used. Continuous historical satellite cloud images are used as input, and the improved self-attention convolutional long short-term memory neural network is used for image prediction to obtain the satellite cloud image at the prediction time. The improved self-attention convolutional long short-term memory neural network is defined as follows: S4. The photovoltaic power sequence is decomposed using the wavelet transform algorithm to remove noise and obtain the inherent characteristic components. The process can be represented as follows: The process of capturing the temporal characteristics of intrinsic feature components based on an improved long short-term memory neural network can be described as follows: Extracting deep features from multi-source weather forecast data using a multilayer perceptron: S5. Using the spatiotemporal characteristics of historical cloud images, the spatial characteristics of predicted cloud images, the temporal characteristics of historical power, and the deep meteorological characteristics as inputs to the multilayer sensor, the photovoltaic power prediction result O is obtained. The process is described as follows: O=[o t+1 ,o t+2 ,...,o t+m ]; Where O represents the multi-step prediction result, [o t+1 ,o t+2 ,...,o t+m [This refers to the photovoltaic power prediction results for each time point from t+1 to t+m;] S6. Based on real-time new data, use incremental update methods to optimize and correct the prediction model in real time; In the formula S1: I, P, and W represent satellite cloud imagery, photovoltaic power, and weather forecast data, respectively; i t is a set of satellite cloud images with multiple wavelength channels at time t, where n is the length of the historical sequence and m is the length of the predicted sequence; In the S3 formula: where x t h t-1 Let be the input value at time t and the output value at time t-1; For x t Updated features obtained through the self-attention module; for h t-1 The updated values ​​obtained by sequentially passing through the dynamic residual module and the self-attention module; c t c t-1 Represents the cell state at time t and time t-1; BDR is the dynamic residual module; SA is the self-attention memory module; i t f t g t These are the input gate, forget gate, and output gate, respectively; W xi W hi b i W xf W hf b f W xc W hc b c W xg W hg b g Here are the learning parameters; tanh is the hyperbolic tangent function, and σ is the activation function. Represents the Hadamard product; h t The state of the cell at time t; p in S4 denoised The result is the denoised photovoltaic power sequence, where t is the time variable. W represents the summation over all scales s and all frequency bands k. k,s These are the wavelet coefficients after thresholding, where k represents the wavelet's frequency band, s represents the scaling parameter, and ψ... k,s It is a wavelet function corresponding to scale s and frequency band k; in, For the time series characteristics of historical photovoltaic power sequences, f IM-LSTM For the improved Long Short-Term Memory neural network model, This is the denoised historical photovoltaic power sequence; For the deep meteorological characteristics of multi-source meteorological forecast data, [w t+1 ,w t+2 ,...,w t+m [f] represents the meteorological forecast data for each predicted time point from t+1 to t+m. MLP This is a multilayer perceptron model.

2. The photovoltaic power prediction method based on deep learning and multi-source data fusion as described in claim 1, characterized in that: In S2, photovoltaic power and multi-source weather forecasts undergo data cleaning, including outlier screening and data imputation. Outliers in the original data are identified and deleted using the box plot method. Missing data are queried along the time series direction and filled using the K-nearest neighbor algorithm.

3. The photovoltaic power prediction method based on deep learning and multi-source data fusion as described in claim 1, characterized in that: Feature extraction of satellite cloud image features at the predicted time is performed using a two-dimensional convolutional neural network model to obtain the spatial distribution features of the predicted cloud image. The process can be described as follows: in, To predict the spatial distribution characteristics of cloud maps, f 2DCNN For a two-dimensional convolutional neural network model, [i t+1 ,i t+2 ,...,i t+m [This is for predicting satellite cloud image sequences.] 4. The photovoltaic power prediction method based on deep learning and multi-source data fusion as described in claim 1, characterized in that: Feature extraction of historical satellite cloud image sequences using a three-dimensional convolutional neural network model yields spatiotemporal dynamic features of clouds. This process can be described as follows: in, For the spatiotemporal dynamic characteristics of historical cloud map sequences, f 3DCNN For a three-dimensional convolutional neural network model, [i t-n ,i t-n+1 ,...,i t [This is a sequence of historical satellite cloud images.] 5. The photovoltaic power prediction method based on deep learning and multi-source data fusion as described in any one of claims 1-4, characterized in that: The steps of the photovoltaic power prediction method driven by deep learning and multi-source data fusion are implemented using a photovoltaic power prediction system driven by deep learning and multi-source data fusion. The photovoltaic power prediction system based on deep learning and multi-source data fusion includes: The data acquisition module collects photovoltaic power output, multi-source meteorological data, and multi-wavelength channel satellite cloud images of the target area to form a data sample library for the prediction model; The data preprocessing module cleans the photovoltaic power output and multi-source meteorological data, and uses the pyramid algorithm to fuse multi-wavelength channel satellite cloud images; The feature extraction module, based on historical multi-step satellite cloud images, uses an improved self-attention convolutional long short-term memory neural network to predict satellite cloud images at future prediction times. It extracts the spatial distribution characteristics of the predicted cloud images and the spatiotemporal dynamic characteristics of historical cloud images through two-dimensional convolutional neural network and three-dimensional convolutional neural network models, respectively. It decomposes the photovoltaic power sequence using wavelet transform algorithm to obtain inherent feature components. Based on the improved long short-term memory neural network, it captures the temporal characteristics of the inherent feature components. Finally, it uses a multilayer perceptron to extract deep features from multi-source weather forecast data. The power prediction module takes the spatiotemporal characteristics of historical cloud images, the spatial characteristics of predicted cloud images, the temporal characteristics of historical power, and the deep meteorological characteristics as inputs to the multilayer sensor to obtain multi-step prediction results. The incremental update module optimizes and corrects the prediction model in real time based on newly added data using the incremental update method.

6. A computer device, characterized in that, include: One or more processors; Memory; And one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by one or more processors, wherein when the programs are executed by the processors, they implement the steps of the photovoltaic power prediction method based on deep learning and multi-source data fusion as described in any one of claims 1-4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the photovoltaic power prediction method based on deep learning and multi-source data fusion as described in any one of claims 1-4.

Images