A photovoltaic ultra-short-term power prediction method, electronic device and storage medium

Abstract

This invention discloses a method, electronic device, and storage medium for predicting ultra-short-term photovoltaic power, relating to the field of photovoltaic power prediction technology. The method includes: constructing hybrid spatial data of a photovoltaic power station using remote sensing data, solar altitude angle, and solar azimuth angle; converting various two-dimensional data structures into one-dimensional data using linear one-dimensionalization and convolutional one-dimensionalization; performing positional encoding on the one-dimensional data to obtain spatial features of the photovoltaic power station, cloud spatial features, and cloud structural features, and inputting these into a two-stage cloud feature dynamic encoding module to extract cloud features that interfere with photovoltaic power generation; converting the cloud features interfering with photovoltaic power generation into interference factors using a feedforward neural network; obtaining the time series of interference factors from time t-T to t+1 through parallel computation, combining it with historical power and timestamp sequences to form multivariate time-series data, and then feeding it into a hybrid neural network to predict power. This invention can solve the problems of current methods that rely too heavily on explicit formulas and lack sufficient spatial modeling precision.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power prediction technology, and in particular to a photovoltaic ultra-short-term power prediction method, electronic device and storage medium. Background Technology

[0002] Solar photovoltaic (PV) power generation plays a vital role in the energy consumption structure. However, influenced by various meteorological factors, PV power generation can fluctuate dramatically, significantly impacting grid operation. Therefore, improving the accuracy of PV power forecasting is a prerequisite for ensuring the future safe and stable operation of the power grid.

[0003] In satellite cloud imagery-based photovoltaic (PV) power prediction research, capturing cloud features that interfere with PV power generation can effectively improve prediction accuracy. However, current methods have the following shortcomings: 1) Non-end-to-end (empirical calculation) methods are difficult to solve. Calculating the intersection point location requires determining the cloud top height at that location, and vice versa, creating a circular dependency and high solution complexity. Introducing empirical values ​​can easily lead to biases. 2) The spatial modeling of PV power plants lacks refinement. In many hilly areas, due to terrain limitations, PV power plants are often irregularly and dispersed, with a large spatial span. The information on obstructing clouds that can be inferred from the fuzzy location of PV power plants is incomplete. 3) The range of interfering clouds is difficult to determine. The solar irradiance received by a PV power plant is not only determined by directly obstructing clouds but also affected by indirectly obstructing clouds around the power plant. Pre-setting the range of obstructing clouds is difficult to adapt to dynamic changes in clouds.

[0004] Therefore, a new technical solution is urgently needed to solve the above problems. Summary of the Invention

[0005] The purpose of this invention is to provide a photovoltaic ultra-short-term power prediction method, electronic device, and storage medium that can solve the problems of current methods relying too much on explicit formulas and insufficient spatial modeling precision.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A method for predicting ultra-short-term photovoltaic power includes:

[0008] Acquire target photovoltaic data; the target photovoltaic data includes remote sensing data, solar altitude angle, solar azimuth angle, cloud top height data, and satellite cloud images;

[0009] Constructing spatial hybrid data for photovoltaic power plants using remote sensing data, solar altitude angle, and solar azimuth angle;

[0010] Linear one-dimensionalization is used to convert two-dimensional photovoltaic power plant spatial hybrid data and cloud top height data into one-dimensional photovoltaic power plant spatial hybrid data and one-dimensional cloud top height data. Convolution one-dimensionalization is used to convert two-dimensional satellite cloud images into satellite cloud images. Figure 1 Dimensional data;

[0011] Location coding was performed on the one-dimensional spatial hybrid data of photovoltaic power plants, cloud top height data, and satellite cloud image data to obtain the spatial characteristics of photovoltaic power plants, cloud spatial characteristics, and cloud structure characteristics.

[0012] The spatial characteristics of photovoltaic power plants, cloud spatial characteristics, and cloud structure characteristics are input into a two-stage cloud feature dynamic coding module to extract cloud features that interfere with photovoltaic power generation.

[0013] The cloud features that interfere with photovoltaic power generation are converted into interference factors using a feedforward neural network.

[0014] The time series of interference factors from time tT to t+1 is obtained through parallel computing, and combined with the historical power and timestamp series to form multivariate time series data; where t is the current time and T is the length of the historical data;

[0015] Multivariate time series data are fed into a hybrid neural network to predict photovoltaic power.

[0016] Optionally, the construction of spatial hybrid data for photovoltaic power plants using remote sensing data, solar altitude angle, and solar azimuth angle includes:

[0017] Based on the remote sensing data, a spatial distribution map of photovoltaic power stations is determined, and the spatial distribution map is mapped onto a spatial grid map of a meteorological satellite. Then, the solar altitude angle and solar azimuth angle are embedded into the spatial network map to obtain spatial hybrid data of photovoltaic power stations.

[0018] Optionally, the linear one-dimensionalization step specifically includes:

[0019] First, the two-dimensional data Reconstructed into a series of data blocks Then, the data blocks are folded using dimensions. Fold into Finally, through a linear layer... The characteristics are mapped to one-dimensional data Where H and W are the height and width of the two-dimensional data, respectively, and N = HW / B 2 The number of data blocks is represented by (B,B), the size of the data block is represented by B, the width of the data block is represented by C, the number of channels is represented by D, and the number of features is represented by D.

[0020] Optionally, the convolution one-dimensionalization step specifically includes:

[0021] First, the two-dimensional data Reconstructed into a series of data blocks Then use a convolutional neural network to process the data blocks The dimension is converted to Then, the dimensions are folded using dimension folding. Fold into Finally, through a linear layer... The characteristics are mapped to

[0022] Optionally, the two-stage cloud feature dynamic encoding module consists of a series of cross-attention encoders, the calculation formula of which is:

[0023]

[0024] Among them, f MCA Indicates multi-head cross-attention, f MLP Denotes a multilayer perceptron, f Norm This represents normalization, where F1 and F2 are the two inputs to the cross-attention mechanism, respectively. N-1 and F1 N These represent the features of the (N-1)th layer and the Nth layer, respectively.

[0025] Optionally, the specific processing steps for extracting cloud features that interfere with photovoltaic power generation include:

[0026] The spatial features of the photovoltaic power station and the cloud spatial features are input into the first encoder to extract the spatial coupling features between the spatial location of the photovoltaic power station and the cloud location.

[0027] The spatial coupling features and cloud structure features are input into the second encoder to extract cloud features that interfere with photovoltaic power generation.

[0028] The present invention also provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the photovoltaic ultra-short-term power prediction method described above.

[0029] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the photovoltaic ultra-short-term power prediction method as described above.

[0030] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0031] This invention discloses a method, electronic device, and storage medium for predicting ultra-short-term photovoltaic power. The method includes constructing spatial hybrid data of a photovoltaic power station using remote sensing data, solar altitude angle, and solar azimuth angle; converting the two-dimensional structured spatial hybrid data of the photovoltaic power station and cloud top height data into one-dimensional spatial hybrid data of the photovoltaic power station and one-dimensional cloud top height data using linear one-dimensionalization; and converting the two-dimensional structured satellite cloud image into satellite cloud image using convolution one-dimensionalization. Figure 1 This invention employs a three-dimensional approach: spatially hybrid data of a photovoltaic (PV) power plant, cloud top height data, and satellite cloud imagery data are encoded to obtain spatial features of the PV power plant, cloud spatial features, and cloud structure features. These features are then input into a two-stage cloud feature dynamic encoding module to extract cloud features that interfere with PV power generation. A feedforward neural network is used to convert these cloud features into interference factors. The interference factor time series from time tT to t+1 is obtained through parallel computation and combined with historical power and timestamp sequences to form multivariate time series data. This multivariate time series data is then fed into a hybrid neural network to predict PV power. This invention addresses the problems of current methods' over-reliance on explicit formulas and insufficient spatial modeling precision. Attached Figure Description

[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0033] Figure 1 This is a flowchart illustrating the photovoltaic ultra-short-term power prediction method of the present invention;

[0034] Figure 2 This is a framework diagram of the photovoltaic ultra-short-term power prediction method in this embodiment;

[0035] Figure 3 This is a schematic diagram of the method for constructing spatial hybrid data for photovoltaic power plants in this embodiment. Detailed Implementation

[0036] 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.

[0037] The purpose of this invention is to provide a photovoltaic ultra-short-term power prediction method, electronic device, and storage medium that can solve the problems of current methods relying too much on explicit formulas and insufficient spatial modeling precision.

[0038] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0039] like Figure 1-Figure 2 As shown, the present invention provides a photovoltaic ultra-short-term power prediction method, comprising:

[0040] S1: Constructing spatial hybrid data for photovoltaic power plants using remote sensing data, solar altitude angle, and solar azimuth angle; combining... Figure 3 The specific steps for constructing spatial hybrid data for photovoltaic power plants are explained below:

[0041] S1.1: Acquire spatial distribution data. Obtain a detailed spatial distribution map of photovoltaic power plants using remote sensing images from high-resolution remote sensing satellites. Figure 3 (The area where photovoltaic panels are installed);

[0042] S1.2: Spatial Information Mapping. Mapping high-resolution spatial distribution maps onto a spatial grid map of meteorological satellites. Figure 3 (The grid area where the photovoltaic power station is located);

[0043] S1.3: Spatial Hybrid Data Construction. Solar elevation angle and solar azimuth angle are embedded into a spatial grid to obtain spatial hybrid data for photovoltaic power plants. Figure 3 (The right side of the middle).

[0044] The dimensions of the constructed spatial hybrid data of photovoltaic power plants are:

[0045] The formulas for calculating the solar altitude angle and solar azimuth angle are as follows:

[0046]

[0047] In the formula: α s γ s These are the solar altitude angle and azimuth angle, respectively; ω and δ are the hour angle and declination angle, respectively; t s This indicates local solar time (24-hour clock), where d represents the day number of the year. Indicates latitude.

[0048] S2: Convert two-dimensional photovoltaic power plant spatial hybrid data and cloud top height data into one-dimensional photovoltaic power plant spatial hybrid data and one-dimensional cloud top height data through linear one-dimensionalization, and convert two-dimensional satellite cloud images into satellite cloud images through convolution one-dimensionalization. Figure 1 Dimensional data;

[0049] The steps of linear one-dimensionalization include:

[0050] S2.1-a: Transfer two-dimensional data Reconstructed into a series of data blocks

[0051] S2.2-a: Data blocks are folded using dimensions Fold into

[0052] S2.3-a: Transform the dimension into a linear layer. The characteristics are mapped to

[0053] The steps of one-dimensional convolution include:

[0054] S2.1-b: Transfer two-dimensional data Reconstructed into a series of data blocks

[0055] S2.2-b: Using a convolutional neural network to process data blocks The dimension is converted to

[0056] This convolutional neural network consists of convolutional layers, residual blocks, and global pooling.

[0057] S2.2-b: Dimensional folding Fold into

[0058] S2.3-b: Transform the dimension into a linear layer. The characteristics are mapped to

[0059] S3: The location coding is performed on the one-dimensional photovoltaic power station spatial hybrid data, cloud top height data and satellite cloud image data to obtain the spatial characteristics of the photovoltaic power station, cloud spatial characteristics and cloud structure characteristics;

[0060] The formula for calculating the position code is:

[0061]

[0062] In the formula: pos is the position index, i is the dimension index, and d is the total dimension of the encoded vector.

[0063] S4: Input the three features into the two-stage cloud feature dynamic coding module to extract cloud features that interfere with photovoltaic power generation;

[0064] The two-stage cloud feature dynamic encoding module consists of cascaded cross-attention encoders. Each encoder contains N stacked encoder layers. The definition for moving from layer N-1 to layer N is as follows:

[0065]

[0066] In the formula: f MCA This represents Multi-headed Cross Attention (MCA), f MLP This refers to a multi-layer perceptron (MLP), f Norm Indicates normalization, F1 N-1 and F1 N These represent the features of the (N-1)th layer and the Nth layer, respectively.

[0067] The calculation method for cross attention is as follows:

[0068]

[0069] in:

[0070]

[0071] In the formula: W is the matrix coefficient, d k Let F be the number of columns in matrix K. A This is the output of the attention mechanism.

[0072] The steps for extracting cloud features are as follows:

[0073] S4.1: Input the spatial features of the photovoltaic power station and the spatial features of the cloud layer into the first encoder to extract the spatial coupling features between the spatial location of the photovoltaic power station and the location of the cloud layer;

[0074] S4.2: Input the spatial coupling features and cloud structure features from S4.1 into the second encoder to extract cloud features that affect photovoltaic power generation.

[0075] S5: Use a feedforward neural network to convert the cloud features in S5 into interference factors;

[0076] S6: Obtain the time series of interference factors from time tT to t+1 through parallel computing, and combine it with the historical power and timestamp series to form multivariate time series data;

[0077] The length of all three types of time series data is T+1, and the power value at time t+1 is set to 0 to maintain sequence alignment.

[0078] S7: Feed multivariate time series data into a hybrid neural network to predict photovoltaic power.

[0079] Hybrid neural networks are a combination of convolutional neural networks and gated recurrent units.

[0080] This embodiment provides an ultra-short-term forecast of the power generation of a photovoltaic power station in a central province for the next hour, using mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²). 2 Evaluate the model's performance:

[0081]

[0082] In the formula: P i and Let be the actual power and the predicted power of the i-th sample, respectively, and n be the total number of samples. This is the sample mean.

[0083] Two sets of comparative analyses were set up:

[0084] In the first group, the Active Cloud Region Selection (ACRS) rule is used as the benchmark method for selecting shading cloud areas, and the required cloud top height is the average cloud top height around the photovoltaic power station. Since this benchmark method does not require the construction of cross-attention, the VIT model is used to encode the local cloud map determined by this method and extract the interference factor.

[0085] In the second set of simulations, the backbone networks of different advanced models were used as benchmarks to replace S2 to S4 in this invention. The purpose was to demonstrate the effectiveness of the two-stage cloud feature dynamic coding module. Meanwhile, as shown in the following equation, X... RAA X CTH and X CI Integrate into X Cat As input, so that the baseline model can process it.

[0086]

[0087] In the formula: f cat This is for splicing operations.

[0088] Table 1 Overall Comparison Results

[0089]

[0090] This invention validated both methods using the VIT model. Except for predictions made 15 minutes in advance, the prediction accuracy of the ACRS-based method was slightly lower than that of the end-to-end method based on multimodal data coupling. This difference mainly stems from the fact that empirical calculations rely on the estimation of a large number of parameters. Introducing empirical values ​​for cloud top heights can easily lead to significant deviations in the ACRS solution results, and the calculated regional cloud map may not represent the occluded area, thus affecting prediction accuracy.

[0091] Although spatially mixed data and cloud top height data can effectively improve the accuracy of photovoltaic power prediction, the self-attention mechanism in VIT does not seem to be suitable for processing multimodal data. Its prediction accuracy is only better than CNN and Resnet-18 with relatively simple functions, but worse than SENet and SKNet with strong feature extraction capabilities. It cannot fully establish the coupling relationship between the spatially mixed data of photovoltaic power plants, cloud top height data and satellite cloud images, and extract cloud features that affect photovoltaic power prediction.

[0092] The two-stage cloud feature dynamic encoding module proposed in this invention effectively solves this problem. It establishes the dependencies between the three elements sequentially through a cross-attention mechanism, adaptively extracting cloud interference variables affecting photovoltaic power generation. Compared to the self-attention mechanism in VIT, the method proposed in this invention reduces MAE and RMSE by an average of 13.38% and 15.67% in multi-step prediction, respectively. 2 On average, it increased by 5.37%.

[0093] To ensure the reliability of the above results, taking 15-minute advance prediction as an example, this invention demonstrates the correlation coefficients between the interference factors generated by different methods and the actual power. Since the stronger the cloud cover effect, the lower the photovoltaic power generation, there is a negative correlation between the interference factor and the actual power; the smaller the value, the higher the correlation. Table 2 shows that the interference factor generated by the method of this invention has the highest correlation with the actual power. This indicates that compared to other advanced models, the method of this invention can more effectively process multimodal data, uncover the dependencies between data, and thus accurately extract the interference factors affecting photovoltaic power generation. Furthermore, this result corroborates the results reported in Table 1, proving the effectiveness of the method proposed in this invention.

[0094] Table 2 Correlation coefficients between interference factor and actual power

[0095]

[0096] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0097] This description uses specific examples to illustrate the principles and implementation methods of the present invention. The above description of the embodiments is only for the purpose of helping to understand the core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. In summary, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for predicting ultra-short-term photovoltaic power, characterized in that, include: Acquire target photovoltaic data; the target photovoltaic data includes remote sensing data, solar altitude angle, solar azimuth angle, cloud top height data, and satellite cloud images; Constructing spatial hybrid data for photovoltaic power plants using remote sensing data, solar altitude angle, and solar azimuth angle; Linear one-dimensionalization is used to convert two-dimensional photovoltaic power plant spatial mixed data and cloud top height data into one-dimensional photovoltaic power plant spatial mixed data and one-dimensional cloud top height data, and convolution one-dimensionalization is used to convert two-dimensional satellite cloud images into one-dimensional satellite cloud images. Location coding was performed on the one-dimensional mixed spatial data of photovoltaic power plants, cloud top height data, and satellite cloud image data to obtain the spatial characteristics of photovoltaic power plants, cloud spatial characteristics, and cloud structure characteristics. The spatial characteristics of photovoltaic power plants, cloud spatial characteristics, and cloud structure characteristics are input into a two-stage cloud feature dynamic coding module to extract cloud features that interfere with photovoltaic power generation. The cloud features that interfere with photovoltaic power generation are converted into interference factors using a feedforward neural network. The time series of interference factors from time tT to t+1 is obtained through parallel computing, and combined with the historical power and timestamp series to form multivariate time series data; where t is the current time and T is the length of the historical data; Multivariate time series data is fed into a hybrid neural network to predict photovoltaic power; The two-stage cloud feature dynamic encoding module consists of a series of cross-attention encoders, and the calculation formula for the cross-attention encoder is as follows: Among them, f MCA Indicates multi-head cross-attention, f MLP Denotes a multilayer perceptron, f Norm This represents normalization, where F1 and F2 are the two inputs to the cross-attention mechanism, respectively. N-1 and F1 N These represent the features of the (N-1)th layer and the Nth layer, respectively; The specific processing steps for extracting cloud features that interfere with photovoltaic power generation include: The spatial features of the photovoltaic power station and the cloud spatial features are input into the first encoder to extract the spatial coupling features between the spatial location of the photovoltaic power station and the cloud location. The spatial coupling features and cloud structure features are input into the second encoder to extract cloud features that interfere with photovoltaic power generation.

2. The photovoltaic ultra-short-term power prediction method according to claim 1, characterized in that, The method of constructing spatial hybrid data for photovoltaic power plants using remote sensing data, solar altitude angle, and solar azimuth angle includes: Based on the remote sensing data, a spatial distribution map of photovoltaic power stations is determined, and the spatial distribution map is mapped onto a spatial grid map of a meteorological satellite. Then, the solar altitude angle and solar azimuth angle are embedded into the spatial network map to obtain spatial hybrid data of photovoltaic power stations.

3. The photovoltaic ultra-short-term power prediction method according to claim 1, characterized in that, The linear one-dimensionalization steps specifically include: First, the two-dimensional data Reconstructed into a series of data blocks Then, the data blocks are folded using dimensions. Fold into Finally, through a linear layer... The characteristics are mapped to one-dimensional data Where H and W are the height and width of the two-dimensional data, respectively, and N = HW / B 2 The number of data blocks is represented by (B,B), the size of the data block is represented by B, the width of the data block is represented by C, the number of channels is represented by D, and the number of features is represented by D.

4. The photovoltaic ultra-short-term power prediction method according to claim 1, characterized in that, The specific steps of convolution one-dimensionalization include: First, the two-dimensional data Reconstructed into a series of data blocks Then use a convolutional neural network to process the data blocks The dimension is converted to Then, the dimensions are folded using dimension folding. Fold into Finally, through a linear layer... The characteristics are mapped to 5. An electronic device, characterized in that, The device includes a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to cause the electronic device to perform the photovoltaic ultra-short-term power prediction method according to any one of claims 1-4.

6. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the photovoltaic ultra-short-term power prediction method as described in any one of claims 1-4.

Images