A photovoltaic power prediction method based on decoupled architecture and adaptive multi-scale convolution

Abstract

This invention discloses a photovoltaic power prediction method based on a decoupled architecture and adaptive multi-scale convolution, comprising: detecting missing values, outliers, and missing time steps in the variables of the photovoltaic dataset, and replacing the detected missing values, outliers, and missing time steps with linear interpolation; dividing the dataset into training, validation, and test sets, and performing max-min normalization on the historical datasets of each variable in the dataset; performing feature selection on the input variables using the Pearson correlation coefficient detection method to select variables with strong correlation to historical photovoltaic power; constructing an adaptive multi-scale convolution model based on a decoupled architecture, and training the model using the training set; selecting the model with the highest accuracy on the validation set, and using this model to test the model's prediction performance on the test set to obtain the predicted photovoltaic power generation value. This invention enables the prediction of photovoltaic power changes over multiple future time steps, improving the stability of the power grid system.

Description

Technical Field

[0001] This invention relates to the field of operation and control of photovoltaic power plants, and specifically to a photovoltaic power prediction method based on decoupled architecture and adaptive multi-scale convolution. Background Technology

[0002] Solar energy, as a widely popular renewable energy source, is highly regarded not only for its abundant resources and ease of access but also for its low operating and maintenance costs. Solar energy not only helps reduce dependence on fossil fuels but also plays a vital role in promoting environmental protection and addressing climate change.

[0003] However, photovoltaic (PV) power generation is significantly affected by weather conditions (such as irradiance, rainfall, humidity, and temperature) and time (such as sunrise and sunset, and seasonal changes), resulting in considerable intermittency, volatility, and uncertainty. The purpose of PV power forecasting is to reduce this uncertainty and provide decision support for grid operators, electricity market participants, and PV power plant managers. By accurately forecasting PV power generation, grid operators can better dispatch power resources, improve grid stability, reduce reliance on backup power sources, lower power system operating costs, and enhance the market competitiveness of PV power. Therefore, accurate forecasting of PV power generation is often essential.

[0004] In recent years, an increasing number of deep learning models have been proposed in the field of photovoltaic power prediction, surpassing previous traditional methods, such as physical and statistical methods, in terms of prediction accuracy. Despite the significant success of deep learning methods in photovoltaic power prediction, existing methods still have some shortcomings. Recurrent neural networks such as LSTM (Long Short-Term Memory) and GRU (Gated Recurrent Unit) have limited memory capacity in practical applications when the dependencies span long time steps. Important historical information may be overwritten, making it difficult for these models to capture long-term temporal relationships. Furthermore, due to their recurrent nature, computation must be performed sequentially according to the time series, making parallel training difficult and resulting in long training times. The Transformer's self-attention mechanism possesses permutation invariance. Although various types of positional encoding techniques can preserve some sequence information, the application of self-attention inevitably results in a loss of temporal information. Time series such as photovoltaic power have inherent temporal characteristics, which is detrimental to time series modeling in photovoltaic power forecasting. Furthermore, the Transformer's self-attention mechanism is point-wise, meaning the semantic information contained in the tokens formed at a single time step is limited. The point-wise nature also makes the self-attention mechanism susceptible to outliers, leading to incorrect weight allocation for each time point. All of these factors contribute to the Transformer's suboptimal accuracy in photovoltaic power forecasting. In recent years, models such as DLinear (which performs trend seasonality decomposition on time series) and PatchTST (patch timeseries Transformer) have adopted channel-independence strategies, failing to model the relationships between variables and neglecting the extraction of dimensional information from variables, thus only capturing temporal dependencies. However, in the field of photovoltaic power prediction, in addition to photovoltaic power as the target variable, other environmental variables, such as radiance (total solar radiance, direct normal radiance, global horizontal radiance), temperature, humidity and other variables are also very important. Relying solely on historical photovoltaic power, the model's ability to capture time-series changes is still limited, and therefore the accuracy of DLinear and PatchTST will also be affected.ModernTCN (Modern Temporal Convolutional Network) features large kernel convolutional blocks, which help convolutions achieve a larger receptive field. It also explicitly designs variable interaction modules to effectively extract information from variable dimensions, achieving good results in the field of time series forecasting. However, ModernTCN's large kernel convolution has high computational cost and complexity, and it does not take into account the multi-scale changes in photovoltaic power and environmental variables over time series. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing models in the field of photovoltaic power prediction by providing a photovoltaic power prediction method based on decoupled architecture and adaptive multi-scale convolution.

[0006] To achieve the above objectives, this invention provides a photovoltaic power prediction method based on a decoupled architecture and adaptive multi-scale convolution, comprising the following steps:

[0007] (1) Obtain the photovoltaic dataset of the photovoltaic power station. The variables of the photovoltaic dataset include historical power generation, radiance, temperature, humidity, air pressure and rainfall. Perform missing value detection, outlier detection and missing time step detection on the variables of the photovoltaic dataset. Replace the detected missing values, outliers and missing time steps with linear interpolation.

[0008] (2) Based on the length of the input history and the range of the prediction, the photovoltaic dataset is divided into a training set, a validation set and a test set in chronological order, with a ratio of 6:2:2; the maximum and minimum values ​​of the training set are used to perform max-min normalization on the historical dataset of each variable in the photovoltaic dataset.

[0009] (3) Use the Pearson correlation coefficient test method to select features of the input variables, select variables that are strongly correlated with historical photovoltaic power, including historical power generation, radiance and temperature; exclude variables that are not strongly correlated with historical photovoltaic power, including air pressure and humidity.

[0010] (4) Construct an adaptive multi-scale convolution model based on a decoupled architecture, train the adaptive multi-scale convolution model using the training set; select the model with the highest accuracy on the validation set based on the accuracy of the adaptive multi-scale convolution model, use the model to test the prediction performance of the model on the test set, and the output of the model is the final photovoltaic power generation prediction value.

[0011] Furthermore, in step (1), the radiance includes total solar radiance, direct normal radiance, and global horizontal radiance.

[0012] Furthermore, in step (4), the adaptive multi-scale convolutional model includes an embedding layer at the input end, stacked sub-blocks, and a projection layer at the output end;

[0013] The embedding layer is used for preliminary processing of the time dimension; that is, it aggregates local contextual semantic information at each time point to achieve dimensionality enhancement of latent features.

[0014] The sub-block includes an adaptive multi-scale convolution module and a group-based convolution gated linear unit. The adaptive multi-scale convolution module includes parallel dilated kernel depthwise convolutions and adaptive aggregation blocks. The adaptive multi-scale convolution model learns features at different scales simultaneously through multiple parallel convolution kernels to enhance the model's representational ability. The adaptive aggregation blocks are used to weight the outputs of these parallel multi-dilated kernel depthwise convolutions, i.e., to weight the features at different scales, so that the model can capture the time series variation characteristics at different scales. According to the number of groups of group convolutions, the group-based convolution gated linear unit is divided into channel convolution gated linear units and variable convolution gated linear units, which are used to learn the cross-channel representation of each variable and the cross-variable representation of each channel, respectively.

[0015] The projection layer is used to set the output step size of the projection layer according to whether the prediction target is single-step prediction or multi-step prediction, thereby realizing single-step prediction and multi-step prediction of the model.

[0016] Furthermore, the forward process of the embedding layer is represented as follows:

[0017] X emb =Embedding(Padding(X) in ))

[0018] Where, X emb X is the output of the embedding layer and also the input to the sub-blocks of the adaptive multi-scale convolutional model. in The input is the original input, Embedding is the embedding operation, and Padding is the padding operation.

[0019] Furthermore, the forward process of the adaptive multi-scale convolution module is represented as follows:

[0020]

[0021] In the formula, Z time X represents the output of the adaptive multi-scale convolution module. emb As the input to the adaptive multi-scale convolution module, DWConv i (X) is the output of each depthwise convolutional branch, BN is the batch regularization layer, N is the number of branches, and ω i These are the corresponding learnable weights.

[0022] Furthermore, based on the number of groups of group convolutions, the group-based convolution gated linear unit includes a channel convolution gated linear unit and a variable linear unit; the channel convolution gated linear unit is used to learn the cross-channel representation of each variable; the variable linear unit is used to learn the cross-variable representation of each channel; the forward propagation formulas for the channel convolution gated linear unit and the variable linear unit are as follows:

[0023] Z channel,1 =ChannelConvGLU(Z time ) groups＝M

[0024] Z channel,2 =Reshape&Permute(Z) channel,1 )

[0025] Z variate,1 =Variate ConvGLU(Z) channel,2 ) groups＝D

[0026] Z variate,2 =Reshape&Permute(Z) variate,1 )

[0027] Among them, Z time It is the input of the channel convolution gated linear unit; Z channel,1 It is the output of the channel convolution gated linear unit, Z channel,2 It is for Z channel,1 reshape and permute, Z channel,2 It is also the input of the variable convolution gated linear unit, Z variate,1 It is the output of the variable convolution gated linear unit, Z variate,2 It is for Z variate,1 The reshape and permute operations ensure that the shape is the same when performing residual joins.

[0028] Furthermore, the forward process of the projection layer is as follows:

[0029]

[0030] Where, X out This represents the output of the projection layer, which is the final predicted photovoltaic power value. The features are learned by stacked L layers of sub-blocks, and then the projection layer maps the learned features to the output X. out This is the final predicted photovoltaic power value.

[0031] To achieve the above objectives, the present invention also provides a photovoltaic power prediction device based on decoupled architecture and adaptive multi-scale convolution, comprising one or more processors for implementing the photovoltaic power prediction method based on decoupled architecture and adaptive multi-scale convolution described above.

[0032] To achieve the above objectives, the present invention also provides an electronic device, including a memory and a processor, wherein the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the above-described photovoltaic power prediction method based on decoupled architecture and adaptive multi-scale convolution.

[0033] To achieve the above objectives, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described photovoltaic power prediction method based on decoupled architecture and adaptive multi-scale convolution.

[0034] Compared with the prior art, the beneficial effects of the present invention are: the present invention can effectively extract information on the time dimension, variable dimension and channel dimension of photovoltaic power and environmental variable sequences, thereby improving the accuracy of photovoltaic power prediction; and the present invention can predict the changes in photovoltaic power in multiple future time steps, which can help grid operators better plan and dispatch power resources in advance, maintain supply and demand balance, improve the stability of the grid system and reduce the operating cost of the grid system. Attached Figure Description

[0035] To more clearly illustrate the technical solution of the present invention, the following will provide a brief introduction to the model and the accompanying drawings used in the specific photovoltaic power prediction process.

[0036] Figure 1 This is a flowchart illustrating the overall process of photovoltaic power prediction in this invention.

[0037] Figure 2 This is a schematic diagram of the overall structure of the prediction model designed in this invention;

[0038] Figure 3 This is a schematic diagram of the adaptive multi-scale convolutional block designed in this invention;

[0039] Figure 4 This is a schematic diagram of the convolutional gated linear unit based on group convolution designed in this invention;

[0040] Figure 5 This is a schematic diagram of the structure of the device of the present invention;

[0041] Figure 6 This is a schematic diagram of an electronic device according to the present invention. Detailed Implementation

[0042] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.

[0043] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The singular forms “a,” “the,” and “the” used in this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0044] It should be understood that although the terms first, second, third, etc., may be used in this invention to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first information may also be referred to as second information without departing from the scope of this invention, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0045] The present invention will now be described in detail with reference to the accompanying drawings. Unless otherwise specified, the features of the following embodiments and implementations can be combined with each other.

[0046] This invention provides an adaptive multi-scale convolutional model based on a decoupled architecture, capable of single-step and multi-step prediction of photovoltaic power generation. In the model, an adaptive multi-scale convolutional block is designed to adaptively assign corresponding weights to each scale of the time series, helping the model learn time dimension information at different scales simultaneously. Considering the excellent performance of gated linear units (GLUs) in fields such as natural language processing, this paper proposes a group convolution-based convolutional gated linear unit (ConvGLU), which helps the model learn the channel dimension and variable dimension information of photovoltaic power and environmental variable sequences respectively. This invention utilizes the idea of ​​a decoupled architecture to avoid the mutual coupling influence between the channel space and the variable space. By decoupling these two spaces, the difficulty of learning the complex changes in photovoltaic power and meteorological variable sequences is reduced, achieving more accurate photovoltaic power prediction.

[0047] See Figure 1The present invention provides a photovoltaic power prediction method based on decoupling architecture and adaptive multi-scale convolution, which specifically includes the following steps:

[0048] (1) Obtain the photovoltaic dataset of the photovoltaic power station. The variables of the photovoltaic dataset include historical power generation, radiance (total solar radiance, direct normal radiance, global horizontal radiance), temperature, humidity, air pressure, rainfall, etc. Perform missing value detection, outlier detection, and missing time step detection on these variables. Replace the detected missing values, outliers, and missing time steps with linear interpolation.

[0049] (2) Based on the length of the historical data input and the range of the prediction, the photovoltaic dataset is divided into training, validation, and test sets in chronological order, with a ratio of 6:2:2. To prevent certain variables in the dataset from dominating in terms of scale and to prevent information leakage, the maximum and minimum values ​​of the training set are used to perform min-max normalization on the historical dataset of each variable in the photovoltaic dataset. The specific normalization formula is as follows:

[0050]

[0051] in, y represents the normalized value of the variable. i Represents the original value of the variable, y min y represents the minimum value of the original data of the variable. max This represents the maximum value of the original dataset containing the variable.

[0052] (3) Use the Pearson correlation coefficient detection method to select features of input variables, select variables that are strongly correlated with historical photovoltaic power, such as historical power generation, radiance (total solar radiance, direct normal radiance, global horizontal radiance), temperature, etc., and exclude variables that are not strongly correlated with historical photovoltaic power, such as air pressure, humidity, etc., so as to reduce the impact of interference and noise on model prediction and balance the performance and efficiency of the model.

[0053] (4) Construct an adaptive multi-scale convolutional model based on a decoupled architecture. Here, L1 error is used as the loss function. The proposed model is trained using the training set. The model with the highest accuracy on the validation set is selected, and then this model is used to test its prediction performance on the test set. The trained model is the prediction model, and its output is the final predicted photovoltaic power generation value. The formula for the L1 error during training is:

[0054]

[0055] in, It is the true value of the i-th normalized photovoltaic power. Let be the i-th normalized photovoltaic power prediction value, and n be the sample size. The L1 error function, or mean absolute error (MAE), describes the degree of fit between the predicted and actual values. The smaller the L1 error, the closer the predicted value is to the actual value.

[0056] Further, see Figure 1 The predicted results and actual values ​​obtained above are used to calculate the indicators MSE, MAE, and R2. MSE stands for Mean Squared Error; MAE for Mean Absolute Error; and R2 for Coefficient of Determination.

[0057] See Figure 2 Slice is a tensor slicing operation. The features learned from the stacked sub-blocks contain the features of all variables. This invention needs to slice to extract the features of the variable corresponding to photovoltaic power. Flatten is a tensor flattening operation, which facilitates the mapping of subsequent projection layers. See also Figure 3 k is the kernel size, k0 is the specific value of the kernel size, d is the dilation rate, and N is the number of parallel convolution branches.

[0058] The adaptive multi-scale convolutional model based on a decoupled architecture proposed in this invention includes an embedding layer at the input, stacked sub-blocks, and a projection layer at the output. A detailed description of each part is as follows:

[0059] For the embedding layer at the input end, the embedding layer performs preliminary processing of the time dimension, aggregating local contextual semantic information at each time point to achieve latent feature dimensionality enhancement. This invention implements the embedding operation through convolution. Before embedding, the original input X is first processed... in Padding is performed to make T = F / / S, where T represents the length of the embedded time dimension, F represents the input length, and S represents the stride of the convolutional layer; specifically, X is first padded to make T = F / / S. in Perform an unsqueeze operation (adding a dimension) to obtain X. in ∈R B×M×1×F Where B represents batch size, M represents the number of variables, and F represents the input length; then in X in X at the tail in The value at the last moment is obtained by repeating it PS times. Finally, fill in the blank. The input is fed into a 1D convolutional layer with kernel size P and stride S. This convolutional layer maps one input channel to D output channels, thus achieving preliminary processing of the time dimension and dimensionality enhancement of latent features. In the above convolution process, the time series of M variables are independently embedded, and the M variables do not affect each other; therefore, we preserve the information of the variable dimension. The forward process of the embedding layer can be represented as:

[0060] X emb =Embedding(Padding(X) in ))

[0061] Where, X emb The output of the embedding layer is also the input to the sub-blocks of the adaptive multi-scale convolutional model; X in The input is the original data; Embedding is the embedding operation, and Padding is the padding operation.

[0062] For the stacked model sub-block, each sub-block comprises three main sub-components. The first sub-component is the Adaptive Multi-Scale Convolution (AMSConv) module designed in this invention. This module includes a series of parallel depthwise convolutions with dilated kernels and an adaptive aggregation block. Each convolutional layer is followed by a batch normalization (BN) layer to accelerate the model's training convergence. Convolutions with different dilation rates help the model extract features at different time scales. Smaller dilation rates can capture short-term patterns, while larger dilation rates can capture long-term patterns. Through multiple parallel convolutional kernels, the model can simultaneously learn features at different scales, enhancing its representational ability and improving its predictive performance. Next, this paper uses the adaptive aggregation block to perform weighted summation of the outputs of these parallel depthwise convolutions with multiple dilated kernels, i.e., weighting the features at different scales instead of simply adding them directly, thereby helping the model better capture the time-series variation characteristics at different scales. The forward process of the adaptive multi-scale convolution module can be represented as:

[0063]

[0064] In the formula, Z time X represents the output of the adaptive multi-scale convolution module. emb As the input to the adaptive multi-scale convolution module, DWConv i (X) is the output of each depthwise convolutional branch, BN is the batch normalization layer, N is the number of branches, and ω i The corresponding learnable weights are used to achieve adaptive weighting. Adaptive multi-scale blocks can effectively learn the temporal dimension information of photovoltaic power and environmental variables.

[0065] See Figure 4 , C in This represents the number of input channels for the convolution. Based on the number of groups in the group convolution, the proposed group convolution-based Convolutional Gated Linear Unit (ConvGLU) serves as the second and third sub-components: the Channel Convolutional Gated Linear Unit (Channel ConvGLU) and the Variational Convolutional Gated Linear Unit (Variate ConvGLU), used to learn the cross-channel representation of each variable and the cross-variable representation of each channel, respectively. The gated linear unit is a channel mixer that has shown better performance than multilayer perceptrons in various natural language processing tasks. Unlike the channel attention (SE) mechanism, the gating signal for each token comes from the token itself, and its receptive field size does not exceed the receptive field of the value branch. The proposed convolutional gated unit consists of three pointwise convolutions (PWConv), with a kernel size of 1. The main function of the pointwise convolution is to change the number of channels in the input data. ConvGLU is an inverted bottleneck structure. The output channels of the two upper PWConv blocks are r times wider than the input channels, and the input channels of the lower PWConv block are r times wider than the output channels. The upper left PWConv block and the sigmoid activation function are combined to generate gate weights with values ​​from 0 to 1. The gate weights are multiplied by the output of the upper right PWConv block, and the result is then passed through the lower PWConv block. This invention uses Channel ConvGLU and Variation ConvGLU to learn channel and variable dimensions, respectively. In Channel ConvGLU, the number of groups of pointwise convolutions is set to the number of variables M. Only D features belonging to the same group can interact with each other. Therefore, Channel ConvGLU is used to learn the cross-channel representation of each variable. In Variation ConvGLU, the number of groups of pointwise convolutions is set to the number of channels D. Only M variables belonging to the same group can interact with each other. Therefore, Variation ConvGLU is used to learn the cross-variable representation of each channel. The forward propagation formulas for Channel ConvGLU and Variation ConvGLU are as follows:

[0066] Z channel,1 =ChannelConvGLU(Z time ) groups＝M

[0067] Z channel,2 =Reshape&Permute(Z) channel,1)

[0068] Z variate,1 =Variate ConvGLU(Z) channel,2 ) groups＝D

[0069] Z variate,2 =Reshape&Permute(Z) variate,1 )

[0070] Among them, Z time It is the input to Channel ConvGLU, i.e., the forward process of the adaptive multi-scale convolutional module; Z channel,1 It is the output of Channel ConvGLU, Z channel,2 It is for Z channel,1 The operations Z involve reshaping and permute. Reshaping deforms the shape of the tensor, while permute permutes two dimensions of the tensor. These operations are necessary for the correct convolutional integral grouping. channel,2 Z is also the input of the Varitae ConvGLU. variate,1 This is the output of VaritaConvGLU, Z. variate,2 It is for Z variate,1 The reshape and permute operations ensure that the shape of the data tensors during residual joins is the same.

[0071] For the projection layer at the output, this invention uses a single linear layer to directly project the features learned from the stacked sub-blocks onto the output. The output X of the projection layer... out This is the final output of the entire model. Here, the output step size of the projection layer can be set according to whether the prediction target is single-step or multi-step prediction, thus achieving single-step or multi-step prediction of the model. The forward process of the projection layer is as follows:

[0072]

[0073] Where, X out This represents the output of the projection layer, which is the final predicted photovoltaic power value. It consists of features learned by stacked L-layer model sub-blocks, and then the projection layer maps the learned features to the output X. out This is the final predicted photovoltaic power value.

[0074] It is worth mentioning that in this invention, the adaptive multi-scale convolutional AMSConv block is responsible for learning the temporal relationships between time points on each feature basis, which plays the same role as the self-attention module in Vision Transformer (ViTs). Channel ConvGLU and Variation ConvGLU are similar to FeedForward Networks (FFNs) in Vision Transformer. Channel ConvGLU is responsible for learning the cross-channel representation of each variable, and Variation ConvGLU is responsible for capturing the cross-variable dependencies of each channel. Channel ConvGLU and Variation ConvGLU can avoid the mutual coupling between the channel space and the variable space. By decoupling these two spaces, the difficulty of learning the complex changes in photovoltaic power sequences and meteorological variable sequences can be reduced. The above three modules operate from the three dimensions of time, channel (or latent feature dimension), and variable, respectively, thereby realizing the idea of ​​decoupling.

[0075] Corresponding to the aforementioned embodiments of the photovoltaic power prediction method based on decoupled architecture and adaptive multi-scale convolution, the present invention also provides embodiments of a photovoltaic power prediction device based on decoupled architecture and adaptive multi-scale convolution.

[0076] See Figure 5 The photovoltaic power prediction device based on decoupled architecture and adaptive multi-scale convolution provided in this embodiment of the invention includes one or more processors for implementing the photovoltaic power prediction method based on decoupled architecture and adaptive multi-scale convolution in the above embodiment.

[0077] The photovoltaic power prediction device based on decoupled architecture and adaptive multi-scale convolution of this invention can be applied to any device with data processing capabilities, such as a computer. The device embodiment can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 5 The diagram shown is a hardware structure diagram of any device with data processing capabilities, including the photovoltaic power prediction device based on decoupling architecture and adaptive multi-scale convolution of this invention. (Except for...) Figure 5In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0078] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0079] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0080] Corresponding to the aforementioned embodiments of the photovoltaic power prediction method based on decoupled architecture and adaptive multi-scale convolution, this application also provides an electronic device, including: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the photovoltaic power prediction method based on decoupled architecture and adaptive multi-scale convolution as described above. Figure 6 The diagram shown is a hardware structure diagram of any device with data processing capabilities for the photovoltaic power prediction method based on decoupled architecture and adaptive multi-scale convolution provided in the embodiments of this application, except for... Figure 6 In addition to the processor, memory, DMA controller, disk, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0081] Corresponding to the aforementioned embodiments of the photovoltaic power prediction method based on decoupled architecture and adaptive multi-scale convolution, this embodiment of the invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements the photovoltaic power prediction method based on decoupled architecture and adaptive multi-scale convolution described in the above embodiments.

[0082] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0083] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0084] The above embodiments are only used to illustrate the design concept and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, all equivalent changes or modifications made based on the principles and design ideas disclosed in the present invention are within the protection scope of the present invention.

Claims

1. A photovoltaic power prediction method based on decoupled architecture and adaptive multi-scale convolution, characterized in that, Includes the following steps: (1) Obtain the photovoltaic dataset of the photovoltaic power station. The variables of the photovoltaic dataset include historical power generation, radiance, temperature, humidity, air pressure and rainfall. Perform missing value detection, outlier detection and missing time step detection on the variables of the photovoltaic dataset. Replace the detected missing values, outliers and missing time steps with linear interpolation. (2) Based on the length of the input history and the range of the prediction, the photovoltaic dataset is divided into a training set, a validation set and a test set in chronological order, with a ratio of 6:2:2; the maximum and minimum values ​​of the training set are used to perform max-min normalization on the historical dataset of each variable in the photovoltaic dataset. (3) Use the Pearson correlation coefficient test method to select features of the input variables, select variables that are strongly correlated with historical photovoltaic power, including historical power generation, radiance and temperature; exclude variables that are not strongly correlated with historical photovoltaic power, including air pressure and humidity. (4) Construct an adaptive multi-scale convolutional model based on a decoupled architecture, train the adaptive multi-scale convolutional model using the training set; select the model with the highest accuracy on the validation set based on the accuracy of the adaptive multi-scale convolutional model, use the model to test the prediction performance of the model on the test set, and the output of the model is the final photovoltaic power generation prediction value. The adaptive multi-scale convolutional model includes an embedding layer at the input, stacked sub-blocks, and a projection layer at the output. The embedding layer is used for preliminary processing of the time dimension; that is, it aggregates local contextual semantic information at each time point to achieve dimensionality enhancement of latent features. The sub-block includes an adaptive multi-scale convolution module and a group-based convolution gated linear unit. The adaptive multi-scale convolution module includes parallel dilated kernel depthwise convolutions and adaptive aggregation blocks. The adaptive multi-scale convolution model learns features at different scales simultaneously through multiple parallel convolution kernels to enhance the model's representational ability. The adaptive aggregation blocks are used to weight the outputs of these parallel multi-dilated kernel depthwise convolutions, i.e., to weight the features at different scales, so that the model can capture the time series variation characteristics at different scales. According to the number of groups of group convolutions, the group-based convolution gated linear unit is divided into channel convolution gated linear units and variable convolution gated linear units, which are used to learn the cross-channel representation of each variable and the cross-variable representation of each channel, respectively. The projection layer is used to set the output step size of the projection layer according to whether the prediction target is single-step prediction or multi-step prediction, thereby realizing single-step prediction and multi-step prediction of the model. The forward process of the adaptive multi-scale convolution module is represented as follows: In the formula, Z time X represents the output of the adaptive multi-scale convolution module. emb As the input to the adaptive multi-scale convolution module, DWConv i (X) is the output of each depthwise convolutional branch, BN is the batch regularization layer, N is the number of branches, and ω i These are the corresponding learnable weights; Depending on the number of groups in the group convolution, the group convolution-based convolution gated linear unit includes a channel convolution gated linear unit and a variable linear unit; the channel convolution gated linear unit is used to learn the cross-channel representation of each variable; The variable linear unit is used to learn the cross-variable representation of each channel; the forward propagation formulas for the channel convolution gated linear unit and the variable linear unit are as follows: WITH channel,1 =ChannelConvGLU(Z time ) groups＝M WITH channel,2 =Reshape&Permute(Z channel,1 ) Z variate,1 =Variate ConvGLU(Z channel,2 ) groups＝D WITH variate,2 =Reshape&Permute(Z variate,1 ) Among them, Z time It is the input of the channel convolution gated linear unit; Z channel,1 It is the output of the channel convolution gated linear unit, Z chamnel,2 It is for Z channel,1 reshape and permute, Z channel,2 It is also the input of the variable convolution gated linear unit, Z variate,1 It is the output of the variable convolution gated linear unit, Z variate,2 It is for Z variate,1 The reshape and permute operations ensure that the shape is the same when performing residual joins.

2. The photovoltaic power prediction method based on decoupled architecture and adaptive multi-scale convolution according to claim 1, characterized in that, In step (1), the radiance includes total solar radiance, direct normal radiance, and global horizontal radiance.

3. The photovoltaic power prediction method based on decoupled architecture and adaptive multi-scale convolution according to claim 1, characterized in that, The forward process of the embedding layer is represented as follows: X emb =Embedding(Padding(X in )) In the formula, X emb X is the output of the embedding layer and also the input to the sub-blocks of the adaptive multi-scale convolutional model. in The input is the original input, Embedding is the embedding operation, and Padding is the padding operation.

4. The photovoltaic power prediction method based on decoupled architecture and adaptive multi-scale convolution according to claim 1, characterized in that, The forward process of the projection layer is as follows: In the formula, X out This represents the output of the projection layer, which is the final predicted photovoltaic power value. The features are learned by stacked L layers of sub-blocks, and then the projection layer maps the learned features to the output X. out This is the final predicted photovoltaic power value.

5. A photovoltaic power prediction device based on a decoupled architecture and adaptive multi-scale convolution, characterized in that, It includes one or more processors for implementing the photovoltaic power prediction method based on decoupled architecture and adaptive multi-scale convolution as described in any one of claims 1-4.

6. An electronic device comprising a memory and a processor, characterized in that, The memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the photovoltaic power prediction method based on decoupled architecture and adaptive multi-scale convolution 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 program is executed by the processor, it implements the photovoltaic power prediction method based on decoupled architecture and adaptive multi-scale convolution as described in any one of claims 1-4.

Images