Short-term prediction method of photovoltaic output based on deep learning

Abstract

The present invention relates to a short-term photovoltaic output prediction method based on deep learning, including: constructing a clustered probability prediction model, the clustered probability prediction model is used to establish a mapping between photovoltaic output and meteorological data and historical output using a conditional generative adversarial network, sampling and generating different prediction results from the potential space under specific conditional constraint labels, and then constructing a prediction result set containing multiple possible scenarios to achieve a probabilistic prediction of photovoltaic output; quantitatively characterizing the positional relationship between photovoltaic units in the cluster through an output characteristic map encoding method; and matrixing the meteorological labels of the time series using a weather characteristic map construction method to maintain the consistency between the conditional labels constructed by the meteorological data and the output data dimension. The present invention can make full use of known information to improve the prediction accuracy of the model.

Description

Technical Field

[0001] The present invention relates to the field of new energy output prediction, and in particular to a photovoltaic output short-term prediction method based on deep learning. Background Art

[0002] Accurate photovoltaic output forecasting can provide reliable reference information for grid dispatching, enabling the system to take regulatory measures such as charging and discharging dispatching of energy storage equipment or starting frequency regulation backup resources in advance, thereby optimizing grid load distribution, balancing power supply and demand, avoiding voltage or frequency fluctuations caused by photovoltaic output fluctuations, and reducing the occurrence of large-scale abandonment of photovoltaic power. Therefore, improving the accuracy and robustness of photovoltaic power generation short-term output forecasting is the key to supporting power system dispatching and operation.

[0003] From the perspective of the presentation of the prediction results, photovoltaic power prediction can be divided into deterministic prediction, interval prediction and probabilistic prediction. Among them, probabilistic prediction can provide the probability distribution of photovoltaic power station output and effectively quantify the uncertainty of photovoltaic power generation, which has attracted more and more attention. The existing probabilistic prediction methods are mainly based on the statistical analysis of the probability distribution of historical prediction errors, and the use of methods such as kernel density estimation or quantile regression to probabilistically expand the deterministic prediction results.

[0004] Ma Lubin et al. divided the photovoltaic power stations in the region into clusters based on the density clustering model, thereby enhancing the consistency of the cluster output characteristics after the photovoltaic power stations were clustered, and used recurrent neural networks to achieve probabilistic prediction; Wang Yubo et al. proposed a method for identifying the fluctuation pattern of power generation in distributed photovoltaic clusters and ultra-short-term probability prediction, integrating satellite cloud images and photovoltaic power data, and constructing a fluctuation pattern recognition model through feature extraction of output fluctuations to achieve mining of fluctuation rules. In addition, Liu Yunchao et al. and Ouyang Yongjian et al. used graph neural networks to construct a static graph network to extract dynamic spatial information between adjacent power stations and achieve prediction of the total output of distributed photovoltaic clusters.

[0005] The above studies have made outstanding contributions to the exploration of the spatiotemporal correlation of distributed photovoltaic clusters, but the analysis and modeling of the spatial location relationship of each site in the cluster is relatively rough, and mainly focuses on the prediction of the total power of the cluster. In addition, due to the high-dimensional, strong nonlinear, and complex correlation characteristics of distributed photovoltaic clusters, the modeling difficulty and computational complexity of analytical methods based on probability statistics have been greatly increased, so the prediction method system and model prediction accuracy need to be further improved. Summary of the invention

[0006] In order to solve the technical problems existing in the above-mentioned prior art, the present invention provides a short-term photovoltaic output prediction method based on deep learning, which can make full use of known information to improve the prediction accuracy of the model.

[0007] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:

[0008] A short-term photovoltaic output prediction method based on deep learning, comprising:

[0009] Constructing a clustered probabilistic prediction model, which is used to establish a mapping between photovoltaic output and meteorological data and historical output using a conditional generative adversarial network (cGAN), sample different prediction results from the latent space under specific conditional constraint labels, and then construct a prediction result set containing multiple possible scenarios to achieve probabilistic prediction of photovoltaic output;

[0010] The positional relationship between photovoltaic units in the cluster is quantitatively characterized by the output characteristic diagram encoding method;

[0011] The weather characteristic map construction method is used to matrix the time series meteorological labels to maintain the consistency between the conditional labels constructed by meteorological data and the output data dimensions.

[0012] In the above scheme, the process of generating prediction data by the cluster probability prediction model includes:

[0013] Obtain real-time output data and meteorological data from photovoltaic power stations, construct historical photovoltaic output characteristic maps and historical weather characteristic maps, and construct inputs with random noise;

[0014] generating short-term output data corresponding to the prediction based on the input;

[0015] Construct multiple sets of prediction results, integrate these prediction results, and form a probability distribution band interval that reflects the possibility of photovoltaic output under conditional constraints, thereby realizing the probabilistic prediction of photovoltaic output.

[0016] In the above scheme, the cluster probability prediction model determines the accuracy of the prediction through the following formula:

[0017]

[0018] Where, L G is the loss function of the generator (G), L D is the loss function of the discriminator (D).

[0019] In the above scheme, the method of quantitatively characterizing the positional relationship between photovoltaic units in the cluster by the output characteristic diagram encoding method adopts a distributed photovoltaic cluster spatial encoding method based on a greedy algorithm, which includes the following steps:

[0020] The space where the distributed photovoltaic cluster is located is gridded, and the maximum grid size that meets the coding principle is found through greedy search;

[0021] On this basis, by cutting off the empty rows / columns around the matrix, the spatial coding matrix representing the location information of the distributed photovoltaic cluster can be obtained.

[0022] In the above scheme, in the method of quantitatively characterizing the positional relationship between photovoltaic units in a cluster by using an output characteristic diagram encoding method, the photovoltaic output characteristic diagram construction method includes the following steps:

[0023] The distributed photovoltaic cluster output data is normalized and preprocessed to the maximum and minimum values, and compressed to the range of [0,1];

[0024] For the cluster output data at a certain moment, fill it into the spatial coding matrix to obtain the photovoltaic output characteristic diagram at that moment.

[0025] In the above scheme, the weather characteristic map encoding method used in the matrixing of the time series meteorological labels by using the weather characteristic map construction method includes the following steps:

[0026] The Pearson Correlation Coefficient is used to quantify the average correlation between each weather variable in the meteorological data and the output of the photovoltaic power station;

[0027] The two selected features are filled into the spatial coding matrix in the form of cross-filling to maintain the consistency of data dimensions.

[0028] The technical solution provided by the present invention has the following beneficial effects:

[0029] 1. The present invention provides a short-term photovoltaic output prediction method based on deep learning. The present invention adopts cGAN to carry out prediction. Its deep neural network architecture can fully explore the high-dimensional nonlinear laws such as the time series change law of photovoltaic output in historical data, the correlation between photovoltaic output and meteorological factors, and the complex spatiotemporal correlation between sites in the photovoltaic cluster, which helps to make full use of known information to improve prediction accuracy.

[0030] 2. The input of the cGAN model contains random vectors, and multiple sets of prediction results can be generated by changing the values ​​of the random vectors, which is naturally suitable for probabilistic prediction work.

[0031] 3. The photovoltaic cluster spatial coding method can effectively quantify and characterize the spatial relationship between photovoltaic units, and retain the spatial mutual information between photovoltaic units to the greatest extent in the form of a data matrix.

[0032] These and other aspects of the present invention will be more clearly understood in the following description of the embodiments. It should be understood that the above general description and the following detailed description are only exemplary and explanatory and cannot limit the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other embodiments can be obtained based on these drawings without paying creative work.

[0034] Figure 1 This is a framework diagram of a short-term photovoltaic output prediction method based on deep learning;

[0035] Figure 2 A framework diagram of a prediction model cGAN in a photovoltaic output short-term prediction method based on deep learning according to an embodiment of the present invention;

[0036] Figure 3 A schematic diagram of a photovoltaic matrix based on relative spatial position coding according to an embodiment of the present invention;

[0037] Figure 4 An output characteristic diagram constructed after filling corresponding power data into a photovoltaic matrix according to an embodiment of the present invention;

[0038] Figure 5 This is a framework diagram of the output characteristic map encoding method and weather characteristic map construction method in the photovoltaic output short-term prediction method based on deep learning;

[0039] Figure 6 This is a schematic diagram of ultra-short-term probability prediction results of a cluster probability prediction model under different weather conditions according to an embodiment of the present invention;

[0040] Figure 7 This is the overall flow chart of the short-term photovoltaic output prediction method based on deep learning. DETAILED DESCRIPTION

[0041] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are 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 creative work are within the scope of protection of the present invention.

[0042] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may also be decomposed, combined or partially merged, so the actual execution order may change according to actual conditions.

[0043] It should be understood that the terms used in this specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0044] The photovoltaic output short-term prediction method based on deep learning of the present invention includes:

[0045] S1. Constructing a clustered probability prediction model, which is used to establish a mapping between photovoltaic output and meteorological data and historical output using a conditional generative adversarial network (cGAN), sample different prediction results from the latent space under specific conditional constraint labels, and then construct a prediction result set containing multiple possible scenarios to achieve probabilistic prediction of photovoltaic output;

[0046] S2, quantitatively characterize the positional relationship between photovoltaic units in the cluster through the output characteristic diagram encoding method;

[0047] S3. Use the weather feature map construction method to matrix the time series meteorological labels to maintain the consistency between the conditional labels constructed by meteorological data and the output data dimensions.

[0048] Furthermore, the process of generating prediction data by the clustered probability prediction model includes:

[0049] Obtain real-time output data and meteorological data from the photovoltaic power station, construct historical photovoltaic output characteristic graph X and historical weather characteristic graph W, and construct input with random noise z;

[0050] generating short-term output data corresponding to the prediction based on the input;

[0051] Construct multiple sets of prediction results, integrate these prediction results, and form a probability distribution band interval that reflects the possibility of photovoltaic output under conditional constraints, thereby realizing the probabilistic prediction of photovoltaic output.

[0052] Furthermore, the cluster probability prediction model determines the accuracy of the prediction by the following formula:

[0053]

[0054] Where, L G is the loss function of the generator (G), L D is the loss function of the discriminator (D).

[0055] Furthermore, a method for quantitatively characterizing the positional relationship between photovoltaic units in a cluster by using an output characteristic diagram encoding method adopts a distributed photovoltaic cluster spatial encoding method based on a greedy algorithm, which includes the following steps:

[0056] The space where the distributed photovoltaic cluster is located is gridded, and the maximum grid size that meets the coding principle is found through greedy search;

[0057] On this basis, by cutting off the empty rows / columns around the matrix, the spatial coding matrix representing the location information of the distributed photovoltaic cluster can be obtained.

[0058] Furthermore, in the method of quantitatively characterizing the positional relationship between photovoltaic units in a cluster by using an output characteristic map encoding method, the photovoltaic output characteristic map construction method includes the following steps:

[0059] The distributed photovoltaic cluster output data is normalized and preprocessed to the maximum and minimum values, and compressed to the range of [0,1];

[0060] For the cluster output data at a certain moment, fill it into the spatial coding matrix to obtain the photovoltaic output characteristic diagram at that moment.

[0061] Furthermore, the weather feature map encoding method used in the matrixing of the time series meteorological labels by using the weather feature map construction method includes the following steps:

[0062] The Pearson Correlation Coefficient is used to quantify the average correlation between each weather variable in the meteorological data and the output of the photovoltaic power station;

[0063] The two selected features are filled into the spatial coding matrix in the form of cross-filling to maintain the consistency of data dimensions.

[0064] Data-driven methods use statistical or machine learning methods to obtain / learn inherent power features from actual measurements to predict short-term output. Generative adversarial networks (GANs) are an emerging branch of data-driven methods in recent years and have been widely used in the field of power systems. GANs can realistically learn complex nonlinear distribution features in data through adversarial training of generators and discriminators, and have a natural advantage in processing high-complexity data of photovoltaic power generation. GANs and their derivative models can adaptively capture implicit patterns in time series during training, especially for feature extraction of long-term trends and short-term fluctuations in photovoltaic data, so they are suitable for prediction work. At present, a large number of studies at home and abroad have applied generative adversarial networks to the field of wind and solar power generation output prediction of new energy sources.

[0065] The training of the GAN model is a substitution and adversarial process: G tries to generate samples G(z) that can deceive D; D learns to distinguish between G(z) and x by assigning a larger probability to x and a smaller probability to G(z). This process is described as a minimax game.

[0066]

[0067] Where P r and P d is the probability distribution of training data and generated data, E is the expectation operator,

[0068] cGAN is an extended version of the Generative Adversarial Network (GAN). By introducing the conditional vector (Condition Vector, c) as prior information, it gives the model the ability to conditionally control the generated results, thereby guiding the generator to generate samples that meet specific conditional constraints.

[0069] The prediction process of the present invention (i.e., the generation process of the generator) can be expressed as:

[0070]

[0071] in, is the generated prediction result, θ G are the network parameters of the generator G, c is the conditional vector, and z is the random vector. The true value (c, X) is used as the input of the discriminator to evaluate the effectiveness of the prediction results, that is:

[0072]

[0073] P r =D(C, X; θ D )∈[0,1] (4)

[0074] where θD is the discriminator parameter, P r is the discriminator score for the real sample, P g Generate a score for the sample.

[0075] Traditional cGAN models usually use binary cross-entropy (BCE) loss. The BCE loss function L of the generator and discriminator is BCE and L D They are:

[0076]

[0077] Among them, p r (c) is the conditional vector probability distribution; p r (c,X) is the joint probability distribution of the conditional vector and its corresponding real data; Is the data generated based on the input (c,z) The probability distribution of .

[0078] It should be noted that the BCE loss focuses on the morphological authenticity of the generated data rather than the prediction accuracy. Therefore, the present invention introduces the root mean square error (RMSE) loss based on the original loss function of the generator:

[0079]

[0080] The final generator loss function LG is the weighted sum of equation (5) and equation (7), that is:

[0081] L G =L BCE +αL RMSE (8)

[0082] Where α is a weight coefficient, which is used to balance the contribution of the two parts of loss. In summary, the training process of the network model of the present invention can be represented by the following minimum and maximum problem:

[0083]

[0084] Specifically, the embodiments of the present invention are further described below in conjunction with the accompanying drawings.

[0085] In one embodiment, see Figure 1 As shown, in an embodiment of the present invention, a short-term photovoltaic output prediction method based on deep learning is also provided. The system includes a clustered probability prediction model, an output characteristic map encoding method, and a weather characteristic map construction method.

[0086] See also Figure 2As shown, the cluster probabilistic prediction model 100 is used to generate photovoltaic output characteristic diagrams for multiple time periods in the future based on historical output characteristic diagrams and weather characteristic diagrams, thereby realizing probabilistic prediction of cluster output.

[0087] The output characteristic diagram encoding method 200 is used to quantitatively characterize the positional relationship between photovoltaic units in a cluster, thereby reflecting the spatial mutual information between photovoltaic units.

[0088] Its core coding principles include:

[0089] (1) Each grid in the matrix contains at most one photovoltaic unit;

[0090] (2) The same photovoltaic unit only exists in a unique matrix grid;

[0091] (3) Under the premise of satisfying the above principles (1) and (2), the dimension of the matrix should be reduced as much as possible to reduce its sparsity.

[0092] To achieve the above coding effect, the present invention proposes an automatic boundary calibration (Automatic Bounding Box) algorithm, and the pseudo code is shown in Algorithm 1. The core idea is to grid the space where the distributed photovoltaic cluster is located, and find the maximum grid size that meets the coding principle by greedy search. On this basis, the empty rows / columns around the matrix are cut off to obtain the spatial coding matrix that represents the location information of the distributed photovoltaic cluster. Based on this matrix, a feature map for prediction can be further constructed.

[0093]

[0094] The encoding result of the characteristic graph is related to the distribution and relative distance of the distributed photovoltaic units in the photovoltaic power station. Figure 3 As shown, this parameter is an example of a distributed photovoltaic power station, which includes 33 photovoltaic units.

[0095] See also Figure 4 As shown in Figure 1, the distributed photovoltaic cluster output data is normalized and preprocessed to the maximum and minimum values, and compressed to the range of [0,1]. For the cluster output data at a certain moment, it is filled into the spatial coding matrix to obtain the photovoltaic output characteristic diagram at that moment.

[0096] The weather characteristic map construction method 300 is used to matrix the time series meteorological labels to match the output characteristic map. The weather characteristic map construction method includes the following steps:

[0097] S301. Use the Pearson correlation coefficient to quantitatively analyze the average correlation between the photovoltaic output and weather variables in the photovoltaic power station data set and the output of the photovoltaic power station:

[0098]

[0099] Where W is the observed value of the meteorological variable, is the average value of meteorological variables, P is the observed value of photovoltaic power, is the average value of PV power, n is the number of observation samples, and m is the number of distributed PV units in the cluster.

[0100] S302, select the two meteorological labels with the highest correlation coefficient as features to construct a weather feature map to provide support for cluster output prediction. Considering that the weather characteristics faced by each photovoltaic unit in the cluster are highly similar, the two selected features are filled into the spatial coding matrix in the form of cross-filling to maintain the consistency of data dimensions.

[0101] In one embodiment, see Figure 5 As shown in Figure 1, the photovoltaic output characteristic map and the weather characteristic map together constitute the data input of the prediction model.

[0102] In an embodiment of the present invention, in order to evaluate the prediction accuracy of the clustered probability prediction model 100, the present invention adopts the mean absolute percentage error (MAPE) and the uncertainty prediction index - quantile score (QS) as the main evaluation indicators.

[0103] It should be noted that the cGAN model in the present invention will generate multiple sets of possible prediction results. Therefore, when calculating the prediction accuracy index, the average of the multiple sets of prediction results will be calculated.

[0104] MAPE is used to measure the deviation between the predicted value and the true value. Its advantage is that it can uniformly measure the error size of data of different dimensions. However, when the true value is low, the denominator in MAPE will significantly amplify the error value, resulting in distortion of its measurement of the overall error. In order to improve the effectiveness of MAPE as an evaluation indicator, the present invention ignores data with true values ​​lower than 0.1 when calculating MAPE, thereby avoiding the influence of too small denominator on the result. The calculation method of MAPE is:

[0105]

[0106] QS evaluates the effect of uncertainty prediction by calculating the coverage of the prediction interval to the true value. This indicator comprehensively evaluates the performance of the prediction model in uncertainty expression by quantifying the width of the prediction interval and its coverage of the true value. The calculation method is as follows:

[0107]

[0108] Among them, X is the actual power value, X i is the predicted value of the i-th quantile, and N is the number of samples. QS evaluates the accuracy of the prediction of a specific quantile.

[0109] In another embodiment of the present invention, the form of the prediction results needs to be consistent with the statistical characteristics and dynamic changes of the actual data. Since the prediction accuracy index cannot reflect the effectiveness of probabilistic prediction, the present invention calculates the confidence interval coverage rate (Coverage Rate, CR) and prediction interval width (Prediction Interval Width, PIW) to comprehensively measure the robustness of the model prediction results.

[0110] CR is defined as the proportion of actual values ​​that fall within the prediction interval. Ideally, CR = 100%, which means that the prediction interval can completely cover all actual values. However, in practical applications, CR often needs to be adjusted according to specific scenarios and fault tolerance requirements to strike a balance between coverage and model accuracy. CR is calculated as follows:

[0111]

[0112] Where CR(1-α) is the coverage rate under the confidence level (1-α), and N(1-α) is the number of samples whose measured photovoltaic power falls within the predicted confidence interval under the confidence level (1-α).

[0113] PIW measures the average width of the prediction interval. A narrower prediction interval indicates that the probability prediction result has a higher certainty, that is, it provides a credible prediction value within a smaller uncertainty range. The calculation method of PIW is as follows:

[0114]

[0115] Where L(1-α) is the lower bound of the prediction interval under the confidence level (1-α), and U(1-α) is the upper bound. It should be noted that in order to accurately measure the prediction effect, PIW needs to be considered together with CR.

[0116] Finally, by comparing the probability distribution of predicted data with that of real data, we can determine whether the predicted results are consistent with the characteristics of the real data at a macro level, and further verify the validity of the predicted results.

[0117] It should be understood that, although described in a certain order, these steps are not necessarily performed in sequence in the above order. Unless there is clear explanation in this article, the execution of these steps does not have strict order restriction, and these steps can be performed in other orders. Moreover, a part of the steps of the present embodiment may include a plurality of steps or a plurality of stages, and these steps or stages are not necessarily performed at the same time, but can be performed at different times, and the execution order of these steps or stages is not necessarily performed in sequence, but can be performed in turn or alternately with at least a part of the steps or stages in other steps or other steps.

[0118] A person skilled in the art can understand that all or part of the processes in the above-mentioned embodiment method can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the process of the embodiment of the above-mentioned method. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include at least one of non-volatile and volatile memory.

[0119] It should be understood that, as used herein, the singular form "a" or "an" is intended to include the plural form as well, unless the context clearly supports an exception. It should also be understood that, as used herein, "and / or" refers to any and all possible combinations of one or more of the items listed in association. The serial numbers of the embodiments disclosed in the above embodiments of the present invention are for description only and do not represent the advantages and disadvantages of the embodiments.

[0120] A person skilled in the art should understand that the discussion of any of the above embodiments is only exemplary and is not intended to imply that the scope of the disclosure of the embodiments of the present invention (including the claims) is limited to these examples; under the concept of the embodiments of the present invention, the technical features in the above embodiments or different embodiments can also be combined, and there are many other changes in different aspects of the embodiments of the present invention as above, which are not provided in detail for the sake of simplicity. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present invention should be included in the protection scope of the embodiments of the present invention.

Claims

1. A short-term photovoltaic output prediction method based on deep learning, characterized in that: include: Constructing a clustered probabilistic prediction model, which is used to establish a mapping between photovoltaic output and meteorological data and historical output using a conditional generative adversarial network, sample and generate different prediction results from the latent space under specific conditional constraint labels, and then construct a prediction result set containing multiple possible scenarios to achieve probabilistic prediction of photovoltaic output; The positional relationship between photovoltaic units in the cluster is quantitatively characterized by the output characteristic diagram encoding method; The weather characteristic map construction method is used to matrix the time series meteorological labels to maintain the consistency between the conditional labels constructed by meteorological data and the output data dimensions.

2. The photovoltaic output short-term prediction method based on deep learning according to claim 1, characterized in that: The process of generating prediction data by the clustered probability prediction model includes: Obtain real-time output data and meteorological data from photovoltaic power stations, construct historical photovoltaic output characteristic maps and historical weather characteristic maps, and construct inputs with random noise; generating short-term output data corresponding to the prediction based on the input; Construct multiple sets of prediction results, integrate these prediction results, and form a probability distribution band interval that reflects the possibility of photovoltaic output under conditional constraints, thereby realizing the probabilistic prediction of photovoltaic output.

3. The photovoltaic output short-term prediction method based on deep learning according to claim 1, characterized in that: The cluster probability prediction model determines the accuracy of the prediction through the following formula: Where, L G is the loss function of the generator, L D is the loss function of the discriminator.

4. The photovoltaic output short-term prediction method based on deep learning according to claim 1, characterized in that: The method for quantifying and characterizing the positional relationship between photovoltaic units in a cluster by encoding an output characteristic graph adopts a distributed photovoltaic cluster spatial encoding method based on a greedy algorithm, which includes the following steps: The space where the distributed photovoltaic cluster is located is gridded, and the maximum grid size that meets the coding principle is found through greedy search; On this basis, by cutting off the empty rows / columns around the matrix, the spatial coding matrix representing the location information of the distributed photovoltaic cluster can be obtained.

5. The photovoltaic output short-term prediction method based on deep learning according to claim 1, characterized in that: In the method of quantitatively characterizing the positional relationship between photovoltaic units in a cluster by encoding an output characteristic map, the photovoltaic output characteristic map construction method includes the following steps: The distributed photovoltaic cluster output data is normalized and preprocessed to the maximum and minimum values, and compressed to the range of [0,1]; For the cluster output data at a certain moment, fill it into the spatial coding matrix to obtain the photovoltaic output characteristic diagram at that moment.

6. The photovoltaic output short-term prediction method based on deep learning according to claim 1, characterized in that: The weather feature map encoding method used in the matrixing of the time series meteorological labels by using the weather feature map construction method includes the following steps: The Pearson correlation coefficient is used to quantitatively analyze the average correlation between each weather variable in the meteorological data and the output of the photovoltaic power station; The two selected features are filled into the spatial coding matrix in the form of cross-filling to maintain the consistency of data dimensions.

Images