Photovoltaic output prediction method of distributed photovoltaic power station

Abstract

The invention relates to a photovoltaic output prediction method for a distributed photovoltaic power station, and belongs to the technical field of fusion technology analysis in electric digital data processing. According to the method, a historical meteorological vector is constructed and clustered, and a result after clustering is used for secondary training of a photovoltaic output prediction model. And when photovoltaic output is predicted in real time, an error prediction vector and photovoltaic prediction output are obtained through the photovoltaic output prediction model after secondary training. And decomposing the error prediction vector to obtain modal components, calculating a joint index value of each modal component, comparing the joint index value with a calculated threshold value, screening all the modal components, and reconstructing to obtain low-frequency and high-frequency error prediction vectors. And respectively constructing matrixes and correcting the low-frequency and high-frequency error prediction vectors to obtain low-frequency and high-frequency error prediction correction vectors. And finally, predicting actual real-time photovoltaic output by combining the calculated weight and the photovoltaic predicted output. The method can accurately predict the photovoltaic output of a photovoltaic power station which is not equipped with electrical acquisition equipment.

Description

Technical Field

[0001] This invention relates to a method for predicting the photovoltaic output of a distributed photovoltaic power station, belonging to the technical field of fusion technology analysis (G06F18 / 25) in electrical digital data processing. Background Technology

[0002] Photovoltaic power generation is a rapidly developing renewable energy technology in recent years. Among them, distributed photovoltaic power stations have been built on a large scale due to their advantages of local consumption, reduced transmission losses, and adaptability to diverse application scenarios.

[0003] Currently, the sheer number and scattered distribution of distributed photovoltaic (PV) power stations make real-time monitoring of their photovoltaic output a challenge. From a technical standpoint, installing electrical data acquisition equipment in each distributed PV power station would suffice, but this would be prohibitively costly and resource-intensive. Consequently, some distributed PV power stations currently lack such equipment, creating monitoring "blind spots."

[0004] To fill data gaps, the industry has gradually adopted photovoltaic output analogy methods, which rely on the operating data of power plants with installed electrical data acquisition equipment in a region to predict the output of power plants without such equipment. However, because photovoltaic output is affected by the coupling of multiple parameters, existing analogy models cannot fully cover these variables, often resulting in large deviations in output curves and insufficient prediction accuracy, which restricts the grid's efficient dispatch of distributed photovoltaic power. Summary of the Invention

[0005] The technical problem to be solved by this invention is: how to accurately predict the photovoltaic output of a photovoltaic power station that is not equipped with electrical data acquisition equipment.

[0006] The technical solution proposed by this invention to solve the above-mentioned technical problems is: a method for predicting the photovoltaic output of a distributed photovoltaic power station, comprising the following steps:

[0007] Step 1: Obtain n historical meteorological data and n historical photovoltaic output data for a photovoltaic power station equipped with electrical data acquisition equipment over n historical periods; the historical meteorological data includes historical temperature, historical cloud cover, and historical irradiance;

[0008] Step 1.1: Standardize the n historical times and n historical meteorological data, and then construct n historical meteorological vectors H from the standardized n historical times and n historical meteorological data. i =(t i x i y i , z i ), i = 1, 2, ..., n; t i It is the i-th historical time, x iIt is the i-th historical temperature, y i This is the i-th historical cloud volume, z i It represents the i-th historical irradiance; each historical meteorological vector corresponds to a historical photovoltaic output.

[0009] Step 1.2: Cluster all historical meteorological vectors to obtain N historical clusters, each of which contains at least two historical meteorological vectors;

[0010] Step 2: Establish a photovoltaic power output prediction model. Input all historical meteorological vectors and their corresponding historical photovoltaic power outputs from one historical cluster into the photovoltaic power output prediction model for training. After training, input all historical meteorological vectors and their corresponding historical photovoltaic power outputs from another historical cluster into the photovoltaic power output prediction model for training, and so on, until all historical meteorological vectors and their corresponding historical photovoltaic power outputs from all historical clusters have been input into the photovoltaic power output prediction model and training is complete, resulting in the photovoltaic power output prediction model after the first training. The input of the photovoltaic power output prediction model after the first training is set as meteorological vectors, and the output is set as photovoltaic predicted power output.

[0011] Step 2.1: Input all historical meteorological vectors into the photovoltaic power output prediction model after the first training, and output n historical photovoltaic power outputs for n historical time periods. Construct a historical error vector by comparing the n historical photovoltaic power outputs for n historical time periods with the n historical photovoltaic power outputs for n historical time periods and the n historical photovoltaic power outputs for n historical time periods. =(x(t1),x(t2),…,x(t n )); x(t) n ) is the nth historical time t n The difference between historical photovoltaic output and historical predicted photovoltaic output;

[0012] Step 2.2: Input the historical error vector and n historical meteorological vectors into the photovoltaic power output prediction model after the first training to train it, and obtain the photovoltaic power output prediction model after the second training. The input of the photovoltaic power output prediction model after the second training is set as the meteorological vector, and the output is set as the error prediction vector and the photovoltaic predicted power output.

[0013] Step 3: Collect T real-time meteorological data for T real-time time periods at a sampling frequency S for a photovoltaic power station that is not equipped with electrical data collection equipment. The real-time meteorological data includes real-time temperature, real-time cloud cover, and real-time irradiance.

[0014] Construct T real-time meteorological vectors j = 1, 2, ..., T; It is the j-th real-time period. It is the j-th real-time temperature. This is the j-th real-time cloud volume. It is the j-th real-time irradiance;

[0015] Step 3.1: Input all real-time meteorological vectors into the photovoltaic power output prediction model after the second training, and output the real-time error prediction vector. And T real-time photovoltaic power output at T real-time times; It is the Tth real-time time The real-time error prediction value;

[0016] Step 3.2: Perform mode decomposition on the real-time error prediction vector to obtain M real-time error prediction mode components;

[0017] The first index value E of each real-time error prediction mode component is calculated according to the following formula (1). i

[0018] (1);

[0019] In equation (1), e i It is the energy of the i-th real-time error prediction mode component; It is the i-th real-time error prediction mode component at the j-th real-time time. The value on; It is the energy of the d-th real-time error prediction mode component;

[0020] Step 3.2: Calculate the second index value of each real-time error prediction mode component according to the following formula (2).

[0021] (2);

[0022] In equation (2), It is the average value of all values ​​in the i-th real-time error prediction mode component; It is the real-time error prediction vector at the j-th real-time time. The value on; It is the average value of all values ​​in the real-time error prediction vector;

[0023] Step 3.3: Calculate the joint index value X of each real-time error prediction mode component according to the following formula (3). i

[0024] (3)

[0025] In equation (3), , and These are the first weight, the second weight, and the third weight, respectively. , and All ≥ 0 and ; It is the first correction factor, 0 < <0.2; It is the Shannon entropy of the i-th real-time error prediction mode component;

[0026] Step 3.4: Calculate the threshold τ using the following formula (4).

[0027] (4);

[0028] In equation (4), It is the average of the combined index values; It is the standard deviation of the joint indicator value; It is the fourth weight, 0.1 < <3;

[0029] Step 3.5: For all joint index values ​​X greater than the threshold τ i Among the corresponding real-time error prediction mode components, the real-time error prediction mode components with a frequency ≤ 0.3S are taken as real-time low-frequency error prediction mode components, and the real-time error prediction mode components with a frequency > 0.3S are taken as real-time high-frequency error prediction mode components.

[0030] All real-time low-frequency error prediction mode components are reconstructed into a real-time low-frequency error prediction vector.

[0031] All real-time high-frequency error prediction mode components are reconstructed into a real-time high-frequency error prediction vector.

[0032] Step 4: Using a sliding window of length m, construct the low-frequency error prediction vector into a low-frequency matrix R1 as shown in equation (5).

[0033] (5);

[0034] In equation (5), It is the real-time low-frequency error prediction vector at the m-th real-time time. The value on;

[0035] Step 4.1: Input the low-frequency matrix R1 into the dynamic mode decomposition model to obtain the real-time low-frequency error prediction correction vector. ;

[0036] Step 4.2: Construct the high-frequency error prediction vector into a high-frequency matrix R2 as shown in equation (6) using a sliding window of length m.

[0037] (6);

[0038] In equation (6), It is the real-time high-frequency error prediction vector at the m-th real-time time. The value on;

[0039] Step 4.3: Input the high-frequency matrix R2 into the extended dynamic mode decomposition model to obtain the real-time high-frequency error prediction correction vector. ;

[0040] Step 5: Calculate the j-th real-time using the following formula (7) Low-frequency prediction error weighting and high-frequency prediction error weights

[0041] (7);

[0042] In equation (7), It is the value of the real-time error prediction vector at the k-th real-time point; It is the value of the real-time low-frequency error prediction correction vector at the k-th real-time time. It is the value of the real-time high-frequency error prediction correction vector at the k-th real-time time. It is the j-th real-time time Low-frequency uncertainty at time; It is the j-th real-time time High-frequency uncertainty at time;

[0043] Step 5.1: Calculate the j-th real-time using the following formula (8). Dynamic weights at time

[0044] (8);

[0045] In equation (8), It is the real-time error prediction vector at the j-th real-time time. The value on; It is the standard deviation of all values ​​in the historical error vector; It is the second correction factor, 0 < <0.25; K is the third correction factor, 1 < K < 20; It is a gated bias, 1 < <3;

[0046] Step 5.2: Predict the value of a photovoltaic power station without electrical data acquisition equipment at the j-th real-time using the following formula (9). Real-time photovoltaic output

[0047] (9);

[0048] In equation (9), It is the j-th real-time time Real-time photovoltaic power output forecast; It is the real-time low-frequency error prediction correction vector over j real-time times. The value on; Real-time high-frequency error prediction correction vector over j real-time times The value on.

[0049] Furthermore, in step 1.2, a density-based spatial clustering algorithm with noise is used for clustering.

[0050] Furthermore, the photovoltaic output prediction model in step 2 is established using a conditional generative adversarial network.

[0051] Furthermore, in step 3.2, mode decomposition is performed using discrete wavelet transform.

[0052] The beneficial effects of this invention are as follows: By introducing an error term into the photovoltaic power output prediction result and performing modal decomposition on the error vector to obtain modal components, the modal components are filtered according to specific screening conditions. Subsequently, the filtered modal components are divided into high-frequency and low-frequency components and reconstructed into high-frequency and low-frequency vectors respectively. Matrices are constructed for each type of vector and corrected to obtain high-frequency and low-frequency corrected vectors. Finally, through adaptive time weighting, the high-frequency and low-frequency corrected vectors are fused with the photovoltaic power output prediction result to output the actual photovoltaic power output. Compared to the low accuracy of traditional analogy methods, this invention can accurately predict the photovoltaic power output of photovoltaic power plants without electrical data acquisition equipment. Attached Figure Description

[0053] The present invention will be further described below with reference to the accompanying drawings.

[0054] Figure 1 This is a schematic diagram of the dynamic mode decomposition (DMD) model in the photovoltaic output prediction method of the distributed photovoltaic power station in the embodiment.

[0055] Figure 2 This is a schematic diagram of the Extended Dynamic Mode Decomposition (EDMD) model in the photovoltaic output prediction method for distributed photovoltaic power stations in the embodiment.

[0056] Figure 3 This is a real-time photovoltaic power output prediction diagram in the photovoltaic power output prediction method of the distributed photovoltaic power station in the embodiment. Detailed Implementation

[0057] Example

[0058] This embodiment of a method for predicting the photovoltaic output of a distributed photovoltaic power station includes the following steps:

[0059] Step 1: Obtain n historical meteorological data and n historical photovoltaic output data for a photovoltaic power station equipped with electrical data acquisition equipment over n historical periods; historical meteorological data includes historical temperature, historical cloud cover, and historical irradiance.

[0060] Step 1.1: Standardize n historical times and n historical meteorological data, and then construct n historical meteorological vectors H from the standardized n historical times and n historical meteorological data. i =(t i x i y i , z i ), i = 1, 2, ..., n; t i It is the i-th historical time, x i It is the i-th historical temperature, y i This is the i-th historical cloud volume, z i It is the i-th historical irradiance; each historical meteorological vector corresponds to a historical photovoltaic output.

[0061] Standardization is an existing technique that eliminates dimensional differences through mathematical transformation. Common methods include range standardization (Min-max) and Z-score standardization. This embodiment uses standardization to normalize time, temperature, cloud cover, and irradiance, eliminating the influence of dimensional differences and data range, making the data comparable, and providing a unified standard data foundation for subsequent clustering, prediction, and other processes.

[0062] Step 1.2: Cluster all historical meteorological vectors to obtain N historical clusters, each of which contains at least two historical meteorological vectors.

[0063] In this embodiment, density-based spatial clustering with noise application algorithm (DBSCAN) is used for clustering. This clustering algorithm is an existing clustering technique in the fields of data mining and machine learning. It can distinguish "noise points", "core points" and "density reachable points" in the data, thereby automatically dividing multiple clusters with similar meteorological characteristics. Each cluster obtained in this embodiment represents a set of data points with similar time, temperature, cloud cover and irradiance.

[0064] Step 2: Establish a photovoltaic power output prediction model. Input all historical meteorological vectors and their corresponding historical photovoltaic power outputs from one historical cluster into the photovoltaic power output prediction model for training. After training, input all historical meteorological vectors and their corresponding historical photovoltaic power outputs from another historical cluster into the photovoltaic power output prediction model for training, and so on, until all historical meteorological vectors and their corresponding historical photovoltaic power outputs from all historical clusters have been input into the photovoltaic power output prediction model and training is complete, resulting in the photovoltaic power output prediction model after the first training. The input of the photovoltaic power output prediction model after the first training is set as meteorological vectors, and the output is set as photovoltaic predicted power output.

[0065] In this embodiment, a photovoltaic power output prediction model is established using a conditional generative adversarial network (CGAN). A CGAN consists of a generator and a discriminator. The generator receives random noise and conditional vectors to generate samples, while the discriminator combines the conditions to distinguish between real and generated samples. The two are optimized through adversarial training for data prediction, representing existing technology in the fields of artificial intelligence and machine learning.

[0066] Step 2.1: Input all historical meteorological vectors into the photovoltaic power output prediction model after the first training, and output n historical photovoltaic power outputs for n historical time periods. Construct a historical error vector by comparing the n historical photovoltaic power outputs for n historical time periods with the n historical photovoltaic power outputs for n historical time periods and the n historical photovoltaic power outputs for n historical time periods. =(x(t1),x(t2),…,x(t n )); x(t) n ) is the nth historical time t n The difference between historical photovoltaic output and historical photovoltaic predicted output.

[0067] Step 2.2: Input the historical error vector and n historical meteorological vectors into the photovoltaic power output prediction model after the first training to train the photovoltaic power output prediction model after the second training. The input of the photovoltaic power output prediction model after the second training is set as the meteorological vector, and the output is set as the error prediction vector and the photovoltaic predicted power output.

[0068] Step 3: Collect T real-time meteorological data for a photovoltaic power station without electrical data acquisition equipment at a sampling frequency S for T real-time periods. The real-time meteorological data includes real-time temperature, real-time cloud cover, and real-time irradiance.

[0069] Construct T real-time meteorological vectors j = 1, 2, ..., T; It is the j-th real-time period. It is the j-th real-time temperature. This is the j-th real-time cloud volume. It is the j-th real-time irradiance.

[0070] Step 3.1: Input all real-time meteorological vectors into the photovoltaic power output prediction model after the second training, and output the real-time error prediction vector. And T real-time photovoltaic power output at T real-time times; It is the Tth real-time time The real-time error prediction value.

[0071] Step 3.2: Perform mode decomposition on the real-time error prediction vector to obtain M real-time error prediction mode components;

[0072] In this embodiment, Discrete Wavelet Transform (DWT) is used for mode decomposition. Discrete Wavelet Transform (DWT) is widely used in scenarios such as multi-scale signal decomposition, error sequence analysis, and data feature extraction. It is an existing technology in the field of signal processing and data processing, aiming to decompose the original signal into modal components of different frequencies through filtering and decomposition operations, thus providing a basis for subsequent feature selection.

[0073] The first index value E of each real-time error prediction mode component is calculated according to the following formula (1). i This is used to measure the "energy contribution" of each real-time error prediction mode component to the real-time error prediction vector.

[0074] (1);

[0075] In equation (1), e i It is the energy of the i-th real-time error prediction mode component; It is the i-th real-time error prediction mode component at the j-th real-time time. The value on; It is the energy of the d-th real-time error prediction mode component;

[0076] Step 3.2: Calculate the second index value of each real-time error prediction mode component according to the following formula (2). This reflects the degree of consistency between each real-time error prediction mode component and the overall error change; the second index value. The higher the value, the more "informative" the real-time error prediction mode component is about the error structure.

[0077] (2);

[0078] In equation (2), It is the average value of all values ​​in the i-th real-time error prediction mode component; It is the real-time error prediction vector at the j-th real-time time. The value on; It is the average value of all values ​​in the real-time error prediction vector;

[0079] Step 3.3: Calculate the joint index value X of each real-time error prediction mode component according to the following formula (3). i

[0080] (3)

[0081] In equation (3), , and These are the first weight, the second weight, and the third weight, respectively. , and All ≥ 0 and ; It is the first correction factor, 0 < <0.2, in this embodiment Take 0.1; It is the Shannon entropy of the i-th real-time error prediction mode component;

[0082] In equation (3), Shannon entropy is introduced to characterize the complexity of the modal components in real-time error prediction from the perspective of probability distribution. The larger the value, the more "disordered" the distribution of the real-time error prediction mode components is, and the closer it is to noise; Shannon entropy A smaller value indicates that the real-time error prediction mode component has a clearer structural feature. The calculation process of Shannon entropy is a current technique, which involves first dividing the amplitude range of the real-time error prediction mode component into several intervals, then statistically analyzing the probability of each interval, and finally substituting it into the Shannon entropy formula to obtain the result.

[0083] Step 3.4: Calculate the threshold τ using the following formula (4).

[0084] (4);

[0085] In equation (4), It is the average of the combined index values; It is the standard deviation of the joint indicator value; It is the fourth weight, 0.1 < <3, This embodiment Take 1.5;

[0086] Step 3.5: For all joint index values ​​X greater than the threshold τ i Among the corresponding real-time error prediction mode components, the real-time error prediction mode components with a frequency ≤ 0.3S are taken as real-time low-frequency error prediction mode components, and the real-time error prediction mode components with a frequency > 0.3S are taken as real-time high-frequency error prediction mode components.

[0087] All real-time low-frequency error prediction mode components are reconstructed into a real-time low-frequency error prediction vector.

[0088] All real-time high-frequency error prediction mode components are reconstructed into a real-time high-frequency error prediction vector.

[0089] This step improves the reliability of subsequent error matrix construction and correction results by eliminating components that contribute little to the prediction error, have weak correlation, and are statistically more like noise.

[0090] Step 4: Construct the low-frequency error prediction vector into a low-frequency matrix R1 as shown in equation (5) using a sliding window of length m.

[0091] (5);

[0092] In equation (5), It is the real-time low-frequency error prediction vector at the m-th real-time time. The value on;

[0093] The real-time low-frequency error prediction vector represents the long-term trend component of a photovoltaic power plant, such as seasonal variations and annual cycles. This low-frequency matrix R1 can capture changes in long-term trends.

[0094] Sliding window is an existing technique for processing time series data. Its core principle is to slide a fixed-length window of length m across continuous time series data at preset step sizes (1 in this embodiment), thereby extracting local subsequences, a total of m-1 subsequences. A matrix R1 is then constructed based on these subsequences. The matrix contains... This corresponds to the last value of the time series data, which is the real-time low-frequency error prediction vector in this embodiment at the T-th real-time time. The value on .

[0095] Step 4.1: Input the low-frequency matrix R1 into the dynamic mode decomposition model, and output the real-time low-frequency error prediction correction vector. .

[0096] Dynamic mode decomposition (DMD) is an existing technology widely used in power systems and signal processing. Its principle is as follows: Figure 1 As shown in the figure. In this embodiment, a dynamic mode decomposition model is used to extract low-frequency dynamic modes from the low-frequency matrix R1 constructed from the low-frequency error prediction vector, capture the changing characteristic values ​​of long-term trends, and then correct slow-varying errors such as seasonality and annual periodicity in photovoltaic power generation.

[0097] Step 4.2: Construct the high-frequency error prediction vector into a high-frequency matrix R2 as shown in equation (6) using a sliding window of length m.

[0098] (6);

[0099] In equation (6), It is the real-time high-frequency error prediction vector at the m-th real-time time. The value on.

[0100] The real-time high-frequency error prediction vector represents the short-term fluctuation component, such as weather changes and irradiance fluctuations. This matrix can capture the dynamic changes of high-frequency fluctuations.

[0101] Step 4.3: Input the high-frequency matrix R2 into the extended dynamic mode decomposition model, and output the real-time high-frequency error prediction correction vector. .

[0102] Extended Dynamic Mode Decomposition (EDMD) is also an existing technology widely used in power systems and signal processing. Its principle is as follows: Figure 2 As shown in the figure. In this embodiment, extended dynamic mode decomposition is used to extract high-frequency dynamic modes from the high-frequency matrix constructed from the high-frequency error prediction vector, and then mapped to a high-dimensional feature space by a kernel function. This captures the changing feature values ​​of high-frequency fluctuations, thereby correcting the short-term rapid errors caused by weather and irradiance fluctuations in photovoltaic power generation.

[0103] Step 5: Calculate the j-th real-time using the following formula (7) Low-frequency prediction error weighting and high-frequency prediction error weights

[0104] (7);

[0105] In equation (7), It is the value of the real-time error prediction vector at the k-th real-time point; It is the value of the real-time low-frequency error prediction correction vector at the k-th real-time time. It is the value of the real-time high-frequency error prediction correction vector at the k-th real-time time. It is the j-th real-time time Low-frequency uncertainty at time; It is the j-th real-time time High-frequency uncertainty at time.

[0106] Step 5.1: Calculate the j-th real-time using the following formula (8). Dynamic weights at time

[0107] (8);

[0108] In equation (8), It is the real-time error prediction vector at the j-th real-time time. The value on; It is the standard deviation of all values ​​in the historical error vector; It is the second correction factor, 0 < <0.25; K is the third correction factor, 1 < K < 20; It is a gated bias, 1 < <3; This embodiment Set K to 0.2 and K to 10. Take 2.

[0109] Step 5.2: Predict the value of a photovoltaic power station without electrical data acquisition equipment at the j-th real-time using the following formula (9). Real-time photovoltaic output

[0110] (9);

[0111] In equation (9), It is the j-th real-time time Real-time photovoltaic power output forecast; It is the real-time low-frequency error prediction correction vector over j real-time times. The value on; Real-time high-frequency error prediction correction vector over j real-time times The value on.

[0112] In this embodiment, the real-time photovoltaic power output of a photovoltaic power station without electrical data acquisition equipment at various real-time times is as follows: Figure 3 As shown by the dashed line, and with the actual photovoltaic output obtained by temporarily installing electrical data acquisition equipment before the forecast, the actual output is as follows: Figure 3 As shown by the solid line, it can be seen that the real-time photovoltaic power output prediction is basically consistent with the actual photovoltaic power output, indicating high accuracy.

[0113] The above description is only a preferred embodiment of the present invention, but the present invention is not limited thereto. All equivalent substitutions or modifications made to the concepts and technical solutions of the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for predicting the photovoltaic output of a distributed photovoltaic power station, characterized in that; Includes the following steps: Step 1: Obtain n historical meteorological data and n historical photovoltaic output data for a photovoltaic power station equipped with electrical data acquisition equipment over n historical periods; the historical meteorological data includes historical temperature, historical cloud cover, and historical irradiance; Step 1.1: Standardize the n historical times and n historical meteorological data, and then construct n historical meteorological vectors H from the standardized n historical times and n historical meteorological data. i =(t i x i y i , z i ), i = 1, 2, ..., n; t i It is the i-th historical time, x i It is the i-th historical temperature, y i This is the i-th historical cloud volume, z i It represents the i-th historical irradiance; each historical meteorological vector corresponds to a historical photovoltaic output. Step 1.2: Cluster all historical meteorological vectors to obtain N historical clusters, each of which contains at least two historical meteorological vectors; Step 2: Establish a photovoltaic power output prediction model. Input all historical meteorological vectors and their corresponding historical photovoltaic power outputs from one historical cluster into the photovoltaic power output prediction model for training. After training, input all historical meteorological vectors and their corresponding historical photovoltaic power outputs from another historical cluster into the photovoltaic power output prediction model for training, and so on, until all historical meteorological vectors and their corresponding historical photovoltaic power outputs from all historical clusters have been input into the photovoltaic power output prediction model and training is complete, resulting in the photovoltaic power output prediction model after the first training. The input of the photovoltaic power output prediction model after the first training is set as meteorological vectors, and the output is set as photovoltaic predicted power output. Step 2.1: Input all historical meteorological vectors into the photovoltaic power output prediction model after the first training, and output n historical photovoltaic power outputs for n historical time periods. Construct a historical error vector by comparing the n historical photovoltaic power outputs for n historical time periods with the n historical photovoltaic power outputs for n historical time periods and the n historical photovoltaic power outputs for n historical time periods. =(x(t1),x(t2),…,x(t n )); x(t) n ) is the nth historical time t n The difference between historical photovoltaic output and historical predicted photovoltaic output; Step 2.2: Input the historical error vector and n historical meteorological vectors into the photovoltaic power output prediction model after the first training to train it, and obtain the photovoltaic power output prediction model after the second training. The input of the photovoltaic power output prediction model after the second training is set as the meteorological vector, and the output is set as the error prediction vector and the photovoltaic predicted power output. Step 3: Collect T real-time meteorological data for T real-time time periods at a sampling frequency S for a photovoltaic power station that is not equipped with electrical data collection equipment. The real-time meteorological data includes real-time temperature, real-time cloud cover, and real-time irradiance. Construct T real-time meteorological vectors j = 1, 2, ..., T; It is the j-th real-time period. It is the j-th real-time temperature. This is the j-th real-time cloud volume. It is the j-th real-time irradiance; Step 3.1: Input all real-time meteorological vectors into the photovoltaic power output prediction model after the second training, and output the real-time error prediction vector. And T real-time photovoltaic power output at T real-time times; It is the Tth real-time time The real-time error prediction value; Step 3.2: Perform mode decomposition on the real-time error prediction vector to obtain M real-time error prediction mode components; The first index E of each real-time error prediction mode component is calculated according to the following formula (1). i （1）； In equation (1), e i It is the energy of the i-th real-time error prediction mode component; It is the i-th real-time error prediction mode component at the j-th real-time time. The value on; It is the energy of the d-th real-time error prediction mode component; Step 3.2: Calculate the second index value of each real-time error prediction mode component according to the following formula (2). （2）； In equation (2), It is the average value of all values ​​in the i-th real-time error prediction mode component; It is the real-time error prediction vector at the j-th real-time time. The value on; It is the average value of all values ​​in the real-time error prediction vector; Step 3.3: Calculate the joint index value X of each real-time error prediction mode component according to the following formula (3). i （3） In equation (3), , and These are the first weight, the second weight, and the third weight, respectively. , and All ≥ 0 and ; It is the first correction factor, 0 < <0.2; It is the Shannon entropy of the i-th real-time error prediction mode component; Step 3.4: Calculate the threshold τ using the following formula (4). （4）； In equation (4), It is the average of the combined index values; It is the standard deviation of the joint indicator value; It is the fourth weight, 0.1 < <3; Step 3.5: For all joint index values ​​X greater than the threshold τ i Among the corresponding real-time error prediction mode components, the real-time error prediction mode components with a frequency ≤ 0.3S are taken as real-time low-frequency error prediction mode components, and the real-time error prediction mode components with a frequency > 0.3S are taken as real-time high-frequency error prediction mode components. All real-time low-frequency error prediction mode components are reconstructed into a real-time low-frequency error prediction vector. All real-time high-frequency error prediction mode components are reconstructed into a real-time high-frequency error prediction vector. Step 4: Construct the low-frequency error prediction vector into a low-frequency matrix R1 as shown in equation (5) using a sliding window of length m. （5）； In equation (5), It is the real-time low-frequency error prediction vector at the m-th real-time time. The value on; Step 4.1: Input the low-frequency matrix R1 into the dynamic mode decomposition model to obtain the real-time low-frequency error prediction correction vector. ; Step 4.2: Construct the high-frequency error prediction vector into a high-frequency matrix R2 as shown in equation (6) using a sliding window of length m. （6）； In equation (6), It is the real-time high-frequency error prediction vector at the m-th real-time time. The value on; Step 4.3: Input the high-frequency matrix R2 into the extended dynamic mode decomposition model to obtain the real-time high-frequency error prediction correction vector. ; Step 5: Calculate the j-th real-time using the following formula (7) Low-frequency prediction error weighting and high-frequency prediction error weights （7）； In equation (7), It is the value of the real-time error prediction vector at the k-th real-time point; It is the value of the real-time low-frequency error prediction correction vector at the k-th real-time time. It is the value of the real-time high-frequency error prediction correction vector at the k-th real-time time. It is the j-th real-time time Low-frequency uncertainty at time; It is the j-th real-time time High-frequency uncertainty at time; Step 5.1: Calculate the j-th real-time using the following formula (8). Dynamic weights at time （8）； In equation (8), It is the real-time error prediction vector at the j-th real-time time. The value on; It is the standard deviation of all values ​​in the historical error vector; It is the second correction factor, 0 < <0.25; K is the third correction factor, 1 < K < 20; It is a gated bias, 1 < <3; Step 5.2: Predict the value of a photovoltaic power station without electrical data acquisition equipment at the j-th real-time using the following formula (9). Real-time photovoltaic output （9）； In equation (9), It is the j-th real-time time Real-time photovoltaic power output forecast; It is the real-time low-frequency error prediction correction vector over j real-time times. The value on; Real-time high-frequency error prediction correction vector over j real-time times The value on.

2. The photovoltaic output prediction method for distributed photovoltaic power stations according to claim 1, characterized in that: In step 1.2, a density-based spatial clustering algorithm with noise is used for clustering.

3. The photovoltaic output prediction method for distributed photovoltaic power stations according to claim 1, characterized in that: The photovoltaic output prediction model in step 2 is established using a conditional generative adversarial network.

4. The photovoltaic output prediction method for distributed photovoltaic power stations according to claim 1, characterized in that: In step 3.2, mode decomposition is performed using discrete wavelet transform.

Images