Distributed power station short-term power prediction method based on data enhancement and MICN

Abstract

The invention relates to the technical field of meteorological monitoring, and discloses a distributed power station short-term power prediction method based on data enhancement and MICN, and the method comprises the following steps: S1, selecting related meteorological factors; s2, inputting meteorological data and distance data of adjacent sites into a Gaussian process regression model; and S3, taking the power data of the target distributed power station and the calculated meteorological data as input and substituting the input into the MICN model to obtain power generation power data in the future 72 hours. The step S1 is specifically as follows: under the condition that internal technical parameters of the photovoltaic power station are basically kept unchanged, main factors influencing the photovoltaic output power are screened out from external meteorological factors, and the correlation degree between the photovoltaic output power and the influence factors is measured by using a Pearson's correlation coefficient R (X, Y). According to the invention, the problem that the meteorological factor prediction of the to-be-predicted distributed photovoltaic power station needs four adjacent meteorological station data is solved, and the prediction accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological monitoring, and specifically provides a short-term power prediction method for distributed power stations based on data augmentation and MICN. Background Art

[0002] Most distributed power stations do not have accurate meteorological monitoring equipment and lack corresponding meteorological data, which brings difficulties to the accurate prediction of distributed photovoltaic power output. To effectively solve the impact of distributed photovoltaic power generation stations on the power grid and ensure the safe operation of the power grid and the accuracy of power system dispatching, it is increasingly necessary to master the power generation situation in advance. Commonly, the meteorological factors of the target meteorological station are calculated by the bilinear interpolation method, and then the power output of distributed photovoltaic power stations is predicted through deep learning algorithms such as LSTNet, Transformer, LSTM, and CNN.

[0003] The existing technologies have the following defects: 1. When calculating the meteorological data of the target distributed power station by the bilinear interpolation method, it is limited by the number of adjacent meteorological stations. There must be at least 4 adjacent meteorological stations, and the actual operation is highly restricted; 2. The bilinear interpolation method has a high error in calculating meteorological factors, and meteorological factors do not change simply linearly; 3. The existing deep learning algorithms such as LSTNet, Transformer, LSTM, and CNN have poor power prediction effects in power stations; Therefore, we propose a short-term power prediction method for distributed power stations based on data augmentation and MICN to solve the problem. Summary of the Invention

[0004] (1) Technical Problems to be Solved Aiming at the deficiencies of the existing technologies, the present invention provides a short-term power prediction method for distributed power stations based on data augmentation and MICN, which solves the problems in the above background art.

[0005] (2) Technical Solutions To achieve the above object, the present invention provides the following technical solutions: A short-term power prediction method for distributed power stations based on data augmentation and MICN, including the following steps: S1: Confirm information such as the installed capacity, power generation rate, and address location to be predicted, confirm the geographical location information of adjacent meteorological stations around the power station to be predicted, and then select relevant meteorological factors; S2: Input the meteorological data and distance data of adjacent stations into the Gaussian process regression model; S3: Take the power data of the target distributed power station and the calculated meteorological data as inputs and bring them into the MICN model to obtain the power generation data for the next 72 hours.

[0006] Preferably, the specific steps of S1 are as follows: Under the condition that the internal technical parameters of the photovoltaic power station remain basically unchanged, it is crucial to screen out the main factors affecting the photovoltaic output power from external meteorological factors. The Pearson correlation coefficient R(X, Y) is used to measure the correlation degree between the photovoltaic output power and each influencing factor. The specific formula is as follows: (1).

[0007] Preferably, and are respectively the output power of the photovoltaic power station sample i and the meteorological factor sample point after standardization; X and Y are respectively the matrices composed of and ; and are respectively and The average value of; i is the number of samples. The value range of the correlation coefficient is [-1, 1]. The larger its absolute value, the stronger the correlation, and the meteorological factors related to the power are screened out.

[0008] Preferably, the specific steps of S2 are as follows: The model expression of Gaussian process regression: (2); In the formula, X is the input vector, is an observation noise vector that is mutually independent and follows a Gaussian distribution, and its variance is , is to describe and The potential function of the mapping relationship between them, where represents the plane coordinate information of the meteorological station, Y is the meteorological data, is the coordinate of the power station, is the meteorological data of the site to be predicted. The Gaussian process can be expressed as: (3); Among them: (4); represents the covariance matrix of the meteorological station coordinates X of the training input samples. Each element k(X i , X j ) represents the correlation degree between the training input samples X i and X j , K(X’, X) = K(X, X’) T is the covariance matrix between the training input sample X and the predicted input sample ; is the coordinate of the predicted input sample site to be predicted The variance. According to the conditional distribution property of the Gaussian distribution, the predicted value can be obtained. The posterior distribution of: (5); (6); (7); x and y are the known coordinate data of meteorological stations, , is the coordinate data of the target station; (8); Using this covariance function to perform GPR modeling on the data, an interpolation model considering the directional change characteristics of geographical phenomena can be obtained, and the meteorological data can be calculated.

[0009] Preferably, the specific steps of S3 are as follows: The model uses a multi-scale hybrid decomposition block to separate the input sequence, then uses a seasonal prediction block to predict seasonal information, uses a trend cycle prediction block to predict trend cycle information, and then adds the prediction results to obtain the final predicted power data.

[0010] (III) Beneficial Effects Compared with the prior art, the present invention provides a short-term power prediction method for distributed power generation stations based on data enhancement and MICN, having the following beneficial effects: (1), The present invention solves the limitation that 4 adjacent meteorological station data are required for predicting meteorological factors of the distributed photovoltaic power generation station to be predicted, and improves the prediction accuracy; (2), The present invention improves the accuracy of power prediction for distributed photovoltaic power generation stations. Description of the Drawings

[0011] Figure 1 is a schematic system flow diagram of the present invention; Figure 2 is a schematic diagram for comparing prediction results of the present invention. Specific Embodiments

[0012] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0013] A short-term power prediction method for distributed power generation stations based on data enhancement and MICN, comprising the following steps: S1: Confirm the information such as the installed capacity of the power station to be predicted, the power generation rate, and the address location. Confirm the geographical location information of the nearby meteorological stations around the power station to be predicted. Then select the relevant meteorological factors. When the internal technical parameters of the photovoltaic power station remain basically unchanged, it is crucial to screen out the main factors affecting the photovoltaic output power from the external meteorological factors, specifically including light irradiation intensity, temperature, humidity, wind speed, etc. Use the Pearson correlation coefficient R(X, Y) to measure the correlation degree between the photovoltaic output power and each influencing factor. The specific formula is as follows: (1); and are respectively the output power of the photovoltaic power station sample i and the meteorological factor sample point after standardization; X and Y are respectively matrices composed of and ; and are respectively and 's average values; i is the number of samples. The value range of the correlation coefficient is [-1, 1]. The larger its absolute value, the stronger the correlation. Screen out the meteorological factors related to power; S2: Input the meteorological data and distance data of the nearby stations into the Gaussian process regression model. The model expression of Gaussian process regression: (2); In the formula, X is the input vector, is an observation noise vector that is mutually independent and follows a Gaussian distribution, and its variance is , is a potential function describing the mapping relationship between and . Among them, represents the plane coordinate information of the meteorological station, Y is the meteorological data, is the coordinate of the power station, is the meteorological data of the site to be predicted. The Gaussian process can be expressed as: (3); Among them: (4); represents the covariance matrix of the meteorological station coordinates X of the training input samples. Each element k(X i , X j ) in the matrix represents the correlation degree between the training input samples X i and X j . K(X’, X) = K(X, X’) T is the covariance matrix between the training input samples X and the predicted input samples ; To predict the variance of the coordinates of the site to be predicted for the input sample According to the conditional distribution property of the Gaussian distribution, the predicted value The posterior distribution of: (5); (6); (7); x and y are the known coordinate data of the meteorological stations, , is the coordinate data of the target site; (8); Using this covariance function to perform GPR modeling on the data, an interpolation model considering the directional change characteristics of geographical phenomena can be obtained, and the meteorological data can be calculated; S3: Taking the power data of the target distributed power generation station and the calculated meteorological data as inputs and bringing them into the MICN model to obtain the power generation power data for the next 72 hours. The model uses a multi-scale hybrid decomposition block to separate the input sequence, then uses a seasonal prediction block to predict seasonal information, uses a trend cycle prediction block to predict trend cycle information, and then adds the prediction results to obtain the final predicted power data.

[0014] The first layer of MICN is a multi-scale hybrid decomposition block, which uses multiple different Avgpool kernels to process the time series of the input meteorological data and power information into seasonal terms and periodic terms , and the process is as shown in formulas (9) and (10): (9); (10); Trend periodic prediction uses a linear regression model to simply predict the power generation power, where is the periodic sequence after multi-scale hybrid decomposition. The specific formula is as shown below (11), represents the power of the distributed power generation station: (11); The seasonal prediction module focuses on more complex seasonality. After the input sequence Xs is embedded, the model uses multi-scale equidistant convolution to predict future information. Each MIC layer has multiple Branches representing different scales to capture local features in the time series of meteorological data information and power information, and by combining the results of different Branches, the comprehensive information utilization of the sequence is completed. The calculation process is as shown below (12), (13), (14), (15): (12); (13); (14); (15); The embedding process fuses the meteorological data features extracted by VE with the TFE - represented time - feature encoding information and the PE position information, and passes the output result to the MIC layer.

[0015] The multi - scale hybrid convolution layer (MIC) simulates different time patterns through multiple branches with different scales, extracts local information features for reasoning. The Local - Global module is formed by connecting in series the Local module that aggregates local features and the Global module that models the relationships between all local features.

[0016] After the local features obtain the corresponding single pattern through Avgpool, one - dimensional convolution is used for downsampling. The process is as follows (16): (16); The Local part first performs average pooling on the input with kernel = i, and then performs downsampling with one - dimensional convolution with stride = kernel = i. This step reduces the sequence to 1 / i of the original. The meteorological data features at each adjacent time point are aggregated into a local feature.

[0017] The Global module models the features of each previous Local part through equidistant convolution on the output of the Local module, then obtains the global relationship, and restores it to the original length through transposed convolution upsampling. Finally, the power matrix is obtained through residual connection and regularization process. As shown in formulas (17) and (18) (17); (18); Finally, the results predicted by the linear branch and the non - linear branch are superimposed to obtain the predicted power of distributed photovoltaic power generation. As shown in formula (19): (19); In summary, 1. Through the processing method of Gaussian process regression in this method, the flexibility of using adjacent meteorological data is improved, and it is not limited to using only 4 adjacent meteorological data. Under the same meteorological data conditions, the accuracy is increased by 5%; 2. By predicting power through the MICN model, the accuracy is increased by nearly 33% compared with models such as LSTM, as shown in the following table:

[0018] Comparison of Prediction Results

[0019] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A short-term power prediction method for distributed power stations based on data enhancement and MICN, characterized in that: The following steps are involved: S1: Confirm the installed capacity, power generation rate, address and other information to be predicted, confirm the geographical location information of the meteorological stations near the power station to be predicted, and then select relevant meteorological factors; S2: input the meteorological data and distance data of nearby sites into the Gaussian process regression model; S3: The power data of the target distributed power station and the calculated meteorological data are taken as input into the MICN model to obtain the power generation data for the next 72 hours.

2. According to claim 1, a method for short-term power prediction of distributed power stations based on data enhancement and MICN is characterized in that: Step S1 is as follows: When the internal technical parameters of the photovoltaic power station remain basically unchanged, it is crucial to screen out the main factors that affect the photovoltaic output power from the external meteorological factors. The Pearson correlation coefficient R (X, Y) is used to measure the correlation between the photovoltaic output power and each influencing factor. The specific formula is as follows: （1）。 3. According to claim 2, a method for short-term power prediction of distributed power stations based on data enhancement and MICN is characterized in that: and are the output power and meteorological factor sample points of the standardized photovoltaic power station sample i; X and Y are and The matrix composed of and They are and The average value of; i is the number of samples, the range of the correlation coefficient is [-1, 1]. The larger the absolute value, the stronger the correlation, and the meteorological factors related to power are screened out.

4. According to claim 3, a method for short-term power prediction of distributed power stations based on data enhancement and MICN is characterized in that: The specific S2 step is: the model expression of Gaussian process regression: （2）； In the formula, X is the input vector. are independent observation noise vectors that follow a Gaussian distribution, and their variance is , For description and The potential function of the mapping relationship between Represents the plane coordinate information of the weather station, For meteorological data, are the power station coordinates, For the meteorological data of the site to be predicted, the Gaussian process can be expressed as: （3）； in: （4）； Represents the covariance matrix of the training input sample weather station coordinates X. Each element k(X i ,X j ) represents the training input sample X i With X j The degree of correlation, K(X',X)=K(X,X') T is the training input sample X and the prediction input sample The covariance matrix between ; Input sample for prediction. Coordinates of the site to be predicted The variance of , according to the conditional distribution property of Gaussian distribution, can get the predicted value The posterior distribution of : （5）； （6）； （7）； x, y are the known weather station coordinate data, , The coordinate data of the target site; （8）； By using this covariance function to model the data using GPR, we can obtain an interpolation model that takes into account the directional change characteristics of geographical phenomena and calculate meteorological data.

5. A method for short-term power prediction of distributed power stations based on data enhancement and MICN according to claim 4, characterized in that: The specific steps of step S3 are as follows: the model uses a multi-scale mixed decomposition block to separate the input sequence, then uses the seasonal prediction block to predict seasonal information, uses the trend cycle prediction block to predict trend cycle information, and then adds the prediction results to obtain the final predicted power data.

Images