Distributed photovoltaic power station cluster power ultra-short-term grid prediction method

Abstract

The application discloses a distributed photovoltaic power station power ultra-short-term grid prediction method, and the technical scheme steps of the application comprise the following steps: firstly, based on an adaptive grid division method, a similar day screening method containing singular spectrum analysis is designed; a three-dimensional convolution network input feature set is constructed based on a Holt-Winters exponential smoothing method; finally, based on the grid division and similar day screening results, and according to the input feature set, a three-dimensional convolution neural network is constructed to establish a distributed photovoltaic power station cluster power ultra-short-term grid prediction model. The prediction method can fully consider the time and space characteristics of the distributed photovoltaic, and effectively improves the accuracy and robustness of the distributed photovoltaic power station cluster power prediction.

Description

Technical Field

[0001] This invention relates to the field of distributed photovoltaic power cluster prediction, and specifically to a gridded method for screening similar days using singular spectrum analysis, particularly an ultra-short-term gridded prediction method for distributed photovoltaic power plant cluster power. Technical Background

[0002] Power prediction for distributed photovoltaic (PV) power plants refers to the prediction of the sum of the day-to-day power curves of all distributed PV power plants within a region. Distributed PV systems are characterized by small individual capacities and dispersed installations, but are numerous and exhibit high penetration and density in some areas. The accuracy of individual distributed PV power predictions is low, and their impact on grid dispatching is minimal. Regional prediction for multiple distributed PV clusters offers significant advantages in accuracy and applicability. To address these issues, a gridded short-term power prediction method for distributed PV power plants is proposed. The key to this method lies in singular spectrum similarity day screening and a prediction model based on a three-dimensional convolutional neural network. This model extracts the spatiotemporal information of the input matrix, establishes nonlinear correlations between local and global data, enriches the input feature space, and thus improves prediction accuracy. Summary of the Invention

[0003] The purpose of this invention is to propose an ultra-short-term gridded prediction method for the power of distributed photovoltaic (PV) power plants. This invention designs an adaptive grid partitioning method based on a Cartesian coordinate system, utilizes singular spectrum analysis to screen similar days, constructs a feature set for a three-dimensional convolutional neural network based on Holt-Winters exponential smoothing, and proposes an ultra-short-term prediction method for distributed PV power plants based on a three-dimensional convolutional neural network, thereby improving the accuracy and robustness of the prediction.

[0004] This invention employs the following technical solution: establishing a method for ultra-short-term gridded prediction of distributed photovoltaic power station cluster power, comprising the following steps:

[0005] (1) Based on the adaptive grid partitioning method, design a similar day screening method with singular spectrum analysis;

[0006] (2) Construct the input feature set of the three-dimensional convolutional network based on the Holt-Winters exponential smoothing method;

[0007] (3) Based on the grid division and similar day screening results, and according to the input feature set, a three-dimensional convolutional neural network is used to establish an ultra-short-term grid prediction model for distributed photovoltaic power stations.

[0008] Specifically, in step (1), the similarity day selection method based on the adaptive grid partitioning method and incorporating singular spectrum analysis is described as follows:

[0009] The specific steps of the proposed adaptive mesh generation are as follows:

[0010] 1) Place the distributed photovoltaic power station cluster in a certain area into a Cartesian coordinate system to obtain the horizontal axis length A and the vertical axis length B, and determine the initial step size of the horizontal axis division. Divide the step size along the vertical axis ,in , ;

[0011] 2) According to , In a two-dimensional plane, a distributed photovoltaic power station cluster is divided into grids. Within each grid, the two most distant distributed photovoltaic power stations are selected, and the cosine similarity of their power output at the same time is calculated. , Power curves of two photovoltaic power plants and The included angle, In the formula: This represents the actual power value at time t, where T represents the total number of times.

[0012] 3) For The grid is then further updated with the step size for the horizontal and vertical axes. , ,in , These represent the step sizes of the x and y axes after the m and nth iterations, respectively. , ;

[0013] 4) When all grid cosine similarity When the time comes, the partition ends;

[0014] The specific implementation steps of the similar day screening method involving singular spectrum analysis are as follows:

[0015] 1) First, obtain power data from the photovoltaic database, and then normalize the power of the photovoltaic power plants in each grid after division to obtain the average daily power curve of the photovoltaic power plants in each grid:

[0016] (1)

[0017] In the above formula, This represents the photovoltaic power station in the k-th grid. The average power at time t, K represents the number of grid cells after partitioning, and J is the number of distributed photovoltaic power stations contained in one grid. and Representing the first The daily power curve and installed capacity of a distributed photovoltaic power station at time t, where t is the time, T is the total number of time moments, and j is the photovoltaic power station number;

[0018] 2) Obtain the power curve within the predicted daily grid using the Holt-Winters exponential smoothing method. We then perform singular spectrum analysis on the samples to find the set of similar days that are closest to the predicted day characteristics.

[0019] Will Singular spectral decomposition is performed, and then the curve is obtained using the Dynamic Time Warping (DTW) algorithm. , Similarity between:

[0020] (2)

[0021] in: This indicates the similarity of the k-th grid curve; but This method can only represent the similarity of a single grid cell, and cannot represent the similarity of power curves across the entire region. This invention further uses a weighted method based on grid similarity to represent the similarity of regional power curves:

[0022] (3)

[0023] In the above formula, m+3 and n+3 represent the number of columns and rows of the grid, respectively; M represents the rated capacity of the k-th grid, and M represents the total rated capacity of the photovoltaic power plants in the entire region. This represents the daily power curve of the distributed photovoltaic system in the i-th row and r-th column and the characteristic distance of the most similar day.

[0024] Specifically, in step (2), the input feature set of the three-dimensional convolutional network constructed based on the Holt-Winters exponential smoothing method is described as follows:

[0025] (4)

[0026] (5)

[0027] (6)

[0028] (7)

[0029] In the above formula, This represents the predicted value at time t+1. , , Do not correspond to the time-time level components, trend components, and seasonal components. All are smoothing coefficients. Let t be the actual value at time t, and u be the seasonal frequency. When predicting data for the next 24 hours, u = 24.

[0030] The final feature set is as follows:

[0031] (8)

[0032] In the above formula, To predict the power at L time points after time t.

[0033] Specifically, in step (3), the establishment of a distributed photovoltaic power station ultra-short-term grid prediction model using a three-dimensional convolutional neural network is described as follows:

[0034] Since the region has been divided into grids, with the longitude and latitude of each region corresponding to the divided W rows and H columns, the input matrix of the 3D convolutional neural network is: Where C is the number of features, specifically including similar days within the grid. The system includes the level component, seasonal component, and photovoltaic power and irradiance data for the predicted day. W is the width of the matrix, H is the height of the matrix, and L represents continuous time intervals, used to characterize the temporal characteristics, as shown below:

[0035] (9)

[0036] (10)

[0037] (11)

[0038] This invention utilizes Taking data from L times around the predicted day t (similar days and predicted day t) as input, the output is: The predicted power of a distributed photovoltaic power station at time t within each grid.

[0039] The technical solution provided by this invention has the following beneficial effects:

[0040] This invention designs an adaptive grid partitioning method based on the Cartesian coordinate system and uses singular spectrum analysis to screen similar days; it constructs a feature set of a three-dimensional convolutional neural network based on Holt-Winters exponential smoothing; finally, it proposes an ultra-short-term power prediction method for distributed photovoltaic power station clusters based on a three-dimensional convolutional neural network, which improves the accuracy and robustness of the prediction effect. Attached Figure Description

[0041] The present invention will be further described below with reference to the accompanying drawings:

[0042] Figure 1 This is a flowchart of the present invention;

[0043] Figure 2 This is a diagram of a 3D convolutional neural network model.

[0044] Figure 3 Filter the charts based on the most similar days;

[0045] Figure 4 The simulation results are shown in the full-grid comprehensive similarity day screening diagram;

[0046] Figure 5 This is a comparison chart of simulation prediction results; Detailed Implementation Plan

[0047] To better understand the purpose, technical solution, and technical effects of this invention, the invention will be further explained below in conjunction with the accompanying drawings.

[0048] This invention proposes an ultra-short-term gridded prediction method for distributed photovoltaic power plants based on singular spectrum analysis. Figure 1 The flowchart of this invention includes the following detailed steps.

[0049] Step 1: Based on the adaptive mesh generation method, design a similarity day selection method incorporating singular spectrum analysis:

[0050] Adaptive mesh generation method:

[0051] A large region is divided into multiple smaller grids, each containing several distributed photovoltaic (PV) systems. Previous grid division methods directly used geographical factors for equal-interval partitioning, which could not guarantee the correlation of the distributed PV systems within each grid. This paper further discusses the correlation of distributed PV systems within each grid using a Cartesian geographic coordinate system, ensuring a high degree of correlation and laying the foundation for accurate prediction later. The specific steps of the adaptive grid division proposed in this paper are as follows:

[0052] 1) Place the distributed photovoltaic power station cluster in a certain area into a Cartesian coordinate system to obtain the horizontal axis length A and the vertical axis length B, and determine the initial step size of the horizontal axis division. Divide the step size along the vertical axis ,in , ;

[0053] 2) According to , In a two-dimensional plane, a distributed photovoltaic power station cluster is divided into grids. Within each grid, the two most distant distributed photovoltaic power stations are selected, and the cosine similarity of their power output at the same time is calculated. , Power curves of two photovoltaic power plants and The included angle, In the formula: This represents the actual power value at time t, where T represents the total number of times.

[0054] 3) For The grid is then further updated with the step size for the horizontal and vertical axes. , ,in , These represent the step sizes of the x and y axes after the m and nth iterations, respectively. , ;

[0055] 4) When all grid cosine similarity When the time comes, the partition ends;

[0056] The specific implementation steps of the similar day screening method involving singular spectrum analysis are as follows:

[0057] 1) First, obtain power data from the photovoltaic database, and then normalize the power of the photovoltaic power plants in each grid after division to obtain the average daily power curve of the photovoltaic power plants in each grid:

[0058] (12)

[0059] In the above formula, This represents the photovoltaic power station in the k-th grid. The average power at time t, K represents the number of grid cells after partitioning, and J is the number of distributed photovoltaic power stations contained in one grid. and Representing the first The daily power curve and installed capacity of a distributed photovoltaic power station at time t, where t is the time, T is the total number of time moments, and j is the photovoltaic power station number;

[0060] 2) Obtain the power curve within the predicted daily grid using the Holt-Winters exponential smoothing method. We then perform singular spectrum analysis on the samples to find the set of similar days that are closest to the predicted day characteristics.

[0061] Will Singular spectral decomposition is performed, and then the curve is obtained using the Dynamic Time Warping (DTW) algorithm. , Similarity between:

[0062] (13)

[0063] in: This indicates the similarity of the k-th grid curve; but This method can only represent the similarity of a single grid cell, and cannot represent the similarity of power curves across the entire region. This invention further uses a weighted method based on grid similarity to represent the similarity of regional power curves:

[0064] (14)

[0065] In the above formula, m+3 and n+3 represent the number of columns and rows of the grid, respectively; M represents the rated capacity of the k-th grid, and M represents the total rated capacity of the photovoltaic power plants in the entire region. This represents the daily power curve of the distributed photovoltaic system in the i-th row and r-th column and the characteristic distance of the most similar day.

[0066] Step 2: Construct the input feature set for the 3D convolutional network based on the Holt-Winters exponential smoothing method.

[0067] (15)

[0068] (16)

[0069] (17)

[0070] (18)

[0071] In the above formula, This represents the predicted value at time t+1. , , Do not correspond to the time-time level components, trend components, and seasonal components. All are smoothing coefficients. Let t be the actual value at time t, and u be the seasonal frequency. When predicting data for the next 24 hours, u = 24.

[0072] The final feature set is as follows:

[0073] (19)

[0074] In the above formula, To predict the power at L time points after time t.

[0075] Step 3: Based on the grid division and similar day selection results, and according to the input feature set, construct a three-dimensional convolutional neural network to establish a distributed photovoltaic power station cluster power ultra-short-term grid prediction model.

[0076] Since the region has been divided into grids, with the longitude and latitude of each region corresponding to the divided W rows and H columns, the input matrix of the 3D convolutional neural network is: Where C is the number of features, specifically including similar days within the grid. The system includes the level component, seasonal component, and photovoltaic power and irradiance data for the predicted day. W is the width of the matrix, H is the height of the matrix, and L represents continuous time intervals, used to characterize the temporal characteristics, as shown below:

[0077] (20)

[0078] (twenty one)

[0079] (twenty two)

[0080] This invention utilizes Taking data from L times around the predicted day t (similar days and predicted day t) as input, the output is: The predicted power of a distributed photovoltaic power station at time t within each grid.

[0081] To further understand this invention and to verify the accuracy and robustness of distributed photovoltaic (PV) forecasting, simulations were conducted. Simulations were performed based on PV data from a portion of northern my country, with each grid cell approximately 20 km × 15 km in size. 121 days were selected from the entire 2019 data set as test days, and the remainder as the training set.

[0082] Option 1: Use correlation coefficients to screen similar days and combine this with the Holt-Winters exponential smoothing method for prediction.

[0083] Option 2: Prediction based on Euclidean distance-based similarity day selection combined with a two-dimensional convolutional neural network.

[0084] Option 3: Use full-grid singular spectrum analysis to screen similar days, combine it with the Holt-Winters exponential smoothing method to construct a feature set, and then use a 3D convolutional neural network for prediction.

[0085] To more intuitively reflect the accuracy of different methods in selecting similar days, simulations were performed on the similar day selection methods of schemes 1, 2, and 3. The results are as follows: Figure 3 As shown in the diagram, taking a single grid as an example, Scheme 1 and Scheme 2 both result in the same most similar day, and their corresponding daily power curves are more similar to the original daily power curve in terms of "average" effect. However, Scheme 3 results in a daily power curve that is more similar to the original daily power curve in terms of local features, while also considering the overall similarity of the curves. The goal of similar day selection is to ensure that the final set of similar days contains daily power curves that reflect both the local and global features of the predicted daily power curve. Therefore, using singular spectrum analysis for similar day selection is more conducive to improving the accuracy of similar day selection.

[0086] To further illustrate the effectiveness of singular spectrum analysis in screening similar days across the entire region, simulation results are as follows: Figure 4 As shown, it is easy to see that similar daily power curves to the original daily power curve can be found in each grid cell across the entire region.

[0087] To demonstrate the predictive performance of the 3D convolutional neural network, simulations were performed on schemes 1, 2, and 3. The results are shown below. Figure 5 See Table 1. Taking the grid in the 3rd row and 1st column of the region as an example, comparing Prediction 1 and Prediction 2 shows that the accuracy of similar day screening significantly affects the prediction effect, and may even lead to the prediction results not converging. Comparing Prediction 2 and Prediction 3 also reveals that the addition of spatiotemporal feature information and 3D convolutional neural network is beneficial to improving the prediction effect.

[0088] Table 1. Daily Mean Absolute Error of Simulation Prediction Results under Different Schemes

[0089]

[0090] The prediction performance of Scheme 2 and Scheme 3 was evaluated using the daily mean absolute error (MAE). Table 1 shows the daily mean absolute error of the simulation prediction results at different grid locations. Based on the error data, Scheme 3 achieves a more stable prediction model compared to Scheme 2. Scheme 2, due to data variability and other factors, exhibits poorer prediction performance in certain grids. Overall, Scheme 3's prediction model demonstrates higher accuracy and greater robustness.

Claims

1. A method for ultra-short-term gridded prediction of distributed photovoltaic power station cluster power, characterized in that, Includes the following steps: (1) Based on the adaptive grid partitioning method, design a similar day screening method with singular spectrum analysis; (2) Construct the input feature set of the three-dimensional convolutional network based on the Holt-Winters exponential smoothing method; (3) Based on the grid division and similar day screening results, and according to the input feature set, a three-dimensional convolutional neural network is used to establish an ultra-short-term grid prediction model for distributed photovoltaic power stations; The adaptive mesh generation method and the similarity day selection method with singular spectrum analysis in step (1) are as follows: The specific steps for adaptive mesh generation are as follows: 1) Place the distributed photovoltaic power station cluster in a certain area into a Cartesian coordinate system to obtain the horizontal axis length A and the vertical axis length B, and determine the initial step size of the horizontal axis division. Divide the step size along the vertical axis ,in , ; 2) According to , In a two-dimensional plane, a distributed photovoltaic power station cluster is divided into grids. Within each grid, the two most distant distributed photovoltaic power stations are selected, and the cosine similarity of their power output at the same time is calculated. , Power curves of two photovoltaic power plants and The included angle, In the formula: This represents the actual power value at time t, where T represents the total number of times. 3) For The grid is then further updated with the step size for the horizontal and vertical axes. , ,in , Let x and y represent the step sizes of the x and y axes after the m-th and n-th iterations, respectively. , ; 4) When all grid cosine similarity When the time comes, the partition ends; The specific implementation steps of the similar day screening method involving singular spectrum analysis are as follows: 1) First, obtain power data from the photovoltaic database, and then normalize the power of the photovoltaic power plants in each grid after division to obtain the average daily power curve of the photovoltaic power plants in each grid: （1） In the above formula, This represents the photovoltaic power station in the k-th grid. The average power at time t, K represents the number of grid cells after partitioning, and J is the number of distributed photovoltaic power stations contained in one grid. and Representing the first The daily power curve and installed capacity of a distributed photovoltaic power station at time t, where t is the time, T is the total number of time moments, and j is the photovoltaic power station number; 2) Obtain the power curve within the predicted daily grid using the Holt-Winters exponential smoothing method. We then perform singular spectrum analysis on the samples to find the set of similar days that are closest to the predicted day characteristics. Will Singular spectral decomposition is performed, and then the curve is obtained using the Dynamic Time Warping (DTW) algorithm. , Similarity between: （2） in: This indicates the similarity of the k-th grid curve; but It can only represent the similarity of a single grid, not the similarity of the power curves over the entire region. Therefore, a weighted method based on grid similarity is used to represent the similarity of the power curves over the region. （3） In the above formula, m+3 and n+3 represent the number of columns and rows of the grid, respectively; M represents the rated capacity of the k-th grid, and M represents the total rated capacity of the photovoltaic power plants in the entire region. This represents the daily power curve of the distributed photovoltaic system in the i-th row and r-th column and the characteristic distance of the most similar day.

2. The method for ultra-short-term gridded prediction of distributed photovoltaic power station cluster power according to claim 1, characterized in that, In step (2), the Holt-Winters exponential smoothing method constructs the input feature set of the three-dimensional convolutional network as follows: （4） （5） （6） （7） In the above formula, This represents the predicted value at time t+1. , , Do not correspond to the time-time level components, trend components, and seasonal components. All are smoothing coefficients. Let t be the actual value at time t, and u be the seasonal frequency. When predicting data for the next 24 hours, u = 24. The final feature set is as follows: （8） In the above formula, To predict the power at L time points after time t.

3. The method for ultra-short-term gridded prediction of distributed photovoltaic power station cluster power according to claim 1, characterized in that, In step (3), the three-dimensional convolutional neural network is used to establish an ultra-short-term grid prediction model for distributed photovoltaic power plants as follows: Since the region has been divided into grids, with the longitude and latitude of each region corresponding to the divided W rows and H columns, the input matrix of the 3D convolutional neural network is: Where C is the number of features, specifically including similar days within the grid. The system includes the level component, seasonal component, and photovoltaic power and irradiance data for the predicted day. W is the width of the matrix, H is the height of the matrix, and L represents continuous time intervals, used to characterize the temporal characteristics, as shown below: （9） （10） （11） use Taking data from L times around the predicted day t (similar days and predicted day t) as input, the output is: The predicted power of a distributed photovoltaic power station at time t within each grid.

Images