Photovoltaic power prediction method based on recent clear sky day decomposition theory and time convolutional network

Abstract

The invention discloses a photovoltaic power prediction method based on a recent clear sky day decomposition theory and a time convolutional network, and belongs to the field of power distribution network photovoltaic power, and the method comprises the following steps: calculating the kurtosis and skewness of daily power data in an original data set, and determining a similar day screening index; wherein the original data set comprises original photovoltaic power data and historical meteorological data; obtaining a clear sky day data set based on the similar day screening index and the meteorological data of the prediction day; inputting the clear sky day data set into a full-connection neural network to obtain an ideal clear sky photovoltaic power prediction result; obtaining a fluctuation data set based on the ideal clear sky photovoltaic power prediction result and the original photovoltaic power data; inputting the fluctuation data set into a time convolutional neural network to obtain a solar photovoltaic power fluctuation prediction result; and obtaining a final photovoltaic power prediction result based on the ideal clear sky photovoltaic power prediction result and the photovoltaic power fluctuation prediction result.

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic power of distribution network, and in particular relates to a photovoltaic power prediction method based on the recent clear sky day decomposition theory and a time convolution network. Background Art

[0002] With the continuous development of photovoltaic energy, many new photovoltaic power generation systems need to be connected to the power grid every year. However, the volatility and instability of photovoltaic power generation have brought huge challenges to the grid dispatching and distribution. The instability of photovoltaic power generation mainly comes from weather changes. Different weather factors affect the solar irradiance and the temperature of photovoltaic panels, which in turn indirectly affects the photovoltaic power generation power, resulting in increased uncertainty in grid operation. Therefore, accurate photovoltaic power prediction, especially photovoltaic power prediction methods that can cope with complex weather scenarios, is crucial to optimizing grid dispatching and improving grid operation stability. Summary of the invention

[0003] In order to solve the above technical problems, the present invention proposes a photovoltaic power prediction method based on the nearest clear sky day decomposition theory and a temporal convolutional network to solve the problems existing in the above-mentioned prior art.

[0004] To achieve the above object, the present invention provides a photovoltaic power prediction method based on the recent clear sky day decomposition theory and a time convolutional network, comprising:

[0005] Calculate the kurtosis and skewness of daily power data in the original data set to determine the similar day screening index; wherein the original data set includes: original photovoltaic power data and historical meteorological data;

[0006] Obtaining a clear sky day data set based on the similar day screening index and the meteorological data of the forecast day;

[0007] Inputting the clear sky day data set into a fully connected neural network to obtain an ideal clear sky photovoltaic power prediction result;

[0008] Obtaining a fluctuation data set based on the ideal clear sky photovoltaic power prediction result and the original photovoltaic power data;

[0009] Inputting the fluctuation data set into a time convolutional neural network to obtain a predicted daily photovoltaic power fluctuation result;

[0010] The final photovoltaic power prediction result is obtained based on the ideal clear sky photovoltaic power prediction result and the predicted daily photovoltaic power fluctuation result.

[0011] Optionally, before determining the similar day screening index, the method further includes: based on the original photovoltaic power data and the historical meteorological data, using the Pearson coefficient method to select meteorological factors with a high correlation with the photovoltaic power to obtain meteorological characteristics.

[0012] Optionally, the process of obtaining a clear sky day dataset includes:

[0013] The sunny day closest to the forecast day is taken as the reference day;

[0014] Calculating the kurtosis and skewness of the reference day;

[0015] The clear sky day data set is screened based on the kurtosis and skewness of the original data set and the reference day.

[0016] Optionally, both the fully connected neural network and the temporal convolutional neural network are dual residual structures.

[0017] Optionally, except for the first fully connected layer, the input of each subsequent fully connected layer of the fully connected neural network is the residual of the input of the previous fully connected layer and the backtracking power prediction result.

[0018] Optionally, the calculation expression of the fluctuation power in the fluctuation data set is:

[0019] W i j =X i j -S j

[0020] Where W i is the fluctuating power on the ith day, X i is the original power of the i-th day, S is the power of the ideal nearest clear sky day, and j represents the j-th data of the day.

[0021] Optionally, each temporal convolutional layer in the temporal convolutional neural network introduces causal dilated convolution.

[0022] Optionally, the calculation expression of the causal dilation convolution is:

[0023]

[0024] In the formula, y t is the convolution output at time step t, k is the convolution kernel size, and w i is the weight of the convolution kernel, x t-i·d is the value of the input sequence at time step ti·d, where d is the dilation factor.

[0025] Compared with the prior art, the present invention has the following advantages and technical effects:

[0026] The present invention combines the recent clear sky day decomposition theory and the temporal convolutional network (TCN) to improve the accuracy of photovoltaic power prediction. By calculating the kurtosis and skewness of the daily power data in the original data set, the similar day screening index is determined, thereby screening out the clear sky day data set. This step can identify historical data with similar meteorological conditions to the predicted day, providing an accurate reference for subsequent predictions. Next, the clear sky day data set is processed using a fully connected neural network to obtain the ideal clear sky photovoltaic power prediction result. Then, through the analysis of the fluctuation data set, the temporal convolutional network can capture the short-term fluctuations of photovoltaic power and further refine the prediction results. Finally, the ideal clear sky photovoltaic power prediction result is combined with the photovoltaic power fluctuation result to obtain a more accurate final photovoltaic power prediction result. This method can effectively deal with the uncertainty and volatility in photovoltaic power prediction, and provide strong technical support for the operation and scheduling of photovoltaic power stations. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0028] Figure 1 This is a flow chart of a photovoltaic power prediction method based on the recent clear sky day decomposition theory and a temporal convolutional network according to an embodiment of the present invention;

[0029] Figure 2 A fully connected neural network diagram of a dual residual structure according to an embodiment of the present invention;

[0030] Figure 3 This is a temporal convolutional neural network diagram of a dual residual structure according to an embodiment of the present invention. DETAILED DESCRIPTION

[0031] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0032] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0033] Embodiment 1

[0034] The purpose of the present invention is to provide a photovoltaic power prediction method based on the recent clear sky day decomposition theory and temporal convolutional network. According to the existing experimental data set, the photovoltaic power in complex weather scenes is predicted by a deep learning method; a clear sky data set is established based on the recent clear sky day data, and the photovoltaic power under ideal clear sky conditions is predicted. The original power data is subtracted from the ideal clear sky power data to obtain the fluctuating power data, so that the meteorological characteristics are more directly mapped to the fluctuating characteristics, and the fluctuating components of the predicted day are obtained, thereby providing the sensitivity of the prediction method to meteorological factors; finally, the ideal clear sky component and the fluctuating component are synthesized to obtain the photovoltaic power prediction result.

[0035] like Figure 1 As shown, this embodiment provides a photovoltaic power prediction method based on the recent clear sky day decomposition theory and a time convolution network, including the following steps:

[0036] The Pearson coefficient method is used to calculate the Pearson coefficient between each meteorological factor and photovoltaic power in the data set, and meteorological features are selected; the kurtosis and skewness of the daily power data in the original data set are calculated as indicators for screening similar days; the sunny day closest to the prediction day is selected, and according to the power kurtosis and skewness of that day, the data in the original data set that differ from it within a certain range are screened out as the clear sky day data set; a fully connected neural network based on a double residual structure is established, and the clear sky day data set is used as a training set to train the photovoltaic power prediction model, and the ideal clear sky photovoltaic power prediction result is obtained through the meteorological data on the prediction day; the power data in the original data set is subtracted from the ideal clear sky photovoltaic power prediction result, and the meteorological data and power difference are combined to obtain the fluctuation data set; a temporal convolutional neural network based on a double residual structure is established, and the fluctuation data set is used as a training set to train the photovoltaic power prediction model, and the photovoltaic power fluctuation result on the prediction day is obtained through the meteorological data on the prediction day; the ideal clear sky photovoltaic power prediction result and the photovoltaic power fluctuation result on the prediction day are synthesized to obtain the final prediction result.

[0037] As a specific implementation of this embodiment, it includes:

[0038] Step 1: Use the Pearson coefficient method to select meteorological factors with a high correlation with photovoltaic power.

[0039] Based on the existing data set, the Pearson coefficient between each meteorological data and power data is calculated, and the meteorological factor with a larger Pearson coefficient is selected as the input meteorological feature of the subsequent method; the calculation formula of the Pearson coefficient method is as follows:

[0040]

[0041] Among them, X i is the ith photovoltaic power data, is the average value of all photovoltaic power data included in the calculation; Y i is the i-th meteorological data, is the average of all meteorological data included in the calculation; r is the Pearson coefficient.

[0042] Step 2: Calculate the skewness and kurtosis of daily PV power of the original data set as the index for screening clear sky similar days. The original data set includes: original PV power data and historical meteorological data.

[0043] The daily output curve of photovoltaic power generation is a parabola that approximates normal distribution, and the skewness can reflect the deviation of the peak of the normal distribution waveform, and the kurtosis can reflect the sharpness of the peak of the normal distribution waveform. The skewness and kurtosis of the photovoltaic output curve on a similar clear day are similar, so it can be used as a screening indicator for similar days; the calculation formulas for skewness and kurtosis are as follows:

[0044]

[0045] Where x is the sample, μ is the mean, σ is the standard deviation, and E is the mean operation.

[0046] Step 3: Obtain the clear sky day data set based on the similar day screening index and the meteorological data of the forecast day. Select the sunny day closest to the forecast day as the reference day, and screen the data whose skewness and kurtosis in the original data are within a certain range from the reference day to form the clear sky day data set.

[0047] Step 4: Establish a fully connected neural network based on a double residual structure, use the clear sky day dataset as the training set, train the photovoltaic power prediction model, and obtain the photovoltaic power prediction results under ideal clear sky conditions through the meteorological data of the prediction day.

[0048] The fully connected neural network with dual residual structure is as follows Figure 2 As shown in the figure, each fully connected layer outputs both the retrospective power prediction result and the future power prediction result; both prediction results are obtained through the currently trained neural network weight prediction, the difference is that the input of the retrospective prediction is the meteorological data of the past few days, and the input of the future prediction is the meteorological data of the current day; except for the first layer, the input of each subsequent fully connected layer is the residual of the input of the previous fully connected layer and the retrospective power prediction result; such a neural network structure makes the training features of each layer more localized, which can promote smoother gradient back propagation.

[0049] Step 5: Obtain a fluctuating data set based on the ideal clear sky PV power prediction results and the original PV power data. Calculate the difference between the original power and the ideal clear sky power, and combine it with the meteorological data to form a fluctuating data set; the fluctuating power calculation formula is as follows:

[0050] W ij =X i j -S j (4)

[0051] Where W i is the fluctuating power on the ith day, X i is the original power of the i-th day, S is the power of the ideal nearest clear sky day, and j represents the j-th data of the day.

[0052] Step 6: Establish a temporal convolutional neural network based on a dual residual structure, use the fluctuation data set as the training set, train the photovoltaic power prediction model, and obtain the photovoltaic power fluctuation results on the predicted day through the meteorological data on the predicted day.

[0053] The temporal convolutional neural network with dual residual structure is as follows Figure 3 As shown in the figure, the dual residual structure is consistent with the fully connected neural network of the dual residual structure. The difference lies in the temporal convolution layer of each layer. The temporal convolutional network introduces causal dilation convolution, which enables the network to have both the multi-channel feature extraction capability of the convolutional neural network and the temporal information mining capability of the temporal neural network. The calculation formula of causal dilation convolution is as follows:

[0054]

[0055] where y t is the convolution output at time step t, k is the convolution kernel size, and w i is the weight of the convolution kernel, x t-i·d is the value of the input sequence at time step ti·d, where d is the dilation factor.

[0056] Step 7: Synthesize the PV power prediction results under ideal clear sky conditions and the predicted daily PV power fluctuation results to obtain the final prediction results.

[0057] The predicted fluctuating power component is superimposed on the ideal clear sky power component to obtain the final photovoltaic power prediction result.

[0058] The present invention does not divide historical data into several simple typical weather scenes, but can establish a prediction model for any weather conditions, which has better universality; the present invention establishes a similar day data set for clear sky conditions, avoiding overfitting caused by too few similar samples when looking for similar days of predicted days under unconventional weather conditions, thereby improving the stability of the model; the present invention establishes a mapping relationship between meteorological characteristics and fluctuation components, allowing the neural network to pay more direct attention to the fluctuation changes in photovoltaic power generation, thereby improving the prediction accuracy of the model in complex weather scenarios.

[0059] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A photovoltaic power prediction method based on the recent clear sky day decomposition theory and time convolution network, characterized in that: The following steps are involved: Calculate the kurtosis and skewness of daily power data in the original data set to determine the similar day screening index; wherein the original data set includes: original photovoltaic power data and historical meteorological data; Obtaining a clear sky day data set based on the similar day screening index and the meteorological data of the forecast day; Inputting the clear sky day data set into a fully connected neural network to obtain an ideal clear sky photovoltaic power prediction result; Obtaining a fluctuation data set based on the ideal clear sky photovoltaic power prediction result and the original photovoltaic power data; Inputting the fluctuation data set into a time convolutional neural network to obtain a predicted daily photovoltaic power fluctuation result; The final photovoltaic power prediction result is obtained based on the ideal clear sky photovoltaic power prediction result and the predicted daily photovoltaic power fluctuation result.

2. The photovoltaic power prediction method based on the recent clear sky day decomposition theory and time convolution network according to claim 1 is characterized in that: Before determining the similar day screening index, it also includes: based on the original photovoltaic power data and historical meteorological data, the Pearson coefficient method is used to select meteorological factors with a high correlation with photovoltaic power to obtain meteorological characteristics.

3. The photovoltaic power prediction method based on the recent clear sky day decomposition theory and time convolution network according to claim 1 is characterized in that: The process of obtaining the clear sky day dataset includes: The sunny day closest to the forecast day is taken as the reference day; Calculating the kurtosis and skewness of the reference day; The clear sky day data set is screened based on the kurtosis and skewness of the original data set and the reference day.

4. The photovoltaic power prediction method based on the recent clear sky day decomposition theory and time convolution network according to claim 1 is characterized in that: The fully connected neural network and the temporal convolutional neural network are both dual residual structures.

5. The photovoltaic power prediction method based on the recent clear sky day decomposition theory and time convolution network according to claim 4 is characterized in that: Except for the first fully connected layer, the input of each subsequent fully connected layer of the fully connected neural network is the residual of the input of the previous fully connected layer and the backtracking power prediction result.

6. The photovoltaic power prediction method based on the recent clear sky day decomposition theory and time convolution network according to claim 1 is characterized in that: The calculation expression of the fluctuation power in the fluctuation data set is: Where W i is the fluctuating power on the ith day, X i is the original power of the i-th day, S is the power of the ideal nearest clear sky day, and j represents the j-th data of the day.

7. The photovoltaic power prediction method based on the recent clear sky day decomposition theory and time convolution network according to claim 1 is characterized in that: Each temporal convolution layer in the temporal convolutional neural network introduces causal dilated convolution.

8. The photovoltaic power prediction method based on the recent clear sky day decomposition theory and time convolution network according to claim 7 is characterized in that: The computational expression of the causal dilation convolution is: In the formula, y t is the convolution output at time step t, k is the convolution kernel size, and w i is the weight of the convolution kernel, x t-i· d is the value of the input sequence at time step ti·d, where d is the dilation factor.

Images