A photovoltaic power prediction method based on recent clear-sky day decomposition theory and temporal convolutional networks

Abstract

This invention discloses a photovoltaic power prediction method based on the recent clear-sky day decomposition theory and a temporal convolutional network, belonging to the field of photovoltaic power in distribution networks. The method includes the following steps: calculating the kurtosis and skewness of daily power data in the original dataset to determine similarity day screening indicators; wherein, the original dataset includes: original photovoltaic power data and historical meteorological data; obtaining a clear-sky day dataset based on the similarity day screening indicators and the meteorological data of the prediction day; inputting the clear-sky day dataset into a fully connected neural network to obtain an ideal clear-sky photovoltaic power prediction result; obtaining a fluctuation dataset based on the ideal clear-sky photovoltaic power prediction result and the original photovoltaic power data; inputting the fluctuation dataset into a temporal convolutional neural network to obtain the photovoltaic power fluctuation result of the prediction day; and obtaining the final photovoltaic power prediction result based on the ideal clear-sky photovoltaic power prediction result and the photovoltaic power fluctuation result of the prediction day.

Description

Technical Field

[0001] This invention belongs to the field of photovoltaic power technology in power distribution networks, and particularly relates to a photovoltaic power prediction method based on the recent clear-sky day decomposition theory and time convolutional network. Background Technology

[0002] With the continuous development of photovoltaic (PV) energy, many new PV power generation systems need to be connected to the grid every year. However, the volatility and instability of PV power generation pose significant challenges to grid dispatch and distribution. The instability of PV power generation mainly stems from weather changes. Different weather factors affect solar irradiance and the temperature of PV panels, indirectly impacting PV power output and increasing the uncertainty of grid operation. Therefore, accurate PV power forecasting, especially methods capable of handling complex weather scenarios, is crucial for optimizing grid dispatch and improving grid operational stability. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention proposes a photovoltaic power prediction method based on the recent clear-sky day decomposition theory and temporal convolutional networks, thereby resolving the issues present in the existing technologies.

[0004] To achieve the above objectives, this invention provides a photovoltaic power prediction method based on the recent clear-sky day decomposition theory and temporal convolutional networks, comprising:

[0005] The kurtosis and skewness of daily power data in the original dataset are calculated to determine the similarity day screening index; wherein, the original dataset includes: original photovoltaic power data and historical meteorological data;

[0006] A dataset of clear-sky days was obtained based on the similarity day screening index and the meteorological data of the predicted days.

[0007] The clear-sky day dataset is input into a fully connected neural network to obtain the ideal clear-sky photovoltaic power prediction result;

[0008] A fluctuation dataset is obtained based on the ideal clear-sky photovoltaic power prediction results and the original photovoltaic power data.

[0009] The fluctuation dataset is input into a temporal convolutional neural network to obtain the predicted daily photovoltaic power fluctuation results;

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

[0011] Optionally, before determining the screening indicators for similar days, the process may also include: using the Pearson coefficient method based on the original photovoltaic power data and historical meteorological data to select meteorological factors with a high correlation to photovoltaic power to obtain meteorological characteristics.

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

[0013] Use the sunny day closest to the predicted date as the reference date;

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

[0015] The dataset of clear-sky days was filtered based on the kurtosis and skewness of the original dataset and the reference date.

[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 in the fully connected neural network is the input of the previous fully connected layer and the residual of the backtracking power prediction result.

[0018] Optionally, the expression for calculating the fluctuation power in the fluctuation dataset is:

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

[0020] In the formula, W i X is the fluctuating power on day i. i S is the raw power of day i, S is the power of the ideal nearest clear sky day, and j represents the j-th data point of a day.

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

[0022] Optionally, the calculation expression for the causal dilated convolution is:

[0023]

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

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

[0026] This invention combines recent clear-sky day decomposition theory and Temporal Convolutional Networks (TCNs) to improve the accuracy of photovoltaic (PV) power prediction. By calculating the kurtosis and skewness of daily power data in the original dataset, similarity day screening criteria are determined, thereby selecting clear-sky day datasets. This step identifies historical data with similar meteorological conditions to the predicted day, providing accurate references for subsequent predictions. Next, a fully connected neural network is used to process the clear-sky day dataset to obtain ideal clear-sky PV power prediction results. Then, through analysis of the fluctuation dataset, the TCN can capture short-term fluctuations in PV power, further refining the prediction results. Finally, the ideal clear-sky PV power prediction results are combined with the PV power fluctuation results to obtain a more accurate final PV power prediction result. This method effectively addresses the uncertainties and fluctuations in PV power prediction, providing strong technical support for the operation and scheduling of PV power plants. Attached Figure Description

[0027] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

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

[0029] Figure 2 This is a fully connected neural network diagram of the dual residual structure in an embodiment of the present invention;

[0030] Figure 3 This is a diagram of a time-convolutional neural network with a dual residual structure according to an embodiment of the present invention. Detailed Implementation

[0031] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

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

[0033] Example 1

[0034] The purpose of this invention is to provide a photovoltaic power prediction method based on the recent clear-sky day decomposition theory and temporal convolutional networks. Based on existing experimental datasets, deep learning is used to predict photovoltaic power under complex weather scenarios. A clear-sky dataset is established based on recent clear-sky day data to predict photovoltaic power under ideal clear-sky conditions. The difference between the original power data and the ideal clear-sky power data is calculated to obtain fluctuating power data, allowing meteorological characteristics to be mapped more directly to fluctuating characteristics, thus obtaining the fluctuating component of the prediction day and improving 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 temporal convolutional networks, including the following steps:

[0036] The Pearson coefficient method was used to calculate the Pearson coefficients between various meteorological factors and photovoltaic power in the dataset, and meteorological features were selected. The kurtosis and skewness of the daily power data in the original dataset were calculated as indicators for screening similar days. The clear day closest to the prediction date was selected, and data in the original dataset that differed from it within a certain range based on the power kurtosis and skewness of that day were selected as the clear sky day dataset. A fully connected neural network based on a dual residual structure was established, and the photovoltaic power prediction model was trained using the clear sky day dataset as the training set. The photovoltaic power prediction result under ideal clear sky conditions was obtained through the meteorological data of the prediction date. The photovoltaic power prediction result under ideal clear sky conditions was subtracted from the power data in the original dataset, and the difference between the meteorological data and the power was combined to obtain the fluctuation dataset. A temporal convolutional neural network based on a dual residual structure was established, and the photovoltaic power prediction model was trained using the fluctuation dataset as the training set. The photovoltaic power fluctuation result on the prediction date was obtained through the meteorological data of the prediction date. The photovoltaic power prediction result under ideal clear sky conditions and the photovoltaic power fluctuation result on the prediction date were synthesized to obtain the final prediction result.

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

[0038] Step 1: Use the Pearson coefficient method to select meteorological factors that are highly correlated with photovoltaic power.

[0039] Based on the existing dataset, the Pearson coefficient between each meteorological data point and the power data is calculated. Meteorological factors with larger Pearson coefficients are selected as input meteorological features for subsequent methods. The formula for calculating the Pearson coefficient is as follows:

[0040]

[0041] Among them, X i This is the i-th photovoltaic power data. It is the average of all photovoltaic power data included in the calculation; Y i This is the i-th meteorological data point. 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 the daily photovoltaic power in the original dataset as indicators for screening clear-sky similar days. The original dataset includes: original photovoltaic power data and historical meteorological data.

[0043] The daily output curve of photovoltaic power generation is an approximately normally distributed parabola. Skewness reflects the skewness of the peak of the normally distributed waveform, while kurtosis reflects the sharpness of the peak. The skewness and kurtosis of photovoltaic output curves on similar clear-sky days are similar, and therefore can be used as screening indicators for similar days. The formulas for calculating 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 dataset based on similarity day screening indicators and meteorological data of the forecast day. Select the clear sky day closest to the forecast day as the reference day, and filter the data in the original data whose skewness and kurtosis differ from the reference day within a certain range to form the clear sky day dataset.

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

[0048] Fully connected neural networks with dual residual structures, such as Figure 2 As shown, each fully connected layer outputs both backtracking power prediction and future power prediction. Both predictions are obtained through the currently trained neural network weights. The difference is that the input for backtracking prediction is meteorological data from the past few days, while the input for future prediction is the meteorological data of the current day. Except for the first layer, the input for each subsequent fully connected layer is the residual between the input of the previous fully connected layer and the backtracking power prediction. This neural network structure makes the training features of each layer more localized, which can promote smoother gradient backpropagation.

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

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

[0051] Among them W i X is the fluctuating power on day i. i S is the raw power of day i, S is the power of the ideal nearest clear sky day, and j represents the j-th data point of a day.

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

[0053] Temporal convolutional neural networks with dual residual structures, such as Figure 3 As shown, the dual residual structure is consistent with the fully connected neural network with dual residual structure, the difference being the temporal convolutional layer in each layer; the temporal convolutional network introduces causal dilated convolution, enabling the network to possess both the multi-channel feature extraction capability of convolutional neural networks and the temporal information mining capability of temporal neural networks. The calculation formula for causal dilated convolution is as follows:

[0054]

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

[0056] Step 7: Combine the photovoltaic power prediction results under ideal clear sky conditions with the photovoltaic power fluctuation results on the predicted day to obtain the final prediction result.

[0057] The final photovoltaic power prediction result is obtained by superimposing the predicted fluctuating power component with the ideal clear-sky power component.

[0058] This invention does not divide historical data into a few simple and typical weather scenarios, but can build a prediction model for any weather condition, thus having better universality. This invention establishes a similar day dataset for clear sky conditions, avoiding overfitting caused by too few similar samples when searching for similar days for prediction days under unconventional weather conditions, thereby improving the stability of the model. This invention establishes a mapping relationship between meteorological features and fluctuation components, allowing the neural network to focus more directly on the fluctuation changes in photovoltaic power generation, thereby improving the prediction accuracy of the model under complex weather scenarios.

[0059] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A photovoltaic power prediction method based on the recent clear-sky day decomposition theory and temporal convolutional networks, characterized in that, Includes the following steps: The kurtosis and skewness of daily power data in the original dataset are calculated to determine the similarity day screening index; wherein, the original dataset includes: original photovoltaic power data and historical meteorological data; A dataset of clear-sky days was obtained based on the similarity day screening index and the meteorological data of the predicted days. The process of obtaining a dataset of clear-sky days includes: Use the sunny day closest to the predicted date as the reference date; Calculate the kurtosis and skewness of the reference day; Clear-sky day datasets were filtered based on the kurtosis and skewness of the original dataset and the reference day; The clear-sky day dataset is input into a fully connected neural network to obtain the ideal clear-sky photovoltaic power prediction result; Both the fully connected neural network and the temporal convolutional neural network have dual residual structures. Each fully connected layer of the fully connected neural network outputs both backtracking power prediction results and future power prediction results. The input for backtracking prediction is past meteorological data, and the input for future prediction is the meteorological data of the current day. Except for the first fully connected layer, the input for each subsequent fully connected layer is the residual between the input of the previous fully connected layer and the backtracking power prediction result. Each temporal convolutional layer in the aforementioned temporal convolutional neural network introduces a causal dilated convolution; the calculation expression for the causal dilated convolution is: ; In the formula, This is the convolution output at time step t, where k is the convolution kernel size. These are the weights of the convolution kernel. The input sequence at time step The value, It is the expansion factor; Based on the ideal clear-sky photovoltaic power prediction results and the original photovoltaic power data, a fluctuation dataset is obtained; the calculation expression for the fluctuation power in the fluctuation dataset is: ; In the formula, It is the first Fluctuating power of the sky It is the first The original power of the day, This is the power output of the ideal, recent clear-sky day. Indicates the first day of the day One data point; The fluctuation dataset is input into a temporal convolutional neural network to obtain the predicted daily photovoltaic power fluctuation results; The final photovoltaic power prediction result is obtained based on the ideal clear sky photovoltaic power prediction result and the photovoltaic power fluctuation result of the prediction day.

2. The photovoltaic power prediction method based on the recent clear-sky day decomposition theory and temporal convolutional network according to claim 1, characterized in that, Before determining the screening indicators for similar days, the following steps are also taken: based on the original photovoltaic power data and historical meteorological data, the Pearson coefficient method is used to select meteorological factors with high correlation to photovoltaic power to obtain meteorological characteristics.

Images