Photovoltaic power generation power fluctuation multi-parameter forecasting method based on multi-task one-dimensional convolutional network

Abstract

The invention discloses a photovoltaic power generation power fluctuation multi-parameter forecasting method based on a multi-task one-dimensional convolutional network, and belongs to the technical field of electric power system control, and the method comprises the steps: firstly, carrying out the selection of forecasting parameters, and selecting a photovoltaic power generation power fluctuation standard deviation and fluctuation variation as the direct description of the photovoltaic power generation power fluctuation range and fluctuation rate; then, carrying out ACF, PACF and MIC time delay correlation analysis on each parameter to carry out predictability analysis on multiple parameters of photovoltaic power generation; and finally, designing a multi-task one-dimensional convolutional neural network to carry out photovoltaic power generation power fluctuation multi-parameter forecasting. According to the method, the photovoltaic power generation power fluctuation standard deviation and fluctuation variation with clear physical significance and good statistical characteristics are introduced to serve as supplements of a traditional power value, the prediction precision of parameters and time steps can be improved through the multi-task one-dimensional convolution network, photovoltaic technology development is promoted, popularization and application of renewable energy sources are promoted, and the method has a wide application prospect. And efficient consumption of photovoltaic power generation is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system control, and in particular relates to a multi-parameter prediction method for photovoltaic power generation power fluctuation based on a multi-task one-dimensional convolutional network. Background Art

[0002] In the context of large-scale photovoltaic power generation, the power grid needs to reasonably arrange the capacity of complementary power sources, reserve sufficient flexibility, ensure that the system regulation speed can quickly respond to changes in wind power, and achieve matching of regulation range and regulation rate to promote efficient consumption of photovoltaic power generation. However, the current photovoltaic power generation forecast objects are basically concentrated on solar radiation and photovoltaic power generation, and there are actually few studies on the forecast of direct characterization indicators of conventional state fluctuation range and fluctuation rate.

[0003] Therefore, the prior art urgently needs a new technical solution to solve the above problems. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a multi-parameter prediction method for photovoltaic power generation power fluctuation based on a multi-task one-dimensional convolutional network. By selecting appropriate photovoltaic power generation fluctuation measurement parameters, analyzing their predictability and mutual coherence, and formulating suitable deterministic forecasting strategies, the traditional single-parameter forecasting mode is expanded to a multi-parameter integrated forecasting mode, thereby providing more diversified information for the peak and frequency regulation of the power grid after a high proportion of photovoltaic power generation is connected.

[0005] The multi-parameter prediction method of photovoltaic power generation fluctuation based on multi-task one-dimensional convolutional network includes the following steps:

[0006] Step 1: Select the standard deviation of photovoltaic power generation fluctuation as the photovoltaic power generation power fluctuation range parameter, and select the photovoltaic power generation power fluctuation variation as the parameter for measuring the photovoltaic power generation power fluctuation rate;

[0007] Step 2: Perform time-delay correlation analysis of the autocorrelation function ACF, partial autocorrelation function PACF and maximum information coefficient MIC on the parameters described in step 1, and perform predictability analysis on multi-parameters of photovoltaic power generation fluctuations;

[0008] Step 3: Establish a multi-task one-dimensional convolutional neural network to predict the multi-parameter fluctuations of photovoltaic power generation described in step 2.

[0009] The autocorrelation function ACF described in step 2 is used to measure the correlation of the time series itself at different lags, and the similarity between the values ​​at each moment in the sequence is displayed by calculating the autocorrelation coefficients of different lag steps:

[0010]

[0011] In the formula, k represents the lag order, X t and X t+k are the observation values ​​at time t and t+k respectively, T is the total length of the sequence, is the mean of the time series, defined as:

[0012] The partial autocorrelation function PACF further considers the direct relationship between each lag step on the basis of ACF, excluding the influence of intermediate lags, and the formula is:

[0013]

[0014] In the formula, φ k,k is the partial autocorrelation coefficient of the sequence at lag k, γ k is the value of the autocorrelation function ACF at lag k, φ k-1,j is the partial autocorrelation coefficient of lag j when the order is k-1;

[0015] The maximum information coefficient MIC is used to detect the nonlinear relationship between two variables, and the formula is:

[0016]

[0017] Where x and y are the sizes of the partitioned grids, and I(X;Y) is the mutual information.

[0018] The multi-task one-dimensional convolutional neural network described in step three is used for time series data modeling, including a convolutional layer, a pooling layer and a fully connected layer. The convolutional layer slides multiple convolution kernels on the layer input to perform corresponding element products and sum operations, and projects the information in the receptive field to generate a set of parallel feature maps; the pooling layer performs nonlinear downsampling on the feature map generated by the convolutional layer; after downsampling, the feature set undergoes a layer of convolution operation and a fully connected operation to obtain the model output.

[0019] Through the above-mentioned design scheme, the present invention can bring the following beneficial effects: a multi-parameter prediction method for photovoltaic power generation power volatility based on a multi-task one-dimensional convolutional network is used to study the direct characterization indicator prediction of the conventional normal state fluctuation range and fluctuation rate, and provide more diversified information for the peak and frequency regulation of the power grid after a high proportion of photovoltaic power generation is connected.

[0020] Furthermore, the present invention introduces the standard deviation and fluctuation variation of photovoltaic power generation with clear physical meaning and good statistical characteristics as a supplement to the traditional power value, directly describes the volatility of photovoltaic power generation, and studies the predictability of each parameter, proving the predictability of the selected parameters within a certain time scale range. At the same time, the multi-task 1dCNN model designed by the present invention is more excellent in the prediction performance of each parameter and each time step compared with the single-task ANN, single-task LSTM, and single-task 1dCNN models. The present invention can improve the accuracy of photovoltaic power generation forecasting, and provides a more direct description of photovoltaic power generation volatility, which promotes the efficient consumption of photovoltaic power generation.

[0021] The present invention was verified on a simulation platform by constructing specific working conditions. It was found that the multi-parameter prediction method of photovoltaic power generation fluctuation based on a multi-task one-dimensional convolutional network can effectively improve the prediction performance of each parameter and each time step, which is helpful to promote the development of photovoltaic technology and the popularization and application of renewable energy. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The present invention is further described below with reference to the accompanying drawings and specific embodiments:

[0023] Figure 1 The present invention is a flow chart of a multi-parameter prediction method for photovoltaic power generation fluctuation based on a multi-task one-dimensional convolutional network.

[0024] Figure 2 Schematic diagram of the multi-task learning mechanism of the multi-parameter prediction method of photovoltaic power generation fluctuation based on the multi-task one-dimensional convolutional network of the present invention.

[0025] Figure 3 This is a schematic diagram of the one-dimensional convolution structure of the multi-parameter prediction method for photovoltaic power generation fluctuation based on a multi-task one-dimensional convolutional network of the present invention.

[0026] Figure 4 This is a schematic diagram of the multi-task one-dimensional convolution structure of the multi-parameter prediction method for photovoltaic power generation fluctuation based on the multi-task one-dimensional convolutional network of the present invention.

[0027] Figure 5 This is a schematic diagram of the high-frequency photovoltaic power generation power multi-parameter sliding window downscaling process of the specific implementation method of the photovoltaic power generation power volatility multi-parameter prediction method based on the multi-task one-dimensional convolutional network of the present invention.

[0028] Figure 6 This is a statistical diagram of the ACF, PACF, and MIC correlation coefficients of the average photovoltaic power generation power P according to a specific implementation method of the photovoltaic power generation power fluctuation multi-parameter prediction method based on a multi-task one-dimensional convolutional network of the present invention.

[0029] Figure 7This is a statistical diagram of the ACF, PACF, and MIC correlation coefficients of the photovoltaic power fluctuation standard deviation σ of the photovoltaic power generation power fluctuation specific implementation method of the present invention.

[0030] Figure 8 This is a statistical diagram of the ACF, PACF, and MIC correlation coefficients of the square root of the photovoltaic power fluctuation variation γ according to the specific implementation method of the photovoltaic power generation power fluctuation multi-parameter prediction method based on the multi-task one-dimensional convolutional network of the present invention.

[0031] Fig. 9 It is a schematic diagram of the comparison of the RMSE of the multi-parameter forecasting of {P, σ, γ} of the specific implementation method of the photovoltaic power fluctuation multi-parameter forecasting method based on the multi-task one-dimensional convolutional network of the present invention. DETAILED DESCRIPTION

[0032] In order to better understand the purpose, structure and function of the present invention, the following is a further detailed description of a multi-parameter prediction method for photovoltaic power fluctuation based on a multi-task one-dimensional convolutional network of the present invention in conjunction with the accompanying drawings.

[0033] The present invention provides a multi-parameter prediction method for photovoltaic power generation fluctuation based on a multi-task one-dimensional convolutional network. Figure 1 As shown, the process comprises the following steps, and the following steps are performed in sequence:

[0034] Step 1: Select the prediction parameters, and choose the standard deviation and fluctuation variation of photovoltaic power generation as the direct characterization of the fluctuation range and fluctuation rate of photovoltaic power generation.

[0035] Step 2: Perform time-delay correlation analysis of ACF, PACF, and MIC on each parameter to conduct predictability analysis of multiple parameters of photovoltaic power generation.

[0036] Step 3: A multi-task one-dimensional convolutional neural network was designed to perform the above-mentioned multi-parameter prediction of photovoltaic power generation power fluctuations.

[0037] In step 1, select the photovoltaic power fluctuation intensity I p To characterize the fluctuation range of photovoltaic power generation, as shown in (1).

[0038]

[0039] In the formula represents the average horizontal radiance, σ P represents the standard deviation of horizontal irradiance, as shown in (2).

[0040]

[0041] P(i) is the photovoltaic power generation at time i, It represents the average power in the time window from time i to i+Δt. n is the total number of data points in the time window, that is, the number of data in the time range from t to t+Δt. t is the starting time point, indicating the starting time of the calculation.

[0042] According to Chebyshev's inequality, the range of positive and negative k times the standard deviation of the mean in any distribution covers at least Therefore, in practical applications, we can use σ P The corresponding multiples of the value are used as thresholds to measure the fluctuation range of photovoltaic power generation within Δt.

[0043] The PV power fluctuation variation is chosen to measure the fluctuation rate of PV power, as shown in (3).

[0044]

[0045] t i represents the i-th time point in the time series, and n represents the total number of n data points in the time series;

[0046] The result means the average change of photovoltaic power generation within the time interval Δt, and the rate of change can be expressed as γ'(Δt) / Δt. Since Δt in the calculation is a constant, γ'(Δt) can be used to measure the rate of change of power.

[0047] In step 2, the time-delay correlation analysis of the autocorrelation function (ACF), partial autocorrelation function (PACF) and maximum information coefficient (MIC) of each parameter is performed, and then the predictability analysis of the parameters selected in step 1 is performed.

[0048] ACF is used to measure the correlation of a time series at different lags. It shows the similarity between the values ​​at each moment in the series by calculating the autocorrelation coefficients at different lag steps, as shown in (4).

[0049]

[0050] Where k represents the lag order, X t and X t+k are the observation values ​​at time t and t+k respectively, and T is the total length of the sequence.

[0051] Based on ACF, PACF further considers the direct relationship between each lag step and excludes the influence of intermediate lags, as shown in (5).

[0052]

[0053] Where φ k,k is the partial autocorrelation coefficient of the sequence at lag k, γ kis the value of the autocorrelation function ACF at lag k, φ k-1,j is the partial autocorrelation coefficient of lag j with order k-1.

[0054] MIC is used to detect the nonlinear relationship between two variables. It measures the dependency between variables by dividing the data into different grids and calculating the mutual information in each division. As shown in (6)

[0055]

[0056] Where x and y are the sizes of the partitioned grids, and I(X;Y) is the mutual information.

[0057] In step 3, multi-task learning is a related task transfer learning method based on shared representation. There are two types of sharing mechanisms here: hard sharing and soft sharing. Hard sharing is very common, and the schematic architecture is as follows Figure 2 The sharing layer describes the relevance of related tasks by directly sharing parameters, while the task layer describes the differences between tasks through independent structures that do not interfere with each other. Soft sharing is also called constraint-based sharing. Each task has its own model and parameters, but the distance between model parameters is regularized to ensure the similarity of the parameters of each task.

[0058] One-dimensional convolutional network (1dCNN) is a one-dimensional form of convolutional neural network. Compared with 2D CNN, which is very good at image processing, 1dCNN is suitable for one-dimensional time series data modeling. It treats different variables in the same period as the attributes of the whole through multi-channel input, uses convolution kernel to extract local information to identify simple patterns in the data, and then generates more complex patterns in the high-level layer through inter-layer propagation, thereby realizing the function of modeling prediction or classification. This method does not have the problem of setting hard rules for dimensional conversion when 2D (or 3D) CNN processes one-dimensional time series data, and may also destroy the problem of time series correlation, but retains the core highlights of CNN local receptive field and weight sharing, and has the advantages of simple network structure, few model parameters, fast training speed, and strong information characterization ability. The structure of 1dCNN mainly includes convolution layer, pooling layer, and fully connected layer, and the latter two can be configured selectively.

[0059] The convolution layer performs the product and sum operation of the corresponding elements by sliding multiple convolution kernels on the layer input (which can be expanded and padded along the boundaries), projecting the information in the receptive field to produce a set of parallel feature maps (one convolution kernel corresponds to one feature map). The convolution kernel is smaller than the input size, and the sliding operation ensures that each feature map shares the same weights and biases. The sliding stride (Stride) and padding (Padding) dimensions are preset hyperparameters that jointly control the size of the output feature map.

[0060] The convolution kernel is a feature finder, and the pooling layer is a nonlinear downsampling of the feature map it produces. It can significantly reduce the size of the data space and the number of parameters, suppress overfitting to a certain extent, and ensure translation invariance. The most common type is maximum pooling, which divides the layer input into several rectangular areas and outputs the maximum value for each sub-area. According to the preset window size and moving stride, the pooling layer calculates the output on one pooling window at a time, and then continues the moving calculation according to the stride setting. However, because the pooling layer reduces the data dimension too quickly, the trend in recent literature is to use smaller pooling sizes or reduce the use of pooling layers.

[0061] like Figure 3 As shown in the figure, the 1dCNN structure consisting of 2 convolutional layers, 1 pooling layer and 1 fully connected layer is schematically shown to give a more intuitive understanding. A certain order of each variable is input into an equal number of channels, and a feature atlas is generated through the convolution layer. After activation processing, it is downscaled by the pooling layer, and then a convolution operation and a fully connected operation are performed to obtain the model output.

[0062] In step 3, the detailed process of multi-task 1dCNN modeling is summarized in Table 1 and Table 2, and the model structure diagram is shown in Figure 4 shown.

[0063] Table 1 is based on {p, σ , γ } Multi-task 1dCNN shared layer part with hard sharing of prediction parameters

[0064]

[0065]

[0066] Table 2 {p, σ , γ } Multi-task 1dCNN task layer part with differentiated prediction parameters

[0067]

[0068] The present invention was verified on a simulation platform by constructing specific working conditions. It was found that the multi-parameter prediction method for photovoltaic power generation fluctuations based on a multi-task one-dimensional convolutional network can effectively improve the prediction performance of each parameter and each time step, which is helpful to promote the development of photovoltaic technology and the popularization and application of renewable energy.

[0069] Embodiment 1:

[0070] Step 1: To match the main time resolution requirements of the power system for photovoltaic power generation forecasting, the modeling data scale used is set to 15 minutes. Since the acquisition of the above-mentioned multiple parameters requires statistical information based on high-frequency photovoltaic power generation details, the data used in this embodiment is derived from the 5-minute power value of the DKA Solar Center as the source data for calculation, and the sampling location is a 1.8 MW solar photovoltaic power station located in Uluru, central Australia;

[0071] The time span of the data is from January 1, 2022 to December 29, 2022. The data is downscaled to 15 minutes in a full window sliding manner to obtain the power mean, fluctuation standard deviation, and fluctuation variation at this scale, such as Figure 5 In order to keep the dimensions consistent, the variance is squared and combined with other variables to form a multi-parameter modeling method, which is abbreviated as {P, σ, γ}.

[0072] Step 2: Perform predictability analysis on the data processed in step 1, and perform Pearson and MIC quantitative analysis on the correlation between parameters. The results are shown in Table 3.

[0073] Table 3

[0074]

[0075] The two coefficients between the three parameters are very significant and the correlation is clear. The delay correlation analysis of ACF, PACF and MIC is performed on each parameter. Since the work in this chapter is based on statistical methods, the predictable length of each parameter will not be too long, so the 16 steps of the rolling scheduling 4h scale are used as the upper limit of the analysis order.

[0076] from Figures 6 to 8 It can be seen that the ACF and MIC coefficients of each parameter show clear significance when the order is small, and the PACF coefficients also show truncation phenomenon including the first-order strong partial autocorrelation. The multi-parameter set consisting of the mean p of photovoltaic power generation, the standard deviation σ of fluctuation, and the square root of variation γ has short-term predictability. Further exploration of the predictable length, the ACF of p at 16 steps is 0.33, which is still very significant. Therefore, based on the above analysis, the subsequent study conservatively adopts a 4-step forecast length of 1h in advance. The predictability of the multi-parameter set within this range is significantly unified, which can ensure the reasonable and reliable forecast effect.

[0077] Step 3: Design a multi-task one-dimensional convolutional network to predict the multi-parameter fluctuation of photovoltaic power generation:

[0078] The collective data {P, σ, γ} is divided into training set and test set according to the ratio of 0.7:0.3, and then the training set is trained. The prediction ability of the model is tested with the test set, and compared with the prediction effects of single-task ANN, single-task LSTM, and single-task 1dCNN. The accuracy level of absolute error is measured by RMSE. The results are as follows: Fig. 9 As shown in Figure 2, it can be seen that for the prediction effect of each time step of {p,σ,γ}, the multi-task 1dCNN has achieved better performance, which shows that the multi-task prediction mode and the local receptive field technology of 1dCNN can more fully extract the complementary information between parameters.

[0079] It is to be understood that the present invention is described by some embodiments, and it is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the scope of protection of the present invention.

Claims

1. A multi-parameter prediction method for photovoltaic power generation fluctuation based on a multi-task one-dimensional convolutional network, characterized by: The following steps are included: Step 1: Select the standard deviation of photovoltaic power generation fluctuation as the photovoltaic power generation power fluctuation range parameter, and select the photovoltaic power generation power fluctuation variation as the parameter for measuring the photovoltaic power generation power fluctuation rate; Step 2: Perform time-delay correlation analysis of the autocorrelation function ACF, partial autocorrelation function PACF and maximum information coefficient MIC on the parameters described in step 1, and perform predictability analysis on multi-parameters of photovoltaic power generation fluctuations; Step 3: Establish a multi-task one-dimensional convolutional neural network to predict the multi-parameter fluctuations of photovoltaic power generation described in step 2.

2. The photovoltaic power generation power fluctuation multi-parameter prediction method based on multi-task one-dimensional convolutional network according to claim 1 is characterized by: The autocorrelation function ACF described in step 2 is used to measure the correlation of the time series itself at different lags, and the similarity between the values ​​at each moment in the sequence is displayed by calculating the autocorrelation coefficients of different lag steps: In the formula, k represents the lag order, X t and X t+k are the observation values ​​at time t and t+k respectively, T is the total length of the sequence, is the mean of the time series, defined as: The partial autocorrelation function PACF further considers the direct relationship between each lag step on the basis of ACF, excluding the influence of intermediate lags, and the formula is: In the formula, φ k,k is the partial autocorrelation coefficient of the sequence at lag k, γ k is the value of the autocorrelation function ACF at lag k, φ k-1,j is the partial autocorrelation coefficient of lag j when the order is k-1; The maximum information coefficient MIC is used to detect the nonlinear relationship between two variables, and the formula is: Where x and y are the sizes of the partitioned grids, and I(X;Y) is the mutual information.

3. The photovoltaic power generation power fluctuation multi-parameter prediction method based on multi-task one-dimensional convolutional network according to claim 1 is characterized by: The multi-task one-dimensional convolutional neural network described in step 3 is used for time series data modeling, including a convolutional layer, a pooling layer, and a fully connected layer. The convolutional layer slides multiple convolution kernels on the layer input to perform the corresponding element product and sum operation, and projects the information in the receptive field to generate a set of parallel feature maps; the pooling layer performs nonlinear downsampling on the feature map generated by the convolutional layer; After downsampling, the feature set undergoes a layer of convolution operation and a full connection operation to obtain the model output.

Images