Distributed photovoltaic power prediction method in low voltage area based on temporal convolutional network

Abstract

A distributed photovoltaic power prediction method for low-voltage areas based on a temporal convolutional network is proposed. By constructing a temporal convolutional network and integrating an efficient channel attention mechanism into the temporal convolutional network, a TCN-ECN photovoltaic power prediction model is formed. Photovoltaic power generation data and multivariate meteorological factors of the prediction day are input into the trained SENET-TCN model to predict the photovoltaic power generation data of the prediction day in the low-voltage area. The present invention can not only integrate the efficient channel attention mechanism into the temporal convolutional network to form a TCN-ECN photovoltaic power prediction model, effectively improving the accuracy of photovoltaic power prediction, but also fully consider the complex characteristics of distributed photovoltaics through the FCM algorithm, effectively reducing the prediction error.

Description

Technical Field

[0001] The present invention relates to a photovoltaic power prediction method, and in particular to a distributed photovoltaic power prediction method in a low-voltage area based on a time convolutional network. Background Art

[0002] The penetration rate of distributed renewable energy represented by photovoltaics in the low-voltage area of ​​the distribution network has increased rapidly. A large number of distributed photovoltaics are connected to the low-voltage distribution network, which has changed the load characteristics of the low-voltage area, aggravated the short-term load fluctuations, and increased the difficulty of distribution network regulation. In order to reduce the impact of photovoltaic processing uncertainty on the power grid, research on distributed photovoltaic power prediction in low-voltage areas has become a hot topic.

[0003] Currently, the model is usually established through convolutional neural networks, long short-term memory, and attention mechanism under multiple correlation and target variable prediction models. The attention mechanism assigns weights to features in the time series according to the time step to avoid interference in traditional LSTM methods.

[0004] Although this photovoltaic power prediction method can improve the accuracy of probabilistic prediction and weaken the impact of strong volatility of distributed photovoltaic power on the power system, it still has the following defects:

[0005] The current prediction method for the low-voltage distribution network composed of distributed photovoltaic units only relies on weights, which has limited energy for extracting the complex characteristics of distributed photovoltaics in the low-voltage area and insufficient prediction accuracy.

[0006] The information disclosed in this background technology section is only intended to increase the understanding of the overall background of the application, and should not be regarded as acknowledging or suggesting in any form that the information constitutes the prior art already known to ordinary technicians in this field. Summary of the invention

[0007] The purpose of the present invention is to overcome the shortcomings of insufficient prediction accuracy in the prior art and to provide a distributed photovoltaic power prediction method in a low-voltage area based on a time convolutional network with higher prediction accuracy.

[0008] To achieve the above objectives, the technical solution of the present invention is:

[0009] A method for predicting distributed photovoltaic power in a low-voltage area based on a temporal convolutional network, the method comprising the following steps:

[0010] S1, constructing a SENet-TCN model, wherein the SENet-TCN model includes a temporal convolutional network and a compression and excitation network, and the compression and excitation network is introduced into the temporal convolutional network to form the SENet-TCN model;

[0011] The temporal convolutional network is used to calculate the photovoltaic output prediction result according to the weather data of the forecast day or the meteorological factors that have a significant impact on the photovoltaic power prediction result;

[0012] The constructed compression and excitation network is used to calculate the weights of virtual signals of different dimensions on the photovoltaic power prediction results, and obtain meteorological factors that have a significant impact on the photovoltaic power prediction results;

[0013] S2, constructing an FCM algorithm model, and constructing an FCM algorithm model through missing data filling, data standardization, outlier processing, data set dimension reduction and fuzzy C-means clustering. The FCM algorithm model is used to process the historical photovoltaic output results and the corresponding weather data;

[0014] S3, data prediction, inputs the historical photovoltaic output results and the corresponding weather into the FCM algorithm model to obtain the photovoltaic output data set of the distributed photovoltaic units in the low-voltage area classified by weather type, and iteratively trains the SENet-TCN model through the obtained photovoltaic output data set. The photovoltaic output data set of the distributed photovoltaic units in the low-voltage area classified by weather type and the multivariate meteorological factors of the forecast day are input into the trained SENet-TCN model to predict the photovoltaic power generation data of the low-voltage area on the forecast day.

[0015] In S1, the expression of the causal convolution model is:

[0016]

[0017] x t =∑(x t-k *w k ),k∈[0,t-1];

[0018] In the above formula, P(x it ) is the prediction probability, T is the total time, Π is the orthogonal operation, represents the meteorological feature sequence of the i-th sample, with a total of m meteorological features, x t =∑(x t-k *w k ), k∈[0,t-1] is the random evolution layer, x t is the meteorological characteristic sequence of the i-th sample at time t, k is an integer between [0, t-1], w k is the filter weight, which is [0,1];

[0019] The expression of the dilated convolution model is:

[0020]

[0021] In the above formula, F(x s) is the input x output by the network at time s s , k is the size of the convolution kernel, d is the dilation coefficient, (sd i ) is the sequence corresponding to the elements in the convolution kernel;

[0022] The expression of the residual connection model is:

[0023] o = RELU[x+F(x s )];

[0024] In the above formula, o is the output of the temporal convolutional network, RELU[] is the rectified linear unit function operation, x is the model input, and F(x s ) is the network output at time s.

[0025] In S1, the compression and excitation network includes a compression network and an excitation network, the compression network is used to compress global spatial information, perform feature learning in the channel dimension, and analyze the importance difference of each channel, and the excitation network is used to assign different weights to each channel, thereby forming a channel attention module;

[0026] The cth element of the data z obtained by the compression network is calculated by the following formula:

[0027]

[0028] In the above formula, H, W, and C are data dimension coefficients, γ and b are constants, and γ is set to 2 and b is set to 1;

[0029] The excitation network adopts an S-type activation gating mechanism:

[0030] S=F ex (z,W);

[0031] In the above formula, F ex (·) is the ReLU function;

[0032] The excitation network obtains the final output by rescaling the result using activation s:

[0033]

[0034] In the above formula F scale (u C ,S C ) is a scalar S C With the feature map u C The channel multiplication between .

[0035] In S2, the weather data corresponding to the historical photovoltaic output result is a discrete value, and the specific operation of the missing data filling is to fill the characteristic variables whose median and mean differ by less than 10% by the mean, and to fill the characteristic variables whose median and mean differ by more than 10% by the median;

[0036] The specific operations of data standardization are:

[0037] The standard score is used to standardize the data, and the expression is:

[0038]

[0039] In the above formula, Z is the standardized data, x is the initial data, μ is the mean of the overall data, and σ is the standard deviation of the overall data;

[0040] The specific operation of the outlier processing is to only retain the values ​​distributed in [μ-3σ,μ+3σ];

[0041] The specific operation of reducing the dimension of the data set includes the following steps:

[0042] Assume that the original data has m samples X 1 ,X 2 ,…,X m , each sample has n-dimensional features Each feature x j All have their own eigenvalues. After standardizing the data, we get the covariance matrix. In the case of three-dimensional data, the expression of the covariance matrix is:

[0043]

[0044] Find the eigenvalue λ and eigenvector u of the covariance matrix C;

[0045] Project the original features onto the selected feature vector to obtain the new k-dimensional features after dimensionality reduction, which is expressed as:

[0046]

[0047] In summary, for each sample X i , will be replaced by the original becomes To achieve the purpose of dimensionality reduction;

[0048] The specific operation of the fuzzy C-means clustering includes the following steps:

[0049] Design the number of clusters and the hyperparameter m, and initialize the membership matrix with random numbers in [0, 1] Where n is the number of data objects, m is the number of clusters, and u ijFor data object x i The degree of membership to the cluster center;

[0050] The updated cluster center C j The expression is:

[0051]

[0052] The expression for updating the membership matrix U is:

[0053]

[0054] The iteration termination condition is that the maximum number of iterations is reached or the error exceeds a threshold. The error expression is:

[0055]

[0056] A distributed photovoltaic power prediction system for low-voltage areas based on a time convolutional network, the system comprising: a SENet-TCN model building module, an FCM algorithm model building module and a data prediction module;

[0057] The SENet-TCN model building module is used to build a temporal convolutional network and a compression and excitation network, and introduce the compression and excitation network into the temporal convolutional network to form a SENet-TCN model;

[0058] The FCM algorithm model construction module is used to construct the FCM algorithm model through missing data filling, data standardization, outlier processing, data set dimension reduction and fuzzy C-means clustering. The FCM algorithm model is used to process the historical photovoltaic output results and the corresponding weather data;

[0059] The data prediction module is used to input the historical photovoltaic output results and the corresponding weather into the FCM algorithm model to obtain the photovoltaic output data set of the distributed photovoltaic units in the low-voltage area classified by weather type, and iteratively train the SENet-TCN model through the obtained photovoltaic output data set. The photovoltaic output data set of the distributed photovoltaic units in the low-voltage area classified by weather type and the multivariate meteorological factors of the forecast day are input into the trained SENet-TCN model to predict the photovoltaic power generation data of the low-voltage area on the forecast day.

[0060] The temporal convolutional network includes a causal convolutional model, a dilated convolutional model and a residual connection model. The expression of the causal convolutional model is:

[0061]

[0062] x t =∑(x t-k *w k),k∈[0,t-1];

[0063] In the above formula, P(x it ) is the prediction probability, T is the total time, Π is the orthogonal operation, represents the meteorological feature sequence of the i-th sample, with a total of m meteorological features, x t =∑(x t-k *w k ), k∈[0,t-1] is the random evolution layer, x t is the meteorological characteristic sequence of the i-th sample at time t, k is an integer between [0, t-1], w k is the filter weight, which is [0,1];

[0064] The expression of the dilated convolution model is:

[0065]

[0066] In the above formula, F(x s ) is the input x output by the network at time s s , k is the size of the convolution kernel, d is the dilation coefficient, (sd i ) is the sequence corresponding to the elements in the convolution kernel;

[0067] The expression of the residual connection model is:

[0068] o = RELU[x+F(x s )];

[0069] In the above formula, o is the output of the temporal convolutional network, RELU[] is the rectified linear unit function operation, x is the model input, and F(x s ) is the network output at time s.

[0070] The compression and excitation network construction module includes a compression network and an excitation network, wherein the compression network is used to compress global spatial information, perform feature learning in the channel dimension, and analyze the importance difference of each channel, and the excitation network is used to assign different weights to each channel, thereby forming a channel attention module;

[0071] The cth element of the data z obtained by the compression network is calculated by the following formula:

[0072]

[0073] In the above formula, H, W, and C are data dimension coefficients, γ and b are constants, and γ is set to 2 and b is set to 1;

[0074] The excitation network adopts an S-type activation gating mechanism:

[0075] S=Fex (z,W);

[0076] In the above formula, F ex (·) is the ReLU function;

[0077] The excitation network obtains the final output by rescaling the result using activation s:

[0078]

[0079] In the above formula, F scale (u C ,S C ) is a scalar S C With the feature map u C The channel multiplication between .

[0080] In the FCM algorithm model, the weather data corresponding to the historical photovoltaic output results are discrete values, and the specific operation of filling the missing data is:

[0081] The feature variables whose median and mean differ by less than 10% are filled by the mean, and the feature variables whose median and mean differ by more than 10% are filled by the median;

[0082] The specific operation of data standardization is: standardize the data using standard scores, and the expression is:

[0083]

[0084] In the above formula, Z is the standardized data, x is the initial data, μ is the mean of the overall data, and σ is the standard deviation of the overall data;

[0085] The specific operation of the outlier processing is to only retain the values ​​distributed in [μ-3σ,μ+3σ];

[0086] The specific operation of reducing the dimension of the data set includes the following steps:

[0087] Assume that the original data has m samples X 1 ,X 2 ,…,X m , each sample has n-dimensional features Each feature x j All have their own eigenvalues. After standardizing the data, we get the covariance matrix. In the case of three-dimensional data, the expression of the covariance matrix is:

[0088]

[0089] Find the eigenvalue λ and eigenvector u of the covariance matrix C;

[0090] Project the original features onto the selected feature vector to obtain the new k-dimensional features after dimensionality reduction, which is expressed as:

[0091]

[0092] In summary, for each sample X i , will be replaced by the original becomes To achieve the purpose of dimensionality reduction;

[0093] The specific operation of the fuzzy C-means clustering includes the following steps:

[0094] Design the number of clusters and the hyperparameter m, and initialize the membership matrix with random numbers in [0, 1] Where n is the number of data objects, m is the number of clusters, and u ij For data object x i The degree of membership to the cluster center;

[0095] The updated cluster center C j The expression is:

[0096]

[0097] The expression for updating the membership matrix U is:

[0098]

[0099] The iteration termination condition is that the maximum number of iterations is reached or the error exceeds a threshold. The error expression is:

[0100]

[0101] A low-voltage area distributed photovoltaic power prediction device based on a time convolutional network, the device comprising a processor and a memory;

[0102] The memory is used to store computer program code and transmit the computer program code to the processor;

[0103] The processor is used to execute the aforementioned low-voltage distributed photovoltaic power prediction method based on a time convolutional network according to the instructions in the computer program code.

[0104] A power balance storage medium for distributed resources, on which a computer program is stored, and when the computer program is executed by a processor, the aforementioned low-voltage distributed photovoltaic power prediction method based on a time convolutional network is implemented.

[0105] Compared with the prior art, the present invention has the following beneficial effects:

[0106] 1. In the distributed photovoltaic power prediction method for low-voltage areas based on a temporal convolutional network of the present invention, a TCN-ECN photovoltaic power prediction model is formed by constructing a temporal convolutional network and integrating an efficient channel attention mechanism into the temporal convolutional network. Photovoltaic power generation data and multivariate meteorological factors of the prediction day are input into the trained SENET-TCN model to predict the photovoltaic power generation data of the prediction day of the low-voltage area. Therefore, this design can integrate the efficient channel attention mechanism into the temporal convolutional network to form a TCN-ECN photovoltaic power prediction model, effectively improving the accuracy of photovoltaic power prediction.

[0107] 2. In the distributed photovoltaic power prediction method for low-voltage areas based on a time convolutional network of the present invention, the FCM algorithm is constructed by missing data completion, data standardization, outlier processing, data set dimension reduction and fuzzy C-means clustering, and the photovoltaic data is classified and processed according to the weather type, and the photovoltaic output data of the distributed photovoltaic units in the low-voltage area classified by the weather type is obtained, so that the photovoltaic power prediction method can fully consider the complex characteristics of distributed photovoltaics and reduce the prediction error. Therefore, this design can fully consider the complex characteristics of distributed photovoltaics through the FCM algorithm and effectively reduce the prediction error.

[0108] 3. In a method for predicting distributed photovoltaic power in low-voltage areas based on a time convolutional network, the present invention further guarantees the generation of xt by adding a random evolution layer to the causal convolutional model. Since the generation of xt only relies on data from time 0 to t-1, it ensures that the current prediction is not affected by future data. BRIEF DESCRIPTION OF THE DRAWINGS

[0109] Figure 1 It is a flow chart of the present invention.

[0110] Figure 2 It is a schematic diagram of the system structure of the present invention.

[0111] Figure 3 It is a schematic diagram of the device structure of the present invention.

[0112] Figure 4 This is the spring forecast result error diagram in Example 2.

[0113] Figure 5 This is the error diagram of the summer prediction results in Example 2.

[0114] Figure 6 This is the error diagram of the autumn forecast results in Example 2.

[0115] Figure 7 This is the winter forecast result error diagram in Example 2. DETAILED DESCRIPTION

[0116] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be described clearly and completely in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0117] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0118] Embodiment 1:

[0119] See also Figure 1 , a distributed photovoltaic power prediction method for low-voltage areas based on a time convolutional network, the photovoltaic power prediction method comprising the following steps:

[0120] S1, constructing a SENet-TCN model, wherein the SENet-TCN model includes a temporal convolutional network and a compression and excitation network, and the compression and excitation network is introduced into the temporal convolutional network to form the SENet-TCN model;

[0121] The temporal convolutional network is used to calculate the photovoltaic output prediction result according to the weather data of the forecast day or the meteorological factors that have a significant impact on the photovoltaic power prediction result;

[0122] The constructed compression and excitation network is used to calculate the weights of virtual signals of different dimensions on the photovoltaic power prediction results, and obtain meteorological factors that have a significant impact on the photovoltaic power prediction results;

[0123] S2, constructing an FCM algorithm model, and constructing an FCM algorithm model through missing data filling, data standardization, outlier processing, data set dimension reduction and fuzzy C-means clustering. The FCM algorithm model is used to process the historical photovoltaic output results and the corresponding weather data;

[0124] S3, data prediction, inputs the historical photovoltaic output results and the corresponding weather into the FCM algorithm model to obtain the photovoltaic output data set of the distributed photovoltaic units in the low-voltage area classified by weather type, and iteratively trains the SENet-TCN model through the obtained photovoltaic output data set. The photovoltaic output data set of the distributed photovoltaic units in the low-voltage area classified by weather type and the multivariate meteorological factors of the forecast day are input into the trained SENet-TCN model to predict the photovoltaic power generation data of the low-voltage area on the forecast day.

[0125] In S1, the expression of the causal convolution model is:

[0126]

[0127] x t =∑(x t-k *w k ),k∈[0,t-1];

[0128] In the above formula, P(x it ) is the prediction probability, T is the total time, Π is the orthogonal operation, represents the meteorological feature sequence of the i-th sample, with a total of m meteorological features, x t =∑(x t-k *w k ), k∈[0,t-1] is the random evolution layer, x t is the meteorological characteristic sequence of the i-th sample at time t, k is an integer between [0, t-1], w k is the filter weight, which is [0,1];

[0129] The filter is used to give data from 0 to t-1 (i.e., affecting x t data) a weighted value, because the data at different times have an impact on x t The degree of influence is different. The closer to time t, the higher the weighted value.

[0130] The expression of the dilated convolution model is:

[0131]

[0132] In the above formula, F(x s ) is the input x output by the network at time s s , k is the size of the convolution kernel, d is the dilation coefficient, (sd i ) is the sequence corresponding to the elements in the convolution kernel;

[0133] The expression of the residual connection model is:

[0134] o = RELU[x+F(x s)];

[0135] In the above formula, o is the output of the temporal convolutional network, RELU[] is the rectified linear unit function operation, x is the model input, and F(x s ) is the network output at time s.

[0136] In S1, the compression and excitation network includes a compression network and an excitation network, the compression network is used to compress global spatial information, perform feature learning in the channel dimension, and analyze the importance difference of each channel, and the excitation network is used to assign different weights to each channel, thereby forming a channel attention module;

[0137] The cth element of the data z obtained by the compression network is calculated by the following formula:

[0138]

[0139] In the above formula, H, W, and C are data dimension coefficients, γ and b are constants, and γ is set to 2 and b is set to 1;

[0140] The excitation network adopts an S-type activation gating mechanism:

[0141] S=F ex (z,W);

[0142] In the above formula, F ex (·) is the ReLU function;

[0143] The excitation network obtains the final output by rescaling the result using activation s:

[0144]

[0145] In the above formula, F scale (u C ,S C ) is a scalar S C With the feature map u C The channel multiplication between .

[0146] In S2, the weather data corresponding to the historical photovoltaic output result is a discrete value, and the specific operation of the missing data filling is to fill the characteristic variables whose median and mean differ by less than 10% by the mean, and to fill the characteristic variables whose median and mean differ by more than 10% by the median;

[0147] The specific operations of data standardization are:

[0148] The standard score is used to standardize the data, and the expression is:

[0149]

[0150] In the above formula, Z is the standardized data, x is the initial data, μ is the mean of the overall data, and σ is the standard deviation of the overall data;

[0151] The specific operation of the outlier processing is to only retain the values ​​distributed in [μ-3σ,μ+3σ];

[0152] The specific operation of reducing the dimension of the data set includes the following steps:

[0153] Assume that the original data has m samples X 1 ,X 2 ,…,X m , each sample has n-dimensional features Each feature x j All have their own eigenvalues. After standardizing the data, we get the covariance matrix. In the case of three-dimensional data, the expression of the covariance matrix is:

[0154]

[0155] Find the eigenvalue λ and eigenvector u of the covariance matrix C;

[0156] Project the original features onto the selected feature vector to obtain the new k-dimensional features after dimensionality reduction, whose expression is:

[0157]

[0158] In summary, for each sample X i , will be replaced by the original becomes To achieve the purpose of dimensionality reduction;

[0159] The specific operation of the fuzzy C-means clustering includes the following steps:

[0160] Design the number of clusters and the hyperparameter m, and initialize the membership matrix with random numbers in [0, 1] Where n is the number of data objects, m is the number of clusters, and u ij For data object x i The degree of membership to the cluster center;

[0161] The updated cluster center C j The expression is:

[0162]

[0163] The expression for updating the membership matrix U is:

[0164]

[0165] The iteration termination condition is that the maximum number of iterations is reached or the error exceeds a threshold. The error expression is:

[0166]

[0167] A distributed photovoltaic power prediction system for low-voltage areas based on a time convolutional network, the system comprising: a SENet-TCN model building module, an FCM algorithm model building module and a data prediction module;

[0168] The SENet-TCN model building module is used to build a temporal convolutional network and a compression and excitation network, and introduce the compression and excitation network into the temporal convolutional network to form a SENet-TCN model;

[0169] The FCM algorithm model construction module is used to construct the FCM algorithm model through missing data filling, data standardization, outlier processing, data set dimension reduction and fuzzy C-means clustering. The FCM algorithm model is used to process the historical photovoltaic output results and the corresponding weather data;

[0170] The data prediction module is used to input the historical photovoltaic output results and the corresponding weather into the FCM algorithm model to obtain the photovoltaic output data set of the distributed photovoltaic units in the low-voltage area classified by weather type, and iteratively train the SENet-TCN model through the obtained photovoltaic output data set. The photovoltaic output data set of the distributed photovoltaic units in the low-voltage area classified by weather type and the multivariate meteorological factors of the forecast day are input into the trained SENet-TCN model to predict the photovoltaic power generation data of the low-voltage area on the forecast day.

[0171] Embodiment 2:

[0172] In order to evaluate the performance of the proposed FCM-SENET-TCN low-voltage distributed photovoltaic prediction model, the mean absolute percentage error (MAPE) and root mean square error (RMSE) were calculated. The photovoltaic power prediction simulation was carried out in the low-voltage distributed photovoltaic system shown in Table 1. The evaluation indicators are defined as follows:

[0173]

[0174] Where: rX' is the actual power generation on the forecast day; fX' is the forecast value of power generation; N is the total number of data.

[0175] Table 1 Photovoltaic system data

[0176] project data Inverter size / type 4×6kW; SMA SMC 6000A Tracker Type DEGE Renergie5000NT; dual axis Array area / m2 4×38.37 Panel Type TRW TSM-195DC01A Number of panels 4×30 Panel Rating / W 195 Array rating / kW 23.4 ;

[0177] In order to verify the effectiveness of the FCM-SENet-TCN model, the influence of the peak changes of meteorological elements under stable weather conditions in different seasons on the forecast performance of the model was first studied. The sample data of typical days in four seasons were set as the forecast date, and LSTM, TCN and SENet-TCN based on the FCM algorithm were selected as comparative models for experiments. The error diagram between the prediction results and the true value can be seen in Figures 4 to 7 .

[0178] Table 2 lists the RMSE values ​​of the FCM-SENet-TCN model and the other two models in different seasons. It can be seen that the RMSE values ​​of the FCM-SENet-TCN model in all seasons are smaller than those of the other models. The average RMSE values ​​of the FCM-SENet-TCN model on the entire dataset are reduced by 54.74%, 29.47%, and 19.86%, respectively, which indicates that the FCM-SENet-TCN model has the smallest deviation from the actual results.

[0179] Table 3 lists the MAPE values ​​of the FCM-SENet-TCN model and the other two models in different seasons. The results show that in all seasons, the MAPE value of the FCM-SENet-TCN model is lower than that of the other models. Compared with the other three models, the average MAPE value of the FCM-SENet-TCN model on the entire dataset decreased by 56.92%, 29.90%, and 18.19%, respectively. Therefore, the MAPE value of the FCM-SENet-TCN model is the best among all models, and the true error of the prediction is the smallest.

[0180] Table 2 Comparison of RMPE (kW) values ​​in different seasons

[0181] Model spring summer Autumn winter average FCM-LSTM 0.9389 1.2985 0.7436 0.4425 0.8558 FCM-TCN 0.6053 0.4833 0.7028 0.4054 0.5492 FCM-SENet-TCN 0.5391 0.3311 0.4369 0.2422 0.3873

[0182] Table 3 Comparison of MAPE (kW) values ​​in different seasons

[0183] Model spring summer Autumn winter average FCM-LSTM 0.7379 1.1692 0.5897 0.3220 0.7047 FCM-TCN 0.4659 0.4443 0.5196 0.3027 0.4331 FCM-SENet-TCN 0.3759 0.3453 0.3291 0.1642 0.3036 ;

[0184] Embodiment 3:

[0185] A low-voltage area distributed photovoltaic power prediction device based on a time convolutional network, the device comprising a processor and a memory;

[0186] The memory is used to store computer program code and transmit the computer program code to the processor;

[0187] The processor is used to execute the low-voltage distributed photovoltaic power prediction method based on a time convolutional network described in Example 1 according to the instructions in the computer program code.

[0188] Embodiment 4:

[0189] A power balance storage medium for distributed resources, on which a computer program is stored. When the computer program is executed by a processor, the method for predicting distributed photovoltaic power in a low-voltage area based on a time convolutional network described in Example 1 is implemented.

[0190] The above description is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiment. Any equivalent modifications or changes made by ordinary technicians in this field based on the contents disclosed by the present invention should be included in the protection scope recorded in the claims.

Claims

1. A method for predicting distributed photovoltaic power in low-voltage areas based on a temporal convolutional network, characterized in that: The photovoltaic power prediction method comprises the following steps: S1, constructing a SENet-TCN model, wherein the SENet-TCN model includes a temporal convolutional network and a compression and excitation network, and the compression and excitation network is introduced into the temporal convolutional network to form the SENet-TCN model; The temporal convolutional network is used to calculate the photovoltaic output prediction result according to the weather data of the forecast day or the meteorological factors that have a significant impact on the photovoltaic power prediction result; The constructed compression and excitation network is used to calculate the weights of virtual signals of different dimensions on the photovoltaic power prediction results, and obtain meteorological factors that have a significant impact on the photovoltaic power prediction results; S2, constructing an FCM algorithm model, and constructing an FCM algorithm model through missing data filling, data standardization, outlier processing, data set dimension reduction and fuzzy C-means clustering. The FCM algorithm model is used to process the historical photovoltaic output results and the corresponding weather data; S3, data prediction, inputs the historical photovoltaic output results and the corresponding weather into the FCM algorithm model to obtain the photovoltaic output data set of the distributed photovoltaic units in the low-voltage area classified by weather type, and iteratively trains the SENet-TCN model through the obtained photovoltaic output data set. The photovoltaic output data set of the distributed photovoltaic units in the low-voltage area classified by weather type and the multivariate meteorological factors of the forecast day are input into the trained SENet-TCN model to predict the photovoltaic power generation data of the low-voltage area on the forecast day.

2. According to claim 1, a method for predicting distributed photovoltaic power in low-voltage areas based on a temporal convolutional network is characterized in that: In S1, the expression of the causal convolution model is: x t =∑(x t-k *w k ),k∈[0,t-1]; In the above formula, P(x it ) is the prediction probability, T is the total time, Π is the orthogonal operation, represents the meteorological feature sequence of the i-th sample, with a total of m meteorological features, x t =∑(x t-k *w k ), k∈[0,t-1] is the random evolution layer, x t is the meteorological characteristic sequence of the i-th sample at time t, k is an integer between [0, t-1], w k is the filter weight, which is [0,1]; The expression of the dilated convolution model is: In the above formula, F(x s ) is the input x output by the network at time s s , k is the size of the convolution kernel, d is the dilation coefficient, (sd i ) is the sequence corresponding to the elements in the convolution kernel; The expression of the residual connection model is: o=RELU[x+F(x s )]: In the above formula, o is the output of the temporal convolutional network, RELU[] is the rectified linear unit function operation, x is the model input, and F(x s ) is the network output at time s.

3. According to claim 2, a method for predicting distributed photovoltaic power in low-voltage areas based on a temporal convolutional network is characterized in that: In S1, the compression and excitation network includes a compression network and an excitation network, the compression network is used to compress global spatial information, perform feature learning in the channel dimension, and analyze the importance difference of each channel, and the excitation network is used to assign different weights to each channel, thereby forming a channel attention module; The cth element of the data z obtained by the compression network is calculated by the following formula: In the above formula, H, W, and C are data dimension coefficients, γ and b are constants, and γ is set to 2 and b is set to 1; The excitation network adopts an S-type activation gating mechanism: S=F ex (z,W); In the above formula, F ex (·) is the ReLU function; The excitation network obtains the final output by rescaling the result using activation s: In the above formula, F scale (u C ,S C ) is a scalar S C With the feature map u C The channel multiplication between .

4. According to claim 1, a method for predicting distributed photovoltaic power in low-voltage areas based on a temporal convolutional network is characterized in that: In S2, the weather data corresponding to the historical photovoltaic output result is a discrete value, and the specific operation of the missing data filling is to fill the characteristic variables whose median and mean differ by less than 10% by the mean, and to fill the characteristic variables whose median and mean differ by more than 10% by the median; The specific operations of data standardization are: The standard score is used to standardize the data, and the expression is: In the above formula, Z is the standardized data, x is the initial data, μ is the mean of the overall data, and σ is the standard deviation of the overall data; The specific operation of the outlier processing is to only retain the values ​​distributed in [μ-3σ,μ+3σ]; The specific operation of reducing the dimension of the data set includes the following steps: Assume that the original data has m samples X 1 ,X 2 ,…,X m , each sample has n-dimensional features Each feature x j All have their own eigenvalues. After standardizing the data, we get the covariance matrix. In the case of three-dimensional data, the expression of the covariance matrix is: Find the eigenvalue λ and eigenvector u of the covariance matrix C; Project the original features onto the selected feature vector to obtain the new k-dimensional features after dimensionality reduction, which is expressed as: In summary, for each sample X i , will be replaced by the original becomes To achieve the purpose of dimensionality reduction; The specific operation of the fuzzy C-means clustering includes the following steps: Design the number of clusters and the hyperparameter m, and initialize the membership matrix with random numbers in [0, 1] Where n is the number of data objects, m is the number of clusters, and u ij For data object x i The degree of membership to the cluster center; The updated cluster center C j The expression is: The expression for updating the membership matrix U is: The iteration termination condition is that the maximum number of iterations is reached or the error exceeds a threshold. The error expression is:

5. A distributed photovoltaic power prediction system for low-voltage areas based on a temporal convolutional network, characterized by: The system includes: a SENet-TCN model construction module, a FCM algorithm model construction module and a data prediction module; The SENet-TCN model building module is used to build a temporal convolutional network and a compression and excitation network, and introduce the compression and excitation network into the temporal convolutional network to form a SENet-TCN model; The FCM algorithm model construction module is used to construct the FCM algorithm model through missing data filling, data standardization, outlier processing, data set dimension reduction and fuzzy C-means clustering. The FCM algorithm model is used to process the historical photovoltaic output results and the corresponding weather data; The data prediction module is used to input the historical photovoltaic output results and the corresponding weather into the FCM algorithm model to obtain the photovoltaic output data set of the distributed photovoltaic units in the low-voltage area classified by weather type, and iteratively train the SENet-TCN model through the obtained photovoltaic output data set. The photovoltaic output data set of the distributed photovoltaic units in the low-voltage area classified by weather type and the multivariate meteorological factors of the forecast day are input into the trained SENet-TCN model to predict the photovoltaic power generation data of the low-voltage area on the forecast day.

6. According to claim 5, a distributed photovoltaic power prediction system for low-voltage areas based on a temporal convolutional network is characterized in that: The temporal convolutional network includes a causal convolutional model, a dilated convolutional model and a residual connection model. The expression of the causal convolutional model is: x t =∑(x t-k *w k ),k∈[0,t-1]; In the above formula, P(x it ) is the prediction probability, T is the total time, Π is the orthogonal operation, represents the meteorological feature sequence of the i-th sample, with a total of m meteorological features, x t =∑(x t-k *w k ), k∈[0,t-1] is the random evolution layer, x t is the meteorological characteristic sequence of the i-th sample at time t, k is an integer between [0, t-1], w k is the filter weight, which is [0,1]; The expression of the dilated convolution model is: In the above formula, F(x s ) is the input x output by the network at time s s , k is the size of the convolution kernel, d is the dilation coefficient, (sd i ) is the sequence corresponding to the elements in the convolution kernel; The expression of the residual connection model is: o=RELU[x+F(x s )]: In the above formula, o is the output of the temporal convolutional network, RELU[] is the rectified linear unit function operation, x is the model input, and F(x s ) is the network output at time s.

7. The low-voltage distributed photovoltaic power prediction system based on a time convolutional network according to claim 6 is characterized by: The compression and excitation network construction module includes a compression network and an excitation network, wherein the compression network is used to compress global spatial information, perform feature learning in the channel dimension, and analyze the importance difference of each channel, and the excitation network is used to assign different weights to each channel, thereby forming a channel attention module; The cth element of the data z obtained by the compression network is calculated by the following formula: In the above formula, H, W, and C are data dimension coefficients, γ and b are constants, and γ is set to 2 and b is set to 1; The excitation network adopts an S-type activation gating mechanism: S=F ex (z,W); In the above formula, F ex (·) is the ReLU function; The excitation network obtains the final output by rescaling the result using activation s: In the above formula, F scale (u C ,S C ) is a scalar S C With the feature map u C The channel multiplication between .

8. According to claim 7, a distributed photovoltaic power prediction system for low-voltage areas based on a temporal convolutional network is characterized in that: In the FCM algorithm model, the weather data corresponding to the historical photovoltaic output results are discrete values, and the specific operation of filling the missing data is: The feature variables whose median and mean differ by less than 10% are filled by the mean, and the feature variables whose median and mean differ by more than 10% are filled by the median; The specific operation of data standardization is: standardize the data using standard scores, and the expression is: In the above formula, Z is the standardized data, x is the initial data, μ is the mean of the overall data, and σ is the standard deviation of the overall data; The specific operation of the outlier processing is to only retain the values ​​distributed in [μ-3σ,μ+3σ]; The specific operation of reducing the dimension of the data set includes the following steps: Assume that the original data has m samples X 1 ,X 2 ,…,X m , each sample has n-dimensional features Each feature x j All have their own eigenvalues. After standardizing the data, we get the covariance matrix. In the case of three-dimensional data, the expression of the covariance matrix is: Find the eigenvalue λ and eigenvector u of the covariance matrix C; Project the original features onto the selected feature vector to obtain the new k-dimensional features after dimensionality reduction, which is expressed as: In summary, for each sample X i , will be replaced by the original becomes To achieve the purpose of dimensionality reduction; The specific operation of the fuzzy C-means clustering includes the following steps: Design the number of clusters and the hyperparameter m, and initialize the membership matrix with random numbers in [0, 1] Where n is the number of data objects, m is the number of clusters, and u ij For data object x i The degree of membership to the cluster center; The updated cluster center C j The expression is: The expression for updating the membership matrix U is: The iteration termination condition is that the maximum number of iterations is reached or the error exceeds a threshold. The error expression is:

9. A distributed photovoltaic power prediction device for low-voltage areas based on a time convolutional network, characterized in that: The device comprises a processor and a memory; The memory is used to store computer program code and transmit the computer program code to the processor; The processor is used to execute the low-voltage distributed photovoltaic power prediction method based on a time convolutional network according to any one of claims 1 to 5 according to the instructions in the computer program code.

10. A storage medium for power balancing of distributed resources, having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for predicting distributed photovoltaic power in a low-voltage area based on a time convolutional network described in any one of claims 1 to 5 is implemented.

Images