A Voltage Situation Awareness Method Based on CNN-GRU

Abstract

The present invention discloses a voltage situation awareness method based on CNN-GRU, comprising the following steps: collecting voltage, active power and solar irradiance data; constructing a voltage time series; understanding the voltage autocorrelation; understanding the correlation of variable factors; preprocessing the time series data; constructing a voltage situation prediction model; and performing short-term voltage situation prediction. By constructing a voltage situation awareness network and using the voltage situation awareness method to mine the time series data, the present invention first collects voltage, active power and solar irradiance data through a power parameter measuring instrument and a sensor unit; then extracts the data to construct a voltage time series and conducts correlation understanding through the maximum information coefficient, etc.; and finally designs the structure of the CNN-GRU situation prediction model, realizing high-precision situation prediction of voltage.

Description

Technical Field

[0001] The present invention relates to the technical field of voltage situation awareness, and particularly relates to a voltage situation awareness method based on CNN-GRU. Background Art

[0002] With the continuous increase in the grid-connected capacity of renewable energy sources such as photovoltaic power, the possible impact on the power quality of the distribution network has also attracted extensive attention. Due to the intermittency and randomness of the output of photovoltaic power plants, power quality problems such as voltage fluctuations and flicker will occur. As one of the power quality indicators, the quality of voltage will directly affect the safe and stable operation of the distribution network.

[0003] Research on mining historical data information of the voltage quality of photovoltaic power plants to pre-perceive the voltage change trend in advance, reserve a certain time for the treatment of voltage quality, and greatly reduce the impact of voltage fluctuations of photovoltaic power plants on the distribution network has become one of the key technologies for the reliable operation of renewable energy grid connection.

[0004] The concept of situation awareness was first applied in the field of aviation and military, referring to the extraction, analysis, and prediction of various factors in a dynamic environment in a specific spatio-temporal environment. Traditional voltage situation awareness methods are insufficient in mining time series data information when facing massive input data, and cannot effectively extract high-dimensional features of the data, resulting in limited situation prediction accuracy. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention provides a voltage situation awareness method based on CNN-GRU with sufficient data information mining and high prediction accuracy.

[0006] The technical solution of the present invention to solve the above problems is: a voltage situation awareness method based on CNN-GRU, comprising the following steps:

[0007] Step 1, collect voltage, active power, and solar irradiance data: collect voltage, active power, and solar irradiance data through a power parameter measuring instrument and a sensor unit;

[0008] Step 2, construct a voltage time series: set an extraction interval time, and extract the collected voltage data as a reference value to construct a voltage time series;

[0009] Step 3, understand the voltage autocorrelation: understand the voltage autocorrelation by calculating the autocorrelation coefficient to obtain the input time step;

[0010] Step 4, understand the correlation of variable factors: understand the correlation degree between variable factors and voltage through the maximum information coefficient method to obtain the input data dimension;

[0011] Step 5, preprocess the time series data: normalize the voltage time series data;

[0012] Step 6, construct a voltage situation prediction model: The first layer of the voltage prediction model is the input layer, the second and fourth layers are convolutional layers, the third and fifth layers are pooling layers, the sixth and seventh layers are GRU layers, the eighth layer is a flattening layer, the ninth layer is a fully connected layer, and the last layer is the output layer. The voltage prediction model uses the ReLU activation function;

[0013] Step 7, conduct short-term voltage situation prediction: Combine the normalized data with the input time step and input data dimension to obtain a data set, and input the data set into the voltage situation prediction model to conduct short-term voltage situation prediction.

[0014] In the above voltage situation perception method based on CNN-GRU, in Step 2, the formula for extracting voltage data is:

[0015]

[0016] where V t_i～t_i+1 is the voltage time series in the time period from t_i to t_i+1, v1 is the voltage data at time t_i, v2 is the voltage data at time t_i+Δt, and v n is the voltage data at time t_i+(n-1)Δt. The extraction interval time Δt is 5 minutes.

[0017] In the above voltage situation perception method based on CNN-GRU, in Step 3, the formula for understanding voltage autocorrelation is:

[0018]

[0019] where C(h) is the autocorrelation coefficient, h is the time lag length, v t represents the voltage value at time t, and v t+h represents the voltage value at time t+h. v is the mean value of the entire voltage sequence, and m is the total number of time points. The maximum value of h such that C(h)≥0.93 is determined as the input time step.

[0020] In the above voltage situation perception method based on CNN-GRU, in Step 4, the formula for calculating the maximum information coefficient is:

[0021]

[0022] Among them, MIC(v; u) is the maximum information coefficient between voltage u and variable factor v, p(v, u) is the joint probability between voltage u and variable factor v, a and b are the number of grids divided in the x and y axis directions, B is the upper limit value of the number of divided grids, the value range of MIC(v; u) is [0, 1], the larger its value, the higher the correlation degree between voltage and variable factor, and the variable factors with MIC(v; u) ≥ 0.6 are included in the input data dimension.

[0023] In the above voltage situation awareness method based on CNN-GRU, in step five, the formula for normalizing voltage time series data is:

[0024]

[0025] Where v t ' represents the normalized voltage value at time t, v t represents the voltage value at time t, v max and v min respectively represent the maximum and minimum values in the voltage time series data.

[0026] In the above voltage situation awareness method based on CNN-GRU, the specific steps of step seven are as follows:

[0027] 7-1) For time series data, select the input time step as j, the input data dimension as q, and the output time step as f, that is, use j×q historical data to predict the future f data, and construct the time series data in the form of where j is generated by understanding the voltage autocorrelation, f is determined according to the control requirements, and q is generated by understanding the variable factor correlation;

[0028] For time t, V x ={V t-j , V t-j+1 ,···, V t},···, R x ={R t-j , R t-j+1 ,···, R t}, T y ={T t+1 , T t+2 ,···, T t+k}, where V x is the voltage time series, P x ,···, R x are time series such as active power and solar irradiance obtained through correlation understanding, V t-j+1 represents the sequence value of the V x sequence at time t-j+1, and T y is the output true voltage time series;

[0029] 7-2) Input V x , P x , ···, R x and other time series input voltage trend prediction models to obtain the prediction result sequence

[0030] The beneficial effects of the present invention are as follows: The present invention provides a voltage trend perception method based on CNN-GRU. By constructing a voltage trend perception network and using the voltage trend perception method to mine time series data, first, voltage, active power, and solar irradiance data are collected through a power parameter measuring instrument and a sensor unit; then, the data is extracted to construct a voltage time series, and the correlation is understood through the maximum information coefficient, etc.; finally, the CNN-GRU trend prediction model structure is designed to achieve high-precision trend prediction of voltage. Brief Description of the Drawings

[0031] Figure 1 is a flowchart of the present invention.

[0032] Figure 2 is a schematic diagram of the CNN-GRU model structure of the present invention. Detailed Embodiments

[0033] The following further describes the present invention with reference to the drawings and embodiments.

[0034] As Figure 1 shown, the voltage trend perception is divided into three stages. Voltage element extraction collects data such as voltage, active power, and solar irradiance through a power parameter measuring instrument and a sensor unit; voltage trend understanding conducts correlation understanding on the time series data extracted from voltage elements through voltage autocorrelation understanding and variable factor correlation understanding; voltage trend prediction inputs the time series data that has undergone correlation understanding and time series data preprocessing into the constructed CNN-GRU voltage trend prediction model for voltage trend prediction.

[0035] A voltage trend perception method based on CNN-GRU, and the specific process is as follows:[[]]

[0036] Step 1, collect voltage, active power, and solar irradiance data: Collect voltage, active power, and solar irradiance data from January 1, 2020 to June 11, 2020 through a power parameter measuring instrument and a sensor unit, and the data is collected at 1s intervals.

[0037] Step 2, construct a voltage time series: Set the extraction interval time, and extract the collected voltage data as a reference value to construct a voltage time series.

[0038] The formula for extracting voltage data is:[[]]

[0039]

[0040] where V t_i～t_i+1 is the voltage time series in the time period from \(t_i\) to \(t_{i + 1}\), \(v1\) is the voltage data at time \(t_i\), \(v2\) is the voltage data at time \(t_i+\Delta t\), and \(v\) n is the voltage data at time \(t_i+(n - 1)\Delta t\). The extraction interval \(\Delta t\) is 5 minutes.

[0041] Step 3, Understanding voltage autocorrelation: Understand the voltage autocorrelation by calculating the autocorrelation coefficient to obtain the input time step.

[0042] The formula for understanding voltage autocorrelation is:

[0043]

[0044] where \(C(h)\) is the autocorrelation coefficient, \(h\) is the time lag length, \(v\) t represents the voltage value at time \(t\), and \(v\) t+h represents the voltage value at time \(t + h\). is the mean value of the entire voltage sequence, \(m\) is the total number of time points, and the maximum value of \(h\) such that \(C(h)\geq0.93\) is determined as the input time step.

[0045] Step 4, Understanding the correlation of variable factors: Understand the correlation degree between variable factors and voltage through the maximum information coefficient method to obtain the input data dimension.

[0046] The formula for the maximum information coefficient is:

[0047]

[0048] where \(MIC(v;u)\) is the maximum information coefficient between voltage \(u\) and variable factor \(v\), \(p(v,u)\) is the joint probability between voltage \(u\) and variable factor \(v\), \(a\) and \(b\) are the numbers of grid divisions in the \(x\) and \(y\) axis directions, \(B\) is the upper limit value of the number of grid divisions, the value range of \(MIC(v;u)\) is \([0,1]\), the larger its value, the higher the correlation degree between voltage and variable factors, and variable factors with \(MIC(v;u)\geq0.6\) are included in the input data dimension.

[0049] Step 5, Preprocessing time series data: Normalize the voltage time series data.

[0050] The formula for normalizing the voltage time series data is:

[0051]

[0052] where \(v\) t ' represents the normalized voltage value at time \(t\), and \(v\)t The voltage value at time t, v max and v min represent the maximum and minimum values in the voltage time series data, respectively.

[0053] Step 6: Construct a voltage trend prediction model. The first layer of the voltage prediction model is the input layer, the second and fourth layers are convolutional layers, the third and fifth layers are pooling layers, the sixth and seventh layers are GRU layers, the eighth layer is a flattening layer, the ninth layer is a fully connected layer, and the last layer is the output layer. The voltage prediction model uses the ReLU activation function.

[0054] Step 7: Perform short-term voltage trend prediction. Combine the normalized data with the input time step and input data dimension to obtain a data set, and input the data set into the voltage trend prediction model to perform short-term voltage trend prediction.

[0055] For time series data, the input time step is selected as 6, the input data dimension is 3, and the output time step is 2. That is, 6×3 historical data are used to predict the data at the next 2 time points, and the time series data is constructed in the form of (V x , P x , R x , T y ). Among them, the input time step is generated by understanding the voltage autocorrelation, the output time step is determined according to the control requirements, and the input data dimension is generated by understanding the variable factor correlation.

[0056] Among them, for time t, V x ={V t-j , V t-j+1 , ···, V t}, P x ={P t-j , P t-j+1 , ···, P t}, R x ={R t-j , R t-j+1 , ···, R t}, T y ={T t+1 , T t+2 , ···, T t+k}, where V x , P x , R x are the voltage, active power, and solar irradiance time series obtained through correlation understanding. V t-j+1 represents the sequence value at time t-j+1 of the V x sequence, and T y is the true value time series of the output;

[0057] Suppose the original time series data has a total of L points and a dimension of q, which is a matrix of size L×q. At this time, the time series data is a matrix of size (L - j - f + 1)×(j×q + f).

[0058] There are a total of 46,944 time points in the time period from January 1, 2020 to June 11, 2020. After processing the time series of this time period as above, a matrix of size 46944×20 is obtained.

[0059] Among them, the data from May 12, 2020 to June 11, 2020 is used as the data for the model prediction stage, and the rest is used as the data for the model training stage.

[0060] A CNN-GRU voltage situation prediction model is established, and its model structure is as Figure 2 . The first layer is the Input input layer, and the input shape is a matrix of 38016×20; the second layer is the Conv1D convolutional layer, with 16 convolutional kernels and a convolutional kernel size of 3; the third layer is the MaxPooling1D max pooling layer, with a pooling size of 1; the fourth layer is the Conv1D convolutional layer, with 32 convolutional kernels and a convolutional kernel size of 3; the fifth layer is the MaxPooling1D max pooling layer, with a pooling size of 1; the sixth layer is the GRU layer, with 40 hidden neurons; the seventh layer is the GRU layer, with 80 hidden neurons; the eighth layer is the Flatten flattening layer; the ninth layer is the Dense fully connected layer, and finally the Output prediction result output layer, using the ReLU activation function f(x) = max(0, x), and the model uses the Adam optimizer. The final output shape is a 38016×2 matrix.

Claims

1. A voltage situation awareness method based on CNN-GRU, characterized in that It includes the following steps: Step 1, collect voltage, active power, and solar irradiance data: Collect voltage, active power, and solar irradiance data through a power parameter measuring instrument and a sensor unit; Step 2, construct a voltage time series: Set the extraction interval time, and extract the collected voltage data as a reference value to construct a voltage time series; Step 3, understand the voltage autocorrelation: Understand the voltage autocorrelation by calculating the autocorrelation coefficient to obtain the input time step; The formula for understanding the voltage autocorrelation is: where C(h) is the autocorrelation coefficient, h is the time lag length, v t represents the voltage value at time t, v t+h represents the voltage value at time t + h, is the mean value of the entire voltage sequence, m is the total time, and the maximum value of h such that C(h) ≥ 0.93 is determined as the input time step; Step 4, understand the correlation of variable factors: Understand the correlation degree between variable factors and voltage through the maximum information coefficient method to obtain the input data dimension; The formula for calculating the maximum information coefficient is: Where MIC(v; u) is the maximum information coefficient between voltage u and variable factor v, p(v, u) is the joint probability between voltage u and variable factor v, a and b are the number of grids divided in the x and y axis directions, B is the upper limit value of the number of divided grids, the value range of MIC(v; u) is [0, 1], the larger its value, the higher the correlation degree between voltage and variable factors, and the variable factors with MIC(v; u) ≥ 0.6 are included in the input data dimension; Step 5, preprocess the time series data: Normalize the voltage time series data; Step 6, construct a voltage trend prediction model: The first layer of the voltage trend prediction model is the input layer, the second and fourth layers are convolutional layers, the third and fifth layers are pooling layers, the sixth and seventh layers are GRU layers, the eighth layer is a flattening layer, the ninth layer is a fully connected layer, and the last layer is the output layer. The voltage trend prediction model uses the ReLU activation function; Step 7, perform short-term voltage trend prediction: Combine the normalized data with the input time step and input data dimension to obtain a data set, and input the data set into the voltage trend prediction model to perform short-term voltage trend prediction. The specific steps are: 7-1) For time series data, the input time step is selected as j, the input data dimension is q, and the output time step is f. That is, using j×q historical data to predict the next f data, and the time series data is constructed as follows where j is generated by understanding the voltage autocorrelation, f is determined according to the control requirements, and q is generated by understanding the variable factor correlation; Among them, for time t, V x ={V t-j , V t-j+1 , ···, V t}, ···, R x ={R t-j , R t-j+1 , ···, R t}, T y ={T t+1 , T t+2 , ···, T t+k}, where V x is the voltage time series, P x , ···, R x are the active power and solar irradiance time series obtained through correlation understanding, V t-j+1 represents the sequence value of the V x series at time t - j + 1, and T y is the output true voltage time series; 7-2) Apply V x , P x , ···, R x The time series input voltage trend prediction model obtains the predicted result sequence 2. The voltage situation awareness method based on CNN-GRU according to claim 1, characterized in that In the above Step 2, the formula for extracting voltage data is: Among which V t_i～t_i+1 is the voltage time series in the time period from \(t_i\) to \(t_{i + 1}\), \(v1\) is the voltage data at time \(t_i\), \(v2\) is the voltage data at time \(t_{i+\Delta t}\), and v n is the voltage data at time \(t_{i+(n - 1)\Delta t}\), and the extraction interval time \(\Delta t\) is 5 minutes.

3. The CNN-GRU-based voltage situation awareness method according to claim 1, wherein In the above Step 5, the formula for normalizing the voltage time series data is: where v t ' represents the normalized voltage value at time t, v t represents the voltage value at time t, v max and v min represent the maximum and minimum values in the voltage time series data, respectively.

Images