阅读说明：本技术 基于智能算法的电子式互感器误差短时预测方法 (Error short-time prediction method for electronic transformer based on intelligent algorithm ) 是由 杨爱超 胡琛 吴宇 李敏 李东江 李嫣 邓小松 于 2020-03-31 设计创作，主要内容包括：本发明涉及一种基于智能算法的电子式互感器误差短时预测方法,属于电能计量领域。该方法包括以下步骤：S1：电子式互感器数据采集；S2：对数据集进行预处理操作；S3：引入箱线图将数据可视化；S4：建立Prophet模型对电子式互感器误差进行预测。本发明利用智能电网中电子式互感器计量数据,通过Prophet方法,对电子式互感器误差发展态势进行分析及仿真研究,以保证电子式互感器的安全稳定运行。本发明充分利用智能电网中电子式互感器计量数据,基于Prophet模型,提出了无标准器的条件下对电子式互感器误差发展态势进行短时预测的方法,以保证电子式互感器的安全稳定运行,可适应电子式互感器规模化应用场景下误差状态分析需求。(The invention relates to an electronic transformer error short-time prediction method based on an intelligent algorithm, and belongs to the field of electric energy metering. The method comprises the following steps: s1: collecting data of the electronic transformer; s2: preprocessing the data set; s3: a drop box plot visualizes the data; s4: and establishing a Prophet model to predict the error of the electronic transformer. According to the method, the error development situation of the electronic transformer is analyzed and simulated by using the metering data of the electronic transformer in the intelligent power grid through a Prophet method, so that the safe and stable operation of the electronic transformer is ensured. The method disclosed by the invention fully utilizes the metering data of the electronic transformer in the intelligent power grid, and provides a method for short-time prediction of the error development situation of the electronic transformer under the condition of no standard device based on a Prophet model, so that the safe and stable operation of the electronic transformer is ensured, and the method can meet the error state analysis requirement of the electronic transformer under the large-scale application scene.)

1. An electronic transformer error short-time prediction method based on an intelligent algorithm is characterized by comprising the following steps: the method comprises the following steps:

s1: collecting data of the electronic transformer;

s2: preprocessing the data set;

s3: a drop box plot visualizes the data;

s4: and establishing a Prophet model to predict the error of the electronic transformer.

2. The intelligent algorithm-based electronic transformer error short-time prediction method according to claim 1, characterized in that: the S1 specifically includes:

the data of the electronic transformer is derived from a 220kV transformer error state monitoring platform; in the platform, electromagnetic current transformers and hollow coil current transformers are installed on a line at intervals, the accuracy of 2 transformers is 0.2 grade, and the rated current is 600A; the rated output of the electromagnetic mutual inductor is 5A, the rated secondary capacity is 25 V.A, the specific difference under rated current is 0.08%, and the angular difference is 6'; the specific difference of the air-core coil current transformer under rated current is 0.12%, the angular difference is 6', and the output follows IEC61850-9-2 protocol; the signal acquisition unit converts an analog signal of the electromagnetic current transformer into a digital signal; the data processing unit receives the output signal of the signal acquisition unit and the sampling value message data, meets the 0.05-level accuracy requirement, and obtains an error comparison result by taking the output of the electromagnetic current transformer as a standard.

3. The intelligent algorithm-based electronic transformer error short-time prediction method according to claim 1, characterized in that: the S2 specifically includes:

setting upper and lower boundaries of data, and deleting abnormal values;

determining the occupation ratio of missing values in the data set, adopting a Lagrange interpolation method or a Newton interpolation method for interpolation when the occupation ratio is large, and not processing when the occupation ratio is small;

thirdly, because the acquisition frequency of the signal acquisition unit of the mutual inductor is high, more than 260 tens of thousands of data are contained in a data sample under one month, and the model is difficult to solve by directly predicting; the average value of the hourly data is obtained by adopting the formula (1), and the average value is used as a training sample, so that the model prediction complexity is reduced, and the prediction efficiency is improved;

in the formula (1)Is the average value of the data per hour,is a data sample per second.

4. The intelligent algorithm-based electronic transformer error short-time prediction method according to claim 1, characterized in that: the S3 specifically includes:

a drop box plot visualizes the data; the box plot utilizes five statistics in the data: the minimum value, the first quartile, the median, the third quartile and the maximum value describe the data.

5. The intelligent algorithm-based electronic transformer error short-time prediction method according to claim 1, characterized in that: the S4 specifically includes:

establishing a Prophet model

The Prophet model adopts a generalized addition model to fit and predict a function, as shown in formula (2);

y(t)＝g(t)+s(t)+h(t)+_t(2)

the model divides the time series into a superposition of 3 parts: trend, period, mutation terms;

if the time series growth trend is non-linear growth, the term g (t) is a logistic regression function; if the growth trend is linear, the term g (t) is a linear function; adopting a linear growth model;

g(t)＝(k+a(t)^T)t+(m_p+a(t)^Tγ) (3)

where k is the growth rate and is the growth rate adjustmentValue, m_pIs an offset parameter, gamma is set to-s_{j j}So that the function is continuous;

(t) is used to represent periodic variation and consists of Fourier series;

in the formula: p represents the period of the target sequence; c. C_nIs the coefficient parameter to be estimated; n is the number of the set approximate items;

the mutation term h (t) represents the change caused by special reasons;

h(t)＝Z(t)k (6)

z(t)＝[l(t∈D₁,...,l(t∈D_L))](7)

in the formula: d_iA date indicating that there is a particular change in the time series; l is used to indicate whether time t belongs to D_iThe indication function of (1); the parameter k follows normal distribution k-N (0, v);

④_twhich is a noise term that follows a normal distribution, and is used to represent random, unpredictable fluctuations.

Technical Field

The invention belongs to the field of electric energy metering, and relates to an electronic transformer error short-time prediction method based on an intelligent algorithm.

Background

In the actual operation process of the electronic transformer, accurate and quick diagnosis can be carried out when the metering error difference of the electronic transformer is normal, and further, the deterioration trend of the metering error of the electronic transformer is required to be predicted timely, so that related operation maintenance personnel can overhaul and maintain the work in time, the loss of electric energy metering is reduced, the normal operation of the measurement and control protection device is ensured, and the method has important significance for the safe, stable and economic operation of an electric power system.

Intelligent algorithms widely used in prediction include neural networks, Prophet models, and the like. Neural networks are used in a variety of contexts, such as power system load prediction, insulation oil breakdown voltage prediction, and rail traffic prediction, among others. The Prophet model also has good effects in the aspects of bank reserve deposit prediction, product sales prediction, daily air quality index prediction and the like.

However, research on the prediction of the metering state of the electronic transformer is still in a blank state at home and abroad at present, and because the state prediction of secondary output information of the electronic transformer is influenced by random fluctuation of the state of a power grid, an effective method for predicting the variation trend of the metering error performance of the electronic transformer is not available.

Disclosure of Invention

In view of the above, the present invention provides an electronic transformer error short-time prediction method based on an intelligent algorithm. The method comprises the steps of analyzing and simulating the error development situation of the electronic transformer by using electronic transformer metering data in the smart grid through a Prophet method, so as to ensure safe and stable operation of the electronic transformer.

In order to achieve the purpose, the invention provides the following technical scheme:

an electronic transformer error short-time prediction method based on an intelligent algorithm comprises the following steps:

s1: collecting data of the electronic transformer;

s2: preprocessing the data set;

s3: a drop box plot visualizes the data;

s4: and establishing a Prophet model to predict the error of the electronic transformer.

Preferably, the S1 is specifically:

the data of the electronic transformer is derived from a 220kV transformer error state monitoring platform; in the platform, electromagnetic current transformers and hollow coil current transformers are installed on a line at intervals, the accuracy of 2 transformers is 0.2 grade, and the rated current is 600A; the rated output of the electromagnetic mutual inductor is 5A, the rated secondary capacity is 25 V.A, the specific difference under rated current is 0.08%, and the angular difference is 6'; the specific difference of the air-core coil current transformer under rated current is 0.12%, the angular difference is 6', and the output follows IEC61850-9-2 protocol; the signal acquisition unit converts an analog signal of the electromagnetic current transformer into a digital signal; the data processing unit receives the output signal of the signal acquisition unit and the sampling value message data, meets the 0.05-level accuracy requirement, and obtains an error comparison result by taking the output of the electromagnetic current transformer as a standard.

Preferably, the S2 is specifically:

setting upper and lower boundaries of data, and deleting abnormal values;

determining the occupation ratio of missing values in the data set, adopting a Lagrange interpolation method or a Newton interpolation method for interpolation when the occupation ratio is large, and not processing when the occupation ratio is small;

thirdly, because the acquisition frequency of the signal acquisition unit of the mutual inductor is high, more than 260 tens of thousands of data are contained in a data sample under one month, and the model is difficult to solve by directly predicting; the average value of the hourly data is obtained by adopting the formula (1), and the average value is used as a training sample, so that the model prediction complexity is reduced, and the prediction efficiency is improved;

in the formula (1)Is the average value of the data per hour,is a data sample per second.

Preferably, the S3 is specifically:

a drop box plot visualizes the data; the box plot utilizes five statistics in the data: the minimum value, the first quartile, the median, the third quartile and the maximum value describe the data.

Preferably, the S4 is specifically:

establishing a Prophet model

The Prophet model adopts a generalized addition model to fit and predict a function, as shown in formula (2);

y(t)＝g(t)+s(t)+h(t)+_t(2)

the model divides the time series into a superposition of 3 parts: trend, period, mutation terms;

if the time series growth trend is non-linear growth, the term g (t) is a logistic regression function; if the growth trend is linear, the term g (t) is a linear function; adopting a linear growth model;

g(t)＝(k+a(t)^T)t+(m_p+a(t)^Tγ) (3)

where k is the growth rate, the growth rate adjustment value, and m_pIs an offset parameter, gamma is set to-s_{j j}So that the function is continuous;

(t) is used to represent periodic variation and consists of Fourier series;

in the formula: p represents the period of the target sequence; c. C_nIs the coefficient parameter to be estimated; n is the number of the set approximate items;

the mutation term h (t) represents the change caused by special reasons;

h(t)＝Z(t)k (6)

z(t)＝[l(t∈D₁,...,l(t∈D_L))](7)

in the formula: d_iA date indicating that there is a particular change in the time series; l is used to indicate whether time t belongs to D_iThe indication function of (1); the parameter k follows normal distribution k-N (0, v);

④_twhich is a noise term that follows a normal distribution, and is used to represent random, unpredictable fluctuations.

The invention has the beneficial effects that:

the method makes up the shortages in the aspect of the forecasting research of the metering state of the electronic transformer. The following conclusions can be obtained through the analysis of the results of the examples: the fitting degree of the BP neural network is in negative correlation with the prediction effect, and the Prophet model can obtain a better fitting effect under the condition of ensuring the prediction accuracy; through image results and quantitative analysis and comparison, the prediction effect of the Prophet model on the error of the electronic transformer is obviously superior to that of a BP neural network model; the error cycle variation of the electronic transformer is mainly based on daily fluctuation, and the reason of the error cycle variation is probably due to the change of the error change rate caused by temperature change in a day.

In conclusion, the method for predicting the error development situation of the electronic transformer in short time under the condition without a standard device is provided by fully utilizing the metering data of the electronic transformer in the intelligent power grid and based on the Prophet model, so that the safe and stable operation of the electronic transformer is ensured, and the error state analysis requirement under the large-scale application scene of the electronic transformer can be met.

Additional advantages, objects, and features of the invention will be set forth in part in the description which follows and in part will become apparent to those having ordinary skill in the art upon examination of the following or may be learned from practice of the invention. The objectives and other advantages of the invention may be realized and attained by the means of the instrumentalities and combinations particularly pointed out hereinafter.

Drawings

For the purposes of promoting a better understanding of the objects, aspects and advantages of the invention, reference will now be made to the following detailed description taken in conjunction with the accompanying drawings in which:

FIG. 1 is a transformer error condition monitoring platform;

FIG. 2 is an original transformer angle difference;

FIG. 3 is a diagram of the angular difference of the transformer after pretreatment;

FIG. 4 is an angular difference box plot;

FIG. 5 is a BP neural network topology;

FIG. 6 is a transformer angular difference prediction; (a) predicting the result for the neural network; (b) predicting results for the Prophet model;

fig. 7 shows the angular difference trend and the periodic component of the transformer.

Detailed Description

The embodiments of the present invention are described below with reference to specific embodiments, and other advantages and effects of the present invention will be easily understood by those skilled in the art from the disclosure of the present specification. The invention is capable of other and different embodiments and of being practiced or of being carried out in various ways, and its several details are capable of modification in various respects, all without departing from the spirit and scope of the present invention. It should be noted that the drawings provided in the following embodiments are only for illustrating the basic idea of the present invention in a schematic way, and the features in the following embodiments and examples may be combined with each other without conflict.

1 data processing

1.1 electronic mutual inductor data acquisition

The electronic transformer data used by the invention is freely sourced from the 220kV transformer error state monitoring platform shown in figure 1. In the platform, electromagnetic current transformers and hollow coil current transformers are installed on a line at intervals, the accuracy of 2 transformers is 0.2 grade, and the rated current is 600A. The rated output of the electromagnetic mutual inductor is 5A, the rated secondary capacity is 25 V.A, the specific difference under rated current is 0.08%, and the angular difference is 6'; the specific difference of the air-core coil current transformer under rated current is 0.12%, the angular difference is 6', and the output conforms to the IEC61850-9-2 protocol. The signal acquisition unit converts the analog signal of the electromagnetic current transformer into a digital signal. The data processing unit receives the output signal of the signal acquisition unit and the sampling value message data, meets the 0.05-level accuracy requirement, and obtains an error comparison result by taking the output of the electromagnetic current transformer as a standard.

The invention takes the angular difference data as an example, the original mutual inductor angular difference data is the actual operation data of the mutual inductors 2018/7/1-2018/7/31, and the sampling time interval is 1s, as shown in fig. 2.

1.2 data preprocessing

Since electronic transformer angular difference data is generally irregular and noisy, the following preprocessing operations need to be performed on the data set before the transformer angular difference is predicted using an intelligent algorithm:

setting upper and lower boundaries of angle difference data, and deleting abnormal values;

determining the occupation ratio of missing values in the data set, adopting a Lagrange interpolation method or a Newton interpolation method for interpolation when the occupation ratio is large, and not processing when the occupation ratio is small;

and thirdly, because the acquisition frequency of the transformer acquisition unit is high, more than 260 tens of thousands of data are contained in a data sample under one month, and the model can be very difficult to solve by directly predicting. The average value of the angular difference per hour can be obtained by adopting the formula (1), and the average value is used as a training sample, so that the model prediction complexity is reduced, and the prediction efficiency is improved.

In the formula (1)Is the average value of the angular difference per hour,sampled as angular difference per second.

After the preprocessing operation, the angle difference data of the transformer are obtained as shown in fig. 3.

1.3 data visualization

In order to visually see whether the angular difference distribution has symmetry, the dispersion degree of the angular difference and the approximate development trend of the angular difference, the invention introduces a box diagram to visualize the angular difference data. The box plot mainly utilizes five statistics in the data: the minimum value, the first quartile, the median, the third quartile and the maximum value are used for describing data, and the method is a common method for data visualization. The preprocessed data sets are divided into 31 data sets at intervals of 24h, each data set comprises 96 data sets corresponding to 7/1-7/31 days, and a box line graph is drawn as shown in fig. 4.

As can be seen from fig. 4, in 7 months, the angular differences of the transformers are distributed uniformly, the dispersion degree is stable, and the abnormal values are more only for a few days (3, 5, 6, 7, 8, 22), and the distribution is concentrated. On days 13-21, the median line was centered, the angle difference was normally distributed, the medians at the beginning and end of the month shifted upward, and the angle difference was distributed eccentrically. The overall trend of the mutual inductor angle difference in the month is fluctuated and reduced can be seen from the change condition of the median line.

2 prediction model

2.1BP neural network rationale

The bp (back propagation) neural network is a multi-layer feedforward neural network trained according to an error back propagation algorithm. According to the Kolmogorov theorem, for a 3-layer BP neural network, any nonlinear continuous function can be approximated on a closed set with any precision as long as the number of hidden nodes is enough.

Fig. 5 shows a general BP neural network topology, which includes an input layer, a hidden layer, and an output layer, where each layer is composed of a certain number of neurons, the neurons in the upper and lower layers are fully interconnected, the connection strength is expressed by a weight value, and the neurons in the same layer are not connected.

In FIG. 5, x₁,...,x_l-1Representing the input, assuming the hidden layer has m neurons, their outputs are x', x₁′,...,x′_m-1The output layer has n neurons for outputting y, y₁,...,y_n-1The weight from the input layer to the hidden layer is w_ijThe threshold value is theta_j(ii) a Weight from hidden layer to output layer is w'_jkThe threshold value is theta'_kThe neuron outputs at each layer are:

wherein f (-) is an activation function, and the S-shaped excitation function is shown as formula (3). The input value is processed by each layer of activation function to obtain an output value, the network corrects each weight value according to the error reduction direction of the output value and the actual value, and the network continuously iterates to enable the output response value to approach the actual value with set precision.

2.2 the basic principle of the Prophet model of the invention

The Prophet model uses a generalized additive model to fit and predict a function, as shown in equation (4).

y(t)＝g(t)+s(t)+h(t)+_t(4)

Here, the model divides the time series into a superposition of 3 parts: trend, period, mutation terms.

If the time series growth trend is non-linear growth, the term g (t) is a logistic regression function; if the growth trend is linear, the term g (t) is a linear function. The increasing trend is linear as can be seen from the angle difference box line of the transformer, so that the invention adopts a linear increasing model.

g(t)＝(k+a(t)^T)t+(m_p+a(t)^Tγ) (5)

Where k is the growth rate, the growth rate adjustment value, and m_pIs an offset parameter, gamma is set to-s_{j j}So that the function is continuous.

And (t) is used to represent the periodic variation and consists of a Fourier series.

In the formula: p represents the period of the target sequence; c. C_nIs the coefficient parameter to be estimated; n is the number of the set approximate terms, and the larger N is, the more complex periodicity can be fitted, but the good filtering effect may not be achieved. The setting of N needs to be considered in conjunction with p, which is set to 7 and N to 3 for each seasonal season.

The mutation term h (t) represents the change caused by special reasons.

h(t)＝Z(t)k (8)

z(t)＝[l(t∈D₁,...,l(t∈D_L))](9)

In the formula: d_iA date indicating that there is a particular change in the time series; l is used to indicate whether time t belongs to D_iThe indication function of (1); the parameter k follows a normal distribution k-N (0, v), with the default value of v set to 10, with larger v allowing the model to accommodate larger fluctuations and smaller values suppressing the influence.

④_tWhich is a noise term that follows a normal distribution, and is used to represent random, unpredictable fluctuations.

3 results of the experiment

3.1 parameter settings

Before prediction by using a neural network, the preprocessed data needs to be standardized, and the Z-score standardization method is adopted in the text according to the formula (10).

In the formula, mu is the mean value of the mutual inductor angular difference data set, and sigma is the standard deviation of the mutual inductor angular difference.

The influence of relevant parameters of the neural network on the final fitting precision and the prediction precision is very large, and the error convergence of the neural network is easy to fall into local optimization due to unreasonable parameter setting, so that the precision requirement cannot be met; meanwhile, the target fitting accuracy is set to be too high, so that the neural network is over-fitted, and the prediction accuracy is reduced.

In the neural network training, the neural network parameters can be continuously modified through the loop statements to obtain multiple groups of experimental results, so that the optimal parameter setting is obtained through comparison. The optimal predicted effect parameter settings obtained after multiple cycles of comparative tests are shown in table 1.

TABLE 1 neural network parameter setting Table

Before the angular difference of the mutual inductor is predicted by using the Prophet model, preprocessed data needs to be adjusted according to the requirements of the Prophet data, and the preprocessed data are changed into two rows of dataframes with fixed names: ds and y, where ds is the date and y is the sequence. And after the input data format is processed, setting the parameters of the Prophet model according to the table 2.

TABLE 2 Prophet model parameter settings

3.2 predictive evaluation criteria

In order to improve the accuracy of the prediction model, the data of the 2018/7 months are divided into a training set and a testing set according to a 7:3 ratio. Wherein, the training set is composed of 7/1-7/21 data, and the testing set is composed of 7/22-7/31 data. In order to evaluate the accuracy of the prediction model, a verification set is formed by actual operation data of 2018/8/1-2018/8/8 mutual inductors, and the prediction data and the verification set are compared by taking the average absolute error (MAE) and the average absolute percentage error (MAPE), the Mean Square Error (MSE) and the Root Mean Square Error (RMSE) as accuracy measurement indexes.

3.3 predictive results analysis

The results of predicting the angular difference of the transformer by using the BP neural network and the Prophet model are respectively shown in fig. 6(a) and 6(b), and the index pair ratio of the two is shown in table 3.

TABLE 3 comparison of fitting and predicted results

In fig. 6(a), since overfitting of the neural network adversely deteriorates the prediction effect, the neural network fitting curve and the actual angular difference curve in the prediction result do not completely fit to each other by appropriate parameter settings. In terms of prediction results, the neural network only predicts the approximate change trend of the mutual inductor angular difference in the future 8 days, and although the average absolute error and the root mean square error are not high, the prediction error in each day is very large.

As can be seen from FIG. 6(b), the Prophet prediction model has a high degree of fitting to the angular difference data and a strong inclusion for the mutation points and outliers. Meanwhile, the Prophet model is very good at predicting time sequences with obvious periodicity, so that a prediction curve of the Prophet model not only accords with the development trend of angular difference, but also has low indexes such as average absolute error, root mean square error and the like, and the prediction error per day is very small and is very close to an actual curve. Therefore, for the transformer angular difference prediction, the Prophet model is obviously superior to the neural network model.

The Prophet model is used for fitting the time sequence according to the formula (4), the angular difference of the mutual inductor is predicted by the Prophet model, and besides a prediction result, the development trend and the periodic fluctuation condition of the angular difference within a period of time can be obtained, as shown in fig. 7.

The trend component in fig. 7 represents the transformer development and prediction trend at 7-month angle difference, and the blue shading represents the confidence range of the prediction. The Weekly and daily components represent the Weekly and daily fluctuation of the transformer angular difference. Comparing the fluctuation amplitude of each component, the trend term (-0.02,0.04), the cycle fluctuation (-0.0025,0.015) and the daily fluctuation (-0.03, 0.03), it can be found that the angular difference change of the mutual inductor is mainly the daily fluctuation change. In the daily component, it can be seen that the angular difference is at a very high level from 9 pm to 8 pm and at a very low level from 10-6 pm, which may be due to the difference in error rate between the air coil transformer and the electromagnetic current transformer caused by the temperature rise, resulting in significant daily fluctuation of the final angular difference data.

Finally, the above embodiments are only intended to illustrate the technical solutions of the present invention and not to limit the present invention, and although the present invention has been described in detail with reference to the preferred embodiments, it will be understood by those skilled in the art that modifications or equivalent substitutions may be made on the technical solutions of the present invention without departing from the spirit and scope of the technical solutions, and all of them should be covered by the claims of the present invention.