A Method for Estimating Loop-Closing Current and Assessing Risks of Medium-Voltage Lines Based on Deep Learning Network

Abstract

The present invention discloses a method for estimating the closed-loop current and risk assessment of medium-voltage lines based on a deep learning network, which includes collecting and obtaining power grid structure parameters, power generation and load level data, distributed power generation output data, historical data of relevant closed-loop feeder currents, etc., constructing neural network input and output data, training the neural network model, and then considering the real-time load prediction value and distributed power generation output prediction value to predict and risk-assess the closed-loop current of medium-voltage lines under expected boundary conditions. The present invention can enhance the adaptability to the modeling complexity of new distribution network feeders and the randomness and uncertainty of boundary conditions, improve the estimation accuracy, enhance the estimation efficiency, fill the gap in risk grading, and add a recommended function for closed-loop operation periods.

Description

Technical Field

[0001] The present invention belongs to the fields of smart grid and digital grid, and particularly relates to a method for estimating loop closing current and risk assessment of medium-voltage lines based on a deep learning network. Background Art

[0002] The distribution network is the closest to users and also the most important system in the power generation, transmission, distribution, and utilization links of the power system. In recent years, in order to reduce the user power outage time, improve power supply reliability, and reduce direct and indirect social and economic losses, most medium-voltage lines or feeders in urban areas of our country have adopted dual-end power supply. The traditional power supply method of "closed-loop design, open-loop operation, and open-loop transfer with short-term power outage" in the distribution network is also evolving into a power supply method of "closed-loop design, open-loop operation, and loop closing transfer without power outage". The "open-loop transfer with short-term power outage" method means that the switch on the power supply side of the line is first disconnected, and then the tie switch is closed to put it into operation with another power supply, that is, there will be a short-term power outage during the operation process. The "loop closing transfer without power outage" method means that when a certain bus, switch, or feeder needs to be repaired or fails, the load with multi-source power supply on the bus, switch, or feeder is transferred through the loop closing operation of the 10kV or 20kV medium-voltage feeder, and the load is transferred to other connected buses or feeders, so as to realize the load transfer operation without power outage. However, during the process of performing the loop closing operation of the 10kV or 20kV medium-voltage feeder, firstly, if there is a large difference in the voltage magnitude and phase of the two ends of the loop closing point before loop closing, it often leads to a large loop closing current, resulting in the action of the 10kV or 20kV feeder protection; secondly, the potential excessive impact current during loop closing may damage the distribution equipment and may also affect the power quality of power users, resulting in the tripping of their electrical equipment; thirdly, during the loop closing process, if a short-circuit fault occurs, it may endanger the personal safety of the operator and affect the safe and stable operation of the power grid. Fourthly, the pre-assessment of the loop closing operation without considering boundary conditions such as the feeder load level and the output of grid-connected new energy sources makes the distribution network dispatching and operation personnel unclear about which day and which time period the medium-voltage feeder has less loop closing risk. Therefore, for the construction of smart grid and digital grid, in order to reduce the average customer power outage time and improve the satisfaction of power users, with the goal of "zero power outage for planned work and zero tendency for fault power outage", it is very necessary to estimate the loop closing current of medium-voltage lines and conduct risk assessment before the loop closing operation of the distribution network feeder.

[0003] At present, actual distribution network operators generally still decide when and whether to perform loop closing operations based on experience, and then formulate loop closing power transfer operation plans. This will lead to a certain blindness in loop closing operations, and there are cases where overcurrent tripping occurs after loop closing, resulting in unsuccessful loop closing operations, reducing the power supply security and reliability of the distribution system. In recent years, there have been some research and application reports on safety analyses such as loop closing current estimation for 10kV or 20kV distribution network loop closing power transfer operations. However, there are the following limitations. First, from the perspective of the research and application object, restricted by the existing main feeder line scenarios of the distribution network, the research and application objects in the existing technologies mainly consider the distribution network feeders without new energy generation access, and do not target the feeders with new energy generation access in the new type of distribution network. Second, from the perspective of the boundary conditions for loop closing current estimation, in the existing technologies, the load level often adopts the current period load or the maximum load value, and does not make load forecasting in combination with the expected period of the feeder loop closing power transfer plan, that is, does not consider the uncertainty of feeder load and new energy generation. Third, from the perspective of the mathematical model or algorithm for loop closing current estimation, the existing technologies mainly adopt model-driven deterministic evaluation methods such as the detailed modeling method and the simplified equivalent method. Among them, the system detailed modeling method is to perform system equivalence on the high-voltage level power grids such as 10kV or 20kV, 35kV, 110kV, and 220kV on both sides of the loop closing point, and use the equivalent power grid for current calculation, and the calculation accuracy is relatively good. However, for large power systems, the model adopted by the system detailed modeling method is complex, with high requirements for parameter accuracy and real-time performance. The operation mode changes in real time, there are many devices, and the parameter maintenance volume is very large. Moreover, there are convergence problems in the power flow calculation iteration process, making this method less feasible in practical applications. The simplified equivalent method has done a lot of simplification processing to enhance the convenience of estimation and the feasibility of application. For example: A method for estimating the loop closing and opening current of a 35kV high-voltage distribution network disclosed in Chinese Patent Publication No. CN104410071A on March 11, 2015 ignores the influence factors of the operation mode change of the 220kV power grid on the lower-level power grid, does not use the traditional power flow calculation method, only adopts the basic addition, subtraction, multiplication, and division of complex numbers, adopts the maximum load, and calculates the loop closing current I θ caused by the voltage difference of 220kV, and calculates the loop closing current I S, and then add the loop currents of the above two parts to obtain the estimated loop current. Obviously, this type of method targets the loop closing of 35 kV lines. Although the calculation is simple, there may be large calculation errors for the loop closing current of medium-voltage feeders in 10 kV or 20 kV distribution networks, which is not conducive to the refined management of the loop closing operation of distribution network feeders. In addition, and more importantly, this method essentially still belongs to a deterministic model-driven method, and has poor adaptability to the modeling complexity and the randomness and uncertainty of boundary conditions such as the topological structure, load type and size, new energy power generation access type and size, and operation mode during the operation of medium-voltage feeders in the distribution network.

[0004] With the development of big data and artificial intelligence technologies, deep learning algorithms based on convolutional neural networks have the ability to efficiently extract features, and long short-term memory networks and their variant networks are specialized in dealing with time series problems, and have been relatively deeply studied and initially applied in fields such as load forecasting and distributed power generation output forecasting. Due to the uncertainty of boundary conditions such as load fluctuations and distributed power generation output and the complexity of distribution network modeling, the estimated results of traditional model-driven deterministic feeder loop closing currents are difficult to meet the needs of the construction of smart grids, digital grids, and the operation of new distribution networks. By means of artificial intelligence technology, the data-driven method based on the probability density prediction of deep learning networks can describe the possible future fluctuation range, uncertainty, and risks of the loop closing current of 10 kV or 20 kV feeders in the new distribution network, and can provide more decision-making information for loop closing operators. Summary of the Invention

[0005] The object of the present invention is to provide a medium-voltage line loop closing current estimation and risk assessment method based on a deep learning network, which can enhance the adaptability to the modeling complexity and the randomness and uncertainty of boundary conditions of new distribution network feeders, improve the estimation accuracy, enhance the estimation efficiency, fill in the missing risk grading, and add a loop closing operation time period recommendation function, so as to overcome the above-mentioned shortcomings.

[0006] 1. A medium-voltage line loop closing current estimation and risk assessment method based on a deep learning network of the present invention includes the following steps:

[0007] (1) Data acquisition: Obtain historical data in Supervisory Control And Data Acquisition (SCADA), etc., including grid structure parameters and operation mode data of transmission lines and transformers in the distribution network dispatching department, power user and grid operation data of the marketing department and production technology department, etc., power generation and load level data, distributed energy output, historical data of relevant loop closing feeder currents, etc.;

[0008] (2) Data preprocessing: Clean the historical data, including deleting outliers, filling in missing values by interpolation method, and normalizing the data to the range of 0 to 1 to remove the unit limitation of the data and convert it into dimensionless pure numerical values;

[0009] (3) Neural network training and generation: Process the cleaned data by the method of time dislocation to form training data, generate the input format suitable for the model structure, adjust the network structure parameters such as the input layer, hidden layer, fully connected layer, convolution step, and convolution kernel of the network to achieve better prediction results. Design a deep learning model based on Convolutional Neural Network (CNN) and long short-term memory network (LSTM) to implement the training of the closed-loop current prediction model and generate its prediction model;

[0010] (4) Determination of closed-loop boundary conditions: Include the operator selecting the line to be closed-loop, specifying the operation time range of the medium-voltage line closed-loop plan, conducting load prediction and distributed power generation output data prediction for this planned operation time range, and determining the operation modes of the high-voltage and medium-voltage distribution networks in the expected time period;

[0011] (5) Closed-loop current probability prediction: Use the prediction model trained in step (3) to carry out closed-loop current probability prediction. Use quantile regression to predict the predicted values under different quantiles of the closed-loop current, and then use kernel density estimation to obtain the probability density distribution of the measured closed-loop current and the cumulative distribution curve of the closed-loop current probability. Calculate the closed-loop safety evaluation index according to the closed-loop current probability distribution curve;

[0012] (6) Output result and display: Compare the closed-loop current with the maximum allowable carrying capacity of the feeder and the setting value of the current protection, comprehensively analyze each risk index of the closed-loop operation, and finally obtain the risk quantification value of the closed-loop operation.

[0013] For the above method for estimating the closed-loop current and risk assessment of medium-voltage lines based on a deep learning network, the missing values are filled in step (2) through the following calculation formula:

[0014]

[0015] In the formula: t represents the moment of the missing data, unit: min; i represents the nearest moment with valid value before the moment t, unit: min; j represents the nearest moment with valid value after the moment t, unit: min; k t represents the missing value at the moment t to be filled; k i represents the data value at the moment i; k j represents the data value at the moment j; k t and ki , k j The data values represented can be voltage, current, active power, reactive power, etc., with units of kV, A, kW, kVar, etc.

[0016] For the above method for estimating loop closing current and risk assessment of medium - voltage lines based on a deep - learning network, in step (2), the data is normalized and transformed into the range of 0 to 1, removing the unit limitation of the data and converting it into a dimensionless pure - numerical calculation formula as follows:

[0017]

[0018] In the formula: k represents the normalized output value, with the unit being dimensionless; k org represents the data value to be processed; k min , k max represent the minimum and maximum values in the feature where the data value to be processed is located, respectively; k org , k min , k max The data values represented can be voltage, current, active power, reactive power, etc., with units of kV, A, kW, kVar.

[0019] For the above method for estimating loop closing current and risk assessment of medium - voltage lines based on a deep - learning network, in step (3), the deep - learning model based on CNN and LSTM consists of an input layer, a CNN layer, a pooling layer, an LSTM layer, a fully - connected layer, and a pooling layer. The optimizer selects the Adam optimizer, and a loss function as shown in formula (3) is constructed:

[0020]

[0021] In the formula: N is the number of test samples, and i is the serial number; X i is the input value for the prediction of the i - th loop closing current; y i is the actual value of the i - th loop closing current prediction; τ represents the quantile; f(W(τ), b(τ), X i ) is the predicted value of the loop closing current at different quantiles output by the i - th loop closing current CNN - LSTM network; W(τ), b(τ) represent the network parameters of the model at different quantiles; ρ τ (x) is the loss function of the loop closing current prediction model at different quantiles, and its calculation formula is as follows:

[0022]

[0023] In the formula: x represents the function variable. All variables in formula (3) and formula (4) are dimensionless.

[0024] The above method for estimating the loop closing current and risk assessment of medium-voltage lines based on a deep learning network, in which the optimal parameters of the model are determined through the following steps in step (3):

[0025] (a) Construct a training sample set and convert it into the input-output format of the neural network model;

[0026] (b) Divide the acquired data into a training set and a test set;

[0027] (c) Initially determine the number of convolutional layers, the number of LSTM layers, and the fully connected fusion layer through multiple trainings;

[0028] (d) After determining the initial number of structural layers, adjust the network parameters of the model structure;

[0029] (e) Finally, re-determine the final structural parameters of the model, and evaluate the performance of the model by checking the evaluation indicators and the number of iterations;

[0030] The above method for estimating the loop closing current and risk assessment of medium-voltage lines based on a deep learning network, in which the specific steps for determining the loop closing boundary conditions in step (4) are as follows:

[0031] (a) The operator selects the line to be looped;

[0032] (b) Specify the operation time range of the medium-voltage loop closing plan;

[0033] (c) Conduct load forecasting and distributed power generation output forecasting for this planned operation time range;

[0034] (d) Determine the operation modes of the high-voltage and medium-voltage distribution networks in the expected time period.

[0035] The above method for estimating the loop closing current and risk assessment of medium-voltage lines based on a deep learning network, in which the loop closing current prediction probability in step (5) includes three aspects, and the specific steps are as follows:

[0036] Firstly, the deterministic prediction of the loop closing current, the steps are as follows:

[0037] (a) Forecast the grid load level and distributed power generation prediction data;

[0038] (b) For the feeder load data and the voltage amplitude, active power, reactive power, etc. of the feeder bus of the high-voltage system directly connected to the loop closing feeder, repeat step 2 to preprocess the data;

[0039] (c) Then process the cleaned data through the time misalignment method to generate the input format suitable for the model structure for loop closing current prediction;

[0040] (d) Determine the operating modes of the high-voltage and medium-voltage distribution networks for the desired time period;

[0041] (e) Calculate the evaluation indices for the deterministic prediction accuracy. The root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) are respectively adopted, and their calculation formulas are as follows:

[0042]

[0043] In the formulas: N represents the number of test samples, and i is the serial number; Y i represents the predicted value of the closed-loop current; y i represents the actual value of the closed-loop current; Y i and y i The units of the variables are both dimensionless;

[0044] (f) Check whether the prediction accuracy indices meet the error requirements. If the prediction accuracy is poor, continue to return and adjust the network parameters and structural parameters of the model to achieve a more accurate prediction result, and determine the optimal parameters of the model;

[0045] Second, the prediction of the closed-loop current under different quantile conditions is as follows:

[0046] Using the training model obtained above, replace its loss function as shown in Equation (3) to conduct the prediction of the closed-loop current under different quantile conditions. Among them, the evaluation indices for the probability prediction of the closed-loop current adopt the reliability index (represented by PICP) and the sharpness index (represented by PINAW), and their calculation formulas are as follows:

[0047] a) Reliability index

[0048]

[0049] b) Sharpness index

[0050]

[0051] In the formulas: N represents the number of samples for the prediction of the closed-loop current, and i is the serial number; λ i represents the number of the predicted values of the closed-loop current belonging to the confidence interval; Y Li represents the lower confidence bound of the probability prediction of the closed-loop current; Y Ui represents the upper confidence bound of the probability prediction of the closed-loop current; R represents the prediction width of the closed-loop current under different quantiles; respectively represent the maximum and minimum values of the quantile prediction at the i-th moment; Y Li 、Y Ui, R, are all dimensionless.

[0052] Thirdly, the closed-loop current kernel density estimation and the calculation of the closed-loop security evaluation index are as follows:

[0053] Firstly, the predicted values of the closed-loop current under the above different quantile conditions are used as the input values of the kernel density estimation. Secondly, the Gaussian kernel density is sampled for estimation to obtain the probability density curve and the cumulative distribution curve of the closed-loop current prediction. Finally, the probability of the feeder overstepping the line is calculated by comparing the obtained probability curve with the maximum allowable current-carrying capacity of the feeder and the current protection setting value; the specific calculation formula is as follows: (a) The kernel density estimation function is expressed by the Gaussian kernel function as Its calculation formula is as follows:

[0054]

[0055] h ≈ 1.06σn -0.2 (13)

[0056]

[0057] where: n is the number of samples, i is the serial number; h is the window width; K(x) is the kernel function; σ is the standard deviation, and x i is the n samples of the closed-loop current prediction, and the units of h, x i and K(x) are all dimensionless.

[0058] (b) Closed-loop security evaluation index; the closed-loop over-limit probability is expressed as P; the maximum over-limit rate of the closed-loop current is expressed as α; the average over-limit rate is expressed as ε; their calculation formulas are as follows:

[0059] I m = F -1 (99.9%) (15)

[0060] P = P(I t ≥ I s ) = 1 - F(I s ) (16)

[0061]

[0062] where: f(x) represents the probability density function of the closed-loop current; F(x) represents the probability cumulative distribution function of the closed-loop current; I s represents the maximum allowable current-carrying capacity of the feeder, unit: A; I m represents the closed-loop current value at the 99.9% probability of the closed-loop current cumulative distribution function, unit: A; I t represents the predicted value of the closed-loop current probability, unit: A.

[0063] For the above method for estimating closed-loop current and risk assessment of medium-voltage lines based on a deep learning network, the output result and display in step (6) are as follows:

[0064] Output result and display: First, denormalize the prediction result, which is achieved through the following formula:

[0065] k o = k p ×(k max - k min ) + k min (19)

[0066] In the formula: k p represents the data output by the network prediction, k min represents the minimum value of the output data, k max represents the maximum value of the output data, k p , k max , k min , all have dimensionless units; k o represents the deterministic prediction result of the closed-loop current after denormalization, unit: A; Then, use kernel density estimation again to obtain the probability density estimation curve of the closed-loop current, compare the obtained prediction curve with the maximum allowable current-carrying capacity of the feeder and the current protection setting value, calculate the closed-loop operation evaluation index, judge whether the closed-loop current will cause misoperation of the overcurrent protection or instantaneous trip protection of the closed-loop switch, and at the same time determine whether the closed-loop can be performed during the expected time period, determine which time periods have the smallest closed-loop current and the smallest operation safety risk, and finally determine the closed-loop risk assessment value. The specific risk assessment is as follows:

[0067] Divide the closed-loop operation of the distribution network into three levels, namely the safe level, the warning level, and the over-limit level, and set the following discriminants: When the risk value > 5%, it is a safe-level operation; when 20% > risk value ≥ 5%, it is a warning-level risk; when 20% > risk value, it is an over-limit level risk; The closed-loop operation risk value of the closed-loop feeder is defined as the product of the "consequence value" and the "probability value"; The consequence value can be expressed as the product result of the "severity of the closed-loop accident", the "social impact factor generated by the closed-loop", and the "importance factor of the closed-loop feeder load", where:

[0068] (a) Severity of the closed-loop accident: Refer to the "Technical Specification for Quantitative Assessment of Operation Safety Risks of Southern Power Grid", calculate 1 point for every 1 thousand kilowatts of lost load as the benchmark score, and the score of the consequence severity is obtained by multiplying the basic score by the regional coefficient, expressed as C a ;

[0069] (b) Social impact factor generated by the closed-loop: According to the user's sensitivity to power outages and the social impact caused by power outages, take values from 0.1 to 0.2. The greater the social impact, the greater the value of this factor, expressed as Cb ;

[0070] (c) Load importance factor: It is taken from 0.1 to 0.4 with reference to factors such as the number of power supply sources, whether a backup automatic transfer device is configured, whether an emergency power supply is configured, the user's sensitivity to power outages, and the level of important users, and is expressed as C c ;

[0071] (d) Risk value of loop closing operation: It is obtained by calculating the cumulative distribution function of loop closing current based on the loop closing current probability density curve. It is the loop closing over-limit probability expression R a ;

[0072] In summary, the closed loop risk assessment value is expressed as R loop , which is calculated as follows:

[0073] R loop =C a ×C b ×C c ×R a (20)

[0074] Compared with the prior art, the present invention has obvious beneficial effects. It can be seen from the above scheme that the present invention uses load forecasting results, distributed power output forecasting results, historical load and other data, adopts artificial intelligence technology, and uses data-driven methods such as deep learning networks to predict the closed-loop current of the medium-voltage feeder of the new distribution network, and finally obtains the probability density of the closed-loop current at any time, thereby more scientifically reflecting the randomness and uncertainty of the current under the closed-loop operation of the medium-voltage feeder of the new distribution network. Starting from the data-driven prediction idea based on artificial intelligence technology, the present invention adopts a deep learning algorithm based on convolutional neural networks and long short-term memory networks to effectively convert the complex model association relationship of the distribution network topology, load, new energy generation and closed-loop current into the data correspondence relationship between convolutional neural networks, effectively solving the existing model-driven deterministic closed-loop current estimation technology for the modeling complexity of the new distribution network feeder and the randomness and uncertainty of boundary conditions. The weak adaptability, poor accuracy, low calculation efficiency, and the problem of solution convergence. In addition, through non-parametric kernel density estimation and closed-loop current probability density calculation, the quantitative assessment and classification of closed-loop operation risks are realized, filling the problem of the lack of risk classification and closed-loop operation period recommendation functions in the prior art. It is a medium voltage line loop current estimation and risk assessment method with good adaptability, high efficiency and good accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 This is a hardware structure system diagram of data acquisition related to closed-loop current prediction of a medium-voltage distribution network according to the present invention;

[0076] Figure 2The technical roadmap for predicting and evaluating the loop - closing current of medium - voltage feeders involved in the present invention;

[0077] Figure 3 The structure diagram of a certain regional high - voltage distribution network system adopted by the present invention;

[0078] Figure 4 The 11 annual load distribution curves of the present invention;

[0079] Figure 5 The display of the specific loop - closing current prediction results implemented by the present invention;

[0080] Figure 6 The probability density curve and probability cumulative distribution curve of different test points of the present invention. Specific implementation method

[0081] The following combines the attached drawings and preferred embodiments to detail the specific implementation manner, structure, features and functions of a method for estimating loop - closing current and risk assessment of medium - voltage lines based on a deep - learning network proposed according to the present invention.

[0082] A method for estimating loop - closing current and risk assessment of medium - voltage lines based on a deep - learning network includes the following steps:

[0083] Refer to Figure 1 , the hardware structure of the loop - closing current prediction system for medium - voltage distribution network applied in the present invention involves: database system, database maintenance terminal, SCADA system data interface, SCADA system data server, dispatching data network, SCADA system sub - station data acquisition sub - station, acquisition sub - station maintenance terminal, marketing system data interface (under development), marketing MIS system server;

[0084] Refer to Figure 2 , the estimation of loop - closing current of medium - voltage distribution network lines applied in the present invention mainly includes: system data acquisition module, loop - closing current prediction model, load prediction model, distributed power source prediction model module, SCADA system real - time data acquisition interface module, online grid real - time mathematical model verification module, power flow calculation module, result output module;

[0085] The present invention uses the DIgSILENT / PowerFactory power system simulation software to establish a quasi - dynamic simulation of SCADA system data with a certain urban power grid in Guizhou as an example. Its grid operation wiring structure diagram is as Figure 3 shown. The specific application steps for realizing loop - closing operation and risk assessment of the distribution network through rapid modeling of grid structure data and loop - closing current prediction analysis are as follows:

[0086] (1) Acquisition of basic data described in Step 1: The data source of the present invention utilizes the DIGSILENT software to simulate the annual historical load data of each node in the SCADA system. There are 11 different annual load distributions built into DIGSILENT. The establishment of each load curve is the same and is divided into 4 seasons, and each season is further divided into working days and non - working days, as shown in Table 1.

[0087] Table 1 Modeling method of annual load

[0088]

[0089] Table 1 shows the modeling method of annual load distribution, including four seasons: winter, spring, summer and autumn. Each season consists of working days and non - working days, and non - working days are further divided into Saturday and Sunday. Therefore, the load of each season is composed of three daily load distributions. For example, the daily load distributions from Monday to Friday in winter are all working - day - winter, the daily load distribution on Saturday is Saturday - winter, and the daily load distribution on Sunday is Sunday - winter. The definitions of the four different seasons are shown in Table 2.

[0090] Table 2 Definitions of seasons

[0091]

[0092] It can be seen from Table 2 that in the DIGSILENT software, from March 21st to May 14th of each year is defined as spring, from May 15th to September 14th is defined as summer, from September 15th to October 31st is defined as autumn, and from November 1st to March 20th of the following year is defined as winter. According to the above - mentioned annual load modeling method, after checking and cleaning the data, 11 annual load distribution curves with a resolution of 15 minutes are obtained. As Figure 4 shown, where the abscissa is the date and the ordinate is the per - unit value of the load, that is, the actual load of the node divided by the rated load of the node. After the power grid model and the load model are established, the BPA software is used to obtain the summer small - operation mode of Guiyang in 2021 as the initial data for quasi - dynamic simulation, so as to obtain the annual operation data with a resolution of 15 minutes, where the number of annual load sampling points n = 35040 (=60 / 15×24×365).

[0093] (2) Data pre - processing described in Step 2: There are cases where the power flow does not converge in the annual power flow data obtained by simulation, which corresponds to possible data missing and other situations in the actual system. For the missing data, programming is implemented using the matlab2021b software, and it is achieved through the following calculation formula:

[0094]

[0095] where: \(t\) represents the moment of missing data, unit: min; \(i\) represents the most recent moment with valid data before moment \(t\), unit: min; \(j\) represents the most recent moment with valid data after moment \(t\), unit: min. \(k\) t represents the missing value at moment \(t\) to be filled; \(k\) i represents the data value at moment \(i\); \(k\) j represents the data value at moment \(j\); \(k\) t and \(k\) i and \(k\) j The data values represented can be voltage, current, active power, reactive power, etc., with units of kV, A, kW, kVar respectively.

[0096] After the data filling is completed, in order to remove the unit limitation of the data, the data is normalized and transformed into the range of 0 to 1 to obtain a dimensionless pure numerical value. The calculation formula is as follows:

[0097]

[0098] where: \(k\) represents the normalized output value, and the unit of the formula symbol after normalization is dimensionless; \(k\) org represents the data value to be processed; \(k\) min and \(k\) max represent the minimum and maximum values in the feature where the data value to be processed is located respectively; \(k\) org and \(k\) min and \(k\) max The data values represented can be voltage, current, active power, reactive power, etc., with units of kV, A, kW, kVar respectively.

[0099] (3) Neural network training and generation described in step 3: Process the cleaned data by the method of time misalignment to form training data, generate the input-output format suitable for the model structure, and construct a CNN-LSTM prediction model, including the division of training samples and test samples, and the determination of the optimal parameters of the model. The specific steps are as follows: Construct a loss function as shown in formula (3), that is:

[0100]

[0101] where: \(N\) is the number of test samples, \(i\) is the serial number; \(X\) i is the input value of the \(i\)-th closed-loop current prediction; \(y\) i is the actual value of the \(i\)-th closed-loop current prediction; \(\tau\) represents the quantile; \(f(W(\tau),b(\tau),X\) i ) is the predicted value of the closed-loop current at different quantiles output by the \(i\)-th closed-loop current CNN-LSTM network; \(W(\tau)\), \(b(\tau)\) represent the network parameters of the model at different quantiles; \(\rho\) τ (x) is the loss function of the closed-loop current prediction model at different quantiles, and its calculation formula is as follows:

[0102]

[0103] Where: x represents the function variable, and all variables in formulas (3) and (4) are dimensionless.

[0104] (4) The neural network training and generation described in step 3 determine the optimal parameters of the model through the following steps:

[0105] (a) Construct a training sample set and convert it into the input-output format of the neural network model;

[0106] (b) Divide the obtained data into a training set and a test set. Due to the nature of the load model established using the DIgSILENT simulation model, since the load fluctuation ranges in spring and autumn are similar in this invention, the last month of each of the three seasons of spring, summer and autumn, and winter is combined as the test set, obtaining 9640 (3×30×96) test samples and 25400 training samples. The input data of this invention includes feeder power flow, bus voltage amplitude, load data, etc., with a total of 47 input features;

[0107] (c) Initially determine the number of convolutional layers, the number of LSTM layers, and the fully connected fusion layer through multiple trainings;

[0108] (d) After determining the initial number of structural layers, adjust the network parameters of the model structure;

[0109] (e) Finally, re-determine the final structural parameters of the model, and evaluate the performance of the model by checking the evaluation index and the number of iterations. Thus, the training and generation of the deep learning network model are completed, and the structural parameters of the obtained deep learning network model are as shown in Table 3 below.

[0110] Table 3 CNN-LSTM Structural Parameters

[0111]

[0112] (5) Determination of the closed-loop boundary conditions described in step 4: Determine that the closed-loop feeders are the Zhongda line and the Fanghua line for closed-loop operation according to the actual closed-loop requirements. The position of the closed-loop point is shown in Figure 3 ; The main parameters of the closed-loop feeders are as shown in Table 4 below.

[0113] Table 4 Main Parameters of Closed-Loop Feeders

[0114]

[0115] (6) Prediction of the loop - closing current probability described in step 5. First, use the prediction model trained in step (3) to carry out the prediction of the loop - closing current probability; secondly, use quantile regression prediction to obtain the predicted values under different quantiles of the loop - closing current; finally, use kernel density estimation to obtain the probability density distribution of the loop - closing current to be measured and the cumulative distribution curve of the loop - closing current probability, and calculate the loop - closing safety evaluation index according to the loop - closing current probability distribution curve. The specific steps are as follows:

[0116] First, the deterministic prediction of the loop - closing current, the steps are as follows:

[0117] (a) Predict the grid load level and distributed power generation prediction data;

[0118] (b) For the feeder load data and the voltage amplitude, active power, reactive power, etc. of the high - voltage system feeder bus directly connected to the loop - closing feeder, repeat step (2) to pre - process the data;

[0119] (c) Then, process the cleaned data by the time - dislocation method to generate the input format suitable for the model structure for loop - closing current prediction;

[0120] (d) Determine the operation modes of the high - voltage and medium - voltage distribution networks in the expected time period;

[0121] (e) Calculate the deterministic prediction accuracy evaluation indicators, respectively using the root - mean - square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE). Their calculation formulas are as follows:

[0122]

[0123] Where: N represents the number of test samples, i is the serial number; Y i represents the predicted loop - closing current value; y i represents the actual loop - closing current value; Y i and y i The units of the variables are all dimensionless;

[0124] (f) Check whether the prediction accuracy index meets the error requirements. If the prediction accuracy is poor, continue to return and adjust the network parameters and structure parameters of the model to achieve a more accurate prediction result, and determine the optimal parameters of the model. The prediction result is as Figure 5 .

[0125] Second, the prediction of the loop - closing current under different quantiles, the steps are as follows:

[0126] Using the training model obtained above, replace its loss function as shown in Equation (3) to predict the closing loop current under different quantile conditions. The evaluation indexes for the prediction of the closing loop current probability are the reliability index (represented by PICP) and the sensitivity index (represented by PINAW), and their calculation formulas are as follows:

[0127] a) Reliability index

[0128]

[0129]

[0130] b) Sensitivity index

[0131]

[0132] In the formula: N represents the number of closing loop current prediction samples, and i is the serial number; λ i represents the number of closing loop current prediction values belonging to the confidence interval; Y Li represents the lower confidence bound of the closing loop current probability prediction; Y Ui represents the upper confidence bound of the closing loop current probability prediction; R represents the prediction width of the closing loop current under different quantiles; respectively represent the maximum and minimum values of the quantile prediction at the i-th moment; Y Li 、Y Ui 、R、 are all dimensionless; the prediction result indexes are shown in Table 5.

[0133] Table 5 Closing loop current probability prediction indexes

[0134]

[0135] Thirdly, the kernel density estimation of the closing loop current and the calculation of the closing loop safety evaluation index are carried out as follows:

[0136] Firstly, take the closing loop current prediction values under the above different quantile conditions as the input values of the kernel density estimation. Secondly, sample the Gaussian kernel density for estimation to obtain the probability density curve and the cumulative distribution curve of the closing loop current prediction. Finally, compare the obtained probability curve with the maximum allowable current carrying capacity of the feeder and the current protection setting value to calculate the probability of the feeder crossing the line; its specific calculation formula is as follows:

[0137] (a) The kernel density estimation function is expressed by the Gaussian kernel function as Its calculation formula is as follows:

[0138]

[0139] h≈1.06σn -0.2 (13)

[0140]

[0141] Where: n is the number of samples, i is the serial number; h is the window width; K(x) is the kernel function; σ is the standard deviation, and x i are the n samples for the loop closing current prediction, and h, x i and K(x) are all dimensionless.

[0142] (b) Loop closing safety evaluation index; the loop closing overlimit probability is expressed as P; the maximum loop closing current overlimit rate is expressed as α; the average overlimit rate is expressed as ε; their calculation formulas are as follows:

[0143] I m = F -1 (99.9%) (15)

[0144] P = P(I t ≥ I s ) = 1 - F(I s ) (16)

[0145]

[0146]

[0147] Where: f(x) represents the loop closing current probability density function; F(x) represents the loop closing current probability cumulative distribution function; I s represents the maximum allowable carrying capacity of the feeder, unit: A; I m represents the loop closing current value at the 99.9% probability of the loop closing current cumulative distribution function, unit: A; I t represents the predicted value of the loop closing current probability, unit: A; the calculation results of its loop closing current probability prediction index are shown in Table 5. The corresponding probability density estimation and cumulative probability distribution results are as Figure 6 ;

[0148] (8) Output result and display: First, denormalize the prediction result through the following formula:

[0149] k o = k p × (k max - k min ) + k min (19)

[0150] Where: k p represents the data output by the network prediction, k min represents the minimum value of the output data, k max represents the maximum value of the output data, k p , kmax , k min , the units are dimensionless; k o It represents the deterministic prediction result of the closed-loop current after inverse normalization, unit: A; the cumulative distribution curve of the closed-loop current at a certain moment of the inverse normalized closed-loop current is obtained, and the closed-loop current probability prediction value is compared with the allowable current carrying capacity of the feeder to obtain the closed-loop safety assessment index, and it is judged whether the closed-loop current will cause the overcurrent protection or quick-break protection of the closed-loop switch to malfunction, and at the same time, it is determined whether the closed-loop can be closed in the expected time period, and which time period has the smallest closed-loop current and the smallest operation safety risk, and finally the closed-loop risk assessment value is determined. The risk assessment is as follows:

[0151] The distribution network closing operation is divided into three levels, namely, safety level, warning level, and over-limit level, and the following judgments are set: 5%> risk value is a safety level operation; 20%> risk value ≥ 5% is a warning level risk; 20%> risk value is an over-limit level risk; the closing operation risk value of the closing feeder is defined as the product of the "consequence value" and the "probability value"; the consequence value can be expressed as the product of "the severity of the closing accident", "the social impact factor of the closing", and "the load importance factor of the closing feeder", where:

[0152] (a) Severity of loop-closing accident: refer to the Technical Specifications for Quantitative Assessment of Operation Safety Risks of China Southern Power Grid and calculate the base score at 1 point per kilowatt of lost load. The score of the severity of the consequence is obtained by multiplying the base score by the regional coefficient, expressed as C a ;

[0153] (b) Social impact factor of the closed loop: The factor is 0.1 to 0.2 based on the user's sensitivity to power outages and the social impact caused by power outages. The greater the social impact, the greater the factor value, expressed as C b ;

[0154] (c) Load importance factor: It is taken from 0.1 to 0.4 with reference to factors such as the number of power supply sources, whether a backup automatic transfer device is configured, whether an emergency power supply is configured, the user's sensitivity to power outages, and the level of important users, and is expressed as C c ;

[0155] (d) Risk value of loop closing operation: It is obtained by calculating the cumulative distribution function of loop closing current based on the loop closing current probability density curve. It is the loop closing over-limit probability expression R a ;

[0156] In summary, the closed loop risk assessment value is expressed as R loop , which is calculated as follows:

[0157] R loop =C a ×C b ×C c ×Ra (20)

[0158] (a) Predicted values of different quantiles of the closed-loop current;

[0159] (b) Probability density curve of the closed-loop current at a certain moment obtained by kernel density estimation;

[0160] (c) Obtaining the probability distribution curve of the closed-loop current by using the kernel density curve;

[0161] (d) Comparing the predicted value of the closed-loop current with the maximum allowable current-carrying capacity of the feeder to obtain the probability of closed-loop over-limit;

[0162] (e) Using the risk assessment formula to obtain the risk quantification value of the closed-loop as shown in Table 6;

[0163] (f) Finally, conduct a safety assessment of the closed-loop operation according to the risk quantification value.

[0164] Table 6 Risk assessment values of closed-loop operation

[0165]

[0166] The prediction results of the closed-loop current of the medium-voltage feeder show that the predicted value of the closed-loop current is relatively small during the low-load period, and the fluctuation range of the closed-loop current is also small. Most of the safe points appear during the low-load period. There are more closed-loop over-limit points during the high-load period, and closed-loop operation is strictly prohibited. The warning points mainly appear when there are individual peak loads on the feeder. At the same time, there may also be a large fluctuation range in the prediction of the closed-loop current, and it is not easy to carry out long-term closed-loop operation. The prediction results show that the present invention can perform closed-loop operation as long as the high-load period is avoided.

[0167] The present invention uses artificial intelligence means to quickly predict the closed-loop current of the urban medium-voltage distribution network, predicts the closed-loop current of different quantiles, and then uses the method of kernel density estimation to obtain the probability density curve and cumulative distribution curve of the closed-loop current, and finally realizes the risk quantification assessment of the closed-loop current. It mainly solves the problems of weak adaptability, poor accuracy, low calculation efficiency, and the problem of solving convergence in the existing estimation of the closed-loop current, fills the gap in the lack of risk classification and recommended function for closed-loop operation time period in the existing technology, and can quickly provide effective decision support for actual closed-loop operation and operators.

[0168] The above is only a preferred embodiment of the present invention, and it does not impose any form of limitation on the present invention. Any simple modification, equivalent change, and modification made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A method for estimating the closed-loop current and risk assessment of medium-voltage lines based on a deep learning network, comprising the following steps: (1) Data collection: Obtain historical data in the data acquisition and monitoring control system, including the grid structure parameters and operation mode data of the transmission lines and transformers in the distribution network dispatching department, the power user and grid operation data of the marketing department and the production technology department, the power generation and load level data, the distributed energy output, and the historical data of the relevant closed-loop feeder current; (2) Data preprocessing: Perform data cleaning on the historical data, including deleting outliers, filling in missing values by interpolation method, and normalizing the data to the range of 0-1, removing the unit limit of the data, and converting it into a dimensionless pure numerical value; (3) Neural network training and generation: Process the cleaned data by the method of time dislocation to form training data, generate the input format suitable for the model structure, adjust the input layer, hidden layer, fully connected layer of the network, and the convolution step size and convolution kernel network structure parameters to achieve better prediction results. Design a deep learning model based on the convolutional neural network CNN and the long short-term memory network LSTM to realize the training of the closed-loop current prediction model and generate its prediction model; wherein the deep learning model based on the convolutional neural network CNN and the long short-term memory network LSTM is composed of an input layer, a CNN layer, a pooling layer, an LSTM layer, a fully connected layer and a pooling layer. The optimizer selects the Adam optimizer and constructs a loss function as shown in formula (3): Where: N is the number of test samples, i is the serial number, and X i is the input value for the prediction of the i-th loop closing current, and y i is the actual value of the i-th loop closing current prediction. τ represents the quantile, and f(W(τ), b(τ), X i ) is the predicted value of the loop closing current at different quantiles output by the CNN-LSTM network for the i-th loop closing current. W(τ) and b(τ) represent the network parameters of the model at different quantiles. ρ τ (x) is the loss function of the loop closing current prediction model at different quantiles, and its calculation formula is as follows: In the formula: x represents the function variable, and all variables in formula (3) and formula (4) are dimensionless; (4) Determination of closed-loop boundary conditions: including the operator selecting the line to be closed-loop, specifying the operation time range of the medium-voltage line closed-loop plan, conducting load prediction for this plan operation time range, and predicting the distributed power output data, and determining the operation modes of the high-voltage and medium-voltage distribution networks in the expected time period; (5) Closed-loop current probability prediction: First, use the prediction model trained in step (3) to carry out the deterministic prediction of the closed-loop current; secondly, use quantile regression prediction to obtain the predicted values under different quantiles of the closed-loop current; finally, use kernel density estimation to obtain the probability density distribution of the measured closed-loop current and the cumulative distribution curve of the closed-loop current probability, and calculate the closed-loop safety assessment index according to the closed-loop current probability distribution curve; wherein the closed-loop current probability prediction includes the following three aspects, and the specific steps are as follows: First, the deterministic prediction of the closed-loop current, the steps are as follows: (a) Predict the grid load level and distributed power prediction data; (b) Preprocess the feeder load data and the voltage amplitude, active power, and reactive power of the feeder bus of the high-voltage system directly connected to the closed-loop feeder, and repeat step 2 for the data; (c) Then process the cleaned data by the method of time dislocation to generate the input format suitable for the model structure for closed-loop current prediction; (d) Determine the operation modes of the high-voltage and medium-voltage distribution networks in the expected time period; (e) Calculate the deterministic prediction accuracy evaluation indexes, and respectively use the root mean square error, the mean absolute error, and the mean absolute percentage error. The calculation formulas are as follows: Where: N represents the number of test samples, i is the serial number; Y i represents the predicted value of the closed-loop current; y i represents the actual value of the closed-loop current; Y i and y i The units of the variables are all dimensionless; (f) Check whether the prediction accuracy index meets the error requirement. If the prediction accuracy is poor, continue to return and adjust the network parameters and structural parameters of the model to achieve a more accurate prediction result, and determine the optimal parameters of the model. Second, the prediction of the closed-loop current under different quantile conditions is as follows: Using the training model obtained above, replace its loss function as shown in Equation (4) to predict the closed-loop current value under different quantile conditions. The evaluation indexes for the probability prediction of the closed-loop current adopt the reliability index represented by PICP and the sensitivity index represented by PINAW. Their calculation formulas are as follows: a) Reliability index b) Sensitivity index R = maxY i α - minY i α (11) Where: N represents the number of predicted samples of the closed-loop current, and i is the serial number; λ i represents the number of predicted values of the closed-loop current belonging to the confidence interval; Y Li represents the lower confidence bound of the closed-loop current probability prediction; Y Ui represents the upper confidence bound of the closed-loop current probability prediction; R represents the prediction width of the closed-loop current at different quantiles; maxY i α , minY i α respectively represent the maximum and minimum values of the quantile prediction at time i; Y Li , Y Ui , R, maxY i α , minY i α are all dimensionless; Third, the kernel density estimation of the closed-loop current and the calculation of the closed-loop safety evaluation index are as follows: First, use the predicted values of the closed-loop current under the above different quantile conditions as the input values for kernel density estimation. Second, sample the Gaussian kernel density for estimation to obtain the probability density curve and cumulative distribution curve of the closed-loop current prediction. Finally, compare the obtained probability curve with the maximum allowable carrying capacity of the feeder and the current protection setting value to calculate the probability of feeder overlimit; its specific calculation formula is as follows: (a) The kernel density estimation function is expressed using a Gaussian kernel function as Its calculation formula is as follows: h≈1.06σn -0.2 (13) Where: n is the number of samples, i is the serial number; h is the window width; K(x) is the kernel function; σ is the standard deviation, and x i are the n samples for predicting the closed-loop current, and the units of h, x i and K(x) are all dimensionless; (b) Closed-loop safety evaluation index; the probability of closed-loop overlimit is denoted as P. The maximum overlimit rate of the closed-loop current is denoted as α; the average overlimit rate is denoted as ε. Their calculation formulas are as follows: I m = F -1 (99.9%) (15) P = P(I t ≥ I s ) = 1 - F(I s ) (16) where: f(x) represents the probability density function of the closed-loop current; F(x) represents the cumulative distribution function of the closed-loop current; I s represents the maximum allowable current-carrying capacity of the feeder, unit: A; I m Indicates the value of the closing loop current at the cumulative distribution function probability of the closing loop current being 99.9%, unit: A; I t Indicates the predicted value of the closed-loop current, unit: A; (6) Output results and display: Compare the closed-loop current with the maximum allowable carrying capacity of the feeder and the current protection setting value, comprehensively analyze each risk index of the closed-loop operation, and finally obtain the risk quantification value of the closed-loop operation.

2. A method for estimating and risk assessing the closed-loop current of a medium-voltage line based on a deep learning network as described in claim 1, wherein in step (2), filling the default value is achieved through the following calculation formula: Where: t represents the moment of the default data, unit: min; i represents the most recent moment with valid data before the moment t, unit: min; j represents the most recent moment with valid data after the moment t, unit: min; k t represents the default value at the moment t to be filled; k i represents the data value at the moment i; k j represents the data value at the moment j; k t and k i and k j The data values represented are voltage, current, active power, and reactive power, with units of kV, A, kW, and kVar respectively.

3. A method for estimating and risk assessing the closed-loop current of a medium-voltage line based on a deep learning network as described in claim 1, wherein in step (2), the data is normalized and converted to the range of 0 to 1, removing the unit limitation of the data, and converting it into a dimensionless pure numerical value through the following calculation formula: Where: k represents the normalized output value, with the unit being dimensionless; k org represents the data value to be processed; k min and k max respectively represent the minimum and maximum values in the feature where the data value to be processed is located; k org and k min and k max The data values represented are voltage, current, active power, and reactive power, with the units being kV, A, kW, and kVar respectively.

4. A method for estimating and risk assessing the closed-loop current of a medium-voltage line based on a deep learning network as described in claim 1, wherein in step (3), the neural network training and generation determine the optimal parameters of the model through the following steps: (a) Construct a training sample set and convert it into the input-output format of the neural network model. (b) Divide the obtained data into a training set and a test set. (c) Initially determine the number of convolutional layers, the number of LSTM layers, and the fully connected fusion layer through multiple trainings. (d) After determining the initial structural layers, adjust the network parameters of the model structure. (e) Finally, determine the final structural parameters of the model again, and evaluate the performance of the model by checking the evaluation indexes and the number of iterations.

5. A method for estimating and risk assessing the closed-loop current of a medium-voltage line based on a deep learning network as described in claim 1, wherein in step (4), the specific steps for determining the closed-loop boundary conditions are as follows: (a) The operator selects the line to be closed-loop. (b) Specify the operation time range of the medium-voltage closed-loop plan. (c) Conduct load prediction and distributed power generation output prediction for this planned operation time range. (d) Determine the operating modes of the high-voltage and medium-voltage distribution networks for the desired time period.

6. The method for estimating the closing loop current and risk assessment of medium voltage lines based on a deep learning network according to claim 1, wherein the output result and display in step (6). The features are as follows: Output results and display: First, denormalize the prediction results, which is achieved through the following formula: k o = k p ×(k max - k min ) + k min (19) Where: k p represents the data output by the network prediction, k min represents the minimum value of the output data, k max represents the maximum value of the output data, k p , k max , k min , and the unit of k o is dimensionless; k represents the deterministic prediction result of the closed-loop current after anti-normalization, unit: A; the probability density estimation curve of the closed-loop current is obtained by using kernel density estimation, and the obtained prediction curve is compared with the maximum allowable carrying capacity of the feeder and the current protection setting value to obtain the closed-loop operation evaluation index, judge whether the closed-loop current will cause misoperation of the overcurrent protection or instantaneous trip protection of the closed-loop switch, and at the same time determine whether the closed-loop can be performed during the expected time period, determine which time periods have the smallest closed-loop current and the smallest operation safety risk, and finally determine the closed-loop risk assessment value. The risk assessment is as follows: The loop closing operation of the distribution network is divided into three levels, namely the safety level, the warning level, and the overlimit level, and the following discriminants are set: When the risk value > 5%, it is a safety-level operation; when 20% > the risk value ≥ 5%, it is a warning-level risk; when 20% > the risk value, it is an overlimit-level risk. The risk value of the loop closing operation of the loop closing feeder is determined as the product of the "consequence value" and the "probability value"; the consequence value is also expressed as the product result of the "severity of the loop closing accident", the "social impact factor generated by loop closing", and the "importance factor of the loop closing feeder load", where: (a) Severity of closed-loop operation accident: Refer to the Technical Specification for Quantitative Assessment of Operational Safety Risks in Southern Power Grid. The basic score is calculated as 1 point for every 1,000 kW of lost load. The score of the consequence severity is obtained by multiplying the basic score by the regional coefficient, denoted as C a ; (b) Social impact factors caused by closing the loop: According to the user's sensitivity to power outages and the social impact caused by power outages, the value ranges from 0.1 to 0.

2. The greater the social impact, the greater the value of this factor, denoted as C b ; (c) Load importance factor: It takes values from 0.1 to 0.4 with reference to factors such as the number of power supply sources, whether there is an automatic switching device, whether there is an emergency power supply, the user's sensitivity to power outages, and the importance level of important users, and is denoted as C c ; (d) Risk occurrence value of loop closing operation: Obtained by calculating the cumulative distribution function of loop closing current based on the probability density curve of loop closing current, which is the loop closing overlimit probability denoted as R a ; In summary, the loop closing risk assessment value is expressed as R loop , and its calculation formula is as follows: R loop = C a × C b × C c × R a (20).

Images