A rapid early warning method for heavy overload of distribution transformers based on feature load prediction
By preprocessing and clustering analysis of historical load data, combined with LSTM model, an offline training-online prediction system is built, which solves the problems of low accuracy and poor efficiency of heavy overload warning of distribution network equipment in the existing technology, and achieves a fast and accurate early warning of distribution and heavy overload risk.
Patent Information
- Application Number
- CN202110651745.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-06-11
- Publication Date
- 2025-07-29
- Estimated Expiration
- 2041-06-11
AI Technical Summary
The load rate prediction and heavy overload warning of existing distribution network equipment mainly rely on expert experience, with low accuracy and poor efficiency, and the deep learning method has insufficient generalization capabilities in large-scale applications, so it is impossible to effectively warning the risk of heavy overload in distribution.
By collecting historical load data, performing data preprocessing and outlier cleaning, using k-means clustering and LSTM model, combining load-daily feature sequence and influencing factors, a re-overload warning system with offline training-online prediction is built, and the input factor with the highest correlation coefficient is selected for quick and accurate warning.
It realizes rapid and accurate classification of matching variables and early warning of heavy overload risks, improves the accuracy and efficiency of the warning system, reduces the computing resource requirements, and meets the real-time early warning needs of large-scale matching variables.
Smart Images
Figure CN113361202B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a distribution transformer heavy overload rapid early warning method based on characteristic load prediction, which is used in the field of distribution transformer prediction. Background Art
[0002] Domestic power transmission and distribution equipment has made great progress in the research of information monitoring, storage and sharing, and has developed status monitoring and fault diagnosis mechanisms for equipment such as transformers. With the continuous growth of distribution network scale and online monitoring equipment, massive data sets have been provided, providing a data foundation for the application of big data technology and deep learning methods in the analysis of heavy overload characteristics and related prediction research of distribution transformers (abbreviated as distribution transformers).
[0003] Currently, load rate prediction and severe overload warning for distribution network equipment are primarily based on manual methods based on expert experience, which suffers from low accuracy and inefficiency. Previous studies have applied deep learning methods to severe overload prediction for distribution transformers, often transforming the severe overload prediction problem into predicting the load status of a substation, the daily peak load of a substation, and the probability of severe overload events. These models are primarily targeted at specific targets and, while achieving good warning results, have poor generalization capabilities and cannot be applied to large-scale distribution transformer risk warning. Summary of the invention
[0004] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a distribution transformer heavy overload rapid warning method based on characteristic load prediction. The method clusters the load daily characteristic sequence, analyzes the influencing factors of each load type to select the corresponding input variables, and on this basis constructs a heavy overload warning system that combines offline training and online prediction.
[0005] A technical solution to achieve the above object is: a rapid early warning method for heavy overload of distribution transformer based on characteristic load prediction, characterized by comprising the following steps:
[0006] Step 1: Collect historical load data sets. The data set format is one data point every 15 minutes, and 96 data points per day.
[0007] Step 2: Preprocess the historical data set, remove missing values and outliers, and replace them with the mean of the previous and next data points;
[0008] Step 3: Average the pre-processed daily load data of the distribution transformer to obtain the typical daily load sequence of the distribution transformer, perform k-means clustering on the typical daily load sequence of all distribution transformers, and select the minimum cluster number k min And the maximum number of clusters k max, perform n times of k-means clustering for each number of centers, select the clustering result with the smallest DBI index among them as the clustering result for that number of centers, and then select the clustering result with the smallest DBI index among different numbers of clustering centers as the final clustering result of the distribution transformer;
[0009] Step 4, select the 3 factors with the highest correlation coefficient and the historical load from the influencing factors of the load as the input factors of the model, and construct the corresponding LSTM model;
[0010] Step 5, evaluate the overload risk level of the distribution transformer according to the maximum load rate in the output prediction value. The evaluation form is shown in the following table.
[0011] Warning level Load rate Display color Transformer status Safety [0,0.6) Blue Light load Orange warning [0.6,0.7) Orange Approaching heavy load Yellow warning [0.7,1) Yellow Heavy load Red warning [1,∞) Red Overload
[0012] Finally, output the warning level of the distribution transformer.
[0013] The existing LSTM algorithm has the characteristics of high accuracy and slow speed, and cannot be directly applied to large-scale overload risk early warning. It requires too much computing resources and cannot meet the daily computing requirements. This patent solves the difficulties in large-scale distribution transformer overload risk early warning, can effectively classify distribution transformers, and on this basis, quickly and accurately warn of the overload risk of distribution transformers and the corresponding moments. Brief Description of the Drawings
[0014] Figure 1 It is a schematic flow chart of a method for quickly warning of overload of distribution transformers based on characteristic load prediction of the present invention. Detailed Embodiment
[0015] In order to better understand the technical solution of the present invention, the following will be described in detail through specific embodiments:
[0016] Please refer to Figure 1 , a method for quickly warning of overload of distribution transformers based on characteristic load prediction of the present invention includes the following steps:
[0017] Step 1, the step of collecting the historical load data set.
[0018] The load data provided by the power company is in the format of one point every 15 minutes, with 96 data points in a day. Let the historical load data set be a matrix X of m*n.
[0019]
[0020] Among them: m is the total number of training samples; n is the number of input features; x ij represents the jth feature of the input data of the ith sample.
[0021] To avoid the impact of the dimension differences and value ranges of various data on model training, it is necessary to perform normalization processing before the data is input into the model.
[0022] Perform normalization processing on each column of the original dataset matrix:
[0023]
[0024] Where: 1 ≤ j ≤ n, minx ij is the minimum value of each column of data, max 1≤i≤m x ij - min 1≤i≤m x ij is the range of each column of data.
[0025] Step 2, preprocess the historical compliance dataset, remove missing values and outliers, and replace them with the mean of the previous and subsequent data points. Although with the popularization of smart meters, the collection of power load data is more accurate and reliable than in the past during manual collection, there may still be cases of abnormal sampling. Some data deviate from the true values, causing abnormal fluctuations in the data change trend. The existence of these data will affect the training of the model, making the model unable to reflect the true load change law. Therefore, it is particularly important to identify and correct these abnormal data. When dealing with abnormal data, mainly utilize the continuity of power load data over time to perform horizontal processing on historical data. Since data may be missing or abnormal due to interference during data collection and transmission. If missing values are directly discarded or outliers and normal values are used as training inputs simultaneously, some important experimental information may be ignored or noise may be introduced, having an adverse impact on the prediction results.
[0026] Data preprocessing mainly includes filling missing values and cleaning outliers. There are various methods for filling missing values, such as filling with the mean, mode, median; weighted processing of adjacent loads, interpolation methods, etc. For the problem of missing data, if the missing time is long, it is difficult to fill, and even if filled, it is difficult to reflect the true data change law. For these long - missing time periods, fill with historical data of the same period. If the missing time is short, interpolation methods can be used to supplement the data, fitting the data of adjacent dates of the missing data. Assume the fitting curve is:
[0027] x d = a n * d n + a n-1 * d n-1 +…+ a1 * d + a0 (3)
[0028] Where n is the highest degree of the fitting regression equation, and xd is the load value of the d - th point in the data of adjacent dates of the missing data.
[0029] There are mainly two methods for the discrimination and elimination of outliers: the physical discrimination method and the statistical discrimination method. Based on the fact that historical load data mostly follows an approximate normal distribution, a statistical discrimination method combining skewness and the 3σ criterion is adopted in this paper to determine and process outliers. Suppose E(n) = e1, e2, e3, …, e n is the load data set. The skewness and the 3σ criterion are defined as:
[0030]
[0031] |v i | = |e i - μ| > 3σ (5)
[0032] where SKEW(E(n)) is the load skewness value; n is the number of load values; e i , μ, σ, v i are the measured load value (i = 1, 2... n), the total load average value, the standard deviation, and the residual error respectively.
[0033] There are mainly three possibilities for skewness in power load: negative skewness, positive skewness, and approximate normality. According to different load skewnesses, the 3σ criterion takes the load data values greater than μ + 3σ or less than μ - 3σ as outliers with large errors and eliminates them. When the difference between the average value and the median is not significant, it is an approximate normal distribution (skewed-0). At this time, the data outside the two extremes (μ + 3σ, μ - 3σ) (that is, 0.15% on each side) are cut off as outliers; in the case of negative skewness (skewed < 0), 0.3% of the data on the left side are cut off; in the case of positive skewness (skewed > 0), 0.3% of the data on the right side are cut off.
[0034] If there are outliers in the data, the average of the loads on the adjacent two days can be used to replace the load value of the current day, that is
[0035] If there is:
[0036] max[|x d - x d-1 |, |x d - x d+1 |] > ε (6)
[0037] where ε is the deviation threshold, and its value is related to a specific date and is determined according to the historical load change trend before the current day and the differences in relevant influencing factors between adjacent days.
[0038] Then the load data of the abnormal day is replaced by:
[0039]
[0040] Step 3: Average the daily load data of the distribution transformer after preprocessing to obtain the typical daily load sequence of the distribution transformer. Perform k-means clustering on the typical daily load sequences of all distribution transformers, and select the minimum number of clusters kmin and the maximum number of clusters kmax. Considering that the k-means clustering result is related to the initial clustering center, n times of k-means clustering are performed for each number of centers, and the clustering result with the minimum DBI index is selected as the clustering result for that number of centers. The calculation formula of the DBI index is shown in Equation (8). Then, select the clustering result with the minimum DBI index among different numbers of clustering centers as the final clustering result of the distribution transformer.
[0041]
[0042] In the formula respectively represent the average distances from the load curves in the K-th class and the h-th class to the class center. d k,h represents the distance between the class centers of the k-th class and the h-th class.
[0043] Build a model for the same type of load. In addition to historical load, select the input parameters of the model. Convert the daily load to the daily average load and analyze its correlation with various factors. The selected factors include daily average temperature, daily average humidity, daily average wind speed, day type, and weather type. Among them, the daily average temperature, daily average humidity, and daily average wind speed are normalized to obtain the corresponding sequences. The day type mainly includes weekdays, weekends, and holidays. The weather type mainly includes sunny, cloudy, overcast, light rain, moderate rain, and heavy rain. The inconvenient travel and increased dehumidification demand on rainy and overcast days will increase the electricity consumption, resulting in higher electricity consumption on rainy and overcast days than on sunny days. The marking values of weather and day type are shown in the following table. When there are multiple weathers in a day, the marking value is taken as the average of the maximum and minimum marking values of all weathers on that day.
[0044]
[0045] Obtain the influence relationship between the load and different factors through the Granger causality test method. Take the daily average load and the daily average temperature as an example. First, obtain the daily average load sequence X and the daily average temperature sequence Y of the same length. Its Granger causality test model is shown in Equations (9) and (10). Among them, Equation (9) is the unconstrained regression model (u), and Equation (10) is the constrained programming model (r).
[0046]
[0047]
[0048] In the formula, α u,0 、α r,0 represent the constant terms; p and q are respectively the maximum lag periods of variables Y and X; α u,i 、α r,i, β u,i are the coefficient estimates of the lagged variables X and Y respectively; X t-i , Y t-i are the (t - i)-th lagged terms of X and Y respectively; ε u,t , ε r,t is the random error term.
[0049] Construct the F-statistic using the unrestricted and restricted regression models as shown in Equation (11).
[0050]
[0051] In the formula, RSS r and RSS u are the sum of squared residuals of the restricted and unrestricted regression models respectively.
[0052] If F ≥ F α (q, n - p - q - 1), it can be considered that the coefficient estimate β u,i is significantly 0, that is, the daily average temperature is not the Granger cause of the change in daily average load, and there is no linkage relationship between them; conversely, if F < F α (q, n - p - q - 1), it proves that the daily average temperature is the Granger cause of the change in daily average load, and there is a linkage relationship between them. Similarly, it should also be tested whether the daily average load is the Granger cause of the daily average temperature. If the test result shows that the daily average load and the daily average temperature are Granger causes of each other, there should be a third variable that simultaneously affects the changes in load and temperature, rather than the temperature and load affecting each other. Therefore, it is considered that the daily average temperature is not an influencing factor of the daily average load.
[0053] Select the 3 factors with the highest correlation coefficients and the historical load among the influencing factors of the load as the input factors of the model, and construct the corresponding LSTM model. The deep LSTM network based on the present invention consists of 1 input layer, multiple hidden layers and 1 output layer. The output of the previous hidden layer is used as the input of the next hidden layer. The input layer and the hidden layers together realize the feature extraction of the input historical load data and influencing factor data. The output of the last hidden layer is a one-dimensional column vector, and the predicted value of the processed load data is obtained through linear regression. The overload risk level of the distribution transformer is evaluated according to the maximum load rate in the output predicted value, and the evaluation criteria are shown in the following table.
[0054] Warning level Load rate Display color Transformer status Safety [0,0.6) Blue Light load Orange warning [0.6,0.7) Orange Approaching heavy load Yellow warning [0.7,1) Yellow Heavy load Red warning [1,∞) Red Overload
[0055] Meanwhile, an offline training-online prediction system is constructed. The offline training module first uses a high-performance server to learn from large-scale historical data, and copies the model with good verification effect to the online prediction module; the online prediction module collects new data, inputs the new data into the model for prediction, and then transmits the new data to the offline training module; when the prediction error continues to be too large or reaches a certain period, the offline training module updates the model using the accumulated new data, and then copies the model with good verification effect to the online prediction module.
[0056] Example:
[0057] After preprocessing the target data group, a typical daily load sequence is obtained. Select kmin = 2 and kmax = 12 for load clustering analysis, and select the clustering result with the lowest DBI value as the optimal clustering result for this number of clustering centers. The DBI results for different numbers of clustering centers are shown in the following table. Select the optimal number of clusters k = 6.
[0058]
[0059] After selecting the corresponding influencing factors for each load class, train its LSTM model. Since general loads are periodic, the input and output lengths are selected as 96*7. The 96*7 historical load data and influencing factor data for seven days are used as the input layer of the LSTM network, and the output layer is the 96*7 load values for the next seven days. The number of cells is 8, the learning rate is 0.002, the network depth is 2, and the optimization method is Adam. Train a set of hyperparameters for each type of distribution transformer. To verify the effectiveness of the method in this paper, a comparison group is set up, and the model used is LSTM: one set of hyperparameters is configured for each distribution transformer, and the LSTM inputs the historical load sequence and all influencing factor sequences.
[0060] To evaluate the prediction accuracy of the model, the mean absolute percentage error and the root mean square error are selected as evaluation indicators, and their calculation formulas are shown in (15)(16).
[0061]
[0062]
[0063] The experimental environment is an i5-9400F processor, 16GB of memory, an NVIDIA GeForce GTX 1660 graphics card, the programming language environment is python3.7, and the software architecture is based on the TensorFlow framework. Apply the method in this paper and the method of the comparison group, train for 50 epochs, and perform seven-day load prediction for all distribution transformers. The average MAPE, RMSE of the prediction results of the six load classes in the experimental group and the comparison group, as well as the training time and generalization time required for the model for 50 epochs are shown in the following table.
[0064] Table 5 Prediction Result Table
[0065]
[0066] The average MAPE value of the control group is 5.42%, and the average RMSE is 0.166 KW. After applying the method of this patent, the MAPE value is reduced by 3.38%, and the RMSE value is reduced by 0.103 KW. The average training efficiency of the distribution transformer is increased by 600% compared with the control group. Based on the accurate distribution transformer load prediction level, the risk of large-scale distribution transformer overload can be effectively warned.
[0067] Those of ordinary skill in the art should recognize that the above embodiments are only used to illustrate the present invention and are not intended to limit the present invention. As long as it is within the spirit of the present invention, changes and modifications to the above embodiments will fall within the scope of the claims of the present invention.
Claims
1. A rapid early warning method for distribution transformer heavy overload based on feature load prediction, characterized in that, The steps include: Step 1: Collect historical load data sets. The load data set format is one data point every 15 minutes, and 96 data points per day. Assume that the historical load data set is an m*n matrix X, Where: m is the total number of training samples; n is the number of input features; x i j represents the j-th feature of the i-th sample input data; Normalize the data in each column of the original data set matrix: Wherein: minx ij is the minimum value of each column of data, and max 1≤i≤ m x ij - min 1≤i≤m x ij is the range of each column of data; Step 2: Preprocess the historical data set, remove missing values and outliers, and replace them with the mean of the previous and next data points. Data preprocessing is to fill missing values and clean outliers: Missing value filling includes mean, mode or median; weighted processing or interpolation method of adjacent loads. If the data is missing for a long time, the historical value data of the same period is used to fill the data; if the missing time is short, interpolation method is used to supplement the data and fit the data of the date close to the missing data. Assume that the fitting curve is: x d = a n * d n + a n-1 * d n-1 +... + a1 * d + a0(3) Where n is the highest degree of the fitted regression equation, and xd is the loading value of the dth point in the data adjacent to the date of missing data; The discrimination and rejection of outliers are carried out by using a statistical discrimination method that combines skewness and the 3σ criterion to determine outliers. Let E(n) = e1, e2, e3, …, e n be the load data set, and the skewness and the 3σ criterion are defined as: |v i | = | e i -μ > 3σ (5) Among them, SKEW(E(n)) is the load skewness value; n is the number of load values; e i , μ, σ, v i are the measured load values (i = 1, 2... n), the total load average value, the standard deviation, and the residual error, respectively. When the mean and median are not much different, the distribution is approximately normal (skewed-0). At this time, the data outside the two extremes (u+3σ, u-3σ) (i.e., 0.15% on each side) are removed as outliers; in the case of negative skewness (skewed<0), 0.3% of the data on the left is removed; in the case of positive skewness (skewed>0), 0.3% of the data on the right is removed. If the data is abnormal, the average of the loads of two consecutive days can be used to replace the load value of the day, that is, If you have: max[|x d -x d-1 |,|x d -x d+1 |]>ε(6) Among them, ε is the deviation threshold, and its value is related to a specific date and is determined based on the historical load change trend before the day and the differences in related influencing factors between adjacent days. Then replace the abnormal daily load data with: Step 3: Average the daily load data of the distribution transformer after preprocessing to obtain the typical daily load sequence of the distribution transformer. Perform k-means clustering on the typical daily load sequences of all distribution transformers, and select the minimum number of clusters k min and the maximum number of clusters k max , perform k-means clustering n times for each number of cluster centers, select the clustering result with the minimum DBI index among them as the clustering result for that number of cluster centers, and then select the clustering result with the minimum DBI index among different numbers of cluster centers as the final clustering result of the distribution transformer In the formula respectively represent the average distances from the load curves in the K-th class and the h-th class to the class centers, d k,h represents the distance between the class centers of the k-th class and the h-th class The Granger causality test method is used to obtain the influence relationship between load and different factors; Step 4: Select the three factors with the highest correlation coefficients and the historical load from the factors affecting the load as the input factors of the model and build the corresponding LSTM model; Step 5: Evaluate the heavy overload risk level of the distribution transformer according to the maximum load rate in the output prediction value. The evaluation table is shown in the following table. The final output is the warning level of the distribution transformer.