Intelligent line loss diagnosis method and device based on comprehensive analysis of machine learning method

Abstract

The invention relates to the technical field of power grid monitoring, in particular to an intelligent line loss diagnosis method and device based on comprehensive analysis of a machine learning method. According to the method, an SHAP algorithm analysis model, a Bayesian network model and other machine learning models are applied, and the experience of service experts is combined, so that a set of quantitative evaluation and optimal treatment scheme for the line loss anomaly of the transformer area is formed, and the key anomaly problem can be efficiently focused and solved on the premise of limited resources, so that the purposes of reducing the line loss and improving the efficiency are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of power grid monitoring, and in particular to a line loss intelligent diagnosis method and device based on comprehensive analysis using a machine learning method. Background Art

[0002] The line loss problem has always been a focus of attention in the power industry. With the continuous expansion of the scale of the power grid and the continuous growth of electricity demand, the line loss problem has become increasingly prominent, directly affecting the economic benefits and operating efficiency of the power grid. Accurate diagnosis and analysis of the causes of line loss are of great significance for reducing the line loss rate and improving the operating efficiency of the power grid. Therefore, the development of advanced line loss intelligent diagnosis algorithms has become a hot area in current power system research.

[0003] Traditional line loss analysis methods mainly rely on manual experience and simple statistical analysis. These methods have problems such as low analysis efficiency, lack of accuracy and objectivity, and difficulty in discovering complex nonlinear relationships. With the advent of the big data era, traditional methods can no longer meet the needs of real-time dynamic analysis of massive data, and new technical means are urgently needed to improve the efficiency and accuracy of line loss diagnosis.

[0004] In recent years, machine learning methods have been widely used in power system analysis. Decision trees and clustering algorithms have been applied to line loss prediction, pattern recognition, influencing factor analysis, and feature extraction. These methods have improved the efficiency and accuracy of line loss analysis to a certain extent, but a single algorithm is often difficult to fully and accurately diagnose complex line loss problems. In view of the complexity of line loss problems, it has become an inevitable trend to organically combine multiple machine learning methods.

[0005] The comprehensive application of multiple algorithms can give full play to the advantages of various algorithms, analyze line loss characteristics from different angles, improve the reliability of results through cross-validation, and realize multi-dimensional and multi-level analysis of line loss problems. This comprehensive research and judgment method is expected to break through the limitations of a single algorithm and provide a more comprehensive and accurate solution for intelligent diagnosis of line loss. However, in practical applications, intelligent diagnosis algorithms for line loss based on multiple machine learning methods still face many challenges.

[0006] The diagnostic results of the existing abnormal analysis and diagnosis models are relatively simple and cannot fully reflect the complex and changeable situations in power grid operations. They perform poorly when dealing with complex and changeable on-site conditions, and the diagnostic accuracy and reliability in complex scenarios are insufficient. The existing models are obviously insufficient in identifying nonlinear relationships in data. There are complex nonlinear relationships between various factors in the power grid system, and traditional models are mostly based on linear assumptions. It is difficult to accurately capture these nonlinear characteristics, thus affecting the accuracy of diagnosis. The existing diagnostic strategies lack sufficient intelligence and adaptability, and it is difficult to meet the growing demand for refined management. Summary of the invention

[0007] The purpose of the present invention is to overcome the defects of the prior art and propose an intelligent line loss diagnosis method based on comprehensive analysis of machine learning methods, which can improve the accuracy, flexibility and adaptability of line loss diagnosis.

[0008] To achieve the above object, the present invention adopts the following specific technical solutions: The line loss intelligent diagnosis method based on comprehensive analysis by machine learning method provided by the present invention comprises the following steps: S1. Construct a classification model for substation unqualified characteristics based on line loss volatility and daily line loss rate, including classifying daily line loss rate types according to the daily line loss rate results of the substation and the theoretical value of an indicator of the substation, and calculating the substation volatility according to the daily line loss results of the substation, which is used to classify abnormal substations; S2. Construct a main cause analysis model for line loss failure, including model training, abnormal sorting, and sorting result summary, to obtain the abnormal causes of line loss failure in the substation area; S3. Construct an abnormal correlation main cause analysis model, and analyze the relationship between abnormal causes by constructing a Bayesian network model to locate the main abnormal cause; S4. Construct an abnormal quantitative impact analysis model, including an abnormal user impact quantitative model and an abnormal station area impact quantitative model, which are used to locate the main abnormal users and main abnormal stations for governance.

[0009] Furthermore, in step S1, the daily line loss rate type is divided according to the daily line loss rate result of the substation area and the theoretical value of an index of the substation area, including ultra-high loss, high loss, small negative loss, negative loss and uncalculated; Ultra-high loss: daily line loss rate ≥ upper threshold + 20%, high loss: upper threshold ≤ daily line loss rate < upper threshold + 20%, small negative loss: -1% ≤ daily line loss rate < 0%, negative loss: daily line loss rate < -1%; Calculate the volatility of the area based on the daily line loss results of the area, including stability, fluctuation and suddenness; Stable: abnormal line loss days in a week ≥ 3 and abnormal dispersion coefficient in the past 60 days < 3, Fluctuation: 1≤ abnormal line loss days in a week < 3 and abnormal dispersion coefficient in the past 60 days ≥ 3, Sudden: abnormal line loss on the day and non-stable and fluctuating areas; The calculation formula of abnormal dispersion coefficient is: , in, Indicates the line loss rate of the area on the i-th day, N indicates the data period required to analyze the volatility of the area, the default is 60 days, and the abnormal dispersion coefficient The higher it is, the greater the abnormal fluctuation of daily line losses in the station area.

[0010] Furthermore, in step S2, the model training is based on the classification model of unqualified characteristics of the substation area and the abnormal factor library. For each type of substation area, a method combining the Pearson correlation coefficient and decision tree regression is used to screen abnormal factors to improve the efficiency and stability of model training. The ensemble learning XGBoost algorithm is used to train the model to realize the line loss prediction of abnormal substation areas, as follows: The training set is defined as a data set containing m features n. , the prediction output of the ensemble learning XGBoost algorithm training model is expressed as ;in, represents the overall prediction function of the model, which is composed of the sum of the prediction outputs of all K trees. represents the regression tree, K is the number of regression trees, Represents the feature vector of a sample of input and outputs the predicted value Add the predicted values ​​of K regression trees; XGBoost's objective function adds a regularization term based on the GBDT loss function. , which is the sum of the regularization values ​​of all regression trees. The objective function of a tree is: ; in, is the L1 regularization coefficient of the number of leaf nodes, is the number of leaf nodes in the tree, is the L2 regularization coefficient of the leaf node weight, is the weight of the jth leaf node of the tth tree; For the The actual line loss rate of samples is For the The sum of the predicted values ​​of the previous trees, For the The feature vector of the samples, For the The predicted value of the tree.

[0011] Furthermore, in step S2, the abnormal sorting constructs a quantitative abnormal sorting model based on the SHAP algorithm, calculates the marginal contribution of each characteristic factor to the line loss rate of the substation area, and quantifies the influence weight of each characteristic factor on the line loss rate, as follows: ; in, It represents the predicted value of line loss in the substation area. is a linear function of the SHAP value, which means mapping the SHAP value to the prediction result of the model. is the baseline value, which indicates the model prediction result when there is no feature influence. represents the SHAP value of the i-th feature, indicating the contribution of this feature to the model prediction results. , which indicates how many of the n features are included in the decision path of the sample; for a sample, if feature k is not in its decision path, the SHAP value of the corresponding feature is 0, that is, , indicating that the feature k will not be attributed to the sample and has no contribution to the final prediction value; The calculation formula of the contribution value of the area characteristic factor is as follows: in, For the The contribution of each characteristic factor to the prediction of the line loss in the substation area is the weighted sum of all possible combinations of characteristic values; is the number of characteristic factors, is a subset of the model application feature factors, is the vector of eigenvalues ​​of all instances, yes Characteristic factors not included in the set, It is a collection and characteristic factors The prediction of the union features of It is a collection The prediction of the eigenvalues ​​in the set Marginalization on features not included in is the number of combinations, indicating the number of combinations of |S| features selected from n−1 features.

[0012] Further, in step S2, the sorting results are summarized according to the calculated abnormal causes of all substations, and the importance of each abnormal cause in each substation class is obtained according to the substation classification, and all important abnormalities under the substation are standardized to obtain the abnormal weight. The specific formula is as follows: ; in, Indicates the importance of abnormal cause i, represents the total number of the e-th type of stations, Indicates the importance of abnormal cause i in the e-th type station area.

[0013] Further, in step S3, the Bayesian network model is an indeterminate causal association model, and the influence degree of abnormal X on abnormal Y under the Bayesian network is defined as the probability of Y when X occurs compared with the probability of Y when X does not occur, and the formula is as follows: ; The Bayesian network model can learn different combinations of anomaly X to anomaly Y under different anomaly combinations group(a), and the occurrence of other anomalies others(o) outside the anomaly group(a) is unknown.

[0014] Furthermore, in step S4, the abnormal user impact quantification model analyzes the abnormal values ​​of the user's historical data, and replaces the data that exceeds the normal range by 1.5 times with the upper and lower limits of the normal range; using the user's historical data, an exponential smoothing algorithm is used to estimate the user's power consumption when the abnormality occurs as the user's loss reduction space; for users who cannot converge, the average power consumption in the same period is used as the user's loss reduction space; The abnormal substation impact quantification model obtains the data on the power consumption affected by all abnormal users in the substation with unqualified line loss based on the quantification output results of the abnormal user impact; calculates the impact quantification of terminals and substations with reference to the abnormal user impact quantification rules, and obtains the data on the power consumption affected by all abnormal terminals and the power consumption affected by substations in the substation with unqualified line loss; obtains all abnormal impacts on the power consumption in the substation with unqualified line loss; obtains the loss reduction space for a substation with unqualified line loss in one indicator.

[0015] The present invention also provides a line loss intelligent diagnosis device based on comprehensive analysis of machine learning methods, using the above-mentioned line loss intelligent diagnosis method based on comprehensive analysis of machine learning methods, the line loss intelligent diagnosis device includes: The unqualified characteristics classification module of the substation area is used to classify the abnormal substation area by dividing the daily line loss rate type according to the daily line loss rate results of the substation area and the theoretical value of an indicator of the substation area, and calculating the substation area volatility according to the daily line loss results of the substation area; The main cause analysis module of line loss failure uses an integrated learning model and SHAP algorithm quantification model to obtain the abnormal causes of line loss failure in the substation area; The abnormal correlation main cause analysis module uses the Bayesian network model to analyze the relationship between abnormal causes and locate the main cause of abnormal correlation; The abnormal quantitative impact analysis module is used to locate abnormal users and abnormal substations for management through the abnormal user impact quantitative model and the abnormal substation impact quantitative model.

[0016] The present invention can achieve the following technical effects: The present invention constructs models such as classification of unqualified characteristics of substations, analysis of main causes of unqualified line losses, analysis of main causes of abnormal correlation, and quantitative impact analysis of abnormalities. It uses machine learning models such as the SHAP algorithm analysis model and the Bayesian network model, and combines the experience of business experts to form a set of quantitative evaluation and optimal management solutions for line loss abnormalities in substations. Under the premise of limited resources, it can efficiently focus on and solve key abnormal problems, so as to achieve the goal of reducing line losses and improving efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of a line loss intelligent diagnosis method based on comprehensive analysis by machine learning method provided in accordance with an embodiment of the present invention. DETAILED DESCRIPTION

[0018] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. In the following description, the same modules are represented by the same reference numerals. In the case of the same reference numerals, their names and functions are also the same. Therefore, the detailed description thereof will not be repeated.

[0019] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and do not constitute a limitation of the present invention.

[0020] The embodiment of the present invention provides a line loss intelligent diagnosis method based on comprehensive analysis by machine learning method, the process is as follows: Figure 1 As shown, the following steps are included: 1. Construct a classification model for unqualified characteristics of substations.

[0021] Since each substation has its own characteristics, classification modeling is used to reduce model complexity and model training calculations, and at the same time improve the application efficiency of the model based on the characteristics of different abnormal substations. Therefore, a substation unqualified characteristic classification model is constructed based on line loss volatility and daily line loss rate, and each abnormal substation is classified into 15 categories, as follows: 11. The daily line loss rate types are divided according to the daily line loss rate results of the substation area and the theoretical value of one index of each substation area, including: 111) Ultra-high loss: daily loss rate ≥ upper threshold + 20%; 112) High loss: upper threshold limit ≤ daily loss rate < upper threshold limit + 20%; 113) Small negative loss: -1%≤Daily loss rate<0%; 114) Negative loss: daily loss rate <-1%; 115) cannot be counted.

[0022] 12. Calculate the volatility of the area based on the daily line loss results of the area, including: 121) Stable: The number of days with abnormal line loss in a week is ≥ 3 and the abnormal dispersion coefficient in the past 60 days is < 3; 122) Fluctuation: 1≤The number of abnormal days of line loss in a week is <3 and the abnormal dispersion coefficient in the past 60 days is ≥3; 123) Breaking news: The daily line loss is abnormal and unstable, with fluctuations.

[0023] The calculation formula of abnormal dispersion coefficient is: , in, Indicates the line loss rate of the area on the i-th day, N indicates the data period required to analyze the volatility of the area, the default is 60 days, and the abnormal dispersion coefficient The higher it is, the greater the abnormal fluctuation of daily line losses in the station area.

[0024] 2. Construct a model to analyze the main causes of line loss failure.

[0025] Construct a line loss failure main cause analysis model to locate the main abnormal causes of line loss failure in the substation area, including model training, abnormal sorting and sorting result summary, as follows: 21. Model training.

[0026] Based on the unqualified characteristic classification model of the substation and the abnormal factor library, for each type of substation, the Pearson correlation coefficient and decision tree regression method are combined to screen the abnormal factors to improve the efficiency of model training and the stability of the model. The integrated learning XGBoost algorithm is used to train the model to realize the line loss prediction of abnormal substations.

[0027] Among them, the Pearson correlation coefficient is a method that can help understand the relationship between features and dependent variables. This method measures the linear correlation between variables, and the result range is [-1, 1]. An obvious defect of this method is that it is only sensitive to linear relationships. If the relationship is nonlinear, even if the two variables have a one-to-one relationship, the correlation may be close to 0; decision tree regression is a tree-based method that can reflect the nonlinear relationship between features and dependent variables, and is better at modeling nonlinear relationships. Each node in the decision tree is a condition about a certain feature. The purpose is to divide the data set into two according to different dependent variables and determine the division node. For decision trees, you can calculate how much impurity each feature has reduced, and use the reduced impurity as the value of feature selection.

[0028] The ensemble learning algorithm is used to realize the prediction of line loss in the substation area. XGBoost (Extreme Gradient Boosting) is an algorithm that improves and efficiently realizes the gradient boosting tree (GBDT). It uses the boosting ensemble idea, including the additive model (a strong learner is composed of a series of weak learners linearly added) and the forward distribution algorithm (the next learner is trained on the basis of the previous learner).

[0029] The training set is defined as a data set containing m features n. , the prediction output of the ensemble learning XGBoost algorithm training model is expressed as ; in, represents the overall prediction function of the model, which is composed of the sum of the prediction outputs of all K trees. represents the regression tree, K is the number of regression trees, Represents the feature vector of a sample of input and outputs the predicted value Add the predicted values ​​of K regression trees; XGBoost's objective function adds a regularization term based on the GBDT loss function. , which is the sum of the regularization values ​​of all regression trees. The objective function of a tree is: ; in, is the L1 regularization coefficient of the number of leaf nodes, is the number of leaf nodes in the tree, is the L2 regularization coefficient of the leaf node weight, is the weight of the jth leaf node of the tth tree; For the The actual line loss rate of samples, For the The sum of the predictions of the previous trees, For the The feature vector of the samples, For the The predicted value of the tree.

[0030] The loss function of the XGBoost algorithm adopts the second-order Taylor expansion and adds a regularization term, which has the characteristics of high accuracy and low overfitting.

[0031] 22. Abnormal sorting.

[0032] Based on the SHAP algorithm, a quantitative anomaly sorting model is constructed to calculate the marginal contribution of each abnormal characteristic factor to the line loss rate of the substation area, and to quantify the influence weight of each abnormal characteristic factor on the line loss rate. The core idea of ​​SHAP is based on the distribution method of Shapley value, which mainly means that the value obtained should be consistent with its own contribution. Through this idea, we can quantify the marginal contribution of each characteristic factor to line loss prediction. Since Shapley satisfies marginality, symmetry, effectiveness and additivity, it has a solid theoretical foundation and ensures that the prediction results are fairly distributed among the abnormal characteristic factors. The specific process is as follows: 221. Input the characteristic factors of each prediction data and the line loss rate value obtained by model training prediction.

[0033] 222. Based on the SHAP algorithm, an additive feature attribution mathematical expression is constructed to calculate the baseline of the entire model when no feature factors are involved in the calculation (the average prediction value, the bias term of the model itself, and the sum of the expected values ​​of all tree root nodes).

[0034] The mathematical expression of the additivity attribute is as follows: ; in It represents the predicted value of line loss in the substation area. is a linear function of the SHAP value, which means mapping the SHAP value to the prediction result of the model. is the baseline value, which indicates the model prediction result when there is no feature influence. represents the SHAP value of the i-th feature, indicating the contribution of this feature to the model prediction results. , which indicates how many of the n features are included in the decision path of the sample; for a sample, if feature k is not in its decision path, the SHAP value of the corresponding feature is 0, that is, , indicating that the feature k will not be attributed to the sample and has no contribution to the final prediction value.

[0035] 223. According to the calculation formula of the contribution value of the characteristic factor of the substation, the contribution value of the characteristic factor is calculated respectively. According to the distribution of the data and the decision tree structure, the difference between the predicted value and the average predicted value is fairly distributed among the characteristic values ​​of each substation to complete the quantification of each characteristic indicator.

[0036] The calculation formula of the contribution value of the area characteristic factor is as follows: in, For the The contribution of each characteristic factor to the prediction of the line loss in the substation area is the weighted sum of all possible combinations of characteristic values; is the number of characteristic factors, is a subset of the model application feature factors, is the vector of eigenvalues ​​of all instances, yes Characteristic factors not included in the set, It is a collection and characteristic factors The prediction of the union features of It is a collection The prediction of the eigenvalues ​​in the set Marginalization on features not included in is the number of combinations, indicating the number of combinations of |S| features selected from n−1 features.

[0037] After determining the contribution value of the abnormal characteristic factor to the line loss rate through the above method, the contribution ranking of various abnormalities is sorted and output. The feature with a contribution greater than a certain threshold is taken as the main cause of the line loss management abnormality.

[0038] 23. Sorting results summary.

[0039] According to the calculated main causes of all the abnormalities in the substation, the importance of each abnormality in each substation class is obtained according to the substation classification. Finally, all important abnormalities in the substation need to be standardized to obtain the abnormality weight. The specific formula is as follows: ; in, Indicates the importance of abnormal cause i, represents the total number of the e-th type of stations, Indicates the importance of abnormal cause i in the e-th type station area.

[0040] 3. Construct an abnormal correlation main cause analysis model.

[0041] By building an association rule model between anomaly types, the "causal" relationship between anomalies is analyzed. Since the causes of major anomalies are diverse and many information cannot be recorded, such as electricity theft, it is impossible to truly determine the causal relationship between anomalies. Therefore, an algorithm is used to analyze the association between anomalies to score the importance of each anomaly cause, helping operation and maintenance personnel determine the main cause.

[0042] Specifically, a Bayesian network model is constructed by combining the abnormal group data with expert opinions. The relationship between the abnormal causes of each user's power consumption is analyzed online, and quantitative analysis results such as the probability of abnormal occurrence and the degree of improvement are given. Different abnormal group models will analyze different results. Analysis of abnormal correlation main causes, a Bayesian network model is constructed by combining the abnormal group data with expert opinions. The relationship between the abnormal causes of each substation or user is analyzed online, and quantitative analysis results such as the probability of abnormal occurrence and the degree of improvement are given. Different abnormal group models will analyze different results.

[0043] Bayesian network, also known as belief network, is a probabilistic graph model. Graph models can be constructed from data or expert opinions. Models are usually applied in predictive analysis, abnormal data detection, data null value completion, etc. Compared with other machine learning models, Bayesian network is good at giving prediction results when information conditions are missing. Mobile phone signal stabilization technology and gene identification, which are closely related to life, are all applications of Bayesian network.

[0044] The Bayesian network is formally a directed acyclic graph, where nodes represent random variables. Directed edges between nodes represent conditional dependencies, from parent nodes to child nodes. Each node is associated with a probability function, the input of which is a set of specific values ​​of the random variable represented by the parent node of the node, and the output is the probability value of the random variable represented by the current node.

[0045] ; The graph structure of the network is calculated by combining the A* algorithm with the d-separation algorithm, and then the conditional probability distribution on each node is obtained by counting.

[0046] Bayesian network is also an indeterminate causal association model that can learn and reason under the condition of limited, incomplete and uncertain information. Therefore, it is widely used in the fields of fault diagnosis and maintenance decision-making. The main use is information completion, such as the probability of another node occurring when any one or several nodes are known. In order to adapt to the analysis of the main causes of abnormalities, the influence of abnormality X on abnormality Y under the Bayesian network is defined as the probability of Y when X occurs divided by the probability of Y when X does not occur. The formula is as follows: ; The network can learn different combinations of anomaly X to anomaly Y under different anomaly combinations group(a). And the occurrence of other anomalies others(o) outside the anomaly group(a) is unknown.

[0047] 4. Construct an abnormal quantitative impact analysis model.

[0048] Abnormal quantitative impact analysis, including the quantitative model of abnormal user impact and the quantitative model of abnormal station impact, locates the main abnormal users and main abnormal stations for governance.

[0049] 41. Quantification model of the impact of abnormal users. With the help of models such as line loss main cause analysis, substation unqualified characteristic classification, and abnormal user loss reduction space analysis, the line loss main cause impact weight coefficient and the abnormal user's loss reduction power are output, and then the line loss impact of each abnormal user is calculated by multiplying the two variables. The user's loss reduction power is analyzed at the user level for the substations with daily line loss abnormalities and abnormal users. Specifically, the abnormal loss power is calculated as the loss reduction space for the meter with abnormal power meter. The time series algorithm is used to calculate the power of the abnormal meter as the loss reduction space for the abnormal collection. The specific analysis process is as follows: 411. Analyze the abnormal values ​​of each user's historical data for the past two months. Replace the data that exceeds the normal range by 1.5 times with the upper and lower limits of the normal range.

[0050] 412. Utilize the user's historical data and use the exponential smoothing algorithm to estimate the user's power consumption when an anomaly occurs as the user's loss reduction space.

[0051] 413. For users who cannot converge, the average of weekly electricity consumption in the past 8 weeks is used as the user's loss reduction space.

[0052] 42. Quantitative model of abnormal substation impact. With the help of the quantitative analysis of abnormal user / substation impact and the one-substation-one-indicator model, the user / substation abnormal loss reduction power and substation line loss target value are output. The impact of substations with unqualified line losses is quantified through these two values, and finally the main abnormal management substation is located according to the size of the loss reduction space. The specific analysis process is as follows: 421. According to the quantified output results of the impact of abnormal users, the power consumption data A of all abnormal users in the substation with unqualified line loss is obtained.

[0053] 422. Calculate the impact of terminals and substations by referring to the quantification rules for the impact of abnormal users, and obtain the power data B affected by all abnormal terminals in substations with unqualified line losses and the power data C affected by substations.

[0054] 423. Obtain all abnormal power consumption in the substation with unqualified line loss: A+B+C.

[0055] 424. Obtain the space for reducing the loss of a substation where the line loss indicator is unqualified.

[0056] By building models such as classification of unqualified characteristics of substations, analysis of main causes of unqualified line losses, analysis of main causes of abnormal correlation, and analysis of quantitative impact of abnormalities, and using machine learning models such as the SHAP algorithm analysis model and the Bayesian network model, combined with the experience of business experts, a set of quantitative evaluation and optimal governance solutions for abnormal line losses in substations is formed. The present invention can achieve efficient focus and resolution of key abnormal problems under the premise of limited resources, so as to achieve the goal of reducing line losses and improving efficiency.

[0057] The present invention also provides an intelligent line loss diagnosis device based on comprehensive analysis by machine learning methods, which adopts the above-mentioned intelligent line loss diagnosis method based on comprehensive analysis by machine learning methods, including: an unqualified characteristic classification module for substations, which divides the daily line loss rate type according to the daily line loss rate results of the substations and the theoretical value of an indicator of the substation, and calculates the substation volatility according to the daily line loss results of the substations, so as to classify abnormal substations; an unqualified line loss main cause analysis module, which is used to obtain the abnormal causes of unqualified line losses in the substations through an integrated learning model and a SHAP algorithm quantification model; an abnormal correlation main cause analysis module, which analyzes the relationship between abnormal causes through a Bayesian network model, so as to locate the main causes of abnormal correlation; an abnormal quantitative impact analysis module, which is used to locate abnormal users and abnormal substations to be managed through an abnormal user impact quantification model and an abnormal substation area impact quantification model.

[0058] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0059] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

[0060] The above specific implementations of the present invention do not constitute a limitation on the protection scope of the present invention. Any other corresponding changes and modifications made based on the technical concept of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A line loss intelligent diagnosis method based on comprehensive analysis of machine learning methods, characterized in that: The steps include: S1. Construct a classification model for substation unqualified characteristics based on line loss volatility and daily line loss rate, including classifying daily line loss rate types according to the daily line loss rate results of the substation and the theoretical value of an indicator of the substation, and calculating the substation volatility according to the daily line loss results of the substation, which is used to classify abnormal substations; S2. Construct a main cause analysis model for line loss failure, including model training, abnormal sorting, and sorting result summary, to obtain the abnormal causes of line loss failure in the substation area; S3. Construct an abnormal correlation main cause analysis model, and analyze the relationship between abnormal causes by constructing a Bayesian network model to locate the main abnormal cause; S4. Construct an abnormal quantitative impact analysis model, including an abnormal user impact quantitative model and an abnormal station area impact quantitative model, which are used to locate the main abnormal users and main abnormal stations for governance.

2. The line loss intelligent diagnosis method based on comprehensive analysis by machine learning method according to claim 1 is characterized in that: In step S1, the daily line loss rate type is divided according to the daily line loss rate result of the area and the theoretical value of an index of an area, including ultra-high loss, high loss, small negative loss, negative loss and uncalculated; Ultra-high loss: daily line loss rate ≥ upper threshold + 20%, high loss: upper threshold ≤ daily line loss rate < upper threshold + 20%, small negative loss: -1% ≤ daily line loss rate < 0%, negative loss: daily line loss rate < -1%; Calculate the volatility of the area based on the daily line loss results of the area, including stability, fluctuation and suddenness; Stable: abnormal line loss days in a week ≥ 3 and abnormal dispersion coefficient in the past 60 days < 3, Fluctuation: 1≤ abnormal line loss days in a week < 3 and abnormal dispersion coefficient in the past 60 days ≥ 3, Sudden: abnormal line loss on the day and non-stable and fluctuating areas; The calculation formula of the abnormal dispersion coefficient is: ， in, Indicates the line loss rate of the area on the i-th day, N indicates the data period required to analyze the volatility of the area, the default is 60 days, and the abnormal dispersion coefficient The higher it is, the greater the abnormal fluctuation of daily line losses in the station area.

3. The line loss intelligent diagnosis method based on comprehensive analysis by machine learning method according to claim 2 is characterized in that: In step S2, the model training is based on the substation unqualified characteristic classification model and the abnormal factor library. For each type of substation, the abnormal factors are screened by combining the Pearson correlation coefficient and the decision tree regression method to improve the efficiency and stability of the model training. The ensemble learning XGBoost algorithm is used to train the model to realize the line loss prediction of the abnormal substation, as follows: The training set is defined as a data set containing m features n. , the prediction output of the ensemble learning XGBoost algorithm training model is expressed as ; in, represents the overall prediction function of the model, which is composed of the sum of the prediction outputs of all K trees. represents the regression tree, K is the number of regression trees, Represents the feature vector of a sample of input and outputs the predicted value Add the predicted values ​​of K regression trees; XGBoost's objective function adds a regularization term based on the GBDT loss function. , which is the sum of the regularization values ​​of all regression trees. The objective function of a tree is: ; in, is the L1 regularization coefficient of the number of leaf nodes, is the number of leaf nodes in the tree, is the L2 regularization coefficient of the leaf node weight, is the weight of the jth leaf node of the tth tree; For the The actual line loss rate of samples, For the The sum of the predictions of the previous trees, For the The feature vector of the samples, For the The predicted value of the tree.

4. The line loss intelligent diagnosis method based on comprehensive analysis by machine learning method according to claim 3 is characterized in that: In step S2, the abnormal sorting constructs a quantitative abnormal sorting model based on the SHAP algorithm, calculates the marginal contribution of each characteristic factor to the line loss rate of the substation area, and quantifies the influence weight of each characteristic factor on the line loss rate, as follows: Input the characteristic factors of each prediction data and the line loss rate value obtained by model training prediction; construct an additive characteristic attribution mathematical expression based on the SHAP algorithm, and calculate the baseline of the entire model when none of the characteristic factors participate in the calculation; calculate the contribution value of the characteristic factors according to the calculation formula of the contribution value of the substation characteristic factors; The mathematical expression of the additivity attribute is as follows: ； in It represents the predicted value of line loss in the substation area. is a linear function of the SHAP value, which means mapping the SHAP value to the prediction result of the model. is the baseline value, which indicates the model prediction result when there is no feature influence. represents the SHAP value of the i-th feature, indicating the contribution of this feature to the model prediction results. , which indicates how many of the n features are included in the decision path of the sample; for a sample, if feature k is not in its decision path, the SHAP value of the corresponding feature is 0, that is, , indicating that the feature k will not be attributed to the sample and has no contribution to the final prediction value; The calculation formula of the contribution value of the area characteristic factor is as follows: in, For the The contribution of each characteristic factor to the prediction of the line loss in the substation area is the weighted sum of all possible combinations of characteristic values; is the number of characteristic factors, is a subset of the model application feature factors, is the vector of eigenvalues ​​of all instances, yes Characteristic factors not included in the set, It is a collection and characteristic factors The prediction of the union features of It is a collection The prediction of the eigenvalues ​​in the set Marginalization on features not included in is the number of combinations, indicating the number of combinations of |S| features selected from n−1 features.

5. The line loss intelligent diagnosis method based on comprehensive analysis by machine learning method according to claim 4 is characterized in that: In step S2, the ranking result is summarized according to the calculated abnormal causes of all substations, and the importance of each abnormal cause in each substation class is obtained according to the substation classification, and all important abnormalities under the substation are standardized to obtain the abnormal weight. The specific formula is as follows: ;in, Indicates the importance of abnormal cause i, represents the total number of the e-th type of stations, Indicates the importance of abnormal cause i in the e-th type station area.

6. The line loss intelligent diagnosis method based on comprehensive analysis by machine learning method according to claim 1, characterized in that: In step S3, the Bayesian network model is an indeterminate causal association model, and the degree of influence of abnormality X on abnormality Y under the Bayesian network is defined as the probability of Y when X occurs compared to the probability of Y when X does not occur, and the formula is as follows: ； The Bayesian network model can learn different combinations of anomaly X to anomaly Y under different anomaly combinations group(a), and the occurrence of other anomalies others(o) outside the anomaly group(a) is unknown.

7. The line loss intelligent diagnosis method based on comprehensive analysis by machine learning method according to claim 1, characterized in that: In step S4, the abnormal user impact quantification model analyzes the abnormal values ​​of the user's historical data, and replaces the data that exceeds the normal range by 1.5 times with the upper and lower limits of the normal range; using the user's historical data, an exponential smoothing algorithm is used to estimate the user's power consumption when the abnormality occurs as the user's loss reduction space; for users who cannot converge, the average power consumption in the same period is used as the user's loss reduction space; The abnormal substation impact quantification model obtains the data on the power consumption affected by all abnormal users in the substation with unqualified line loss according to the quantification output results of the abnormal user impact; calculates the impact quantification of terminals and substations with reference to the quantification rules of the abnormal user impact, and obtains the data on the power consumption affected by all abnormal terminals and the power consumption affected by the substation in the substation with unqualified line loss; obtains the power consumption affected by all abnormalities in the substation with unqualified line loss; and obtains the loss reduction space for the substation with unqualified line loss in one indicator.

8. A line loss intelligent diagnosis device based on comprehensive analysis by machine learning method, adopting the line loss intelligent diagnosis method based on comprehensive analysis by machine learning method as claimed in any one of claims 1 to 7, characterized in that: The device comprises: The unqualified characteristics classification module of the substation area is used to classify the abnormal substation area by dividing the daily line loss rate type according to the daily line loss rate results of the substation area and the theoretical value of an indicator of the substation area, and calculating the substation area volatility according to the daily line loss results of the substation area; The main cause analysis module of line loss failure uses an integrated learning model and SHAP algorithm quantification model to obtain the abnormal causes of line loss failure in the substation area; The abnormal correlation main cause analysis module uses the Bayesian network model to analyze the relationship between abnormal causes and locate the main cause of abnormal correlation; The abnormal quantitative impact analysis module is used to locate abnormal users and abnormal substations for management through the abnormal user impact quantitative model and the abnormal substation impact quantitative model.

Images