A power grid power distribution overload prediction method

Abstract

The application provides a power grid power distribution heavy overload prediction method, and belongs to the technical field of power grid operation and maintenance. The power grid power distribution heavy overload prediction method comprises the following steps: S1, building a system structure; S2, improving an HTAP database; S3, starting prediction execution; S4, randomly selecting and calling a feature engineering algorithm by using SQL; S5, selecting an algorithm for prediction, and executing to obtain a prediction model M; S6, predicting production domain synchronized load and environmental data to obtain an early warning result R; S7, pushing R and making decisions and maintenance; S8, comparing R; S9, adjusting the existing feature engineering algorithm and machine learning algorithm; and S10, re-performing the prediction execution process. The method does not need any data migration, can more efficiently and accurately obtain a prediction result, thereby more effectively assisting in the formulation of a power distribution network work plan and a planning scheme, reducing the probability of heavy overload of a distribution transformer, and improving the power grid operation safety and power supply service quality.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power grid operation and maintenance, in particular to a power grid power distribution heavy overload prediction method. BACKGROUND

[0002] The heavy overload prediction of the distribution network is an important technical field in the field of power grid operation and maintenance. Traditionally, there are various methods to deal with the prediction problem. For a considerable period of time, the artificial method based on expert experience is mainly used. This method has a serious change in effect depending on personal experience, low accuracy and poor efficiency. In recent years, with the continuous growth and upgrading of online monitoring equipment of distribution transformers and other equipment, the big data analysis and prediction method based on massive historical load and heavy overload sample data has developed rapidly. Various statistical analysis methods and machine learning algorithms are used for strong correlation load feature selection and heavy overload prediction, which plays a great role in promoting the development and implementation of maintenance work plans based on the heavy overload early warning results of the distribution network.

[0003] The current big data analysis technology method has the following problems: 1. The historical data of the prediction analysis domain needs to be passively collected from the transaction production domain regularly and timely. In addition to the huge workload, the real-time performance of the prediction cannot be guaranteed, and the deviation between the prediction results and the actual events cannot be fed back in real time; 2. The traditional big data analysis tools and methods have long iteration period and low efficiency, and are separated from the data storage domain (database). In most cases, related data needs to be moved to a special machine learning or analysis domain for analysis, resulting in two migrations of historical data from the production domain to the storage domain (database) and from the storage domain to the analysis domain, which is almost impossible to bear in the case of large data volume; 3. The selected related load features and machine learning algorithms for prediction analysis execution have various attempts and different effects, but once selected, they cannot be changed and cannot be adjusted in real time according to the actual prediction effect. However, the actual situation is that different feature sets and algorithms cannot achieve the same effect as research and testing in practice for different situations and environments, and need to be dynamically adjusted; 4. The prediction based on big data analysis generally needs a certain number of historical samples to improve accuracy. However, most cases of heavy overload of the distribution network are small sample events, and it is difficult to fundamentally improve the prediction accuracy by selecting algorithms. Therefore, there is an urgent need for more efficient and accurate prediction methods in practice. SUMMARY

[0004] In order to make up for the above shortcomings, the present application provides a power grid power distribution heavy overload prediction method, which aims to improve the problems raised in the above background technology.

[0005] The present application is implemented as follows: a power grid power distribution heavy overload prediction method, comprising the following steps:

[0006] S1: Build a distribution network overload prediction system structure based on the HTAP database, place the load production and transaction data of each district power grid in the transaction domain, place the prediction analysis data in the analysis domain, and complete the real-time synchronization of large amounts of data from the transaction domain to the analysis domain;

[0007] S2: Improve the HTAP database, in the analysis domain, in the form of a database SQL user-defined function UDF, use the MPP feature to increase the implementation of feature engineering algorithms and machine learning algorithms;

[0008] S3: According to the typical event occurring in the production domain, the prediction execution starts in real time or at regular intervals;

[0009] S4: In the analysis domain of the HTAP database, randomly select and use SQL to call the feature engineering algorithm, select the relevant load and environmental feature set for prediction, and select the global real-time and historical data involved;

[0010] S5: In the analysis domain of the HTAP database, randomly select and use SQL to call the machine learning algorithm, select the algorithm for prediction, and perform prediction based on the full data set selected in step S4 to obtain the prediction model M;

[0011] S6: Use the prediction model M to predict the load and environmental data of a certain period of time synchronized in real time from the production domain to obtain the early warning result R;

[0012] S7: Push the early warning result R to the business end for decision-making and maintenance;

[0013] S8: Compare the early warning result R with the overload occurrence of each district distribution transformer in the business end;

[0014] S9: In the analysis domain of the HTAP database, adjust the existing feature engineering algorithm and machine learning algorithm according to the feedback difference set;

[0015] S10: According to the new typical event occurring in the production domain or at regular intervals, the prediction execution process is performed again.

[0016] In one preferred technical solution of the present application, in S2, when the feature engineering algorithm is added, initially, principal component analysis PCA and association rule method Apriori are added, respectively denoted as FUDF_PCA and FUDF_APRIORI, and later increased according to demand.

[0017] In one preferred technical solution of the present application, in S2, when the machine learning algorithm is added, initially, logistic regression, support vector machine, and typical deep learning CNN algorithm are added, respectively denoted as MUDF_LR, MUDF_SVM, and MUDF_RESNET, and later increased according to demand.

[0018] In a preferred technical solution of the present application, in step S4, the feature engineering algorithm called by SQL is selected as FUDF_PCA.

[0019] In a preferred technical solution of the present application, in step S5, the machine learning algorithm called by SQL is selected as MUDF_LR.

[0020] In a preferred technical solution of the present application, in step S8, the method for comparing the warning result R with the occurrence of the heavy overload of each transformer in the business end area is to perform a difference between the predicted occurrence probability value P and the actual value, and feed the difference set to the HTAP database analysis domain.

[0021] In a preferred technical solution of the present application, in step S9, the adjustment rule is: if the average value of the feedback difference set is greater than 0.5, the operator is replaced, otherwise, the operator is not replaced, and the replacement of the operator adopts a random method.

[0022] The beneficial effects of the present application are: 1. The integrated prediction system structure design based on the HTAP database is completed, which is simple and convenient, and avoids the migration process of a large amount of data in the traditional method;

[0023] 2. Real-time active synchronization of batch load and environmental production data is not needed, the data set is naturally and automatically formed, the real-time and accuracy are increased;

[0024] 3. The limit analysis and incremental update characteristics of the HTAP database analysis domain greatly improve the execution efficiency of the analysis and prediction;

[0025] 4. The difference feedback algorithm adjustment strategy of the prediction result and the actual event makes up for the accuracy deficiency of small sample prediction and the applicability of the algorithm, and greatly improves the accuracy;

[0026] 5. The feature engineering and machine learning model prediction method embedded in the database directly realizes all the model prediction processes by SQL, which is easy to realize and convenient to master, and has great practicability and popularization in practice.

[0027] The present application provides a new system structure and method for distribution network heavy overload prediction in practice, which is completely realized in the database without any data migration, is convenient and easy to implement, and can more efficiently and accurately obtain the prediction result, thereby more effectively assisting the formulation of distribution network work plan and planning scheme, reducing the heavy overload occurrence probability of distribution transformers, and improving the power grid operation safety and power supply service quality. BRIEF DESCRIPTION OF DRAWINGS

[0028] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation to the scope, and for those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0029] Figure 1 is a flow chart provided by the embodiments of the present application;

[0030] Figure 2 is a logic function chart provided by the embodiments of the present application. DETAILED DESCRIPTION

[0031] In order to make the objects, technical solutions and advantages of the embodiments of the present application more clear, the following will combine the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0032] EMBODIMENT

[0033] The traditional big data analysis technology method generally has many problems. Therefore, the inventors have found, through long-term observation and research, a new production domain + analysis domain, real-time data + historical data, and real-time event feedback integrated distribution network overload prediction architecture model method.

[0034] Figure 1 A flow chart of a power grid distribution overload prediction method is provided. Please refer to Figure 1 The present application provides a power grid distribution overload prediction method, which comprises the following steps:

[0035] S1: Build a distribution network overload prediction system structure based on an HTAP database, place the load production (transaction) data of each district power grid in the transaction domain, place the prediction analysis data in the analysis domain, and complete real-time synchronization of large amounts of data from the transaction domain to the analysis domain;

[0036] S2: Improve the HTAP database in its analysis domain, in the form of database SQL user-defined functions UDF, use MPP features, increase the implementation of feature engineering algorithms (initially increase principal component analysis PCA and association rule method Apriori, respectively denoted as FUDF_PCA, FUDF_APRIORI, and later can be increased according to demand), machine learning algorithms (initially increase logistic regression, support vector machine, typical deep learning CNN algorithm such as ResNet, etc. Denoted as MUDF_LR, MUDF_SVM, MUDF_RESNET, and later can be increased according to demand);

[0037] S3: According to the typical events (which can be customized according to business needs) occurring in the production domain, real-time or timed (such as every 15 minutes) driving prediction execution starts;

[0038] S4: In the HTAP database analysis domain, randomly select and use SQL to call feature engineering algorithms (such as FUDF_PCA) to select relevant load and environmental feature sets for prediction, and select the global real-time and historical data involved;

[0039] S5: In the HTAP database analysis domain, randomly select and use SQL to call machine learning algorithms (such as MUDF_LR) to select algorithms for prediction, and perform prediction based on the full data set selected in step S4, to obtain a prediction model M;

[0040] S6: Use the prediction model M to predict the load and environmental data of a certain period of time in the production domain in real time, to obtain a warning result R;

[0041] S7: Push the warning result R to the business end for decision-making and maintenance;

[0042] S8: Compare the warning result R with the occurrence of heavy overload of each transformer in the business end, that is, use the predicted probability value P and the actual value (0 or 1) to find the difference, and feed the difference set back to the HTAP database analysis domain;

[0043] S9: In the HTAP database analysis domain, according to the feedback difference set, adjust the existing feature engineering algorithms and machine learning algorithms. The invention designs a simple rule: if the average of the feedback difference set is greater than 0.5, replace the operator, otherwise do not replace; The replacement of the operator also uses a random method;

[0044] S10: According to the new typical events occurring in the production domain or the timing of re-prediction execution process.

[0045] It should be noted that the S9 and S10 steps above can be performed in parallel, that is, in the process of operator replacement, a new prediction event starts to be executed. In this case, the new prediction event is executed first, and then the difference operator replacement is performed.

[0046] Under this model:

[0047] 1. Use the bulk transaction data synchronization features of the multi-modal, mixed load HTAP database to synchronize the load data to the prediction analysis domain in real time, without the need for self-collection, and to conveniently and efficiently realize the prediction analysis of full-quantity real-time data, thereby greatly improving the prediction efficiency and accuracy;

[0048] 2. In the database analysis domain, a feature analysis and machine learning method library is designed to be embedded in the database, and a prediction model is generated by directly using SQL statements in the database analysis domain to complete the load feature selection and machine learning algorithm execution, and the real-time load and environmental data are predicted according to the obtained prediction model in the database to obtain early warning results, thereby further improving the efficiency;

[0049] 3. The column storage, large-scale parallel processing MPP and incremental limit analysis features of the HTAP database are used to realize strong correlation load feature selection and machine learning model execution, thereby achieving an extremely efficient prediction process execution and further improving the efficiency;

[0050] 4. The difference between the prediction result and the actual event is fed back to the prediction analysis domain in real time, and the related load feature set and machine learning algorithm of the previous round of prediction are automatically updated and adjusted, thereby to a large extent, the accuracy defects of traditional distribution network overload small sample prediction are compensated;

[0051] 5. For subsequent execution of the same machine learning algorithm, only the real-time updated sample parameters are used for incremental model updating, thereby maximizing the avoidance of full-quantity iterative prediction each time;

[0052] 6. The prediction model is executed to obtain the prediction result by using the method of driving the designed prediction model in real time combined with the timing driving method, thereby more effectively improving the real-time performance of the prediction support at critical moments.

[0053] The above only describes the preferred embodiments of the present application and is not used to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for predicting heavy overload in power grid distribution, characterized in that, Includes the following steps: S1: Build a distribution network heavy overload prediction system structure based on HTAP database, place the power grid load production and transaction data of each distribution area in the transaction domain, place the prediction analysis data in the analysis domain, and complete the real-time synchronization of large batches of data from the transaction domain to the analysis domain. S2: Improve the HTAP database by adding feature engineering and machine learning algorithms in its analysis domain, using database SQL user-defined functions (UDFs) and leveraging MPP features. S3: Based on the occurrence of typical events in the production domain, predictive execution is initiated in real time or at scheduled intervals. S4: In the HTAP database analysis domain, randomly select and call the feature engineering algorithm using SQL to select the relevant load and environmental feature set for prediction, and select the real-time and historical data of the entire domain involved. S5: In the HTAP database analysis domain, randomly select and call a machine learning algorithm using SQL, select the algorithm for prediction, and perform prediction based on the full dataset selected in step S4 to obtain the prediction model M. S6: Use the prediction model M to predict the load and environmental data synchronized from the production domain in real time for a certain period of time, and obtain the early warning result R; S7: Push the warning result R to the business end for decision-making and maintenance; S8: Compare the early warning result R with the heavy overload occurrence of distribution network transformers in each service area; S9: In the HTAP database analysis domain, adjust the existing feature engineering algorithm and machine learning algorithm based on the feedback difference set; S10: Re-execute the prediction process based on new typical events occurring in the production domain or at regular intervals.

2. The power grid distribution heavy overload prediction method according to claim 1, characterized in that, In S2, when adding feature engineering algorithms, two methods were initially added: Principal Component Analysis (PCA) and Association Rule Method Apriori, denoted as FUDF_PCA and FUDF_APRIORI, respectively. More will be added later as needed.

3. The power grid distribution heavy overload prediction method according to claim 1, characterized in that, In S2, when adding machine learning algorithms, initially add logistic regression, support vector machine, and typical deep learning CNN algorithms, denoted as MUDF_LR, MUDF_SVM, and MUDF_RESNET respectively. More can be added later as needed.

4. The power grid distribution heavy overload prediction method according to claim 1, characterized in that, In step S4, the feature engineering algorithm selected and called using SQL is FUDF_PCA.

5. The power grid distribution heavy overload prediction method according to claim 1, characterized in that, In step S5, the algorithm for machine learning that is selected and called using SQL is MUDF_LR.

6. The power grid distribution heavy overload prediction method according to claim 1, characterized in that, In step S8, the method for comparing the early warning result R with the heavy overload occurrence of distribution transformers in each area of ​​the business terminal is to calculate the difference between the predicted occurrence probability value P and the actual value, and then feed the difference set back to the HTAP database analysis domain.

7. The power grid distribution heavy overload prediction method according to claim 1, characterized in that, In step S9, the adjustment rule is as follows: if the mean of the feedback difference set is greater than 0.5, then the operator is replaced; otherwise, it is not replaced. The replacement of the operator is done using a random method.

Images