Distribution network common transformer voltage out-of-limit prediction method based on LightGBM ensemble learning algorithm

Abstract

The invention relates to a distribution network public transformer voltage out-of-limit prediction method based on a LightGBM ensemble learning algorithm. The method comprises the following steps of 1, performing data preprocessing on marketing distribution transformer outlet operation data, historical voltage out-of-limit data, scheduling line outlet operation data and power supply station operation and maintenance shift data; 2, establishing a machine learning model based on a LightGBM ensemble learning algorithm, and training historical data; and step 3, using the trained model to complete the prediction of the distribution transformer outlet voltage in a future time period, and according to a prediction result. According to the technical scheme, a more scientific and accurate quantitative result is provided to support the daily management work of the voltage of the transformer area.

Description

Technical Field

[0001] This invention relates to the field of power grid technology, and in particular to a method for predicting voltage overruns of distribution network transformers based on the LightGBM ensemble learning algorithm. Background Technology

[0002] Voltage is a crucial indicator of power quality and a vital economic and technical metric for the company's power grid operation. With the rapid expansion of the power grid and the continuous commissioning of new energy power generation equipment, the total power load has increased dramatically. In some areas, the load fluctuates significantly seasonally, and some low-voltage distribution lines are aging or have excessively long supply radii. This leads to situations where the voltage at both the beginning and end of distribution substations cannot simultaneously meet requirements, resulting in frequent voltage exceedances. Voltage exceedances not only affect customers' normal electricity use, but long-term overvoltage supply will also reduce the operating efficiency and lifespan of user equipment, with particularly pronounced impacts on industrial and agricultural production sectors that require high voltage stability.

[0003] Currently, the diagnosis and treatment of voltage overruns in distribution transformers are usually based on a small amount of past measured data, relying solely on manual experience for assessment and judgment, and then taking subsequent measures such as adjusting the voltage rating, cutting off the load, and adjusting reactive power compensation.

[0004] The shortcomings of existing methods for diagnosing and managing voltage over-limit issues in distribution transformers: (1) It is greatly affected by subjective human factors. Relying on human experience for evaluation cannot fully consider all aspects, resulting in poor accuracy and low work efficiency. (2) No definitive conclusion can be drawn because multiple influencing variables are involved, and the degree of influence of voltage influencing factors under different conditions in different transformer areas cannot be accurately determined. (3) Voltage fluctuations cannot be predicted in advance. Traditional methods for analyzing data exceeding limits are inherently lagging in terms of data sources and cannot predict reasonable adjustment ranges. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a method for predicting voltage overruns of distribution network transformers based on the LightGBM ensemble learning algorithm, providing more scientific and accurate quantitative results to support the daily management of transformer voltage.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for predicting voltage over-limit of distribution network transformers based on the LightGBM ensemble learning algorithm, comprising the following steps: Step 1: Perform data preprocessing on marketing distribution transformer outlet operation data, historical voltage over-limit data, dispatch line outlet operation data, and power supply station operation and maintenance data. Step 2: Build a machine learning model based on the LightGBM ensemble learning algorithm and train it on historical data; Step 3: Use the trained model to predict the distribution transformer outlet voltage in the future time period, and then analyze the prediction results.

[0007] In a preferred embodiment, the data in step 1 comes from three main sources: 10kV busbar and feeder data, low-voltage distribution area basic measured data, and 10kV feeder and tap information data. The 10kV busbar and feeder data comes from the power energy system and D5000 of the control center. The low-voltage distribution area basic measured data comes from the marketing electricity consumption information collection system. The 10kV feeder and tap information data comes from the basic archive data of the operation and maintenance PMS system, and the tap information comes from collection and organization.

[0008] In a preferred embodiment, in step 1, the data from the three main parts of the CSV format data table are dynamically acquired, and irrelevant data such as transformer area number and transformer area name are removed. Abnormal data such as negative active power and power factor and load rate of distribution transformers and lines, as well as invalid data with missing active power and load rate data, are deleted to obtain the source data. The formula for calculating the overall dataset completeness is: Data standardization was achieved through data cleaning and feature extraction. Outliers with three features were handled using the quartile method, and only when two or more features were found to be outliers were they identified and deleted.

[0009] In a preferred embodiment, step 2 uses Bayesian dynamic parameter tuning as an optimization method for model parameter tuning; and trains the model using five-fold cross-validation to find the optimal model.

[0010] In a preferred embodiment, step 3 uses running speed and modeling time as selection criteria to determine whether to use the Dash framework as the application interface building framework.

[0011] Compared with existing technologies, this invention has the following advantages: Compared with previous methods for diagnosing and managing voltage exceedances in distribution transformers, the proposed technology can provide scientific and quantifiable data and analysis. It employs the LightGBM lightweight gradient boosting tree model, resulting in relatively high accuracy. The model utilizes an iterative prediction error method to connect multiple decision tree models for joint decision-making, significantly improving accuracy. Furthermore, the algorithm has undergone numerous optimizations, supporting efficient parallel training and offering advantages such as faster training speed, lower memory consumption, better accuracy, and support for distributed processing to quickly handle massive amounts of data.

[0012] Furthermore, this proposal utilizes a Bayesian optimization hyperparameter tuning algorithm for dynamic model parameter tuning, allowing for the addition of more feature variables for training at any time, automatically generating the optimal model and demonstrating strong scalability. The model also adopts a modular development approach, resulting in robust program performance and providing a solid foundation for future in-depth development. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating a preferred embodiment of the present invention. Figure 2 This is a statistical result chart of missing values ​​in the source data according to a preferred embodiment of the present invention; Figure 3 Here are histograms showing the distribution of key data in a preferred embodiment of the present invention; Figure 4 This is a diagram of the outlier data detection code according to a preferred embodiment of the present invention; Figure 5 This is a correlation matrix diagram of a preferred embodiment of the present invention; Figure 6 This is a schematic diagram of the preliminary model construction result of a preferred embodiment of the present invention; Figure 7 This is a Bayesian optimization parameter diagram of a preferred embodiment of the present invention; Figure 8 To build the optimal model and achieve the optimal result in the preferred embodiment of the present invention, R 2 A diagram showing a value of 0.82; Figure 9 A comparison chart of prediction results after model evaluation in a preferred embodiment of the present invention; Figure 10 This is a comparison chart of the importance scores of feature values ​​in the regression model of a preferred embodiment of the present invention; Figure 11 This is a schematic diagram of the changes in various feature variables in a preferred embodiment of the present invention. Figure 12 This is a schematic diagram of the visual display interface of a preferred embodiment of the present invention; Figure 13 This is a result data diagram of a preferred embodiment of the present invention. Detailed Implementation

[0014] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0015] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0016] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations according to this application; as used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise; furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0017] refer to Figure 1-13 This invention utilizes "data mining" to predict distribution transformer outlet voltage values. By leveraging massive amounts of data, including marketing distribution transformer outlet operation data, historical voltage limit exceedance data, dispatch line outlet operation data, and power supply station operation and maintenance data, a machine learning model is established. This model is trained on historical data and used to predict the distribution transformer outlet voltage for future periods. Based on the prediction results, adjustments are made to optimize the reactive power operation of the distribution transformer. The flowchart of the overall scheme is shown below. Figure 1 As shown: Step 1: Data Preprocessing The model uses data from three main sources: 10kV busbar and feeder data, measured data of low-voltage distribution area foundations, and 10kV feeder and tap information data.

[0018] The 10kV busbar and feeder data are sourced from the power energy system and D5000 of the control center. The extracted data range includes the 10kV busbar voltage, 10kV feeder active power value, and power factor at 12:00 noon every day throughout 2021 for the entire region. The basic measured data of low-voltage distribution areas are sourced from the marketing electricity consumption information collection system. The extracted data range includes the basic measured data of all low-voltage distribution areas at 12:00 noon every day throughout 2021 for the entire region. The 10kV feeder and tap position information data are sourced from the basic archive data of the operation and maintenance PMS system. The tap position information is collected and organized. The extracted data range includes the 10kV feeder length information and tap position information at 12:00 noon every day throughout 2020-2021 for the entire region.

[0019] The data is dynamically acquired from a CSV format data table, integrating data from three main parts. Irrelevant data such as transformer substation numbers and names are removed, along with abnormal data (negative active power, power factor, and load rate for transformers and lines) and invalid data (missing active power and load rate data), resulting in the source data. The statistical results of missing values ​​in the source data are shown below. Figure 2 As shown: The formula for calculating the overall dataset completeness is: Data integrity analysis shows that the source data integrity is 97.27%, which meets the requirements of this model.

[0020] Data standardization primarily involves data cleaning and feature extraction. Data cleaning mainly involves using histograms to examine data distribution characteristics and removing redundant features such as active power. Histograms of key data distributions are shown below. Figure 3 As shown: Because power factor, active power, load factor, and current have similar distribution characteristics, outlier detection using the quartile method is employed. Furthermore, only when two or more features jointly constitute an outlier are they considered outliers and removed. An outlier detection code diagram is shown below. Figure 4 As shown: Feature extraction primarily uses the feature method to draw a correlation matrix, initially determining that the active power of the distribution transformer, load factor, bus voltage, and tap position have a strong correlation. A correlation matrix diagram is shown below. Figure 5 As shown: Step 2: Model Building Current data modeling methods include machine learning, deep learning, and reinforcement learning. Among them, machine learning excels at analyzing multi-dimensional feature data and making predictions about target data, making it suitable for this model. Based on factors such as prediction performance, overfitting probability, and the amount of supporting data, the LightGBM ensemble learning algorithm was chosen as the data modeling method.

[0021] Data prediction programming based on LightGBM was used to perform regression predictions on a subset of sample data, achieving high accuracy. Preliminary model results were obtained without parameter tuning, and are shown in R... 2 The value is 0.68, such as Figure 6 As shown: Based on the predicted results, the model can be optimized to reduce the number of voltage adjustments, the number of voltage limit violations, and to save costs and manpower. Using factors such as the accuracy improvement, the existence of an optimal solution, and computation time as selection criteria, Bayesian dynamic parameter tuning was chosen as the optimization method for model tuning. Five-fold cross-validation was used for training to find the optimal model. The Bayesian optimization parameter diagram is shown below. Figure 7 As shown: Build the optimal model, achieve the optimal results, and use R. 2 The value is 0.82, a significant improvement, such as Figure 8 As shown: After model evaluation, the prediction results were accurate, and the prediction results were as follows: Figure 9 As shown: We can also derive the importance scores of the feature values ​​in this regression model, such as... Figure 10 As shown: Depend on Figure 10 It can be seen that line length, bus voltage, tap position, line active power, and line power factor have relatively high importance scores. This is similar to the results of the correlation matrix diagram mentioned earlier.

[0022] Step 3: Voltage Over-Limit Prediction and Application: Based on factors such as running speed and modeling time, the Dash framework was chosen as the application interface framework. To facilitate the addition, deletion, querying, and modification of sample data, three major application modules were integrated to realize functions such as intuitive viewing of the quantitative impact factors of voltage change, ranking the importance of the impact factors of voltage change, and calculating the optimal voltage level suggestion.

[0023] The visualization of changes in various feature variables is achieved through line graphs, such as... Figure 11 As shown: Visual display interface, such as Figure 12 As shown: For the four features with the highest model importance, voltage ranges for each level were calculated under a reasonable fluctuation range, with 215V-235.4V defined as the reasonable range. The results are as follows: Figure 13 As shown: The key technical aspects and points to be protected in this model compared to existing solutions are as follows: (1) Assistance on the causes of voltage exceedance. By importing recent operating data of the transformer area that exceeds the limit, comparative analysis is performed to calculate the influence weight of various characteristic variables on the causes of voltage exceedance in the transformer area, and the contribution of each characteristic change is calculated and presented in a Pareto-like form to provide quantitative analysis results for the causes of voltage exceedance, so as to guide the key directions of voltage management.

[0024] (2) Quantitative change trend of characteristics. Based on the historical operation data of different transformer substations, the model is fitted to the quantitative change curve of the influence of various characteristic variables on the voltage of the target transformer substation, so as to understand the change trend of characteristic variables and their sensitivity and influence on voltage values, and provide a reliable decision-making basis for subsequent targeted management of transformer substations that exceed the limit.

[0025] (3) Intelligent Diagnosis. The model can predict the voltage fluctuation range at each level based on the characteristic fluctuation range, thus realizing dynamic prediction of the voltage in the distribution area. For predicted changes in characteristic variables, such as load changes and bus voltage, the reasonable operating range of the distribution transformer can be predicted. In addition, the reasonable operating range of the characteristic variable can be deduced in reverse based on the predicted reasonable operating range of the operating range of the operating range, guiding the management work, turning passive into active, and arranging in advance the management measures such as distribution network project establishment and maintenance, distribution line optimization and reconnection, installation of 10kV line reactive power compensators, distribution transformer level adjustment, load balancing, and distribution transformer reactive power input and output, so as to take the initiative in the voltage management of the distribution area.

[0026] (4) Collaborative optimization of AVC control strategy. Taking the 10kV busbar as the topology unit as the research object, the optimal section of the 10kV busbar in the substation is calculated in reverse by combining the voltage operating range of the distribution area under the topology unit for a certain period of time. Collaborative analysis and collaborative management and control method of substation busbar-distribution area voltage quality is proposed, and AVC optimization control strategy is reasonably implemented.

Claims

1. A method for predicting voltage exceedances in distribution network transformers based on the LightGBM ensemble learning algorithm, characterized in that, Includes the following steps: Step 1: Perform data preprocessing on marketing distribution transformer outlet operation data, historical voltage over-limit data, dispatch line outlet operation data, and power supply station operation and maintenance data. Step 2: Build a machine learning model based on the LightGBM ensemble learning algorithm and train it on historical data; Step 3: Use the trained model to predict the distribution transformer outlet voltage in the future time period, and then analyze the prediction results.

2. The method for predicting voltage exceedance of distribution network transformers based on the LightGBM ensemble learning algorithm according to claim 1, characterized in that, The data in step 1 comes from three main sources: 10kV busbar and feeder data, low-voltage distribution area basic measured data, and 10kV feeder and tap information data. The 10kV busbar and feeder data comes from the power energy system and D5000 of the control center; the low-voltage distribution area basic measured data comes from the marketing electricity consumption information collection system; and the 10kV feeder and tap information data comes from the basic archive data of the operation and maintenance PMS system, while the tap information comes from collection and organization.

3. The method for predicting voltage exceedance of distribution network transformers based on the LightGBM ensemble learning algorithm according to claim 2, characterized in that, In step 1, the data from the three main parts of the CSV format data table are dynamically acquired and integrated. Data irrelevant to the transformer area number and name is removed. Abnormal data with negative active power, power factor, and load rate of the transformer and line are deleted, as well as invalid data with missing active power and load rate data, to obtain the source data. The formula for calculating the overall dataset completeness is: Data standardization is achieved through data cleaning and feature extraction. Outliers with three features are handled using the quartile method, and only when two or more features are found to be outliers are they identified and deleted.

4. The method for predicting voltage exceedance of distribution network transformers based on the LightGBM ensemble learning algorithm according to claim 1, characterized in that, In step 2, Bayesian dynamic parameter tuning is used as an optimization method for model parameter tuning; and five-fold cross-validation is used for training to find the optimal model.

5. The method for predicting voltage exceedance of distribution network transformers based on the LightGBM ensemble learning algorithm according to claim 1, characterized in that, In step 3, the selection criteria were based on running speed and modeling time, and the Dash framework was chosen as the framework for building the application interface.

Images