Small current grounding line selection method and system based on CatBoost algorithm

Abstract

The invention relates to the technical field of small-current grounding line selection, and discloses a small-current grounding line selection method and system based on a CatBoost algorithm, and the method comprises the steps: firstly, collecting the zero-sequence voltage and zero-sequence current of each line through a data collection splitter, and transmitting the collected data to a line selection host through the data collection splitter; the line selection host trained by the CatBoost algorithm selects a faulty line according to the collected data, the data collection splitter preprocesses the collected data, converts a collected analog signal into a digital signal, and transmits the digital signal to the line selection host through an optical fiber. Data of various identification methods such as amplitude comparison and phase comparison, first half wave, power direction and higher harmonic are fused together, the fused data drive a fault line selection model, fault information in multiple aspects is fully utilized, various line selection criteria are integrated and complemented and fused with one another, the fault judgment capability is improved, and the method has wider adaptability.

Description

A method and system for selecting low-current grounding lines based on the CatBoost algorithm Technical Field

[0001] This invention relates to the field of low-current grounding fault location technology, and in particular to a low-current grounding fault location method and system based on the CatBoost algorithm. Background Technology

[0002] The location of single-phase grounding faults in neutral-point non-effectively grounded systems has long been a major unresolved problem both domestically and internationally. Various fault location methods have been proposed, but none are fully adaptable to diverse power grid structures and complex fault conditions. Hubei's power grid has over 800 35kV substations, many of which are equipped with low-current grounding fault location devices. However, due to low accuracy and unsatisfactory performance, users still rely on manual circuit breaking and testing for single-phase grounding fault location. This method is not only time-consuming and labor-intensive, but also frequently results in accidents causing injury or death to people and livestock due to delayed fault handling, negatively impacting the power company's public image. Furthermore, this approach is incompatible with the requirements of smart grids.

[0003] The main reason is that the line selection device has a single line selection method and poor adaptability, and cannot adapt to various power grid structures and complex fault conditions at the same time. Therefore, a line selection method that integrates multiple line selection principles is needed for line selection. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for selecting low-current grounding faults based on the CatBoost algorithm. It integrates a data-driven fault selection model, makes full use of multiple fault information, and combines multiple fault selection criteria to complement and integrate them, thereby improving fault identification capabilities and having wider applicability.

[0005] The above-mentioned technical objective of the present invention is achieved through the following technical solution:

[0006] A method for selecting faulty grounding lines based on the CatBoost algorithm first uses a data acquisition branch to collect the zero-sequence voltage and zero-sequence current of each line. The data acquisition branch then sends the collected data to the fault selection host. After training, the fault selection host selects the faulty line based on the collected data. The data acquisition branch preprocesses the collected data, converting the collected analog signals into digital signals, and then transmits the digital signals to the fault selection host via optical fiber.

[0007] By adopting the above technical solution, this method uses a host computer, a data acquisition processor, and a high-precision zero-sequence CT to form a complete system, ensuring the timeliness and accuracy of line selection. The branching device can be digitized locally, reducing the attenuation of the zero-sequence signal.

[0008] A further provision of the present invention is that the host computer is trained using the Catboost algorithm, and the specific steps include obtaining the original data table, preprocessing the data, and offline training of the Catboost algorithm.

[0009] By adopting the above technical solutions, CatBoost can handle more data types without requiring preprocessing of non-numerical features. Using symmetric decision trees as its base learners not only improves the efficiency of processing categorical features but also solves problems such as gradient bias and prediction offset, while more effectively preventing overfitting during the learning process.

[0010] The host training method of this method uses the CatBoost artificial intelligence classification algorithm to fuse data from multiple route selection methods, which overcomes the problem that a single method has certain effects but poor adaptability, and solves the problem of route selection very well.

[0011] The invention further includes the following: the original data table includes actual field operation data and power simulation data; the data preprocessing includes synchronizing, filtering, denoising, truncating and normalizing zero-sequence voltage and zero-sequence current; and it also includes feature extraction of the data.

[0012] By adopting the above technical solution, the original fault data comes from two sources: one is the waveform data recorded during actual power grid operation, and the other is the operational data collected based on the power grid simulation model. The actual power grid operational data is limited to a specific operational scenario. Simulation data is used to supplement the dataset with scenario-specific data. Furthermore, the original fault data is decomposed and extracted using a variational mode decomposition algorithm to obtain feature-enhanced datasets at different frequency scales.

[0013] The present invention is further configured to: comprehensively employ the principles and methods of fundamental wave group amplitude and phase ratio algorithm, first half wave algorithm, fifth harmonic algorithm and power direction algorithm for feature extraction of data.

[0014] A further provision of the present invention is that the offline training step of the Catboost algorithm includes:

[0015] Step 1. Divide the preprocessed data samples into two parts: a training set and a test set.

[0016] Step 2. Build a CatBoost model, input the training set data to train the model, and optimize the model parameters to obtain the CatBoost model with the optimal parameter combination;

[0017] Step 3. Input the test set into the trained CatBoost model for validation, and evaluate the performance of the line selection model according to the evaluation metrics.

[0018] A further provision of this invention is as follows: during the training process of the selected host using the Catboost algorithm, the first sub-classifier is initially trained using the complete dataset. Based on the residual between the prediction result of the first sub-classifier and the true label, a new target variable is constructed. The new target variable is used to train the next sub-classifier to further reduce the residual error of the model. The above steps are repeated continuously until the predetermined number of iterations is reached.

[0019] By adopting the above technical solutions, this method comprehensively uses multiple criteria such as "fundamental wave amplitude and phase ratio algorithm", "first half wave algorithm", "fifth harmonic algorithm", "power direction algorithm" and "transient analysis method", and uses a combination of integrated learning model and operational experience to perform data fusion, thereby achieving comprehensive route selection;

[0020] This method applies a fusion data-driven fault route selection model, making full use of multiple fault information and integrating various route selection criteria. These criteria complement and integrate with each other, which can significantly improve the fault discrimination capability and make this method more widely applicable.

[0021] A further embodiment of the present invention is a system for implementing the above method, comprising a zero-sequence CT, a data acquisition branch, and a line selection host. Several zero-sequence CTs are directly connected to a data branch. The zero-sequence CTs directly transmit analog signals to the data acquisition branch. The data acquisition branch is connected to the line selection host via a communication line or optical fiber. Each data acquisition branch has a built-in independent CPU, which can synchronously analyze, pre-screen, and preprocess the acquired data. The data acquisition branch has a built-in A / D converter and a 5th harmonic frequency selection circuit.

[0022] By adopting the above technical solution, compared with the line selection devices of other domestic manufacturers, this system uses a distributed network structure consisting of a line selection host, data acquisition branch units, and high-precision zero-sequence current transformers. The addition of the data acquisition branch unit as an intermediate link has the main advantage of each unit having an independent industrial CPU. This CPU can detect zero-sequence signals in real time and refresh and store them in the data buffer. Upon a ground fault, it receives a broadcast command from the host, enabling each CPU to simultaneously perform analysis, preliminary screening, and preprocessing, thereby reducing the load on the host and improving the timeliness of line selection. Furthermore, the use of digital signal transmission helps reduce the number of signal cables and minimize signal transmission errors.

[0023] Each branch unit can monitor three outgoing lines. All branch units are connected from the high-voltage room to the main unit in the control room via a single 485 communication line or fiber optic cable. This eliminates the need to run the secondary cables of each zero-sequence current transformer to the main unit in the control room, saving costs, reducing installation workload, and minimizing problems such as frequent failures, high maintenance, and aesthetic impact caused by excessive wiring to the main unit. It also facilitates future capacity expansion.

[0024] The line selection host can monitor the voltage of 4 PT segments and display, upload, and control the tripping of grounding lines on each bus segment. It can adapt to the segmented or parallel operation of the bus segment and has a maximum of 90 line selection circuits, which can meet the needs of large substations.

[0025] The beneficial effects of this invention are:

[0026] 1. This invention develops a line selection host based on the Catboost algorithm of artificial intelligence, which integrates data from multiple identification methods such as amplitude and phase comparison, first half-wave, power direction and higher harmonics. It overcomes the problem that although a single line selection method has a certain effect, it has poor adaptability. It has been widely used in a large number of substations with "neutral point ungrounded system" and has achieved a good result with an overall line selection accuracy of over 98%.

[0027] 2. This invention employs a host computer, a data acquisition processor, and a high-precision zero-sequence CT to form a complete system, ensuring the timeliness and accuracy of route selection. The branching device can be digitized locally, reducing the attenuation of the zero-sequence signal. The route selection host computer is trained using the CatBoost algorithm. Multiple criteria are comprehensively employed, including the "fundamental wave amplitude-phase ratio algorithm," "first half-wave algorithm," "fifth harmonic algorithm," "power direction algorithm," and "transient analysis method." Data fusion is achieved using an integrated learning model combined with operational experience, thereby realizing comprehensive route selection. This data-driven fault route selection model fully utilizes multifaceted fault information and integrates various route selection criteria. These criteria complement and fuse with each other, significantly improving fault discrimination capabilities and giving the method broader applicability.

[0028] 3. Addressing the issue of arc suppression coils altering the amplitude and phase characteristics of the power frequency state, and the low content and small amplitude of the 5th harmonic, making it difficult to distinguish from noise, the data acquisition splitter of this invention divides each zero-sequence signal into two paths after acquisition and transformation. One path directly enters the A / D converter, while the other path is amplified to the same order of magnitude as the fundamental frequency after being selected by the 5th harmonic frequency selection circuit, and then sent to the A / D converter. This ensures that the signal-to-noise ratio of the fundamental frequency and the 5th harmonic is comparable. The amplified 5th harmonic data is then spliced ​​together with the zero-sequence current data and the zero-sequence voltage transient data to form a characteristic sample set under the arc suppression coil grounding system, improving the effectiveness of line selection for the arc suppression coil grounding system.

[0029] Furthermore, the line selection host based on the Catboost algorithm of artificial intelligence better solves the problem of data fusion between the fifth harmonic algorithm of the arc suppression coil system and other algorithms, and greatly improves the accuracy of line selection.

[0030] 5. This system can meet the protection line selection requirements when multiple lines experience simultaneous grounding during operation of a 4-section busbar combination or segmented operation; it can also ensure accuracy when the transition resistance during grounding is high or when system resonance occurs. This meets the reliability requirements of the power system. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 is a schematic diagram of the CatBoost decision model of the present invention.

[0033] Figure 2 is a schematic diagram of the training process of the line selection host of the present invention.

[0034] Figure 3 is a flowchart of the online usage process of this system.

[0035] Figure 4 is a block diagram of the system structure. Detailed Implementation

[0036] The technical solution of the present invention will now be clearly and completely described with reference to specific embodiments. Obviously, the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0037] A method for selecting faulty grounding lines based on the CatBoost algorithm first uses a data acquisition branch to collect the zero-sequence voltage and zero-sequence current of each line. The data acquisition branch then sends the collected data to the fault selection host. After training, the fault selection host selects the faulty line based on the collected data. The data acquisition branch preprocesses the collected data, converting the collected analog signals into digital signals, and then transmits the digital signals to the fault selection host via optical fiber.

[0038] The training of the host computer utilizes the Catboost algorithm. The specific steps include obtaining the original data table, data preprocessing, and offline training of the Catboost algorithm. The original data table includes actual field operation data and power simulation data. Data preprocessing involves synchronizing, filtering, denoising, truncating, and normalizing zero-sequence voltage and zero-sequence current, as well as extracting features from the data. The feature extraction of the data comprehensively adopts the principles and methods of the fundamental wave group amplitude and phase comparison algorithm, the first half-wave algorithm, the fifth harmonic algorithm, and the power direction algorithm.

[0039] CatBoost can handle more data types and does not require preprocessing of non-numerical features. Using symmetric decision trees as its base learners not only improves the efficiency of handling categorical features but also addresses issues such as gradient bias and prediction offset, while more effectively preventing overfitting during the learning process.

[0040] The host training method of this method uses the CatBoost artificial intelligence classification algorithm to fuse data from multiple route selection methods, which overcomes the problem that a single method has certain effects but poor adaptability, and solves the problem of route selection very well.

[0041] The original fault data comes from two sources: one is the waveform data recorded during actual power grid operation, and the other is the operational data collected based on the power grid simulation model. The actual power grid operation data is limited to a specific operating scenario. The simulation data is used to supplement the dataset with scenario information, and the original fault data is decomposed and extracted based on the variational mode decomposition algorithm to obtain feature-enhanced datasets at different frequency scales.

[0042] The offline training steps for the Catboost algorithm include:

[0043] Step 1. Divide the preprocessed data samples into two parts: a training set and a test set.

[0044] Step 2. Build a CatBoost model, input the training set data to train the model, and optimize the model parameters to obtain the CatBoost model with the optimal parameter combination;

[0045] Step 3. Input the test set into the trained CatBoost model for validation, and evaluate the performance of the line selection model according to the evaluation metrics.

[0046] In the training process of the Catboost algorithm for the selected host, the first sub-classifier is initially trained using the complete dataset. Based on the residual between the prediction results of the first sub-classifier and the true label, a new target variable is constructed. This new target variable is then used to train the next sub-classifier to further reduce the residual error of the model. This process is repeated until the predetermined number of iterations is reached.

[0047] A system for implementing the above method includes a zero-sequence CT, a data acquisition splitter, and a line selection host. Several zero-sequence CTs are directly connected to one data splitter. The zero-sequence CTs directly transmit analog signals to the data acquisition splitter. The data acquisition splitter is connected to the line selection host via a communication line or optical fiber. Each data acquisition splitter has a built-in independent CPU, capable of synchronously analyzing, initially screening, and preprocessing the acquired data. The data acquisition splitter incorporates an A / D converter and a 5th harmonic frequency selection circuit.

[0048] Compared to line selection devices from other domestic manufacturers, this system employs a distributed network structure consisting of a line selection host, data acquisition branch units, and high-precision zero-sequence current transformers. The addition of the data acquisition branch units as an intermediate link offers a significant advantage: each branch unit has an independent industrial CPU capable of real-time detection of zero-sequence signals and updating and storing them in a data buffer. Upon a ground fault, it receives a broadcast command from the host, enabling each CPU to synchronously perform analysis, preliminary screening, and preprocessing, thereby reducing the load on the host and improving line selection timeliness. Furthermore, the use of digital signal transmission helps reduce the number of signal cables and minimize signal transmission errors.

[0049] Each branch unit can monitor three outgoing lines. All branch units are connected from the high-voltage room to the main unit in the control room via a single 485 communication line or fiber optic cable. This eliminates the need to run the secondary cables of each zero-sequence current transformer to the main unit in the control room, saving costs, reducing installation workload, and minimizing problems such as frequent failures, high maintenance, and aesthetic impact caused by excessive wiring to the main unit. It also facilitates future capacity expansion.

[0050] The line selection host can monitor the voltage of 4 PT segments and display, upload, and control the tripping of grounding lines on each bus segment. It can adapt to the segmented or parallel operation of the bus segment and has a maximum of 90 line selection circuits, which can meet the needs of large substations.

[0051] This invention develops a line selection host based on the Catboost algorithm of artificial intelligence. It integrates data from multiple identification methods, such as amplitude and phase comparison, first half-wave, power direction, and higher harmonics, overcoming the problem that single line selection methods, while effective, have poor adaptability. It has been widely used in a large number of substations with "neutral point ungrounded systems," achieving an overall line selection accuracy of over 98%. By combining various data collected during grounding status with operational experience, and using fuzzy theory and computer technology, a unique algorithm program was independently developed using a combination of logical relationships and fuzzy inference. In terms of hardware, it uses large-scale integrated circuits and microcontrollers imported from the United States. Structurally, it features a distributed network structure with multiple processors working simultaneously, enabling real-time tracking, synchronous acquisition and processing of multiple data, and timely release of accurate results.

[0052] The specific working principle is as follows: adopting the working mode of "distributed acquisition and centralized decision-making", the distributed acquisition system monitors and uploads the operating data signals of each outgoing line, including zero-sequence voltage, zero-sequence current, etc. The central decision processor decomposes the signals, enhances the data features in the form of multi-modal and multi-frequency, and recombines them to form a decision dataset, which is input into the trained data-driven fault line selection decision module and outputs the line selection result.

[0053] This system applies a data-driven fault route selection model, making full use of fault information from multiple aspects and integrating various route selection criteria. These criteria complement and integrate with each other, which can significantly improve fault identification capabilities and make the system more adaptable.

[0054] The connection of the arc suppression coil alters the amplitude and phase angle of the steady-state zero-sequence current, rendering it devoid of fault characteristics, while the transient quantity remains unaffected by the neutral grounding method. This invention derives the composition of the fault transient zero-sequence current, concluding that the transient zero-sequence current exhibits component differences under different fault phase angles, transition resistances, and fault types. Furthermore, variational mode decomposition is introduced into the zero-sequence current feature extraction process, and combined with the conclusion of component differences in the fault zero-sequence current, the zero-sequence current signal is decomposed and its features extracted.

[0055] Since the arc suppression coil is designed for 50Hz power frequency signals, it mainly alters the amplitude and phase characteristics of the power frequency state. While it still exhibits characteristics of amplitude and phase changes in the zero-sequence current of the faulty line for the 5th harmonic, its content is low and its amplitude is small, making it difficult to distinguish from noise. To address this issue, this invention utilizes a data acquisition splitter that can better handle 5th harmonic signals. The processor works by acquiring and transforming each zero-sequence signal, then splitting it into two paths. One path goes directly to the A / D converter, while the other path is amplified to the same order of magnitude as the fundamental frequency after being selected by a 5th harmonic frequency selection circuit before being sent to the A / D converter. This makes the signal-to-noise ratio of the fundamental frequency and the 5th harmonic comparable. The amplified 5th harmonic data is then spliced ​​with the zero-sequence current data and zero-sequence voltage transient data to form a characteristic sample set for the arc suppression coil grounding system, improving the effectiveness of line selection for arc suppression coil grounding systems.

[0056] In addition, the line selection host based on the Catboost algorithm of artificial intelligence has better solved the problem of data fusion between the fifth harmonic algorithm of the arc suppression coil system and other algorithms, and has greatly improved the accuracy of line selection.

[0057] This invention employs a host computer, a data acquisition processor, and a high-precision zero-sequence CT to form a complete system, ensuring the timeliness and accuracy of route selection. The branching device can be digitized locally, reducing the attenuation of the zero-sequence signal. The route selection host computer is trained using the CatBoost algorithm. Multiple criteria are comprehensively adopted, including the "fundamental wave amplitude-phase ratio algorithm," "first half-wave algorithm," "fifth harmonic algorithm," "power direction algorithm," and "transient analysis method." Data fusion is performed using an integrated learning model combined with operational experience, thereby achieving comprehensive route selection. The fused data-driven fault route selection model makes full use of multifaceted fault information and integrates multiple route selection criteria. These criteria complement and fuse with each other, significantly improving fault discrimination capabilities and making this method more widely applicable.

[0058] This invention has been described through several embodiments. Those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of this invention, modifications can be made to these features and embodiments to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, this invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are protected by this invention.

Claims

1. A low-current grounding line selection method based on the CatBoost algorithm, characterized in that: First, the zero-sequence voltage and zero-sequence current of each line are collected using a data acquisition branch. The data acquisition branch then sends the collected data to the line selection host. After training, the line selection host selects the faulty line based on the collected data.

2. The method for selecting a low-current grounding line based on the CatBoost algorithm according to claim 1, characterized in that: The data acquisition splitter preprocesses the acquired data, converting the acquired analog signals into digital signals, and then transmits the digital signals to the line selection host via optical fiber.

3. The low-current grounding line selection method based on the CatBoost algorithm according to claim 2, characterized in that: The host computer is trained using the Catboost algorithm, and the specific steps include obtaining the original data table, data preprocessing, and offline training of the Catboost algorithm.

4. The low-current grounding line selection method based on the CatBoost algorithm according to claim 3, characterized in that: The original data table includes actual on-site operating data and power simulation data.

5. The low-current grounding line selection method based on the CatBoost algorithm according to claim 4, characterized in that: The data preprocessing includes synchronizing, filtering, denoising, truncating, and normalizing zero-sequence voltage and zero-sequence current, as well as extracting features from the data.

6. The low-current grounding line selection method based on the CatBoost algorithm according to claim 5, characterized in that: The feature extraction of the data adopts the principles and methods of the fundamental wave group amplitude and phase ratio algorithm, the first half-wave algorithm, the fifth harmonic algorithm, and the power direction algorithm.

7. A low-current grounding line selection method based on the CatBoost algorithm according to claim 6, characterized in that: The offline training steps of the CatBoost algorithm include: Step 1. Dividing the preprocessed data samples into two parts: a training set and a test set; Step 2. Building a CatBoost model, inputting the training set data for model training, and optimizing the model parameters to obtain the CatBoost model with the optimal parameter combination; Step 3. Inputting the test set into the trained CatBoost model for verification, and evaluating the performance of the line selection model according to the evaluation metrics.

8. The low-current grounding line selection method based on the CatBoost algorithm according to claim 7, characterized in that: During the training process of the selected host using the Catboost algorithm, the first sub-classifier is initially trained using the complete dataset. Based on the residual between the prediction result of the first sub-classifier and the true label, a new target variable is constructed. This new target variable is then used to train the next sub-classifier to further reduce the residual error of the model. The above steps are repeated until the predetermined number of iterations is reached.

9. A system for implementing the method of any one of claims 1-8, characterized in that: The system includes a zero-sequence CT, a data acquisition splitter, and a line selection host. Several zero-sequence CTs are directly connected to a data splitter. The zero-sequence CTs directly transmit analog signals to the data acquisition splitter. The data acquisition splitter is connected to the line selection host via a communication line or optical fiber. Each data acquisition splitter has a built-in independent CPU, which can synchronously analyze, pre-screen, and preprocess the acquired data.

10. A low-current grounding line selection system based on the CatBoost algorithm according to claim 9, characterized in that: The data acquisition branch unit has a built-in A / D converter and a 5th harmonic frequency selection circuit.

Images