Flexible smart power grid fault severity determination method based on physical and data dual drive

Abstract

The invention discloses a flexible intelligent power grid fault severity judgment method based on physical and data dual drive. The method comprises the following steps: 1, continuously monitoring real-time values of voltage and current of a power grid terminal by using intelligent electric meters and PMU devices which are arranged at a feeder line and a bus; 2, after fault detection is triggered, the system synchronously collects two types of key data to construct a fusion feature database; 3, inputting the constructed fusion feature library data into the model by adopting an improved graph neural network; 4, starting targeted load flow calculation based on a primary judgment result output by machine learning; and 5, after power flow analysis and verification, integrating judgment information by the system, and giving a corresponding disposal strategy. According to the method, physical drive and data drive are fused, that is, machine learning is utilized to quickly process a large amount of data and capture complex fault features, and the accuracy and adaptability of data processing are improved by means of power flow analysis based on a physical principle verification result; the traditional method is difficult to consider both judgment speed and result reliability.

Description

Technical Field

[0001] This invention relates to the field of flexible smart grid technology, and in particular to a method for determining the severity of faults in flexible smart grids based on both physical and data-driven approaches. Background Technology

[0002] Flexible smart grids represent an intelligent and advanced form of traditional power grids. They integrate advanced sensing, communication, control, and decision support systems (such as big data and artificial intelligence), enabling the power grid to be observable, controllable, and adaptive. This allows it to flexibly, efficiently, and resiliently respond to various uncertainties, achieving optimal energy allocation and reliable supply.

[0003] With the high proportion of new energy sources being integrated and the large-scale application of power electronic equipment, flexible smart grids exhibit core characteristics of "multi-source, dynamic topology, and complex fault characteristics." Existing technologies with purely physical models have limitations such as slow response speed, poor adaptability, and weak generalization ability, while purely data-driven models have limitations such as insufficient reliability, lack of physical constraints, and poor interpretability. Neither of these methods can adapt to the "multi-source interactive characteristics" of flexible smart grids, resulting in a high risk of fault propagation.

[0004] Therefore, those skilled in the art provide a method for determining the severity of faults in flexible smart grids based on both physical and data-driven approaches to address the problems mentioned in the background section. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method for determining the severity of faults in flexible smart grids based on both physical and data-driven approaches, characterized by the following steps: 1. Utilize smart meters and PMU devices installed at feeders and busbars to continuously monitor the real-time values ​​of voltage and current at the power grid terminals. Set the voltage threshold to ±5% - 10% of the rated voltage and the current threshold to 1.2 - 1.5 times the rated current. Compare the measured voltage and current with the rated values. Once the voltage or current exceeds the threshold range, the fault detection process is automatically triggered. This step is the starting signal for the entire fault determination, timely capturing abnormal power grid operation and preparing for subsequent in-depth analysis of the nature and level of the fault. This differs from the crude handling of traditional protection that simply trips the circuit breaker, aiming to provide a starting point for refined fault management. Second, after fault detection is triggered, the system synchronously collects two types of key data to construct a fusion feature database of "transient features + power flow features". On the one hand, transient data is collected, namely the high-frequency instantaneous waveforms of voltage and current at the moment of fault occurrence and within a short period of time, from which transient features are extracted. These features can characterize the nonlinear abrupt changes in the early stage of the fault. On the other hand, steady-state data is collected, namely the power flow information of the power grid during stable operation about 1 second before the fault occurs. Third, an improved graph neural network (GNN) is adopted, which takes advantage of its natural adaptation to the "node-edge" structure of the power grid topology. The constructed fusion feature library data is input into the model. The model learns the connection relationship of the equipment in the power grid topology and the correlation between various feature data. It can quickly calculate within 5ms and output the fault severity judgment result, while giving the feature importance ranking. IV. Based on the initial judgment results output by machine learning, initiate targeted power flow calculations, strictly follow the physical laws of the power system, simulate the power flow distribution after the fault, and verify whether the initial judgment results conform to the actual operating constraints of the power grid. V. After power flow analysis and verification, the system integrates the judgment information within 15ms, outputs the final fault severity level, and provides corresponding handling strategies based on power grid operation and maintenance requirements, such as prioritizing energy storage regulation for minor faults and triggering selective tripping of circuit breakers for severe faults.

[0006] Preferably, in step two, the power flow information during stable grid operation includes active / reactive power of lines, effective values ​​of node voltages, and grid topology connections, forming steady-state power flow characteristics mainly based on line load rates and dynamic power flow characteristics mainly based on power imbalance. These multi-dimensional characteristics simultaneously serve subsequent machine learning model training and power flow analysis verification, laying a solid data foundation for accurate fault determination.

[0007] Preferably, in step three, the model uses the efficient pattern recognition capability of machine learning to achieve preliminary and rapid screening of faults, and uses the advantages of data-driven approach to make up for the shortcomings of slow calculation in traditional physical models. However, at this time the results have not yet been verified by physical laws and belong to the preliminary judgment of "data first".

[0008] Preferably, in step four, the physical laws of the power system that need to be followed include Kirchhoff's current / voltage law and the power balance principle.

[0009] Preferably, in step four, the simulation calculation includes calculating whether the line load rate exceeds the equipment capacity after the fault, whether the node voltage deviation is within the safe range, and whether the power transfer will trigger a cascading overload risk. By using physical rules as a safety net, it corrects unreasonable judgments that machine learning may make due to data noise or missing data under extreme conditions, ensuring that the results are both consistent with data patterns and in line with the objective operating logic of the power system of "power conservation and equipment capacity limitation". It is a supplementary correction step of "physical constraints" to "data-driven" processing.

[0010] Preferably, in step five, the handling strategy is directly connected to the power grid energy management system to provide a basis for decision-making for subsequent fault isolation and load transfer operations, thereby achieving precise, rapid and reasonable control of faults in the flexible smart grid and improving the intelligence and efficiency of power grid fault response.

[0011] The technical effects and advantages of this invention are as follows: Compared with the prior art, the beneficial effects of the present invention are: 1. This invention integrates both physical and data-driven approaches. It utilizes machine learning to rapidly process large amounts of data and capture complex fault characteristics, while leveraging power flow analysis to verify results based on physical principles, thus improving the accuracy and adaptability of data processing. Traditional methods struggle to balance judgment speed and result reliability. This invention achieves rapid capture and preliminary judgment of fault characteristics through machine learning, combined with power flow analysis for physical-level verification and correction, reducing the judgment time to within 15ms and achieving a false judgment rate of ≤5%. It can promptly and accurately determine the severity of faults in their early stages. Attached Figure Description

[0012] Figure 1 This is a flowchart of a method for determining the severity of faults in a flexible smart grid based on both physical and data-driven approaches, as provided in an embodiment of this application. Detailed Implementation

[0013] The present invention will now be described in further detail with reference to specific embodiments. The embodiments of the present invention are given for illustrative and descriptive purposes only, and are not intended to be exhaustive or to limit the invention to the disclosed forms. Many modifications and variations will be apparent to those skilled in the art. The embodiments were chosen and described to better illustrate the principles and practical application of the invention, and to enable those skilled in the art to understand the invention and design various embodiments with various modifications suitable for a particular purpose. Example 1

[0014] Please see Figure 1 This embodiment provides a method for determining the severity of faults in a flexible smart grid based on both physical and data-driven approaches, characterized by the following steps: Step 1: Utilize smart meters and PMU devices installed at feeders and busbars to continuously monitor the real-time values ​​of voltage and current at the power grid terminals. Set the voltage threshold to ±5% - 10% of the rated voltage and the current threshold to 1.2 - 1.5 times the rated current. Compare the measured voltage and current with the rated values. Once the voltage or current exceeds the threshold range, the fault detection process is automatically triggered. This step is the starting signal for the entire fault determination process, promptly capturing abnormal power grid operation and preparing for subsequent in-depth analysis of the fault nature and level. This differs from the crude handling of traditional protection that simply trips the circuit breaker, aiming to provide a starting point for refined fault management.

[0015] Step Two: After triggering fault detection, the system simultaneously collects two types of key data to construct a fusion feature database of "transient features + power flow features". On one hand, transient data is collected, namely the high-frequency instantaneous waveforms of voltage and current at the moment of fault occurrence and within a short period, from which transient features such as current change rate and harmonic energy torque are extracted. These features can characterize the nonlinear abrupt changes in the early stages of a fault. On the other hand, steady-state data is collected, namely the power flow information of the grid during stable operation approximately 1 second before the fault occurs, including line active / reactive power, effective values ​​of node voltages, and grid topology connections, forming steady-state power flow features mainly based on line load rates, and dynamic power flow features mainly based on power imbalance. These multi-dimensional features simultaneously serve subsequent machine learning model training and power flow analysis verification, laying a solid data foundation for accurate fault determination.

[0016] Step 3: An improved Graph Neural Network (GNN) is employed, leveraging its natural adaptation to the "node-edge" structure of the power grid topology. The constructed fusion feature library data is input into the model. The model learns the connection relationships between devices in the power grid topology and the correlations between various feature data, rapidly calculating within 5ms and outputting the fault severity judgment result. It also provides a ranking of feature importance. Utilizing the efficient pattern recognition capabilities of machine learning, preliminary and rapid fault screening is achieved. This data-driven approach compensates for the slow computation of traditional physical models. However, at this stage, the results have not yet been verified by physical laws; it is a preliminary judgment based on "data first."

[0017] Step 4: Based on the initial judgment results from machine learning, initiate targeted power flow calculations. Strictly adhering to the physical laws of the power system, such as Kirchhoff's current / voltage law and power balance principles, simulate the power flow distribution after the fault to verify whether the initial judgment results conform to the actual operating constraints of the power grid. For example, calculate whether the line load rate exceeds equipment capacity after the fault, whether the node voltage deviation is within a safe range, and whether power transfer will trigger cascading overload risks. Through physical rules as a safety net, correct unreasonable judgments that machine learning may make due to data noise or missing data under extreme conditions, ensuring that the results both conform to data patterns and the objective operating logic of the power system, such as "power conservation and equipment capacity limitations." This is a supplementary correction step between "physical constraints" and "data-driven" approaches.

[0018] Step 5: After power flow analysis and verification, the system integrates the judgment information within 15ms, outputs the final fault severity level, and, combined with grid operation and maintenance requirements (such as prioritizing energy storage regulation for minor faults and selectively tripping circuit breakers for severe faults), provides corresponding handling strategies. These results are directly connected to the grid energy management system, providing decision-making basis for subsequent fault isolation, load transfer, and other operations, enabling precise, rapid, and reasonable control of faults in the flexible smart grid, and improving the intelligence and efficiency of grid fault response.

[0019] The aforementioned system includes the collaborative workflow of the machine learning module and the power flow analysis module; protect the system architecture and devices: protect the system architecture that implements the determination method, including the composition and connection relationships of devices such as the data acquisition unit, feature extraction and fusion module, machine learning model, power flow analysis module, result output and decision-making unit; protect the application scenarios and effects of the technology.

[0020] The specific implementation method based on the flowchart is as follows: The system consists of a 6MW distributed photovoltaic array (40% penetration rate), a 3MWh lithium battery energy storage system, and one 5MVA grid-connected converter. The rated voltage of the distribution line is 380V (three-phase low voltage), and the rated current is 1200A. The load mainly consists of production equipment in the industrial park (total load 5MW). One day, a sudden strong gust of wind shaded the photovoltaic array, causing the photovoltaic output to drop sharply from 4.5MW to 2.8MW (triggering grid disconnection risk), resulting in a voltage dip in the distribution area. The following analysis is performed according to the established procedures.

[0021] Step 1: The intelligent measurement terminal (PMU, sampling frequency 10kHz) installed at the transformer substation's incoming line monitors the three-phase voltage RMS value in real time, outputting one set of data every 0.1ms. The measurement terminal recorded that the RMS value of phase A voltage suddenly dropped from 380V (rated) to 304V (measured), and maintained this value for three consecutive sampling cycles (0.3ms). Calculate the voltage sag rate:

[0022] Given that the rated voltage U is 380V and the measured voltage U during a fault is 304V, substituting these values, we get: voltage sag rate = 20% > 5%, triggering the fault detection process.

[0023] Transient data acquisition was performed (0-100ms after the fault), collecting the converter output current, photovoltaic inverter low-voltage ride-through current, and energy storage charging / discharging current. Current waveforms within 0-1ms after the fault were collected (10 sampling points, 0.1ms interval), with the data as follows: [1050A, 1080A, 1120A, 1170A, 1230A, 1290A, 1350A, 1410A, 1470A, 1530A]. Sampling points within 0.2ms were also collected: 1050A at t=0.1ms, 1290A at t=0.3ms, a difference of 240A, with a time interval of 0.2ms.

[0024] Calculate the rate of change of current:

[0025] (Approximately 50V / ms under normal operating conditions); Steady-state data acquisition (1 second before fault): Actual line power: 4.8MW, Rated line power: 6MW; Load rate:

[0026] (Normal operating range 50%-90%) Power imbalance ΔP = 2.8MW (output after photovoltaic grid disconnection) - 5MW (load) = -2.2MW (supply and demand difference).

[0027] Step Two: Constructing a "transient + current flow" integrated feature library

[0028] Step 3: An improved graph neural network (GNN) was used, with training data containing 1000+ historical fault scenarios (such as photovoltaic grid disconnection, line short circuits, etc.). Each scenario was labeled "minor / serious / extremely serious". A power flow constraint regularization term was added to the loss function (e.g., strengthening the weight of "serious fault" when the load factor > 120%). After the model inputs features, the output result is given within 5ms: Fault Severity: Serious Fault (confidence 88%). Step Four: Based on the initial assessment, the power flow calculation tool was activated to simulate the grid state after the fault. After detecting the power imbalance, the discharge was initiated within 0.5ms, with a discharge power of 0.9MW (75% of the rated discharge power of 1.2MW). Actual line power after fault: P after fault = 5MW (load) - 2.8MW (photovoltaic) - 0.9MW (energy storage discharge) = 1.3MW; Load rate:

[0029] No risk of overload; Converter dynamic reactive power compensation: After detecting a voltage sag, the reactive power generated increases from 1Mvar to 2.5Mvar within 2ms; Voltage recovery data: 50ms after the fault, the voltage sampled by the measurement terminal rose to 362V; Voltage deviation rate:

[0030] The voltage can be restored to the acceptable range.

[0031] Step 5: Final output (completed within 15ms): Fault severity: Moderate to severe fault; Judgment criteria: The initial judgment was triggered by a sudden current change (120A / ms) and a power imbalance (-2.2MW), but the power flow verification showed that the load rate was 21.7% and the voltage deviation was 4.7%, both of which were controllable.

[0032] Linkage effect: The result was connected to the EMS system. No circuit breaker tripping was triggered. The voltage was restored to 375V after 30 seconds through energy storage and converter regulation. The load rate was stabilized at 25%. There was no power outage loss, which verified the accuracy of the "dual drive" judgment.

[0033] This method integrates both physical and data-driven approaches. It utilizes machine learning to rapidly process large amounts of data and capture complex fault characteristics, while power flow analysis verifies results based on physical principles, improving the accuracy and adaptability of data processing. Traditional methods struggle to balance speed and reliability. This invention achieves rapid capture and initial assessment of fault characteristics through machine learning, combined with power flow analysis for physical-level verification and correction. This reduces the judgment time to within 15ms with a false positive rate of ≤5%, enabling timely and accurate assessment of fault severity in the early stages of a fault. Judgment speed is improved by over 50%. Through a dual closed-loop mechanism of "5ms initial judgment via machine learning + 10ms verification via power flow analysis," the total time is ≤15ms, meeting the requirements of flexible smart grids for "rapid fault propagation and millisecond-level response," thus gaining crucial time for subsequent fault isolation. The false positive rate is reduced to below 5%. Machine learning captures the correlation between "transient characteristics + power flow state," reducing data noise interference, while power flow analysis uses physical constraints to correct model biases, ultimately achieving an overall false positive rate of ≤5% for fault severity, significantly superior to existing technologies.

[0034] Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art and related fields based on the embodiments of the present invention without inventive effort should fall within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described and explained in the present invention, unless otherwise specified or limited, shall be implemented according to conventional means in the art.

Claims

1. A method for determining the severity of faults in a flexible smart grid based on both physical and data-driven approaches, characterized in that, Includes the following steps:

1. Utilize smart meters and PMU devices installed at feeders and busbars to continuously monitor the real-time values ​​of voltage and current at the power grid terminals. Compare the measured voltage and current with the rated values. Once the voltage or current is detected to exceed the threshold range, the fault detection process is automatically triggered. Second, after triggering fault detection, the system synchronously collects two types of key data to construct a fusion feature database of "transient features + power flow features". On the one hand, transient data is collected, namely the high-frequency instantaneous waveforms of voltage and current at the moment of fault occurrence and within a short period of time, from which transient features are extracted. These features can characterize the nonlinear sudden changes in the early stage of the fault. On the other hand, steady-state data is collected, namely the power flow information of the power grid when it is running stably for about 1 second before the fault occurs. Third, an improved graph neural network is adopted, which takes advantage of its natural adaptation to the "node-edge" structure of the power grid topology. The constructed fusion feature library data is input into the model. The model learns the connection relationship of the equipment in the power grid topology and the correlation between various feature data. It can quickly calculate within 5ms and output the fault severity judgment result, while giving the feature importance ranking. IV. Based on the initial judgment results output by machine learning, initiate targeted power flow calculations, strictly follow the physical laws of the power system, simulate the power flow distribution after the fault, and verify whether the initial judgment results conform to the actual operating constraints of the power grid. V. After power flow analysis and verification, the system integrates the judgment information within 15ms, outputs the final fault severity level, and provides corresponding handling strategies based on power grid operation and maintenance requirements, such as prioritizing energy storage regulation for minor faults and triggering selective tripping of circuit breakers for severe faults.

2. The method for determining the severity of faults in a flexible smart grid based on both physical and data-driven approaches as described in claim 1, characterized in that, In step two, the power flow information during stable grid operation includes active / reactive power of lines, effective values ​​of node voltages, and grid topology connections, forming steady-state power flow characteristics mainly based on line load rates and dynamic power flow characteristics mainly based on power imbalance. These multi-dimensional characteristics serve the subsequent training of machine learning models and verification of power flow analysis, laying a solid data foundation for accurate fault determination.

3. The method for determining the severity of faults in a flexible smart grid based on both physical and data-driven approaches as described in claim 1, characterized in that... In step three, the model uses the efficient pattern recognition capabilities of machine learning to achieve preliminary and rapid screening of faults, and leverages the advantages of data-driven approaches to compensate for the slow computation of traditional physical models.

4. The method for determining the severity of faults in a flexible smart grid based on both physical and data-driven approaches as described in claim 1, characterized in that, Step four requires adherence to the physical laws of power systems, including Kirchhoff's current / voltage law and the power balance principle.

5. The method for determining the severity of faults in a flexible smart grid based on both physical and data-driven approaches as described in claim 1, characterized in that... In step four, the simulation calculation includes determining whether the line load rate exceeds the equipment capacity after the fault, whether the node voltage deviation is within the safe range, and whether the power transfer will trigger a cascading overload risk.

6. The method for determining the severity of faults in a flexible smart grid based on both physical and data-driven approaches as described in claim 1, characterized in that, In step five, the handling strategy is directly integrated into the power grid energy management system to provide decision-making basis for subsequent operations such as fault isolation and load transfer, so as to achieve accurate, rapid and reasonable control of faults in the flexible smart grid and improve the intelligence and efficiency of power grid fault response.

Images