Transformer surge current and internal fault distinguishing method

Abstract

The invention discloses a transformer surge current and internal fault distinguishing method. The method comprises the following steps of obtaining three-phase current data of a transformer; performing difference calculation on the current data; when the differential value is abnormal, features are extracted and decomposed into multiple energy levels, and surge current or internal faults are distinguished through an LSTM network; in the invention, the LSTM network receives the D7 energy level as input, and processes the time sequence features to identify surge current and internal faults. The LSTM network optimizes the performance of the model by learning the long-term and short-term dependency relationship in the data and adjusting the weight and deviation thereof, and these adjustments allow the LSTM network to effectively capture the dynamic change of the differential current, thereby improving the accuracy and reliability of fault detection, and improving the reliability of fault detection. The trained LSTM network can identify the surge current and the internal fault of the transformer with high precision and speed, and when the network identifies that the mode of the differential current accords with the characteristics of the internal fault, the network will automatically send out a tripping signal.

Description

Technical Field

[0001] The invention relates to the field of relay protection, and in particular to a method for distinguishing between a transformer surge current and an internal fault. Background Art

[0002] In order to ensure the safe operation of the power system and transformers, relay protection devices are generally installed in various parts of the power system. Once an internal fault of the transformer is detected, the relay protection device will quickly cut off the transformer from the system to avoid serious consequences.

[0003] Both inrush current and internal fault can cause large transient current and affect differential protection, but they differ in causes, current waveform and duration. Inrush current is caused by magnetic saturation when the transformer is closed under no-load condition, and has significant non-periodic components and a shorter duration. Internal fault is usually caused by transformer winding short circuit or insulation fault, and the current is continuous and highly destructive. Relay protection device detects the amplitude and harmonic components of differential current, distinguishes between the two and decides whether to trigger tripping to ensure safe operation of the system.

[0004] Internal faults are the root cause of nearly 70% to 80% of damage in transformers, leading to serious problems in a short period of time. In this regard, differential protection can correctly detect internal faults and issue a trip command in a short period of time to protect the safety of the transformer. However, when the transformer is energized, it will receive a large current (inrush current), which will magnetize the transformer winding. The main existing problem of differential protection is the false operation caused by the inrush current. The amplitude of the inrush current mainly depends on the transformer capacity, the core characteristics, the residual magnetic flux in the core, and the voltage level when the power is turned on. However, the inrush current is usually high enough to cause the false operation of the differential protection. Therefore, various schemes are needed in the field of differential protection to distinguish between inrush current and internal faults. Summary of the invention

[0005] The object of the present invention is to provide a method for distinguishing surge current from internal faults of a transformer, so as to solve the problem in the prior art that various solutions are needed to distinguish surge current from internal faults.

[0006] In order to solve the above technical problems, the present invention is achieved through the following technical solutions:

[0007] The present invention is a method for distinguishing transformer surge current from internal fault, comprising the following steps:

[0008] S1. Use a current transformer to measure the three-phase current of the transformer and obtain the current data of each phase;

[0009] S2, performing differential operation on the three phase currents to obtain the calculated transformer differential current;

[0010] Differential current ID (t) is defined in time as the current entering the differential relay I P (t) and I s The difference between (t):

[0011] I D (t) = I P (t)-I s (t)

[0012] The formula for the constraint current is as follows:

[0013]

[0014] Internal faults and inrush currents in transformers are differentiated based on the following factors:

[0015] I D (t)≥S D ×I R (t)

[0016]

[0017] S D is a normal value that determines the slope of the characteristic;

[0018] S3. When the value of the differential current exceeds the preset threshold, feature value extraction begins. The feature value is extracted using the signal decomposition method. This method is performed by MRA technology. MRA uses wavelet transform to classify signals. Wavelet transform is formed by expanding and translating an original function. Discrete wavelet transformation can be achieved by the following equation:

[0019]

[0020] S4, normalizing waveforms of different amplitudes to a uniform range, completing waveform standardization of the differential current;

[0021] S5, applying MODWT (maximum overlap wavelet transform) to decompose the normalized differential current waveform into multiple energy levels;

[0022] MODWT is applied to the main data using basic wavelet functions. The coefficients of the high-pass filter h[k] and the low-pass filter g[k] are implemented by the following formulas:

[0023]

[0024]

[0025] MODWT uses basic wavelet functions and high-pass and low-pass filters to decompose the input signal (differential current) into expansion coefficients and wavelet coefficients;

[0026] S6, select D7 energy level as feature input to LSTM (long short-term memory) network;

[0027] The LSTM network processes the input D7 energy level features and captures the short-term and long-term relationships of the differential current in the time series by adjusting the trained weights and biases;

[0028] S7, LSTM network distinguishes the inrush current and internal fault of the transformer with high accuracy and speed. If LSTM detects an internal fault, it sends a trip signal to avoid accidents.

[0029] Each LSTM unit in the LSTM network updates its state using the following formula:

[0030] Input Layer:

[0031] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0032] Forgetting layer:

[0033] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0034] Candidate unit status:

[0035]

[0036] Unit status update:

[0037]

[0038] Output layer:

[0039] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0040] Hidden state:

[0041] h t =o t tanh(C t )

[0042] The LSTM network calculates the activation value i of the input layer t , control the impact of new information on the current unit state, and then calculate the activation value f of the forgetting layert , decide how much previous cell state information to keep, and then calculate the candidate cell state Prepare a new candidate state for the current time, then update the unit state C t , based on the results of the forgetting layer and the input layer combined with the updated state and candidate state of the previous time, calculate the activation value o of the output layer t , controls the influence of the current cell state on the hidden state, and finally, updates the hidden state h t , ready for the next step and final output;

[0043] The last layer of the LSTM network is a fully connected layer, which generates the final classification probability through linear transformation and Sigmoid activation function:

[0044]

[0045] in, represents the probability of internal failure, according to the set threshold, if If it is greater than the set threshold, it is judged as an internal fault, otherwise it is an inrush current.

[0046] Preferably: In step S2, under normal circumstances, I D (t) is equal to 0. When the transformer has an inrush current or an internal fault, this value will be greater than S D and the constraint current, so I D (t) will become a non-zero value, which may cause the differential relay to malfunction. The equation in step S2 only shows the abnormal situation and cannot distinguish between inrush current and internal fault. The differential relay should not operate in the case of inrush current, but must operate quickly when an internal fault occurs. However, in both cases, the value of the differential current will be higher than the threshold level, which will cause the differential relay to issue an unwanted tripping command. Therefore, the proposed method must be able to identify internal faults and inrush currents to prevent the differential relay from malfunctioning.

[0047] Preferably: in step S3, the feature extraction process includes calculating time domain features, frequency domain features, wavelet features, time-frequency features, and statistical features.

[0048] Preferably: the step S3 further comprises a differential feature extraction module, wherein the differential feature extraction module comprises a current measurement unit, a differential current calculation unit, a feature extraction unit and a waveform standardization unit;

[0049] The current measuring unit measures the three-phase current of the transformer using a current transformer, and the differential current calculating unit calculates the differential current of the transformer;

[0050] The feature extraction unit extracts multiple features from the differential current;

[0051] The waveform normalization unit normalizes the waveform of the differential current to a finite energy.

[0052] Preferably: in step S5, MODWT uses mother wavelet transform and corresponding high-pass filtering and low-pass filtering to decompose the input differential current into expansion coefficients and wavelet coefficients. Both the high-pass filter and the low-pass filter can be applied to the input signal of each level, and the differential current can be decomposed into n energy levels D1, D2, D3, ... Dn.

[0053] Preferably: the step S5 further includes a wavelet decomposition unit and an LSTM network input unit;

[0054] The wavelet decomposition unit applies maximum overlap wavelet transform (MODWT) to decompose the normalized differential current waveform into multiple energy levels;

[0055] The LSTM network processing unit captures the short-term, long-term and dependency relationships of the differential current in the time series by adjusting the trained weights and biases.

[0056] Preferably: the step S6 further includes a fault detection unit and a trip signal output unit;

[0057] The fault detection unit uses the LSTM network to analyze the results and distinguish between inrush current and internal faults of the transformer with high accuracy and speed;

[0058] When the LSTM network detects an internal fault, the trip signal output unit sends a trip signal.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] 1. In the present invention, the LSTM network receives the D7 energy level as input and processes these time series features to identify inrush current and internal faults. The LSTM network optimizes the performance of the model by learning the long-term and short-term dependencies in the data and adjusting its weights and biases. These adjustments allow the LSTM network to effectively capture the dynamic changes of the differential current, thereby improving the accuracy and reliability of fault detection;

[0061] 2. In the present invention, the trained LSTM network can identify the surge current and internal faults of the transformer with high accuracy and speed. When the network recognizes that the pattern of the differential current meets the characteristics of the internal fault, the system will automatically send a trip signal. This signal triggers the protection mechanism of the transformer, thereby quickly cutting off the power supply, avoiding potential electrical accidents and equipment damage, and ensuring the safety and stability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1A logic diagram of a method for distinguishing between transformer surge current and internal fault according to the present invention;

[0063] Figure 2 This is a schematic diagram of the LSTM network structure of the present invention; DETAILED DESCRIPTION

[0064] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0065] See also Figure 1-2 , the figure shows that the present invention is a method for distinguishing between surge current and internal fault of a transformer, comprising the following steps;

[0066] S1. Use a current transformer to measure the three-phase current of the transformer and obtain current data of each phase, which includes the amplitude and phase information of the phase current;

[0067] S2. Perform differential operation on the three phase currents to obtain the calculated transformer differential current. The differential current is an important indicator for detecting unbalanced load or fault. It can help identify abnormal conditions of the transformer. The calculation result of the differential current will serve as the basic data for subsequent processing.

[0068] Differential current I D (t) is defined in time as the current entering the differential relay I P (t) and I s The difference between (t):

[0069] I D (t) = I P (t)-I s (t)

[0070] The formula for the constraint current is as follows:

[0071]

[0072] Internal faults and inrush currents in transformers are differentiated based on the following factors:

[0073] I D (t)≥S D ×I R (t)

[0074]

[0075] S DIt is a normal value that determines the slope of the characteristic, usually 0.15-0.4. Under normal circumstances, I D (t) is 0, but when the transformer has surge current or internal fault, I D (t) will become greater than S D and the constraint current, so I D (t) will become a non-zero value, which may cause the differential relay to malfunction. The above equation only shows the abnormal situation and cannot distinguish between inrush current and internal fault. The differential relay should not operate under inrush current, and must operate quickly when an internal fault occurs. However, in both cases, the value of the differential current will be higher than the threshold level, which will cause the differential relay to issue an unwanted trip command. Therefore, the proposed method must be able to identify internal faults and inrush currents to prevent the differential relay from malfunctioning. The calculated differential current data also needs to be further processed;

[0076] S3. When the value of the differential current is abnormal and exceeds the preset threshold, it indicates that there may be an internal fault or surge current. At this time, feature value extraction begins. In this stage, the relevant data of the differential current are first collected and processed to extract meaningful features. The feature extraction process includes calculating time domain features, frequency domain features, wavelet features, time-frequency features, and statistical features to provide further understanding of the signal. The extracted features will be used in subsequent classification and judgment steps.

[0077] Specifically, the eigenvalues ​​are extracted by using the signal decomposition method, which is performed by MRA technology. It can provide frequency-time information of a given signal in two dimensions and multiple energy levels at the same time. MRA applies wavelet transform for signal classification. Wavelet transform is formed by expanding and translating an original function. Discrete wavelet transformation can be realized by the following equation:

[0078]

[0079] S4, normalizing waveforms of different amplitudes to a uniform range, completing waveform standardization of the differential current;

[0080] Specifically, the energy of the differential current waveform is adjusted to a limited range, which is usually achieved by calculating the energy of the waveform and scaling it to a standard range. This standardization not only reduces the impact due to amplitude differences, but also improves the stability and accuracy of subsequent algorithms.

[0081] S5. Apply MODWT (maximum overlap wavelet transform) to decompose the standardized differential current waveform into multiple energy levels, among which the D7 energy level contains important features for distinguishing surge current from internal faults. The D7 energy level is particularly important for capturing the key features of surge current and internal faults, because surge current usually manifests as high-frequency components. By analyzing the characteristics of the D7 energy level, the two fault modes can be effectively distinguished.

[0082] MODWT is applied to the main data using basic wavelet functions. The coefficients of the high-pass filter h[k] and the low-pass filter g[k] are implemented by the following formulas:

[0083]

[0084]

[0085] MODWT uses basic wavelet functions and high-pass and low-pass filters to decompose the input signal (differential current) into expansion coefficients and wavelet coefficients. In fact, at each level, high-pass and low-pass filters are applied to the input signal and can be performed up to the nth level. Therefore, the differential current is decomposed into 9 energy levels, D1, D2, …, D9, to provide high resolution;

[0086] S6, select D7 energy level as a feature input to the LSTM (Long Short-Term Memory) network. By inputting D7 energy level into the LSTM network, the LSTM network can learn the pattern and change of the differential current signal in the time series, thereby effectively distinguishing the surge current and internal fault of the transformer. If an internal fault is detected, a trip signal is issued;

[0087] The LSTM network processes the input D7 energy level features and captures the short-term and long-term relationships of the differential current in the time series by adjusting the trained weights and biases;

[0088] S7, LSTM network distinguishes the inrush current and internal fault of the transformer with high accuracy and speed. If LSTM detects an internal fault, it sends a trip signal to avoid accidents.

[0089] Each LSTM unit in the LSTM network updates its state using the following formula:

[0090] Input Layer:

[0091] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0092] Forgetting layer:

[0093] f t=σ(W f ·[h t-1 ,x t ]+b f )

[0094] Candidate unit status:

[0095]

[0096] Unit status update:

[0097]

[0098] Output layer:

[0099] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0100] Hidden state:

[0101] h t =o t tanh(C t )

[0102] The LSTM network calculates the activation value i of the input layer t , control the impact of new information on the current unit state, and then calculate the activation value f of the forgetting layer t , decide how much previous cell state information to keep, and then calculate the candidate cell state Prepare a new candidate state for the current time, then update the unit state C t , based on the results of the forgetting layer and the input layer combined with the updated state and candidate state of the previous time, calculate the activation value o of the output layer t , controls the influence of the current cell state on the hidden state, and finally, updates the hidden state h t , ready for the next step and final output;

[0103] The last layer of the LSTM network is a fully connected layer, which generates the final classification probability through linear transformation and Sigmoid activation function:

[0104]

[0105] in, represents the probability of internal failure, according to the set threshold, if If it is greater than the set threshold, it is judged as an internal fault, otherwise it is an inrush current.

[0106] Wherein, the step S3 further includes a differential feature extraction module, and the differential feature extraction module includes a current measurement unit, a differential current calculation unit, a feature extraction unit and a waveform normalization unit;

[0107] The current measuring unit measures the three-phase current of the transformer using a current transformer, and the differential current calculating unit calculates the differential current of the transformer;

[0108] The feature extraction unit extracts multiple features from the differential current;

[0109] The waveform normalization unit normalizes the waveform of the differential current to finite energy for the convenience of subsequent processing and analysis.

[0110] Specifically, in step S5, MODWT uses mother wavelet transform and corresponding high-pass filtering and low-pass filtering to decompose the input differential current into expansion coefficients and wavelet coefficients. Both the high-pass filter and the low-pass filter can be applied to the input signal of each level, and the differential current can be decomposed into n energy levels D1, D2, D3, ... Dn.

[0111] It can be understood that the step S5 further includes a wavelet decomposition unit and an LSTM network input unit;

[0112] The wavelet decomposition unit applies the maximum overlap wavelet transform (MODWT) to decompose the normalized differential current waveform into multiple energy levels, among which the D7 energy level is selected as an important feature for distinguishing inrush current from internal faults. The D7 energy level feature is then input into the LSTM network.

[0113] The LSTM network processing unit captures the short-term, long-term and dependency relationships of the differential current in the time series by adjusting the trained weights and biases.

[0114] Wherein, the step S6 further includes a fault detection unit and a trip signal output unit;

[0115] The fault detection unit uses the LSTM network to analyze the results and distinguish between inrush current and internal faults of the transformer with high accuracy and speed;

[0116] When the LSTM network detects an internal fault, the trip signal output unit sends out a trip signal to avoid potential accidents.

[0117] The above content is a further detailed description of the present invention in combination with specific implementation methods. It cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as belonging to the scope of protection determined by the claims submitted for the present invention.

Claims

1. A method for distinguishing transformer surge current from internal fault, characterized in that: The method comprises the following steps: S1. Use a current transformer to measure the three-phase current of the transformer and obtain the current data of each phase; S2, performing differential operation on the three phase currents to obtain the calculated transformer differential current; The differential current ID(t) is defined in time as the difference between the currents entering the differential relay IP(t) and IS(t): I D (t)=I P (t)-I s (t) The formula for the constraint current is as follows: Internal faults and inrush currents in transformers are differentiated based on the following factors: I D (t)≥S D ×I R (t) S D is a normal value that determines the slope of the characteristic; S3. When the value of the differential current exceeds the preset threshold, feature value extraction begins. The feature value is extracted using the signal decomposition method. This method is performed by MRA technology. MRA uses wavelet transform to classify signals. Wavelet transform is formed by expanding and translating an original function. Discrete wavelet transformation can be achieved by the following equation: S4, normalizing waveforms of different amplitudes to a uniform range, completing waveform standardization of the differential current; S5, applying MODWT (maximum overlap wavelet transform) to decompose the normalized differential current waveform into multiple energy levels; MODWT is applied to the main data using basic wavelet functions. The coefficients of the high-pass filter h[k] and the low-pass filter g[k] are implemented by the following formulas: MODWT uses basic wavelet functions and high-pass and low-pass filters to decompose the input signal (differential current) into expansion coefficients and wavelet coefficients; S6, select D7 energy level as feature input to LSTM (long short-term memory) network; The LSTM network processes the input D7 energy level features and captures the short-term and long-term relationships of the differential current in the time series by adjusting the trained weights and biases; S7, LSTM network distinguishes the inrush current and internal fault of the transformer with high accuracy and speed. If LSTM detects an internal fault, it sends a trip signal; Each LSTM unit in the LSTM network updates its state using the following formula: Input Layer: i t =σ(W i ·[h t-1 ,x t ]+b i ) Forgetting layer: f t =σ(W f ·[h t-1 ,x t ]+b f ) Candidate unit status: Unit status update: Output layer: the t =σ(W o ·[h t-1 ,x t ]+b o ) Hidden state: h t =o t ·tanh(C t ) The LSTM network calculates the activation value i of the input layer t , control the impact of new information on the current unit state, and then calculate the activation value f of the forgetting layer t , decide how much previous cell state information to keep, and then calculate the candidate cell state Prepare a new candidate state for the current time, then update the unit state C t , based on the results of the forgetting layer and the input layer combined with the updated state and candidate state of the previous time, calculate the activation value o of the output layer t , controls the influence of the current cell state on the hidden state, and finally, updates the hidden state h t , ready for the next step and final output; The last layer of the LSTM network is a fully connected layer, which generates the final classification probability through linear transformation and Sigmoid activation function: in, represents the probability of internal failure, according to the set threshold, if If it is greater than the set threshold, it is judged as an internal fault and the responsible one is the inrush current.

2. A method for distinguishing transformer surge current from internal fault according to claim 1, characterized in that: In step S2, normally I D (t) is equal to 0. When the transformer has an inrush current or an internal fault, this value will be greater than S D The product of the constraint current.

3. A method for distinguishing transformer surge current from internal fault according to claim 1, characterized in that: In step S3, the feature extraction process includes calculating time domain features, frequency domain features, wavelet features, time-frequency features, and statistical features.

4. A method for distinguishing transformer surge current from internal fault according to claim 3, characterized in that: The step S3 further includes a differential feature extraction module, which includes a current measurement unit, a differential current calculation unit, a feature extraction unit and a waveform standardization unit; The current measuring unit measures the three-phase current of the transformer using a current transformer, and the differential current calculating unit calculates the differential current of the transformer; The feature extraction unit extracts multiple features from the differential current; The waveform normalization unit normalizes the waveform of the differential current to a finite energy.

5. A method for distinguishing transformer surge current from internal fault according to claim 1, characterized in that: In step S5, MODWT uses mother wavelet transform and corresponding high-pass filtering and low-pass filtering to decompose the input differential current into expansion coefficients and wavelet coefficients. Both the high-pass filter and the low-pass filter can be applied to the input signal of each level, and the differential current can be decomposed into n energy levels D1, D2, D3, ... Dn.

6. A method for distinguishing transformer surge current from internal fault according to claim 5, characterized in that: In the step S5, a wavelet decomposition unit and an LSTM network input unit are also included; The wavelet decomposition unit applies maximum overlap wavelet transform (MODWT) to decompose the normalized differential current waveform into multiple energy levels; The LSTM network processing unit captures the short-term, long-term and dependency relationships of the differential current in the time series by adjusting the trained weights and biases.

7. A method for distinguishing transformer surge current from internal fault according to claim 1, characterized in that: The step S6 also includes a fault detection unit and a trip signal output unit; The fault detection unit uses the LSTM network to analyze the results and distinguish between inrush current and internal faults of the transformer with high accuracy and speed; When the LSTM network detects an internal fault, the trip signal output unit sends a trip signal.

Images