A short-circuit current zero-point prediction system fused with a deep learning algorithm

Abstract

The application provides a short-circuit current zero point prediction system fusing a deep learning algorithm, relates to the field of phase control breaking of a large-capacity vacuum circuit breaker, and comprises a current signal acquisition module, a model construction processing module, a deep learning prediction module and a phase control instruction execution module. The current signal acquisition module is used for collecting short-circuit current data of the system in real time and performing time synchronization and cache processing. The model construction processing module initializes model parameters by constructing a short-circuit current mathematical model containing direct current and alternating current components. The deep learning prediction module adopts an alternating optimization method fusing a recursive least square algorithm and a self-adaptive matrix estimation algorithm, fits a current waveform in real time and predicts a zero point moment. The phase control instruction execution module generates a control instruction to realize phase control breaking according to a prediction result and a circuit breaker opening time. The system can effectively improve breaking synchronization and breaking reliability of a multi-break vacuum circuit breaker, is suitable for real-time zero point prediction of a power grid system fault short-circuit current, and is especially suitable for protection and breaking of a large-capacity power generator motor, such as a pumped storage power station, in a high-frequency opening and closing scene.

Description

Technical Field

[0001] This invention relates to the field of phase-controlled breaking of high-capacity vacuum circuit breakers, and specifically to a short-circuit current zero-point prediction system that integrates deep learning algorithms. Background Technology

[0002] Multi-break parallel vacuum circuit breakers must avoid asynchronous arc extinction at zero-crossing points during current interruption, which would put arc-extinguishing and breaking pressure on the vacuum interrupter chamber after the initial break. Therefore, phase-controlled interruption is one of the core technologies of this type of circuit breaker. Furthermore, to fully utilize the intermittent, fluctuating, and random nature of wind power, photovoltaic, and other new energy sources, pumped storage power stations frequently switch between generator and pumping modes daily. Multi-break parallel vacuum circuit breakers interrupt load current 5 to 10 times daily. Without phase-controlled technology, the arc-extinguishing chamber suffers severe cumulative erosion, significantly affecting the circuit breaker's breaking reliability and electrical life. The standard time constant of the short-circuit current from the system source in a pumped storage power station is 45 ms, with a DC component accounting for 20% to 75%; the standard time constant of the short-circuit current from the generator source is 150 ms, with a DC component accounting for 20% to 130%. In actual operation, the system and unit parameters vary widely, and the short-circuit current model parameters under the two types of faults are highly uncertain. In addition, the nonlinear characteristics of the short-circuit process make it difficult to predict the zero point of the current quickly and accurately.

[0003] Many zero-crossing prediction systems have been developed. Extensive research and reference have revealed that existing zero-crossing prediction systems, such as the one disclosed in CN112347415B, generally include the following methods: determining whether a short-circuit fault has occurred by combining fault current detection methods with collected fault current values; if a fault has occurred, the invention obtains a constant coefficient matrix based on the sampling frequency and sampling time window length, and initializes the covariance matrix and parameter matrix; further sampling fault current data to update the estimation error, continuously updating the estimation error, variable forgetting factor, gain matrix, parameter matrix, posterior error, and covariance matrix sequentially; finally, constructing a mathematical model of the fault current based on the updated parameter matrix and the sampling time window length, and using Newton's iteration method to obtain the predicted zero-crossing point, thus predicting the zero-crossing time of the short-circuit current. For example, the zero-crossing prediction algorithm disclosed in CN115389812B proposes a short-circuit current zero-crossing prediction method based on a long short-term memory network, and directly learns current sequence features through a long short-term memory (LSTM) model. However, the effectiveness of this algorithm depends on the richness of the offline training samples of the fault current. The aforementioned publicly available algorithms or systems all employ classical or deep learning algorithms to study the zero-point prediction problem, without fully integrating the characteristics and advantages of different algorithms, nor simultaneously considering the impact of wide DC time constant and high proportion of DC component on short-circuit current zero-point prediction. Their adaptability and generalization ability are generally weak. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings by proposing a short-circuit current zero-point prediction system that integrates deep learning algorithms.

[0005] The present invention adopts the following technical solution:

[0006] A short-circuit current zero-point prediction system integrating deep learning algorithms includes a current signal acquisition module, a model building and processing module, a deep learning prediction module, and a phase control command execution module;

[0007] The current signal acquisition module is used to acquire system fault short-circuit current information; the model building and processing module is used to build and process the short-circuit current model; the deep learning prediction module is used to optimize the current fitting curve and predict the current zero point; and the phase control command execution module is used to control the command to disconnect the circuit.

[0008] The current signal acquisition module includes a data sampling unit and a data buffering unit. The data sampling unit is used to collect system fault short-circuit current data, and the data buffering unit is used to buffer the sampled data for subsequent processing.

[0009] The model building and processing module includes a short-circuit current modeling unit and a parameter initialization unit. The short-circuit current modeling unit is used to build a mathematical model of a short-circuit circuit and perform model calculations. The parameter initialization unit is used to initialize the parameters in the model.

[0010] The deep learning prediction module includes an optimization process control unit, a recursive least squares optimization unit, and a deep learning parameter optimization unit. The optimization process control unit is used to manage and control the optimization process, the recursive least squares optimization unit is used to fit linear parameters, and the deep learning parameter optimization unit is used to iteratively update nonlinear parameters.

[0011] The phase control command execution module includes a circuit breaker tripping time preset unit, a phase control interruption judgment unit, and a control command generation unit. The circuit breaker tripping time preset unit is used to set the tripping action duration of the circuit breaker, the phase control interruption judgment unit is used to generate the optimal interruption interval, and the control command generation unit is used to generate commands to control the circuit breaker to disconnect the circuit.

[0012] Furthermore, the short-circuit current modeling unit includes a nonlinear expression generator, a function expansion calculator, and a modeling and simulation processor. The nonlinear expression generator is used to generate a current model containing a DC component and a sinusoidal AC term. The function expansion calculator is used to perform a linear combination approximate expansion of the original model. The modeling and simulation processor is used to establish a mathematical model based on the current data and output simulation waveforms.

[0013] The short-circuit circuit model constructed by the nonlinear expression generator is as follows:

[0014] ;

[0015] , , ;

[0016] Where t represents time, and I(t) represents the fitted current at time t. p This represents the AC amplitude of the short-circuit current. The initial phase angle of the short-circuit current AC is A. sy Indicates the percentage of DC component. is the DC time constant of the short-circuit current.

[0017] Furthermore, the optimization process control unit includes an iteration scheduler, a termination condition judger, and a time manager. The iteration scheduler is used to set and control the iteration rhythm of alternating optimizations. The termination condition judger is used to determine whether the current optimization meets the termination condition. The time manager is used to ensure that the optimization process is completed within a specified time window.

[0018] The iterative scheduler divides the parameter vector into linear parameter blocks. and nonlinear parameter blocks Iterative scheduling is performed according to the following process:

[0019] S1. Set the iteration count n=0 and initialize the parameters. ;

[0020] S2, Fixed Call the recursive least squares optimization unit to update ;

[0021] S3, Fixed Call the deep learning parameter optimization unit to update ;

[0022] S4. Let n = n + 1, call the termination condition checker to determine whether to terminate the iteration. If not, return to step S2.

[0023] Furthermore, the recursive least squares optimization unit includes a gain calculator, a covariance matrix manager, and a linear parameter estimator. The gain calculation is used to dynamically adjust the gain factor during the fitting process, the covariance matrix manager is used to maintain the error covariance matrix during the optimization process, and the linear parameter estimator is used to recursively optimize linear parameters such as AC amplitude and initial phase angle.

[0024] The gain calculator updates the gain matrix K according to the following formula. n+1 :

[0025] ;

[0026] The covariance matrix manager updates the covariance matrix P according to the following formula. n+1 :

[0027] ;

[0028] The linear parameter estimator updates the parameter vector according to the following formula: :

[0029] ;

[0030] in, T is the regression vector. s The sampling interval is... I0 is the damping factor, a diagonal matrix.

[0031] Furthermore, the deep learning parameter optimization unit includes a gradient update processor, a momentum estimator, and a parameter correction calculator. The gradient update processor is used to calculate and update the gradient information of the nonlinear parameters, the momentum estimator is used to capture the gradient change trend and control the convergence direction, and the parameter correction calculator is used to correct the nonlinear parameters.

[0032] The beneficial effects achieved by this invention are:

[0033] This system employs an alternating optimization strategy that combines recursive least squares method with deep learning adaptive optimization algorithm. This strategy can simultaneously take into account the dynamic updates of linear and nonlinear parameters. Even under conditions of high DC component and large time constant, it can still achieve zero-point prediction within the millisecond error range, significantly improving prediction accuracy and algorithm robustness.

[0034] To further understand the features and technical content of the present invention, please refer to the following detailed description and drawings of the present invention. However, the drawings provided are for reference and illustration only and are not intended to limit the present invention. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the overall structural framework of the present invention;

[0036] Figure 2 This is a schematic diagram of the current signal acquisition module of the present invention;

[0037] Figure 3 This is a schematic diagram of the model construction and processing module of the present invention;

[0038] Figure 4 This is a schematic diagram of the deep learning prediction module of the present invention;

[0039] Figure 5 This is a schematic diagram of the phase control instruction execution module of the present invention;

[0040] Figure 6 This is a comparison chart of the actual test curves of the present invention;

[0041] Figure 7 This is a diagram showing the error magnitude and three-dimensional scatter distribution of the zero point prediction in this invention. Detailed Implementation

[0042] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can understand the advantages and effects of the present invention from the content disclosed in this specification. The present invention can be implemented or applied through other different specific embodiments, and various details in this specification can also be modified and changed based on different viewpoints and applications without departing from the spirit of the present invention. Furthermore, the accompanying drawings of the present invention are for simple illustrative purposes only and are not depictions of actual dimensions; this is stated beforehand. The following embodiments will further describe the relevant technical content of the present invention in detail, but the disclosed content is not intended to limit the scope of protection of the present invention.

[0043] Example 1.

[0044] This embodiment provides a short-circuit current zero-point prediction system that integrates deep learning algorithms, combined with... Figure 1 It includes a current signal acquisition module, a model building and processing module, a deep learning prediction module, and a phase control command execution module;

[0045] The current signal acquisition module is used to acquire the current information of the circuit; the model building and processing module is used to build and process the short-circuit current model; the deep learning prediction module is used to optimize the current fitting curve and predict the current zero point; and the phase control command execution module is used to control the command to disconnect the circuit.

[0046] Combination Figure 2 The current signal acquisition module includes a data sampling unit and a data buffering unit. The data sampling unit is used to collect system fault short-circuit current data, and the data buffering unit is used to buffer the sampled data for subsequent processing.

[0047] Combination Figure 3 The model building and processing module includes a short-circuit current modeling unit and a parameter initialization unit. The short-circuit current modeling unit is used to build a mathematical model of a short-circuit circuit and perform model calculations. The parameter initialization unit is used to initialize the parameters in the model.

[0048] Combination Figure 4The deep learning prediction module includes an optimization process control unit, a recursive least squares optimization unit, and a deep learning parameter optimization unit. The optimization process control unit is used to manage and control the optimization process, the recursive least squares optimization unit is used to fit linear parameters, and the deep learning parameter optimization unit is used to iteratively update nonlinear parameters.

[0049] Combination Figure 5 The phase control command execution module includes a circuit breaker tripping time preset unit, a phase control interruption judgment unit, and a control command generation unit. The circuit breaker tripping time preset unit is used to set the tripping action duration of the circuit breaker. The phase control interruption judgment unit is used to generate the optimal interruption interval. The control command generation unit is used to generate commands to control the circuit breaker to disconnect the circuit.

[0050] The short-circuit current modeling unit includes a nonlinear expression generator, a function expansion calculator, and a modeling and simulation processor. The nonlinear expression generator is used to generate a current model containing a DC component and a sinusoidal AC term. The function expansion calculator is used to perform a linear combination approximate expansion of the original model. The modeling and simulation processor is used to establish a mathematical model based on the current data and output simulation waveforms.

[0051] The short-circuit circuit model constructed by the nonlinear expression generator is as follows:

[0052] ;

[0053] , , ;

[0054] Where t represents time, and I(t) represents the fitted current at time t. p This represents the AC amplitude of the short-circuit current. The initial phase angle of the short-circuit current AC is A. sy Indicates the percentage of DC component. is the DC time constant of the short-circuit current.

[0055] The optimization process control unit includes an iteration scheduler, a termination condition judger, and a time manager. The iteration scheduler is used to set and control the iteration rhythm of alternating optimizations. The termination condition judger is used to determine whether the current optimization meets the termination condition. The time manager is used to ensure that the optimization process is completed within a specified time window.

[0056] The termination condition determiner defines the loss function according to the following formula. :

[0057] ;

[0058] in, Let N be the parameter vector, and y be the number of data points. i This represents the i-th current value actually collected;

[0059] The iterative scheduler divides the parameter vector into linear parameter blocks. and nonlinear parameter blocks Iterative scheduling is performed according to the following process:

[0060] S1. Set the iteration count n=0 and initialize the parameters. ;

[0061] S2, Fixed Call the recursive least squares optimization unit to update ;

[0062] S3, Fixed Call the deep learning parameter optimization unit to update ;

[0063] S4. Let n = n + 1, call the termination condition checker to determine whether to terminate the iteration. If not, return to step S2.

[0064] The recursive least squares optimization unit includes a gain calculator, a covariance matrix manager, and a linear parameter estimator. The gain calculation is used to dynamically adjust the gain factor during the fitting process. The covariance matrix manager is used to maintain the error covariance matrix during the optimization process. The linear parameter estimator is used to recursively optimize linear parameters such as AC amplitude and initial phase angle.

[0065] The gain calculator updates the gain matrix K according to the following formula. n+1 :

[0066] ;

[0067] The covariance matrix manager updates the covariance matrix P according to the following formula. n+1 :

[0068] ;

[0069] The linear parameter estimator updates the parameter vector according to the following formula: :

[0070] ;

[0071] in, T is the regression vector. s The sampling interval is... I0 is the damping factor, a diagonal matrix.

[0072] The deep learning parameter optimization unit includes a gradient update processor, a momentum estimator, and a parameter correction calculator. The gradient update processor is used to calculate and update the gradient information of the nonlinear parameters, the momentum estimator is used to capture the gradient change trend and control the convergence direction, and the parameter correction calculator is used to correct the nonlinear parameters.

[0073] The gradient update processor calculates the gradient of the loss function according to the following formula. :

[0074] ;

[0075] Among them, y n =I(t n ), Indicates y n The correction value, t n Indicates the nth sampling time;

[0076] The momentum estimator calculates the first moment M according to the following formula. n Second moment V n Correction value of the first moment and the correction value of the second moment :

[0077] ;

[0078] ;

[0079] ;

[0080] ;

[0081] in, and This is the momentum parameter, with a default value of: , ;

[0082] The parameter correction calculator calculates the new nonlinear parameters according to the following formula:

[0083] ;

[0084] ;

[0085] in, For learning rate, The minimum value is used to prevent division by zero;

[0086] The short-circuit current zero-point prediction method of this invention, which integrates deep learning algorithms, uses any combination of one or more programming languages ​​to write code for performing the operations disclosed herein. These programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as "C++" or similar languages. The code is executed on the operating system, partially on the user device, or as a programming language. Partial code information of this system is shown below:

[0087] class ZeroPointPredictionSystem:

[0088] def __init__(self):

[0089] self.data_sampling = DataSamplingUnit()

[0090] self.time_sync = TimeSyncUnit()

[0091] self.data_cache = DataCacheUnit()

[0092] self.window_cutter = SignalWindowCutter()

[0093] self.model_builder = ShortCircuitModelBuilder()

[0094] self.param_initializer = ParameterInitializer()

[0095] self.flow_controller = OptimizationFlowController()

[0096] self.linear_optimizer = RDLSOptimizer()

[0097] self.nonlinear_optimizer = DeepLearningOptimizer()

[0098] self.breaker_coordinator = MultiBreakerCoordinator()

[0099] self.zero_point_unit = ZeroPointDecisionUnit()

[0100] self.command_generator = ControlCommandGenerator()

[0101] def run_system(self):

[0102] self.data_sampling.collect_current_data()

[0103] self.time_sync.synchronize_time()

[0104] self.data_cache.cache_data()

[0105] self.window_cutter.cut_window(None)

[0106] self.model_builder.build_model()

[0107] self.param_initializer.initialize_parameters()

[0108] self.flow_controller.control_flow()

[0109] self.linear_optimizer.optimize_linear_parameters()

[0110] self.nonlinear_optimizer.optimize_nonlinear_parameters()

[0111] self.breaker_coordinator.evaluate_contact_difference()

[0112] self.zero_point_unit.generate_optimal_window()

[0113] self.command_generator.generate_command()

[0114] if __name__ == "__main__":

[0115] system = ZeroPointPredictionSystem()

[0116] system.run_system().

[0117] The system was used for actual testing to obtain a comparison waveform between the fitted curve and the curve containing noise and harmonics. The waveform was then processed to obtain... Figure 6 and Figure 7 . Figure 6 It reflects the similarity between the predicted waveform and the fitted waveform; Figure 7 It reflects the magnitude of the error in predicting the zero point and the distribution of three-dimensional scattered points.

[0118] The content disclosed above is only a preferred and feasible embodiment of the present invention, and is not intended to limit the scope of protection of the present invention. Therefore, all equivalent technical changes made based on the content of the present invention specification and drawings are included within the scope of protection of the present invention. Furthermore, the elements therein can be updated as technology develops.

Claims

1. A short circuit current zero point prediction system fused with deep learning algorithm, characterized in that, The current signal acquisition module, the model construction processing module, the deep learning prediction module and the phase control instruction execution module are included. The current signal acquisition module is used for acquiring system short-circuit current information, the model construction processing module is used for constructing a short-circuit current model and initializing, the deep learning prediction module is used for optimizing a current fitting curve and predicting a current zero point, and the phase control instruction execution module is used for controlling a breaking circuit. The current signal acquisition module includes a data sampling unit and a data cache unit, the data sampling unit is used for acquiring system short-circuit current data, and the data cache unit is used for caching the sampling data for subsequent processing. The model construction processing module includes a short-circuit current modeling unit and a parameter initialization unit, the short-circuit current modeling unit is used for constructing a short-circuit current mathematical model and performing model calculation, and the parameter initialization unit is used for initializing parameters in the model. The deep learning prediction module includes an optimization process control unit, a recursive least squares optimization unit and a deep learning parameter optimization unit, the optimization process control unit is used for managing and controlling the optimization process, the recursive least squares optimization unit is used for fitting linear parameters, and the deep learning parameter optimization unit is used for iteratively updating nonlinear parameters. The phase control instruction execution module includes a breaking time presetting unit, a phase control breaking judgment unit and a control instruction generating unit, the breaking time presetting unit is used for setting the breaking action time length of the circuit breaker, the phase control breaking judgment unit is used for generating an optimal breaking interval, and the control instruction generating unit is used for generating an instruction to control the circuit breaker to disconnect the circuit. The short-circuit current modeling unit includes a nonlinear expression generator, a function expansion calculator and a modeling simulation processor, the nonlinear expression generator is used for generating a current model containing a direct current component and a sinusoidal alternating current term, the function expansion calculator is used for linearly combining and approximating the original model, and the modeling simulation processor is used for establishing a mathematical model according to the current data and outputting a simulation waveform. The short-circuit current model constructed by the nonlinear expression generator is as follows: ； ， ， ； where t represents time, I(t) represents the fitted current at time t, I p is the short circuit current AC amplitude, is the short circuit current AC initial phase angle, A sy represents the DC component percentage, is the short circuit current DC time constant; The optimization process control unit includes an iteration scheduler, a termination condition judge and a time manager, the iteration scheduler is used for setting and controlling the iteration rhythm of alternating optimization, the termination condition judge is used for judging whether the current optimization meets the end condition, and the time manager is used for ensuring that the optimization process is completed within a specified time window. The termination condition determiner defines a loss function according to the following equation : ； wherein, is the parameter vector, N is the number of data points, y i denotes the i-th actually acquired current value; The iterative scheduler separates the parameter vector into linear parameter blocks and non-linear parameter blocks and iteratively schedules according to the following procedure: S1, set iteration count value n = 0, initialize parameters ; S2, fix , call a recursive least squares optimization unit update ; S3, fixing , calling a deep learning parameter optimization unit to update ; S4, let n=n+1, call the termination condition judge to judge whether to terminate iteration, if not, return to step S2; The recursive least squares optimization unit includes a gain calculator, a covariance matrix manager and a linear parameter estimator, the gain calculator is used for dynamically adjusting the gain factor in the fitting process, the covariance matrix manager is used for maintaining the error covariance matrix in the optimization process, and the linear parameter estimator is used for recursively optimizing linear parameters such as alternating current amplitude and initial phase angle. The gain calculator updates the gain matrix K according to the following equation n+1 : ； The covariance matrix manager updates the covariance matrix P according to n+1 : ； The linear parameter estimator updates the parameter vector : ； wherein is the regression vector, T s is the sampling interval, is the damping factor, I0is a diagonal matrix.

2. The short circuit current zero point prediction system fused with deep learning algorithm of claim 1, wherein, The deep learning parameter optimization unit comprises a gradient update processor, a momentum estimator and a parameter correction calculator, the gradient update processor is used for calculating and updating gradient information of the nonlinear parameter, the momentum estimator is used for capturing a gradient change trend to control a convergence direction, and the parameter correction calculator is used for correcting the nonlinear parameter; The gradient update processor calculates the gradient according to the following formula: ; The gradient is used for reflecting a current optimization direction; The momentum estimator calculates a first moment, a second moment and a corresponding correction value according to the following formula; Compute the first moment: ; The first moment is used for smoothing current gradient fluctuation; Computing the second moment: ; The second moment is used for describing a gradient amplitude change trend; Because the momentum estimation in the initial stage has a small problem, the first moment and the second moment need to be corrected for deviation; Corrected first moment: ; Corrected second moment: ; The parameter correction calculator adjusts the nonlinear parameter according to the following formula Compute nonlinear parameter iteration step: ; Optimize the non-linear parameters of this step: .

3. The short circuit current zero point prediction system fused with deep learning algorithm of claim 2, wherein, A damping factor is used in the recursive least square method to suppress parameter divergence, the damping factor is initialized as 1e-4, the stability of the optimization process is significantly improved by introducing a penalty term of parameter change rate in the least square parameter optimization objective function, the explosion phenomenon caused by the dramatic increase of parameter update amount when the covariance matrix decreases is prevented, and the optimization process is prevented from converging prematurely due to improper initial value setting of the parameter vector and the covariance matrix.

4. The short circuit current zero point prediction system fused with deep learning algorithm of claim 3, wherein, The learning rate is dynamically adjusted in an exponential decay manner in the deep learning parameter optimization process, the learning rate: step size, is an important hyperparameter of the neural network optimization algorithm, the step size satisfies 0 < a < 1 / L, L is the gradient Lipschitz constant, in the deep learning algorithm, the value of the learning rate a is very important, if it is too large, it will not converge, if it is too small, the convergence speed is too slow, the learning rate should be kept large at the beginning to ensure the convergence speed, and should be reduced when converging to the vicinity of the optimal point to avoid oscillation, the learning rate adjustment can be performed through a decay manner, which is called learning rate annealing, the present zero point prediction algorithm adopts exponential decay, wherein β < 1 is the decay rate, n is the recursive calculation step, the learning rate initialization = 1e-2, = 0.

99.

5. The short circuit current zero point prediction system fused with deep learning algorithm as claimed in claim 4, wherein, The deep learning prediction module normalizes and pretreats the short-circuit current sampling data before processing, the normalization operation adjusts the numerical interval of the sample data, unifies the dimension difference, improves the balance of the model parameter update, effectively alleviates the problems of gradient explosion or disappearance and the like, optimizes the stability and generalization ability in the network training process, suppresses the internal covariant offset, stabilizes the network training, meets the gradient change requirement of the Adam moment estimation algorithm for optimizing the direct current decay time constant τ.

6. The short circuit current zero point prediction system fused with deep learning algorithm as claimed in claim 5, wherein, In the process based on gradient descent, if the gradient increases sharply, updating the parameter with a large gradient will cause it to deviate from the optimal point. In order to avoid this situation, when the modulus of the gradient is greater than a certain threshold, the gradient is intercepted, which is called gradient clipping. When the kth iteration is performed, the gradient is Gn, a interval [a, b] is given, and Gn is: Gn = max(min(Gn, b), a).

Images