Low-frequency load shedding method and system based on zero crossing point real-time frequency tracking

Abstract

The invention discloses a low-frequency load shedding method and system based on zero crossing point real-time frequency tracking, and the method specifically comprises the steps: setting a dynamic time window according to the power frequency characteristics of a sensing signal, and removing an abnormal zero crossing point exceeding the dynamic time window; matching a pre-stored harmonic feature library, establishing a mapping relation between a harmonic mode and a phase offset, and performing adaptive phase compensation on a pseudo zero crossing point of the sensing signal; a sensing signal, PMU phasor data and meteorological parameters are input into a three-dimensional convolutional neural network, a cross-modal incidence matrix is calculated through a self-attention mechanism, a future frequency change trajectory is predicted in combination with an LSTM unit, and an optimal load shedding strategy is output; and based on the optimal load shedding strategy, in combination with a forward acceleration criterion and reverse fault locking, executing emergency load shedding according to load importance grading. The technical bottleneck of a traditional low-frequency load shedding device is effectively solved, and the performance and reliability of frequency stability control of the power system are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power distribution networks, and in particular to a low-frequency load reduction method and system based on zero-crossing real-time frequency tracking. Background Art

[0002] With the transformation of the global energy mix and the rapid development of renewable energy generation technologies, the integration of a high proportion of renewable energy into the power grid has become an inevitable trend in power system development. Renewable energy sources such as wind power and photovoltaics are characterized by intermittency, volatility, and uncertainty. Their large-scale grid integration poses unprecedented challenges to the frequency stability of the power system. In the event of rapid fluctuations in renewable energy output or sudden disconnection from the grid, the grid frequency can drop in milliseconds or even faster. If effective control measures are not implemented in a timely manner, this could trigger a system frequency collapse, leading to widespread power outages and seriously impacting the safe and stable operation of the power system.

[0003] Traditional low-frequency load shedding devices, as an important means of controlling power system frequency stability, play a key role in ensuring system security. However, faced with the complex operating conditions of scenarios with a high proportion of renewable energy access, traditional low-frequency load shedding devices have exposed many technical bottlenecks and limitations, which are specifically reflected in the following aspects:

[0004] 1. Insufficient frequency measurement accuracy: Traditional low-frequency load shedding devices mostly use software algorithms based on fixed delay period measurement for frequency calculation. When the frequency changes rapidly, due to the limitations of the algorithm itself, cumulative errors are easily generated, resulting in deviations in the frequency calculation results, which in turn affects the accuracy of the load shedding decision.

[0005] 2. Sensitivity to harmonic interference: When the power grid has high harmonic content (such as total harmonic distortion (THD) > 8%), the zero-crossing detection algorithm of traditional low-frequency load shedding devices is susceptible to harmonic interference, and the misjudgment rate increases significantly (up to 15%), resulting in a frequency calculation error exceeding ±0.2Hz, seriously affecting the reliability and effectiveness of the load shedding device.

[0006] 3. Slow response speed: Traditional low-frequency load-shedding devices typically require a delay of 300-500ms from detecting a frequency over-limit to executing a load-shedding operation. This delay is clearly insufficient for the rapid response required for millisecond-level frequency drops caused by renewable energy disconnections, potentially causing the system frequency to drop to a dangerous level before the load-shedding operation is initiated.

[0007] 4. Single criterion protection is prone to false operation: Traditional low-frequency load shedding devices often only rely on a single frequency criterion when determining whether load shedding operations are necessary, without fully considering the impact of factors such as voltage disturbances on frequency measurement. They are prone to false operation under the illusion of frequency fluctuations caused by voltage disturbances, resulting in unnecessary load shedding and affecting the economy and power supply reliability of the power system. Summary of the Invention

[0008] The purpose of the present invention is to provide a low-frequency load reduction method and system based on real-time frequency tracking of zero-crossing points, which effectively solves the technical bottlenecks of traditional low-frequency load reduction devices, significantly improves the performance and reliability of power system frequency stability control, and thus better adapts to the power system frequency stability control requirements under high-proportion new energy access scenarios, so as to solve at least one of the above-mentioned existing technical problems.

[0009] In a first aspect, the present invention provides a low-frequency load reduction method based on zero-crossing real-time frequency tracking, the method specifically comprising:

[0010] The high-voltage signal of the power grid is converted into a low-voltage signal through a voltage transformer, a standard square wave signal is generated through a zero-crossing comparison circuit, and a high-speed comparator is used to build a hysteresis comparison circuit to eliminate signal jitter and form a sensing signal;

[0011] According to the power frequency characteristics of the sensor signal, a dynamic time window is set to eliminate abnormal zero-crossing points that exceed the dynamic time window;

[0012] The sliding discrete Fourier transform is used to calculate the total harmonic distortion rate of the sensor signal in real time, and the activation of the dynamic weighted average algorithm is determined by comparing the total harmonic distortion rate of the signal with the preset distortion rate threshold;

[0013] Matching the pre-stored harmonic feature library, using the pre-trained deep generative adversarial network to establish the mapping relationship between harmonic patterns and phase offsets, and adaptively compensate for the pseudo zero-crossing points of the sensor signal;

[0014] The sensor signals, PMU phasor data, and meteorological parameters are input into a three-dimensional convolutional neural network. The cross-modal correlation matrix is calculated through the self-attention mechanism. Combined with the LSTM unit, the future frequency change trajectory is predicted and the optimal load reduction strategy is output.

[0015] Based on the optimal load shedding strategy, combined with the forward acceleration criterion and reverse fault blocking, emergency load shedding is performed according to the load importance classification, and the load shedding amount is dynamically adjusted through the Nash equilibrium model to minimize load loss and frequency overshoot.

[0016] In a second aspect, the present invention provides a low-frequency load reduction system based on zero-crossing real-time frequency tracking, the system specifically comprising:

[0017] The signal processing module is used to convert the high-voltage signal of the power grid into a low-voltage signal through a voltage transformer, generate a standard square wave signal through a zero-crossing comparison circuit, and use a high-speed comparator to build a hysteresis comparison circuit to eliminate signal jitter to form a sensing signal;

[0018] The primary filtering module is used to set a dynamic time window according to the power frequency characteristics of the sensor signal and eliminate abnormal zero-crossing points that exceed the dynamic time window;

[0019] The secondary verification module is used to calculate the total harmonic distortion rate of the sensor signal in real time using a sliding discrete Fourier transform, and to determine whether to activate the dynamic weighted average algorithm by comparing the total harmonic distortion rate of the signal with a preset distortion rate threshold;

[0020] The three-stage compensation module is used to match the pre-stored harmonic feature library, establish a mapping relationship between harmonic patterns and phase offsets using a pre-trained deep generative adversarial network, and perform adaptive phase compensation for the pseudo zero-crossing points of the sensor signal;

[0021] The load shedding prediction module is used to input sensor signals, PMU phasor data, and meteorological parameters into a three-dimensional convolutional neural network. It calculates the cross-modal association matrix through a self-attention mechanism, combines it with LSTM units to predict future frequency change trajectories, and outputs the optimal load shedding strategy.

[0022] The composite criterion module is used to perform emergency load shedding based on the optimal load shedding strategy, combined with the forward acceleration criterion and reverse fault blocking, according to the load importance classification, and dynamically adjust the load shedding amount through the Nash equilibrium model to minimize load loss and frequency overshoot.

[0023] In a third aspect, the present invention provides a computer device comprising: a memory and a processor and a computer program stored in the memory. When the computer program is executed on the processor, the low-frequency load reduction method based on zero-crossing real-time frequency tracking as described in any one of the above methods is implemented.

[0024] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the low-frequency load reduction method based on zero-crossing real-time frequency tracking as described in any one of the above methods is implemented.

[0025] Compared with the prior art, the present invention has at least one of the following technical effects:

[0026] 1. The present invention effectively solves the technical bottlenecks of traditional low-frequency load reduction devices, significantly improves the performance and reliability of power system frequency stability control, and thus better adapts to the power system frequency stability control needs in scenarios with a high proportion of new energy access.

[0027] 2. The present invention effectively improves the response speed and accuracy of low-frequency load shedding by real-time tracking of the zero-crossing frequency, realizes the optimization of intelligent load shedding strategy under multimodal data fusion, and significantly enhances the frequency stability of the power system.

[0028] 3. The dynamic time window setting and abnormal zero-crossing point elimination technology of the present invention effectively filters out measurement errors and noise interference, ensures the accuracy and stability of frequency measurement, and provides a reliable data basis for subsequent processing.

[0029] 4. The sliding discrete Fourier transform combined with the dynamic weighted average algorithm of the present invention realizes real-time monitoring and adaptive adjustment of the harmonic distortion rate, effectively suppresses the influence of harmonic interference on frequency measurement, and improves measurement accuracy.

[0030] 5. The present invention realizes intelligent mapping of harmonic patterns and phase offsets by applying deep generative adversarial networks, and adaptive phase compensation for pseudo zero crossings, further improving the accuracy and reliability of frequency measurement.

[0031] 6. By combining a three-dimensional convolutional neural network with an LSTM unit, the present invention achieves deep fusion of multimodal data and accurate prediction of future frequencies, providing a scientific basis for the formulation of optimal load shedding strategies and enhancing the intelligence and foresight of load shedding decisions.

[0032] 7. The present invention combines the forward acceleration criterion with the reverse fault blocking, and applies the Nash equilibrium model to achieve dynamic optimization of load shedding and minimization of load loss, effectively balancing system stability and economy.

[0033] 8. The present invention constructs the objective function through the Nash equilibrium model, fully considers the load weight and frequency stability requirements, realizes the optimization of load reduction decision-making, and provides strong support for the frequency stability control of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0035] Figure 1 1 is a flow chart of a low-frequency load reduction method based on zero-crossing real-time frequency tracking provided by one embodiment of the present invention;

[0036] Figure 2 1 is a structural diagram of a low-frequency load reduction system based on zero-crossing real-time frequency tracking provided by one embodiment of the present invention;

[0037] Figure 3 It is a structural diagram of a computer device provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0038] In the following description, specific details such as specific system structures and techniques are provided for purposes of illustration rather than limitation to facilitate a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid obscuring the description of the present application with unnecessary detail.

[0039] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or collections thereof.

[0040] It will also be understood that the term "and / or" used in this specification and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0041] As used in this specification and the appended claims, the term "if" can be interpreted as "when" or "upon" or "in response to determining" or "in response to detecting," depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "upon determination" or "in response to determining" or "upon detection of [described condition or event]" or "in response to detecting [described condition or event]," depending on the context.

[0042] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0043] References to "one embodiment" or "some embodiments" in this specification mean that a particular feature, structure, or characteristic described in conjunction with that embodiment is included in one or more embodiments of the present application. Thus, phrases such as "in one embodiment," "in some embodiments," "in other embodiments," and "in other embodiments" appearing in various places in this specification do not necessarily refer to the same embodiment, but rather mean "one or more but not all embodiments," unless otherwise specifically emphasized. The terms "including," "comprising," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0044] In the embodiments of the present application, the execution subject of the process includes a terminal device, which includes but is not limited to: a server, a computer, a smart phone, a tablet computer, and other devices capable of executing the method disclosed in the present application. Figure 1 A flow chart of a low-frequency load reduction method based on zero-crossing real-time frequency tracking disclosed in the first embodiment of the present invention is shown, and is described in detail as follows:

[0045] S101 converts the high-voltage signal of the power grid into a low-voltage signal through a voltage transformer, generates a standard square wave signal through a zero-crossing comparison circuit, and uses a high-speed comparator to build a hysteresis comparison circuit to eliminate signal jitter and form a sensing signal.

[0046] In this embodiment, a high-precision voltage transformer (such as the XX type) with a rated transformation ratio of 10kV / 0.1kV is used to convert the grid's high-voltage (10kV) signal to a low-voltage (0.1kV) signal. The primary and secondary sides of the voltage transformer are insulated with epoxy resin and have a withstand voltage rating of ≥30kV, ensuring secondary side safety in the event of a grid fault. The transformation ratio error is ≤0.2%, and the angular error is ≤10°, ensuring that the low-voltage signal accurately reflects the grid voltage amplitude and phase.

[0047] The high-speed comparator input is connected to the low-voltage signal (0.1 kV) from the secondary side of the voltage transformer. The signal amplitude is adjusted to the ±5 V input range of the high-speed comparator (such as the LM319) using voltage divider resistors (R1 = 10 kΩ, R2 = 1 kΩ). The high-speed comparator output is connected to an external interrupt pin of a microcontroller (such as the STM32F4 series) to capture zero-crossing events. When the input voltage changes from negative to positive, the comparator output transitions from low to high, triggering a microcontroller interrupt and recording a timestamp of T1. When the input voltage changes from positive to negative, the comparator output transitions from high to low, triggering an interrupt and recording a timestamp of T2. The time difference between two consecutive interrupts, ΔT = T2 - T1, is used to calculate the current signal period, T = 2ΔT (assuming sinusoidal symmetry).

[0048] A positive feedback resistor (Rf = 100kΩ) is added to the high-speed comparator input to create a hysteresis characteristic. The hysteresis voltage threshold is set to ±100mV to suppress glitches in the input signal (such as ±50mV noise). When the input voltage rises above the positive threshold (such as +100mV), the comparator outputs a high level; when the input voltage drops below the negative threshold (such as -100mV), the comparator outputs a low level. The hysteresis characteristic ensures that the comparator output remains stable when the input signal fluctuates near the threshold, eliminating signal jitter.

[0049] The microcontroller records the zero-crossing timestamps at a 100kHz sampling rate, forming the raw sensor signal sequence {T1, T2, T3, ...}. The signal period sequence {T1-T0, T2-T1, T3-T2, ...} is calculated by the difference between adjacent timestamps and converted into a frequency sequence {f1, f2, f3, ...}.

[0050] In this embodiment, through the coordinated design of the voltage transformer, the zero-crossing comparison circuit and the hysteresis comparison circuit, high-precision, interference-resistant sensor signal generation is achieved, providing a reliable data basis for subsequent frequency calculation and load reduction decision-making, and effectively solving the technical bottleneck of traditional low-frequency load reduction devices in scenarios with a high proportion of new energy.

[0051] S102: setting a dynamic time window according to the power frequency characteristics of the sensing signal, and eliminating abnormal zero-crossing points that exceed the dynamic time window.

[0052] In this embodiment, the power frequency (e.g., 50 Hz or 60 Hz) is the core characteristic of power grid signals. However, the actual collected signals often contain harmonics (e.g., third and fifth harmonics) or noise, which can result in false zero crossings and affect the frequency measurement accuracy of traditional zero-crossing methods. Prior art fixed time window or simple threshold methods cannot adapt to the dynamic changes in power frequency and may misjudge or miss valid zero crossings. Therefore, the present invention eliminates abnormal zero crossings by adjusting the time window range in real time, ensuring the robustness of frequency calculation.

[0053] Specifically, a sliding window method or zero-crossing method is used to perform preliminary analysis on the collected signal to determine the current signal's power frequency range. For example, if the signal's main frequency is 50 Hz, the theoretical time interval between adjacent zero crossings is 20 ms (1 / 50 Hz). Considering that the actual power frequency may vary due to grid fluctuations (e.g., 45 Hz to 55 Hz), a dynamic time window center value and tolerance range are set: the center value calculates the theoretical zero crossing time interval based on the current estimated power frequency; the tolerance range sets a time window width based on the power frequency fluctuation range, for example, a ±10% tolerance. Assuming the current estimated power frequency is 50 Hz, the theoretical zero crossing time interval is 20 ms. The time window range is set to 18 ms to 22 ms (with a ±10% tolerance). If a zero crossing time interval is not within the 18 ms to 22 ms range, it is identified as an abnormal zero crossing and discarded.

[0054] Preprocess the collected signal using a bandpass filter (e.g., 40Hz-60Hz) to suppress high-frequency noise and harmonics. Record the time points at which the signal crosses zero from negative to positive (or vice versa) to form a zero-crossing time series. After each zero-crossing is detected, dynamically update the time window based on the current signal power frequency estimate. For example, if the power frequency estimate changes to 52Hz, the theoretical zero-crossing interval becomes 19.23ms, and the time window is updated to 17.31ms-21.15ms. Check the time intervals between adjacent zero-crossings one by one. If a time interval does not fall within the current time window, remove the corresponding zero-crossing. Only zero-crossings that pass the time window check are retained to form a valid zero-crossing time series. Calculate the signal frequency based on the valid zero-crossing time series. For example, if the valid zero-crossing interval is 20ms, the signal frequency is 50Hz.

[0055] In this embodiment, the dynamic time window method effectively eliminates abnormal zero-crossing points by adjusting the time window range in real time, thereby improving the accuracy of frequency calculation, especially in scenarios of harmonic interference and frequency mutation.

[0056] S103 , using sliding discrete Fourier transform to calculate the total harmonic distortion rate of the sensor signal in real time, and determining whether to activate the dynamic weighted average algorithm by comparing the total harmonic distortion rate of the signal with a preset distortion rate threshold.

[0057] In this embodiment, harmonic interference and signal fluctuations in power system and industrial sensor signals can increase errors in traditional signal analysis methods (such as fixed-window Fourier transforms). For example, fixed-window methods struggle to track dynamic signal changes in real time, resulting in significant deviations between THD (total harmonic distortion) calculation results and actual values. Noise or transient interference can also generate false signal signatures, affecting subsequent signal analysis (such as frequency calculation and waveform recognition). To address these issues, the present invention improves the accuracy and robustness of signal processing by calculating THD in real time and adaptively adjusting the signal analysis strategy.

[0058] Specifically, the continuously acquired sensor signal is segmented into segments of fixed length (e.g., N sampling points), each segment being called an "analysis window." After each analysis is completed, the window slides forward one sampling point to form a new analysis window, enabling real-time updating of the spectrum. A Fourier transform is performed on each analysis window to extract the amplitudes of the fundamental wave (e.g., 50 Hz) and each harmonic (e.g., 100 Hz, 150 Hz, etc.). Using the fundamental wave amplitude as a reference, the relative ratios of the amplitudes of each harmonic to the fundamental wave amplitude are calculated, providing basic data for THD calculation. Based on the fundamental wave amplitude and each harmonic amplitude, THD is calculated using the following logical steps: The square sum of all harmonic amplitudes is calculated; this square sum is divided by the square of the fundamental wave amplitude to obtain a dimensionless value for the distortion ratio; and this dimensionless value is converted to a percentage, which is then output as the real-time THD.

[0059] Set a preset THD threshold (e.g., 5%) based on the application scenario. For example, in power systems, 5% THD is generally considered the acceptable limit for power quality. The real-time calculated THD is compared with the preset threshold. If the THD exceeds the threshold, the signal is determined to have significant harmonic interference. When the THD exceeds the threshold, a dynamic weighted averaging algorithm is activated to weight subsequent signal analysis results (e.g., frequency and amplitude). Weights are assigned based on the reliability of the signal. For example, data from periods with lower THD are given a higher weight, while data from periods with higher THD are given a lower weight.

[0060] Maintain a fixed-length data cache queue (e.g., data from the most recent 10 analysis windows) to store historical signal analysis results. Calculate a corresponding weight based on the THD value of each analysis window. For example, the weight can be inversely proportional to the THD value (lower THD, higher weight). Perform a weighted average of the data in the cache queue based on the weights. For example, if the weight of one window is 0.8 and the weight of another window is 0.2, the final result is the weighted sum of the data from both windows. This weighted average is used as the final analysis result for the current signal and is used for subsequent signal processing or decision-making.

[0061] In this embodiment, the real-time THD calculation and dynamic adjustment of the signal analysis strategy through SDFT significantly improve the accuracy and stability of signal processing, especially in harmonic interference and noise scenarios.

[0062] S104, matching the pre-stored harmonic feature library, using the pre-trained deep generative adversarial network to establish a mapping relationship between the harmonic pattern and the phase offset, and performing adaptive phase compensation on the pseudo zero-crossing point of the sensor signal.

[0063] In this embodiment, during sensor signal processing, harmonic interference can cause false zero crossings (i.e., zero crossings that are not at the true zero position), thereby affecting the accuracy of key parameters such as frequency calculation and phase measurement. Traditional methods typically rely on fixed thresholds or empirical formulas for phase compensation, but these methods suffer from the following drawbacks: harmonic patterns vary significantly across scenarios, making fixed methods incapable of covering all cases; noise or transient interference can produce false zero crossings, leading to biased compensation results; and traditional methods are computationally complex, making it difficult to meet real-time requirements. To address these issues, the present invention achieves precise correction of sensor signals by matching a harmonic feature library, training a GAN model, and adaptively compensating for phase offset.

[0064] Specifically, sensor signal samples are collected from different scenarios (such as power system harmonic signals and industrial sensor output signals). These samples must cover typical harmonic patterns (such as third and fifth harmonics) and their phase offset ranges. Time and frequency domain analysis is performed on the sample signals to extract key features (such as harmonic amplitude, frequency, and phase). These features are categorized and stored by harmonic pattern to form a pre-stored harmonic feature library. The feature library is regularly updated based on newly collected signal samples to ensure coverage of the latest harmonic patterns. Clustering algorithms (such as K-means) are used to merge features to reduce redundant data. A feature index is established to improve feature matching efficiency.

[0065] Harmonic patterns (input data) and their corresponding phase offsets (labeled data) are extracted from a pre-stored harmonic feature library to construct a training set. The data is stratified and sampled based on dimensions such as harmonic order and amplitude range to ensure representativeness. Data augmentation (such as adding noise and random phase offsets) is performed on the training set to improve the generalization ability of the GAN (Generative Adversarial Network). The GAN model consists of a generator and a discriminator. The generator takes harmonic pattern features (such as amplitude and frequency) as input and outputs predicted phase offsets. The discriminator takes the combination of harmonic pattern features and phase offsets as input and outputs a plausibility score (0–1) for the combined data. The generator and discriminator are trained alternately, with the generator attempting to generate realistic phase offsets to deceive the discriminator, while the discriminator strives to distinguish between real and generated data. The training objective is to minimize the loss function (such as cross-entropy loss) for both the generator and discriminator. The GAN model is considered converged when the discriminator cannot distinguish between real and generated data.

[0066] Perform zero-crossing detection on the real-time sensor signal and mark all zero-crossing locations. Combined with the signal's harmonic pattern characteristics, determine whether the zero-crossing is a pseudo-zero-crossing (e.g., by matching it with a pre-stored feature library). Extract the current signal's harmonic pattern characteristics (e.g., amplitude, frequency); search the pre-stored harmonic feature library for the most similar feature pattern. Calculate the similarity (e.g., cosine similarity) between the current feature and the features in the feature library. If the similarity exceeds a preset threshold, the match is considered successful. Input the matched harmonic pattern features into a trained GAN generator to predict the corresponding phase offset; the generator outputs an adaptive phase compensation value. Perform phase correction on the pseudo-zero-crossing based on the predicted phase offset; the corrected signal zero-crossing position is closer to the true zero point, improving the accuracy of signal analysis.

[0067] In this embodiment, through the combined application of a pre-stored harmonic feature library and GAN, adaptive mapping of harmonic patterns and phase offsets is achieved, effectively eliminating the pseudo zero-crossing error of the sensor signal and improving the accuracy and robustness of signal processing.

[0068] In step S105, the sensor signals, PMU phasor data, and meteorological parameters are input into a three-dimensional convolutional neural network. The cross-modal correlation matrix is calculated through the self-attention mechanism. Combined with the LSTM unit, the future frequency change trajectory is predicted and the optimal load reduction strategy is output.

[0069] In this embodiment, power system frequency stability is a key indicator for ensuring safe grid operation. Traditional frequency prediction methods have the following shortcomings: sensor signals (such as current and voltage waveforms), PMU phasor data (such as amplitude and phase), and meteorological parameters (such as wind speed and light intensity) are typically analyzed independently, making it difficult to explore correlations between multimodal data; existing methods have limited ability to model the temporal dependencies of frequency changes, resulting in insufficient prediction accuracy; and load shedding strategies based on static thresholds cannot adapt to the dynamically changing grid operating conditions and are prone to overload or underload. To address these issues, the present invention achieves accurate prediction and stable control of power system frequency by fusing multimodal data, exploring temporal correlations, and dynamically optimizing load shedding strategies.

[0070] Specifically, the high-precision clock synchronization function of the PMU is used to obtain synchronized phasor data (such as voltage amplitude and phase angle) from grid nodes. This phasor data is then processed using a sliding window (window length 10 seconds, step size 1 second) to generate a time series feature sequence. The phasor data is converted into a two-dimensional feature map (such as an amplitude-phase heat map) as another input to the 3D-CNN. Meteorological parameters (such as wind speed, light intensity, and temperature) for the grid coverage area are obtained from a weather station or weather forecast platform. These parameters are normalized to eliminate dimensional differences. These parameters are then mapped into a two-dimensional feature map (such as a wind speed-time curve) as the third input to the 3D-CNN (three-dimensional convolutional neural network).

[0071] The input layer of the 3D-CNN receives time-frequency maps of sensor signals, PMU phasor feature maps, and meteorological parameter feature maps, generating three-dimensional input data (temporal, spatial, and feature dimensions). The convolutional layer uses a three-dimensional convolution kernel (e.g., 3×3×3) to extract the spatial-temporal features of multimodal data. Multiple convolutional layers are used to gradually expand the receptive field and capture global features. The pooling layer uses max pooling or average pooling to reduce feature dimensionality and computational complexity. A self-attention mechanism is applied to the 3D-CNN output feature maps to calculate the correlation weight matrix between the different modal data. The correlation weight matrix represents the contribution of each modal data to frequency prediction. Based on the correlation weight matrix, the multimodal features are weightedly fused to generate a fused feature vector.

[0072] The input layer of the LSTM unit receives the fused feature vector output by the 3D-CNN, forming a time series input sequence. The time series layer uses multiple layers of LSTM units to capture the temporal dependencies of frequency changes. Forget gates, input gates, and output gates are set to control the retention and updating of historical information. The output layer outputs the frequency change trajectory over a period of time (e.g., the frequency forecast for the next 10 minutes). The predicted frequency trajectory is compared with the actual frequency data to calculate the prediction error (e.g., mean squared error, mean absolute error). Based on the prediction error, the parameters of the LSTM unit (e.g., learning rate, number of hidden layer nodes) are dynamically adjusted to optimize prediction performance.

[0073] Define a load shedding strategy space, including parameters such as load shedding amount and priority for different nodes. Generate a set of candidate load shedding strategies based on the grid topology and load characteristics. Set evaluation metrics (such as frequency stability, load loss, and economic cost) to quantify the pros and cons of each strategy. Use the LSTM-predicted frequency trajectory as a constraint to select load shedding strategies that meet frequency stability requirements. Use multi-objective optimization algorithms (such as genetic algorithms and particle swarm optimization) to search for the optimal solution in the strategy space. Dynamically adjust the load shedding strategy based on real-time frequency changes to ensure grid frequency stability.

[0074] In this embodiment, by fusing multimodal data, mining cross-modal associations and dynamically optimizing load reduction strategies, accurate prediction and stable control of the power system frequency are achieved, significantly improving the anti-interference ability and operating efficiency of the power grid.

[0075] S106, based on the optimal load reduction strategy, combined with the forward acceleration criterion and reverse fault blocking, performs emergency load reduction according to the load importance classification, and dynamically adjusts the load reduction amount through the Nash equilibrium model to minimize load loss and frequency overshoot.

[0076] In this embodiment, conventional power system emergency load shedding methods have the following shortcomings: fixed-threshold-based load shedding strategies cannot adapt to the dynamic state of the grid, easily leading to overload or underload; they fail to consider load priorities (e.g., industrial loads, residential loads, and critical infrastructure loads), potentially causing unnecessary socioeconomic losses; the frequency may fluctuate significantly during load shedding, impacting grid stability; and the load shedding amount is typically statically set and cannot be dynamically adjusted based on real-time grid conditions. To address these issues, the present invention uses a positive acceleration criterion to preemptively trigger load shedding and a reverse fault lockout to prevent misoperation. Combining load importance grading with a Nash equilibrium model, this method dynamically optimizes the load shedding amount, minimizing load losses and frequency overshoot.

[0077] Specifically, the grid frequency change rate (df / dt) is monitored in real time. When the frequency change rate exceeds the positive acceleration threshold (e.g., df / dt > 0.5 Hz / s), an emergency load shedding process is triggered. Using the positive acceleration criterion, the load shedding process is initiated early to prevent further frequency deterioration and buy time for subsequent load shedding operations. During the load shedding process, the grid status is monitored in real time. If a reverse fault (e.g., a short circuit or malfunctioning switch) is detected, the load shedding operation is immediately blocked to prevent accidental load shedding. The reverse fault blocking mechanism compares the current grid status with the pre-fault state to determine if there are any abnormal changes.

[0078] Loads are categorized into three groups based on their socioeconomic impact and tolerance for power outages: Category I loads: critical infrastructure (such as hospitals and transportation hubs) and important industrial loads; Category II loads: general industrial and commercial loads; and Category III loads: residential and non-critical agricultural loads. Category I loads have the highest priority, while Category III loads have the lowest. During load shedding, low-priority loads are prioritized, while high-priority loads are retained. Load shedding is carried out in ascending order of priority. After each round of load shedding, the grid frequency is reassessed. If the frequency has not stabilized, the next round of load shedding is continued.

[0079] Each load node in the power grid is defined as a game agent. Each agent chooses a load reduction based on its own interests (e.g., load loss and frequency stability). The goal of the game is to minimize the total losses of all agents while maintaining grid frequency stability. Each agent's strategy space consists of a set of optional load reductions (e.g., 0%, 10%, 20%). Through multiple rounds of the game, each agent gradually adjusts its strategy until a Nash equilibrium is reached. During the load reduction process, the grid frequency and load status are monitored in real time, and feedback is fed into the Nash equilibrium model. Based on this feedback, the load reduction of each agent is dynamically adjusted to ensure grid frequency stability. When all agents' strategies no longer change, or the magnitude of the change is less than a set threshold, the Nash equilibrium is considered reached and adjustments cease.

[0080] In some embodiments, in the above step S102, setting a dynamic time window according to the power frequency characteristics of the sensor signal and eliminating abnormal zero-crossing points that exceed the dynamic time window specifically includes:

[0081] An initial time window is calculated based on the allowable range of the power frequency of the sensing signal, and a dynamic time window range is set in combination with a time margin for compensating for measurement errors. The time margin is dynamically adjusted by the real-time frequency change rate.

[0082] Collect any two timestamps of adjacent zero-crossing points and calculate the time interval between the two timestamps. If the time interval exceeds the dynamic time window range, mark the zero-crossing point as an outlier and remove it.

[0083] Use a sliding window to store historical period data, calculate the weighted average period through the exponential decay weight formula, and adjust the dynamic time window range according to the weighted average period;

[0084] If the continuous anomaly manifests as a rapid frequency drop, emergency control is triggered; otherwise, the zero-crossing sequence is reconstructed using weighted average periodic interpolation.

[0085] In this embodiment, the zero crossings of sensor signals are used for frequency calculation and phase synchronization in power system monitoring. However, due to harmonics, noise, or transient faults, these zero crossings may exhibit abnormal offsets, leading to frequency measurement errors. Traditional fixed time window methods struggle to adapt to rapid frequency fluctuations, necessitating a dynamic adjustment of the time window to achieve high-precision rejection.

[0086] Specifically, based on power grid standards (e.g., 50 Hz ± 0.5 Hz), the power frequency range is determined to be 49.5 Hz to 50.5 Hz. 50 Hz corresponds to a period of 20 ms, 49.5 Hz to approximately 20.2 ms, and 50.5 Hz to approximately 19.8 ms. The initial time window is set to 19.8 ms to 20.2 ms to cover the power frequency fluctuation range.

[0087] To compensate for measurement errors, the initial time window is extended by 0.1ms at both ends (for example, 19.7ms to 20.3ms). The frequency change rate is monitored in real time (for example, once per second). If the frequency change rate exceeds a threshold (for example, ±0.1 Hz / s), the time margin is dynamically increased: as the frequency increases, the lower limit of the time window is shortened (for example, by adding a 0.02ms margin for every 0.1 Hz / s). As the frequency decreases, the upper limit of the time window is extended (for example, by adding a 0.02ms margin for every 0.1 Hz / s).

[0088] Record zero-crossing timestamps at a fixed sampling rate (e.g. 10kHz). and , calculate the time interval .like If the current dynamic time window is exceeded, it will be marked as abnormal. For example, if the current time window is 19.8ms~20.2ms, =21ms, it is judged as abnormal and rejected. If there are three consecutive zero-crossing abnormalities, the emergency control process will be entered.

[0089] The most recent 50 valid zero-crossing periods are stored. New data is weighted more heavily (e.g., 0.9) and older data is weighted less heavily (e.g., 0.1). A weighted average period is calculated using an exponential decay mechanism. Assuming the historical periods are 19.9ms, 20.0ms, and 20.1ms (with decreasing weights), the weighted average period is 20.0ms. Based on the weighted average period, the time window range is adjusted (e.g., ±0.2ms), resulting in a new time window of 19.8ms to 20.2ms.

[0090] If the interval between five consecutive zero-crossings shows a decreasing trend (e.g., from 20ms to 18ms), a rapid frequency drop is detected. An alarm is triggered, the current time window is frozen, and dynamic adjustments are suspended. The higher-level system is notified to take protective measures (such as load shedding). For any abnormal zero-crossings that are removed, the weighted average of the preceding and following valid points is used for interpolation. For example, if the 10th point is abnormal, the linear interpolation of the 9th and 11th points is used instead.

[0091] In this embodiment, the time window can be adjusted in real time with frequency fluctuations, adapting to the ±0.5Hz operating frequency range. Harmonic interference is suppressed through weighted averaging, improving zero-crossing detection accuracy. An emergency control mechanism can handle extreme operating conditions such as frequency drops, ensuring system safety.

[0092] In some embodiments, in step S103, the method of calculating the total harmonic distortion rate of the sensor signal in real time by using a sliding discrete Fourier transform and determining whether to activate the dynamic weighted average algorithm by comparing the total harmonic distortion rate of the signal with a preset distortion rate threshold may specifically include:

[0093] The sensor signal is sampled according to a preset sampling frequency, and the kth harmonic component is calculated by a recursive formula, wherein the kth harmonic component includes a fundamental component and odd harmonic components;

[0094] Based on the kth harmonic component, the sliding discrete Fourier transform is used to calculate the total harmonic distortion rate of the signal in the current window;

[0095] Compare the total harmonic distortion rate of the signal with the preset distortion rate threshold. If the total harmonic distortion rate of the signal is greater than the preset distortion rate threshold, adaptively allocate weights according to harmonic energy;

[0096] Based on the weights after adaptive allocation, the fundamental components are weighted averaged to obtain weighted fundamental components, the amplitudes of the weighted fundamental components are used to reconstruct the zero-crossing sequence, and the corrected power frequency is output.

[0097] In this embodiment, harmonic contamination in sensor signals during power system monitoring can lead to increased total harmonic distortion (THD), which in turn affects frequency calculation and phase synchronization accuracy. Traditional fixed threshold methods struggle to adapt to dynamic harmonic changes. Therefore, high-precision signal processing requires real-time THD monitoring and adaptive filtering strategies.

[0098] Specifically, based on the power signal frequency (e.g., 50 Hz), the sampling frequency is set to an integer multiple of the power frequency (e.g., 10 kHz, corresponding to 200 points / cycle). A ring buffer is used to store the latest N sampling points (e.g., N = 256) to ensure real-time performance. A recursive algorithm (e.g., a simplified version of the Goertzel algorithm) is used to calculate the fundamental (1st harmonic) and odd harmonics (3rd, 5th, 7th, 9th, 11th, etc.) point by point.

[0099] Select a window length relative to the power frequency period. With each new sampling point, the window slides forward one point for real-time updates. Calculate the sum of the squares of all harmonic components (fundamental and odd harmonics) within the window and use this sum to calculate the THD. Set a THD threshold (e.g., 5%) based on system requirements. If the current window THD exceeds the threshold, activate the dynamic weighted averaging algorithm; otherwise, maintain the normal filtering strategy. Assign weights to each harmonic component based on their energy contribution, weighting the fundamental component.

[0100] The zero-crossing sequence is reconstructed based on the phase information of the weighted fundamental component. The reciprocal of the time interval between adjacent zero-crossing points is calculated based on the reconstructed zero-crossing sequence to obtain the corrected power frequency.

[0101] If the THD values for M consecutive windows (e.g., M=5) exceed the threshold, emergency control is triggered. The current filter parameters are frozen, the backup filter is activated, and the higher-level system is notified to take protective measures (such as load shedding). Based on historical THD data, the threshold and window length are dynamically optimized. The harmonic model is regularly updated to improve the accuracy of weight allocation.

[0102] In this embodiment, a sliding window mechanism ensures that THD calculations are updated synchronously with the weighted averaging algorithm. Dynamic weight allocation adapts to changes in harmonic energy, improving the accuracy of fundamental component extraction. An emergency control mechanism can address persistent harmonic pollution and ensure stable system operation.

[0103] In some embodiments, in the above step S104, the matching of the pre-stored harmonic feature library, the use of a pre-trained deep generative adversarial network to establish a mapping relationship between the harmonic pattern and the phase offset, and the adaptive phase compensation of the pseudo zero-crossing point of the sensor signal specifically include:

[0104] The amplitude distribution of each odd harmonic component is statistically analyzed through historical fault data, and the initial harmonic phase angle and harmonic frequency of each odd harmonic component are collected to establish a pre-stored typical harmonic pattern parameter set and form a harmonic feature library;

[0105] Construct a deep generative adversarial network, based on the harmonic feature library, and use the adversarial loss function to train the generator and recognizer of the deep generative adversarial network;

[0106] The current harmonic energy ratio is calculated. If the current harmonic energy ratio is greater than the preset energy ratio threshold, it is determined that there is a pseudo zero-crossing point. The current harmonic parameters are input into the deep generative adversarial network, and the phase compensation amount is output.

[0107] In this embodiment, in the power system, odd harmonics can cause phase shifts at the zero crossings of the sensor signal, forming pseudo-zero crossings, which in turn affects the accuracy of frequency calculations. Traditional methods rely on fixed compensation parameters and are difficult to adapt to the dynamic changes in harmonic patterns. The present invention achieves dynamic mapping between harmonic patterns and phase offsets by constructing a harmonic feature library and a deep generative adversarial network, thereby improving the accuracy of pseudo-zero-crossing compensation.

[0108] Specifically, harmonic data from historical grid fault records (such as transformer switching and capacitor bank connection) is collected. The amplitude distribution range of each odd-order harmonic component (3rd, 5th, 7th, etc.) is calculated to form an amplitude probability density map. For example, the amplitude distribution of the 3rd harmonic is 0.1% to 10% of the fundamental amplitude; the amplitude distribution of the 5th harmonic is 0.05% to 5% of the fundamental amplitude.

[0109] Record the initial phase angle of each odd harmonic (the phase difference relative to the fundamental) to create a phase angle statistics table. For example, the phase angle range for the third harmonic is -30° to 30°, and the phase angle range for the fifth harmonic is -15° to 15°. Extract the deviation of the harmonic frequency from the power frequency (e.g., ±0.5 Hz) to create a frequency offset distribution chart.

[0110] Combine parameters such as amplitude, phase angle, and frequency into typical harmonic patterns to form parameter sets. For example, Pattern 1: 3rd harmonic amplitude 5%, phase angle 10°, frequency offset 0.2Hz; Pattern 2: 5th harmonic amplitude 2%, phase angle -5°, frequency offset 0.1Hz. Store parameter sets in a database or file to create a pre-stored harmonic signature library.

[0111] A deep generative adversarial network (GAN) was constructed, consisting of a generator and a discriminator. The generator's input layer receives harmonic parameters (amplitude, phase angle, and frequency) as input features, and the hidden layer uses a multi-layer fully connected network (e.g., 3 layers, 128 neurons per layer) with a Reluctant Unit (ReLU) activation function. The output layer outputs a phase compensation value (e.g., -15° to 15°). The discriminator's input layer receives a combined vector of harmonic parameters and phase compensation values, and the hidden layer uses a convolutional neural network (CNN) to extract features. This is followed by a fully connected layer for binary classification (real / fake), and the output layer outputs a probability value (0 to 1).

[0112] The generator and discriminator are trained using an adversarial loss function. The generator's training goal is to make it impossible for the discriminator to distinguish between the generated phase compensation and the real compensation. The discriminator's training goal is to accurately distinguish between the real phase compensation and the forged phase compensation generated by the generator. The parameters in the harmonic feature library are randomly perturbed (e.g., amplitude ±5%, phase angle ±3°) to generate training samples. The real samples in the training samples are the phase compensation corresponding to the actual harmonic data, and the forged samples are the phase compensation generated by the generator.

[0113] The ratio of the total harmonic energy to the fundamental energy is calculated. If the energy ratio exceeds a preset threshold (e.g., 10%), a false zero crossing is detected. The amplitude, phase angle, and frequency parameters of the current harmonic are fed into a trained deep generative adversarial network. The generator outputs the corresponding phase compensation (e.g., -8°). The phase angle of the harmonic component is adjusted based on the compensation. Based on the adjusted harmonic component, the zero crossing timestamp is recalculated to correct the false zero crossing.

[0114] Regularly add newly collected harmonic data and compensation results to the training set to perform incremental training on the generative adversarial network. Update the weight parameters of the generator and discriminator at regular intervals to improve model adaptability. Compare the frequency calculation error before and after compensation to evaluate the effectiveness of phase compensation. If the error exceeds the threshold after multiple consecutive compensations, an abnormality alarm is triggered, notifying operations and maintenance personnel.

[0115] In this embodiment, a deep generative adversarial network (GAN) achieves precise compensation for pseudo-zero crossings by learning the mapping between harmonic patterns and phase offsets. This online learning mechanism enables the system to adapt to dynamic changes in harmonic patterns, improving long-term operational stability. Harmonic energy ratio calculations are performed simultaneously with phase compensation output, ensuring real-time compensation.

[0116] In some embodiments, in step S105, the sensor signals, PMU phasor data, and meteorological parameters are input into a three-dimensional convolutional neural network, a cross-modal correlation matrix is calculated through a self-attention mechanism, and the future frequency change trajectory is predicted in combination with an LSTM unit to output an optimal load reduction strategy, specifically including:

[0117] Collect PMU phasor data and meteorological parameters, align the sensor signals, PMU phasor data and meteorological parameters in time, and then splice them into a three-dimensional tensor along the channel dimension;

[0118] A three-dimensional convolutional neural network is used to extract features from the three-dimensional tensor, and a high-order feature map is generated through the ReLU activation function;

[0119] Perform linear projection on the high-order feature map to obtain the query matrix, key matrix and value matrix. Based on the query matrix, key matrix and value matrix, calculate the cross-modal association matrix and perform weighted fusion to obtain weighted fusion features.

[0120] The weighted fusion features are input into the LSTM unit, the hidden state is updated through the state equation, and the frequency prediction value within the preset future time period is output.

[0121] In this embodiment, traditional power system frequency prediction methods often rely on single-modal data (such as PMU phasors), making it difficult to capture the impact of multi-source data (such as weather changes and equipment status) on frequency under complex operating conditions. This invention utilizes a three-dimensional convolutional neural network and a self-attention mechanism to perform cross-modal correlation analysis of multimodal data. This, combined with LSTM units, predicts future frequency change trajectories, providing a decision-making basis for dynamic load shedding.

[0122] Specifically, synchronized phasor measurement units (PMUs) acquire phasor data, such as voltage amplitude, phase angle, and frequency, at power grid nodes at a refresh rate of 50 frames per second. Real-time meteorological data, including temperature (range: -20°C to 50°C), humidity (range: 10% to 100%), and wind speed (range: 0 m / s to 30 m / s), is acquired from weather stations. The sensor signals, PMU phasor data, and meteorological parameters are timestamp-aligned to ensure temporal consistency across all modal data. These three types of data are concatenated into a three-dimensional tensor based on the channel dimension, with the structure [time step × number of channels × feature dimension], where the time step covers the past hour of data (e.g., 3600 time points with a step of 1 second), the number of channels is 3 (sensor signals, PMU phasors, and meteorological parameters), and the feature dimension is the sum of the feature counts for each modal data (e.g., two-dimensional sensor signals: voltage and current; six-dimensional PMU phasors: amplitude, phase angle, and frequency; and three-dimensional meteorological parameters: temperature, humidity, and wind speed).

[0123] Three 3D convolutional layers are used, with a kernel size of 3×3×3, a stride of 1, padding of "same," and a Reluctant Unified Unit (ReLU) activation function. Each convolution layer is followed by a batch normalization layer to accelerate training convergence. Every two convolutional layers are followed by a max pooling layer with a pooling window of 2×2×2 to reduce feature dimensionality. After the convolution and pooling operations, a high-order feature map is generated with the structure [timestep / compression ratio × number of channels / compression ratio × feature dimension / compression ratio]. For example, after three convolutional layers on an input tensor of [3600×3×11], the output feature map is [225×1×16] (with a compression ratio of 16).

[0124] Linearly project the high-order feature maps to generate query matrices (Q), key matrices (K), and value matrices (V), all with dimensions [time steps / compression ratio × feature dimension / compression ratio × number of attention heads]. For example, a feature map [225 × 1 × 16] is projected to Q, K, and V [225 × 1 × 8] (with 8 attention heads).

[0125] The dot product of the query matrix and the key matrix is calculated to generate a cross-modal correlation matrix. The correlation matrix has dimensions [time step / compression ratio × time step / compression ratio × number of attention heads], which represents the correlation strength of each modal feature at different time steps. Softmax normalization is performed on the correlation matrix to obtain attention weights, which are multiplied by the value matrix to generate weighted fusion features. The structure of the weighted fusion feature is [time step / compression ratio × feature dimension / compression ratio × number of attention heads].

[0126] The weighted fusion features are input into an LSTM unit with 64 hidden layer neurons, the activation function is Tanh, and the output layer is a fully connected layer. The unit outputs the predicted future frequency value. The hidden state is updated using the LSTM state equation to capture time series dependencies. The system outputs the frequency change trajectory for a preset time period (e.g., 5 minutes, 10 minutes) in the future, with a sampling frequency of 1 second. Based on the predicted frequency change trajectory and combined with grid safety constraints (e.g., a lower frequency limit of 49.5 Hz), the optimal load reduction amount is calculated. The load reduction strategy is output, including the load reduction node, load reduction amount, and execution time. For example, node A: load reduction of 50 MW, executed immediately; node B: load reduction of 30 MW, executed after 30 seconds.

[0127] Regularly add newly collected multimodal data and prediction results to the training set to perform incremental training on the 3D-CNN and LSTM models. Update model parameters at regular intervals to improve prediction accuracy. Compare the predicted frequency with the actual frequency, calculating the mean square error (MSE) and mean absolute error (MAE). Simulate the effectiveness of load shedding strategies in a simulation environment to evaluate frequency recovery speed and stability. If the prediction error exceeds the threshold multiple times in a row, an abnormality alarm is triggered, notifying operations and maintenance personnel.

[0128] In this implementation, a three-dimensional convolutional neural network and a self-attention mechanism are used to effectively capture cross-modal correlations between sensor signals, PMU phasors, and meteorological parameters. An LSTM unit combines historical data with real-time input to accurately predict future frequency changes. Based on these predictions, an optimal load reduction strategy is generated, improving grid frequency stability.

[0129] In some embodiments, in step S106, the optimal load shedding strategy, combined with the forward acceleration criterion and the reverse fault blocking, performs emergency load shedding according to load importance classification, and dynamically adjusts the load shedding amount through the Nash equilibrium model to minimize load loss and frequency overshoot, specifically including:

[0130] Calculate the frequency change rate and voltage change rate in real time, and generate a primary load reduction instruction when the frequency change rate is greater than the frequency change rate threshold and the voltage change rate is less than the voltage change rate threshold;

[0131] The positive sequence impedance mutation rate is calculated using PMU data. If the positive sequence impedance mutation rate is greater than the mutation rate threshold, the primary load reduction instruction is immediately blocked.

[0132] Based on the preset load weight matrix, the objective function of minimizing load loss and frequency overshoot is constructed through the Nash equilibrium model;

[0133] Based on the objective function and the primary load shedding instruction, the optimal load shedding amount of each load is obtained by solving the Lagrange multiplier method;

[0134] Based on the optimal load reduction, each load is removed in stages starting from the high-weight load.

[0135] In this embodiment, traditional grid load shedding strategies often rely on fixed thresholds or single criteria, making them difficult to adapt to the rapid frequency changes and dynamic load characteristics under complex operating conditions. This invention dynamically adjusts load shedding by monitoring the frequency and voltage change rates, as well as the positive-sequence impedance mutation rate, in real time. This combines load importance grading with Nash equilibrium optimization to improve grid frequency stability.

[0136] Specifically, the frequency rate of change (Δf / Δt) and voltage rate of change (ΔU / Δt) are calculated, and data fluctuations are smoothed using a sliding window method (with a window length of 1 second). When the frequency rate of change Δf / Δt exceeds threshold A (e.g., 0.2 Hz / s) and the voltage rate of change ΔU / Δt is less than threshold B (e.g., -5 V / s), a primary load shedding command is generated. For example, if Δf / Δt = 0.3 Hz / s and ΔU / Δt = -3 V / s, primary load shedding is triggered.

[0137] The PMU acquires three-phase voltage and current data and calculates the positive-sequence impedance (Z1 = U1 / I1). The positive-sequence impedance mutation rate (ΔZ1 / Δt) is calculated using the first-order difference method. When the positive-sequence impedance mutation rate ΔZ1 / Δt exceeds threshold C (e.g., 10Ω / s), the primary load shedding command is immediately blocked. For example, if ΔZ1 / Δt = 15Ω / s, primary load shedding is blocked to prevent accidental load shedding.

[0138] Based on the load importance classification (e.g., primary, secondary, and tertiary loads), a preset weight matrix W is set. For example, primary load weight = 1.0 (highest priority); secondary load weight = 0.5; and tertiary load weight = 0.2. The objective function is set to minimize the weighted sum of load loss (L) and frequency overshoot (Δf_max), with the constraint that the total load reduction does not exceed the grid's load shedding capacity. The total load reduction constraint L ≤ L_max is converted into a Lagrangian function. This Lagrangian function is optimized using an iterative algorithm (e.g., gradient descent) to determine the optimal load reduction for each load. Based on the load weight matrix, load shedding is performed in a tiered manner, starting with the highest-weight loads: tertiary loads are prioritized; if the frequency remains unstable, some secondary loads are removed; and primary loads (e.g., hospitals and data centers) are removed only in extreme cases.

[0139] In this embodiment, the Nash equilibrium model and Lagrange multiplier method are used to dynamically adjust load shedding, minimizing load losses. The positive-sequence impedance mutation rate criterion effectively avoids misdirected load shedding and improves strategy reliability. Load importance grading ensures power supply to critical loads and safeguards social stability.

[0140] Furthermore, the objective function of minimizing load loss and frequency overshoot is constructed based on the preset load weight matrix through the Nash equilibrium model, specifically including:

[0141] The load loss term is constructed according to the load priority weight in the preset load weight matrix, and the objective function is formed by combining the frequency overshoot integral term;

[0142] Set power balance constraints and frequency safety constraints, and determine the solution range of the objective function based on the power balance constraints and frequency safety constraints;

[0143] The weight coefficients of the load loss term and the frequency overshoot integral term are adjusted according to the frequency change rate.

[0144] In this embodiment, traditional grid load shedding strategies typically rely on fixed weights or a single objective function to optimize load loss and frequency overshoot, making them difficult to adapt to the rapid frequency changes and dynamic load characteristics under complex operating conditions. This invention achieves multi-objective collaborative optimization and improves grid frequency stability by using a preset load weight matrix, a Nash equilibrium model, and dynamic weight adjustment.

[0145] Specifically, the objective function satisfies

[0146]

[0147]

[0148]

[0149] in, represents the objective function value, Indicates the load reduction. represents the load loss term, represents the frequency overshoot integral term, and Represent the weight coefficients of the load loss term and the frequency overshoot integral term, represents the priority weight of the i-th type of load, n represents the number of loads, represents the removal ratio of the i-th type load, represents the frequency safety constraint, t represents the optimization time window, Indicates the system rated frequency reference value, and Indicates the two endpoints of the optimization time window.

[0150] Specifically, the power balance constraint satisfies

[0151]

[0152] in, represents the reference power of the i-th type load, Indicates the total amount of load reduction required by the system.

[0153] Specifically, the frequency safety constraint satisfies

[0154]

[0155] Constraints ensure that all participants (load nodes) cannot unilaterally improve their own costs under a given strategy.

[0156] Furthermore, the dynamic adjustment formula for adjusting the weight coefficients of the load loss term and the frequency overshoot integral term according to the frequency change rate satisfies

[0157]

[0158] in, represents the frequency change rate, f represents the frequency, k represents the sensitivity coefficient, represents the hyperbolic tangent function.

[0159] Combining power balance and frequency safety constraints, the feasible solution space for the objective function—that is, the range of load shedding—is determined. The frequency rate of change is collected in real time through the PMU, and a dynamic adjustment threshold is set. When the frequency rate of change exceeds the threshold, the weight of the frequency overshoot integral term, β, is increased, and the weight of the load loss term, α, is decreased. Conversely, α is increased and β is decreased. Based on the adjusted weight coefficients, the objective function is recalculated, and the optimal load shedding is determined.

[0160] Reference Figure 2An embodiment of the present invention provides a low-frequency load reduction system 2 based on zero-crossing real-time frequency tracking, and the system 2 specifically includes:

[0161] The signal processing module 201 is used to convert the high-voltage signal of the power grid into a low-voltage signal through a voltage transformer, generate a standard square wave signal through a zero-crossing comparison circuit, and use a high-speed comparator to construct a hysteresis comparison circuit to eliminate signal jitter to form a sensing signal;

[0162] The primary filtering module 202 is used to set a dynamic time window according to the power frequency characteristics of the sensor signal and eliminate abnormal zero-crossing points that exceed the dynamic time window;

[0163] The secondary verification module 203 is used to calculate the total harmonic distortion rate of the sensor signal in real time by using a sliding discrete Fourier transform, and determine whether to activate the dynamic weighted average algorithm by comparing the total harmonic distortion rate of the signal with a preset distortion rate threshold;

[0164] The third-level compensation module 204 is used to match the pre-stored harmonic feature library, establish a mapping relationship between harmonic patterns and phase offsets using a pre-trained deep generative adversarial network, and perform adaptive phase compensation on the pseudo zero-crossing points of the sensor signal;

[0165] The load shedding prediction module 205 is used to input sensor signals, PMU phasor data, and meteorological parameters into a three-dimensional convolutional neural network, calculate the cross-modal correlation matrix through the self-attention mechanism, combine with the LSTM unit to predict the future frequency change trajectory, and output the optimal load shedding strategy;

[0166] The composite criterion module 206 is used to perform emergency load shedding based on the optimal load shedding strategy, combined with the forward acceleration criterion and the reverse fault blocking, according to the load importance classification, and dynamically adjust the load shedding amount through the Nash equilibrium model to minimize load loss and frequency overshoot.

[0167] It is understandable that if Figure 1 The contents of the embodiment of the low-frequency load shedding method based on zero-crossing real-time frequency tracking shown in FIG. 1 are all applicable to the embodiment of the low-frequency load shedding system based on zero-crossing real-time frequency tracking. The functions specifically implemented by the embodiment of the low-frequency load shedding system based on zero-crossing real-time frequency tracking are the same as those in FIG. Figure 1 The low frequency load reduction method embodiment based on zero-crossing real-time frequency tracking is the same as that shown in FIG. Figure 1 The beneficial effects achieved by the embodiment of the low-frequency load reduction method based on zero-crossing real-time frequency tracking are also the same.

[0168] It should be noted that the information interaction, execution process and other contents between the above-mentioned systems are based on the same concept as the embodiment of the method of the present invention. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.

[0169] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0170] Reference Figure 3 An embodiment of the present invention further provides a computer device 3, comprising: a memory 302, a processor 301, and a computer program 303 stored in the memory 302. When the computer program 303 is executed on the processor 301, the low-frequency load reduction method based on zero-crossing real-time frequency tracking as described in any one of the above methods is implemented.

[0171] The computer device 3 may be a desktop computer, a notebook computer, a PDA, a cloud server or other computing devices. The computer device 3 may include, but is not limited to, a processor 301 and a memory 302. Those skilled in the art will understand that Figure 3 This is merely an example of the computer device 3 and does not constitute a limitation on the computer device 3 . The computer device 3 may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, the computer device 3 may also include input and output devices, network access devices, etc.

[0172] The processor 301 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor may be a microprocessor or any conventional processor.

[0173] In some embodiments, the memory 302 may be an internal storage unit of the computer device 3, such as a hard drive or memory of the computer device 3. In other embodiments, the memory 302 may also be an external storage device of the computer device 3, such as a plug-in hard drive, a Smart Media Card (SMC), a Secure Digital (SD) card, a flash memory card, etc. equipped on the computer device 3. Furthermore, the memory 302 may include both an internal storage unit of the computer device 3 and an external storage device. The memory 302 is used to store an operating system, application programs, a boot loader, data, and other programs, such as the program code of the computer program. The memory 302 may also be used to temporarily store data that has been output or is about to be output.

[0174] An embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the low-frequency load reduction method based on zero-crossing real-time frequency tracking as described in any one of the above methods is implemented.

[0175] In this embodiment, if the integrated unit is implemented as a software functional unit and sold or used as a standalone product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application can implement all or part of the process steps in the above-mentioned method embodiments by using a computer program to instruct the relevant hardware. The computer program can be stored in a computer-readable storage medium. When executed by a processor, the computer program can implement the steps of each of the above-mentioned method embodiments. The computer program includes computer program code, which can be in source code form, object code form, executable file, or some intermediate form. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a camera / terminal device, recording medium, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signals, telecommunication signals, and software distribution media. Examples include USB flash drives, removable hard drives, magnetic disks, or optical disks. In some jurisdictions, based on legislation and patent practice, computer-readable media cannot be electric carrier signals or telecommunication signals.

[0176] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant description of other embodiments.

[0177] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0178] In the embodiments disclosed in the present application, it should be understood that the disclosed devices / terminal equipment and methods can be implemented in other ways. For example, the device / terminal equipment embodiments described above are merely schematic. For example, the division of the modules or units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0179] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

Claims

1. A low-frequency load reduction method based on real-time zero-crossing frequency tracking, characterized in that: The method specifically includes: The high-voltage signal of the power grid is converted into a low-voltage signal through a voltage transformer, a standard square wave signal is generated through a zero-crossing comparison circuit, and a high-speed comparator is used to build a hysteresis comparison circuit to eliminate signal jitter and form a sensing signal; According to the power frequency characteristics of the sensor signal, a dynamic time window is set to eliminate abnormal zero-crossing points that exceed the dynamic time window; The sliding discrete Fourier transform is used to calculate the total harmonic distortion rate of the sensor signal in real time, and the activation of the dynamic weighted average algorithm is determined by comparing the total harmonic distortion rate of the signal with the preset distortion rate threshold; Matching the pre-stored harmonic feature library, using the pre-trained deep generative adversarial network to establish the mapping relationship between harmonic patterns and phase offsets, and adaptively compensate for the pseudo zero-crossing points of the sensor signal; The sensor signals, PMU phasor data, and meteorological parameters are input into a three-dimensional convolutional neural network. The cross-modal correlation matrix is calculated through the self-attention mechanism. Combined with the LSTM unit, the future frequency change trajectory is predicted and the optimal load reduction strategy is output. Based on the optimal load shedding strategy, combined with the forward acceleration criterion and reverse fault blocking, emergency load shedding is performed according to the load importance classification, and the load shedding amount is dynamically adjusted through the Nash equilibrium model to minimize load loss and frequency overshoot.

2. The method according to claim 1, characterized in that The method of setting a dynamic time window according to the power frequency characteristics of the sensor signal and eliminating abnormal zero-crossing points that exceed the dynamic time window specifically includes: An initial time window is calculated based on the allowable range of the power frequency of the sensing signal, and a dynamic time window range is set in combination with a time margin for compensating for measurement errors. The time margin is dynamically adjusted by the real-time frequency change rate. Collect any two timestamps of adjacent zero-crossing points and calculate the time interval between the two timestamps. If the time interval exceeds the dynamic time window range, mark the zero-crossing point as an outlier and remove it. Use a sliding window to store historical period data, calculate the weighted average period through the exponential decay weight formula, and adjust the dynamic time window range according to the weighted average period; If the continuous anomaly manifests as a rapid frequency drop, emergency control is triggered; otherwise, the zero-crossing sequence is reconstructed using weighted average periodic interpolation.

3. The method according to claim 1, characterized in that The method of using sliding discrete Fourier transform to calculate the total harmonic distortion rate of the sensor signal in real time and determining whether to activate the dynamic weighted average algorithm by comparing the total harmonic distortion rate of the signal with a preset distortion rate threshold specifically includes: The sensor signal is sampled according to a preset sampling frequency, and the kth harmonic component is calculated by a recursive formula, wherein the kth harmonic component includes a fundamental component and odd harmonic components; Based on the kth harmonic component, the sliding discrete Fourier transform is used to calculate the total harmonic distortion rate of the signal in the current window; Compare the total harmonic distortion rate of the signal with the preset distortion rate threshold. If the total harmonic distortion rate of the signal is greater than the preset distortion rate threshold, adaptively allocate weights according to harmonic energy; Based on the weights after adaptive allocation, the fundamental components are weighted averaged to obtain weighted fundamental components, the amplitudes of the weighted fundamental components are used to reconstruct the zero-crossing sequence, and the corrected power frequency is output.

4. The method according to claim 3, characterized in that The matching of the pre-stored harmonic feature library and the use of a pre-trained deep generative adversarial network to establish a mapping relationship between the harmonic pattern and the phase offset, and the adaptive phase compensation of the pseudo zero-crossing point of the sensor signal specifically include: The amplitude distribution of each odd harmonic component is statistically analyzed through historical fault data, and the initial harmonic phase angle and harmonic frequency of each odd harmonic component are collected to establish a pre-stored typical harmonic pattern parameter set and form a harmonic feature library; Construct a deep generative adversarial network, based on the harmonic feature library, and use the adversarial loss function to train the generator and recognizer of the deep generative adversarial network; The current harmonic energy ratio is calculated. If the current harmonic energy ratio is greater than the preset energy ratio threshold, it is determined that there is a pseudo zero-crossing point. The current harmonic parameters are input into the deep generative adversarial network, and the phase compensation amount is output.

5. The method according to claim 1, wherein The sensor signals, PMU phasor data, and meteorological parameters are input into a three-dimensional convolutional neural network. The cross-modal correlation matrix is calculated through the self-attention mechanism. The LSTM unit is combined to predict the future frequency change trajectory and output the optimal load reduction strategy. Specifically, the following steps are involved: Collect PMU phasor data and meteorological parameters, align the sensor signals, PMU phasor data and meteorological parameters in time, and then splice them into a three-dimensional tensor along the channel dimension; A three-dimensional convolutional neural network is used to extract features from the three-dimensional tensor, and a high-order feature map is generated through the ReLU activation function; Perform linear projection on the high-order feature map to obtain the query matrix, key matrix and value matrix. Based on the query matrix, key matrix and value matrix, calculate the cross-modal association matrix and perform weighted fusion to obtain weighted fusion features. The weighted fusion features are input into the LSTM unit, the hidden state is updated through the state equation, and the frequency prediction value within the preset future time period is output.

6. The method according to claim 1, characterized in that The optimal load shedding strategy, combined with the forward acceleration criterion and reverse fault blocking, performs emergency load shedding according to load importance classification and dynamically adjusts the load shedding amount through the Nash equilibrium model to minimize load loss and frequency overshoot. Specifically, it includes: Calculate the frequency change rate and voltage change rate in real time, and generate a primary load reduction instruction when the frequency change rate is greater than the frequency change rate threshold and the voltage change rate is less than the voltage change rate threshold; The positive sequence impedance mutation rate is calculated using PMU data. If the positive sequence impedance mutation rate is greater than the mutation rate threshold, the primary load reduction instruction is immediately blocked. Based on the preset load weight matrix, the objective function of minimizing load loss and frequency overshoot is constructed through the Nash equilibrium model; Based on the objective function and the primary load shedding instruction, the optimal load shedding amount of each load is obtained by solving the Lagrange multiplier method; Based on the optimal load reduction, each load is removed in stages starting from the high-weight load.

7. The method according to claim 6, characterized in that The objective function of minimizing load loss and frequency overshoot is constructed based on the preset load weight matrix through the Nash equilibrium model, specifically including: The load loss term is constructed according to the load priority weight in the preset load weight matrix, and the objective function is formed by combining the frequency overshoot integral term; Set power balance constraints and frequency safety constraints, and determine the solution range of the objective function based on the power balance constraints and frequency safety constraints; The weight coefficients of the load loss term and the frequency overshoot integral term are adjusted according to the frequency change rate.

8. A low-frequency load reduction system based on real-time frequency tracking of zero-crossing points, characterized in that: The system specifically includes: The signal processing module is used to convert the high-voltage signal of the power grid into a low-voltage signal through a voltage transformer, generate a standard square wave signal through a zero-crossing comparison circuit, and use a high-speed comparator to build a hysteresis comparison circuit to eliminate signal jitter to form a sensing signal; The primary filtering module is used to set a dynamic time window according to the power frequency characteristics of the sensor signal and eliminate abnormal zero-crossing points that exceed the dynamic time window; The secondary verification module is used to calculate the total harmonic distortion rate of the sensor signal in real time using a sliding discrete Fourier transform, and to determine whether to activate the dynamic weighted average algorithm by comparing the total harmonic distortion rate of the signal with a preset distortion rate threshold; The three-stage compensation module is used to match the pre-stored harmonic feature library, establish a mapping relationship between harmonic patterns and phase offsets using a pre-trained deep generative adversarial network, and perform adaptive phase compensation for the pseudo zero-crossing points of the sensor signal; The load shedding prediction module is used to input sensor signals, PMU phasor data, and meteorological parameters into a three-dimensional convolutional neural network. It calculates the cross-modal association matrix through a self-attention mechanism, combines it with LSTM units to predict future frequency change trajectories, and outputs the optimal load shedding strategy. The composite criterion module is used to perform emergency load shedding based on the optimal load shedding strategy, combined with the forward acceleration criterion and reverse fault blocking, according to the load importance classification, and dynamically adjust the load shedding amount through the Nash equilibrium model to minimize load loss and frequency overshoot.

9. A computer device, characterized in that: include: A memory, a processor, and a computer program stored in the memory, when the computer program is executed on the processor, implements the low-frequency load reduction method based on zero-crossing real-time frequency tracking as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that A computer program is stored thereon, and when the computer program is executed by a processor, the low-frequency load reduction method based on zero-crossing real-time frequency tracking according to any one of claims 1 to 7 is implemented.

Images