Load response method based on gru-cnn multi-modal learning

Abstract

The application discloses a load response method based on GRU-CNN multi-modal learning. The method comprises the following steps: acquiring a plurality of voltage values and power system parameters in a fault time period; inputting the plurality of voltage values and the power system parameters in the fault time period into a trained GRU-CNN load model respectively, and obtaining a physical power value. The method has the characteristics of high fault load response prediction efficiency, strong reliability, timely updating and the like.

Description

Technical Field

[0001] This invention relates to the field of fault load response technology, and more specifically, to a load response method based on GRU-CNN multimodal learning. Background Technology

[0002] Distribution network load refers to the total power demand carried by electrical equipment in a distribution system, including both active and reactive power. Therefore, load characteristic analysis is a crucial foundation for distribution network planning, helping to scientifically analyze load variation patterns and providing a basis for the design and optimization of the distribution network.

[0003] In the existing technology, the load modeling of the distribution network mainly relies on manual industry surveys or identification of massive waveform data, which has problems such as long modeling cycle, low accuracy of dynamic characteristic fitting, inability of model to be adaptively updated, and difficulty in supporting the stability analysis of new power systems. The current load modeling of the distribution network mainly adopts the statistical synthesis method and the overall measurement method. It requires manual detailed statistics on the composition of power equipment in various industries, or complex parameter verification by collecting fault waveform data with high sampling rate. This method has the following problems: (1) Low modeling efficiency and high dependence on manual statistics: The statistical synthesis method requires a large-scale survey of the composition of power equipment in power industries, which has high data collection costs, long cycle and poor timeliness; The overall measurement method faces problems such as tedious cleaning of massive raw data, data missing or significant noise interference, which leads to long modeling process and modeling accuracy is significantly affected by data processing experience. (2) Unreliable physical model fitting: With the large-scale access of new energy, the load exhibits complex characteristics such as strong nonlinearity, randomness and power backflow. Traditional ZIP models or induction motor models are too simplified. When the power grid experiences large disturbances, the fixed physical model cannot accurately reproduce the dynamic response process of frequency and power angle, which can easily lead to simulation conclusions deviating from reality and pose a risk of inaccurate safety analysis. (3) Lack of dynamic perception and online update mechanism: Most existing models are "offline trained and fixed in use", lacking the ability to perceive seasonal, regional and load structure evolution. The model parameters cannot be fine-tuned in real time according to the power grid operating status, making it difficult to achieve rapid migration from "general model" to "specific scenario model". The model has poor adaptability in humid environments or extreme operating conditions, affecting the accuracy of power grid dispatching decisions.

[0004] Existing technology discloses a CNN-based The power load forecasting method using GRU and ARIMA models covers the field of power load forecasting in power systems. Specifically, it includes: S1: Extracting power load data and performing data preprocessing; S2: Constructing a GRU model... The algorithm consists of five modules: S1: CNN module for extracting time-series features from power load data; S2: ARIMA module for building a time series model and making predictions; S3: Model training and evaluation; S4: Model fusion and prediction output. This method is computationally complex and has low modeling efficiency. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies, such as low modeling efficiency, unreliable model fitting, and the lack of dynamic perception and online update mechanisms, by providing a load response method based on GRU-CNN multimodal learning. This method features high efficiency, strong reliability, and timely updates in fault load response prediction.

[0006] The primary objective of this invention is to solve the aforementioned technical problems. The technical solution of this invention is as follows:

[0007] Load response methods based on GRU-CNN multimodal learning include: S1: Obtain multiple voltage values ​​and power system parameters during the fault period; S2: Input the multiple voltage values ​​and power system parameters during the fault period into the trained GRU-CNN load model to obtain the physical power value.

[0008] Furthermore, the GRU-CNN load model includes: a normalization layer, an input layer, a gated recurrent unit layer, a first fully connected layer, a second fully connected layer, a third fully connected layer, an attention mechanism layer, a convolutional layer, and an output layer; Multiple voltage values ​​during the fault period are input to the input terminal of the normalization layer. The output terminal of the normalization layer is connected to the input terminal of the input layer. The output terminal of the input layer is connected to the input terminal of the gated loop unit layer. The output terminal of the gated loop unit layer is connected to the input terminal of the first fully connected layer. The output terminal of the first fully connected layer is connected to the input terminal of the second fully connected layer. The power system parameters are input to the input terminal of the convolutional layer. The output terminal of the convolutional layer is connected to the input terminal of the third fully connected layer. The output terminal of the third fully connected layer is connected to the input terminal of the attention mechanism layer. The output terminals of the attention mechanism layer and the second fully connected layer are connected to the input terminal of the output layer. The output terminal of the output layer outputs the physical power value.

[0009] Furthermore, the formula for the output layer is as follows:

[0010] This represents the output of the attention mechanism layer. This represents the output of the second fully connected layer. This represents a fully connected fusion function;

[0011] Indicates active power. Indicates reactive power. This represents the activation function. , Represents a preset matrix. Indicates fused input. , This represents a preset constant. This represents the physical power value.

[0012] Furthermore, step S2 includes the following: S3: Input multiple voltage values ​​and power system parameters within the fault time period into a preset physical hybrid model for processing to obtain theoretical output values; S4: If the deviation between the theoretical output value and the physical power value exceeds a preset threshold, the physical power value is weighted and corrected according to the theoretical output value to obtain the final physical power value.

[0013] A method for training a GRU-CNN load model includes: S01: Obtain the training dataset, which includes multiple sets of data; each set of data includes: multiple voltage values, power system parameters, active power, and reactive power during the fault period; construct the GRU-CNN load model; S02: Select a set of data from the training dataset as the first set of data; S03: Input multiple voltage values ​​and power system parameters during the fault time period in the first data into the GRU-CNN load model to obtain the response active power value and the response reactive power value; S04: Calculate the instantaneous power loss value based on the response active power value, the response reactive power value, the active power of the first data, and the reactive power of the first data; S05: Optimize the GRU-CNN load model based on the instantaneous power loss value; S06: Select another set of data from the training dataset as the new first data, and repeat steps S03 to S05 until the preset conditions are met to obtain the trained GRU-CNN load model.

[0014] Furthermore, the formula for calculating the instantaneous power loss value is as follows:

[0015] The weight representing active power. The weighting of reactive power This indicates the active power value of the response. This indicates the reactive power value in response. This represents the active power value of the first data point. This represents the reactive power value of the first data point. This represents the L2 norm regularization term.

[0016] Furthermore, the preset conditions include: the average absolute percentage error growth is less than a preset value; The formula for calculating the mean absolute percentage error is as follows:

[0017] Indicates the mean absolute percentage error. This represents the total number of iterations, where t represents one iteration. This represents the active power value of the response in one cycle. This represents the reactive power value in one cycle. This represents the active power value of the first data point in one cycle. This represents the reactive power value of the first data in one cycle.

[0018] Furthermore, in step S05, optimizing the GRU-CNN load model only includes: optimizing the parameters of the first fully connected layer, optimizing the parameters of the second fully connected layer, and optimizing the parameters of the third fully connected layer.

[0019] A load response system based on GRU-CNN multimodal learning includes: Data acquisition module: Acquires multiple voltage values ​​and power system parameters during the fault period; Power inference module: Input multiple voltage values ​​and power system parameters during the fault period into the trained GRU-CNN load model to obtain physical power values.

[0020] A GRU-CNN load model training system includes: Dataset Acquisition Module: Acquires the training dataset, which includes multiple sets of data; each set of data includes: multiple voltage values, power system parameters, active power, and reactive power during the fault period; constructs a GRU-CNN load model; Data selection module: Selects a set of data from the training dataset as the first set of data; Response module: Input multiple voltage values ​​and power system parameters within the fault time period from the first data into the GRU-CNN load model to obtain the response active power value and the response reactive power value; Loss calculation module: Based on the response active power value, response reactive power value, active power of the first data, and reactive power of the first data, calculate the instantaneous power loss value; Optimization module: Optimizes the GRU-CNN load model based on the instantaneous power loss value; Conditional judgment module: Select another set of data from the training dataset as the new first data, repeat until the preset condition is met, and obtain the trained GRU-CNN load model.

[0021] Compared with the prior art, the beneficial effects of the present invention are: This invention employs a GRU-CNN load model, which can deeply capture the strong nonlinear and stochastic characteristics of loads under fault disturbances. Experiments demonstrate that this model significantly outperforms traditional basic recurrent neural network (RNN) and standard artificial neural network (ANN) models in terms of fitting accuracy for active and reactive power, and can accurately reproduce the dynamic evolution of frequency and power angle. Compared with traditional ZIP or induction motor models, it greatly reduces the risk of simulation conclusions deviating from actual operating conditions. Therefore, this method features high efficiency, high reliability, and timely updates in fault load response prediction. Attached Figure Description

[0022] Figure 1 The flowchart shows the load response method based on GRU-CNN multimodal learning provided in Example 1.

[0023] Figure 2 This is a schematic diagram of the structure of the GRU-CNN load model provided in Example 1.

[0024] Figure 3 This is a schematic diagram of the structure of the gated loop unit layer provided in Example 1.

[0025] Figure 4 This is a schematic diagram of the structure of the convolutional layer provided in Example 1.

[0026] Figure 5 The flowchart is provided for a GRU-CNN load model training method in Example 1.

[0027] Figure 6 A line graph comparing the actual and predicted values ​​of active power provided in Example 1.

[0028] Figure 7 The normalized absolute error of active power provided in Example 1 is shown in a line graph. Detailed Implementation

[0029] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent. To better illustrate this embodiment, some parts in the accompanying drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions; It will be understood by those skilled in the art that certain well-known structures and their descriptions may be omitted in the accompanying drawings.

[0030] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0031] Example 1 like Figure 1 As shown, the load response method based on GRU-CNN multimodal learning includes: S1: Obtain multiple voltage values ​​and power system parameters during the fault period; S2: Input the multiple voltage values ​​and power system parameters during the fault period into the trained GRU-CNN load model to obtain the physical power value.

[0032] It should be noted that modeling refers to building a model that can provide a dynamic response like a real load when a fault occurs in the power grid (voltage / frequency fluctuations) (for security and stability analysis).

[0033] In the field of distribution network load modeling, the following traditional methods are typically relied upon: (1) Industry Statistical Synthesis Method: This method relies on periodic surveys and statistics on the composition ratio of electrical equipment (such as lighting, air conditioning, and electric motors) in different industries (such as industry, commerce, and residential). Based on the typical characteristic curves of various types of electrical equipment, a static load model reflecting the characteristics of a specific industry is constructed by weighted superposition. The industry statistical synthesis method is highly dependent on manual surveys and classification statistics of the composition of electrical equipment. Its data acquisition cost is high and the update cycle is long, resulting in poor model timeliness and difficulty in capturing the dynamic evolution characteristics of new loads, including distributed power sources and electric vehicles.

[0034] (2) Physical parameter identification method: Based on bus disturbance data collected by fault recording devices (DFR) or phasor measurement units (PMU), a fixed physical model structure (such as ZIP model, induction motor model, or a combination thereof) is set. Through nonlinear optimization algorithms, the unknown physical parameters in the model are fitted and calibrated so that the simulated power curve approaches the measured curve under specific disturbances. The physical parameter identification method has extremely high requirements for the quality of fault recording data. However, in the actual operation of the power grid, the massive amount of raw data often contains noise or is missing, resulting in a cumbersome parameter identification process and low computational efficiency. At the same time, traditional fixed physical structures such as ZIP or induction motors are difficult to accurately fit the complex nonlinear power response and power feedback phenomenon under the new power system.

[0035] (3) Empirical Proportion Allocation Method: In the absence of real-time monitoring methods or detailed survey data, operators, based on historical experience, literature research, or typical industry examples, assign a fixed ZIP component ratio and motor proportion (such as the common combination of 60% constant impedance and 40% induction motor), and apply it directly as a general parameter in system simulation analysis. The empirical proportion allocation method lacks scientific basis for specific scenarios and cannot take into account the load differences of different seasons, regions, and areas. When conducting power system safety and stability analysis under large disturbances, the simulation conclusions often deviate from reality due to the "mismatch" between model parameters and actual operating conditions, which poses a risk of misleading dispatch decisions.

[0036] In the field of data-driven and AI-based load modeling, existing technologies typically rely on the following methods: (1) Standard Artificial Neural Network (ANN) Modeling: The distribution network load is treated as a "black box". A three-layer BP neural network or radial basis function (RBF) neural network is used for training with a large amount of voltage, frequency input and power output sample data to try to construct a nonlinear mapping relationship between voltage / frequency disturbances and load power changes. Standard artificial neural network (ANN) modeling lacks the ability to process time-series characteristics: Since ANN is essentially a static nonlinear mapping, it cannot effectively extract the time-series evolution characteristics of the load during the fault process, which limits its accuracy in describing dynamic characteristics such as load differential algebraic equations and makes it difficult to reproduce the transient evolution process of the power system under large disturbances.

[0037] (2) Basic Recurrent Neural Network (RNN) Modeling: This method utilizes the recursive structure of neural networks to process time-series data of the power system. By passing hidden states, it attempts to capture the evolution of load power in the time dimension, thereby describing the dynamic response characteristics of the load during power grid faults. Basic Recurrent Neural Network (RNN) modeling is difficult to take into account spatial features and has poor training stability: Ordinary RNNs are prone to gradient vanishing or gradient exploding problems when processing long sequence data, resulting in unstable modeling effects. At the same time, since it only focuses on the information transmission in the time dimension and ignores the spatial topological features in the power flow calculation of the distribution network, the model is not comprehensive enough in feature extraction and the fitting effect is not ideal when facing complex and ever-changing power grid operating conditions.

[0038] (3) Traditional machine learning regression modeling: Classic machine learning algorithms such as Support Vector Machine (SVM), Random Forest, or Extreme Gradient Boosting Tree (XGBoost) are used to extract and regress small-scale disturbance features, aiming to establish a correspondence model between system state variables and equivalent load parameters. Traditional machine learning regression modeling has insufficient generalization ability and scalability: Such models are often "black box" models, lacking physical mechanism support, and are usually trained offline for specific sample sets. When the load structure of the distribution network changes significantly due to seasonal changes, distributed power source access, etc., the model cannot be fine-tuned or incrementally updated online, and must be retrained, resulting in low modeling efficiency and difficulty in adapting to the real-time response requirements of new power systems.

[0039] Furthermore, such as Figure 2 As shown, the GRU-CNN load model includes: a normalization layer, an input layer, a gated recurrent unit layer, a first fully connected layer, a second fully connected layer, a third fully connected layer, an attention mechanism layer, a convolutional layer, and an output layer; Multiple voltage values ​​during the fault period are input to the input terminal of the normalization layer. The output terminal of the normalization layer is connected to the input terminal of the input layer. The output terminal of the input layer is connected to the input terminal of the gated loop unit layer. The output terminal of the gated loop unit layer is connected to the input terminal of the first fully connected layer. The output terminal of the first fully connected layer is connected to the input terminal of the second fully connected layer. The power system parameters are input to the input terminal of the convolutional layer. The output terminal of the convolutional layer is connected to the input terminal of the third fully connected layer. The output terminal of the third fully connected layer is connected to the input terminal of the attention mechanism layer. The output terminals of the attention mechanism layer and the second fully connected layer are connected to the input terminal of the output layer. The output terminal of the output layer outputs the physical power value.

[0040] It should be noted that this invention significantly improves modeling efficiency and effectively reduces reliance on manual labor: through the automatic feature extraction capabilities of deep learning, the model can directly learn load characteristics from massive amounts of waveform data or simulation data, eliminating the need for large-scale, high-cost manual industry electricity consumption surveys. The modeling cycle is shortened from the traditional months to days or even hours, effectively solving the bottleneck problems of slow data updates and cumbersome manual verification in statistical synthesis methods.

[0041] The CNN module extracts spatial topological features by simulating power flow calculations, while the GRU module captures temporal evolution patterns by representing differential-algebraic equations. This spatiotemporal decoupling and fusion design effectively solves the gradient vanishing problem of basic recurrent neural networks when processing long sequence data, enabling the system to maintain extremely high training stability and generalization performance when facing complex and ever-changing power grid topologies and large-scale disturbance conditions.

[0042] In one specific embodiment, such as Figure 3As shown, the gated cyclic unit layer uses gated cyclic units (GRUs) to process time series data and use differential algebraic equations (DAEs) to characterize the load, accurately capturing the time evolution of voltage and frequency during disturbances.

[0043] The dynamic response of electrical loads is essentially governed by a complex set of differential algebraic equations (DAEs). To accurately characterize the inertial characteristics of induction motors and the recovery properties of loads, this patent employs a gated recurrent unit (GRU) as the core of the time-series processing. Compared to traditional RNNs, GRUs effectively solve the gradient vanishing problem in long-sequence training through a simplified gating structure.

[0044] Its state evolution logic is as follows: Reset Gate: Determines the influence of the previous state on the current candidate state.

[0045]

[0046] Update Gate: Controls the fusion ratio of historical information and current input information, corresponding to the "memory" length of the load transition from transient to steady state in the power system.

[0047]

[0048] Candidate hidden state: The nonlinear transient response of the simulated load when subjected to a momentary voltage drop.

[0049]

[0050] Ultimately hidden state: Enables smooth iterative updates of state variables.

[0051]

[0052] like Figure 4 As shown, the convolutional layer extracts the spatial topological features of the distribution network under different operating conditions through a convolutional neural network (CNN). Leveraging the similarity between CNN convolutional computation and power flow calculation (NRPFC), the node voltage correction process is simulated to extract key power system features caused by disturbances.

[0053] In active distribution networks (ADNs), load distribution exhibits distinct spatial topological characteristics. This patent utilizes the local perception characteristics of convolutional neural networks (CNNs) to extract the topological correlation between voltage distribution and power flow, which mathematically simulates the Jacobian matrix mapping process of Newton-Raphson power flow calculation (NRPFC).

[0054] Let the input feature map be This represents the voltage distribution at different nodes in the distribution network. The operation of a convolutional layer is defined as follows:

[0055] Among them, convolution kernel By sliding across the feature map, the interaction between adjacent nodes is captured. Through multi-layer convolution stacking, CNN can extract power spatial coupling features caused by grid disturbances. This method avoids the need to explicitly construct complex node admittance matrices in traditional modeling. The cumbersome process of [the process] greatly improves the model's adaptability to changes in distribution network topology.

[0056] Furthermore, the formula for the output layer is as follows:

[0057] This represents the output of the attention mechanism layer. This represents the output of the second fully connected layer. This represents a fully connected fusion function;

[0058] Indicates active power. Indicates reactive power. This represents the activation function. , Represents a preset matrix. Indicates fused input. , This represents a preset constant. This represents the physical power value.

[0059] It should be noted that the output layer deeply fuses the GRU branch (representing the load differential algebra equation) with the CNN branch (simulating power flow calculations to extract features). Through a fully connected layer (FC), multidimensional features are mapped to specific active power (P) and reactive power (Q) output curves, solving the challenge of dynamic equivalent modeling in ADN (Active Distribution Network). This "spatiotemporal decoupling-feature fusion" design ensures that the model can reflect both the proportional relationships of the load composition and reproduce complex fault response trajectories.

[0060] Furthermore, step S2 includes the following: S3: Input multiple voltage values ​​and power system parameters within the fault time period into a preset physical hybrid model for processing to obtain theoretical output values; S4: If the deviation between the theoretical output value and the physical power value exceeds a preset threshold, the physical power value is weighted and corrected according to the theoretical output value to obtain the final physical power value.

[0061] It should be noted that in step S4, the weighted correction includes: Define the equivalent load ratio vector The strategy of "preliminary boundary determination through literature review + iterative optimization through trial and error" was used for refined adjustments: Optimize the objective function:

[0062] Through multiple rounds of fault fitting analysis, this patent determined the optimal ratio scheme adapted to the new power system: the active component ratio of ZIP is maintained at 1:0:0, the reactive component is allocated at 0.3:0.4:0.3, and the proportion of induction motors is set to 0.6. This parameter scheme significantly improves the transient fitting accuracy of the model under extreme faults.

[0063] Its core lies in constructing a hybrid modeling logic of "mechanism-based delimitation and data-driven optimization." By quantitatively analyzing the sensitivity of ZIP model components (active power 1:0:0, etc.) and the proportion of induction motors (0.6) to the stability of system frequency and power angle, the traditional statistical synthesis ratio is used as the "physical anchor point" of the deep learning model. This not only ensures that the model output strictly follows the basic physical constraints of the power system, but also achieves an organic unity between the fitting efficiency of the black-box model and the interpretability of the white-box physical model.

[0064] like Figure 5 As shown, a GRU-CNN load model training method includes: S01: Obtain the training dataset, which includes multiple sets of data; each set of data includes: multiple voltage values, power system parameters, active power, and reactive power during the fault period; construct the GRU-CNN load model; S02: Select a set of data from the training dataset as the first set of data; S03: Input multiple voltage values ​​and power system parameters during the fault time period in the first data into the GRU-CNN load model to obtain the response active power value and the response reactive power value; S04: Calculate the instantaneous power loss value based on the response active power value, the response reactive power value, the active power of the first data, and the reactive power of the first data; S05: Optimize the GRU-CNN load model based on the instantaneous power loss value; S06: Select another set of data from the training dataset as the new first data, and repeat steps S03 to S05 until the preset conditions are met to obtain the trained GRU-CNN load model.

[0065] In one specific embodiment, the preset condition is that when the MAPE improvement is less than 0.01% for 20 consecutive training epochs, or when the validation set loss shows a continuous upward trend, the early stopping mechanism is triggered.

[0066] Furthermore, the formula for calculating the instantaneous power loss value is as follows:

[0067] The weight representing active power. The weighting of reactive power This indicates the active power value of the response. This indicates the reactive power value in response. This represents the active power value of the first data point. This represents the reactive power value of the first data point. This represents the L2 norm regularization term.

[0068] In one specific embodiment, typically take Ensure stable simulation accuracy of the power angle.

[0069] Furthermore, the preset conditions include: the average absolute percentage error growth is less than a preset value; The formula for calculating the mean absolute percentage error is as follows:

[0070] Indicates the mean absolute percentage error. This represents the total number of iterations, where t represents one iteration. This represents the active power value of the response in one cycle. This represents the reactive power value in one cycle. This represents the active power value of the first data point in one cycle. This represents the reactive power value of the first data in one cycle.

[0071] Furthermore, in step S05, optimizing the GRU-CNN load model only includes: optimizing the parameters of the first fully connected layer, optimizing the parameters of the second fully connected layer, and optimizing the parameters of the third fully connected layer.

[0072] The overall deployment process is as follows: 1. Offline evolutionary pre-training based on high-fidelity simulation: Before deploying the algorithm, offline training with a high-fidelity model is required in the Simulink environment. The Adam optimization algorithm with a momentum term is used for weight updates. First-moment estimate of gradient calculation and second-order moment estimation .

[0073] Parameter update rules:

[0074] Learning rate Set as This stage aims to enable the model to grasp the general knowledge of power load response, laying the foundation for subsequent online migration.

[0075] In the pre-training process, step S05 optimizes the GRU-CNN load model by optimizing all layers.

[0076] 2. Online Adaptive Fine-tuning Mechanism: Addressing the significant differences in load composition between different areas (e.g., industrial areas and residential areas), this patent introduces an online adaptive algorithm based on fine-tuning technology. When new measured waveform data is acquired... At this time, the algorithm keeps the underlying convolutional features unchanged and only modifies the high-level perceptual parameters:

[0077] Fine-tuning step size Take the smaller value (e.g.) This ensures that the local characteristics of a specific power grid can be quickly captured without compromising the learned physical stability, thus achieving a precise conversion of the model from "general" to "scenario-specific".

[0078] In the fine-tuning, step S05, optimizing the GRU-CNN load model only includes: optimizing the parameters of the first fully connected layer, optimizing the parameters of the second fully connected layer, and optimizing the parameters of the third fully connected layer.

[0079] If the predicted power deviation after the online update instantaneously exceeds 15%, the system will automatically discard the current weights, revert to the optimal model verified in the previous version, and trigger an alarm log for operation and maintenance personnel to check.

[0080] It should be noted that the system introduces an update mechanism based on fine-tuning technology, which can perceive the seasonal evolution and regional differences in the distribution network load structure. By introducing small-sample measured waveform data for rapid fine-tuning, the model can achieve seamless migration from a "general model" to a "scenario-specific model," ensuring the continued effectiveness of the simulation model under distributed power source fluctuations or the access of new equipment.

[0081] A load response system based on GRU-CNN multimodal learning includes: Data acquisition module: Acquires multiple voltage values ​​and power system parameters during the fault period; Power inference module: Input multiple voltage values ​​and power system parameters during the fault period into the trained GRU-CNN load model to obtain physical power values.

[0082] A GRU-CNN load model training system includes: Dataset Acquisition Module: Acquires the training dataset, which includes multiple sets of data; each set of data includes: multiple voltage values, power system parameters, active power, and reactive power during the fault period; constructs a GRU-CNN load model; Data selection module: Selects a set of data from the training dataset as the first set of data; Response module: Input multiple voltage values ​​and power system parameters within the fault time period from the first data into the GRU-CNN load model to obtain the response active power value and the response reactive power value; Loss calculation module: Based on the response active power value, response reactive power value, active power of the first data, and reactive power of the first data, calculate the instantaneous power loss value; Optimization module: Optimizes the GRU-CNN load model based on the instantaneous power loss value; Conditional judgment module: Select another set of data from the training dataset as the new first data, repeat until the preset condition is met, and obtain the trained GRU-CNN load model.

[0083] In one specific embodiment, the output layer further includes a model fine-tuning update module: This module addresses the data distribution offset problem by introducing model fine-tuning techniques. Based on the pre-trained model, it performs short-cycle retraining using small-sample waveform data from new scenarios to quickly correct biases. The prediction results are as follows... Figure 6 , Figure 7 As shown.

[0084] This invention utilizes Simulink to build a detailed power distribution network simulation model (such as the Jinghu area in Dongguan) for model training. Subsequently, actual fault waveform data (such as measured data from Beizha and Gukeng substations) are introduced for field verification to ensure high reliability of the model in real-world environments. Furthermore, a strategy combining literature review and trial-and-error is employed to optimize and adjust the proportions of each component and the percentage of induction motors in the ZIP model. Ultimately, the optimal scheme adapted to the Guangdong power grid is determined (ZIP active power 1:0:0, reactive power 0.3:0.4:0.3, induction motor percentage 0.6), effectively improving the simulation accuracy of frequency and power angle stability.

[0085] The same or similar labels correspond to the same or similar parts; The terms used to describe positional relationships in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent. Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A load response method based on GRU-CNN multimodal learning, characterized in that, include: S1: Obtain multiple voltage values ​​and power system parameters during the fault period; S2: Input the multiple voltage values ​​and power system parameters during the fault period into the trained GRU-CNN load model to obtain the physical power value.

2. The load response method based on GRU-CNN multimodal learning according to claim 1, characterized in that, The GRU-CNN load model includes: a normalization layer, an input layer, a gated recurrent unit layer, a first fully connected layer, a second fully connected layer, a third fully connected layer, an attention mechanism layer, a convolutional layer, and an output layer; Multiple voltage values ​​during the fault period are input to the input terminal of the normalization layer. The output terminal of the normalization layer is connected to the input terminal of the input layer. The output terminal of the input layer is connected to the input terminal of the gated loop unit layer. The output terminal of the gated loop unit layer is connected to the input terminal of the first fully connected layer. The output terminal of the first fully connected layer is connected to the input terminal of the second fully connected layer. The power system parameters are input to the input terminal of the convolutional layer. The output terminal of the convolutional layer is connected to the input terminal of the third fully connected layer. The output terminal of the third fully connected layer is connected to the input terminal of the attention mechanism layer. The output terminals of the attention mechanism layer and the second fully connected layer are connected to the input terminal of the output layer. The output terminal of the output layer outputs the physical power value.

3. The load response method based on GRU-CNN multimodal learning according to claim 2, characterized in that, The formula for the output layer is as follows: This represents the output of the attention mechanism layer. This represents the output of the second fully connected layer. This represents a fully connected fusion function; Indicates active power. Indicates reactive power. This represents the activation function. , Represents a preset matrix. Indicates fused input. , This represents a preset constant. This represents the physical power value.

4. The load response method based on GRU-CNN multimodal learning according to claim 1, characterized in that, Step S2 is followed by: S3: Input multiple voltage values ​​and power system parameters within the fault time period into a preset physical hybrid model for processing to obtain theoretical output values; S4: If the deviation between the theoretical output value and the physical power value exceeds a preset threshold, the physical power value is weighted and corrected according to the theoretical output value to obtain the final physical power value.

5. A method for training a GRU-CNN load model, characterized in that, include: S01: Obtain the training dataset, which includes multiple sets of data; each set of data includes: multiple voltage values, power system parameters, active power, and reactive power during the fault period; construct the GRU-CNN load model; S02: Select a set of data from the training dataset as the first set of data; S03: Input multiple voltage values ​​and power system parameters during the fault time period in the first data into the GRU-CNN load model to obtain the response active power value and the response reactive power value; S04: Calculate the instantaneous power loss value based on the response active power value, the response reactive power value, the active power of the first data, and the reactive power of the first data; S05: Optimize the GRU-CNN load model based on the instantaneous power loss value; S06: Select another set of data from the training dataset as the new first data, and repeat steps S03 to S05 until the preset conditions are met to obtain the trained GRU-CNN load model.

6. The GRU-CNN load model training method according to claim 5, characterized in that, The formula for calculating the instantaneous power loss is as follows: The weight representing active power. The weight representing reactive power. This indicates the active power value of the response. This indicates the reactive power value in response. This represents the active power value of the first data point. This represents the reactive power value of the first data point. This represents the L2 norm regularization term.

7. The GRU-CNN load model training method according to claim 5, characterized in that, The preset conditions include: the average absolute percentage error growth is less than a preset value; The formula for calculating the mean absolute percentage error is as follows: Indicates the mean absolute percentage error. This represents the total number of iterations, where t represents one iteration. This represents the active power value of the response in one cycle. This represents the reactive power value in one cycle. This represents the active power value of the first data point in one cycle. This represents the reactive power value of the first data in one cycle.

8. The GRU-CNN load model training method according to claim 5, characterized in that, In step S05, optimizing the GRU-CNN load model only includes: optimizing the parameters of the first fully connected layer, optimizing the parameters of the second fully connected layer, and optimizing the parameters of the third fully connected layer.

9. A load response system based on GRU-CNN multimodal learning, applied to the response method described in any one of claims 1 to 4, characterized in that, include: Data acquisition module: Acquires multiple voltage values ​​and power system parameters during the fault period; Power inference module: Input multiple voltage values ​​and power system parameters during the fault period into the trained GRU-CNN load model to obtain physical power values.

10. A GRU-CNN load model training system, applied to the training method described in any one of claims 5 to 8, characterized in that, include: Dataset Acquisition Module: Acquires the training dataset, which includes multiple sets of data; each set of data includes: multiple voltage values, power system parameters, active power, and reactive power during the fault period; constructs a GRU-CNN load model; Data selection module: Selects a set of data from the training dataset as the first set of data; Response module: Input multiple voltage values ​​and power system parameters within the fault time period from the first data into the GRU-CNN load model to obtain the response active power value and the response reactive power value; Loss calculation module: Based on the response active power value, response reactive power value, active power of the first data, and reactive power of the first data, calculate the instantaneous power loss value; Optimization module: Optimizes the GRU-CNN load model based on the instantaneous power loss value; Conditional judgment module: Select another set of data from the training dataset as the new first data, repeat until the preset condition is met, and obtain the trained GRU-CNN load model.

Images