Active fluctuation collaborative stabilizing method and system for high-proportion distributed new energy power grid

Abstract

The invention discloses an active fluctuation collaborative stabilizing method and system for a high-proportion distributed new energy power grid, and relates to the technical field of power grid dispatching. According to the method, a physically consistent weather prediction model is established through multi-source meteorological data fusion, and a regional fluctuation propagation rule is accurately captured; identifying a high-risk fluctuation cluster based on dynamic time warping and spectral clustering, simulating a fluctuation propagation path by using a digital twin platform, and quantifying resource requirements; energy storage resource configuration is optimized by adopting mixed integer programming and a column generation algorithm, and multi-dimensional stability verification is carried out through a digital twin environment; a self-adaptive optimization mechanism based on reinforcement learning is established, continuous evolution of the system is realized, the technical bottlenecks of a traditional method in the aspects of fluctuation perception, resource allocation, system self-adaption and the like are solved, a collaborative stabilization mechanism with accurate prediction, intelligent recognition and decision optimization is formed, and a complete solution is provided for safe and stable operation of a high-proportion new energy power grid.

Description

Technical Field

[0001] This invention relates to the field of power grid dispatching technology, specifically to a method and system for collaborative mitigation of active power fluctuations in high-proportion distributed renewable energy power grids. Background Technology

[0002] Against the backdrop of energy transition, building a stable and efficient power system has become a core task for ensuring energy security and achieving green development. With the large-scale integration of distributed renewable energy sources such as photovoltaics and wind power, the power grid is facing unprecedented challenges, the importance of which is self-evident. Especially in scenarios with a high proportion of renewable energy connected to the grid, power fluctuations have become a key factor affecting grid stability, directly related to power supply quality and the absorption capacity of renewable energy. However, current methods for dealing with power fluctuations are often limited to passive adjustments in local links, lacking a comprehensive understanding of the root causes of fluctuations and the ability to coordinate systematically. When faced with complex and changeable weather conditions and distributed energy sources connected to multiple points, this approach is unable to effectively capture the underlying correlation patterns of fluctuations, nor can it achieve reasonable resource allocation at different time and spatial scales, resulting in a significant reduction in the effectiveness of fluctuation mitigation.

[0003] A deeper technical challenge lies in the fact that the power fluctuations of distributed renewable energy sources are highly uncertain and regionally correlated. The output power of renewable energy sources is greatly affected by weather changes, and weather itself is a dynamic process. Changes in wind speed or sunlight in one area often affect surrounding areas, forming a chain reaction. This regional correlation makes fluctuation prediction extremely complex. Traditional prediction methods are unable to accurately grasp the fluctuation trend. Furthermore, this inadequacy in prediction directly affects the timeliness and accuracy of resource allocation. For example, within a certain period of time, the power grid in a certain area may experience power shortages or surpluses due to failure to predict fluctuations in advance, thereby causing system instability. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for collaborative mitigation of active power fluctuations in high-proportion distributed renewable energy power grids, which realizes multi-level collaborative mitigation of active power fluctuations in high-proportion distributed renewable energy power grids.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] This application provides a method for coordinated mitigation of active power fluctuations in high-proportion distributed renewable energy power grids, including the following steps:

[0007] S1. Real-time collection of multi-regional weather data through satellite remote sensing and ground sensor networks; embedding meteorological physical equations as constraints into the data fusion process through physical information neural networks; integrating into a unified spatiotemporal sequence dataset; obtaining a physically consistent weather change trend sequence.

[0008] S2. Based on the weather change trend sequence, apply the spatiotemporal graph attention network to process time series features, extract inter-regional propagation patterns and potential chain reaction features, and determine the regional correlation fluctuation prediction model.

[0009] S3. If the chain reaction characteristics in the regional correlation fluctuation prediction model exceed the preset threshold, then the dynamic time warping is used as a similarity measure for the spectral clustering algorithm to group the regions based on the morphological similarity of the fluctuation curves, thereby obtaining potential high-risk fluctuation clusters.

[0010] S4. Based on information on potential high-risk fluctuation clusters and historical power data, construct a digital twin model to simulate the propagation path of power fluctuations at different time scales and obtain a distribution map of resource coordination demand.

[0011] S5. Based on the specific requirements quantified by the resource coordination demand distribution map, as well as the power grid physical model and operation rules, a mixed integer programming model with the goal of minimizing total cost and transmission loss is constructed, and a column generation algorithm is used for efficient solution to calculate the real-time resource allocation scheme for backup power and energy storage equipment.

[0012] S6. Update the power grid control system parameters through the determined real-time resource allocation scheme, simulate the execution of coordinated operations, obtain the system stability indicators after feedback, and determine whether iterative adjustments are needed.

[0013] S7. If the system stability indicators do not meet the preset standards, the adaptive optimization and adjustment mechanism will be activated to dynamically correct the resource allocation scheme and reassess the system stability.

[0014] This application also provides a system for coordinated active power fluctuation mitigation in high-proportion distributed renewable energy power grids, and a method for coordinated active power fluctuation mitigation in high-proportion distributed renewable energy power grids, including:

[0015] The multi-source meteorological fusion module integrates satellite remote sensing data and ground sensor networks, and uses physical information neural network technology to embed meteorological physical equations as constraints into the data fusion process to generate physically consistent weather change trend sequences.

[0016] The regional fluctuation prediction module constructs a regional spatiotemporal map structure based on weather change trend sequences, applies a spatiotemporal map attention network to extract inter-regional propagation patterns and chain reaction characteristics, establishes a regionally related fluctuation prediction model, and outputs a risk-calibrated fluctuation prediction distribution.

[0017] The fluctuation cluster identification module, when the detected chain reaction characteristics exceed the preset threshold, uses dynamic time warping algorithm and spectral clustering method to group regions based on the similarity of fluctuation curve morphology, identify potential high-risk fluctuation clusters, and generate a control priority sequence for regional classification.

[0018] The digital twin simulation module combines high-risk fluctuation cluster information and power grid topology parameters to construct a digital twin model of the power grid, simulate the propagation path of power fluctuations at different time scales, and generate a resource coordination demand distribution map through quantitative analysis of resource demand.

[0019] The resource optimization decision module, based on the resource coordination demand distribution map, constructs a mixed integer programming model with the goal of minimizing total cost and transmission loss, and uses a column generation algorithm for efficient solution to calculate the real-time resource allocation scheme for backup power and energy storage equipment.

[0020] The control strategy verification module injects the resource allocation scheme into the digital twin system, simulates the execution of coordination operations, uses a deep neural network to evaluate the system stability indicators in real time, and generates verified executable control instructions.

[0021] The adaptive optimization and evolution module activates an adaptive optimization and adjustment mechanism when the system stability indicators fail to meet preset standards. It dynamically corrects the resource allocation scheme through a multi-objective optimization algorithm and establishes an intelligent decision support mechanism based on case reasoning.

[0022] The beneficial effects of this invention are as follows:

[0023] By constructing a dual mechanism of multi-source meteorological fusion and regional fluctuation prediction, the problem of insufficient perception of new energy power fluctuations by traditional methods is effectively solved. The physical information neural network is used to embed meteorological physical equations into the data fusion process, and the spatiotemporal graph attention network is combined to capture the propagation law of fluctuations between regions, realizing the transformation from single-point isolated prediction to regional collaborative prediction, which significantly improves the accuracy and foresight of fluctuation prediction.

[0024] By establishing a collaborative system for wave cluster identification and digital twin simulation, the spatial and temporal limitations of traditional methods in resource allocation are overcome. The intelligent clustering technology based on wave morphology similarity can accurately identify high-risk areas. Combined with the simulation of wave propagation path by digital twin model, a resource optimization allocation scheme under multiple time scales is constructed, realizing cross-regional and multi-resource collaborative mitigation.

[0025] By constructing a closed-loop system for control strategy verification and adaptive optimization, the problem of insufficient adaptive capability of traditional methods is solved. The control strategy is verified in multiple dimensions in a digital twin environment. By combining reinforcement learning and multi-objective optimization algorithms, the online adjustment and continuous optimization of the system strategy are realized, forming an intelligent decision-making mechanism with self-learning capabilities. Attached Figure Description

[0026] To better understand and implement this application, the technical solution is described in detail below with reference to the accompanying drawings.

[0027] Figure 1A flowchart illustrating the active power fluctuation collaborative mitigation method for a high-proportion distributed renewable energy power grid provided in Embodiment 1 of this application;

[0028] Figure 2 A flowchart illustrating step S2 in the active power fluctuation collaborative mitigation method for a high-proportion distributed renewable energy power grid provided in Embodiment 1 of this application;

[0029] Figure 3 A flowchart illustrating step S3 in the active power fluctuation collaborative mitigation method for a high-proportion distributed renewable energy power grid provided in Embodiment 1 of this application.

[0030] Figure 4 This is a schematic diagram of the active power fluctuation collaborative mitigation system for a high-proportion distributed renewable energy grid provided in Embodiment 2 of this application. Detailed Implementation

[0031] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, exemplary embodiments will be described in detail below, examples of which are illustrated in the accompanying drawings. In the following description, when referring to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0032] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0033] The following detailed description of the specific implementation methods, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided in detail.

[0034] Example 1

[0035] Please see Figures 1-3 This embodiment provides a method for collaborative mitigation of active power fluctuations in a high-proportion distributed renewable energy grid, including the following steps:

[0036] S1. Real-time collection of multi-regional weather data, including parameters such as wind speed, light intensity, and temperature, is achieved through satellite remote sensing and ground sensor networks. Meteorological physical equations are then embedded as constraints into the data fusion process using a physical information neural network, integrating them into a unified spatiotemporal sequence dataset to obtain a physically consistent weather change trend sequence.

[0037] Further, step S1 specifically includes:

[0038] Wide-area meteorological observation data (including cloud distribution, surface temperature, and aerosol concentration) are acquired through a satellite remote sensing platform, while meteorological parameters (covering wind speed, light intensity, and temperature) are collected at ground-based sensor networks. A unified coordinate framework is established using a spatiotemporal registration algorithm, and data quality is controlled through Kalman filtering technology to construct an initial dataset with a unified spatiotemporal reference.

[0039] Based on standardized datasets, fluid dynamics equations and thermodynamic laws are embedded as constraints into a neural network model; multi-source data fusion is achieved through a physics-guided training strategy to generate a standardized spatiotemporal sequence dataset that combines observational accuracy and physical consistency.

[0040] Based on a standardized spatiotemporal sequence dataset, a spatiotemporal convolutional network is used to extract multi-parameter joint distribution features; an anomaly detection mechanism based on physical laws is established. When parameter fluctuations exceed the range of physical interpretation, a graph neural network is used for feature reconstruction, and a physically verified spatiotemporal pattern dataset is output.

[0041] Based on spatiotemporal model datasets, a multi-scale trend analysis model is constructed, which decomposes features into physical process components at different scales. Then, based on physical constraints, component recombination and uncertainty propagation analysis are performed to generate weather change trend sequences with clear physical meaning and verified reliability.

[0042] The process of constructing an initial dataset with a unified spatiotemporal reference includes: establishing a unified coordinate framework through a spatiotemporal registration algorithm; utilizing a geographic information system to spatially align wide-area raster data acquired by satellite remote sensing with point data from ground sensors; mapping all data to a unified geographic grid and timestamp sequence; and addressing the heterogeneity issues of different source data in terms of resolution, sampling rate, and coverage. Kalman filtering is then employed for data quality control: meteorological parameters (such as wind speed and temperature) at each grid point are treated as state variables of a dynamic system. The values ​​at the current moment are predicted from the previous moment's state and then weighted and fused with the actual observed values ​​at the current moment. By recursively calculating the residuals between the predicted and measured values, the reliability of the observed data is dynamically evaluated, effectively filtering out outliers caused by sensor noise, transient interference, or transmission errors, and filling in temporary data gaps. The final output is an initial dataset with a unified spatiotemporal reference and controlled data quality.

[0043] The process of generating a standardized spatiotemporal sequence dataset includes: constructing a physical information network architecture, connecting differential equation operators to the output layer of the neural network to directly output physical field quantities such as fluid velocity, pressure, and temperature; subsequently, designing a physical constraint loss function, introducing a strong physical regularization term in addition to the traditional observation data fitting loss term, substituting the field quantities output by the network into the Navier-Stokes equations, the continuity equation, and the law of conservation of energy, calculating the residuals (i.e., the degree to which the equations are not satisfied), and using these residuals as the physical loss term to jointly weight and sum with the data loss term to form the total loss function; during the physics-guided training process, backpropagation is used to simultaneously minimize data errors and physical equation residuals, not only utilizing the data itself but also forcing the network learning results to comply with the underlying physical laws, thereby providing reliable inference basis from physical laws in regions with sparse data or noise, and finally generating a deeply fused multi-source meteorological analysis field that maintains physical consistency even under incomplete observations.

[0044] The output is a physically validated spatiotemporal pattern dataset, which includes: modeling the spatiotemporal data domain as a graph structure, where each geographical location's meteorological station or grid point serves as a graph node. Node features contain multi-parameter historical observation data for that point, while edges are constructed based on geographical distance, wind field correlation, or topological connectivity to quantify spatial dependencies. Utilizing the message passing mechanism of a graph neural network, each node updates its embedding representation by aggregating information from its neighboring nodes, thereby learning complex spatial correlation patterns. When a physical anomaly is detected in a node parameter, the GNN, with its powerful spatial relationship modeling capabilities, uses information from surrounding normal nodes to infer and reconstruct physically reasonable values ​​for the anomaly point. To ensure the physical authenticity of the reconstruction results, the initial reconstructed data output by the GNN is substituted again into simplified physical conservation equations (such as mass conservation or energy conservation) for residual calculation. If the residual exceeds the allowable range, it is fed back to the GNN for iterative optimization. The final output is a rigorously physically validated spatiotemporal pattern dataset that fills in data gaps, corrects anomalies, and maintains consistency in spatial relationships and physical laws.

[0045] Generating weather change trend sequences includes: employing adaptive signal decomposition algorithms (such as empirical mode decomposition or wavelet transform) to decompose the original complex spatiotemporal weather feature sequence into a series of intrinsic mode functions (EMFs) representing different time scales. High-frequency components correspond to rapidly changing processes such as local turbulence, mid-frequency components correspond to weather system evolution, and low-frequency components correspond to slowly changing processes such as the climate background field, thus decoupling the observational data into flow components with clear physical meaning. Subsequently, component recombination and uncertainty propagation analysis are performed based on physical constraints: according to the physical characteristics of the current weather system (such as quasi-geostrophic equilibrium or boundary layer similarity theory), constructing... By rationally selecting components and applying weighted fusion rules, physically incompatible noise modes are suppressed. Meanwhile, Monte Carlo simulation or ensemble transformation Kalman filtering methods are used to quantify the uncertainty of each component in the prediction process, and uncertainty propagation and synthesis are strictly performed according to physical laws. Finally, through a reliability verification mechanism, the recombined trend sequence is compared with independent observations or high-resolution numerical model results, and its confidence interval and skill score are calculated. This generates a rigorously verified weather change trend sequence that is both physically interpretable and contains quantitative uncertainty information, providing a highly reliable input for new energy power prediction.

[0046] Specifically, by constructing a multi-source meteorological data fusion system based on physical information neural networks, the problem of insufficient perception of new energy power fluctuations by traditional methods has been solved. This has enabled a transformation from simple data superposition to deep fusion driven by physical laws, significantly improving the accuracy and reliability of weather trend forecasting and providing high-quality input for new energy power forecasting, thus laying a solid foundation for subsequent fluctuation smoothing.

[0047] S2. Based on the weather change trend sequence, apply the spatiotemporal graph attention network to process time series features, extract inter-regional propagation patterns and potential chain reaction features, and determine the regional correlation fluctuation prediction model.

[0048] Further, step S2 specifically includes:

[0049] S21. Based on the weather change trend sequence and combined with historical weather data records, a regionalized spatiotemporal graph structure is constructed. In this structure, nodes represent geographical regions, node features are time series data of weather parameters, and edge weights are calculated based on the meteorological correlation and geographical distance between regions. Through sliding window sampling and Z-score standardization, a structured spatiotemporal graph dataset suitable for deep learning models is generated.

[0050] S22. Input the structured spatiotemporal graph dataset into the spatiotemporal graph attention network. In the spatial dimension, use the graph attention mechanism to dynamically learn the mutual influence weights between regions and identify key propagation paths. In the temporal dimension, capture the dynamic evolution of weather parameters through gated temporal convolution and quantify the correlation patterns between regions through the collaborative extraction of spatiotemporal features.

[0051] S23. Based on the extracted spatiotemporal features, a wave propagation inference model is constructed. Through an encoder-decoder architecture, the propagation process of weather waves in the regional network is simulated: the encoder maps the current spatiotemporal state to a potential representation, and the decoder recursively predicts the regional state at multiple future time steps based on the representation, thereby obtaining potential chain reaction paths.

[0052] S24. Establish a risk assessment mechanism based on prediction uncertainty. When the predicted volatility and uncertainty of a certain region exceed a preset threshold, it will be marked as a high-risk region, and the model will be activated for adaptive optimization: the graph attention weights will be adjusted through reinforcement learning strategy to prioritize the prediction accuracy of high-risk regions, and finally the predicted distribution of regional volatility after risk calibration will be output.

[0053] The construction of a regionalized spatiotemporal map structure includes: dividing the geographical area covered by the target power grid into several regular grid units or partitioning it according to natural geographical units such as administrative divisions and watershed boundaries. Each unit is defined as a graph node, and each node is assigned a feature, which is a multi-dimensional time-series tensor containing a sequence of physically consistent weather parameters for the region within a rolling time window, such as wind speed, light intensity, and temperature. Then, based on the geographic information system and historical meteorological data, the weights of the connecting edges between nodes are calculated: on the one hand, the spatial proximity weight is calculated based on the geographical distance decay using a Gaussian kernel function; on the other hand, the meteorological correlation weight is obtained by analyzing historical weather data to calculate the Pearson correlation coefficient or mutual information of key meteorological parameters between regions. Finally, these two weights are linearly or nonlinearly fused, and a dynamically weighted structured spatiotemporal map dataset suitable for deep learning models is generated through sliding window sampling and Z-score normalization techniques.

[0054] By collaboratively extracting spatiotemporal features, the correlation patterns between regions are quantified. This includes: In the spatial dimension, a graph attention mechanism is used to dynamically calculate the attention weights of all neighboring nodes for each target node. The features of each node are mapped to a high-dimensional space through a shared trainable weight matrix. Then, the correlation score between the features of the target node and its neighboring nodes is calculated using an attention function, and normalized using a Softmax function. This allows the model to automatically focus on key neighbors that have strong meteorological influences upstream or are highly correlated, accurately identifying key spatial paths for wave propagation. In the temporal dimension, in sync with spatial processing, a gated temporal convolutional network or a temporal attention mechanism is used to perform one-dimensional convolution operations along the time axis at each node. The gating mechanism can effectively capture the long-term dependencies and dynamic patterns of weather parameter evolution, such as the movement and evolution speed of weather systems. Through multi-layered stacked spatiotemporal graph attention layers, deep coupling between spatial correlation and temporal evolution is achieved, accurately quantifying the dynamic and nonlinear correlation patterns between regions.

[0055] A wave propagation inference model is constructed, including: using an encoder-decoder architecture to simulate the spatiotemporal propagation process of weather waves in a regional network; the encoder receives the spatiotemporal graph sequence of the past several time steps as input, and is composed of graph neural network layers and temporal coding layers alternately. It is responsible for abstracting and fusing the spatial dependencies (through graph attention mechanism) and temporal dynamic features (through GRU or TCN) of each time step layer by layer, and finally encodes the entire input sequence into a fixed-dimensional context vector that condenses the historical state of the system; the decoder performs multi-step forward inference based on this context vector: at each prediction time step, the decoder uses its internal state (initial state is the context vector) and the prediction output of the previous time step to reconstruct the prediction graph structure of the current step; subsequently, the graph is also processed by graph neural network and temporal network, on the one hand outputting the wave prediction values ​​of all regions at the current time step, and on the other hand updating its internal state to be passed to the next time step.

[0056] Specifically, by constructing a spatiotemporal graph attention network and an encoder-decoder architecture, the problem of traditional methods' difficulty in capturing the propagation patterns of regional fluctuations is effectively solved. A dynamic graph attention mechanism is used to quantify spatial correlations, and gated temporal convolution is combined to capture temporal evolution characteristics, achieving accurate prediction of regional fluctuation propagation paths and chain reactions. Through uncertainty-based risk assessment and adaptive optimization mechanisms, the accuracy and reliability of fluctuation prediction are significantly improved, providing crucial support for grid coordinated control.

[0057] S3. If the chain reaction characteristics in the regional correlation fluctuation prediction model exceed the preset threshold, then the dynamic time warping is used as a similarity measure for spectral clustering algorithm to group regions based on the morphological similarity of the fluctuation curves, thereby obtaining potential high-risk fluctuation clusters.

[0058] Further, step S3 specifically includes:

[0059] S31. When the intensity of the chain reaction characteristics output by the regional correlation fluctuation prediction model exceeds the preset threshold, the risk cluster identification process is automatically triggered to extract the complete fluctuation time series curve of new energy power in each region from the model output.

[0060] S32. The dynamic time warping algorithm is used as the core similarity measure to calculate the morphological distance between the fluctuation curves of each region. By constructing a fluctuation morphology similarity matrix, the similarity characteristics of different regions in fluctuation amplitude, phase and waveform are captured.

[0061] S33. Based on the DTW similarity matrix, the spectral clustering algorithm is used to group regions. Through Laplace matrix eigenvalue decomposition and k-means clustering, regions with similar wave patterns are automatically divided to obtain potential high-risk wave clusters with common wave characteristics.

[0062] S34. Taking into account the characteristics of fluctuation amplitude, duration and spatial clustering, the entropy weight-TOPSIS method is used to quantitatively assess the risk level of potential high-risk fluctuation clusters and generate a control priority sequence for partitioning and classification.

[0063] The process of capturing the similarities in amplitude, phase, and waveform across different regions involves extracting complete time-series fluctuation curves of renewable energy power from regional correlation fluctuation prediction models. These curves may exhibit phase differences (i.e., asynchronous peaks and troughs) on the time axis. The DTW algorithm constructs an Euclidean distance matrix between points in the two sequences and uses dynamic programming to find a curved path that minimizes the cumulative distance, thereby nonlinearly aligning the two sequences on the time axis. This alignment process eliminates the influence of time offset and directly compares the intrinsic similarities in amplitude and waveform (e.g., spikes and gentle slopes). By calculating the pairwise DTW distances between all regions, a symmetrical fluctuation morphology similarity matrix can be constructed. Each element in this matrix no longer represents the traditional point-to-point distance but rather the degree of difference in the overall shape of the two fluctuation curves, thus accurately capturing the deep similarities in amplitude, phase, and waveform across different regions.

[0064] Based on the DTW similarity matrix, a spectral clustering algorithm is used for region grouping. This includes: converting the DTW distance matrix into an affinity matrix (usually using a Gaussian kernel function) to characterize the similarity connection strength between regions; calculating the Laplacian matrix corresponding to this affinity matrix (usually using a symmetric normalized Laplacian matrix under the normalized cut criterion); and obtaining the eigenvectors corresponding to the k smallest eigenvalues ​​through eigenvalue decomposition. These eigenvectors constitute a new, low-dimensional feature space, transforming the complex nonlinear clustering structure implicit in the original DTW similarity matrix into a more easily separable linear structure in this space. Each region is mapped to a point in this new feature space, and the k-means clustering algorithm is used to partition this low-dimensional feature vector set, thereby completing the automatic grouping of regions with similar fluctuation patterns and obtaining several potential high-risk fluctuation clusters with common fluctuation characteristics.

[0065] The entropy-weighted TOPSIS method is used to quantitatively assess the risk level of potential high-risk volatility clusters. This includes: constructing a multi-indicator assessment system, including key characteristics such as volatility amplitude (e.g., maximum wave power), volatility duration (e.g., duration exceeding a threshold), and spatial clustering (e.g., electrical proximity within the cluster); objectively determining the weight of each indicator using the entropy-weighted method: by calculating the information entropy of each indicator across all clusters, and based on the principle that the smaller the entropy value, the greater the degree of variation of the indicator and the higher its weight, objective weights are assigned to each indicator to avoid subjective bias; then, a comprehensive assessment is performed using the TOPSIS method: by calculating the Euclidean distance between each assessment object (volatile cluster) and the positive ideal solution (the virtual optimal solution with the highest risk) and the negative ideal solution (the virtual worst solution with the lowest risk), relative proximity is used as the evaluation criterion. All volatility clusters are ranked according to relative proximity, generating a zoned and categorized control priority sequence, providing a decision-making basis for subsequent coordinated mitigation resource allocation.

[0066] Specifically, by using dynamic time warping algorithms and spectral clustering technology, the key technical challenge of identifying regional clusters with similar fluctuation patterns, which is difficult to achieve with traditional methods, has been effectively solved. It can accurately identify potential high-risk fluctuation clusters and generate scientific regional classification and control priorities, providing a key basis for subsequent precise allocation of mitigation resources and implementation of differentiated control strategies, and significantly improving the power grid's ability to identify and control regional fluctuation risks.

[0067] S4. Based on information on potential high-risk fluctuation clusters and historical power data, construct a digital twin model to simulate the propagation path of power fluctuations at different time scales and obtain a distribution map of resource coordination demand.

[0068] Further, step S4 specifically includes:

[0069] Based on information on potential high-risk fluctuation clusters, and combined with historical power data and power grid topology parameters, a digital twin model of the power grid is constructed. Through data cleaning and outlier correction techniques, a dynamic simulation dataset containing spatiotemporal characteristics is established to provide a high-fidelity virtual environment for fluctuation propagation simulation.

[0070] In a digital twin environment, a physical constraint-based propagation algorithm is used to simulate the propagation process of power fluctuations in the power grid topology at different time scales. The spatiotemporal dependence of fluctuation propagation is captured by a temporal convolutional network, and the propagation path and impact range of fluctuations between regions are predicted.

[0071] Based on the simulation results of wave propagation, a resource demand quantification model is established. By analyzing key indicators such as power deficit and voltage stability margin of each node, the intensity of resource coordination demand in different regions is calculated, and the analytic hierarchy process is used to determine the priority sequence of resource allocation.

[0072] Based on demand priority, a resource coordination demand distribution map is generated, and a dynamic update mechanism is established. When the fluctuation intensity of a certain area exceeds the preset threshold, the resource allocation ratio reconstruction algorithm is automatically triggered. Through reinforcement learning strategy, resource allocation is optimized to form a resource coordination scheme with adaptive capabilities.

[0073] The construction of a digital twin model of the power grid includes: integrating multi-source heterogeneous data through a data bus, the core of which is to integrate real-time measurement data (SCADA, PMU), static parameters (power grid topology, line impedance, transformer nameplate parameters) from the physical entity side of the power grid, as well as the spatial distribution and temporal characteristics of identified potential high-risk fluctuation clusters; using a deep neural network trained with historical power data and meteorological data to perform behavioral calibration on components that are difficult to model accurately, such as equipment aging characteristics and dynamic response of new energy power plants, to make up for the shortcomings of pure physical models; and constructing its digital mapping in virtual space based on graph theory and differential-algebraic equations. This mapping not only includes a static topology network composed of nodes and branches, but also achieves synchronous iterative updates with the physical power grid by introducing dynamic state estimators (such as the Kalman filter family). This forms a real-time synchronous virtual power grid system that is jointly constrained by data-driven and physical laws, and can simulate dynamic processes from transient to quasi-steady state at multiple time scales with high fidelity, providing a highly reliable simulation environment for subsequent fluctuation propagation simulation and resource collaborative optimization.

[0074] Predicting the propagation path and impact range of fluctuations across regions involves: using rigorous power grid physics laws as a foundation, embedding AC power flow equations and generator swing equations as hard constraints into the simulation kernel to ensure that the propagation of power fluctuations in the simulation strictly follows fundamental principles such as energy conservation and Ohm's law, thereby guaranteeing the physical realism and reliability of the simulation results. To overcome the shortcomings of pure physical models in computational efficiency and capturing complex nonlinear spatiotemporal relationships, a temporal convolutional network (TCN) is introduced as an intelligent enhancement: this network, through causal convolution and dilated convolution in the time dimension, can efficiently learn the causal relationships and long-term dependency patterns of fluctuation events in historical data, accurately capturing the evolution of fluctuations in the time dimension; simultaneously, by treating the power grid topology as a graph structure, the combination of TCN and graph neural networks can effectively learn the propagation inertia, attenuation characteristics, and chain reaction patterns of fluctuations along specific electrical paths in the spatial dimension. Ultimately, the hybrid model can extrapolate, at different time scales such as seconds, minutes, and hours, how power disturbances originating from high-risk fluctuation clusters will form specific transmission paths between regions along lines and through nodes based on the electrical characteristics and historical patterns of the power grid, and accurately quantify the impact range of power overruns, voltage deviations, frequency shifts, etc. on each region and line, providing a dynamic and predictable decision-making basis for subsequent precise resource coordination.

[0075] Specifically, by constructing a digital twin model that integrates physical constraints and data-driven approaches, and combining temporal convolutional networks with multi-timescale simulation technology, the technical challenge of accurately simulating the propagation path of power fluctuations using traditional methods has been effectively solved. This has enabled accurate prediction of the fluctuation propagation process and assessment of its impact range. Furthermore, by establishing a quantitative model of resource demand and a dynamic update mechanism, a scientific resource allocation scheme has been formed, significantly improving the power grid's ability to accurately control fluctuations and its resource utilization efficiency.

[0076] S5. Based on the specific requirements quantified by the resource coordination demand distribution map, as well as the power grid physical model and operation rules, a mixed integer programming model with the goal of minimizing total cost and transmission loss is constructed, and a column generation algorithm is used for efficient solution to calculate the real-time resource allocation scheme for backup power and energy storage equipment.

[0077] Further, step S5 specifically includes:

[0078] Based on the spatiotemporal demand data quantified by the resource coordination demand distribution map, and combined with power grid flow constraints, equipment operation limits and safety criteria, a mixed integer programming model is constructed with the goal of weighted minimization of total operating cost, transmission loss and reliability indicators, unifying discrete decision-making (equipment start-up and shutdown) and continuous variables (power allocation) within the optimization framework.

[0079] A column generation algorithm is used to decompose and solve large-scale mixed integer programming problems. The main problem handles resource allocation decisions, while the sub-problems dynamically generate feasible equipment combinations and transmission paths, which significantly reduces the solution complexity and ensures that an economically feasible near-optimal solution is obtained within a limited time.

[0080] The Dijkstra algorithm is embedded in the column generation solution process to optimize the transmission path in real time. At the same time, a hard constraint verification mechanism for node capacity and line transmission capacity is introduced to ensure that the obtained resource transfer path meets all physical constraints.

[0081] Based on the optimization results, a real-time allocation scheme for backup power and energy storage devices in each region is generated, including device switching sequence, power commands and transmission plans, and an executable list and contingency plan for handling anomalies are established.

[0082] The column generation algorithm is used to decompose and solve large-scale mixed-integer programming problems. This algorithm is an efficient solution strategy for large-scale mixed-integer programming problems. It decomposes the original problem into a main problem and several subproblems, and iteratively solves both to approximate the optimal solution. Specifically, the main problem handles macro-level resource allocation decisions. Initially, it contains only a finite set of feasible equipment combinations (called "columns"). By solving this simplified model, the current optimal solution and the dual variable (i.e., shadow price) are obtained. The subproblems (or pricing problems) dynamically search for and generate new equipment combinations and power transmission schemes (i.e., new "columns") that can further reduce the total cost based on the dual cost passed from the main problem. After this new column is added to the main problem, the main problem is solved again. This process is iterated until the subproblems can no longer generate any new columns that can reduce costs. At this point, the solution to the main problem is a high-quality near-optimal solution to the original large-scale problem. This approach avoids directly enumerating all possible combinations of devices and transmission paths. By generating data on demand, it significantly reduces computational complexity, making it possible to provide an economically feasible solution for real-time scheduling within a limited computation time.

[0083] The Dijkstra algorithm is embedded for real-time transmission path optimization. This involves a sub-problem modeling the power grid as a weighted graph, where nodes represent buses and edges represent transmission lines. The edge weights are determined by transmission losses, dual costs (from the main problem), and congestion levels. The Dijkstra algorithm is used to find the lowest-cost transmission path from each resource source (e.g., an activated backup power source) to a resource sink (i.e., a high-demand node). However, finding only the lowest-cost path is insufficient; a hard-constraint verification mechanism must be introduced. Before a path is ultimately adopted, the system verifies whether the transmission power of all lines along the path exceeds their transmission capacity limits, and checks whether the injected power of the nodes involved in the path is within the allowable range of node capacity (e.g., generator output limits, load demand). If the verification fails, the weight of the path or congested element is temporarily set to infinity, forcing the Dijkstra algorithm to find another feasible path that is both low-cost and satisfies all physical and security constraints. This ensures that every generated "column" (i.e., resource transfer scheme) is truly feasible.

[0084] Specifically, by constructing a mixed-integer programming model and using a column generation algorithm for efficient solution, the shortcomings of traditional resource allocation methods in terms of computational efficiency and global optimization are effectively addressed. Discrete device start-up and shutdown decisions and continuous power allocation are unified within the optimization framework. Real-time optimization of transmission paths is achieved by combining the Dijkstra algorithm, and the feasibility of the scheme is ensured through hard constraint verification. This breaks through the limitations of traditional local optimization and human experience, realizing the transformation from experience-based decision-making to intelligent decision-making. Under the premise of ensuring the safe operation of the power grid, resource utilization efficiency is significantly improved and the total operating cost of the system is reduced.

[0085] S6. Update the power grid control system parameters according to the determined real-time resource allocation scheme, simulate the execution of coordinated operations, obtain the system stability indicators after feedback, and determine whether iterative adjustments are needed.

[0086] Further, step S6 specifically includes:

[0087] The real-time resource partitioning scheme is converted into a set of control parameters and injected into the power grid digital twin system through a data bus. A synchronously updated simulation environment is built based on the real-time operating status of the power grid to ensure the consistency between the digital model and the physical system.

[0088] Perform graded loading tests in a digital twin environment, including transient stability simulation (second-level), voltage stability analysis (minute-level), and frequency stability assessment (hour-level), and accelerate the simulation process through parallel computing technology to obtain system dynamic response data in real time;

[0089] A pre-trained deep neural network model is used to analyze the simulation results in real time. Through feature extraction and pattern recognition, the relative position of the system's operating state and stability boundary is quickly determined, and the safety margin of the scheme is accurately evaluated.

[0090] When insufficient stability margin is detected, a reinforcement learning-based optimization algorithm is launched to automatically adjust the resource allocation strategy while ensuring the constraints. Through multiple iterations, an optimization scheme that meets all stability requirements is obtained, and the final executable control instructions are generated.

[0091] When the system stability reaches the preset standard, the current resource allocation scheme is locked, a complete set of control parameter configurations is generated, and the verified control strategies are packaged into an executable instruction set, ready to be sent to the actual power grid control system.

[0092] Among them, a pre-trained deep neural network model is used to analyze the simulation results in real time. This includes: after performing hierarchical simulation tests in the digital twin environment, a spatiotemporal evolution sequence containing tens of thousands of data points such as voltage, phase angle, frequency, and line power of each node will be generated. If traditional methods are used to analyze each criterion one by one, it will be extremely time-consuming. A pre-trained deep neural network model (such as a long short-term memory network or a convolutional neural network) is used to directly process these high-dimensional time series data. The model has been trained on a large scale in the offline stage using massive historical simulation data and known stability boundaries. It has learned to directly extract key features (such as voltage collapse precursors and frequency oscillation modes) from the system's dynamic response curve and map them to a comprehensive stability margin quantification index. When applied online, the DNN model can bypass complex numerical calculations and complete the slice-by-slice diagnosis of the simulation results in milliseconds, quickly determining the relative distance between the current operating state of the system and the instability boundary.

[0093] The algorithm employs a reinforcement learning-based optimization mechanism. When the DNN model diagnoses insufficient stability margin, an online reinforcement learning-based optimization algorithm is immediately triggered to dynamically adjust the resource allocation strategy. Here, the agent represents the decision-making system, the environment is a digital twin model of the power grid, the action is fine-tuning the power commands or switching states of backup power and energy storage devices, the state is the real-time operating condition of the power grid, and the reward is a function that comprehensively considers stability margin improvement, cost control, and constraint violation. The algorithm iterates continuously within the digital twin environment: the agent attempts an adjustment action, the environment provides new states and rewards, and the agent updates its strategy accordingly. This cycle continues until a new resource allocation strategy is found that significantly improves system stability (i.e., yields a high reward) while strictly guaranteeing all physical constraints (such as power flow limits and equipment capacity).

[0094] Specifically, by constructing a closed-loop verification system in a digital twin environment, the shortcomings of traditional methods in assessing the security of control strategies are effectively addressed. This achieves a shift from offline analysis to online verification and from manual judgment to intelligent decision-making, significantly improving the reliability and execution effectiveness of control schemes and providing strong support for the safe and stable operation of the power grid.

[0095] S7. If the system stability indicators do not meet the preset standards, the adaptive optimization and adjustment mechanism will be activated to dynamically correct the resource allocation scheme and reassess the system stability.

[0096] Further, step S7 specifically includes:

[0097] When the system stability indicators fail to meet the preset standards, based on the multi-dimensional simulation data such as transient stability, voltage stability and frequency stability output by the digital twin environment, a correlation graph between stability defects and resource allocation schemes is established. By introducing a dynamic risk propagation model, the chain reaction risks caused by insufficient resource allocation or unreasonable configuration are analyzed, the weak links and key influencing factors of system stability are identified, and a risk assessment report and improvement priority sequence are formed.

[0098] Based on the risk assessment results, a stability risk minimization objective is added to the original economic objective, a multi-objective optimization framework is established, and the improved NSGA-II algorithm is used to search for Pareto optimal solutions. Through dynamic constraint processing technology and adaptive crossover mutation operator, multiple non-dominated solutions that take into account both economy and stability are generated, forming a candidate solution set.

[0099] Based on the fuzzy comprehensive evaluation method, candidate solutions are comprehensively evaluated from dimensions such as the degree of stability improvement, the rate of economic retention, and the feasibility of implementation. At the same time, the parameter adjustment strategies, constraint handling experience, and solution evaluation results during the optimization process are accumulated into the system knowledge base, and a case-based intelligent decision support mechanism is established to provide a reference for the rapid optimization of similar scenarios in the future.

[0100] Among them, by introducing a dynamic risk propagation model, the risk of cascading failures caused by insufficient resource allocation or unreasonable configuration is analyzed. This includes: the dynamic risk propagation model abstracts the power grid as a complex network that evolves dynamically, where nodes are generators, loads, and critical buses, and edges are transmission lines with real-time power flow. When the digital twin simulation identifies an initial stability defect (such as a voltage dip in a certain area), the model does not treat it as an isolated event, but as an initial disturbance. Based on predefined fault propagation rules (for example, after a line trips due to overload, its power flow will transfer to adjacent lines, which may lead to new overloads), the model simulates the subsequent cascading failure development path in the virtual network with a time step of milliseconds / seconds. Through thousands of Monte Carlo simulations, the model can statistically determine which lines or transformers are most frequently involved in cascading failures, and which generator trip or load area instability will trigger the most severe system collapse.

[0101] The improved NSGA-II algorithm introduces two core enhancements: First, dynamic constraint handling, which uses an adaptive penalty function or feasibility rule to dynamically weigh the degree of constraint violation during evolution, thereby more intelligently utilizing effective information from infeasible solutions to guide the population towards the feasible region boundary; Second, adaptive crossover and mutation operators, whose crossover and mutation probabilities are no longer fixed but automatically adjusted based on the diversity of the current generation of the population (such as the dispersion of distances between individuals). When diversity is high, the probabilities are reduced to promote convergence, and when diversity is low, the probabilities are increased to escape local optima. These two improvements work together to significantly improve the algorithm's convergence speed, solution set uniformity, and global search capability when solving high-dimensional, strongly constrained power grid resource optimization problems.

[0102] Specifically, by constructing a multi-objective optimization framework and intelligent decision-making mechanism, the problem of lacking effective adjustment strategies when the system stability fails to meet the standards in traditional methods is effectively solved. Based on digital twin simulation data, a correlation graph between stability defects and resource allocation is established. An improved NSGA-II algorithm is used for multi-objective optimization. Through dynamic constraint processing and adaptive genetic operators, a Pareto optimal solution set that balances economy and stability is generated. The best solution is selected using a fuzzy comprehensive evaluation method. This realizes the transformation from single optimization to continuous evolution and from empirical decision-making to intelligent reasoning, significantly improving the system's adaptive capability and decision-making level, and providing reliable technical support for power grid stability control in complex operating scenarios.

[0103] Example 2

[0104] Please see Figure 4 This embodiment provides a collaborative active power fluctuation mitigation system for a high-proportion distributed renewable energy grid, used to implement a collaborative active power fluctuation mitigation method for a high-proportion distributed renewable energy grid, including:

[0105] The multi-source meteorological fusion module integrates satellite remote sensing and ground sensor data, uses spatiotemporal registration and Kalman filtering for quality control, and forms an initial dataset with a unified spatiotemporal benchmark. Fluid dynamics and thermodynamic equations are embedded as constraints into the neural network training to ensure the physical consistency of the data. Through multi-scale trend decomposition and uncertainty quantification, it generates weather change trend sequences with clear physical meaning and that have been verified, providing accurate input for new energy power prediction.

[0106] The regional fluctuation prediction module constructs a regionalized spatiotemporal graph structure, dynamically learns spatial correlation weights using a graph attention mechanism, captures temporal evolution patterns using gated temporal convolution, simulates the chain propagation path of weather fluctuations based on an encoder-decoder architecture, and establishes an uncertainty risk assessment mechanism. When the predicted fluctuation exceeds a threshold, the model's attention weights are automatically adjusted, and the predicted regional fluctuation distribution is output after risk calibration.

[0107] The fluctuation cluster identification module automatically extracts the power fluctuation curves of each region when strong chain reaction characteristics are detected. It constructs a morphological similarity matrix using a dynamic time warping algorithm, automatically groups regions with similar fluctuation patterns through spectral clustering to form high-risk fluctuation clusters, and quantifies the risk level by using the entropy weight-TOPSIS method to generate a control priority sequence for different regions.

[0108] The digital twin simulation module integrates high-risk cluster information, historical power data, and power grid topology parameters to establish a high-fidelity simulation platform. It uses physical constraint propagation algorithm and temporal convolutional network to simulate the propagation path and impact range of fluctuations at multiple time scales. By analyzing indicators such as node power deficit and voltage stability margin, it calculates the resource demand intensity of each region and uses the analytic hierarchy process to determine allocation priorities, generating a dynamically updated resource coordination demand distribution map.

[0109] The resource optimization decision module constructs a mixed integer programming model with the goal of minimizing total cost and transmission loss based on the distribution of resource demand. It uses a column generation algorithm for efficient solution, generates economically feasible equipment combinations through master-subproblem iteration, embeds Dijkstra's algorithm for real-time optimization of transmission paths during the solution process, and introduces a hard constraint verification mechanism. Finally, it generates a real-time allocation scheme and executable list of backup power and energy storage equipment.

[0110] The control strategy verification module injects the resource allocation scheme into the digital twin system, performs multi-timescale stability simulation through parallel computing, and uses a pre-trained deep neural network to analyze the simulation results in real time to quickly assess the safety margin. When the margin is insufficient, a reinforcement learning algorithm is activated to automatically adjust the resource allocation strategy. Through multiple iterations, an optimized scheme that meets the stability requirements is obtained, and the final executable control instructions are generated.

[0111] The adaptive optimization and evolution module, when stability does not meet the standard, establishes a defect-resource correlation map based on simulation data, identifies weak links through a dynamic risk propagation model, establishes a multi-objective optimization framework, searches for Pareto optimal solutions using an improved NSGA-II algorithm, selects the best solution using fuzzy comprehensive evaluation, and simultaneously stores the optimization process and results in a knowledge base, establishes a case reasoning mechanism, and provides intelligent decision support for subsequent optimization.

[0112] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for coordinated mitigation of active power fluctuations in high-proportion distributed renewable energy power grids, characterized by: Includes the following steps: S1. Real-time collection of multi-regional weather data through satellite remote sensing and ground sensor networks; embedding meteorological physical equations as constraints into the data fusion process through physical information neural networks; integrating into a unified spatiotemporal sequence dataset; obtaining a physically consistent weather change trend sequence. S2. Based on the weather change trend sequence, apply the spatiotemporal graph attention network to process time series features, extract inter-regional propagation patterns and potential chain reaction features, and determine the regional correlation fluctuation prediction model. S3. If the chain reaction characteristics in the regional correlation fluctuation prediction model exceed the preset threshold, then the dynamic time warping is used as a similarity measure for the spectral clustering algorithm to group the regions based on the morphological similarity of the fluctuation curves, thereby obtaining potential high-risk fluctuation clusters. S4. Based on information on potential high-risk fluctuation clusters and historical power data, construct a digital twin model to simulate the propagation path of power fluctuations at different time scales and obtain a distribution map of resource coordination demand. S5. Based on the specific requirements quantified by the resource coordination demand distribution map, as well as the power grid physical model and operation rules, a mixed integer programming model with the goal of minimizing total cost and transmission loss is constructed, and a column generation algorithm is used for efficient solution to calculate the real-time resource allocation scheme for backup power and energy storage equipment. S6. Update the power grid control system parameters through the determined real-time resource allocation scheme, simulate the execution of coordinated operations, obtain the system stability indicators after feedback, and determine whether iterative adjustments are needed. S7. If the system stability indicators do not meet the preset standards, the adaptive optimization and adjustment mechanism will be activated to dynamically correct the resource allocation scheme and reassess the system stability.

2. The method for coordinated active power fluctuation mitigation in a high-proportion distributed renewable energy power grid according to claim 1, characterized in that: Step S1 specifically includes: Wide-area meteorological observation data are acquired through satellite remote sensing platforms, while meteorological parameters at various locations are collected using ground sensor networks. A unified coordinate framework is established using a spatiotemporal registration algorithm, and data quality is controlled using Kalman filtering technology to construct an initial dataset with a unified spatiotemporal reference. Based on standardized datasets, fluid dynamics equations and thermodynamic laws are embedded as constraints into a neural network model; multi-source data fusion is achieved through a physics-guided training strategy to generate a standardized spatiotemporal sequence dataset that combines observational accuracy and physical consistency. Based on a standardized spatiotemporal sequence dataset, a spatiotemporal convolutional network is used to extract multi-parameter joint distribution features; an anomaly detection mechanism based on physical laws is established. When parameter fluctuations exceed the range of physical interpretation, a graph neural network is used for feature reconstruction, and a physically verified spatiotemporal pattern dataset is output. Based on spatiotemporal model datasets, a multi-scale trend analysis model is constructed, which decomposes features into physical process components at different scales. Then, based on physical constraints, component recombination and uncertainty propagation analysis are performed to generate weather change trend sequences with clear physical meaning and verified reliability.

3. The method for coordinated active power fluctuation mitigation in a high-proportion distributed renewable energy power grid according to claim 1, characterized in that: Step S2 specifically includes: S21. Based on the weather change trend sequence and combined with historical weather data records, a regionalized spatiotemporal graph structure is constructed. In this structure, nodes represent geographical regions, node features are time series data of weather parameters, and edge weights are calculated based on the meteorological correlation and geographical distance between regions. Through sliding window sampling and Z-score standardization, a structured spatiotemporal graph dataset for a deep learning model is generated. S22. Input the structured spatiotemporal graph dataset into the spatiotemporal graph attention network. In the spatial dimension, use the graph attention mechanism to dynamically learn the mutual influence weights between regions and identify key propagation paths. In the temporal dimension, capture the dynamic evolution of weather parameters through gated temporal convolution and quantify the correlation patterns between regions through the collaborative extraction of spatiotemporal features. S23. Based on the extracted spatiotemporal features, a wave propagation inference model is constructed. Through an encoder-decoder architecture, the propagation process of weather waves in the regional network is simulated: the encoder maps the current spatiotemporal state to a potential representation, and the decoder recursively predicts the regional state at multiple future time steps based on the representation to obtain potential chain reaction paths. S24. Establish a risk assessment mechanism based on prediction uncertainty. When the predicted volatility and uncertainty of a certain region exceed a preset threshold, it will be marked as a high-risk region, and the model will be activated for adaptive optimization: the graph attention weights will be adjusted through reinforcement learning strategy to prioritize the prediction accuracy of high-risk regions and output the predicted distribution of regional volatility after risk calibration.

4. The method for coordinated active power fluctuation mitigation in a high-proportion distributed renewable energy power grid according to claim 1, characterized in that: Step S3 specifically includes: S31. When the intensity of the chain reaction characteristics output by the regional correlation fluctuation prediction model exceeds the preset threshold, the risk cluster identification process is automatically triggered to extract the complete fluctuation time series curve of new energy power in each region from the model output. S32. The dynamic time warping algorithm is used as the core similarity measure to calculate the morphological distance between the fluctuation curves of each region. By constructing a fluctuation morphology similarity matrix, the similarity characteristics of different regions in fluctuation amplitude, phase and waveform are captured. S33. Based on the DTW similarity matrix, the spectral clustering algorithm is used to group regions. Through Laplace matrix eigenvalue decomposition and k-means clustering, regions with similar wave patterns are automatically divided to obtain potential high-risk wave clusters with common wave characteristics. S34. Taking into account the characteristics of fluctuation amplitude, duration and spatial clustering, the entropy weight-TOPSIS method is used to quantitatively assess the risk level of potential high-risk fluctuation clusters and generate a control priority sequence for partitioning and classification.

5. The method for coordinated active power fluctuation mitigation in a high-proportion distributed renewable energy power grid according to claim 4, characterized in that: Based on the DTW similarity matrix, a spectral clustering algorithm is used for region grouping, including: converting the DTW distance matrix into an affinity matrix to characterize the similarity connection strength between regions; calculating the Laplacian matrix corresponding to the affinity matrix; and obtaining the eigenvectors corresponding to the k smallest eigenvalues ​​through eigenvalue decomposition. The eigenvectors constitute a new, low-dimensional feature space, transforming the complex nonlinear clustering structure implicit in the original DTW similarity matrix into a linear structure in the space; mapping each region to a point in the new feature space; and using the k-means clustering algorithm to partition the low-dimensional feature vector set to complete the automatic grouping of regions with similar fluctuation patterns, resulting in several potential high-risk fluctuation clusters with common fluctuation characteristics.

6. The method for coordinated active power fluctuation mitigation in a high-proportion distributed renewable energy power grid according to claim 1, characterized in that: Step S4 specifically includes: Based on information on potential high-risk fluctuation clusters, and combined with historical power data and power grid topology parameters, a digital twin model of the power grid is constructed. Through data cleaning and outlier correction techniques, a dynamic simulation dataset containing spatiotemporal characteristics is established. In a digital twin environment, a physical constraint-based propagation algorithm is used to simulate the propagation process of power fluctuations in the power grid topology at different time scales. The spatiotemporal dependence of fluctuation propagation is captured by a temporal convolutional network, and the propagation path and impact range of fluctuations between regions are predicted. Based on the simulation results of wave propagation, a resource demand quantification model is established to calculate the intensity of resource coordination demand in different regions, and the analytic hierarchy process is used to determine the priority sequence of resource allocation. Based on demand priority, a resource coordination demand distribution map is generated, and a dynamic update mechanism is established. When the fluctuation intensity of a certain area exceeds the preset threshold, the resource allocation ratio reconstruction algorithm is automatically triggered. Through reinforcement learning strategy, resource allocation is optimized to form a resource coordination scheme with adaptive capabilities.

7. The method for coordinated active power fluctuation mitigation in a high-proportion distributed renewable energy power grid according to claim 1, characterized in that: Step S5 specifically includes: Based on the spatiotemporal demand data quantified by the resource coordination demand distribution map, and combined with power grid flow constraints, equipment operation limits and safety criteria, a mixed integer programming model is constructed with the goal of weighted minimization of total operating cost, transmission loss and reliability indicators, unifying discrete decision-making and continuous variables within the optimization framework. A column generation algorithm is used to decompose and solve large-scale mixed integer programming problems. The main problem handles resource allocation decisions, while the sub-problems dynamically generate feasible equipment combinations and transmission paths. The Dijkstra algorithm is embedded in the column generation solution process to optimize the transmission path in real time, and a hard constraint verification mechanism for node capacity and line transmission capacity is introduced. Based on the optimization results, a real-time allocation scheme for backup power and energy storage devices in each region is generated, including device switching sequence, power commands and transmission plans, and an executable list and contingency plan for handling anomalies are established.

8. The method for coordinated active power fluctuation mitigation in a high-proportion distributed renewable energy power grid according to claim 1, characterized in that: Step S6 specifically includes: The real-time resource partitioning scheme is converted into a set of control parameters and injected into the power grid digital twin system through the data bus. A synchronously updated simulation environment is built based on the real-time operating status of the power grid. Perform graded loading tests in a digital twin environment, including transient stability simulation, voltage stability analysis and frequency stability assessment, and accelerate the simulation process through parallel computing technology to obtain system dynamic response data in real time; A pre-trained deep neural network model is used to analyze the simulation results in real time. Through feature extraction and pattern recognition, the relative position of the system's operating state and stability boundary is quickly determined, and the safety margin of the scheme is evaluated. When insufficient stability margin is detected, a reinforcement learning-based optimization algorithm is launched to automatically adjust the resource allocation strategy while ensuring the constraints. Through multiple iterations, an optimization scheme that meets all stability requirements is obtained, and the final executable control instructions are generated. When the system stability reaches the preset standard, the current resource allocation scheme is locked, a complete set of control parameter configurations is generated, and the verified control strategies are packaged into an executable instruction set, ready to be sent to the actual power grid control system.

9. The method for coordinated suppression of active power fluctuations in a high-proportion distributed renewable energy power grid according to claim 1, characterized in that: Step S7 specifically includes: When the system stability indicators fail to meet the preset standards, a correlation graph between stability defects and resource allocation schemes is established based on the multi-dimensional simulation data output by the digital twin environment. By introducing a dynamic risk propagation model, the chain reaction risks caused by insufficient resource allocation or unreasonable configuration are analyzed, the weak links and key influencing factors of system stability are identified, and a risk assessment report and improvement priority sequence are formed. Based on the risk assessment results, a stability risk minimization objective is added to the original economic objective, a multi-objective optimization framework is established, and the improved NSGA-II algorithm is used to search for Pareto optimal solutions. Through dynamic constraint processing technology and adaptive crossover mutation operator, multiple non-dominated solutions that take into account both economy and stability are generated, forming a candidate solution set. Based on the fuzzy comprehensive evaluation method, candidate solutions are comprehensively evaluated. At the same time, the parameter adjustment strategies, constraint handling experience and solution evaluation results in the optimization process are accumulated into the system knowledge base, and a case-based intelligent decision support mechanism is established.

10. A system for coordinated active power fluctuation mitigation in a high-proportion distributed renewable energy grid, applied to the method for coordinated active power fluctuation mitigation in a high-proportion distributed renewable energy grid as described in any one of claims 1-9, characterized in that: include: The multi-source meteorological fusion module integrates satellite remote sensing data and ground sensor networks, and uses physical information neural network technology to embed meteorological physical equations as constraints into the data fusion process to generate physically consistent weather change trend sequences. The regional fluctuation prediction module constructs a regional spatiotemporal map structure based on weather change trend sequences, applies a spatiotemporal map attention network to extract inter-regional propagation patterns and chain reaction characteristics, establishes a regionally related fluctuation prediction model, and outputs a risk-calibrated fluctuation prediction distribution. The fluctuation cluster identification module, when the detected chain reaction characteristics exceed the preset threshold, uses dynamic time warping algorithm and spectral clustering method to group regions based on the similarity of fluctuation curve morphology, identify potential high-risk fluctuation clusters, and generate a control priority sequence for regional classification. The digital twin simulation module combines high-risk fluctuation cluster information and power grid topology parameters to construct a digital twin model of the power grid, simulate the propagation path of power fluctuations at different time scales, and generate a resource coordination demand distribution map through quantitative analysis of resource demand. The resource optimization decision module, based on the resource coordination demand distribution map, constructs a mixed integer programming model with the goal of minimizing total cost and transmission loss, and uses a column generation algorithm for efficient solution to calculate the real-time resource allocation scheme for backup power and energy storage equipment. The control strategy verification module injects the resource allocation scheme into the digital twin system, simulates the execution of coordination operations, uses a deep neural network to evaluate the system stability indicators in real time, and generates verified executable control instructions. The adaptive optimization and evolution module activates an adaptive optimization and adjustment mechanism when the system stability indicators fail to meet preset standards. It dynamically corrects the resource allocation scheme through a multi-objective optimization algorithm and establishes an intelligent decision support mechanism based on case reasoning.

Images