A Method and System for Identifying Extreme Wind Power Climbing Events under Regional Coupled Spatiotemporal Evolution
By constructing a spatiotemporal coupling graph sequence of regional wind farms and a dual-stream adaptive spatiotemporal supergraph network, the problem of the spatiotemporal coupling relationship of wind farms not being considered in existing technologies is solved. This enables accurate identification of extreme ramping events and full-dimensional risk warning, improving the accuracy of wind farm power prediction and grid security.
Patent Information
- Application Number
- CN202611285534.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-24
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies fail to effectively integrate the spatiotemporal coupling relationship between regional wind farms, resulting in insufficient accuracy in predicting extreme power fluctuations and a lack of confidence in determining extreme ramp events, making it difficult to meet the diverse risk warning needs of refined grid dispatch.
By acquiring historical power data and weather forecast data of each wind farm in the target area, an initial adjacency matrix is constructed. Weighted correction is performed based on time-delay cross-correlation to generate a dynamic spatiotemporal coupled graph sequence. Spatiotemporal fusion features are extracted using a dual-stream adaptive spatiotemporal hypergraph network, and the prediction quantiles and confidence of extreme ramp events are calculated. The dynamic ramp threshold is then used for judgment.
It significantly improves the accuracy and robustness of identifying extreme wind power ramp events, and provides multi-dimensional quantitative outputs of event confidence, intensity range, and occurrence time window, supporting grid dispatch departments in formulating differentiated risk response strategies to ensure the safe and stable operation of the power grid.
Smart Images

Figure CN123070403A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wind farm extreme event identification technology, specifically to a method and system for identifying wind farm extreme ramping events under regional coupled spatiotemporal evolution. Background Technology
[0002] With the continuous increase in the penetration rate of new energy sources, wind power generation is significantly affected by weather systems, resulting in power fluctuations. Extreme wind power ramp-up events can easily impact grid frequency stability, line transmission capacity, and dispatch plans, and even trigger safety accidents. Therefore, accurate identification and early warning of these events have become a critical requirement for the safe operation of the power grid. Existing technologies mostly analyze ramp-up events based on historical power data from a single wind farm, failing to fully consider the spatiotemporal coupling relationship between wind farms within the target area caused by wind energy propagation. This leads to insufficient accuracy in predicting extreme power fluctuations. Furthermore, traditional methods often rely on fixed thresholds to determine events, lacking quantitative outputs on the confidence level of the prediction results, the range of event intensity, and the occurrence time window, making it difficult to meet the diverse risk warning needs of refined power grid dispatch.
[0003] In summary, existing technologies have technical problems such as difficulty in combining the spatiotemporal coupling relationship of regional wind farms, resulting in insufficient accuracy in predicting extreme power fluctuations and a lack of confidence in determining extreme ramp events. Summary of the Invention
[0004] To overcome the aforementioned deficiencies in the prior art, the purpose of this application is to provide a method and system for identifying extreme wind power ramping events under regional coupled spatiotemporal evolution. This addresses the technical problems in the prior art, such as the difficulty in combining the spatiotemporal coupling relationship of regional wind farms, leading to insufficient accuracy in extreme power fluctuation prediction and a lack of confidence in the determination of extreme ramping events. The goal is to achieve accurate identification of extreme wind power ramping events and comprehensive risk warning. To this end, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for identifying extreme wind power ramping events under regional coupled spatiotemporal evolution, which includes the following steps: S100: Obtain historical power data and historical weather forecast data for each wind farm within the target area; S200: Based on the historical weather forecast data, calculate the propagation path and theoretical time delay of wind energy between any two wind farms, and construct an initial adjacency matrix with wind farms as nodes; S300: Calculate the time-delay cross-correlation between the power data of each wind farm based on the historical power data, and perform weighted correction on the initial adjacency matrix according to the time-delay cross-correlation calculation results to generate a dynamically evolving spatiotemporal coupling graph sequence. S400: Input the spatiotemporal coupling graph sequence and the node features of each wind farm into a pre-trained dual-stream adaptive spatiotemporal supergraph network, extract spatiotemporal fusion features, and obtain the predicted quantiles of the power of each wind farm in the future time period based on the spatiotemporal fusion features. The predicted quantiles include the 10th quantile, the 50th quantile and the 90th quantile. S500: Calculate the corresponding power ramp rate based on the predicted quantile, and determine whether an extreme ramp event has occurred by combining the dynamic ramp threshold. At the same time, output the confidence level, intensity range and expected occurrence time window of the event.
[0005] This technical solution effectively overcomes the shortcomings of existing technologies that rely solely on historical data from a single wind farm and neglect the spatiotemporal coupling relationship between wind farms in a region due to wind energy propagation, resulting in insufficient accuracy in predicting extreme power fluctuations. Through a dual fusion modeling approach combining physical mechanisms and data-driven methods, it significantly improves the accuracy and robustness of wind farm power prediction. Furthermore, based on a multi-quantile ramp rate calculation and dynamic threshold determination mechanism, it not only achieves high-precision identification of extreme ramp events but also provides, for the first time, multi-dimensional quantitative outputs of event confidence, intensity range, and occurrence time window. This solves the problem of traditional fixed threshold determination methods lacking confidence assessment, providing a more comprehensive and reliable decision-making basis for grid dispatch departments to formulate differentiated risk response strategies, and effectively ensuring the safe and stable operation of the power grid under high-proportion wind power integration.
[0006] As a preferred technical means, step S200 specifically includes: S210: For any two wind farms within the target area, denoted as the first wind farm and the second wind farm respectively, based on the wind direction field at time t of the historical weather forecast data, determine whether there is a wind energy propagation path from the first wind farm to the second wind farm. S220: If the propagation path is determined to exist, the theoretical time delay Δt for wind energy to propagate from the first wind farm to the second wind farm is calculated based on the geographical distance between the first wind farm and the second wind farm and the average wind speed corresponding to the propagation path. S230: Using the first wind farm and the second wind farm as nodes, and the theoretical time delay Δt as the core weight of the directed edge from the first wind farm to the second wind farm, construct the corresponding directed edge; S240: Traverse all wind farm node pairs within the target area, complete the construction of all directed edges, and generate the initial adjacency matrix.
[0007] This technical solution is based on the physical mechanism of wind energy propagation. Using wind direction as the core criterion, it determines the propagation path between wind farms. Combining geographical distance and average wind speed to quantify theoretical time delays, it constructs a directed graph topology, giving the initial adjacency matrix clear physical interpretability. This restores the temporal and spatial correlation logic of wind energy propagation within the region, laying an accurate and reliable prior topological foundation for subsequent modeling of the spatiotemporal coupling relationship of wind farms. It characterizes the directionality and time delay features of wind energy propagation at the meteorological mechanism level, providing a physically consistent structured foundation for the subsequent construction of the spatiotemporal coupling graph. This avoids the spurious correlations that may be introduced by relying solely on data correlation, improving the physical interpretability and accuracy of regional wind farm coupling relationship modeling.
[0008] As a preferred technical means, step S300 specifically includes: S310: Divide the historical power data into multiple consecutive sub-time slices according to a preset time interval to obtain the wind farm power data under each sub-time slice; S320: For each sub-time slice, perform the following operations: S321: Calculate the cross-correlation coefficient ρ(τ) of the historical power sequences of any two wind farms within this sub-time slice under time delay τ; S322: Find the time delay τ_max that maximizes the cross-correlation coefficient ρ(τ), compare the time delay τ_max with the theoretical time delay Δt of the corresponding node pair, and calculate the absolute difference between the two. S323: Determine the weighting factor based on the comparison result between the absolute difference and the preset tolerance threshold, and fuse the weighting factor with the weight of the corresponding directed edge in the initial adjacency matrix to obtain the fused edge weight set under the sub-time slice; S324: Based on the fused edge weight set, correct the initial adjacency matrix and generate the dynamic spatiotemporal coupling graph corresponding to the sub-time slice; S330: Sequentially combine the dynamic spatiotemporal coupling diagrams corresponding to all sub-time slices in chronological order to obtain the dynamic evolution spatiotemporal coupling diagram sequence.
[0009] This technical solution achieves a refined, time-segmented characterization of the spatiotemporal coupling relationship between wind farms by dividing the data into sub-time slices. Based on the time-delay cross-correlation of power sequences, it mines the actual power correlation patterns between wind farms, thereby weighting and correcting the prior topology. This accurately captures the time-varying characteristics of the coupling relationship between wind farms, solving the problem that traditional static topologies cannot adapt to the dynamic changes in wind energy propagation. It preserves the temporal logic of wind energy propagation while reflecting the dynamic correlation patterns of actual power fluctuations. The resulting spatiotemporal coupling graph sequence more realistically depicts the evolution of the coupling relationship between regional wind farms over time, providing a high-quality dynamic graph structure input for subsequent spatiotemporal feature extraction.
[0010] As a preferred technical approach: the weighting factor is determined based on the comparison between the absolute difference and the preset tolerance threshold, specifically as follows: If the absolute difference is less than or equal to the preset tolerance threshold, then the weighting factor is determined to be the maximum value of the cross-correlation coefficient ρ(τ); If the absolute difference is greater than the preset tolerance threshold, the weighting factor is determined as the product of the reverse inhibition coefficient and the maximum value of the cross-correlation coefficient ρ(τ), wherein the reverse inhibition coefficient is a value less than 1 and is inversely proportional to the absolute difference.
[0011] This technical solution sets a tolerance threshold and determines the weighting factor based on the degree of agreement between the data-driven time delay and the theoretical time delay. When the two are in agreement, the maximum cross-correlation coefficient is directly used to strengthen the high-confidence association. When deviations exist, an inverse suppression coefficient, inversely proportional to the degree of deviation, is introduced to weaken the weights. This effectively reduces the adverse impact of time delay deviations caused by local meteorological disturbances, terrain influences, and other factors on the graph structure edge weights, ensuring the rationality and robustness of the edge weights in the spatiotemporal coupled graph, thereby improving the quality of input data for subsequent extreme slope event prediction models.
[0012] As a preferred technical means: the dual-stream adaptive spatiotemporal hypergraph network includes a temporal branch and a spatial branch, wherein the temporal branch uses causal dilated convolution to capture long-range extreme temporal dependencies, and the spatial branch uses hypergraph convolution to mine multi-scale regional coupling patterns; in step S400, the extraction of spatiotemporal fusion features specifically includes: S410: Input the spatiotemporal coupling graph sequence and the node features of each wind farm into the temporal branch, and capture the temporal dependency features at different time scales through multiple parallel causal-hole convolutional layers with different hole rates and splice them to obtain the multi-scale temporal features of each wind farm node. S420: Input the spatiotemporal coupling graph sequence and the node features of each wind farm into the spatial branch, mine the multi-scale spatial coupling pattern of the regional wind farm through hypergraph convolution operation, and output the spatial features of each wind farm node after hypergraph convolution enhancement. S430: Adaptively weightedly fuse the multi-scale temporal features and spatial features corresponding to each wind farm node to generate the spatiotemporal fusion feature.
[0013] This technical solution adopts a dual-stream parallel network architecture, extracting features of corresponding dimensions independently through temporal and spatial branches, thus avoiding the feature interference problem caused by serial architecture. Causal dilated convolution can capture the long-range temporal dependence of extreme ramp events without introducing future information. Hypergraph convolution breaks through the limitations of traditional graph convolution and can effectively mine high-order spatial coupling relationships between multiple wind farms. Finally, the spatiotemporal fusion features generated by adaptive fusion have both temporal dynamics and spatial correlation, which greatly improves the accuracy and effectiveness of feature expression, providing richer and more accurate feature expressions for accurate prediction of multiquantile power, and effectively improving the ability to perceive extreme power fluctuation events.
[0014] As a preferred technical means: Step S400 obtains the predicted quantiles of the power of each wind farm in the future time period, specifically including: S440: Determine the preset target quantiles as the 10th quantile, 50th quantile, and 90th quantile; S450: Construct a regression output layer for a dual-stream adaptive spatiotemporal hypergraph network. The regression output layer is based on an asymmetric quantile regression loss function and is pre-trained using the preset target quantile. S460: The pre-trained regression output layer performs nonlinear mapping analysis on the spatiotemporal fusion features to generate the 10th, 50th, and 90th quantile prediction results of the power of each wind farm in the future time period.
[0015] This technical solution sets the 10th, 50th, and 90th quantiles as prediction targets and pre-trains the regression output layer using an asymmetric quantile regression loss function. This allows the network to simultaneously output the central trend and extreme fluctuation boundaries of future power. Extending traditional point prediction to probabilistic interval prediction not only provides the most likely power change scenarios but also quantifies the power fluctuation range under pessimistic and optimistic extreme scenarios. This provides a quantitative basis for subsequently determining extreme power surge events from multiple scenario perspectives, enhancing the comprehensiveness of event identification and risk warning capabilities.
[0016] As a preferred technique: Step S450 trains the regression output layer using an asymmetric quantile regression loss function, specifically including: A training sample set covering scenarios of normal power fluctuations and extreme ramping events was collected. The training sample set includes historical spatiotemporal coupling graph sequences, wind farm node characteristics, and actual power values for the corresponding time periods. A network structure for the regression output layer is constructed. The spatiotemporal coupling graph sequence of the training sample set and the node features of the wind farm are used as inputs, and the actual power value of the corresponding time period is used as the label for training. The quantile loss between the predicted value and the true value under each preset target quantile is calculated. Calculate the instantaneous ramp rate at each time point based on historical power data in the training sample set; When the absolute value of the instantaneous ramp rate exceeds the preset sensitivity threshold, a penalty weight greater than 1 is assigned to the quantile loss at that time point. An asymmetric quantile regression loss function is constructed by weighting each quantile loss with its corresponding penalty weight, and the regression output layer is iteratively trained until convergence.
[0017] This technical solution identifies extreme climbing moments in historical data by calculating the instantaneous climbing rate during the training of the regression output layer, and applies additional penalty weights to the prediction loss of these key moments. This makes the model more sensitive to extreme climbing event samples during training, enabling it to prioritize learning feature mapping patterns under extreme scenarios. This significantly improves the model's prediction accuracy for future extreme power fluctuations (especially climbing events), overcoming the deficiency of traditional loss functions that treat regular and extreme samples equally, leading to insufficient prediction ability for extreme events.
[0018] As a preferred technical means, step S500 specifically includes: S510: Obtain the dynamic ramp threshold that adapts to the power grid operating conditions and wind farm operating characteristics of the target area. S520: Calculate the baseline ramp rate R50 based on the predicted power of the 50th quantile, calculate the pessimistic ramp rate R10 based on the predicted power of the 10th quantile, and calculate the optimistic ramp rate R90 based on the predicted power of the 90th quantile. S530: Extreme climbing events are determined according to the following logic: If the absolute value of the baseline ramp rate R50 exceeds the dynamic ramp threshold, it is determined to be a high-confidence extreme ramp event. If the absolute value of the baseline ramp rate R50 does not exceed the dynamic ramp threshold, but the absolute value of the pessimistic scenario ramp rate R10 or the optimistic scenario ramp rate R90 exceeds the dynamic ramp threshold, it is determined to be a low-confidence, high-risk extreme ramp event. If none of the above conditions are met, it is determined that there are no extreme climbing events. S540: Based on the judgment result, synchronously output the confidence level of the event, the range of event intensity represented by the numerical ranges of R10, R50, and R90, and the expected occurrence time window of the event.
[0019] This technical solution, based on multiquantile prediction results, calculates the power ramp-up rate under baseline, pessimistic, and optimistic scenarios, and performs tiered judgment by combining dynamic ramp-up thresholds. It can not only identify extreme ramp-up events with high certainty (high confidence), but also provide early warnings of potential events that, while the baseline scenario is normal, pose a risk of severe fluctuations under extreme scenarios (low confidence, high risk), while simultaneously outputting the confidence level, intensity range, and expected time window. This tiered judgment and multi-dimensional information output mechanism provides power grid dispatching departments with comprehensive decision support, from "certainty warning" to "probability warning," which helps in formulating differentiated risk response strategies and improving the safety and economy of power grid operation.
[0020] Secondly, this invention provides a wind power extreme ramping event identification system under regional coupled spatiotemporal evolution, comprising: The data acquisition module is used to acquire historical power data and historical weather forecast data of each wind farm in the target area; The initial adjacency matrix construction module is used to calculate the propagation path and theoretical time delay of wind energy between wind farms based on the historical weather forecast data, and to construct an initial adjacency matrix with wind farms as nodes. The spatiotemporal coupling graph sequence generation module is used to calculate the time-delay cross-correlation between the power data of each wind farm based on the historical power data, perform weighted correction on the initial adjacency matrix, and generate a dynamically evolving spatiotemporal coupling graph sequence. The spatiotemporal fusion feature extraction module is used to input the spatiotemporal coupling graph sequence and the node features of each wind farm into a pre-trained dual-stream adaptive spatiotemporal hypergraph network, extract spatiotemporal fusion features, and obtain the predicted quantiles of the power of each wind farm in the future time period. The predicted quantiles include the 10th quantile, the 50th quantile, and the 90th quantile. The extreme ramp event determination module is used to calculate the power ramp rate based on the predicted quantile, and determine whether an extreme ramp event has occurred by combining the dynamic ramp threshold. At the same time, it outputs the confidence level, intensity range and expected occurrence time window of the event.
[0021] This technical solution achieves fully automated processing from data acquisition, spatiotemporal coupling relationship modeling, multiquantile power prediction to the hierarchical determination of extreme slope events through the organic combination of a data acquisition module, an initial adjacency matrix construction module, a spatiotemporal coupling graph sequence generation module, a spatiotemporal fusion feature extraction module, and an extreme slope event determination module. Taking a regional wind farm cluster as the analysis object, the solution integrates the initial adjacency matrix construction module and the spatiotemporal coupling graph sequence generation module to fuse wind energy propagation meteorological laws and actual power operation data, thus characterizing the dynamic spatiotemporal coupling relationship between wind farms in the region. This solves the core problem of traditional wind power slope recognition systems that only analyze single wind farms and ignore the correlation characteristics of power fluctuations between wind farms, resulting in missed detections and high false detection rates of extreme events. This significantly improves the accuracy of wind power extreme slope event identification. Unlike traditional identification systems that can only output a single judgment result of "whether a ramp event has occurred", this system, through an extreme ramp event judgment module, can simultaneously output three core early warning information: the confidence level of the event, the intensity range, and the expected occurrence time window. This provides power grid dispatching departments with full-dimensional decision support covering risk probability, impact degree, and occurrence time, and can directly serve core businesses such as power grid power balance regulation and operation risk prevention and control.
[0022] As a preferred technical means: the dual-stream adaptive spatiotemporal hypergraph network includes a temporal branch and a spatial branch; the temporal branch uses multiple parallel causal dilated convolutional layers with different dilation rates to capture long-range extreme temporal dependency features at different time scales; the spatial branch uses hypergraph convolutional layers to mine multi-scale high-order spatial coupling patterns between wind farms in the region. The spatiotemporal fusion feature extraction module has a built-in pre-trained convergent regression output layer. The regression output layer is pre-trained with an asymmetric quantile regression loss function and is used to perform nonlinear mapping analysis on the input spatiotemporal fusion features, and output the prediction results of the 10th quantile, 50th quantile and 90th quantile of the power of each wind farm in the future time period. The extreme ramp event determination module is specifically used to calculate the baseline ramp rate R50, the pessimistic scenario ramp rate R10, and the optimistic scenario ramp rate R90 based on the prediction results of the 10th quantile, 50th quantile, and 90th quantile, respectively, and execute the following determination logic: If the absolute value of the baseline ramp rate R50 exceeds the dynamic ramp threshold, it is determined to be a high-confidence extreme ramp event. If the absolute value of the baseline ramp rate R50 does not exceed the dynamic ramp threshold, but the absolute value of the pessimistic scenario ramp rate R10 or the optimistic scenario ramp rate R90 exceeds the dynamic ramp threshold, it is determined to be a low-confidence, high-risk extreme ramp event. If none of the above conditions are met, it is determined that there is no extreme climbing event.
[0023] This technical solution refines the internal execution logic and network structure of the system's core modules. It enhances the system's ability to extract the spatiotemporal coupling characteristics of wind farms through a dedicated dual-stream network architecture, ensures the accuracy of power quantile prediction through a pre-trained regression output layer, and standardizes the identification rules for extreme events through clear hierarchical judgment logic. This further improves the stability of system operation, the reliability and consistency of identification results, and enables the stable output of accurate and comprehensive early warning information for extreme wind power ramping events.
[0024] Beneficial effects: 1. Breaking through the limitations of existing technologies that rely solely on wind farm analysis, this technology constructs a dual-constraint modeling system driven by "meteorological and physical mechanisms + operational data," accurately depicting the dynamic evolution characteristics of the spatiotemporal coupling relationship between regional wind farms. This fundamentally solves the core pain point of insufficient accuracy in predicting extreme power fluctuations caused by neglecting the inter-site correlation characteristics of existing technologies, significantly improving the accuracy and robustness of wind farm power prediction.
[0025] 2. A multi-quantile probability prediction system was constructed, upgrading the traditional single-point power prediction to a full-range probability prediction covering the baseline trend, pessimistic extreme boundary, and optimistic extreme boundary. Combined with the hierarchical judgment logic of dynamic ramp threshold, high-precision identification of extreme ramp events was achieved. At the same time, multi-dimensional quantitative output of event confidence, intensity range, and occurrence time window was completed, solving the problem of lack of confidence assessment in the traditional fixed threshold judgment method. This provides a comprehensive and reliable decision-making basis for power grid dispatching departments to formulate differentiated risk response strategies.
[0026] 3. A dual-stream adaptive spatiotemporal hypergraph network architecture was designed. By using parallel temporal and spatial branches to independently extract features of corresponding dimensions, the problem of mutual interference of features caused by serial architecture is avoided. Causal dilated convolution can accurately capture the long-range temporal dependency features of extreme climbing events without introducing future information. Hypergraph convolution breaks through the limitations of traditional graph convolution and can effectively explore the high-order spatial coupling relationship between multiple wind farms, which greatly improves the accuracy of feature representation and the ability to perceive extreme power fluctuation events.
[0027] 4. During the model training phase, extreme climbing scenarios are labeled by instantaneous climbing rate, and additional penalty weights are applied to the prediction loss of extreme scenarios. This effectively solves the defects of the model learning dominated by a large number of regular operating samples and insufficient learning of extreme event features in small samples in conventional training. It significantly enhances the model's sensitivity and fitting accuracy to extreme climbing events in wind power and greatly reduces the probability of missed or misjudged extreme events. Attached Figure Description
[0028] Figure 1 This is a schematic diagram of the process of the present invention.
[0029] Figure 2This is a schematic diagram illustrating the process of constructing the initial adjacency matrix with wind farms as nodes in this invention.
[0030] Figure 3 This is a schematic diagram of the wind power extreme slope climbing event identification system under regional coupled spatiotemporal evolution provided in an embodiment of the present invention.
[0031] Figure labeling: 10, Data acquisition module; 20, Initial adjacency matrix construction module; 30, Spatiotemporal coupling graph sequence generation module; 40, Spatiotemporal fusion feature extraction module; 50, Extreme climbing event determination module. Detailed Implementation
[0032] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.
[0033] Example 1 like Figure 1 As shown, this embodiment provides a method for identifying extreme wind power ramping events under regional coupled spatiotemporal evolution, which includes the following steps: S100: Obtain historical power data and historical weather forecast data for each wind farm within the target area.
[0034] Specifically, the scope of the target area needs to be clearly defined, that is, the geographical area corresponding to the wind farm cluster covered by the extreme uphill event needs to be identified to ensure that all wind farms in the area are included in the data collection scope; then, for each wind farm in the area, two types of core data are collected simultaneously: historical power data and historical weather forecast data.
[0035] Historical power data must include real-time power output records of each wind farm over continuous time periods. The time granularity must meet the requirements of subsequent time-delay cross-correlation calculations and time-series feature extraction, such as minute-level or hour-level continuous power values. It must also cover different seasons and weather conditions, such as operational data under strong winds, light winds, and extreme weather conditions, to reflect the dynamic changes in power output. Historical weather forecast data must fully match the time dimension of historical power data, covering key meteorological parameters such as wind direction distribution, average wind speed, and air pressure within the corresponding time period. These parameters are the core basis for subsequent calculations of wind energy propagation paths and theoretical time delays between wind farms. Finally, through data cleaning and normalization, it is ensured that both types of data are free of missing data and outliers, forming a dataset with a unified structure.
[0036] S200: Based on historical weather forecast data, calculate the propagation path and theoretical time delay of wind energy between any two wind farms, and construct an initial adjacency matrix with wind farms as nodes.
[0037] This step targets any two wind farms within the target area, designated as Wind Farm 1 and Wind Farm 2. Using wind direction data at each time point t from historical weather forecasts as the core basis, it determines whether a physical path for wind energy propagation exists between the two wind farms. Specifically, it uses wind direction distribution characteristics to determine whether wind energy possesses the natural propagation conditions from Wind Farm 1 to Wind Farm 2. If such a propagation path is determined, the geographical distance between the two wind farms is further extracted. Combined with the average wind speed data at the corresponding time t, the theoretical time delay Δt = geographical distance between the two wind farms / average wind speed is used to calculate the wind energy propagation path from Wind Farm 1 to Wind Farm 2. The theoretical time difference of electric field propagation to the second wind farm is calculated. Then, the first and second wind farms are used as nodes in the graph structure. The calculated theoretical time delay Δt is used as the core weight of the directed edge between the two nodes. The direction of the directed edge is from the first wind farm to the second wind farm, which clarifies the correlation strength and temporal relationship between the two wind farms in the wind energy propagation dimension. Finally, all node pairs of wind farms in the target area are traversed, and the above path judgment, theoretical time delay calculation and directed edge construction operations are repeated to complete the correlation characterization between wind farm nodes in the entire area. Finally, an initial adjacency matrix is generated with wind farms as nodes and the theoretical time delay of wind energy propagation as edge weight.
[0038] In one feasible way, such as Figure 2 As shown, step S200 specifically includes: S210: For any two wind farms within the target area, denoted as the first wind farm and the second wind farm respectively, based on the wind direction field at time t of the historical weather forecast data, determine whether there is a wind energy propagation path from the first wind farm to the second wind farm.
[0039] For any two wind farms arbitrarily selected within the target area, denoted as Wind Farm 1 and Wind Farm 2 respectively, the wind direction field data at a specific time node t in historical weather forecast data is used as the core analysis basis. By extracting the airflow motion vector information of the wind direction field at that time t, including wind direction angle and wind speed direction distribution, and combining it with the geographical coordinates of the two wind farms, it is determined whether there is a natural propagation trend of airflow from the spatial location of Wind Farm 1 to the spatial location of Wind Farm 2. If the wind direction field shows that the angle between the main direction of airflow and the direction of the geographical line connecting the two wind farms is within a preset threshold range (e.g., within 30°), and there is no significant terrain obstruction or meteorological disturbance causing a sudden change in airflow direction, then it is determined that there is a wind energy propagation path from Wind Farm 1 to Wind Farm 2; otherwise, it is determined that there is no propagation path in that direction. This provides the prerequisite for subsequent theoretical time delay calculation and directed edge construction.
[0040] S220: If the propagation path is determined to exist, the theoretical time delay Δt for wind energy to propagate from the first wind farm to the second wind farm is calculated based on the geographical distance between the first wind farm and the second wind farm, and the average wind speed corresponding to the propagation path.
[0041] When a wind energy propagation path from the first wind farm to the second wind farm is determined, the precise geographical coordinates (such as latitude and longitude) of the two wind farms are first obtained. The straight-line geographical distance between the two wind farms is calculated using a geographic information calculation tool based on the Haversine formula, ensuring that the distance data accurately reflects the actual spatial span of wind energy propagation. Next, the average wind speed data corresponding to time t of the determined propagation path is extracted from historical weather forecast data. This wind speed should be the average wind speed value along the airflow propagation path between the two wind farms, rather than the wind speed of a single wind farm, to avoid the influence of local wind speed fluctuations on the calculation results. Finally, based on the physical laws of wind energy propagation, the theoretical time difference Δt required for wind energy to propagate from the first wind farm to the second wind farm is calculated using the formula: theoretical time delay Δt = geographical distance between the first and second wind farms / average wind speed of the corresponding propagation path. This time difference will serve as the key basis for constructing the core weights of the directed edges in the initial adjacency matrix, quantifying the temporal correlation strength between the two wind farms in the wind energy propagation dimension.
[0042] S230: Using the first wind farm and the second wind farm as nodes, and the theoretical time delay Δt as the core weight of the directed edge from the first wind farm to the second wind farm, construct the corresponding directed edge.
[0043] Using the first and second wind farms as two independent nodes in the graph structure, the corresponding identities of the two nodes are clearly defined: the first wind farm is the starting node for wind energy propagation, and the second wind farm is the ending node. Subsequently, the calculated theoretical time delay Δt is used as the core weight of the directed edge connecting these two nodes. The direction of the directed edge strictly follows the direction of wind energy propagation, that is, from the starting node, the first wind farm, to the ending node, the second wind farm, thus reflecting the temporal logic of wind energy propagation from the former to the latter. The magnitude of the edge weight directly corresponds to the theoretical time delay Δt, quantitatively reflecting the time required for wind energy to propagate from the first wind farm to the second wind farm. This allows the graph structure to not only reflect the relationship between wind farms, but also to characterize the temporal characteristics of the relationship through the edge weights, providing structured edge weight data support for the subsequent construction of the initial adjacency matrix.
[0044] S240: Traverse all wind farm node pairs within the target area, complete the construction of all directed edges, and generate an initial adjacency matrix; where the weight of the directed edge corresponding to node pairs without wind energy propagation paths is set to 0.
[0045] After constructing the directed edge for a single node pair (the first wind farm and the second wind farm), using all wind farms within the target area as the complete node set, all node pairs are traversed according to the principle of pairwise combination and verification. That is, for each new node pair, the two directed cases of all unordered node combinations in the set must be covered (e.g., A to B and B to A need to be determined separately). The propagation path judgment, theoretical time delay Δt calculation, and directed edge construction operations are repeatedly performed to ensure that no potential wind energy propagation association between any two wind farms is missed. During the traversal, each valid edge is recorded simultaneously. The starting node, ending node, and corresponding core weight (theoretical time delay Δt) of each directed edge are defined. If the path does not exist, the weight is set to 0 or null. After all node pairs have been traversed, the core weight of each directed edge is filled into the intersection of the corresponding row and column of the matrix, with the wind farm as the rows and columns (rows representing starting nodes and columns representing ending nodes). This forms a matrix with the same dimension as the number of nodes. This matrix is the initial adjacency matrix that can completely depict the temporal relationship of wind energy propagation between wind farms in the target area, laying a structured foundation for subsequent dynamic correction based on power data.
[0046] S300: Based on historical power data, calculate the time-delay cross-correlation between power data of each wind farm, perform weighted correction on the initial adjacency matrix, and generate a dynamically evolving spatiotemporal coupling graph sequence.
[0047] This step divides the acquired historical power data into multiple consecutive sub-time slices according to a preset time interval. The length of each time slice must be adapted to the accuracy of subsequent time-delay cross-correlation calculations to ensure that each sub-time slice contains enough data points to reflect short-term power change characteristics, laying the foundation for subsequent time-segmented analysis of power correlation between wind farms. For each sub-time slice, for any two wind farms within the target area, the cross-correlation coefficient ρ(τ) of their historical power sequences under different time delays is calculated, and the time delay τ_max that makes ρ(τ) reach its maximum value is located. This time delay represents the time difference in which the power changes of the two wind farms are most closely correlated. Subsequently, τ_max is compared and verified with the theoretical time delay Δt calculated based on meteorological data, and the absolute difference between the two is calculated. The weighting factor is determined based on the relationship between the absolute difference and the preset tolerance threshold, and then the initial adjacency matrix is corrected to generate the dynamic spatiotemporal coupling graph corresponding to a single sub-time slice. Finally, the dynamic spatiotemporal coupling graphs of all sub-time slices are serialized and combined in chronological order to form a dynamically evolving spatiotemporal coupling graph sequence. This sequence can reflect the dynamic coupling relationship between wind farms based on the dual correlation of meteorology and power at different time stages in real time, providing structured data support for subsequent input of the dual-stream adaptive spatiotemporal hypergraph network.
[0048] In one feasible manner, step S300 specifically includes: S310: Divide the historical power data into multiple consecutive sub-time slices according to a preset time interval to obtain the wind farm power data under each sub-time slice.
[0049] Specifically, continuous historical power data is broken down into short-term data units with time independence to capture the dynamic changes in power correlations between wind farms over different time periods. Next, the division criteria are determined: combining the time scale of wind energy propagation, such as short-term propagation patterns within a few hours, and data validity requirements, ensuring that each unit contains sufficient data points to support correlation calculations. A fixed time interval is preset as the sub-time slice length, such as 1 hour or 30 minutes, which needs to be adjusted according to the power fluctuation frequency of wind farms in the target area to avoid the time slice being too long to mask short-term correlation changes or too short to result in insufficient data. Subsequently, following the principle of continuous non-overlapping, the acquired full-time historical power data is divided into segments according to preset time intervals to form multiple continuous sub-time slices. Each sub-time slice contains complete power sequence data for all wind farms within that time period. For example, a 1-hour sub-time slice contains power values for each wind farm every 5 minutes, ultimately resulting in multiple sets of sub-time slice power data covering the entire historical period and arranged in chronological order.
[0050] S320: For each sub-time slice, perform the following operations: S321: Calculate the cross-correlation coefficient ρ(τ) of the historical power sequences of any two wind farms within this sub-time slice under time delay τ.
[0051] For each sub-time slice, focus on the historical power data of all wind farms in the target area within that time slice. For any two wind farms, denoted as wind farm A and wind farm B, extract their complete power sequences within that sub-time slice, ensuring precise timestamp alignment, such as power records every minute. Subsequently, set the range of the time delay τ to cover the possible propagation time interval of wind energy, such as 0 to 120 minutes, with a step size of 5 minutes to balance calculation accuracy and efficiency. For each τ value, advance the power sequence of wind farm A. After time unit τ, cross-correlation analysis is performed with the original power sequence of wind farm B. The Pearson correlation coefficient correlation degree calculation method is used to obtain the cross-correlation coefficient ρ(τ) under the corresponding time delay τ. The numerical range of ρ(τ) is [-1, 1]. The larger the absolute value, the closer the correlation between the power changes of the two wind farms under the time delay. A positive value indicates a positive correlation, with the power increasing or decreasing together. A negative value indicates a negative correlation, with the power change trends being opposite. By traversing all τ values, the correlation strength of the wind farm pair under different time delays within this sub-time slice is quantified.
[0052] S322: Find the time delay τ_max that maximizes the cross-correlation coefficient ρ(τ), compare and verify the time delay τ_max with the theoretical time delay Δt of the corresponding node pair, and calculate the absolute difference between the two.
[0053] After calculating the cross-correlation coefficient ρ(τ) for different time delays τ of any two wind farms within the current sub-time slice, the time delay τ_max that maximizes ρ(τ) is selected from all τ values and their matching ρ(τ). This τ_max represents the actual time difference where the power changes of the two wind farms are most significantly correlated and have the strongest synchronicity, directly reflecting the true propagation time sequence characteristics of power fluctuations between the two wind farms. Subsequently, the selected τ_max is compared and verified with the theoretical time delay Δt: by calculating the absolute difference between τ_max and Δt, the degree of deviation between the two is quantified, thereby judging the matching degree between the actual power correlation time delay between wind farms and the time delay derived from meteorological theory. If the absolute difference is small, it indicates that the power correlation law and the meteorological propagation law are highly consistent; if the absolute difference is large, it indicates that factors such as terrain interference and local airflow anomalies may cause the deviation between the two, providing a core judgment basis for subsequent determination of weighting factors and correction of adjacency matrix edge weights.
[0054] S323: Determine the weighting factor based on the comparison result between the absolute difference and the preset tolerance threshold, and fuse the weighting factor with the weight of the corresponding directed edge in the initial adjacency matrix to obtain the fused edge weight set under this sub-time slice.
[0055] The weighting factor is determined based on the comparison and verification results of τ_max and Δt. First, a tolerance threshold is preset, which is set according to the historical data of wind farms in the target area and the meteorological accuracy requirements. This threshold is used to judge the matching degree between the two. If the absolute difference is less than or equal to the preset tolerance threshold, it indicates that the actual power correlation law is highly consistent with the meteorological theoretical derivation. In this case, the weighting factor is directly taken as the maximum value of the cross-correlation coefficient ρ(τ) in the sub-time slice to strengthen the correlation weight when the two are matched. If the absolute difference is greater than the preset tolerance threshold, it indicates that there is a deviation between the two, which may be affected by factors such as local airflow and terrain obstruction. In this case, the weighting factor is set as the product of the maximum value of the cross-correlation coefficient ρ(τ) and the reverse suppression coefficient. The reverse suppression coefficient is a value less than 1 and is inversely proportional to the absolute difference. That is, the larger the deviation, the smaller the suppression coefficient. This weakens the correlation weight when the deviation is large and reduces abnormal interference. After determining the weighting factor, for the edge weight Δt of the corresponding wind farm node pair in the initial adjacency matrix, a preset fusion rule is adopted: fused edge weight = Δt × weighting factor. This calculation preserves the time-series information of meteorological theoretical time lag and incorporates the dynamic characteristics of power correlation. The process is repeated for all wind farm node pairs in the current sub-time slice, and the above weighting factor determination and edge weight fusion operation is repeated to finally form a fused edge weight set covering all node pairs in the sub-time slice.
[0056] S324: Based on the fusion edge weight set, modify the initial adjacency matrix and generate the corresponding dynamic spatiotemporal coupling graph for this sub-time slice.
[0057] Using the fused edge weight set of the current sub-time slice as the core correction basis, for each pair of wind farm nodes in the initial adjacency matrix, corresponding to a directed edge, the fused edge weight of that node pair in the fused edge weight set replaces the corresponding original edge weight in the initial adjacency matrix, which is only the meteorological theoretical time delay Δt. This completes the local correction of the initial adjacency matrix. The corrected matrix not only retains the temporal logic of wind energy propagation but also incorporates the dynamic characteristics of the actual power correlation between wind farms within this sub-time slice, forming a dynamic adjacency matrix that corresponds only to the current sub-time slice. Next, using wind farms as nodes and the fused edge weights in the corrected matrix as directed edge weights, a dynamic spatiotemporal coupling graph under this sub-time slice is constructed. The connection relationships and time delay weights of the nodes in the graph both reflect the real spatiotemporal correlation state between wind farms in the current time period.
[0058] S330: Sequentially combine the dynamic spatiotemporal coupling diagrams corresponding to all sub-time slices in chronological order to obtain a dynamically evolving spatiotemporal coupling diagram sequence.
[0059] After all sub-time slices have completed the operation of fusing edge weight correction matrices and generating dynamic spatiotemporal coupling graphs, all independent dynamic spatiotemporal coupling graphs are sequentially arranged according to the time sequence of the sub-time slices, from the earliest time period to the latest time period of the historical data. Finally, a spatiotemporal coupling graph sequence that can completely depict the spatiotemporal relationship between wind farms in the target area over time is formed.
[0060] S400: Input the spatiotemporal coupled graph sequence and the node features of each wind farm into the pre-trained dual-stream adaptive spatiotemporal hypergraph network, extract the spatiotemporal fusion features, and obtain the predicted quantiles of the power of each wind farm in the future time period based on the spatiotemporal fusion features. The predicted quantiles include the 10th quantile, the 50th quantile and the 90th quantile.
[0061] This step employs a dual-branch input mechanism, using the generated dynamic evolutionary spatiotemporal coupled graph sequence and the characteristics of each wind farm node, including static and dynamic attributes such as installed capacity, geographical coordinates, and historical average power, as joint inputs. The spatiotemporal coupled graph sequence must maintain temporal continuity, and the node features must correspond one-to-one with the wind farm nodes in each time slice of the graph sequence. The temporal branch of the network uses multiple parallel causal dilated convolutional layers, each with different dilation rates to capture short, medium, and long-term multi-scale temporal dependencies. It extracts the power correlation temporal features of each wind farm node in the spatiotemporal coupled graph sequence hierarchically, and outputs the multi-scale temporal features of each node after feature concatenation. The spatial branch, through hypergraph convolution operations, mines high-order spatial coupling relationships between wind farms within the region based on the edge weight information in the spatiotemporal coupled graph sequence, such as multi-wind farm cooperative fluctuation patterns, generating enhanced spatial features for each node. Subsequently, the network dynamically weights and fuses the multi-scale temporal features and enhanced spatial features of each node through an adaptive fusion module, forming a spatiotemporal fusion feature that combines temporal evolution patterns and spatial correlation characteristics. This fusion feature is mapped through a fully connected layer and then input to the output layer pre-trained with an asymmetric quantile regression loss function. During training, higher weights are given to the prediction loss of historical extreme ramp periods to enhance the model's sensitivity to extreme events. Finally, the network outputs three types of prediction quantiles for the power of each wind farm in the future preset time period: the 10th quantile (lowest power prediction under pessimistic scenario), the 50th quantile (most likely power prediction under baseline scenario), and the 90th quantile (highest power prediction under optimistic scenario), providing key quantitative indicators for subsequent calculation of power ramp rate and determination of extreme events.
[0062] In one feasible approach, the dual-stream adaptive spatiotemporal hypergraph network comprises a temporal branch and a spatial branch, wherein the temporal branch employs causal dilated convolutions to capture long-range extreme dependencies, and the spatial branch uses hypergraph convolutions to mine multi-scale regional coupling patterns.
[0063] Specifically, the dual-stream adaptive spatiotemporal hypergraph network achieves in-depth mining of the spatiotemporal correlation features of wind farms through the collaborative work of temporal and spatial branches. The specific structure and functions are as follows: The temporal branch uses causal dilated convolution as its core structure. This convolution expands the receptive field without increasing the amount of computation and parameters by introducing "holes" in the convolution kernel, i.e., skipping some input pixels. It can effectively capture the temporal dependence of wind farm power correlation over a long period of time. Especially for phenomena with long-range time characteristics such as extreme hill climbing events, causal dilated convolution can avoid the excessive focus of traditional convolution on short-term data and accurately capture the long-range temporal pattern of wind farm power fluctuations under different sub-time slices, such as the impact of the power change of one wind farm on another wind farm a few hours later. At the same time, the "causal" characteristic ensures that only historical and current data are used, which is consistent with the logic of actual time series prediction.
[0064] Spatial branching uses hypergraph convolution as its core operation. Unlike traditional graph convolution, which can only depict the direct pairwise relationships between nodes, hypergraph convolution constructs "hyperedges" to connect the edges of multiple nodes. This allows it to uncover multi-scale regional coupling patterns between wind farms within a target area. For example, it can not only capture the direct relationship between two wind farms, but also depict the clustered relationship formed by three or more wind farms. This includes the coordinated power fluctuations of multiple wind farms in a certain area affected by the same airflow system. This achieves full coverage of the spatial relationship of wind farms from "local pairwise relationships" to "global multi-scale coupling", accurately restoring the collaborative or restrictive relationships of wind farms in different spatial ranges.
[0065] In one implementable manner, step S400 specifically includes: S410: Input the spatiotemporal coupled graph sequence and the features of each node into the temporal branch of the dual-stream adaptive spatiotemporal hypergraph network, and use multiple parallel causal dilated convolutional layers to capture temporal dependency features at different time scales and splice them to obtain the multi-scale temporal features of each node.
[0066] Specifically, the generated dynamic spatiotemporal coupled graph sequence and the static and dynamic attributes of each wind farm node, such as installed capacity, geographical location, and historical average power, are used as joint inputs and fed into the temporal branch of the dual-stream adaptive spatiotemporal hypergraph network. This branch is configured with multiple parallel causal dilated convolutional layers. Each layer, with preset different dilation rates (e.g., 1, 6, 24), accurately captures temporal dependencies across multiple time scales. For example, convolutional layers with smaller dilation rates focus on short-term power correlation fluctuations between adjacent sub-time slices, while convolutional layers with larger dilation rates can span multiple sub-time slices to capture extreme dependency features in the long-term time dimension, such as the temporal precursors before extreme hill-climbing events. Simultaneously, the "causal" characteristic ensures that the convolutional operations rely only on historical and current data, avoiding future information leakage and conforming to actual prediction logic. After each causal-hole convolutional layer outputs the temporal features corresponding to the time scale, the features are then merged according to the channel dimension through feature splicing operation to generate a set of temporal features covering short, medium and long time scales for each wind farm node. This multi-scale temporal feature of each node comprehensively describes the power correlation evolution law under different time spans.
[0067] The temporal branch incorporates multiple parallel causal dilated convolutional layers. By pre-setting differentiated dilation rates, these layers enable targeted extraction of features across different time spans. The first convolutional layer uses a dilation rate of 1, with its receptive field focused on adjacent sub-time slices, such as hourly or half-hourly units. This allows for precise capture of rapid power fluctuations in wind farms over hourly timescales, such as sudden increases or decreases in power caused by short-term gusts. The second convolutional layer uses a dilation rate of 6, extending its receptive field to several hours, such as 6-12 hours. This is specifically designed to capture typical uphill characteristics at this timescale, such as continuous increases or decreases in power caused by changes in wind speed gradients. The third convolutional layer uses a dilation rate of 24, with its receptive field covering the entire day and beyond. This effectively captures long-range temporal dependence features influenced by continuous weather processes, such as cold waves and monsoons, over day-scale timescales. Through this parallel design of multiple porosities, each convolutional layer focuses on key features at different time scales, ensuring that the multi-scale temporal features formed by subsequent splicing can fully cover the temporal patterns of the entire range from short-term fluctuations to long-term weather processes, providing complete temporal dimension feature support for the accurate identification of extreme uphill events.
[0068] S420: Input the spatiotemporal coupled graph sequence and the features of each node into the spatial branch of the dual-stream adaptive spatiotemporal hypergraph network for convolution operation, mine the spatial patterns of regional coupling, and output the spatial features of each node after hypergraph convolution enhancement.
[0069] The dynamic spatiotemporal coupling graph sequence, originating from the temporal branch, and the features of each wind farm node are input into the spatial branch of the dual-stream adaptive spatiotemporal hypergraph network. This branch uses hypergraph convolution as its core computational unit. By constructing "hyperedges," hypergraph convolution can simultaneously connect multiple wind farm nodes, breaking through the limitations of local correlations and deeply mining the multi-scale regional coupling spatial patterns of wind farms within the target area. This includes both the local proximity coupling relationship between a single wind farm and its surrounding wind farms, and the cluster-like global coupling relationship formed by multiple wind farms being affected by the airflow system and terrain conditions, such as the spatial coordination pattern of power fluctuations occurring simultaneously in multiple wind farms within a certain area. During the hypergraph convolution operation, the network dynamically adjusts the weights according to the node features and the strength of spatial correlation, strengthening the feature expression of key coupling relationships. Finally, it outputs a feature vector enhanced by spatial correlation for each wind farm node, that is, the hypergraph convolution-enhanced spatial features of each node, accurately reflecting its positional importance and correlation strength in the regional spatial coupling network.
[0070] S430: Generate spatiotemporal fusion features by fusing the multi-scale temporal features and spatial features of each node.
[0071] This paper addresses the multi-scale temporal features of each wind farm node in the temporal branch output of a dual-stream adaptive spatiotemporal hypergraph network. These features encompass power correlation patterns across short, medium, and long time scales. The spatial features are enhanced by convolution with the corresponding node hypergraph output from the spatial branch, reflecting the spatial correlation patterns of multi-scale coupling within the region. An adaptive fusion mechanism built into the network is employed for feature integration. This mechanism dynamically calculates the attention weights of the two types of features; for example, features with a more significant impact from extreme hill-climbing events are assigned higher weights. Multi-scale temporal and spatial features are then fused element-wise according to their weights, with simultaneous channel dimension concatenation operations. This achieves deep integration of the two types of features, ensuring that the fused features retain both the temporal evolution trajectory of wind farm power changes and their spatial correlation attributes within the regionally coupled network. Ultimately, a comprehensive feature vector with both temporal dynamics and spatial correlation is generated for each wind farm node—a spatiotemporally fused feature that comprehensively characterizes the core features of the wind farm in the regionally coupled spatiotemporal evolution process.
[0072] In one possible implementation, step S400 further includes: S440: Determine the preset target quantiles, including the 10th quantile, the 50th quantile, and the 90th quantile.
[0073] Specifically, considering the actual needs of wind farm power forecasting, it is necessary to cover both the range of normal power fluctuations and the key boundary of extreme power changes. Preset target quantiles are determined as the 10th, 50th, and 90th quantiles. The 50th quantile, or median, serves as the core benchmark, reflecting the most likely power value of each wind farm in the future period and characterizing the central trend of power changes. The 10th and 90th quantiles correspond to the extreme boundaries of low and high power, respectively. The former defines the possible range of abnormally low power, while the latter locks in the potential upper limit related to significant power surges and extreme ramp-up events. Together, these three quantiles construct a power probability forecasting framework covering the entire range from normal to extreme, providing a clear quantitative benchmark for subsequently identifying extreme ramp-up events through quantile differences.
[0074] S450: The regression output layer is trained based on a preset target quantile using an asymmetric quantile regression loss function.
[0075] Using the 10th, 50th, and 90th quantiles as target benchmarks, an asymmetric quantile regression loss function is employed to train the network's regression output layer. The core feature of this loss function is the asymmetric penalty applied to the deviation between predicted and true values. For the 10th quantile, a larger penalty weight is given to positive deviations where the predicted value is higher than the true value, strengthening the capture of abnormally low power values. For the 90th quantile, a heavier penalty is applied to negative deviations where the predicted value is lower than the true value, focusing on optimizing the prediction accuracy for extreme power spikes. The 50th quantile, however, uses a symmetric penalty mechanism to ensure the stability of the central trend prediction. This differentiated penalty strategy forces the regression output layer to prioritize learning the feature mapping patterns at extreme quantiles during training, improving its sensitivity to extreme power values and prediction accuracy. Ultimately, this enables the output layer to stably fit the nonlinear relationship between spatiotemporal fusion features and each target quantile, laying the model foundation for generating reliable power prediction quantiles.
[0076] In one feasible approach, the regression output layer is trained using an asymmetric quantile regression loss function, specifically including: S451: Collect training sample set. Collect a training sample set covering the historical operating cycles of wind farms in the target area. The sample set should include historical spatiotemporal coupling graph sequences, characteristics of each node, such as installed capacity, geographical location, and actual power values for the corresponding time period. Ensure that the sample set can cover both normal power fluctuations and extreme ramp-up events, providing comprehensive data support for model training.
[0077] S452: Construct the network structure of the regression output layer, train it using the training sample set, and calculate the quantile loss between the predicted value and the true value under the preset target quantile.
[0078] A network structure for the regression output layer is constructed using a multilayer perceptron, comprising an input layer, hidden layers, and an output layer. The number of neurons in the output layer is matched with the number of preset target quantiles. The spatiotemporal coupling graph sequence and node features in the training sample set are used as inputs, and the actual power values are used as labels to initially train the model. During the training process, for the determined 10th, 50th, and 90th quantiles, the basic quantile loss between the model's predicted value and the actual power value at each quantile is calculated.
[0079] S453: Calculate the instantaneous ramp rate at each time point based on the training sample set.
[0080] Based on historical actual power data of each wind farm in the collected training sample set, for each wind farm node, at preset time intervals consistent with the granularity of sub-time slices (e.g., 15 minutes, 30 minutes), the instantaneous ramp rate is calculated using the formula: (actual power value at the current time point - actual power value at the previous adjacent time point) / time interval. This indicator directly quantifies the severity of power changes at each time point. A positive instantaneous ramp rate indicates that the power is increasing (climbing), while a negative value indicates that the power is decreasing (descending). The larger the absolute value, the more drastic the power change. By traversing all time points and wind farm nodes in the training sample set, full-sample-coverage instantaneous ramp rate data is obtained, accurately marking time points where extreme ramp events may occur. This provides a quantitative basis for subsequent dynamic adjustment of loss penalty weights and strengthening the model's learning of extreme scenarios.
[0081] Step S454: When the absolute value of the instantaneous ramp rate exceeds the preset sensitivity threshold, assign a penalty weight greater than 1 to the loss at that time point, establish an asymmetric quantile regression loss function by weighting the loss of each quantile with the corresponding penalty weight, and train the regression output layer until convergence.
[0082] A sensitivity threshold is preset, set based on the statistical characteristics of historical extreme climbing events, to distinguish between normal power fluctuations and extreme changes. For each time point in the training sample set, the calculated absolute value of the instantaneous climbing rate is compared with this threshold. If the absolute value of the instantaneous climbing rate exceeds the sensitivity threshold, the time point is determined to correspond to an extreme climbing scenario, including sharp ascents or descents. A penalty weight greater than 1 (e.g., 1.2~2.0) is assigned to the quantile loss at that time point. The weight is positively correlated with the degree of extremeness to strengthen the model's learning focus on extreme scenarios. If the threshold is not exceeded, the penalty weight remains at 1, with normal penalty intensity. Subsequently, an asymmetric quantile regression loss function is constructed, where total loss = sum of each quantile loss × corresponding penalty weight. This function guides the regression output layer to impose a more severe penalty on the prediction error at extreme climbing moments during training, forcing the model to prioritize fitting the characteristic patterns of extreme scenarios. Based on this loss function, the regression output layer is continuously trained iteratively, and the network parameters are continuously adjusted through backpropagation until the total loss value tends to stabilize and convergence conditions are met. Finally, a regression output layer model with significantly improved prediction accuracy for extreme climbing events is obtained.
[0083] S460: By analyzing the spatiotemporal fusion characteristics through the regression output layer, the predicted quantiles of the power of each wind farm in the future time period are generated.
[0084] The generated spatiotemporal fusion features, which integrate multi-scale temporal features and spatial coupling features, are input into the regression output layer, which is trained and optimized using an asymmetric quantile regression loss function. This output layer performs nonlinear transformation and deep analysis on the input spatiotemporal fusion features using a network structure such as a multilayer perceptron. Based on the learned feature-quantile mapping relationship, it generates predicted values for the 10th, 50th, and 90th quantiles of power for each wind farm node in the future, for preset time periods such as the next 1 hour and 6 hours. These predicted quantiles not only quantify the central trend of future power (i.e., the 50th quantile) but also define the extreme range of possible power fluctuations through upper and lower quantiles, providing a direct probabilistic predictive basis for subsequent judgments based on quantile differences regarding the occurrence of extreme ramp events.
[0085] S500: Calculates the power ramp rate based on the predicted quantile, and determines whether an extreme ramp event has occurred by combining the dynamic ramp threshold. It also outputs the confidence level, intensity range, and expected occurrence time window of the event.
[0086] Specifically, the dynamic ramp-up threshold for the target area is retrieved. This threshold needs to be set in conjunction with the historical extreme ramp-up event records of the wind farm, the grid's safety carrying capacity, and dispatch requirements, and must be adaptable to different seasons and grid operating conditions. Simultaneously, the functional positioning of the three types of prediction quantiles is clarified: the 50th quantile is the baseline power prediction value, the 10th quantile is the pessimistic scenario low power fluctuation prediction value, and the 90th quantile is the optimistic scenario high power fluctuation prediction value. Next, according to the calculation rules of the power ramp-up rate (i.e., the ratio of power change per unit time to the baseline power), the power ramp-up rate for the corresponding future period is calculated based on the three types of prediction quantiles: the baseline ramp-up rate R50 is calculated using the predicted power of the 50th quantile, and the pessimistic scenario ramp-up rate R10 and the optimistic scenario ramp-up rate R90 are calculated using the predicted power of the 10th and 90th quantiles, respectively. Subsequently, extreme ramp events are identified based on a pre-defined judgment logic: if the absolute value of the baseline ramp rate R50 exceeds the dynamic ramp threshold, it indicates that the extreme ramp conditions are met under the baseline scenario, and the event is judged as a high-confidence extreme ramp event; if the absolute value of the baseline ramp rate R50 does not exceed the dynamic ramp threshold, but the absolute values of the pessimistic scenario ramp rate R10 or the optimistic scenario ramp rate R90 exceed the dynamic ramp threshold, it indicates that although there are no extreme events under the baseline scenario, there is a risk under the extreme fluctuation scenario, and the event is judged as a low-confidence high-risk extreme ramp event; if neither of the above two conditions is met, the event is judged as having no extreme ramp events. Finally, the judgment results are integrated to output the core information of the event: clearly marking the event confidence level and intensity range, represented by the ramp rate value ranges of R10, R50, and R90, and combining the time dimension information corresponding to the predicted quantiles to determine the expected occurrence time window of the extreme ramp event, providing accurate decision-making basis for the power grid dispatching department to formulate power balance strategies and avoid operational risks.
[0087] In one feasible manner, step S500 specifically includes: S510: Obtain the preset dynamic climbing threshold.
[0088] Based on the installed capacity of wind farms in the target area, grid safety constraints such as frequency stability limits, line transmission capacity, intensity distribution characteristics of historical extreme ramping events, and grid regulation needs at different times, such as peak / off-peak electricity consumption and seasonal transitions, a dynamic ramping threshold is pre-set. This threshold is not a fixed value but is adaptively adjusted according to the actual operating scenario. For example, during peak load periods or when grid reserve capacity is low, the threshold is appropriately lowered to improve early warning sensitivity; during equipment maintenance periods or when renewable energy penetration is low, the threshold can be moderately increased to reduce false alarms. By obtaining this dynamically updated threshold, a quantitative benchmark matching the actual operating scenario is provided for subsequent extreme event judgments based on ramp rate, ensuring that the judgment results meet both grid safety requirements and wind farm operating characteristics.
[0089] S520: Calculate the baseline ramp rate R50 based on the predicted power at the 50th quantile.
[0090] Based on the generated power prediction quantiles for each wind farm within the future time period, the 50th quantile predicted power value, which best reflects the power center trend, is extracted and used as the calculation benchmark. Subsequently, the benchmark ramp rate R50 is calculated for each wind farm according to the formula: R50 = (50th quantile predicted power at the future target time - actual power at the current time) / time interval between the future target time and the current time. This indicator quantifies the most likely rate and direction of change in wind farm power within the future time period. If R50 is positive, it indicates that the future power is likely to increase; if it is negative, it indicates that the future power is likely to decrease. Its absolute value directly reflects the severity of this most likely trend.
[0091] S530: Calculate the pessimistic scenario ramp rate R10 and the optimistic scenario ramp rate R90 based on the predicted power of the 10th and 90th quantiles.
[0092] Based on the output power prediction quantiles for each wind farm in the future time period, the 10th and 90th quantile predicted power values, reflecting the extreme boundaries of power fluctuations, are extracted. The 10th quantile represents the potentially low extreme level of future power, corresponding to the pessimistic scenario of the slowest power growth or the most drastic decline. The 90th quantile represents the potentially high extreme level of future power, corresponding to the optimistic scenario of the fastest power growth or the slowest decline. Subsequently, following the calculation logic of the benchmark ramp rate, the pessimistic scenario ramp rate R10 and the optimistic scenario ramp rate R90 for each wind farm are calculated using the formula: Scenario ramp rate = (Predicted power at the corresponding quantile of the future target time - Actual power at the current time) / Time interval between the future target time and the current time. These two indicators quantify the most drastic trends that may occur in future power changes from the perspective of extreme boundaries. R10 focuses on extreme ramp-down or weak ramp-up, while R90 focuses on extreme ramp-up or weak ramp-down, providing a judgment basis covering extreme scenarios for subsequent identification of potential high-risk extreme events.
[0093] S540: The logic for determining extreme climbing events is as follows: If the absolute value of the baseline climb rate R50 exceeds the dynamic climb threshold, it is determined to be a high-confidence extreme climb event.
[0094] If the absolute value of the baseline ramp rate R50 does not exceed the dynamic ramp threshold, but the absolute value of the pessimistic scenario ramp rate R10 or the optimistic scenario ramp rate R90 exceeds the dynamic ramp threshold, it is determined to be a low-confidence, high-risk extreme ramp event.
[0095] Otherwise, it is determined as no extreme climbing event.
[0096] If the absolute value of the baseline ramp rate R50 exceeds the dynamic ramp threshold, it indicates that the most likely future power change has reached an extreme level. This judgment, based on quantiles reflecting the central trend, has high reliability and is therefore directly classified as a high-confidence extreme ramp event. Secondly, if the absolute value of the baseline ramp rate R50 does not exceed the threshold, but the absolute values of the pessimistic scenario ramp rate R10 or the optimistic scenario ramp rate R90 exceed the threshold, it indicates that although the most likely power change is within the normal range, there is still a risk of drastic power fluctuations in extreme scenarios. Since this type of judgment is based on extreme quantiles and has relatively low certainty, it is classified as a low-confidence, high-risk extreme ramp event. Finally, if neither of the above two conditions is met—that is, the absolute values of R50, R10, and R90 do not exceed the dynamic ramp threshold—it indicates that future power changes are within the normal fluctuation range and is classified as a non-extreme ramp event. This hierarchical judgment logic accurately captures high-certainty extreme events while providing early warnings of potential risks, offering comprehensive decision support for grid dispatch from deterministic to probabilistic perspectives.
[0097] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.
[0098] Example 2: Based on the same inventive concept as the wind power extreme ramping event identification method under regional coupled spatiotemporal evolution in the aforementioned embodiments, such as Figure 3 As shown, this application provides a wind power extreme ramping event identification system under regional coupled spatiotemporal evolution. The system and method embodiments in this application are based on the same inventive concept. The system includes: Data acquisition module 10 is used to acquire historical power data and historical weather forecast data of each wind farm in the target area; The initial adjacency matrix construction module 20 is used to calculate the propagation path and theoretical time delay of wind energy between wind farms based on historical weather forecast data, and to construct an initial adjacency matrix with wind farms as nodes. The spatiotemporal coupling graph sequence generation module 30 is used to calculate the time-delay cross-correlation between the power data of each wind farm based on historical power data, perform weighted correction on the initial adjacency matrix, and generate a dynamically evolving spatiotemporal coupling graph sequence. The spatiotemporal fusion feature extraction module 40 is used to input the spatiotemporal coupling graph sequence and the node features of each wind farm into the pre-trained dual-stream adaptive spatiotemporal hypergraph network, extract the spatiotemporal fusion features, and obtain the predicted quantiles of the power of each wind farm in the future time period. The predicted quantiles include the 10th quantile, the 50th quantile, and the 90th quantile. The extreme ramp event determination module 50 is used to calculate the power ramp rate based on the predicted quantile and determine whether an extreme ramp event has occurred by combining the dynamic ramp threshold. At the same time, it outputs the confidence level, intensity range and expected occurrence time window of the event.
[0099] The dual-stream adaptive spatiotemporal hypergraph network includes a temporal branch and a spatial branch. The temporal branch uses multiple parallel causal dilated convolutional layers with different dilation rates to capture long-range extreme temporal dependency features at different time scales. The spatial branch uses hypergraph convolutional layers to mine multi-scale high-order spatial coupling patterns between wind farms in the region. The spatiotemporal fusion feature extraction module has a built-in pre-trained convergent regression output layer. The regression output layer is pre-trained with an asymmetric quantile regression loss function and is used to perform nonlinear mapping analysis on the input spatiotemporal fusion features, outputting the 10th, 50th, and 90th quantile prediction results of the power of each wind farm in the future time period. The extreme ramp event determination module is specifically used to calculate the baseline ramp rate R50, the pessimistic scenario ramp rate R10, and the optimistic scenario ramp rate R90 based on the prediction results of the 10th quantile, 50th quantile, and 90th quantile, respectively, and execute the following determination logic: If the absolute value of the baseline climb rate R50 exceeds the dynamic climb threshold, it is determined to be a high-confidence extreme climb event. If the absolute value of the baseline ramp rate R50 does not exceed the dynamic ramp threshold, but the absolute value of the pessimistic scenario ramp rate R10 or the optimistic scenario ramp rate R90 exceeds the dynamic ramp threshold, it is judged as a low-confidence, high-risk extreme ramp event. If none of the above conditions are met, it is determined that there is no extreme climbing event.
[0100] Furthermore, the system is also used to implement the following functions: For any two wind farms within the target area, denoted as the first wind farm and the second wind farm, based on the wind direction field at time t according to the historical weather forecast data, it is determined whether there is a wind energy propagation path from the first wind farm to the second wind farm. If the propagation path is determined to exist, the theoretical time delay Δt for wind energy propagation from the first wind farm to the second wind farm is calculated based on the geographical distance between the first wind farm and the second wind farm, and the average wind speed corresponding to the propagation path. Using the first wind farm and the second wind farm as nodes, and the theoretical time delay Δt as the core weight of the directed edge from the first wind farm to the second wind farm, the corresponding directed edge is constructed. All wind farm node pairs within the target area are traversed to complete the construction of all directed edges, generating the initial adjacency matrix.
[0101] Furthermore, the system is also used to implement the following functions: The historical power data is divided into multiple consecutive sub-time slices according to a preset time interval to obtain wind farm power data for each sub-time slice. For each sub-time slice, the following operations are performed: calculate the cross-correlation coefficient ρ(τ) of the historical power sequences of any two wind farms within the sub-time slice under a time delay τ; find the time delay τ_max that maximizes the cross-correlation coefficient ρ(τ), compare and verify the time delay τ_max with the theoretical time delay Δt of the corresponding node pair, and calculate the absolute difference between the two; determine the weighting factor based on the comparison result of the absolute difference with a preset tolerance threshold, and fuse the weighting factor with the weight of the corresponding directed edge in the initial adjacency matrix to obtain the fused edge weight set for the sub-time slice; modify the initial adjacency matrix based on the fused edge weight set to generate the dynamic spatiotemporal coupling graph corresponding to the sub-time slice; and serialize and combine the dynamic spatiotemporal coupling graphs corresponding to all sub-time slices in chronological order to obtain the dynamically evolving spatiotemporal coupling graph sequence.
[0102] Furthermore, the system is also used to implement the following functions: If the absolute difference is less than or equal to the preset tolerance threshold, the weighting factor is determined to be the maximum value of the cross-correlation coefficient ρ(τ); if the absolute difference is greater than the preset tolerance threshold, the weighting factor is determined to be the product of the inverse inhibition coefficient and the maximum value of the cross-correlation coefficient ρ(τ), wherein the inverse inhibition coefficient is a value less than 1 and is inversely proportional to the absolute difference.
[0103] Furthermore, the system is also used to implement the following functions: The spatiotemporal coupling graph sequence and the node features of each wind farm are input into the temporal branch. Multiple parallel causal-dilated convolutional layers with different void ratios are used to capture and concatenate temporal dependency features at different time scales to obtain multi-scale temporal features for each wind farm node. The spatiotemporal coupling graph sequence and the node features of each wind farm are input into the spatial branch. Hypergraph convolution operations are used to mine the multi-scale spatial coupling patterns of regional wind farms, and the spatial features of each wind farm node after hypergraph convolution enhancement are output. The multi-scale temporal features and spatial features corresponding to each wind farm node are adaptively weighted and fused to generate the spatiotemporal fused features.
[0104] Furthermore, the system is also used to implement the following functions: The preset target quantiles are determined as the 10th, 50th, and 90th quantiles; a regression output layer of a dual-stream adaptive spatiotemporal hypergraph network is constructed, which is pre-trained using the preset target quantiles based on an asymmetric quantile regression loss function; the pre-trained regression output layer is used to perform nonlinear mapping analysis on the spatiotemporal fusion features to generate the 10th, 50th, and 90th quantile prediction results of the power of each wind farm in the future time period.
[0105] Furthermore, the system is also used to implement the following functions: A training sample set covering both regular power fluctuations and extreme ramp events is collected. This training sample set includes historical spatiotemporal coupling graph sequences, wind farm node features, and actual power values for corresponding time periods. A network structure for the regression output layer is constructed, using the spatiotemporal coupling graph sequences and wind farm node features from the training sample set as input and the actual power values for corresponding time periods as labels for training. The quantile loss between the predicted and actual values at each preset target quantile is calculated. Based on the historical power data in the training sample set, the instantaneous ramp rate at each time point is calculated. When the absolute value of the instantaneous ramp rate exceeds a preset sensitivity threshold, a penalty weight greater than 1 is assigned to the quantile loss at that time point. An asymmetric quantile regression loss function is constructed by weighting each quantile loss with its corresponding penalty weight, and the regression output layer is iteratively trained until convergence.
[0106] Furthermore, the system is also used to implement the following functions: Obtain dynamic ramp thresholds that adapt to the operating conditions of the power grid and the operating characteristics of the wind farm in the target area; calculate the baseline ramp rate R50 based on the 50th quantile predicted power, the pessimistic scenario ramp rate R10 based on the 10th quantile predicted power, and the optimistic scenario ramp rate R90 based on the 90th quantile predicted power; determine extreme ramp events according to the following logic: if the absolute value of the baseline ramp rate R50 exceeds the dynamic ramp threshold, it is determined as a high-confidence extreme ramp event; if the absolute value of the baseline ramp rate R50 does not exceed the dynamic ramp threshold, but the absolute value of the pessimistic scenario ramp rate R10 or the optimistic scenario ramp rate R90 exceeds the dynamic ramp threshold, it is determined as a low-confidence high-risk extreme ramp event; if none of the above conditions are met, it is determined as no extreme ramp event; based on the determination results, simultaneously output the event's confidence level, the event intensity range represented by the R10, R50, and R90 value intervals, and the event's expected occurrence time window.
[0107] In this embodiment, the specific implementation methods of each module of the system are the same as those in Embodiment 1, and the repeated content will not be repeated here.
[0108] The above are specific embodiments of the present invention, which demonstrate the substantial features and progress of the present invention. Equivalent modifications can be made to them according to actual usage needs, based on the teachings of the present invention, and all such modifications are within the scope of protection of this solution.
Claims
1. A method for identifying extreme wind power ramping events under regional coupled spatiotemporal evolution, characterized in that, Includes the following steps: S100: Obtain historical power data and historical weather forecast data for each wind farm within the target area; S200: Based on the historical weather forecast data, calculate the propagation path and theoretical time delay of wind energy between any two wind farms, and construct an initial adjacency matrix with wind farms as nodes; S300: Calculate the time-delay cross-correlation between the power data of each wind farm based on the historical power data, and perform weighted correction on the initial adjacency matrix according to the time-delay cross-correlation calculation results to generate a dynamically evolving spatiotemporal coupling graph sequence. S400: Input the spatiotemporal coupling graph sequence and the node features of each wind farm into a pre-trained dual-stream adaptive spatiotemporal supergraph network, extract spatiotemporal fusion features, and obtain the predicted quantiles of the power of each wind farm in the future time period based on the spatiotemporal fusion features. The predicted quantiles include the 10th quantile, the 50th quantile and the 90th quantile. S500: Calculate the corresponding power ramp rate based on the predicted quantile, and determine whether an extreme ramp event has occurred by combining the dynamic ramp threshold. At the same time, output the confidence level, intensity range and expected occurrence time window of the event.
2. The method for identifying extreme wind power ramping events under regional coupled spatiotemporal evolution according to claim 1, characterized in that, Step S200 specifically includes: S210: For any two wind farms within the target area, denoted as the first wind farm and the second wind farm respectively, based on the wind direction field at time t of the historical weather forecast data, determine whether there is a wind energy propagation path from the first wind farm to the second wind farm. S220: If it is determined that the propagation path exists, then the theoretical time delay Δt of wind energy propagating from the first wind farm to the second wind farm is calculated based on the geographical distance between the first wind farm and the second wind farm and the average wind speed corresponding to the propagation path. S230: Using the first wind farm and the second wind farm as nodes, and the theoretical time delay Δt as the core weight of the directed edge from the first wind farm to the second wind farm, construct the corresponding directed edge; S240: Traverse all wind farm node pairs within the target area, complete the construction of all directed edges, and generate the initial adjacency matrix.
3. The method for identifying extreme wind power ramping events under regional coupled spatiotemporal evolution according to claim 2, characterized in that, Step S300 specifically includes: S310: Divide the historical power data into multiple consecutive sub-time slices according to a preset time interval to obtain the wind farm power data under each sub-time slice; S320: For each sub-time slice, perform the following operations: S321: Calculate the cross-correlation coefficient ρ(τ) of the historical power sequences of any two wind farms within this sub-time slice under time delay τ; S322: Find the time delay τ_max that maximizes the cross-correlation coefficient ρ(τ), compare the time delay τ_max with the theoretical time delay Δt of the corresponding node pair, and calculate the absolute difference between the two. S323: Determine the weighting factor based on the comparison result between the absolute difference and the preset tolerance threshold, and fuse the weighting factor with the weight of the corresponding directed edge in the initial adjacency matrix to obtain the fused edge weight set under the sub-time slice; S324: Based on the fused edge weight set, correct the initial adjacency matrix and generate the dynamic spatiotemporal coupling graph corresponding to the sub-time slice; S330: Sequentially combine the dynamic spatiotemporal coupling diagrams corresponding to all sub-time slices in chronological order to obtain the dynamic evolution spatiotemporal coupling diagram sequence.
4. The method for identifying extreme wind power ramping events under regional coupled spatiotemporal evolution according to claim 3, characterized in that, The weighting factor is determined based on the comparison between the absolute difference and the preset tolerance threshold, specifically as follows: If the absolute difference is less than or equal to the preset tolerance threshold, then the weighting factor is determined to be the maximum value of the cross-correlation coefficient ρ(τ); If the absolute difference is greater than the preset tolerance threshold, the weighting factor is determined as the product of the reverse inhibition coefficient and the maximum value of the cross-correlation coefficient ρ(τ), wherein the reverse inhibition coefficient is a value less than 1 and is inversely proportional to the absolute difference.
5. The method for identifying extreme wind power ramping events under regional coupled spatiotemporal evolution according to claim 1, characterized in that, The dual-stream adaptive spatiotemporal hypergraph network includes a temporal branch and a spatial branch. The temporal branch uses causal dilated convolution to capture long-range extreme temporal dependencies, while the spatial branch uses hypergraph convolution to mine multi-scale regional coupling patterns. In step S400, the extraction of spatiotemporal fusion features specifically includes: S410: Input the spatiotemporal coupling graph sequence and the node features of each wind farm into the temporal branch, and capture the temporal dependency features at different time scales through multiple parallel causal-hole convolutional layers with different hole rates and splice them to obtain the multi-scale temporal features of each wind farm node. S420: Input the spatiotemporal coupling graph sequence and the node features of each wind farm into the spatial branch, mine the multi-scale spatial coupling pattern of the regional wind farm through hypergraph convolution operation, and output the spatial features of each wind farm node after hypergraph convolution enhancement. S430: Adaptively weightedly fuse the multi-scale temporal features and spatial features corresponding to each wind farm node to generate the spatiotemporal fusion feature.
6. The method for identifying extreme wind power ramping events under regional coupled spatiotemporal evolution according to claim 5, characterized in that, Step S400 obtains the predicted quantiles of power for each wind farm in the future time period, specifically including: S440: Determine the preset target quantiles as the 10th quantile, 50th quantile, and 90th quantile; S450: Construct a regression output layer for a dual-stream adaptive spatiotemporal hypergraph network. The regression output layer is based on an asymmetric quantile regression loss function and is pre-trained using the preset target quantile. S460: The pre-trained regression output layer performs nonlinear mapping analysis on the spatiotemporal fusion features to generate the 10th, 50th, and 90th quantile prediction results of the power of each wind farm in the future time period.
7. The method for identifying extreme wind power ramping events under regional coupled spatiotemporal evolution according to claim 6, characterized in that, Step S450 involves training the regression output layer using an asymmetric quantile regression loss function, specifically including: A training sample set covering scenarios of normal power fluctuations and extreme ramping events was collected. The training sample set includes historical spatiotemporal coupling graph sequences, wind farm node characteristics, and actual power values for the corresponding time periods. A network structure for the regression output layer is constructed. The spatiotemporal coupling graph sequence of the training sample set and the node features of the wind farm are used as inputs, and the actual power value of the corresponding time period is used as the label for training. The quantile loss between the predicted value and the true value under each preset target quantile is calculated. Calculate the instantaneous ramp rate at each time point based on historical power data in the training sample set; When the absolute value of the instantaneous ramp rate exceeds the preset sensitivity threshold, a penalty weight greater than 1 is assigned to the quantile loss at that time point. An asymmetric quantile regression loss function is constructed by weighting each quantile loss with its corresponding penalty weight, and the regression output layer is iteratively trained until convergence.
8. The method for identifying extreme wind power ramping events under regional coupled spatiotemporal evolution according to claim 1, characterized in that, Step S500 specifically includes: S510: Obtain the dynamic ramp threshold that adapts to the power grid operating conditions and wind farm operating characteristics of the target area; S520: Calculate the baseline ramp rate R50 based on the predicted power of the 50th quantile, calculate the pessimistic ramp rate R10 based on the predicted power of the 10th quantile, and calculate the optimistic ramp rate R90 based on the predicted power of the 90th quantile. S530: Extreme climbing events are determined according to the following logic: If the absolute value of the baseline ramp rate R50 exceeds the dynamic ramp threshold, it is determined to be a high-confidence extreme ramp event. If the absolute value of the baseline ramp rate R50 does not exceed the dynamic ramp threshold, but the absolute value of the pessimistic scenario ramp rate R10 or the optimistic scenario ramp rate R90 exceeds the dynamic ramp threshold, it is determined to be a low-confidence, high-risk extreme ramp event. If none of the above conditions are met, it is determined that there are no extreme climbing events. S540: Based on the judgment result, synchronously output the confidence level of the event, the range of event intensity represented by the numerical ranges of R10, R50, and R90, and the expected occurrence time window of the event.
9. A wind power extreme ramping event identification system under regional coupled spatiotemporal evolution, characterized in that, include: The data acquisition module is used to acquire historical power data and historical weather forecast data of each wind farm in the target area; The initial adjacency matrix construction module is used to calculate the propagation path and theoretical time delay of wind energy between wind farms based on the historical weather forecast data, and to construct an initial adjacency matrix with wind farms as nodes. The spatiotemporal coupling graph sequence generation module is used to calculate the time-delay cross-correlation between the power data of each wind farm based on the historical power data, perform weighted correction on the initial adjacency matrix, and generate a dynamically evolving spatiotemporal coupling graph sequence. The spatiotemporal fusion feature extraction module is used to input the spatiotemporal coupling graph sequence and the node features of each wind farm into a pre-trained dual-stream adaptive spatiotemporal hypergraph network, extract spatiotemporal fusion features, and obtain the predicted quantiles of the power of each wind farm in the future time period. The predicted quantiles include the 10th quantile, the 50th quantile, and the 90th quantile. The extreme ramp event determination module is used to calculate the power ramp rate based on the predicted quantile, and determine whether an extreme ramp event has occurred by combining the dynamic ramp threshold. At the same time, it outputs the confidence level, intensity range and expected occurrence time window of the event.
10. The wind power extreme ramping event identification system under regional coupled spatiotemporal evolution according to claim 9, characterized in that, The dual-stream adaptive spatiotemporal hypergraph network includes a temporal branch and a spatial branch; the temporal branch employs multiple parallel causal dilated convolutional layers with different dilation rates to capture long-range extreme temporal dependency features at different time scales. The spatial branch employs a hypergraph convolutional layer to mine multi-scale high-order spatial coupling patterns between wind farms within the region; The spatiotemporal fusion feature extraction module has a built-in pre-trained convergent regression output layer. The regression output layer is pre-trained with an asymmetric quantile regression loss function and is used to perform nonlinear mapping analysis on the input spatiotemporal fusion features, and output the prediction results of the 10th quantile, 50th quantile and 90th quantile of the power of each wind farm in the future time period. The extreme ramp event determination module is specifically used to calculate the baseline ramp rate R50, the pessimistic scenario ramp rate R10, and the optimistic scenario ramp rate R90 based on the prediction results of the 10th quantile, 50th quantile, and 90th quantile, respectively, and execute the following determination logic: If the absolute value of the baseline ramp rate R50 exceeds the dynamic ramp threshold, it is determined to be a high-confidence extreme ramp event. If the absolute value of the baseline ramp rate R50 does not exceed the dynamic ramp threshold, but the absolute value of the pessimistic scenario ramp rate R10 or the optimistic scenario ramp rate R90 exceeds the dynamic ramp threshold, it is determined to be a low-confidence, high-risk extreme ramp event. If none of the above conditions are met, it is determined that there is no extreme climbing event.
