A power grid stability margin prediction method and device based on a multi-source model stack

Abstract

The application discloses a power grid stability margin prediction method and device based on a multi-source model stack, and belongs to the technical field of power grid stable operation. The method comprises the following steps: performing two-stage weighted fusion on multi-source weather data of a new energy grid-connected node to generate a day-ahead power prediction result of the new energy grid-connected node; constructing a stability margin state feature vector and a support response feature, and calculating a node sensitivity index; weighting the stability margin state feature vector according to the node sensitivity index to obtain a weighted feature vector; inputting the weighted feature vector into multiple primary base models and then into a secondary meta-learning model to obtain an initial prediction result of the power grid stability margin; and performing consistency correction on the initial prediction result of the power grid stability margin to obtain a final prediction result of the power grid stability margin. The application can directly output a stability margin prediction value in a day-ahead stage, and realizes fast prediction of a deterministic stability margin for day-ahead scheduling.

Description

Technical Field

[0001] This invention relates to the field of power grid stable operation technology, and in particular to a method and apparatus for predicting power grid stability margin based on multi-source model stacking. Background Technology

[0002] As the penetration rate of new energy sources such as wind power and photovoltaics in the power system continues to rise, dispatching departments need to predict the trend of stability margin changes under different operating modes in advance during the day-ahead phase, so as to arrange unit start-up and shutdown, reactive power allocation, and section control strategies. If stability margin prediction is not carried out in sync with new energy output prediction, it often leads to a disconnect between day-ahead plans and actual stability constraints.

[0003] In existing technologies, one type of method focuses on predicting renewable energy power output, only outputting wind and solar power output results, which are then used by dispatchers to assess stability risks based on experience. Another type of method focuses on mechanism modeling or simulation solutions, requiring repeated power flow analysis, small disturbance analysis, or time-domain scanning for each candidate operating condition. This results in a large computational load and makes it difficult to use quickly under day-ahead batch operating conditions. Especially under conditions of high-proportion renewable energy integration, the number of operating modes is large and operating conditions change rapidly; relying solely on offline repetitive simulations is insufficient to meet the timeliness requirements of dispatching.

[0004] Furthermore, traditional single-model forecasting methods fail to adequately utilize multi-source meteorological information, making it difficult to consider the adaptability of different weather sources and algorithms to local weather conditions, complex terrain, and extreme operating conditions. Existing stability assessment methods typically lack explicit identification and error correction mechanisms for weak points, resulting in insufficient interpretability of forecast results and difficulty in directly guiding dispatch and response. Some probabilistic risk analysis schemes also rely on extensive scene sampling, distribution reconstruction, and tail fraction calculation, which face challenges such as large sample sizes, long modeling chains, and high online refresh costs during engineering deployment. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a power grid stability margin prediction method and device based on multi-source model stacking, which can directly output stability margin prediction values ​​at the day-ahead stage and realize deterministic stability margin prediction for day-ahead scheduling.

[0006] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0007] On the one hand, this invention provides a power grid stability margin prediction method based on multi-source model stacking, comprising:

[0008] Based on the preset data base model, the multi-source meteorological data of the new energy grid-connected nodes are weighted and fused in two levels to generate the day-ahead power prediction results of the new energy grid-connected nodes;

[0009] Based on the day-ahead power prediction results, a stability margin state feature vector is constructed. Based on the electrical distance of the new energy grid-connected nodes, short-circuit support strength, conventional unit support coverage, and power fluctuation degree, node sensitivity index is calculated.

[0010] Based on the node sensitivity index, the stability margin state feature vector is weighted to obtain a weighted feature vector.

[0011] Based on the day-ahead power prediction results, short-circuit support information, reactive power support information, and reactive power reserve information, construct the support response characteristics;

[0012] The weighted feature vectors are input into multiple first-level basis models to obtain multiple stability margin sub-prediction results; the support response features, node sensitivity indices and the stability margin sub-prediction results are input into a second-level meta-learning model to obtain the initial prediction result of the power grid stability margin.

[0013] Based on pre-constructed power reserve constraints, stability margin boundary constraints, and adjacent time period smoothing constraints, the initial prediction results of the power grid stability margin are corrected for consistency, and the final prediction results of the power grid stability margin are obtained.

[0014] Optionally, the formula for calculating the day-ahead power prediction result of the new energy grid-connected node is as follows: ; ; The intra-source weights and inter-source weights are obtained by minimizing the objective function, which is expressed as: ; Based on the support response characteristics and the final prediction results of the power grid stability margin, the training sample weights are updated using the following formula: ; in, express New energy grid connection nodes The The combined forecast results within each meteorological data source; Indicates the grid connection node of new energy sources The The first meteorological data source Source weights of each data base model; express New energy grid connection nodes The The first meteorological data source The prediction results of the data-based model; Indicates the number of data base models; express New energy grid connection nodes The day-ahead power forecast results; Indicates the grid connection node of new energy sources The Inter-source weights of meteorological data sources; Indicates the number of meteorological data sources; Represent the objective function; Represents the set of training sample time periods; express Training sample weights at any time; express New energy grid connection nodes Actual output; All represent regularization coefficients; This represents the updated training sample weights; Both indicate updating the weight coefficients; This indicates updating the training sample time period set; They represent time, Time-based historical stability margin label samples; They represent time, The final prediction result of the power grid stability margin at time t; They represent time, Weighted support response characteristics of all sensitive nodes across the network at all times; Represents positive numbers; This indicates taking the maximum value.

[0015] Optionally, the formula for calculating the stability margin state feature vector is: ; ; ; in, These represent new energy grid connection node 1 and new energy grid connection node, respectively. New energy grid connection nodes The change in power between adjacent time periods; They represent New Energy Grid Connection Node 1, New Energy Grid Connection Node The day-ahead power forecast results; They represent time, New energy grid connection nodes The day-ahead power prediction results; They represent New Energy Grid Connection Node 1, New Energy Grid Connection Node New energy grid connection nodes Historical prediction residual statistics; Indicates the length of the residual statistical time window; express New energy grid connection nodes Actual output; express New energy grid connection nodes The day-ahead power forecast results; express Constantly rotating standby feature; express Constant reactive power support characteristics; express Key cross-sectional tidal characteristics at all times; express Time-based network structure characteristics; This represents the statistical characteristics of historical stability margin; express Time-series stability margin state feature vector; This indicates the matrix transpose.

[0016] Optionally, the formula for calculating the node sensitivity index is: ; ; ; ; ; in, Indicates the grid connection node of new energy sources To strong support busbar assembly The equivalent electrical distance; Indicates strong support busbar Weighting coefficients; Indicates the grid connection node of new energy sources With strong support busbar Equivalent impedance elements between; Indicates the grid connection node of new energy sources The short-circuit support coefficient; Indicates the grid connection node of new energy sources The short-circuit capacity; Indicates the grid connection node of new energy sources Rated grid-connected capacitor; Represents positive numbers; Indicates the grid connection node of new energy sources The coverage of conventional unit support; Indicates the number of generating units; Indicates the unit Available spare capacity; Indicates the grid connection node of new energy sources to the unit Electrical distance; express New energy grid connection nodes The power fluctuation coefficient; express New energy grid connection nodes The day-ahead power forecast results; express New energy grid connection nodes In length of The average predicted output within the time window; express New energy grid connection nodes Node sensitivity index; The weighting coefficients representing the node sensitivity index; express The normalized value; express The normalized value; express The normalized value; express The normalized value.

[0017] Optionally, the formula for calculating the weighted feature vector is: ; ; in, express Time and A feature node sensitivity index vector of the same dimension; express Time-of-flight feature correlation matrix; express Time-node sensitivity index vector; express Time-weighted feature vector; express Time-series stability margin state feature vector; Indicates the sensitivity amplification factor; This indicates element-wise multiplication.

[0018] Optionally, the formula for calculating the support response characteristics is: ; ; in, express New energy grid connection nodes Support response characteristics; express New energy grid connection nodes The day-ahead power forecast results; Indicates the grid connection node of new energy sources The short-circuit capacity; All represent adjustment coefficients; Represents positive numbers; express New energy grid connection nodes Reactive power support margin in the region; express The system continuously rotates the total reserve capacity. express Weighted support response characteristics of all sensitive nodes across the network at all times; express New energy grid connection nodes Node sensitivity index; This indicates the number of new energy grid-connected nodes.

[0019] Optionally, the stability margin sub-prediction result is expressed as: ; The initial prediction result of the power grid stability margin is expressed as follows: ; Multiple first-level base models and second-level meta-learning models are jointly trained using a loss function, which is expressed as follows: ; in, They represent The first-order basis model, the second-order basis model, and the first time step are all based on the first-order basis model. The first-level basic model, the first The stability margin sub-prediction results output by each first-level basis model; Indicates the first One first-level basis model; express Time-weighted feature vector; express Initial prediction results of grid stability margin at time point; express Weighted support response characteristics of all sensitive nodes across the network at all times; express Time-node sensitivity index vector; This represents a two-dimensional meta-learning model; Represents the loss function; Represents the set of training sample time periods; express Time-based historical stability margin label samples; All of these represent the weighting coefficients of the loss function; Represents the set of model parameters; This represents the L2 norm.

[0020] Optionally, the data base model includes a categorical feature enhancement model, an extreme gradient enhancement model, a random forest model, and a generalized additive model; The first-level base model includes at least two of the following: the class feature boosting model, the extreme gradient boosting model, the random forest model, and the generalized additive model. The second-order meta-learning model includes any one of the following: linear regression model, ridge regression model, Lasso model, support vector regression model, and lightweight neural network model.

[0021] Optionally, the formula for calculating the final prediction result of the power grid stability margin is as follows: ; ; ; ; in, They represent The lower boundary of the stability margin and the upper boundary of the stability margin at any given time; Both represent the coefficients of the lower boundary of the stability margin; express The system continuously rotates the total reserve capacity. express Real-time system reactive power support margin; express Standard unit support capacity indicators at all times; express Total power injected by new energy sources at all times; Both represent the coefficients of the upper boundary of the stability margin; express Time-of-flight system short-circuit support indicators; express The target stability margin value corresponding to the power supply backup constraint at any given time; All of these represent coefficients for power reserve constraints; They represent time, The final prediction result of the power grid stability margin at time t; This represents the stability margin of the independent variable that minimizes the objective function. The value of ; express Initial prediction results of grid stability margin at time point; All of these represent coefficients representing the final predicted results of the power grid stability margin; This indicates taking the maximum value.

[0022] Optional, also includes; Based on the node sensitivity masking analysis, the contribution ranking results of each new energy grid-connected node to the stability margin are generated; Based on the final prediction results of the grid stability margin and the ranking results of the contribution of each new energy grid-connected node to the stability margin, a stability early warning level and a weak node identification result are generated.

[0023] Optionally, the formula for calculating the contribution of each new energy grid-connected node to the stability margin is as follows: ; The formula for calculating the stability warning level is as follows: ; The contribution of each new energy grid-connected node to the stability margin is ranked from largest to smallest, and the top A new energy grid-connected nodes in the ranking result are identified as weak nodes. in, express New energy grid connection nodes Contribution to stability margin; express New energy grid connection nodes Node sensitivity index; express Initial prediction results of grid stability margin at time point; They represent The first-order basis model, the second-order basis model, and the first time step are all based on the first-order basis model. The stability margin sub-prediction results output by each first-level basis model; express Weighted support response characteristics of all sensitive nodes across the network at all times; express Always keep the new energy grid connection nodes The sensitivity vector after the corresponding sensitivity components are set to zero; This represents a two-dimensional meta-learning model; express The level of stability warning at any given moment; express The final prediction result of the power grid stability margin at time t; All represent preset thresholds; A represents the preset number of recognitions.

[0024] On the other hand, the present invention provides a power grid stability margin prediction device based on multi-source model stacking, for performing the method described in the first aspect, comprising:

[0025] The data fusion module is used to: perform two-level weighted fusion of multi-source meteorological data of new energy grid-connected nodes according to a preset data base model, and generate day-ahead power prediction results for new energy grid-connected nodes;

[0026] The sensitivity calculation module is used to: construct a stability margin state feature vector based on the day-ahead power prediction results, and calculate the node sensitivity index based on the electrical distance, short-circuit support strength, conventional unit support coverage and power fluctuation degree of the new energy grid-connected nodes;

[0027] The feature weighting module is used to: weight the stability margin state feature vector according to the node sensitivity index to obtain a weighted feature vector;

[0028] The feature construction module is used to: construct support response features based on the day-ahead power prediction results, short-circuit support information, reactive power support information, and reactive power reserve information;

[0029] The model prediction module is used to: input the weighted feature vector into multiple first-level base models to obtain multiple stability margin sub-prediction results; and input the support response features, node sensitivity index and the stability margin sub-prediction results into a second-level meta-learning model to obtain the initial prediction result of the power grid stability margin.

[0030] The prediction correction module is used to: perform consistency correction on the initial prediction result of the power grid stability margin based on pre-constructed power reserve constraints, stability margin boundary constraints and adjacent time period smoothing constraints, so as to obtain the final prediction result of the power grid stability margin.

[0031] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0032] This invention fully utilizes multi-source meteorological data and multi-model information through two-level fusion to improve the accuracy and stability of new energy power prediction; by introducing node sensitivity indicators and support response characteristics, it enables stability margin prediction to explicitly characterize the impact of weak nodes on stability margin, improving the interpretability of the results; through a consistency correction mechanism, it ensures that the prediction results meet the requirements of reserve constraints, support constraints, and time series continuity, enhancing engineering practicality; and by directly outputting stability margin prediction values, it is more suitable for day-ahead operation mode verification and rapid response. Attached Figure Description

[0033] Figure 1A flowchart illustrating the power grid stability margin prediction method based on multi-source model stacking provided in an embodiment of the present invention;

[0034] Figure 2 A flowchart illustrating the day-ahead power prediction results for new energy grid-connected nodes provided in an embodiment of the present invention;

[0035] Figure 3 A schematic diagram showing the comparison between the day-ahead power prediction result and the actual value of node 1 provided in an embodiment of the present invention;

[0036] Figure 4 A schematic diagram showing the comparison between the day-ahead power prediction result and the actual value of node 2 provided in this embodiment of the invention;

[0037] Figure 5 A schematic diagram showing the comparison between the day-ahead power prediction result and the actual value of node 3 provided in this embodiment of the invention;

[0038] Figure 6 A schematic diagram showing the comparison between the day-ahead power prediction result and the actual value of node 4 provided in this embodiment of the invention;

[0039] Figure 7 This is a schematic diagram comparing the day-ahead power prediction result of node 5 with the actual value provided in an embodiment of the present invention. Detailed Implementation

[0040] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0041] The term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0042] Example 1

[0043] This embodiment introduces a power grid stability margin prediction method based on multi-source model stacking, including:

[0044] Based on the preset data base model, the multi-source meteorological data of the new energy grid-connected nodes are weighted and fused in two levels to generate the day-ahead power prediction results of the new energy grid-connected nodes;

[0045] Based on the day-ahead power prediction results, a stability margin state feature vector is constructed. Based on the electrical distance of the new energy grid-connected nodes, short-circuit support strength, conventional unit support coverage, and power fluctuation degree, node sensitivity index is calculated.

[0046] Based on the node sensitivity index, the stability margin state feature vector is weighted to obtain a weighted feature vector.

[0047] Based on the day-ahead power prediction results, short-circuit support information, reactive power support information, and reactive power reserve information, construct the support response characteristics;

[0048] The weighted feature vectors are input into multiple first-level basis models to obtain multiple stability margin sub-prediction results; the support response features, node sensitivity indices and the stability margin sub-prediction results are input into a second-level meta-learning model to obtain the initial prediction result of the power grid stability margin.

[0049] Based on pre-constructed power reserve constraints, stability margin boundary constraints, and adjacent time period smoothing constraints, the initial prediction results of the power grid stability margin are corrected for consistency, and the final prediction results of the power grid stability margin are obtained.

[0050] like Figure 1 As shown in this embodiment, a power grid stability margin prediction method based on multi-source model stacking includes the following steps:

[0051] Step 1: Generate day-ahead power prediction results using a two-level fusion strategy, specifically as follows:

[0052] Multi-source meteorological data includes multi-source meteorological forecast data corresponding to each new energy grid-connected node, as well as historical output data, historical dispatch plans, power grid topology, conventional unit operation mode, system reserve information, reactive power support information, and historical stability margin label samples.

[0053] Historical stability margin label samples can be derived from production system safety verification results, offline simulation platform calculation results, or historical operating results confirmed by scheduling. They are calculated by an offline stability verification program. Specifically, the renewable energy injection plan, conventional unit output, system reserve information, network topology, and reactive power support configuration for a historical period are input into a preset stability verification program to obtain historical stability margin label samples for the corresponding period. The stability verification program can be a stability verification module in the production scheduling environment, or it can be an offline analysis module consistent with the production model.

[0054] Data from different sources are processed through time alignment, missing value imputation, outlier removal, and uniform standardization to obtain training and prediction sample sets. Preferably, missing values ​​are imputed by combining interpolation of nearby time periods with the mean of similar weather samples; outliers are corrected by quantile threshold truncation or isolated forest identification.

[0055] Based on a pre-defined data base model, multi-source meteorological data from renewable energy grid-connected nodes are fused using a two-level weighted fusion method. A stacked fusion strategy with both intra-source and inter-source weighted fusion is employed to predict the day-ahead power output, photovoltaic output, or integrated renewable energy injection power of each renewable energy grid-connected node, generating high-precision day-ahead power prediction results for the renewable energy grid-connected nodes. In this embodiment, the data base model includes Categorical Boosting (CatBoost), Extreme Gradient Boosting (XGBoost), Generalized Additive Model (GAM), and Random Forest models. The multi-source meteorological data includes data from the German Meteorological Service (DWD) and the Global Forecast System (GFS).

[0056] like Figure 2 As shown, first prepare the data source model pool. DWD and GFS each have 4 models. DWD contains CatBoost, XGBoost, GAM and Random Forest models. GFS contains CatBoost, XGBoost, GAM and Random Forest models.

[0057] First, advanced DWD (Digital Dynamics) and GFS (Geometry and Precision Fiber) source fusions are performed to obtain DWD combined prediction results and GFS combined prediction results. Then, the DWD combined prediction results and GFS combined prediction results are fused between sources to obtain the final point prediction, that is, the day-ahead power prediction result of the new energy grid-connected node.

[0058] The formula for calculating the day-ahead power prediction results of renewable energy grid-connected nodes is as follows:

[0059] ;

[0060] ;

[0061] satisfy, ;

[0062] The intra-source weights and inter-source weights are obtained by minimizing the objective function, which is expressed as:

[0063] ;

[0064] in, express New energy grid connection nodes The The combined forecast results within each meteorological data source; Indicates the grid connection node of new energy sources The The first meteorological data source Source weights of each data base model; express New energy grid connection nodes The The first meteorological data source The prediction results of the data-based model; Indicates the number of data base models; express New energy grid connection nodes The day-ahead power forecast results; Indicates the grid connection node of new energy sources The Inter-source weights of meteorological data sources; Indicates the number of meteorological data sources; Represent the objective function; Represents the set of training sample time periods; express The weights of training samples can be set according to weather type, degree of extreme fluctuation, or key periods in the day before, so that the model pays more attention to high-risk working conditions; express New energy grid connection nodes Actual output; Both represent regularization coefficients.

[0065] Step 2: Construct the stability margin state feature vector, specifically as follows:

[0066] A stability margin state feature vector is constructed based on the day-ahead power prediction results of new energy grid connection nodes.

[0067] First, renewable energy injection features are extracted, including node power level, ramp rate, changes between adjacent time periods, mean of historical prediction residuals, and variance of historical prediction residuals. Among these, the renewable energy injection features used to construct the stability margin state feature vector include changes between adjacent time periods and historical prediction residuals, as shown in the formula:

[0068] ;

[0069] ;

[0070] Simultaneously, operational characteristics such as the start-up combination, spinning reserve level, reactive power support capacity, tie line power flow, and transmission power at key sections of conventional power sources are extracted; further structural characteristics such as grid short-circuit capacity, electrical distance, network partitioning characteristics, topology switching identifiers, and historical stability margin statistics are extracted.

[0071] The above information is concatenated to form a stability margin state feature vector, as shown in the formula:

[0072] ;

[0073] in, These represent new energy grid connection node 1 and new energy grid connection node, respectively. New energy grid connection nodes The change in power between adjacent time periods; They represent New Energy Grid Connection Node 1, New Energy Grid Connection Node The day-ahead power forecast results; express New energy grid connection nodes The day-ahead power forecast results; They represent New Energy Grid Connection Node 1, New Energy Grid Connection Node New energy grid connection nodes Historical prediction residual statistics; Indicates the length of the residual statistical time window; express New energy grid connection nodes Actual output; express New energy grid connection nodes The day-ahead power forecast results; express Constantly rotating standby feature; express Constant reactive power support characteristics; express Key cross-sectional tidal characteristics at all times; express Time-based network structure characteristics; This represents the statistical characteristics of historical stability margin; express Time-series stability margin state feature vector; This indicates the matrix transpose.

[0074] Step 3: Calculate the node sensitivity index, specifically:

[0075] To enhance the responsiveness of stability margin prediction to weak nodes, a node sensitivity index is calculated based on the electrical distance, short-circuit support strength, conventional unit support coverage, and power fluctuation of the new energy grid-connected nodes. The higher the node sensitivity index, the more likely that node is to dominate the narrowing of the system stability margin.

[0076] First, the equivalent electrical distance from the node to the set of strongly supported buses is calculated based on the network equivalent impedance matrix, using the following formula:

[0077] ;

[0078] Secondly, the short-circuit support factor is calculated based on the short-circuit capacity and the rated capacity of the node, using the following formula:

[0079] ;

[0080] Then, the support coverage of conventional units is calculated based on the relationship between the available reserve capacity of conventional units and the electrical location of nodes, using the following formula:

[0081] ;

[0082] Then, the power fluctuation coefficient is calculated based on the degree of fluctuation of the predicted output within the sliding time window. The formula is as follows:

[0083] ;

[0084] Finally, after normalizing the above indicators, a node sensitivity index is synthesized according to preset weights. The formula is as follows:

[0085] ;

[0086] The normalized index is obtained in the following way:

[0087] ;

[0088] in, Indicates the grid connection node of new energy sources To strong support busbar assembly The equivalent electrical distance; Indicates strong support busbar Weighting coefficients; Indicates the grid connection node of new energy sources With strong support busbar Equivalent impedance elements between; Indicates the grid connection node of new energy sources The short-circuit support coefficient; Indicates the grid connection node of new energy sources The short-circuit capacity; Indicates the grid connection node of new energy sources Rated grid-connected capacitor; To represent positive numbers, to prevent the denominator from being zero; Indicates the grid connection node of new energy sources The coverage of conventional unit support; Indicates the number of generating units; Indicates the unit Available spare capacity; Indicates the grid connection node of new energy sources to the unit Electrical distance; express New energy grid connection nodes The power fluctuation coefficient; express New energy grid connection nodes The day-ahead power forecast results; express New energy grid connection nodes In length Average predicted output within the time window; express New energy grid connection nodes Node sensitivity index; All node sensitivity indexes are represented by weight coefficients, and satisfy the following conditions: ; express The normalized value; express The normalized value; express The normalized value; express The normalized value; express New energy grid connection nodes To strong support busbar assembly Equivalent electrical distance / new energy grid connection node Short-circuit support coefficient / new energy grid connection node Short-circuit capacity / new energy grid connection nodes The normalized value of the power fluctuation coefficient; express New energy grid connection nodes To strong support busbar assembly Equivalent electrical distance / new energy grid connection node Short-circuit support coefficient / new energy grid connection node Short-circuit capacity / new energy grid connection nodes The power fluctuation coefficient before normalization; express time The minimum and maximum values.

[0089] Step 4: Feature weighting and constructing supporting response features, specifically:

[0090] Based on the node sensitivity index, the stability margin state feature vector is weighted to obtain the weighted feature vector.

[0091] Construct a feature correlation matrix, mapping the node sensitivity index to a feature sensitivity vector of the same dimension as the stability margin state feature vector, as shown in the formula:

[0092] ;

[0093] Then, a weighted feature vector is generated based on the feature sensitivity vector, using the following formula:

[0094] ;

[0095] in, express Time and A feature node sensitivity index vector of the same dimension; express The time-feature correlation matrix, and satisfies ; The dimension of the stability margin state eigenvector is represented when the eigenvector is... The stability margin state feature vector corresponds to the new energy grid-connected node. When considering local power, local residual, or local fluctuation characteristics, matrix elements Take 1 if the first value is 1, otherwise take 0. If the stability margin state feature vector is a system-level feature, then the sensitivity of multiple nodes is linearly combined using preset weights; express Time-node sensitivity index vector; express Time-weighted feature vector; Indicates the sensitivity amplification factor; This indicates element-wise multiplication.

[0096] Based on the day-to-day power forecast, short-circuit support information, reactive power support information, and reactive power reserve information, support response characteristics are constructed. To further characterize the matching relationship between renewable energy injection and support resources, support response characteristics are constructed for each node, with the following formula:

[0097] ;

[0098] The weighted support response characteristics of all nodes are obtained by summing the support response characteristics of all sensitive nodes in the network according to their node sensitivity. The formula is as follows:

[0099] ;

[0100] in, express New energy grid connection nodes Support response characteristics; All represent adjustment coefficients; Represents positive numbers; express New energy grid connection nodes Reactive power support margin in the area; express The system continuously rotates the total reserve capacity. express Weighted support response characteristics of all sensitive nodes across the network at all times; This indicates the number of new energy grid-connected nodes.

[0101] Step 5: Stability margin prediction based on a multi-level model, specifically:

[0102] First-level basis models are used to predict the stability margin of weighted feature vectors. The model type can be the same as the data basis model, but the training objective, input features and output results are different. First-level basis models include at least two of the following: class feature boosting model, extreme gradient boosting model, random forest model and generalized additive model. Second-level meta-learning models include any one of the following: linear regression model, ridge regression model, Least Absolute Shrinkage and Selection Operator (Lasso) model, support vector regression model and lightweight neural network model.

[0103] By inputting the weighted feature vectors into multiple first-level basis models, multiple stability margin sub-prediction results are obtained, represented as follows:

[0104] ;

[0105] By inputting the support response features, node sensitivity indices, and the stability margin sub-prediction results into the two-dimensional learning model, the initial prediction result of the power grid stability margin is obtained, expressed as:

[0106] ;

[0107] Multiple first-level base models and second-level meta-learning models are jointly trained using a loss function, which is expressed as follows:

[0108] ;

[0109] in, They represent The first-order basis model, the second-order basis model, and the first time step are all based on the first-order basis model. The first-level basic model, the first The stability margin sub-prediction results output by each first-level basis model; Indicates the first One first-level basis model; express Initial prediction results of grid stability margin at time point; express Time-node sensitivity index vector; This represents a two-dimensional meta-learning model; Represents the loss function; Represents the set of training sample time periods; All of these represent the weighting coefficients of the loss function; Represents the set of model parameters; This represents the L2 norm.

[0110] In this embodiment, cross-validation is first used to generate the out-of-bounds prediction results of the first-level base model, and then the out-of-bounds prediction results are used to train the second-level meta-learning model to reduce the risk of overfitting.

[0111] Step Six: Perform consistency correction on the initial prediction results of the power grid stability margin, specifically as follows:

[0112] Based on pre-constructed power reserve constraints, stability margin boundary constraints, and adjacent time period smoothing constraints, the initial prediction results of the power grid stability margin are consistent to obtain the final prediction results. Specifically, the stability margin boundary constraints are implemented through upper and lower stability margin boundaries, the power reserve constraints are implemented through reserve constraint correction terms, and the adjacent time period smoothing constraints are implemented by incorporating the final prediction results of the power grid stability margin from the previous time period. To ensure that the prediction results meet engineering operation constraints, upper and lower stability margin boundaries are constructed, using the following formulas:

[0113] ;

[0114] ;

[0115] The power reserve constraint formula is:

[0116] Based on this, a consistency correction is performed on the initial predicted value of the stability margin, using the following formula:

[0117] ;

[0118] in, They represent The lower boundary and upper boundary of the stability margin at any given time. ; All represent the coefficients of the lower boundary of the stability margin, and satisfy... ; express Real-time system reactive power support margin; express Standard unit support capacity indicators at all times; express Total power injected by new energy sources at all times; All represent the coefficients of the upper boundary of the stability margin, and satisfy... ; express Time-of-flight system short-circuit support indicators; express The target stability margin value corresponding to the power supply backup constraint at any given time; All of these represent coefficients for power reserve constraints; They represent time, The final prediction result of the power grid stability margin at time t; This represents the stability margin of the independent variable that minimizes the objective function. The value of ; All represent coefficients of the final predicted result of the power grid stability margin, and satisfy the following conditions: ; This indicates taking the maximum value.

[0119] Step 7: Generate stability warning levels and weak node identification results, specifically:

[0120] Based on the node sensitivity masking analysis, a ranking of the contributions of each renewable energy grid-connected node to the stability margin is generated. The formula for calculating the contribution of each renewable energy grid-connected node to the stability margin is as follows:

[0121] ;

[0122] in, express New energy grid connection nodes Contribution to stability margin; express Always keep the new energy grid connection nodes The sensitivity vector after the corresponding sensitivity components are set to zero.

[0123] Based on the final power grid stability margin prediction results and the ranking of the contributions of each renewable energy grid-connected node to the stability margin, stability warning levels and weak node identification results are generated, namely:

[0124] The contribution of each new energy grid-connected node to the stability margin is ranked from largest to smallest to obtain the time... The weak node ranking results are used to identify the top A new energy grid-connected nodes as weak nodes. This ranking result can be directly used by the dispatching side to locate areas sensitive to stability risks.

[0125] The final predicted power grid stability margin is compared with a preset threshold to determine the early warning level, using the following formula:

[0126] ;

[0127] in, express The level of stability warning at any given moment; All represent preset thresholds, satisfying A represents the preset number of recognitions.

[0128] In engineering implementation, the above thresholds can be adjusted offline according to the dispatching department's needs for handling different risk levels.

[0129] At the same time, the grid-connected new energy nodes with the highest contribution ranking are identified as the set of weak nodes, and corresponding scheduling prompts are output, such as strengthening reactive power support, adjusting the combination of conventional units, optimizing cross-sectional power flow, or restricting the output of local new energy.

[0130] After the target scheduling cycle ends, the actual renewable energy output, actual operating mode, and the final grid stability margin prediction results obtained from the review are fed back into the sample database. For new samples, the training sample weights are updated based on the supporting response characteristics and the final grid stability margin prediction results, using the following formula:

[0131] ;

[0132] in, This represents the updated training sample weights; Both indicate updating the weight coefficients; This indicates updating the training sample time period set; express Time-based historical stability margin label samples; express The final prediction result of the power grid stability margin at time t; express Weighted support response characteristics of all sensitive nodes across the network at all times.

[0133] Samples with larger prediction errors or higher support response characteristics will receive higher update weights for subsequent incremental training. This strategy allows the model to maintain higher sensitivity to recent weak operating conditions and key operating modes.

[0134] This embodiment achieves critical node impact correction through node sensitivity indicators and support response characteristics. The overall goal is to directly output stability margin prediction values, weak node ranking, and graded early warning results during the day-ahead scheduling phase, rather than first reconstructing the stability margin probability distribution and then performing tail statistical analysis. The stability margin is preferably the margin indicator of the target operating mode's distance from the stability constraint boundary under the preset stability verification scenario. Therefore, the technical approach of this embodiment emphasizes deterministic prediction chains, consistency correction mechanisms, and the engineering interpretation capability of the results.

[0135] Example 2

[0136] Based on Example 1, this example introduces an experimental example of a power grid stability margin prediction method based on multi-source model stacking:

[0137] Figure 2 The day-ahead power prediction process for renewable energy grid-connected nodes was demonstrated. Figures 3 to 7 The figures show a comparison between the final predicted grid stability margins and the actual values ​​for five grid-connected nodes. The predicted stability margins include wind power and solar power forecasts. As can be seen from the figures, the multi-source model stacking method can effectively track the power change trends of each node, providing reliable input for subsequent stability margin predictions.

[0138] In this example, the final predicted 24-hour grid stability margin as shown in Table 1 is used as the output. For ease of scheduling interpretation, the stability margin can be divided into three intervals: when the stability margin is greater than 3.0, the system is considered to be in a state of sufficient safety margin; when the stability margin is between 2.0 and 3.0, the system is considered to have entered the range of concern; when the stability margin is less than 2.0, the system is considered to have entered the warning range.

[0139] Table 1 Final prediction results of power grid stability margin

[0140]

[0141] As shown in Table 1, the stability margin is relatively low between 11:00 and 15:00, with a significant narrowing at 13:00 and 14:00. This indicates a significant decrease in system stability margin under the operating mode of higher renewable energy injection and relatively limited conventional support during midday. The ranking of node contributions further identifies the top-ranked key renewable energy grid-connected nodes. These nodes typically possess large electrical distances, low short-circuit support strength, and strong power fluctuation characteristics.

[0142] For example, during the critical midday period, if the short-circuit support coefficient of a node decreases while the power fluctuation coefficient increases, the sensitivity calculation formula will lead to an increase in its node sensitivity index. This, in turn, amplifies the node's impact on the overall stability margin through support response characteristics and stacking prediction models, making the final stability margin prediction more consistent with actual dispatching experience. Dispatchers can then prioritize measures such as enhanced reactive power support, cross-sectional power flow redistribution, or improved support for conventional units in areas with high-ranking nodes.

[0143] Example 3

[0144] Based on Embodiment 1 or 2, this embodiment introduces a power grid stability margin prediction device based on multi-source model stacking, comprising:

[0145] The data fusion module is used to: perform two-level weighted fusion of multi-source meteorological data of new energy grid-connected nodes according to a preset data base model, and generate day-ahead power prediction results for new energy grid-connected nodes;

[0146] The sensitivity calculation module is used to: construct a stability margin state feature vector based on the day-ahead power prediction results, and calculate the node sensitivity index based on the electrical distance, short-circuit support strength, conventional unit support coverage and power fluctuation degree of the new energy grid-connected nodes;

[0147] The feature weighting module is used to: weight the stability margin state feature vector according to the node sensitivity index to obtain a weighted feature vector;

[0148] The feature construction module is used to: construct support response features based on the day-ahead power prediction results, short-circuit support information, reactive power support information, and reactive power reserve information;

[0149] The model prediction module is used to: input the weighted feature vector into multiple first-level base models to obtain multiple stability margin sub-prediction results; and input the support response features, node sensitivity index and the stability margin sub-prediction results into a second-level meta-learning model to obtain the initial prediction result of the power grid stability margin.

[0150] The prediction correction module is used to: perform consistency correction on the initial prediction result of the power grid stability margin based on pre-constructed power reserve constraints, stability margin boundary constraints and adjacent time period smoothing constraints, so as to obtain the final prediction result of the power grid stability margin.

[0151] The specific functions of each module described above are explained in the relevant content of Embodiment 1 or 2, and will not be repeated here.

[0152] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0153] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0154] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0155] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0156] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for predicting power grid stability margin based on multi-source model stacking, characterized in that, include: Based on the preset data base model, the multi-source meteorological data of the new energy grid-connected nodes are weighted and fused in two levels to generate the day-ahead power prediction results of the new energy grid-connected nodes; Based on the day-ahead power forecast results, a stability margin state feature vector is constructed. Based on the electrical distance, short-circuit support strength, conventional unit support coverage, and power fluctuation degree of the new energy grid-connected nodes, the node sensitivity index is calculated. Based on the node sensitivity index, the stability margin state feature vector is weighted to obtain a weighted feature vector. Based on the day-ahead power prediction results, short-circuit support information, reactive power support information, and reactive power reserve information, construct the support response characteristics; The weighted feature vectors are input into multiple first-level basis models to obtain multiple stability margin sub-prediction results; the support response features, node sensitivity indices and the stability margin sub-prediction results are input into a second-level meta-learning model to obtain the initial prediction result of the power grid stability margin. Based on the pre-constructed power reserve constraints, stability margin boundary constraints, and adjacent time period smoothing constraints, the initial prediction results of the power grid stability margin are corrected for consistency, and the final prediction results of the power grid stability margin are obtained.

2. The power grid stability margin prediction method based on multi-source model stacking according to claim 1, characterized in that, The formula for calculating the day-ahead power prediction result of the new energy grid-connected node is as follows: ； ； The intra-source weights and inter-source weights are obtained by minimizing the objective function, which is expressed as: ； Based on the support response characteristics and the final prediction results of the power grid stability margin, the training sample weights are updated using the following formula: ； in, express New energy grid connection nodes The The combined forecast results within each meteorological data source; Indicates the grid connection node of new energy sources The The first meteorological data source Source weights of each data base model; express New energy grid connection nodes The The first meteorological data source The prediction results of the data-based model; Indicates the number of data base models; express New energy grid connection nodes The day-ahead power prediction results; Indicates the grid connection node of new energy sources The Inter-source weights of meteorological data sources; Indicates the number of meteorological data sources; Represent the objective function; Represents the set of training sample time periods; express Training sample weights at any time; express New energy grid connection nodes Actual output; All represent regularization coefficients; This represents the updated training sample weights; Both indicate updating the weight coefficients; This indicates updating the training sample time period set; They represent time, Time-based historical stability margin label samples; They represent time, The final prediction result of the power grid stability margin at time t; They represent time, Weighted support response characteristics of all sensitive nodes across the network at all times; Represents positive numbers; This indicates taking the maximum value.

3. The power grid stability margin prediction method based on multi-source model stacking according to claim 1, characterized in that, The formula for calculating the stability margin state feature vector is as follows: ； ； ； in, Representing new energy grid connection node 1 and new energy grid connection node respectively. New energy grid connection nodes The change in power between adjacent time periods; They represent New Energy Grid Connection Node 1, New Energy Grid Connection Node The day-ahead power prediction results; They represent time, New energy grid connection nodes The day-ahead power prediction results; They represent New Energy Grid Connection Node 1, New Energy Grid Connection Node New energy grid connection nodes Historical prediction residual statistics; Indicates the length of the residual statistical time window; express New energy grid connection nodes Actual output; express New energy grid connection nodes The day-ahead power prediction results; express Constantly rotating standby feature; express Constant reactive power support characteristics; express Key cross-sectional tidal characteristics at all times; express Time-based network structure characteristics; This represents the statistical characteristics of historical stability margin; express Time-series stability margin state feature vector; This indicates the matrix transpose.

4. The power grid stability margin prediction method based on multi-source model stacking according to claim 1, characterized in that, The formula for calculating the node sensitivity index is as follows: ； ； ； ； ； in, Indicates the grid connection node of new energy sources To the strong support busbar assembly The equivalent electrical distance; Indicates strong support busbar Weighting coefficients; Indicates the grid connection node of new energy sources With strong support busbar Equivalent impedance elements between; Indicates the grid connection node of new energy sources The short-circuit support coefficient; Indicates the grid connection node of new energy sources The short-circuit capacity; Indicates the grid connection node of new energy sources Rated grid-connected capacitor; Represents positive numbers; Indicates the grid connection node of new energy sources The coverage of conventional unit support; Indicates the number of generating units; Indicates the unit Available spare capacity; Indicates the grid connection node of new energy sources to the unit Electrical distance; express New energy grid connection nodes The power fluctuation coefficient; express New energy grid connection nodes The day-ahead power prediction results; express New energy grid connection nodes In length Average predicted output within the time window; express New energy grid connection nodes Node sensitivity index; The weighting coefficients representing the node sensitivity index; express The normalized value; express The normalized value; express The normalized value; express The normalized value.

5. The power grid stability margin prediction method based on multi-source model stacking according to claim 1, characterized in that, The formula for calculating the weighted feature vector is: ； ； in, express Time and A feature node sensitivity index vector of the same dimension; express Time-of-time feature correlation matrix; express Time-node sensitivity index vector; express Time-weighted feature vector; express Time-series stability margin state feature vector; Indicates the sensitivity amplification factor; This indicates element-wise multiplication.

6. The power grid stability margin prediction method based on multi-source model stacking according to claim 1, characterized in that, The formula for calculating the support response characteristics is as follows: ； ； in, express New energy grid connection nodes Support response characteristics; express New energy grid connection nodes The day-ahead power prediction results; Indicates the grid connection node of new energy sources The short-circuit capacity; All represent adjustment coefficients; Represents positive numbers; express New energy grid connection nodes reactive power support margin in the region; express The system continuously rotates the total reserve capacity. express Weighted support response characteristics of all sensitive nodes across the network at all times; express New energy grid connection nodes Node sensitivity index; This indicates the number of new energy grid-connected nodes.

7. The power grid stability margin prediction method based on multi-source model stacking according to claim 1, characterized in that, The stability margin sub-prediction result is expressed as follows: ； The initial prediction result of the power grid stability margin is expressed as follows: ； Multiple first-level base models and second-level meta-learning models are jointly trained using a loss function, which is expressed as follows: ； in, They represent The first first-order basis model, the second first-order basis model, and the first... The first-level basic model, the first The stability margin sub-prediction results output by each first-level basis model; Indicates the first One first-level basis model; express Time-weighted feature vector; express Initial prediction results of power grid stability margin at time point; express Weighted support response characteristics of all sensitive nodes across the network at all times; express Time-node sensitivity index vector; This represents a two-dimensional meta-learning model; Represents the loss function; Represents the set of training sample time periods; express Time-based historical stability margin label samples; All of these represent the weighting coefficients of the loss function; Represents the set of model parameters; This represents the L2 norm.

8. The power grid stability margin prediction method based on multi-source model stacking according to claim 1, characterized in that, The data base models include the category feature enhancement model, the extreme gradient enhancement model, the random forest model, and the generalized additive model; The first-level base model includes at least two of the following: the class feature boosting model, the extreme gradient boosting model, the random forest model, and the generalized additive model. The second-order meta-learning model includes any one of the following: linear regression model, ridge regression model, Lasso model, support vector regression model, and lightweight neural network model.

9. The power grid stability margin prediction method based on multi-source model stacking according to claim 1, characterized in that, The formula for calculating the final prediction result of the power grid stability margin is as follows: ； ； ； ； in, They represent The lower boundary of the stability margin and the upper boundary of the stability margin at any given time; Both represent the coefficients of the lower boundary of the stability margin; express The system continuously rotates the total reserve capacity. express Real-time system reactive power support margin; express Standard unit support capacity indicators at all times; express Total power injected by new energy sources at all times; Both represent the coefficients of the upper boundary of the stability margin; express Time-of-flight system short-circuit support indicators; express The target stability margin value corresponding to the power supply backup constraint at any given time; All of these represent coefficients for power reserve constraints; They represent time, The final prediction result of the power grid stability margin at time t; This represents the stability margin of the independent variable that minimizes the objective function. The value of ; express Initial prediction results of power grid stability margin at time point; All of these represent coefficients representing the final predicted results of the power grid stability margin; This indicates taking the maximum value.

10. The power grid stability margin prediction method based on multi-source model stacking according to claim 1, characterized in that, Also includes; Based on the node sensitivity masking analysis, the contribution ranking results of each new energy grid-connected node to the stability margin are generated; Based on the final prediction results of the grid stability margin and the ranking results of the contribution of each new energy grid-connected node to the stability margin, a stability early warning level and a weak node identification result are generated.

11. The power grid stability margin prediction method based on multi-source model stacking according to claim 10, characterized in that, The formula for calculating the contribution of each new energy grid-connected node to the stability margin is as follows: ； The formula for calculating the stability warning level is as follows: ； The contribution of each new energy grid-connected node to the stability margin is ranked from largest to smallest, and the top A new energy grid-connected nodes in the ranking result are identified as weak nodes. in, express New energy grid connection nodes Contribution to stability margin; express New energy grid connection nodes Node sensitivity index; express Initial prediction results of power grid stability margin at time point; They represent The first first-order basis model, the second first-order basis model, and the first... The stability margin sub-prediction results output by each first-level basis model; express Weighted support response characteristics of all sensitive nodes across the network at all times; express Always keep the new energy grid connection nodes The sensitivity vector after the corresponding sensitivity components are set to zero; This represents a two-dimensional meta-learning model; express The level of stability warning at any given moment; express The final prediction result of the power grid stability margin at time t; All represent preset thresholds; A represents the preset number of recognitions.

12. A power grid stability margin prediction device based on multi-source model stacking, characterized in that, For performing the method according to any one of claims 1 to 11, comprising: The data fusion module is used to: perform two-level weighted fusion of multi-source meteorological data of new energy grid-connected nodes according to a preset data base model, and generate day-ahead power prediction results for new energy grid-connected nodes; The sensitivity calculation module is used to: construct a stability margin state feature vector based on the day-ahead power prediction results, and calculate the node sensitivity index based on the electrical distance, short-circuit support strength, conventional unit support coverage and power fluctuation degree of the new energy grid-connected nodes; The feature weighting module is used to: weight the stability margin state feature vector according to the node sensitivity index to obtain a weighted feature vector; The feature construction module is used to: construct support response features based on the day-ahead power prediction results, short-circuit support information, reactive power support information, and reactive power reserve information; The model prediction module is used to: input the weighted feature vector into multiple first-level base models to obtain multiple stability margin sub-prediction results; and input the support response features, node sensitivity index and the stability margin sub-prediction results into a second-level meta-learning model to obtain the initial prediction result of the power grid stability margin. The prediction correction module is used to: perform consistency correction on the initial prediction result of the power grid stability margin based on pre-constructed power reserve constraints, stability margin boundary constraints and adjacent time period smoothing constraints, so as to obtain the final prediction result of the power grid stability margin.

Images