An artificial intelligence-based power grid fault rapid prediction alarm method

Abstract

This invention discloses an artificial intelligence-based method for rapid prediction and alarm of power grid faults, comprising the following steps: collecting multi-source monitoring data and preprocessing it to obtain a standardized state time series; constructing a power grid operating state vector and node correlation matrix to obtain a power grid operating state dataset; performing phase space reconstruction to obtain a power grid operating state trajectory sequence; constructing a candidate nonlinear dynamic function library; inputting an improved sparse nonlinear dynamics identification algorithm, dynamically adjusting the sparse screening intensity, and performing differentiated screening to obtain the nonlinear dynamic equations of the power grid operating state; constructing a stable operating state manifold and extrapolating it to obtain a future state trajectory sequence; generating a fault risk index based on the future state trajectory sequence; determining the risk level of the power grid operating state and outputting fault prediction results and alarm information. This invention utilizes an improved sparse dynamics modeling method to achieve early prediction of power grid faults, possessing the advantages of high accuracy and strong early warning capability.

Description

Technical Field

[0001] This invention relates to the field of power grid operation monitoring technology, and in particular to a method for rapid prediction and alarm of power grid faults based on artificial intelligence. Background Technology

[0002] As the power grid continues to expand and its operational structure becomes increasingly complex, its operating status exhibits strong nonlinearity, strong coupling, and dynamic time-varying characteristics. Traditional fault early warning methods based on threshold judgment or empirical rules are no longer sufficient to meet practical needs. In recent years, some technologies have employed machine learning or deep learning methods to model power grid operating data, training predictive models using historical data to achieve fault identification and early warning, thereby improving automation levels and predictive capabilities to some extent.

[0003] However, most existing technologies treat power grid fault prediction as a data fitting or classification problem, lacking a characterization of the inherent dynamic evolution of the power grid's operating state and failing to effectively reflect the subtle changes in the system's evolution from a stable state to a fault state. Furthermore, existing methods typically ignore the influence of the power grid topology on state evolution and lack a comprehensive mechanism for judging the degree of deviation from the operating trajectory in a multi-dimensional state space, resulting in insufficient early identification capability for potential faults and requiring improvement in the stability and reliability of prediction results.

[0004] Therefore, how to provide a method for rapid prediction and alarm of power grid faults based on artificial intelligence is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method for rapid prediction and alarm of power grid faults based on artificial intelligence. This invention utilizes an improved sparse dynamics modeling method to achieve early prediction of power grid faults, and has the advantages of high accuracy and strong early warning capability.

[0006] A method for rapid prediction and alarm of power grid faults based on artificial intelligence according to an embodiment of the present invention includes the following steps: Multi-source monitoring data are collected during the continuous operation of the power grid and preprocessed to obtain a standardized state time series; A power grid operation state vector is constructed based on standardized state time series, and a node correlation matrix is ​​constructed to obtain a power grid operation state dataset; Phase space reconstruction is performed on the power grid operation state dataset to map multiple values ​​of the power grid operation state vector to a high-dimensional state space, thereby obtaining the power grid operation state trajectory sequence; A candidate nonlinear dynamics function library is constructed based on the power grid operation state trajectory sequence and node correlation matrix; The power grid operation state trajectory sequence is input into the improved sparse nonlinear dynamics identification algorithm, the sparse screening intensity is dynamically adjusted, and the topological coupling terms and non-topological coupling terms are differentiated by combining the node correlation matrix to obtain the nonlinear dynamic equation of the power grid operation state. Based on the nonlinear dynamic equation of power grid operation state, the historical stable operation state is mapped to construct the stable operation state manifold, and the future operation state is deduced to obtain the future state trajectory sequence; Fault risk indicators are generated based on the distance, direction deviation, cumulative deviation, and duration of continuous deviation between the future state trajectory sequence and the stable operating state manifold. The system determines the risk level of the power grid operation status based on fault risk indicators and outputs corresponding fault prediction results and alarm information.

[0007] Optionally, the multi-source monitoring data includes node voltage, branch current, system frequency, active power, reactive power, node phase angle, circuit breaker status, and protection operation status. Preprocessing includes time alignment, missing value completion, outlier removal, noise suppression, dimension unification, and normalization.

[0008] Optionally, the generation of the power grid operation status dataset includes: The standardized state time series is read in the order of sampling time, and the node voltage, system frequency, active power, reactive power and node phase angle corresponding to each node are extracted at each sampling time, as well as the branch current, circuit breaker status and protection action status of the lines connected to each node, to form the node state data corresponding to each sampling time. Arrange the node status data corresponding to each node at the same sampling time according to the node number order to obtain the power grid operation state vector corresponding to each sampling time. The connection status between any two nodes is determined based on the connection relationships between power grid nodes and lines, and a node association matrix is ​​constructed according to the node numbering order. The power grid operation state vectors corresponding to each sampling time are arranged in the order of sampling time and combined with the node association matrix to obtain the power grid operation state dataset.

[0009] Optionally, the generation of the power grid operation state trajectory sequence includes: Read the power grid operation status vector in the power grid operation status dataset according to the sampling time order, and use the power grid operation status vector corresponding to the current sampling time as the current state vector, and use the power grid operation status vectors corresponding to multiple consecutive sampling times before the current sampling time as the historical state vector to construct a continuous time series segment corresponding to the current sampling time. The current state vector and each historical state vector in the continuous time series segment are arranged sequentially according to the sampling time order, and then sequentially spliced ​​to obtain a high-dimensional reconstructed state vector corresponding to the current sampling time. The process of constructing continuous time series segments and generating high-dimensional reconstructed state vectors is repeatedly executed for each sampling time to obtain multiple high-dimensional reconstructed state vectors that correspond one-to-one with each sampling time. The multiple high-dimensional reconstructed state vectors are arranged in the order of sampling time to map the values ​​of the power grid operation state vector at multiple consecutive sampling times to the high-dimensional state space. By sequentially connecting multiple high-dimensional reconstructed state vectors arranged in the order of sampling time in the high-dimensional state space, a power grid operation state trajectory sequence is obtained.

[0010] Optionally, the construction of the candidate nonlinear dynamics function library includes: Read the power grid operation status trajectory sequence in the order of sampling time, and extract the state variable values ​​in the power grid operation status trajectory sequence corresponding to each sampling time to form a set of state variables corresponding to each sampling time. Construct a set of linear terms of state variables, a set of quadratic terms of state variables, a cross-coupled term of state variables, and a time-delay term of state variables coupled with topology based on the set of state variables; The first-order terms, second-order terms, cross-coupled terms, time-delay terms, and topological coupling terms of the state variables are arranged in a unified order to form a candidate nonlinear dynamics function library.

[0011] Optionally, the generation of the nonlinear dynamic equations for the power grid operating state includes: The power grid operation state trajectory sequence is input into the improved sparse nonlinear dynamics identification algorithm according to the sampling time order. The first-order terms, second-order terms, cross-coupled terms, time-delay terms, topological coupling terms, and non-topological coupling terms of the state variables in the candidate nonlinear dynamics function library are matched with the power grid operation state trajectory sequence to obtain the candidate dynamics term set corresponding to each sampling time. The dynamic sparse constraint is improved to:

[0012] The improved objective function is:

[0013] The topological difference screening has been improved to:

[0014] The topological consistency constraint is:

[0015]

[0016] Calculate the state residual change, local trajectory curvature change and state fluctuation intensity between adjacent sampling times based on the power grid operation state trajectory sequence; The state residual change, local trajectory curvature change, and state fluctuation intensity are combined in the order of sampling time to obtain the sparse screening intensity corresponding to each sampling time, and the sparse screening intensity is applied to the candidate dynamics term set at the corresponding sampling time. Based on the node correlation matrix, the candidate dynamics items in the candidate dynamics item set are classified into categories. Candidate dynamics items formed by the combination of node state variables with node connection relationships are identified as topologically coupled items, and candidate dynamics items without node connection relationships are identified as non-topologically coupled items. Differential sparse screening is performed on topologically coupled and non-topologically coupled terms based on the sparse screening strength, retaining candidate dynamic terms that meet the screening conditions and removing candidate dynamic terms that do not meet the screening conditions. For topological coupling terms, the selection criteria are:

[0017] For non-topological coupling terms, the selection criteria are as follows:

[0018] Furthermore, for topological coupling terms, topological consistency constraints must also be satisfied:

[0019] The retained candidate dynamic terms are combined in a unified order to obtain the nonlinear dynamic equations of the power grid operation state.

[0020] Optionally, the generation of the future state trajectory sequence includes: Extract the power grid operation state trajectory sequence corresponding to the historical stable operation state, and input the power grid operation state trajectory sequence corresponding to the historical stable operation state into the nonlinear dynamic equation of the power grid operation state according to the sampling time order. Perform state evolution constraint mapping on the power grid operation state trajectory sequence corresponding to each sampling time to obtain the stable operation state mapping point corresponding to each sampling time. Arrange the stable operating state mapping points corresponding to each sampling time in the order of sampling time, and connect the stable operating state mapping points corresponding to adjacent sampling times in sequence to obtain the stable operating state mapping trajectory sequence. Construct a stable operating state manifold based on the trajectory sequence mapped from the stable operating state; The power grid operating state corresponding to the current sampling time is input into the nonlinear dynamic equation of the power grid operating state, and the power grid operating state of subsequent consecutive sampling times is deduced according to the sampling time sequence to obtain the future state point corresponding to each future sampling time. Arrange the future state points corresponding to each future sampling time in the order of sampling time, and connect the future state points corresponding to adjacent future sampling times in sequence to obtain the future state trajectory sequence.

[0021] Optionally, the generation of the fault risk indicator includes: Read each future state point in the future state trajectory sequence according to the sampling time order, and read the stable operating state mapping points in the stable operating state manifold. For each future state point, calculate the state space distance between the future state point and each stable operating state mapping point, and calculate the minimum manifold distance sequence. Differential calculation is performed on two adjacent future state points to obtain the trajectory evolution direction within the corresponding sampling interval. The stable operating state mapping point corresponding to the current future state point is determined in the stable operating state manifold. The manifold evolution direction between the stable operating state mapping point and its adjacent stable operating state mapping points is extracted, and the trajectory evolution direction deviation sequence is calculated. The minimum manifold distances corresponding to multiple consecutive future state points are accumulated according to the sampling time sequence to obtain the deviation cumulative quantity sequence; Based on the sampling time sequence, determine whether the minimum manifold distance corresponding to each future state point is continuously greater than zero, and accumulate the number of consecutive sampling points with a continuous minimum manifold distance greater than zero to obtain a continuous deviation time sequence; The fault risk index is obtained by combining the minimum manifold distance sequence, trajectory evolution direction deviation sequence, cumulative deviation sequence, and continuous deviation duration sequence according to the sampling time order.

[0022] Optionally, the generation of the fault prediction results and alarm information includes: Read the fault risk indicators in the order of sampling time, and extract the fault risk indicator values ​​corresponding to each sampling time. The fault risk index value corresponding to each sampling time is compared with multiple risk level intervals, and the risk level corresponding to each sampling time is determined based on the comparison results. Based on the risk level at each sampling time, generate corresponding fault prediction results, and output alarm information of the corresponding level based on the fault prediction results at each sampling time.

[0023] The beneficial effects of this invention are: This invention transforms the power grid fault prediction problem from the traditional data fitting and threshold determination method into a nonlinear dynamic evolution modeling problem of power grid operating state. Based on the power grid operating state trajectory sequence, a candidate nonlinear dynamic function library is constructed. Combined with an improved sparse nonlinear dynamic identification algorithm, it can automatically identify the nonlinear dynamic equations of power grid operating state from multi-source monitoring data, realizing an explicit characterization of the power grid operating mechanism. Compared with methods that rely solely on historical data statistical patterns or black box models for prediction, this invention can directly reflect the coupling relationship and evolution trend between power grid state variables, thereby improving the modeling ability of complex operating state changes and providing a more physically meaningful and interpretable foundation for subsequent fault prediction.

[0024] Furthermore, this invention constructs a stable operating state manifold by mapping the state evolution constraints of historical stable operating states, and continuously extrapolates future operating states based on the nonlinear dynamic equations of the power grid operating state, forming a future state trajectory sequence. Based on this, a multi-dimensional fusion fault risk index is constructed by calculating the minimum manifold distance, trajectory evolution direction deviation, cumulative deviation, and duration of continuous deviation of the future state trajectory sequence relative to the stable operating state manifold, thereby achieving a quantitative characterization of the degree to which the power grid operating state deviates from the stable region. This method can comprehensively assess system stability changes from multiple dimensions, including distance, direction, and evolution process. Compared to single indexes or simple threshold judgments, it can detect weak precursors of potential faults earlier, improving the lead time and sensitivity of fault prediction.

[0025] Furthermore, this invention introduces a node correlation matrix during the dynamic modeling process to perform differentiated sparse screening of topologically coupled and non-topologically coupled terms in the candidate nonlinear dynamic function library. This results in the final nonlinear dynamic equations of the power grid operation state not only possessing sparsity but also exhibiting physical constraint characteristics consistent with the actual topology of the power grid. This effectively reduces the impact of redundant terms on model accuracy and improves the model's stability and generalization ability. This invention enables high-precision dynamic modeling and real-time risk assessment of the power grid operation state, providing early warnings before faults occur and enhancing the safety, reliability, and intelligence level of power grid operation. Attached Figure Description

[0026] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a method for rapid prediction and alarm of power grid faults based on artificial intelligence proposed in this invention; Figure 2This is a flowchart illustrating the process of generating nonlinear dynamic equations of power grid operating state using an improved sparse nonlinear dynamics identification algorithm in an artificial intelligence-based rapid prediction and alarm method for power grid faults proposed in this invention. Figure 3 This is a schematic diagram illustrating the calculation process of the fault risk index based on manifold deviation in the AI-based rapid prediction and alarm method for power grid faults proposed in this invention. Detailed Implementation

[0027] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0028] refer to Figures 1-3 A method for rapid prediction and alarm of power grid faults based on artificial intelligence includes the following steps: Multi-source monitoring data is collected during the continuous operation of the power grid. The multi-source monitoring data includes node voltage, branch current, system frequency, active power, reactive power, node phase angle, circuit breaker status and protection operation status. Time alignment is performed, missing value completion, outlier removal, noise suppression, dimension unification and normalization are performed to obtain a standardized state time series. A power grid operation state vector is constructed based on standardized state time series, and a node correlation matrix is ​​constructed based on the connection relationship between power grid nodes and the connection relationship between lines to obtain a power grid operation state dataset. Phase space reconstruction is performed on the power grid operation state dataset to map the values ​​of the power grid operation state vector at multiple consecutive sampling times to a high-dimensional state space, thereby obtaining the power grid operation state trajectory sequence; A candidate nonlinear dynamics function library is constructed based on the power grid operation state trajectory sequence and node correlation matrix. The candidate nonlinear dynamics function library includes state variable linear terms, state variable quadratic terms, state variable cross-coupling terms, state variable time delay terms, and topology coupling terms. The power grid operation state trajectory sequence is input into the improved sparse nonlinear dynamics identification algorithm. The sparse screening intensity is dynamically adjusted based on the state residual change, local trajectory curvature change and state fluctuation intensity of the power grid operation state trajectory sequence between adjacent sampling times. The algorithm also performs differentiated sparse screening on topological coupling terms and non-topological coupling terms in the candidate nonlinear dynamics function library in combination with the node correlation matrix to obtain the nonlinear dynamic equation of the power grid operation state. Based on the nonlinear dynamic equation of power grid operation state, the power grid operation state trajectory sequence corresponding to the historical stable operation state is mapped by state evolution constraint to obtain the stable operation state mapping trajectory sequence. Based on the stable operation state mapping trajectory sequence, the stable operation state manifold is constructed. Then, based on the nonlinear dynamic equation of power grid operation state, the power grid operation state in future continuous time periods is deduced to obtain the future state trajectory sequence. Calculate the minimum manifold distance, trajectory evolution direction deviation, and cumulative deviation of the future state trajectory sequence relative to the stable operating manifold, and generate a fault risk index by combining the continuous deviation duration of the future state trajectory sequence; The system determines the risk level of the power grid operation status based on fault risk indicators and outputs corresponding fault prediction results and alarm information.

[0029] In this embodiment, the multi-source monitoring data includes node voltage, branch current, system frequency, active power, reactive power, node phase angle, circuit breaker status, and protection operation status. Preprocessing includes time alignment, missing value completion, outlier removal, noise suppression, dimension unification, and normalization.

[0030] In this embodiment, the generation of the power grid operation status dataset includes: The standardized state time series is read in the order of sampling time, and the node voltage, system frequency, active power, reactive power and node phase angle corresponding to each node are extracted at each sampling time, as well as the branch current, circuit breaker status and protection action status of the lines connected to each node, to form the node state data corresponding to each sampling time. Arrange the node status data corresponding to each node at the same sampling time according to the node number order, and then concatenate the node voltage, system frequency, active power, reactive power, node phase angle, branch current, circuit breaker status and protection action status corresponding to each node in sequence to obtain the power grid operation state vector corresponding to each sampling time. The connection status between any two nodes is determined based on the connection relationship between power grid nodes and the connection relationship between lines. A node association matrix is ​​constructed according to the node number order. When there is a line connection relationship between two nodes, a connection value is assigned to the corresponding position in the node association matrix. When there is no line connection relationship between two nodes, a zero connection value is assigned to the corresponding position in the node association matrix. The power grid operation state vectors corresponding to each sampling time are arranged in the order of sampling time and combined with the node association matrix to obtain the power grid operation state dataset.

[0031] In this embodiment, the generation of the power grid operation state trajectory sequence includes: Read the power grid operation status vector in the power grid operation status dataset according to the sampling time order, and use the power grid operation status vector corresponding to the current sampling time as the current state vector, and use the power grid operation status vectors corresponding to multiple consecutive sampling times before the current sampling time as the historical state vector to construct a continuous time series segment corresponding to the current sampling time. The current state vector and each historical state vector in the continuous time series segment are arranged in order of sampling time and then concatenated sequentially to obtain the high-dimensional reconstructed state vector corresponding to the current sampling time. The process of obtaining the high-dimensional reconstructed state vector corresponding to the current sampling time specifically includes: taking the current sampling time as the reference, selecting the power grid operation state vectors corresponding to multiple consecutive historical sampling times, arranging the power grid operation state vector corresponding to the current sampling time and the selected power grid operation state vectors corresponding to each historical sampling time in chronological order, and performing sequential concatenation on the arranged power grid operation state vectors to obtain the high-dimensional reconstructed state vector corresponding to the current sampling time. The process of constructing continuous time series segments and generating high-dimensional reconstructed state vectors is repeatedly executed for each sampling time to obtain multiple high-dimensional reconstructed state vectors that correspond one-to-one with each sampling time. The multiple high-dimensional reconstructed state vectors are arranged in the order of sampling time to map the values ​​of the power grid operation state vector at multiple consecutive sampling times to the high-dimensional state space. By sequentially connecting multiple high-dimensional reconstructed state vectors arranged in the order of sampling time in the high-dimensional state space, a power grid operation state trajectory sequence reflecting the continuous change of power grid operation state with sampling time is obtained.

[0032] In this embodiment, the construction of the candidate nonlinear dynamics function library includes: Read the power grid operation state trajectory sequence according to the sampling time order, and extract the state variable values ​​in the power grid operation state trajectory sequence corresponding to each sampling time. According to the sampling time order, extract the state variable values ​​at each dimension position in the high-dimensional reconstructed state vector corresponding to each sampling time. According to the dimensional arrangement order of the high-dimensional reconstructed state vector, map the state variable values ​​in the corresponding dimension to form a set of state variables corresponding to each sampling time. Constructing a set of linear terms of state variables based on a set of state variables involves taking the values ​​of each state variable at each sampling time as independent candidate terms and arranging them in the order of sampling time to obtain a set of linear terms of state variables. Based on the set of state variables, we construct quadratic terms and cross-coupled terms of state variables. Specifically, we perform combination processing on each state variable with itself to obtain quadratic terms of state variables, and perform pairwise combination processing on different state variables to obtain cross-coupled terms of state variables. The state variable time delay term is constructed based on the set of state variables. Specifically, it includes extracting the state variable values ​​corresponding to multiple consecutive sampling times before the current sampling time, and arranging the state variable values ​​corresponding to multiple consecutive sampling times in the order of sampling time to obtain the state variable time delay term. The construction of topological coupling terms based on the node association matrix specifically includes combining the values ​​of the state variables corresponding to the nodes with connection relationships in the node association matrix to obtain the topological coupling terms. The first-order terms, second-order terms, cross-coupled terms, time-delay terms, and topological coupling terms of the state variables are arranged in a unified order to form a candidate nonlinear dynamics function library.

[0033] In this embodiment, the generation of the nonlinear dynamic equations for the power grid operating state includes: The power grid operation state trajectory sequence is input into the improved sparse nonlinear dynamics identification algorithm according to the sampling time order. The first-order terms, second-order terms, cross-coupled terms, time-delay terms, topological coupling terms, and non-topological coupling terms of the state variables in the candidate nonlinear dynamics function library are matched with the power grid operation state trajectory sequence to obtain the candidate dynamics term set corresponding to each sampling time. The improved sparse nonlinear dynamics identification algorithm is improved by introducing dynamic sparse constraints and topology difference screening mechanism. The sparse screening intensity is adaptively adjusted according to the state residual change, local trajectory curvature change and state fluctuation intensity of the power grid operation state trajectory. The algorithm also combines the node correlation matrix to perform differential screening of topology coupling terms and non-topology coupling terms, and finally obtains the nonlinear dynamic equation of power grid operation state with sparse structure and topology consistency. The dynamic sparse constraint is improved to:

[0034] , and All are positive numbers, and , and The sum is 1; The state residual changes are as follows:

[0035] The local trajectory curvature change is as follows:

[0036] The intensity of the state fluctuation is:

[0037] in, ; The improved objective function is:

[0038]

[0039] The coefficients corresponding to the candidate kinetic terms are: ; The candidate nonlinear dynamics function library is as follows:

[0040] The topological difference screening has been improved to:

[0041] in , here Represents the set of topological coupling terms. This represents the set of non-topologically coupled items, meaning that topologically coupled items are selected with lower weights and non-topologically coupled items are selected with higher weights. The topological consistency constraint is:

[0042]

[0043] The state residual change, local trajectory curvature change and state fluctuation intensity between adjacent sampling times are calculated based on the power grid operation state trajectory sequence. The state residual change is determined by the fitting difference of the power grid operation state trajectory sequence corresponding to adjacent sampling times. The local trajectory curvature change is determined by the degree of inflection of the power grid operation state trajectory sequence corresponding to three consecutive sampling times. The state fluctuation intensity is determined by the change amplitude of the power grid operation state trajectory sequence corresponding to multiple consecutive sampling times. The state residual change, local trajectory curvature change, and state fluctuation intensity are combined in the order of sampling time to obtain the sparse screening intensity corresponding to each sampling time, and the sparse screening intensity is applied to the candidate dynamics term set at the corresponding sampling time. Based on the node correlation matrix, the candidate dynamics items in the candidate dynamics item set are classified into categories. Candidate dynamics items formed by the combination of node state variables with node connection relationships are identified as topologically coupled items, and candidate dynamics items without node connection relationships are identified as non-topologically coupled items. Differential sparse screening is performed on topologically coupled and non-topologically coupled terms based on the sparse screening strength, retaining candidate dynamic terms that meet the screening conditions and removing candidate dynamic terms that do not meet the screening conditions. Differentiated sparse screening is specifically manifested as follows: for topologically coupled terms, connection constraints based on the node association matrix are introduced during the sparse screening process, so that the corresponding coefficients satisfy the consistency condition of node connection relationship, and a lower screening strength threshold is used for retention; for non-topologically coupled terms, no topological constraints are introduced, and a higher screening strength threshold is used for screening, thereby realizing differentiated processing of different types of candidate dynamic terms in terms of screening strength and constraint method. The specific screening criteria include: based on the sparse screening intensity corresponding to each sampling time, sparse constraints are applied to the coefficients of candidate dynamics terms. When the absolute value of the coefficient of a candidate dynamics term is less than the sparse screening intensity at the corresponding sampling time, the candidate dynamics term is removed. For topologically coupled terms, based on the connection relationship between corresponding nodes in the node association matrix, topological consistency constraints are applied to their corresponding coefficients. When the coefficient of a topologically coupled term does not satisfy the trend of change consistent with the node connection relationship, the topologically coupled term is removed. For non-topologically coupled terms, when their corresponding coefficient is less than the corresponding sparse screening intensity in multiple consecutive sampling times, the non-topologically coupled term is removed. Candidate dynamics terms that simultaneously satisfy both sparse constraints and topological constraints are retained as the screening results. The absolute values ​​of the coefficients corresponding to the candidate kinetic terms are evaluated, and the sparse screening intensity at the corresponding sampling time is assumed to be... The coefficients corresponding to the candidate kinetic terms are

[0044] For topological coupling terms, the selection criteria are:

[0045] Where α is an adjustment coefficient less than 1, which makes the screening threshold corresponding to the topological coupling term lower than the basic screening intensity, thereby preferentially retaining the dynamic terms related to the node connection relationship; For non-topological coupling terms, the selection criteria are as follows:

[0046] Wherein, β is an adjustment coefficient greater than 1, which makes the screening threshold corresponding to non-topological coupling terms higher than the basic screening intensity, thereby enhancing the elimination of non-critical dynamic terms. Furthermore, for topological coupling terms, topological consistency constraints must also be satisfied:

[0047] in, This indicates the connection relationship between corresponding nodes in the node association matrix, used to ensure that the retained topology coupling terms are consistent with the actual power grid topology. By using the above screening criteria, topologically coupled terms are preferentially retained under lower thresholds and topological constraints, while non-topologically coupled terms are strictly screened under higher thresholds, thus achieving differentiated sparsity screening in a mathematical sense. The retained candidate dynamic terms are combined in a unified order to obtain the nonlinear dynamic equations of the power grid operation state.

[0048] In this embodiment, the generation of the future state trajectory sequence includes: Extract the power grid operation state trajectory sequence corresponding to the historical stable operation state, and input the power grid operation state trajectory sequence corresponding to the historical stable operation state into the nonlinear dynamic equation of the power grid operation state according to the sampling time order. Perform state evolution constraint mapping on the power grid operation state trajectory sequence corresponding to each sampling time to obtain the stable operation state mapping point corresponding to each sampling time. Arrange the stable operating state mapping points corresponding to each sampling time in the order of sampling time, and connect the stable operating state mapping points corresponding to adjacent sampling times in sequence to obtain the stable operating state mapping trajectory sequence. Constructing a stable operating state manifold based on a stable operating state mapping trajectory sequence specifically involves embedding and arranging the stable operating state mapping points in the stable operating state mapping trajectory sequence into a manifold according to the sampling time order, so that the stable operating state mapping trajectory sequence forms a continuous manifold structure in a high-dimensional state space, thus obtaining a stable operating state manifold. The construction of the stable operating state manifold specifically includes: using each stable operating state mapping point in the stable operating state mapping trajectory sequence as a manifold embedding point to construct a set of mapping points in a high-dimensional state space; calculating the distance relationship between each stable operating state mapping point based on the mapping point set, and selecting several mapping points with the smallest distance as neighborhood points to construct an adjacency relationship; performing local linear reconstruction on each stable operating state mapping point and its neighborhood points so that each stable operating state mapping point can be linearly represented by its neighborhood points; and unifying and integrating the local linear reconstruction relationships of all stable operating state mapping points to obtain a stable operating state manifold that reflects the evolution structure of the stable operating state. The power grid operating state corresponding to the current sampling time is input into the nonlinear dynamic equation of the power grid operating state, and the power grid operating state of subsequent consecutive sampling times is deduced according to the sampling time sequence to obtain the future state point corresponding to each future sampling time. Arrange the future state points corresponding to each future sampling time in the order of sampling time, and connect the future state points corresponding to adjacent future sampling times in sequence to obtain the future state trajectory sequence.

[0049] In this embodiment, the generation of fault risk indicators includes: Read each future state point in the future state trajectory sequence according to the sampling time order, and read the stable operating state mapping points in the stable operating state manifold. For each future state point, calculate the state space distance between the future state point and each stable operating state mapping point, and determine the minimum value among the calculated state space distances as the minimum manifold distance corresponding to the future state point, thus obtaining the minimum manifold distance sequence. The calculation of state space distance specifically includes: for each future state point, extracting the state variable values ​​of the future state point at each dimension position in turn, and extracting the state variable values ​​of each stable operating state mapping point at the corresponding dimension position, calculating the difference of the state variable values ​​at the corresponding dimension position, squaring and summing the differences at each dimension position respectively, and then taking the square root to obtain the state space distance between the future state point and each stable operating state mapping point. Differential calculation is performed on two adjacent future state points to obtain the trajectory evolution direction within the corresponding sampling interval. The stable operating state mapping point corresponding to the current future state point is determined in the stable operating state manifold. The manifold evolution direction between the stable operating state mapping point and its adjacent stable operating state mapping points is extracted. The angle between the trajectory evolution direction and the manifold evolution direction is calculated to obtain the trajectory evolution direction deviation sequence. The minimum manifold distances corresponding to multiple consecutive future state points are accumulated according to the sampling time sequence to obtain the deviation cumulative quantity sequence; Based on the sampling time sequence, determine whether the minimum manifold distance corresponding to each future state point is continuously greater than zero, and accumulate the number of consecutive sampling points with a continuous minimum manifold distance greater than zero to obtain a continuous deviation time sequence; The fault risk index is obtained by combining the minimum manifold distance sequence, trajectory evolution direction deviation sequence, cumulative deviation sequence, and continuous deviation duration sequence according to the sampling time order.

[0050] In this embodiment, the generation of fault prediction results and alarm information includes: Read the fault risk indicators in the order of sampling time, and extract the fault risk indicator values ​​corresponding to each sampling time. The fault risk index value corresponding to each sampling time is compared with multiple risk level intervals, and the risk level corresponding to each sampling time is determined based on the comparison results. The comparison results are as follows: the fault risk index value corresponding to each sampling time is compared one by one with the upper and lower boundary values ​​of multiple risk level intervals. When the fault risk index value is between the lower and upper boundary values ​​of a certain risk level interval, the risk level corresponding to that sampling time is determined to be the risk level corresponding to that risk level interval; when the fault risk index value is greater than the upper boundary value of the highest risk level interval, the risk level corresponding to that sampling time is determined to be the highest risk level; when the fault risk index value is less than the lower boundary value of the lowest risk level interval, the risk level corresponding to that sampling time is determined to be the lowest risk level. Based on the risk level at each sampling time, generate corresponding fault prediction results, and output alarm information of the corresponding level based on the fault prediction results at each sampling time.

[0051] Example 1: To verify the feasibility of this invention in practice, it was applied to a distribution network system in actual operation in a certain region. This regional power grid includes multiple feeders and load nodes, characterized by frequent load fluctuations, complex inter-node coupling relationships, and rapid changes in operating status. Long-term operation revealed that existing fault early warning methods based on threshold monitoring or traditional machine learning are insufficient to promptly identify subtle changes in the power grid's evolution from normal to abnormal states. Alarms are often triggered only after a fault has clearly occurred, making it difficult for maintenance personnel to take timely control measures, thus affecting the safety and stability of the power grid. This invention addresses these problems by constructing a nonlinear dynamic model of the power grid's operating state and combining it with manifold deviation analysis to achieve early identification of potential faults.

[0052] In practical applications, monitoring equipment deployed at each node continuously collects multi-source monitoring data during power grid operation, including node voltage, branch current, system frequency, active power, reactive power, node phase angle, and switchgear status information. The collected data is aligned according to a unified time scale and undergoes missing value completion, outlier removal, noise suppression, and dimensional standardization to form a stable and reliable standardized state time series. Based on this, the operational data of each node and its related lines are fused to construct a power grid operation state vector that reflects the overall operational characteristics of the power grid. Simultaneously, a node correlation matrix is ​​generated based on the power grid topology to describe the connection relationships and mutual influences between nodes.

[0053] Subsequently, phase space reconstruction is performed on the power grid operating state vector, combining the state information from multiple consecutive sampling moments and mapping it to a high-dimensional state space to form a power grid operating state trajectory sequence. This method transforms the previously difficult-to-observe dynamic evolution process into a trajectory form, making the changing trends of the power grid operating state clearer. After obtaining the trajectory sequence, a candidate nonlinear dynamic function library is constructed based on the combination relationships between state variables. This library includes not only the basic terms of each state variable but also coupling terms between variables and combination terms related to the power grid topology, thus providing sufficient expressive power for subsequent dynamic modeling.

[0054] In the dynamic modeling process, the power grid operating state trajectory sequence is input into an improved sparse nonlinear dynamic identification algorithm. By analyzing the state changes between adjacent sampling times and the local variation characteristics of the trajectory, the sparsity screening intensity is adaptively adjusted. Simultaneously, combined with the node correlation matrix, different types of candidate dynamic terms are distinguished and processed, ensuring that the coupling relationships related to the actual power grid structure are effectively preserved. Through multiple rounds of screening and optimization, a nonlinear dynamic equation describing the evolution of the power grid operating state is finally obtained. This equation not only possesses strong expressive power but also good structural simplicity, which is beneficial for subsequent analysis and application.

[0055] After obtaining the nonlinear dynamic equation of the power grid operation state, the state trajectory corresponding to the historical stable operation stage is input into the equation for constraint mapping to construct the stable operation state manifold. This manifold reflects the state distribution and evolution path of the power grid under normal operating conditions. During real-time operation, the current state is input into the dynamic equation to extrapolate the power grid operation state in the future period and form a future state trajectory sequence. By comparing the future state trajectory with the stable operation state manifold, it can be determined whether the power grid operation state has a trend of deviating from the stable region.

[0056] During the risk assessment phase, the power grid's operating status is comprehensively evaluated by calculating the distance changes, evolution direction differences, and cumulative deviations of the future state trajectory relative to the stable operating manifold. When the trajectory gradually deviates from the stable manifold and the deviation trend continues to increase, the system can identify potential abnormal risks. Furthermore, based on the obtained risk indicators, the power grid's operating status is classified and determined, and corresponding fault prediction results and alarm information are output, enabling maintenance personnel to receive early warnings before faults occur.

[0057] As can be seen from the above implementation process, this invention can capture subtle changes in the system evolution process before obvious anomalies appear in the power grid's operating state, thus improving the advance detection capability of faults compared to traditional methods. Furthermore, by incorporating power grid topology information and dynamic modeling mechanisms, the prediction results are more consistent with actual operating patterns, improving the stability and reliability of the predictions. In addition, continuous analysis of the operating state trajectory can effectively reduce the interference of occasional fluctuations on the judgment results, improving the consistency of early warning results, thereby providing a more reliable decision-making basis for power grid operation and scheduling.

[0058] Table 1. Performance Comparison Experiment Results of Power Grid Fault Prediction Methods

[0059] As can be seen from Table 1, the method of the present invention outperforms traditional threshold methods, traditional machine learning methods, and conventional deep learning methods in key indicators such as prediction accuracy, early warning capability, false alarm control, false alarm control, and processing response efficiency. This indicates that the present invention can not only identify potential faults more accurately, but also provide early warning results earlier and more stably.

[0060] Regarding early warning time, traditional threshold methods can only issue warnings 2.1 minutes in advance, traditional machine learning methods 4.5 minutes, and deep learning methods 6.8 minutes, while the method of this invention achieves 12.6 minutes. This result demonstrates that this invention can identify abnormal signs before a fault fully manifests by observing the deviation trend between the future state trajectory sequence and the stable operating state manifold, thereby improving the early warning time. This effect is particularly important for power grid operation, as a longer early warning time means that maintenance personnel or the dispatching system have more time to perform load transfer, protection setting adjustments, or fault isolation.

[0061] Regarding the false alarm rate, the traditional threshold method has a rate of 9.8%, the traditional machine learning method has a rate of 7.2%, the deep learning method has a rate of 5.9%, and the method of this invention reduces it to 3.1%. This demonstrates that while improving sensitivity, this invention does not introduce more false alarms; on the contrary, it effectively suppresses the occurrence of false alarms. This invention does not directly trigger an alarm based on instantaneous fluctuations at a particular moment, but rather makes a comprehensive judgment based on continuous deviation trends and evolutionary directions. Therefore, it can effectively reduce false alarms caused by short-term noise, transient impacts, or occasional fluctuations.

[0062] Regarding response time, the traditional threshold method takes 5.6 seconds, the traditional machine learning method takes 4.2 seconds, the deep learning method takes 3.8 seconds, and the method of this invention takes 2.5 seconds. From the results, this invention not only improves the prediction accuracy and early warning lead time, but also maintains faster processing efficiency. This invention performs differentiated sparse screening of candidate dynamic terms by improving the sparse nonlinear dynamic identification algorithm, eliminating redundant terms and retaining key dynamic terms, so that the final nonlinear dynamic equation of the power grid operation state has good sparsity and structural compactness. Therefore, it has higher computational efficiency in the online simulation and risk assessment stages.

[0063] Based on the data in Table 1, it can be concluded that the present invention can more effectively solve the problem of difficulty in timely identifying subtle changes in the power grid's operating state during the transition from stable operation to fault state in the existing technology, and has practical application value.

[0064] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for rapid prediction and alarm of power grid faults based on artificial intelligence, characterized in that, Includes the following steps: Multi-source monitoring data are collected during the continuous operation of the power grid and preprocessed to obtain a standardized state time series; A power grid operation state vector is constructed based on standardized state time series, and a node correlation matrix is ​​constructed to obtain a power grid operation state dataset; Phase space reconstruction is performed on the power grid operation state dataset to map multiple values ​​of the power grid operation state vector to a high-dimensional state space, thereby obtaining the power grid operation state trajectory sequence; A candidate nonlinear dynamics function library is constructed based on the power grid operation state trajectory sequence and node correlation matrix; The power grid operation state trajectory sequence is input into the improved sparse nonlinear dynamics identification algorithm, the sparse screening intensity is dynamically adjusted, and the topological coupling terms and non-topological coupling terms are differentiated by combining the node correlation matrix to obtain the nonlinear dynamic equation of the power grid operation state. Based on the nonlinear dynamic equation of power grid operation state, the historical stable operation state is mapped to construct the stable operation state manifold, and the future operation state is deduced to obtain the future state trajectory sequence; Fault risk indicators are generated based on the distance, direction deviation, cumulative deviation, and duration of continuous deviation between the future state trajectory sequence and the stable operating state manifold. The system determines the risk level of the power grid operation status based on fault risk indicators and outputs corresponding fault prediction results and alarm information.

2. The method for rapid prediction and alarm of power grid faults based on artificial intelligence according to claim 1, characterized in that, The multi-source monitoring data includes node voltage, branch current, system frequency, active power, reactive power, node phase angle, circuit breaker status, and protection operation status. Preprocessing includes time alignment, missing value completion, outlier removal, noise suppression, dimension unification, and normalization.

3. The method for rapid prediction and alarm of power grid faults based on artificial intelligence according to claim 1, characterized in that, The generation of the power grid operation status dataset includes: The standardized state time series is read in the order of sampling time, and the node voltage, system frequency, active power, reactive power and node phase angle corresponding to each node are extracted at each sampling time, as well as the branch current, circuit breaker status and protection action status of the lines connected to each node, to form the node state data corresponding to each sampling time. Arrange the node status data corresponding to each node at the same sampling time according to the node number order to obtain the power grid operation state vector corresponding to each sampling time. The connection status between any two nodes is determined based on the connection relationships between power grid nodes and lines, and a node association matrix is ​​constructed according to the node numbering order. The power grid operation state vectors corresponding to each sampling time are arranged in the order of sampling time and combined with the node association matrix to obtain the power grid operation state dataset.

4. The method for rapid prediction and alarm of power grid faults based on artificial intelligence according to claim 1, characterized in that, The generation of the power grid operation status trajectory sequence includes: Read the power grid operation status vector in the power grid operation status dataset according to the sampling time order, and use the power grid operation status vector corresponding to the current sampling time as the current state vector, and use the power grid operation status vectors corresponding to multiple consecutive sampling times before the current sampling time as the historical state vector to construct a continuous time series segment corresponding to the current sampling time. The current state vector and each historical state vector in the continuous time series segment are arranged sequentially according to the sampling time order, and then sequentially spliced ​​to obtain a high-dimensional reconstructed state vector corresponding to the current sampling time. The process of constructing continuous time series segments and generating high-dimensional reconstructed state vectors is repeatedly executed for each sampling time to obtain multiple high-dimensional reconstructed state vectors that correspond one-to-one with each sampling time. The multiple high-dimensional reconstructed state vectors are arranged in the order of sampling time to map the values ​​of the power grid operation state vector at multiple consecutive sampling times to the high-dimensional state space. By sequentially connecting multiple high-dimensional reconstructed state vectors arranged in the order of sampling time in the high-dimensional state space, a power grid operation state trajectory sequence is obtained.

5. The method for rapid prediction and alarm of power grid faults based on artificial intelligence according to claim 1, characterized in that, The construction of the candidate nonlinear dynamics function library includes: Read the power grid operation status trajectory sequence in the order of sampling time, and extract the state variable values ​​in the power grid operation status trajectory sequence corresponding to each sampling time to form a set of state variables corresponding to each sampling time. Construct a set of linear terms of state variables, a set of quadratic terms of state variables, a cross-coupled term of state variables, and a time-delay term of state variables coupled with topology based on the set of state variables; The first-order terms, second-order terms, cross-coupled terms, time-delay terms, and topological coupling terms of the state variables are arranged in a unified order to form a candidate nonlinear dynamics function library.

6. The method for rapid prediction and alarm of power grid faults based on artificial intelligence according to claim 1, characterized in that, The generation of the nonlinear dynamic equations for the power grid operating state includes: The power grid operation state trajectory sequence is input into the improved sparse nonlinear dynamics identification algorithm according to the sampling time order. The first-order terms, second-order terms, cross-coupled terms, time-delay terms, topological coupling terms, and non-topological coupling terms of the state variables in the candidate nonlinear dynamics function library are matched with the power grid operation state trajectory sequence to obtain the candidate dynamics term set corresponding to each sampling time. The dynamic sparse constraint is improved to: ； The improved objective function is: ； The topological difference screening has been improved to: ； The topological consistency constraint is: ； ； Calculate the state residual change, local trajectory curvature change and state fluctuation intensity between adjacent sampling times based on the power grid operation state trajectory sequence; The state residual change, local trajectory curvature change, and state fluctuation intensity are combined in the order of sampling time to obtain the sparse screening intensity corresponding to each sampling time, and the sparse screening intensity is applied to the candidate dynamics term set at the corresponding sampling time. Based on the node correlation matrix, the candidate dynamics items in the candidate dynamics item set are classified into categories. Candidate dynamics items formed by the combination of node state variables with node connection relationships are identified as topologically coupled items, and candidate dynamics items without node connection relationships are identified as non-topologically coupled items. Differential sparse screening is performed on topologically coupled and non-topologically coupled terms based on the sparse screening strength, retaining candidate dynamic terms that meet the screening conditions and removing candidate dynamic terms that do not meet the screening conditions. For topological coupling terms, the selection criteria are: ； For non-topological coupling terms, the selection criteria are as follows: ； Furthermore, for topological coupling terms, topological consistency constraints must also be satisfied: ； The retained candidate dynamic terms are combined in a unified order to obtain the nonlinear dynamic equations of the power grid operation state.

7. The method for rapid prediction and alarm of power grid faults based on artificial intelligence according to claim 1, characterized in that, The generation of the future state trajectory sequence includes: Extract the power grid operation state trajectory sequence corresponding to the historical stable operation state, and input the power grid operation state trajectory sequence corresponding to the historical stable operation state into the nonlinear dynamic equation of the power grid operation state according to the sampling time order. Perform state evolution constraint mapping on the power grid operation state trajectory sequence corresponding to each sampling time to obtain the stable operation state mapping point corresponding to each sampling time. Arrange the stable operating state mapping points corresponding to each sampling time in the order of sampling time, and connect the stable operating state mapping points corresponding to adjacent sampling times in sequence to obtain the stable operating state mapping trajectory sequence. Construct a stable operating state manifold based on the trajectory sequence mapped from the stable operating state; The power grid operating state corresponding to the current sampling time is input into the nonlinear dynamic equation of the power grid operating state, and the power grid operating state of subsequent consecutive sampling times is deduced according to the sampling time sequence to obtain the future state point corresponding to each future sampling time. Arrange the future state points corresponding to each future sampling time in the order of sampling time, and connect the future state points corresponding to adjacent future sampling times in sequence to obtain the future state trajectory sequence.

8. The method for rapid prediction and alarm of power grid faults based on artificial intelligence according to claim 1, characterized in that, The generation of the fault risk indicators includes: Read each future state point in the future state trajectory sequence according to the sampling time order, and read the stable operating state mapping points in the stable operating state manifold. For each future state point, calculate the state space distance between the future state point and each stable operating state mapping point, and calculate the minimum manifold distance sequence. Differential calculation is performed on two adjacent future state points to obtain the trajectory evolution direction within the corresponding sampling interval. The stable operating state mapping point corresponding to the current future state point is determined in the stable operating state manifold. The manifold evolution direction between the stable operating state mapping point and its adjacent stable operating state mapping points is extracted, and the trajectory evolution direction deviation sequence is calculated. The minimum manifold distances corresponding to multiple consecutive future state points are accumulated according to the sampling time sequence to obtain the deviation cumulative quantity sequence; Based on the sampling time sequence, determine whether the minimum manifold distance corresponding to each future state point is continuously greater than zero, and accumulate the number of consecutive sampling points with a continuous minimum manifold distance greater than zero to obtain a continuous deviation time sequence; The fault risk index is obtained by combining the minimum manifold distance sequence, trajectory evolution direction deviation sequence, cumulative deviation sequence, and continuous deviation duration sequence according to the sampling time order.

9. The method for rapid prediction and alarm of power grid faults based on artificial intelligence according to claim 1, characterized in that, The generation of the fault prediction results and alarm information includes: Read the fault risk indicators in the order of sampling time, and extract the fault risk indicator values ​​corresponding to each sampling time. The fault risk index value corresponding to each sampling time is compared with multiple risk level intervals, and the risk level corresponding to each sampling time is determined based on the comparison results. Based on the risk level at each sampling time, generate corresponding fault prediction results, and output alarm information of the corresponding level based on the fault prediction results at each sampling time.

Images