Power system state evaluation and early warning method based on dynamic Bayesian network

Abstract

The invention provides an electric power system state evaluation and early warning method based on a dynamic Bayesian network, which relates to the technical field of electric power, and comprises the following steps: obtaining time sequence historical data and topological relation data of an electric power system, extracting time sequence characteristics from the time sequence historical data based on a hierarchical sliding window method of an alarm level, and calculating the state of the electric power system according to the extracted time sequence characteristics; on the basis of the topological relation data, constructing a domain structure in a topological conversion mode; performing dynamic Bayesian network modeling according to the time sequence characteristics and the domain structure; according to the dynamic Bayesian network, constructing a conditional probability table and a state transition matrix by using a cross-domain prior probability in combination with a maximum likelihood estimation method; calculating a prediction state of the domain structure based on the conditional probability table and the state transition matrix in combination with monitoring data of each node in the domain structure; and performing power system state evaluation and early warning according to the prediction state of each domain structure in the power system. According to the invention, a new methodology is provided for fault prediction and operation state management of the intelligent power grid.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electric power, in particular, and particularly relates to a power system state assessment and early warning method based on a dynamic Bayesian network. BACKGROUND

[0002] With the advancement of new power system construction, the scale of the power grid is continuously expanding, and the high proportion of renewable energy access and multi-element load characteristics significantly increase the complexity of system operation. The intelligent development of the power grid makes the multi-source data generated in the power system (such as alarm data, operation data, operation logs, and environmental data) present the characteristics of massiveness, high dimensionality, and dynamic correlation. In the prior art, power grid state assessment is mostly based on static rules or single-dimensional statistical analysis, which has the problems of insufficient time sequence correlation, weak multi-source data fusion capability, and low early warning accuracy. The cross-domain coupling relationship between device alarms, power parameters, operation behaviors, and environmental factors is not fully explored, and it cannot adapt to real-time dynamic scenarios. Dynamic Bayesian networks have the advantages of handling time-dependent dependencies and uncertain reasoning, but the prior art does not combine them with power grid multi-source data and device topology for system state assessment and early warning.

[0003] The current power system state assessment and early warning method has the following technical problems:

[0004] (1) Insufficient time sequence dynamic modeling capability: traditional methods (such as static rules or single-dimensional statistics) cannot effectively capture the dynamic correlation of the evolution of the power grid state over time (such as the continuous fluctuation of the device state with the change of the load and the environment), resulting in lagging assessment and lack of foresight in early warning.

[0005] (2) Defects in multi-source heterogeneous data fusion: the prior art does not fully explore the cross-domain coupling relationship between device alarms, power parameters, operation logs, and environmental data, making it difficult to support complex causal reasoning.

[0006] (3) Lack of system-level cascading risk prediction: existing methods focus on single devices or local states, lack the ability to model coupled risks across time slices (such as fault propagation timing) and across levels (such as main transformers to feeders to users), and are difficult to predict large-scale cascading failures.

[0007] Specifically, patent application No. CN202510225388.3 discloses a bridge intelligent inspection evaluation system and an inspection evaluation method, which comprises: a cluster bridge multi-source information database building module for building a cluster bridge multi-source information database according to collected bridge design, construction, maintenance, disease and service environment information; an intelligent inspection device for identifying and positioning the bridge to be evaluated, taking photos of the disease, and identifying the type and location of the disease, and uploading the identified disease information to the cluster bridge multi-source information database; and a cloud evaluation module for performing risk dynamic evaluation on the bridge to be evaluated based on a Bayesian network using the cluster bridge multi-source information stored in the cluster bridge multi-source information database.

[0008] The above-mentioned bridge intelligent inspection evaluation system does not integrate the network topology as a constraint into the Bayesian network, and does not form a cross-level system risk transmission path.

[0009] Patent application No. CN201410027588.X discloses a kind of overhead transmission line operating state evaluation method based on bidirectional Bayesian network, can carry out real-time evaluation to overhead transmission line operating state.It constructs the Bayesian network structure of transmission line operating state evaluation by taking various factors influencing transmission line operating state as condition attribute set, the operating state of line as decision attribute, obtains condition probability table according to sample training, utilizes the bidirectional reasoning technique special for Bayesian network, not only can judge the operating state of line by cause-effect reasoning, but also can identify state hidden danger by diagnostic reasoning;When there is evaluation deviation, it can be prewarned and corrected through self-feedback system, real-time dynamic correction evaluation database and network structure and parameter to adapt to update, and practically guarantee the healthy operation of transmission line.

[0010] The transmission line operating state evaluation method does not use multi-source heterogeneous data as sample training, does not integrate topology as a hard constraint into the Bayesian network, and finally does not perform system-wide state evaluation.

[0011] In view of the problems in the related art, no effective solution has been proposed so far. SUMMARY

[0012] Therefore, the present application provides a power system state evaluation and early warning method based on dynamic Bayesian network to solve the above-mentioned problems.

[0013] To solve the above-mentioned problems, the present application adopts the following specific technical solutions:

[0014] A power system state evaluation and early warning method based on dynamic Bayesian network, comprising the following steps:

[0015] S1, acquire time sequence historical data and topological relationship data of the power system, extract time sequence features from the time sequence historical data based on a hierarchical sliding window method of alarm levels, and construct a domain structure based on the topological relationship data and by using a topological conversion method;

[0016] S2, model a dynamic Bayesian network according to the time sequence features and the domain structure, and construct a conditional probability table and a state transition matrix according to the dynamic Bayesian network and by using a cross-domain prior probability combined with a maximum likelihood estimation method;

[0017] S3, calculate a predicted state of the domain structure based on the conditional probability table and the state transition matrix and in combination with monitoring data of each node in the domain structure, and perform power system state assessment and early warning according to the predicted states of each domain structure in the power system.

[0018] Preferably, the acquiring of the time sequence historical data and the topological relationship data of the power system, the extraction of the time sequence features from the time sequence historical data based on the hierarchical sliding window method of alarm levels, and the construction of the domain structure based on the topological relationship data and by using the topological conversion method include the following steps:

[0019] S11, acquire time sequence historical data and topological relationship data respectively based on a time sequence database and a relational database of a substation monitoring system;

[0020] S12, construct a hierarchical sliding window based on an alarm level based on a sliding window method, and perform discretization processing on the time sequence historical data according to the alarm level;

[0021] S13, extract time sequence features within the hierarchical sliding window according to the discretized time sequence historical data;

[0022] S14, determine topological connection relationships between power equipment according to the topological relationship data, and convert the power system into a domain structure by using a topological conversion method.

[0023] Preferably, the modeling of the dynamic Bayesian network according to the time sequence features and the domain structure, and the construction of the conditional probability table and the state transition matrix according to the dynamic Bayesian network and by using the cross-domain prior probability combined with the maximum likelihood estimation method include the following steps:

[0024] S21, perform time slicing on the time sequence features based on a hierarchical dynamic sliding window to obtain a plurality of time slices;

[0025] S22, determine nodes of the dynamic Bayesian network and a dependency relationship of the time slices according to the domain structure and the time slices, and construct the dynamic Bayesian network based on the nodes and the dependency relationship;

[0026] S23, learning parameters within the domain by using the dynamic Bayesian network, and constructing a conditional probability table and a state transition matrix based on cross-domain prior probability and state transition analysis.

[0027] Preferably, the nodes of the dynamic Bayesian network include observation nodes, hidden state nodes, intervention nodes, covariate nodes, and domain nodes; the observation nodes are power operation data; the hidden state nodes are operation states of the domains; the intervention nodes are system operation logs; the covariate nodes are environmental data; and the domain nodes are inter-domain connection relationships.

[0028] Preferably, the learning parameters within the domain by using the dynamic Bayesian network, and constructing a conditional probability table and a state transition matrix based on cross-domain prior probability and state transition analysis includes the following steps:

[0029] S231, constructing a conditional probability table based on the dynamic Bayesian network, cross-domain prior probability, and maximum likelihood estimation method;

[0030] S232, performing state transition analysis on the hidden state nodes based on the dynamic Bayesian network, and constructing a state transition matrix according to the analysis results.

[0031] Preferably, the constructing a conditional probability table based on the dynamic Bayesian network, cross-domain prior probability, and maximum likelihood estimation method includes the following steps:

[0032] S2311, determining a step influence coefficient matrix according to the domain structure, and generating a domain structure topology prior probability according to the step influence coefficient matrix;

[0033] S2312, calculating an original conditional probability table based on the maximum likelihood estimation method according to the co-occurrence frequency of parent node combinations and child node states within the domain structure;

[0034] S2313, correcting and fusing the original conditional probability table by using the domain structure topology prior probability to obtain a conditional probability table.

[0035] Preferably, the performing state transition analysis on the hidden state nodes based on the dynamic Bayesian network, and constructing a state transition matrix according to the analysis results includes the following steps:

[0036] S2321, based on time series historical data of the power system, counting the number of changes of the hidden state nodes between adjacent time slices;

[0037] S2322, calculating a transition probability according to the number of changes of the hidden state nodes between adjacent time slices;

[0038] S2323, generating a state transition matrix according to the calculated transition probability.

[0039] Preferably, the conditional probability table and the state transition matrix are used to calculate the predicted state of the domain structure based on the monitoring data of each node in the domain structure; and the power system state assessment and early warning based on the predicted state of each domain structure in the power system comprises the following steps:

[0040] S31, obtaining the current monitoring data of each node in the domain structure and initializing the current state distribution of the domain structure;

[0041] S32, calculating and correcting the posterior probability of the current hidden state of the domain structure based on the current monitoring data of each node in the domain structure, the state transition matrix and the conditional probability table;

[0042] S33, performing inter-domain risk superposition early warning analysis based on the posterior probability of the current hidden state of the domain structure and a preset hierarchical early warning scheme to obtain a warning result.

[0043] Preferably, the calculation and correction of the posterior probability of the current hidden state of the domain structure based on the current monitoring data of each node in the domain structure, the state transition matrix and the conditional probability table comprises the following steps:

[0044] S321, calculating the initial state prediction probability based on the current monitoring data of each node in the domain structure, the state transition matrix and a prediction state calculation formula;

[0045] S322, obtaining the likelihood probability based on the conditional probability table and correcting the prediction state based on the likelihood probability to obtain the corrected state prediction probability;

[0046] S323, performing secondary correction on the state prediction probability based on the inter-domain connection relationship to obtain the posterior probability of the current hidden state of the domain structure.

[0047] Preferably, the inter-domain risk superposition early warning analysis based on the posterior probability of the current hidden state of the domain structure and a preset hierarchical early warning scheme to obtain a warning result comprises the following steps:

[0048] S331, determining the risk values of different domains in the domain structure based on the posterior probability of the current hidden state of the domain structure;

[0049] S332, calculating the hierarchical risk aggregation value based on the risk values of different domains and the inter-domain aggregation logic;

[0050] S333, generating a final warning signal based on the hierarchical early warning scheme and the hierarchical risk aggregation value.

[0051] The application has the beneficial effect that the application upgrades the traditional passive data analysis to active causal exploration, is particularly suitable for the complex scene of multi-factor coupling in the power system, and is difficult to directly infer the causal relationship through observation data, and provides a new methodology for fault prediction and operation state management of the smart grid. BRIEF DESCRIPTION OF DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description only constitute some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort based on these drawings. In the drawings:

[0053] Figure 1 is a flowchart of a power system state evaluation and early warning method based on a dynamic Bayesian network according to an embodiment of the present application;

[0054] Figure 2 is a domain structure example schematic diagram in a power system state evaluation and early warning method based on a dynamic Bayesian network according to an embodiment of the present application;

[0055] Figure 3 is a dependency relationship schematic diagram within a time slice in a power system state evaluation and early warning method based on a dynamic Bayesian network according to an embodiment of the present application;

[0056] Figure 4 is a dependency relationship example schematic diagram between time slices in a power system state evaluation and early warning method based on a dynamic Bayesian network according to an embodiment of the present application. DETAILED DESCRIPTION

[0057] In order to enable those skilled in the art to better understand the technical solutions in the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments only constitute some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without any creative effort should fall within the scope of protection of the present application.

[0058] According to an embodiment of the present application, a power system state evaluation and early warning method based on a dynamic Bayesian network is provided.

[0059] The present application will be further described in conjunction with the drawings and specific embodiments, as shown in Figure 1 The power system state evaluation and early warning method based on a dynamic Bayesian network according to the embodiment of the present application includes the following steps:

[0060] S1, acquire time sequence historical data and topological relationship data of a power system, extract time sequence features of the time sequence historical data based on a hierarchical sliding window method of alarm levels, and construct a domain structure based on the topological relationship data and by using a topological conversion method;

[0061] As a preferred embodiment, the acquiring of the time sequence historical data and the topological relationship data of the power system, the extraction of the time sequence features of the time sequence historical data based on the hierarchical sliding window method of alarm levels, and the construction of the domain structure based on the topological relationship data and by using the topological conversion method include the following steps:

[0062] S11, acquire time sequence historical data and topological relationship data based on a time sequence database and a relational database of a substation monitoring system, respectively;

[0063] Specifically, historical alarm data, power operation data, system operation logs, environmental data, and device topological relationship data are acquired from a time sequence database and a relational database of a monitoring system of a substation or a distribution network as data samples. The specific data acquisition is shown in Table 1.

[0064] Table 1: Multi-source data sources

[0065] S12, based on a sliding window method, construct a hierarchical sliding window based on alarm levels, and discretize the time sequence historical data according to the alarm levels;

[0066] Specifically, based on the sliding window method, a hierarchical sliding window method based on alarm levels is designed, and historical alarm data, power data operation data, and system operation logs are discretized according to different alarm levels.

[0067] The hierarchical sliding window design formula is as follows:

[0068] ;

[0069] In the above formula, Window alarm represents the hierarchical sliding window, t a represents the alarm trigger time, T alarm represents the sliding window time corresponding to the alarm level, and a represents the alarm level, wherein the specific alarm level and the coefficient relationship are divided as shown in Table 2:

[0070] Table 2: Alarm level coefficient division

[0071] The historical alarm data, power operation data, and other time sequence data in the multi-source data are discretized according to the alarm levels, and the data features are extracted within the hierarchical alarm window.

[0072] S13, extracting time sequence features in the hierarchical sliding window according to the time sequence historical data after the discretization processing;

[0073] It should be noted that the extraction of time sequence data features based on the alarm level in the hierarchical sliding window includes: feature extraction of power operation data and system operation log based on the alarm level, dynamic extraction of power operation data features through the alarm grading weighting method, and extraction of operation log through the operation log grading backtracking method.

[0074] Among them, the historical alarm data extracts the alarm time, alarm device ID, alarm type and alarm level;

[0075] For power operation data: first, design the weight coefficient: dynamically adjust the feature extraction weight according to the alarm level (1-7) level, the formula is as follows:

[0076] ;

[0077] In the formula, α represents the alarm level, δ α represents the feature weight.

[0078] The power operation feature such as voltage, current, active power and other electrical quantity data feature extraction formula:

[0079] ;

[0080] In the formula, F w represents the electrical quantity data feature, μ α represents the mean value of electrical quantity w in the time window, σ w represents the standard deviation, and dynamic feature extraction is used to strengthen the fluctuation characteristics under high-level alarm.

[0081] For system operation log: backtracking operation log to construct operation time sequence chain according to alarm level. Among them, alarm levels 1-2 are not backtracked, alarm levels 3-5 are backtracked for 1 window, and alarm levels greater than 5 are backtracked for 3 windows.

[0082] S14, determining the topological connection relationship between power equipment according to the topological relationship data, and converting the power system into a domain structure through topological conversion.

[0083] Specifically, the device topological connection relationship is converted into a domain (Region) structure, and a three-level domain structure based on functional roles: a main transformer domain (the main transformer is the core), a protection and control domain (the circuit breaker is the core) and a transmission channel domain (the line is the core). The power system topology is converted into a functional domain structure, each domain is composed of a key node and its direct downstream non-key equipment, and the specific steps include the following:

[0084] (1) Key node definition and topological role, as shown in Table 3;

[0085] Table 3 Key node definition and topology role

[0086] (2) Non-key node definition and attribution rule, as shown in Table 4;

[0087] Table 4 Non-key node definition and attribution rule

[0088] (3) Domain structure construction, for example, as shown in the following formula. Figure 2

[0089] S2, according to the time sequence characteristics and the domain structure, dynamic Bayesian network modeling is carried out; according to the dynamic Bayesian network, and by using the cross-domain prior probability combined with the maximum likelihood estimation method, the conditional probability table and the state transition matrix are constructed;

[0090] Specifically, the dynamic Bayesian network (DBN) modeling includes two steps of Bayesian network design and parameter learning based on the domain structure.

[0091] As a preferred embodiment, the dynamic Bayesian network modeling according to the time sequence characteristics and the domain structure; according to the dynamic Bayesian network, and by using the cross-domain prior probability combined with the maximum likelihood estimation method, the conditional probability table and the state transition matrix are constructed, including the following steps:

[0092] S21, based on the hierarchical dynamic sliding window, the time sequence characteristics are time-sliced to obtain a plurality of time slices;

[0093] S22, according to the domain structure and the time slice, the nodes of the dynamic Bayesian network and the dependence relationship of the time slice are determined respectively, and the dynamic Bayesian network is constructed based on the nodes and the dependence relationship;

[0094] As a preferred embodiment, the nodes of the dynamic Bayesian network include observation nodes, hidden state nodes, intervention nodes, covariate nodes and domain nodes; the observation nodes are power operation data; the hidden state nodes are the operation states of the domains; the intervention nodes are system operation logs; the covariate nodes are environmental data; and the domain nodes are inter-domain connection relationships.

[0095] S23, using the dynamic Bayesian network to learn the parameters within the domain, and based on the cross-domain prior probability, combining with the state transition analysis, the conditional probability table and the state transition matrix are constructed.

[0096] As a preferred embodiment, the dynamic Bayesian network is used to learn the parameters within the domain, and based on the cross-domain prior probability, combining with the state transition analysis, the conditional probability table and the state transition matrix are constructed, including the following steps: ​

[0097] S231, constructing a conditional probability table based on a dynamic Bayesian network and using cross-domain prior probability and maximum likelihood estimation method;

[0098] As a preferred embodiment, the step of constructing a conditional probability table based on a dynamic Bayesian network and using cross-domain prior probability and maximum likelihood estimation method comprises the following steps:

[0099] S2311, determining a step influence coefficient matrix according to the domain structure, and generating a domain structure topology prior probability according to the step influence coefficient matrix;

[0100] S2312, calculating an original conditional probability table based on a maximum likelihood estimation method according to the co-occurrence frequency of parent node combinations and child node states in the domain structure;

[0101] S2313, modifying and fusing the original conditional probability table using the domain structure topology prior probability to obtain a conditional probability table.

[0102] S232, based on a dynamic Bayesian network, performing state transition analysis on the hidden state node, and constructing a state transition matrix according to the analysis result.

[0103] As a preferred embodiment, the step of performing state transition analysis on the hidden state node based on a dynamic Bayesian network and constructing a state transition matrix according to the analysis result comprises the following steps:

[0104] S2321, based on the time sequence historical data of the power system, counting the number of changes of the hidden state node between adjacent time slices;

[0105] S2322, calculating a transition probability according to the number of changes of the hidden state node between adjacent time slices;

[0106] S2323, generating a state transition matrix according to the calculated transition probability.

[0107] It should be noted that the specific implementation steps of the Bayesian network design based on the domain structure are as follows:

[0108] (1) Time slicing, using hierarchical dynamic sliding window to slice time, and the observation variables inside each time window are the historical alarm data, power operation data and system operation log data after domain feature extraction.

[0109] (2) Defining nodes, including: observation nodes, hidden state nodes, intervention nodes, covariate nodes and domain nodes;

[0110] The observation nodes are power operation data; including current, voltage, active power, load rate;

[0111] The hidden state nodes are the running states of the domains;

[0112] Intervening nodes are system operation logs;

[0113] Covariate nodes are environmental data;

[0114] Domain nodes are inter-domain connection relationships.

[0115] (3) Define the dependency relationship. Within the same time slice, the dependency relationship within the domain is the causal relationship of variables (for example, in the power supply domain: temperature → main variable load rate → main variable health status); the dependency relationship between domains is the connection relationship, for example, the protection control domain device is only affected by the state of the directly connected power supply domain and the operation record of each device within the domain, and the line voltage is jointly determined by the upstream circuit breaker and the current load rate.

[0116] Between time slices, the dependency relationship is the state transition matrix.

[0117] Specific implementation examples:

[0118] Specific variables and examples of node definition are shown in Table 5.

[0119] Table 5 Node definition

[0120] The dependency relationship within the time slice is established, as shown in Table 6, and the dependency relationship between time slices is established, as shown in Table 7. Figure 3 Figure 4 It should be noted that the domain parameter learning mainly includes: designing the conditional probability table within the domain, calculating and generating the transition matrix, and the specific implementation process is as follows:

[0121] It should be noted that the domain parameter learning mainly includes: designing the conditional probability table within the domain, calculating and generating the transition matrix, and the specific implementation process is as follows:

[0122] 1. Design the adjustment probability table, adopt the cross-domain prior probability combined with the maximum likelihood estimation method (Maximum Likelihood Estimation, MLE) to construct the conditional probability table; specifically, construct the Bayesian network with the domain as the smallest unit, and adopt the cross-domain prior probability combined with the maximum likelihood estimation method in the construction process of the conditional probability table.

[0123] (1) According to the domain structure (main variable domain → protection control domain → transmission channel domain), define the topological influence coefficient matrix W topo :

[0124] ;

[0125] Subsequently, the domain structure topological prior probability is generated, and the calculation formula is as follows:

[0126] ;

[0127] In the formula, P (W) represents the topological prior probability, ​​P (x | pa) = N (x, pa) / N (pa)

[0128] (2) Construct the original conditional probability table of the current domain structure by using the maximum likelihood estimation method. The core idea of the traditional maximum likelihood estimation method is to select the parameter with the maximum probability of observation data. For the conditional probability table (CPT), the MLE calculates the conditional probability by counting the co-occurrence frequency of the parent node combination and the state of the child node in the historical data, and the specific formula is as follows:

[0129] ;

[0130] In the formula, P (x | pa) represents the conditional probability, pa represents the set of other nodes in the domain structure that can affect the state of the child node x. N (x, pa) represents the frequency when the state of the child node is x and the state of the parent node combination is pa.

[0131] (3) Fuse and correct the conditional probability table by using the topological prior probability and the maximum likelihood estimation method, and the calculation formula is as follows:

[0132] ;

[0133] In the formula, represents the corrected conditional probability, and a is a normalization coefficient.

[0134] Specific implementation example: taking the operating state of the main transformer domain as an example, first, pre-process to obtain discrete data based on the time window, collect related devices and nodes in the domain structure, and the example is shown in Table 6.

[0135] Table 6 State of nodes in the domain structure

[0136] Among them, the load rate of the main transformer includes high (more than 90%), medium (70%-90%), and low (less than 70%); the temperature is divided into high (40 degrees Celsius), medium (30-40 degrees Celsius), and low (less than 30 degrees Celsius); the circuit breaker operation has two states of on and off; the operating state of the main transformer is derived from the alarm data, and is mainly divided into normal, warning, and accident.

[0137] Generate the topological prior probability between domains: π topo = 0.7;

[0138] According to the frequency statistics of the historical data, as shown in Table 7;

[0139] Table 7 Frequency statistics of history

[0140] The conditional probability can be calculated by combining the frequency statistics;

[0141] ;

[0142] The final result is .

[0143] Similarly, the conditional probability of the transformer operating state under different scenarios can be calculated. Laplace smoothing is used to process the parent node combination that has never appeared in the historical data. The calculation formula of Laplace smoothing is as follows:

[0144] ;

[0145] In the formula, P(X=x|pa) represents the probability of the child node state being x under the premise that the parent node state combination is pa, ∂ represents the smoothing coefficient, |X| represents the possible state data, Count(X=x, pa) represents the number of samples in which the parent node state combination is pa and the child node state is x, and Count(pa) represents the number of samples in which the parent node state combination is pa in the sample.

[0146] The conditional probability is finally generated by using the prior probability combined with the maximum likelihood estimation method, as shown in Table 8.

[0147] Table 8 Conditional probability table

[0148] 2. Transition probability: The hidden state node changes between adjacent time slices are counted from the time series data. The transition probability is calculated and the transition matrix is generated by the following formula:

[0149] ;

[0150] In the formula, represents the probability of transitioning from state i to state j.

[0151] Specific implementation example: The number of node state changes is counted, as shown in Table 9.

[0152] Table 9 Number of node state changes

[0153] The transition probability is calculated, and a specific example is as follows:

[0154] ;

[0155] According to the calculated transition probability, the state transition matrix is generated, and the generation result is shown in Table 10.

[0156] Table 10 State transition matrix

[0157] S3, based on the conditional probability table and the state transition matrix, combining the monitoring data of each node in the domain structure, calculating the predicted state of the domain structure; according to the predicted state of each domain structure in the power system, performing power system state assessment and early warning.

[0158] As a preferred embodiment, the based on the conditional probability table and the state transition matrix, combining the monitoring data of each node in the domain structure, calculating the predicted state of the domain structure; according to the predicted state of each domain structure in the power system, performing power system state assessment and early warning comprises the following steps:

[0159] S31, obtaining the current monitoring data of each node in the domain structure, and initializing the current state distribution of the domain structure;

[0160] S32, according to the current monitoring data of each node in the domain structure, combining the state transition matrix and the conditional probability table, calculating and correcting the posterior probability of the current hidden state of the domain structure;

[0161] As a preferred embodiment, the according to the current monitoring data of each node in the domain structure, combining the state transition matrix and the conditional probability table, calculating and correcting the posterior probability of the current hidden state of the domain structure comprises the following steps:

[0162] S321, according to the current monitoring data of each node in the domain structure, combining the state transition matrix, using the prediction state calculation formula to calculate the initial state prediction probability;

[0163] S322, obtaining the likelihood probability based on the conditional probability table, and correcting the prediction state through the likelihood probability to obtain the corrected state prediction probability;

[0164] S323, based on the inter-domain connection relationship, the state prediction probability is corrected again to obtain the posterior probability of the current hidden state of the domain structure.

[0165] S33, according to the posterior probability of the current hidden state of the domain structure, combining the preset hierarchical early warning scheme, performing inter-domain risk superposition early warning analysis to obtain the early warning result.

[0166] As a preferred embodiment, the according to the posterior probability of the current hidden state of the domain structure, combining the preset hierarchical early warning scheme, performing inter-domain risk superposition early warning analysis to obtain the early warning result comprises the following steps:

[0167] S331, according to the posterior probability of the current hidden state of the domain structure, determining the risk value of different domains in the domain structure;

[0168] S332, according to the risk value of different domains, combining the inter-domain aggregation logic, calculating the hierarchical risk aggregation value;

[0169] S333, based on the hierarchical early warning scheme and the hierarchical risk aggregation value, generating the final early warning signal.

[0170] It should be noted that the posterior probability of the current hidden state node (device health state) is calculated based on the observation data of the nodes in the current domain structure, and the specific implementation steps are as follows:

[0171] Initialization, from the state distribution at initial time t=0 Start;

[0172] Prediction step, use transition matrix to calculate the predicted state of the hidden state node, the calculation formula is as follows:

[0173] ;

[0174] In the formula, A represents the state transition matrix, wherein represents the state change data of the hidden state node of the domain structure in the period from 1 to t-1. represents the probability predicted state based on the state transition matrix, represents the predicted state of t-1 period based on historical data.

[0175] Update step: correct the hidden node state by combining the observation data of each node in the domain structure in the current time window, the calculation formula is as follows:

[0176] ;

[0177] In the formula, represents the corrected posterior probability. represents the likelihood probability, which is obtained through the conditional probability table CPT.

[0178] In addition, the probability correction based on the inter-domain relationship includes:

[0179] (1) Input: the predicted probability of the hidden node state obtained;

[0180] (2) Find the adjacent domain according to the inter-domain connection relationship table and obtain the adjacent domain state, combine the inter-domain state decay coefficient to calculate the domain state, obtain the corrected posterior domain state, the formula is as follows:

[0181] ;

[0182] In the formula, the inter-domain state influence decay coefficient is λ cluster , wherein the decay coefficient rule is shown in Table 11;

[0183] Table 11 Decay coefficient rule

[0184] Specific implementation example: assuming that the initial period observation node (load rate) = high, the covariate node (temperature) = high, the intervention node (operation) is no operation, and the current device running state is normal:

[0185] ;

[0186] Prediction step: calculate the prediction step, that is, the state transition probability, and the result is as follows:

[0187] ;

[0188] Update step: obtain the conditional probability of load rate = high, temperature = high, and operation = no operation from the CPT.

[0189] ;

[0190] Calculate the posterior probability;

[0191] ;

[0192] Based on the inter-domain relationship, the probability is corrected.

[0193] Specifically, when the hierarchical risk is aggregated, the aggregated risk is calculated according to different domain roles in the system, and the calculation formula is as follows:

[0194] ;

[0195] Wherein, the aggregation logic is shown in Table 12;

[0196] Table 12 Aggregation logic example

[0197] In addition, the hierarchical early warning scheme is shown in Table 13;

[0198] Table 13 Hierarchical early warning scheme

[0199] Specifically, the early warning signal displays and alarms the warning information through the alarm list and the alarm sound, and pushes the key device information and core device information early warning through the form of short message.

[0200] In summary, with the above technical solutions of the present application, the traditional passive data analysis is upgraded to active causal exploration, which is especially suitable for the complex scene of multi-factor coupling in the power system, and the causal relationship is difficult to directly infer through observation data, providing a new methodology for fault prediction and operation state management of smart grid.

[0201] Those skilled in the art will appreciate that embodiments of the present application can be readily used as a method, a system or a computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer readable storage media (including, but not limited to, disk memory, optical storage medium and the like) embodying computer readable program code thereon.

[0202] The above description is provided as an enabling teaching of the application. It is to be understood that this description is not intended to limit the application. Modifications and variations are possible in light of the above teachings or can be acquired from practice of the application. For example, while a particular feature of the application can have been disclosed with respect to only one of several implementations, other implementations can include the feature. It is also possible for a feature to be added to or removed from the present application without departing from the scope of the application. Modifications, variations, and improvements can be incorporated as well.

Claims

1. A power system state assessment and early warning method based on dynamic Bayesian networks, characterized in that, Includes the following steps: S1. Obtain the time-series historical data and topology relationship data of the power system, extract time-series features from the time-series historical data based on the hierarchical sliding window method of alarm level, and construct the domain structure based on the topology relationship data and using topology transformation. S2. Based on temporal characteristics and domain structure, perform dynamic Bayesian network modeling; Based on dynamic Bayesian networks, and using cross-domain prior probabilities combined with maximum likelihood estimation, conditional probability tables and state transition matrices are constructed. S3. Based on the conditional probability table and state transition matrix, and combined with the monitoring data of each node in the domain structure, calculate the predicted state of the domain structure; and perform power system state assessment and early warning based on the predicted state of each domain structure in the power system.

2. The power system state assessment and early warning method based on dynamic Bayesian networks according to claim 1, characterized in that, The process of acquiring time-series historical data and topology data of the power system, extracting time-series features from the time-series historical data using a hierarchical sliding window method based on alarm levels, and constructing the domain structure based on the topology data and using topology transformation includes the following steps: S11. Based on the time-series database and relational database of the substation monitoring system, acquire time-series historical data and topology relationship data respectively; S12. Based on the sliding window method, construct a hierarchical sliding window based on alarm level, and discretize the time-series historical data according to the alarm level. S13. Extract time-series features within a hierarchical sliding window based on the discretized historical time-series data. S14. Determine the topological connection relationship between power equipment based on the topological relationship data, and transform the power system into a domain structure through topological transformation.

3. The power system state assessment and early warning method based on dynamic Bayesian networks according to claim 1, characterized in that, The dynamic Bayesian network modeling is performed based on temporal characteristics and domain structure. Based on the dynamic Bayesian network, and utilizing cross-domain prior probabilities combined with maximum likelihood estimation, the conditional probability table and state transition matrix are constructed, including the following steps: S21. Based on a hierarchical dynamic sliding window, time slices are made on the time series features to obtain several time slices; S22. Based on the domain structure and time slices, determine the dependencies between nodes and time slices in the dynamic Bayesian network, and construct the dynamic Bayesian network based on the nodes and dependencies. S23. Use dynamic Bayesian networks to learn intra-domain parameters, and construct conditional probability tables and state transition matrices based on cross-domain prior probabilities and state transition analysis.

4. The power system state assessment and early warning method based on dynamic Bayesian networks according to claim 3, characterized in that, The nodes of the dynamic Bayesian network include: observation nodes, hidden state nodes, intervention nodes, covariate nodes, and domain nodes; the observation nodes represent power operation data; the hidden state nodes represent the operating status of the domains; the intervention nodes represent system operation logs; the covariate nodes represent environmental data; and the domain nodes represent inter-domain connections.

5. The power system state assessment and early warning method based on dynamic Bayesian networks according to claim 3, characterized in that, The process of learning intra-domain parameters using dynamic Bayesian networks and constructing conditional probability tables and state transition matrices based on cross-domain prior probabilities and state transition analysis includes the following steps: S231. Based on dynamic Bayesian networks, a conditional probability table is constructed using cross-domain prior probability and maximum likelihood estimation. S232. Based on dynamic Bayesian networks, perform state transition analysis on hidden state nodes and construct a state transition matrix based on the analysis results.

6. The power system state assessment and early warning method based on dynamic Bayesian networks according to claim 5, characterized in that, The construction of the conditional probability table based on dynamic Bayesian networks and using cross-domain prior probability and maximum likelihood estimation includes the following steps: S2311. Determine the topology influence coefficient matrix based on the domain structure, and generate the domain structure topological prior probability based on the topology influence coefficient matrix. S2312. Based on the co-occurrence frequency of parent node combinations and child node states within the domain structure, calculate the original conditional probability table using the maximum likelihood estimation method. S2313. The original conditional probability table is modified and fused using the prior probability of the domain structure topology to obtain the conditional probability table.

7. The power system state assessment and early warning method based on dynamic Bayesian networks according to claim 5, characterized in that, The process of performing state transition analysis on hidden state nodes based on dynamic Bayesian networks and constructing a state transition matrix based on the analysis results includes the following steps: S2321. Based on the time-series historical data of the power system, count the number of hidden state node changes between adjacent time slices; S2322. Calculate the transition probability based on the number of hidden state node changes between adjacent time slices; S2323. Generate the state transition matrix based on the calculated transition probabilities.

8. The power system state assessment and early warning method based on dynamic Bayesian networks according to claim 1, characterized in that, The process of calculating the predicted state of a domain structure based on conditional probability tables and state transition matrices, combined with monitoring data from each node in the domain structure, and then assessing and issuing early warnings for the power system state based on the predicted states of each domain structure within the power system includes the following steps: S31. Obtain the current monitoring data of each node within the domain structure and initialize the current state distribution of the domain structure; S32. Based on the current monitoring data of each node in the domain structure, combined with the state transition matrix and the conditional probability table, calculate and correct the posterior probability of the current hidden state of the domain structure. S33. Based on the posterior probability of the current hidden state of the domain structure, and combined with the preset hierarchical early warning scheme, perform inter-domain risk superposition early warning analysis to obtain the early warning result.

9. A power system state assessment and early warning method based on dynamic Bayesian networks according to claim 8, characterized in that, The process of calculating and correcting the posterior probability of the current hidden state of the domain structure based on the current monitoring data of each node within the domain structure, combined with the state transition matrix and the conditional probability table, includes the following steps: S321. Based on the current monitoring data of each node in the domain structure, combined with the state transition matrix, calculate the initial state prediction probability using the prediction state calculation formula. S322. Obtain the likelihood probability based on the conditional probability table, and correct the predicted state using the likelihood probability to obtain the corrected state prediction probability. S323. Based on the inter-domain connectivity, the state prediction probability is modified a second time to obtain the posterior probability of the current hidden state of the domain structure.

10. A power system state assessment and early warning method based on dynamic Bayesian networks according to claim 8, characterized in that, The step of performing inter-domain risk superposition early warning analysis based on the posterior probability of the current hidden state of the domain structure and in conjunction with a preset hierarchical early warning scheme to obtain the early warning result includes the following steps: S331. Determine the risk value of different domains in the domain structure based on the posterior probability of the current hidden state of the domain structure. S332. Based on the risk values ​​of different domains, and combined with the inter-domain aggregation logic, calculate the hierarchical risk aggregation value; S333. Based on the graded early warning scheme and the layered risk aggregation value, the final early warning signal is generated.

Images