Fault prediction and early warning system for energy storage box type transformer substation

Abstract

The invention relates to the technical field of energy storage equipment monitoring, and discloses a fault prediction and early warning system for an energy storage box-type substation. The multi-modal data acquisition and preprocessing module, the macroscopic operation mode vector extraction module, the real-time causal calculation engine module, the real-time dynamic causal entropy map construction module and the working condition causal baseline dynamic mapping module generate a theoretical health baseline map corresponding to the current working condition in real time according to the macroscopic operation mode vector; an atlas topology difference and anomaly detection module generates a difference atlas by comparing the real-time dynamic causal entropy atlas with the theoretical health baseline atlas; the topology fingerprint and fault diagnosis module analyzes the difference atlas to extract topology fingerprint vectors and matches a fault knowledge base to output a diagnosis conclusion. By constructing the health base line which is dynamically adjusted along with the working condition, normal operation fluctuation can be effectively filtered out, the accuracy of fault detection is improved, and diagnosis from abnormal detection to specific fault attribution is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of energy storage device monitoring, in particular to a fault prediction and early warning system for energy storage box-type substations. BACKGROUND

[0002] As a key unit for energy deployment and stable operation of modern power systems, the safety and reliability of energy storage box-type substations are of great importance. To ensure their stable operation and avoid economic losses or safety accidents caused by equipment failure, developing effective fault prediction and early warning technology is a common demand in the industry.

[0003] Some existing fault detection technologies rely on setting fixed alarm thresholds for key physical quantities such as temperature, voltage, and current. However, the physical operating conditions of energy storage box-type substations are not constant, and they will change dynamically with factors such as state of charge, load size, and ambient temperature. Fixed alarm thresholds are difficult to adapt to such changes, and may frequently trigger false alarms during high-load operation, while missing real early-stage faults due to insufficient sensitivity during low-load operation.

[0004] Some technical solutions introduce models based on statistics or machine learning to establish health benchmarks. However, the health benchmarks relied on by these models are usually static or only correspond to a limited number of discrete operating conditions. For the actual operation process of energy storage systems, which continuously and smoothly transitions between different operating conditions, these static or semi-static benchmark models still have the problem of insufficient adaptability, which affects the accuracy of their diagnostic conclusions.

[0005] Most existing technologies focus on the numerical changes of monitoring data, and less on the changes in the causal transmission relationship between internal components of the system. The evolution process of equipment failure is essentially a deviation from the internal physical laws of the system, which will manifest as a change in the information flow transmission pattern between components. Existing methods have limitations in early-stage fault tracing and machine interpretation because they do not go deep enough. SUMMARY

[0006] To address the shortcomings of existing technologies, the present application provides a fault prediction and early warning system for energy storage box-type substations, which solves the problem that the health baseline used by existing fault diagnosis methods is difficult to adapt to the dynamic changes in the operating conditions of energy storage box-type substations, leading to low fault warning accuracy and frequent false positives and false negatives.

[0007] To achieve the above purpose, the present application realizes the following technical solutions: a fault prediction and early warning system for energy storage box-type substations, comprising: a multi-modal data acquisition and preprocessing module for acquiring original time series data and processing to output synchronized and standardized multivariate time series data sets; a macro-operation mode vector extraction module connected with the multi-modal data acquisition and preprocessing module, configured to generate a macro-operation mode vector quantifying a current physical operation condition based on the multivariate time series dataset; a real-time causal computation engine module connected with the multi-modal data acquisition and preprocessing module, configured to compute a real-time transfer entropy matrix based on the multivariate time series dataset; a real-time dynamic causal entropy graph construction module connected with the real-time causal computation engine module, configured to convert the real-time transfer entropy matrix into a real-time dynamic causal entropy graph; a condition causal baseline dynamic mapping module connected with the macro-operation mode vector extraction module, configured to generate a theoretical health baseline graph corresponding to the current condition according to the macro-operation mode vector; a graph topology difference and anomaly detection module connected with the real-time dynamic causal entropy graph construction module and the condition causal baseline dynamic mapping module, configured to receive the real-time dynamic causal entropy graph and the theoretical health baseline graph, compute a topology structure difference therebetween to generate a difference graph, and detect whether a topology anomaly exists based on the difference graph; a topology fingerprint and fault diagnosis module connected with the graph topology difference and anomaly detection module, configured to, when the topology anomaly is detected, analyze the difference graph to extract a topology fingerprint vector, and match the topology fingerprint vector with a preset fault knowledge base to output a fault diagnosis conclusion of the energy storage box-type substation.

[0008] Preferably, the condition causal baseline dynamic mapping module comprises: a pre-trained nonlinear mapping model is used to receive the macro-operation mode vector as input and output a theoretical health transfer entropy matrix; and based on the theoretical health transfer entropy matrix, a noise threshold processing is applied to generate the theoretical health baseline graph.

[0009] Further, the nonlinear mapping model can be represented as: wherein, M TE,base (t) is F map is a nonlinear mapping model, V om (t) is the macro-operation mode vector at the current time provided by the macro-operation mode vector extraction module, and t is a time variable.

[0010] Preferably, the nonlinear mapping model is obtained by training a training dataset containing paired macro-operation mode vectors and health transfer entropy matrices in an offline training stage.

[0011] Further, the loss function of the offline training stage can be defined as:​ ; In the formula, This represents the value of the loss function. Represents the training dataset. This represents the total number of samples in the training dataset. Represents the macroscopic operating mode vector. Represents the true health delivery matrix. This represents the output matrix of the mapping model.

[0012] Preferably, the real-time dynamic causal entropy map construction module includes: A specific noise threshold is applied to the elements in the real-time transfer entropy matrix for comparison; When an element is greater than the specific noise threshold, a corresponding weighted directed edge is established in the real-time dynamic causal entropy graph.

[0013] Preferably, the specific noise threshold is determined individually for each pair of time series using a proxy data method.

[0014] Preferably, the map topology difference and anomaly detection module includes: The difference map is generated by calculating the difference between the weight matrix of the real-time dynamic causal entropy map and the theoretical health baseline map. The difference map is then examined according to the topological anomaly determination rules to detect topological anomalies.

[0015] Furthermore, the difference map Any two nodes V j and V i directed edge weights between The calculation is as follows: ; In the formula, For difference spectrum From node V j Point to V i The weight of the edge. For real-time dynamic causal entropy map From node V j Point to V i The weight of the edge. For theoretical health baseline map From node V j Point to V i The weight of the edges.

[0016] Preferably, the topology fingerprinting and fault diagnosis module includes: When the topological anomaly is detected, a topological fingerprint vector is extracted from the differential map. And the topology fingerprint vector is calculated with the pre-stored typical topology fingerprint vector in the fault knowledge base to match and output a fault diagnosis conclusion.

[0017] Preferably, the topology fingerprint and fault diagnosis module extracts the topology fingerprint vector by extracting and combining at least one of the following features: Node-level topology features, including weighted in-degree deviation or weighted out-degree deviation; Graph-level topology features, including network centrality changes of key nodes; Subgraph-level topology features, including the count of specific network motifs.

[0018] Preferably, the topology fingerprint and fault diagnosis module includes the following after similarity calculation: Compare the maximum similarity value with a preset diagnosis confidence threshold; If the maximum similarity value is greater than the diagnosis confidence threshold, output the corresponding fault label; If the maximum similarity value is less than or equal to the diagnosis confidence threshold, output the diagnosis conclusion as unknown fault.

[0019] Preferably, the macro operating mode vector extraction module includes: Perform normalization processing on the continuous operating condition features in the preset operating condition features; Perform one-hot encoding processing on the discrete operating condition features in the preset operating condition features; And the processed continuous operating condition features and discrete operating condition features are spliced to construct the macro operating mode vector.

[0020] The present application provides a fault prediction and early warning system for energy storage tank substation. Has the following beneficial effects: 1、The present application sets up the operating condition causal baseline dynamic mapping module, which can generate a theoretical health baseline graph that accurately corresponds to the real-time changing macro operating mode vector, so that the system can effectively filter out the causal topology structure fluctuations caused by the normal operating condition changes of the energy storage tank substation when performing topology difference comparison, thereby significantly reducing the false positive rate of fault warning and improving the accuracy of anomaly detection.

[0021] 2、The present application constructs a real-time dynamic causal entropy graph, which improves the relationship between each monitoring point in the system from numerical correlation to causal information flow level. Unlike traditional methods that rely only on data thresholds, this scheme can intuitively reveal the propagation path and impact range of the fault by analyzing the causal paths of new, disappearing or weight changes in the difference graph, providing a clear physical basis for fault tracing.

[0022] 3、The present application can extract a topological fingerprint vector containing node level, graph level and subgraph level features from the difference graph through the topological fingerprint and fault diagnosis module, and match the vector with the preset fault knowledge base, establish a mapping relationship between the detected topological anomaly pattern and the specific fault type, realize the leap from anomaly detection to fault diagnosis, can output clear fault conclusion, not only can diagnose known faults, but also can identify unknown fault types through the matching degree result, and provide more accurate decision support. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 A method flowchart of the present application; Figure 2 A working condition causal baseline dynamic mapping principle schematic diagram of the present application. DETAILED DESCRIPTION

[0024] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the specification of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0025] Please refer to the drawings in the specification of the present application Figure 1 - the drawings in the specification of the present application Figure 2 The embodiment of the present application provides a fault prediction and early warning system for energy storage box-type substation, which comprises: A multi-modal data acquisition and preprocessing module is used to obtain original time series data from the energy storage box-type substation and process the original time series data to output a synchronized and standardized multi-element time series data set. Specifically, the multi-modal data acquisition and preprocessing module of the present embodiment is used to obtain original time series data from a plurality of subsystems in the energy storage box-type substation. The original time series data specifically includes: each battery monomer voltage, monomer temperature, battery cluster total voltage, total current, state of charge (SOC) collected by the BMS; AC side voltage, current, active power, reactive power, and DC side voltage, current collected by the PCS; winding temperature, oil level collected by the transformer subsystem; and physical quantities such as cabinet internal environment temperature and humidity collected by the environment control unit.

[0026] The multi-modal data acquisition and preprocessing module performs a series of processing on the obtained original time series data to generate a synchronized and standardized multi-element time series data set. In one embodiment, the processing specifically includes the following steps: Timestamp alignment: Since there are differences between the data reporting period of each subsystem and the internal clock, the multimodal data acquisition and preprocessing module establishes a unified time reference. For out-of-sync raw time series data, interpolation methods (such as linear interpolation) or forward / backward filling methods can be used to resample them to a fixed time frequency (such as once per second), forming a time-aligned data matrix.

[0027] Missing value processing: For missing values that may exist in the data matrix after timestamp alignment, a pre-set filling strategy is used for processing. The specific filling method for missing values.

[0028] Data normalization: To eliminate the influence of different physical dimensions on subsequent causal calculations, the multimodal data acquisition and preprocessing module normalizes each time series Normalization is performed. In one embodiment, the Z-Score normalization method is used, whose calculation formula is: ; In the formula, is the normalized time series, is the value of sensor i at time t, is the historical mean value of sensor i, is the historical standard deviation of sensor i.

[0029] and The historical period is statistically obtained to maintain the stability of the system during operation.

[0030] After the above processing, the multimodal data acquisition and preprocessing module 10 finally outputs an N-dimensional, synchronized and normalized multivariate time series data set D(t), where N is the total number of selected sensors.

[0031] Macroscopic operating mode vector extraction module, which is connected with the multimodal data acquisition and preprocessing module, is used to generate a macroscopic operating mode vector quantifying the current physical operating condition based on pre-set operating condition characteristics in the multivariate time series data set; Specifically, the macroscopic operating mode vector extraction module in this embodiment selects pre-set operating condition characteristics from the multivariate time series data set received from the multimodal data acquisition and preprocessing module, and generates a macroscopic operating mode vector quantifying the current physical operating condition by processing the pre-set operating condition characteristics.

[0032] Pre-set working condition features are used to represent the overall operation state and energy exchange mode of the energy storage box-type substation at a specific time. In an embodiment, the pre-set working condition features can be divided into continuous working condition features and discrete working condition features.

[0033] The continuous working condition features can specifically include the total active power P(t) and the total reactive power Q(t) reported by the PCS. The total active power P(t) reflects whether the system is currently in a charging state or a discharging state and the power size thereof, and the total reactive power Q(t) reflects the reactive power support provided by the system.

[0034] The discrete working condition features can specifically include the system operation state flag M(t) provided by the energy management system or the BMS. The operation state flag M(t) explicitly indicates the current pre-set working mode of the system, for example, the values 0, 1 and 2 of the operation state flag M(t) can correspond to the standby mode, the constant-current charging mode and the constant-voltage discharging mode, respectively.

[0035] The macroscopic operation mode vector extraction module performs feature engineering processing on the selected pre-set working condition features to construct a macroscopic operation mode vector . The feature engineering processing specifically includes; For the continuous working condition features (for example, P(t) and Q(t)), normalization processing is performed. This processing aims to eliminate the dimensional differences of different physical quantities and scale the values thereof to a unified interval. In an embodiment, the min-max normalization method can be adopted: ; In the formula, is the normalized active power value, P(t) is the real-time active power value at t, and are the rated maximum output power and the rated maximum input power of the energy storage box-type substation, respectively. Similar normalization processing is performed on the total reactive power Q(t) to obtain .

[0036] For the discrete working condition features (for example, M(t)), one-hot encoding processing is performed. This processing converts the classification values into a multi-dimensional binary vector to eliminate the ordinal relationship between different state values. If the operation state flag M(t) has K possible states, it is converted into a K-dimensional vector . For example, if K = 3 and the states are standby, charging and discharging, respectively, when M(t) is in the charging state, can be expressed as .

[0037] After the above processing is completed, the macroscopic operation mode vector extraction module splices all the processed features to construct the final d-dimensional macroscopic operation mode vector ; ; In the formula, denotes the macro-operation mode vector at time t, denotes the active power normalized eigenvalue at time t, denotes the reactive power normalized eigenvalue at time t, denotes the multi-dimensional operation eigen vector at time t, denotes the transposition operation of a matrix or a vector.

[0038] The macro-operation mode vector uniquely represents the coordinates of the system in the d-dimensional operating space at time t. The macro-operation mode vector extraction module will generate the macro-operation mode vector output.

[0039] A real-time causal calculation engine module connected with the multi-modal data acquisition and preprocessing module, for calculating the directional information flow strength between each time series in the data set based on the multi-dimensional time series data set, to generate a real-time transfer entropy matrix; Specifically, the real-time causal calculation engine module of the embodiment receives the synchronized and normalized multi-dimensional time series data set D(t), and calculates the directional information flow strength between any two time series and in the data set based on the multi-dimensional time series data set, to generate a real-time transfer entropy matrix. In one embodiment, the calculation of the directional information flow strength uses the transfer entropy algorithm. The transfer entropy is a nonlinear, asymmetric information theory measure, and its physical meaning is that it quantifies the additional amount of information that the historical information of the source time series can provide to the state of at a future time under the condition that the historical information of the target time series is known. If the value of is significantly greater than zero, it indicates that there is a directional information flow from to .

[0040] The specific formula for calculating by the real-time causal calculation engine module is as follows: ; In the formula, denotes the future state of the time series at time , denotes the k-order historical state vector of the time series at time t, denotes the k-order historical state vector of the time series at time , Order-order historical state vector Represents conditional probability. express , and The joint probability distribution, Indicates that in a given past state and Under the past state conditions, The conditional probability, Indicates only when given In its past state, The conditional probability, Represents a probability function or a conditional probability function. Let k represent the logarithm to the base 2, and k represent the time series. Historical order Representing time series Historical order Representing time series Time delay, Representing time series Time delay, Represents different variables or nodes in the system, where i represents the target variable, j represents the source variable, and t represents the time index.

[0041] The historical order k, and time delay , These are key parameters for calculating transfer entropy. In one embodiment, these parameters are predetermined based on the physical characteristics of the energy storage box-type substation and the data sampling frequency.

[0042] To achieve real-time computation, the real-time causal computation engine module performs the transfer entropy calculation over a sliding time window. The sliding window has a preset time length L. In each analysis cycle (e.g., the sliding window slides forward one step), the module acquires the multivariate time series data within the current window and calculates all... The transfer entropy value between time series pairs.

[0043] After the calculation is completed, the real-time causal computation engine module outputs a... Real-time transfer entropy matrix of dimension , where the matrix's first Line number Column elements The value is The real-time transfer entropy matrix The actual causal topology that is transmitted to build the system.

[0044] a real-time dynamic causal entropy graph construction module, connected with the real-time causal computation engine module, for converting the real-time transfer entropy matrix into a real-time dynamic causal entropy graph representing the current actual information flow topology of the system; In particular, the real-time dynamic causal entropy graph construction module of the present embodiment receives the real-time transfer entropy matrix output by the real-time causal computation engine module , and based on the real-time transfer entropy matrix , converts it into a real-time dynamic causal entropy graph representing the current actual information flow topology of the system .

[0045] In one embodiment, the real-time dynamic causal entropy graph is a weighted directed graph, which is defined as: ; wherein, represents the real-time dynamic causal entropy graph at time t, v represents the node set in the graph, represents the directed edge set at time t, represents the weight set at time t, and t represents the time index.

[0046] The elements of the graph are defined as follows: Node set : The node set contains N nodes. The N nodes correspond one-to-one to the N sensors (i.e., N time series) in the multivariate time series data set, and each node represents a specific sensor monitoring point.

[0047] Edge set and weight set : The generation of the edge set and the weight set is accomplished by performing thresholding processing on the real-time transfer entropy matrix .

[0048] Specifically, for any element in the real-time transfer entropy matrix , whose value represents , the real-time dynamic causal entropy graph construction module compares it with a pre-set noise threshold .

[0049] If , it is determined that there is a significant directed information flow from node to node that exceeds the system noise level, and a directed edge from to is established in the edge set .directed edge . Meanwhile, a weight is assigned to the directed edge , whose value is equal to the transfer entropy value, i.e. wherein denotes the true or actual weight of the given indices j and i at time t, denotes a matrix containing transformation or adjustment coefficients related to time t, denotes the element in the jth row and the ith column of the matrix .

[0050] If , it is determined that the directed information flow from node to node is weak and is considered as random noise of the system, which is not statistically significant, and no directed edge is established between node and node .

[0051] The determination of the noise threshold has an important impact on the accuracy of subsequent topology analysis. In an embodiment, in order to improve the reliability of the graph, the noise threshold is not a global uniform fixed value, but a specific noise threshold is determined for each pair of time series .

[0052] The specific determination method of the specific noise threshold adopts a proxy data method. The specific steps of the proxy data method are as follows: The target time series is kept unchanged; the time order of the source time series is randomly permuted multiple times (for example, 1000 times) to generate a group of proxy time series . The random permutation process destroys the original time correlation between and , but retains the statistical properties of itself; For each group of proxy time series , the transfer entropy of the group to is calculated; this process is repeated to obtain a transfer entropy empirical distribution under the zero hypothesis that has no causal effect on ; The 95th percentile or the 99th percentile of the empirical distribution is taken as the specific noise threshold ​​.

[0053] Through the above steps, the real-time dynamic causal entropy graph construction module completes the construction of the real-time dynamic causal entropy graph , which is a sparse topological structure representing the current actual internal information flow of the system.

[0054] The working condition causal baseline dynamic mapping module is connected with the macro operation mode vector extraction module, and is used for receiving the macro operation mode vector and generating a theoretical health baseline graph corresponding to the current working condition according to the macro operation mode vector. Specifically, the working condition causal baseline dynamic mapping module receives the macro operation mode vector output by the macro operation mode vector extraction module, and generates a theoretical health baseline graph corresponding to the current working condition according to the macro operation mode vector.

[0055] Nonlinear mapping from continuous working condition space to theoretical health causal graph space. The running state of the energy storage box-type substation is dynamically changing, and the causal relationship strength and topological structure between its internal components will change accordingly. The traditional baseline modeling method often divides the complex running working condition into a limited and discrete mode, and constructs a fixed health baseline graph for each discrete mode.

[0056] The working condition causal baseline dynamic mapping module adopts a continuous mapping method, and directly maps the macro operation mode vector (representing a point in the continuous working condition space) to its corresponding theoretical health transfer entropy matrix through a pre-trained nonlinear mapping model . The theoretical health transfer entropy matrix is then converted into a theoretical health baseline graph . This continuous mapping method can realize real-time and dynamic adjustment of the system health baseline, so as to accurately match the current actual running working condition.

[0057] The mapping model can be expressed as: ; In the formula, is at time t, is a nonlinear mapping model, is the macro operation mode vector at the current time provided by the macro operation mode vector extraction module, and t is a time variable. The nonlinear mapping model is obtained through an offline training stage before the online running of the system. The offline training stage aims to construct the nonlinear mapping model ​This enables it to accurately learn and reproduce the complex mapping relationships from macroscopic operating conditions to internal causal graphs under healthy conditions. The offline training phase specifically includes steps such as training dataset construction, model selection, and training. The construction of the training dataset is fundamental to the offline training phase. This involves acquiring a large amount of historical time-series data on the energy storage substation operating in a confirmed healthy state. This confirmed health data can originate from factory testing, acceptance data from the early stable operation phase, or long-term operation records confirmed by experts to be fault-free. The historical time-series data should cover as wide a range of operating conditions as possible.

[0058] Historical time-series data is processed to generate paired input-output data. Specifically, within each time window... Using the same processing logic as the macro-operation mode vector extraction module, the macro-operation mode vector corresponding to this window is extracted from historical time series data. , as input features of the training dataset; Using the same processing logic as the real-time causal computation engine module, the health transit entropy matrix corresponding to this window is calculated. , which serves as the output label for the training dataset.

[0059] The above process is performed on all historical health data, ultimately constructing a training dataset containing a large number of paired samples. .

[0060] Nonlinear mapping model The selection of the model needs to be able to handle data from a low-dimensional vector (macroscopic operating mode vector). From a high-dimensional matrix (health transfer entropy matrix) to a high-dimensional matrix (a) nonlinear regression task.

[0061] In one embodiment, a nonlinear mapping model This can be achieved using deep neural networks. For example, a multilayer perceptron can be constructed, whose input layer neurons are related to the macroscopic operating mode vector. The dimension d matches; the number of neurons in its output layer matches the healthy transfer entropy matrix. The total number of elements in Matching.

[0062] Nonlinear mapping model The training process utilizes paired datasets. Minimize a pre-defined loss function through an optimization algorithm. Loss function The transfer entropy matrix used to measure model predictions With the real health transfer entropy matrix The difference between them. In one embodiment, the loss function The mean squared error function can be used, and its form can be based on the definition of the F-norm of the matrix: ; In the formula, This represents the value of the loss function. Represents the training dataset. This represents the total number of samples in the training dataset. Represents the macroscopic operating mode vector. Represents the true health delivery matrix. This represents the output matrix of the mapping model.

[0063] The obtained nonlinear mapping model It is fixed and deployed in the dynamic mapping module of causal baseline of operating conditions.

[0064] During the online operation phase of the energy storage box-type substation, the operating condition causal baseline dynamic mapping module utilizes the nonlinear mapping model obtained during the offline training phase. It performs real-time generation of theoretical health baseline maps.

[0065] At any online analysis time t, the dynamic mapping module for the causal baseline of the operating condition first receives the macroscopic operating mode vector of the current time from the macroscopic operating mode vector extraction module. The dynamic mapping module for the causal baseline of operating conditions will map the macroscopic operating mode vector. As input, it is fed into a pre-trained nonlinear mapping model. In the process, a forward propagation calculation is performed. The output of the forward propagation calculation is the theoretical health propagation entropy matrix. .

[0066] In obtaining the theoretical health transport entropy matrix Next, the dynamic mapping module for causal baselines under operating conditions needs to convert it into a theoretical health baseline map. In one embodiment, the method used in the transformation process, along with the real-time dynamic causal entropy graph construction module, will transmit the entropy matrix in real time. Transformed into a real-time dynamic causal entropy map The methods used are completely identical to ensure that the two are comparable.

[0067] The transformation process is as follows: the dynamic mapping module for the causal baseline of the operating condition adopts the same noise threshold as the real-time dynamic causal entropy map construction module (e.g., a specific noise threshold determined through a proxy data method). ), for the theoretical health transfer entropy matrix Each element in Perform thresholding.

[0068] like Then in the theoretical health baseline map Establish a path from node point to Given directed edges and assigning them weights. .like If the condition is met, then no directed edge is established.

[0069] Through the above steps, the working condition causal baseline dynamic mapping module finally generates a baseline with respect to the current time. Real-time operating conditions Precisely Corresponding Theoretical Health Baseline Map .

[0070] The topology difference and anomaly detection module, together with the real-time dynamic causal entropy map construction module and the working condition causal baseline dynamic mapping module, is used to receive the real-time dynamic causal entropy map and the theoretical health baseline map, calculate the topological differences between the two to generate a difference map, and detect whether there is a topological anomaly based on the difference map. Specifically, this embodiment receives the real-time dynamic causal entropy map output by the real-time dynamic causal entropy map construction module. And the theoretical health baseline map output by the working condition causal baseline dynamic mapping module, which precisely corresponds to the current working condition. .

[0071] The graph topology difference and anomaly detection module calculates the topological differences between the two received graphs to generate a difference graph. The difference map Used to quantify the degree and pattern by which the actual information flow topology of the current system deviates from its healthy baseline.

[0072] In one embodiment, the difference map Similarly, it is a weighted directed graph, defined as follows: ; Node set The difference spectrum node set With the real-time dynamic causal entropy map and the theoretical health baseline map The node sets are completely identical, both representing Nodes of a sensor .

[0073] Edge set With weight set The difference spectrum The topology and weights are determined by the topology and weights of the topology. and The weight matrix is ​​obtained by difference calculation.

[0074] Specifically, the difference map Any two nodes and directed edge weights between The calculation is as follows: ; In the formula, For difference spectrum From the node point to The weight of the edge. For real-time dynamic causal entropy map From the node point to The weight of the edge. If If the directed edge does not exist in the given path, then... The value is 0. For theoretical health baseline map From the node point to The weight of the edge, if If the directed edge does not exist in the given path, then... The value is 0.

[0075] Based on the above calculations, the difference spectrum... weight The patterns of topological deviation were accurately characterized: like and This indicates that a new edge has appeared in the real-time graph that did not exist in the healthy baseline, with a weight of . .

[0076] like and This represents an edge that exists in the healthy baseline but disappears in the real-time graph, with a difference weight of . .

[0077] like and This indicates that the actual information flow intensity is enhanced when an edge exists within a healthy baseline.

[0078] like and This indicates that the actual information flow intensity is weakened when there is an edge present in a healthy baseline.

[0079] like This indicates that the causal path remains unchanged.

[0080] Difference map (It can be obtained from its difference weight matrix) Once generated, the (representation) will be used for subsequent topological anomaly detection. This is done after obtaining the difference map. (or its corresponding difference weight matrix) After that, the topological difference and anomaly detection module analyzes the difference map according to a series of preset topological anomaly judgment rules. The system is analyzed to determine whether it is currently deviating from a healthy state.

[0081] In one embodiment, the topology anomaly determination rules specifically include the following types: Determining the causal boundary between new and disappearance; The judgment rules focus on the real-time dynamic causal entropy graph. It appears in the theoretical health baseline map, but in the theoretical health baseline map. Strong causal relationships that should not exist in the baseline, or conversely, key causal relationships that should exist in the baseline of health, disappear in the real-time graph.

[0082] New strong causal edges: if difference graph There is an edge in Its theoretical health baseline map The corresponding weights =0 (or below a specific noise threshold) However, in the real-time dynamic causal entropy map... The corresponding weights Significantly greater than 0, i.e. .in This is a preset threshold for determining newly created edges. This situation indicates that an abnormal information flow path has occurred within the system.

[0083] Key causal edges disappear: if in the theoretical health baseline map There is a weight The key edge, but in the real-time dynamic causal entropy graph Its weight Approaching 0. At this point, the difference plot... The corresponding weights And its absolute value is greater than the preset vanishing edge determination threshold. This situation indicates that the system's critical information transmission path may have been interrupted.

[0084] The decision-making rules focus on those in the real-time dynamic causal entropy graph. and theoretical health baseline map Causal edges exist in all cases, but their actual weights are different. With theoretical weight A significant deviation occurred between them.

[0085] Absolute value deviation: difference plot absolute value of the weight of the middle edge Exceeding a preset absolute deviation threshold .

[0086] Determination of causal reversal; The judgment rules focus on fundamental changes in the direction of information flow between two nodes. For example, in the theoretical health baseline map... In, there exists from arrive strong information flow (i.e.) (Very large) and reverse information flow Very weak; but in real-time dynamic causal entropy graphs In the middle, the situation reversed from arrive Information flow It becomes very weak, while the reverse information flow It became very strong.

[0087] This situation is observed in the difference graph. The middle is manifested as It is a significantly negative value, and It is a significantly positive value.

[0088] The key node topology attribute deviation judgment rule starts from the node level and focuses on whether the overall information flow input or output of the key node has changed significantly.

[0089] Weighted in-degree deviation: Calculating the real-time dynamic causal entropy map Middle node Weighted in-degree and theoretical health baseline map Theoretical weighted in-degree .

[0090] If the absolute value of the difference between the two Exceeding the node Preset in-degree deviation threshold .

[0091] Weighted out-degree deviation: Similarly, calculate the absolute value of the difference between the weighted out-degrees.

[0092] and deviates from the preset out-degree threshold. Comparison. This situation indicates that the role and status of key nodes in the information flow network have undergone an abnormal change.

[0093] The topology difference and anomaly detection module performs analysis on the difference map at each analysis time t, based on one or more of the aforementioned topology anomaly determination rules. The system will be checked. If any rule in the system is triggered, the module will determine that the system is in an abnormal state and generate an abnormal alarm signal.

[0094] The topology fingerprint and fault diagnosis module is connected to the topology differential and anomaly detection module. When a topology anomaly is detected, it analyzes the differential map to extract the topology fingerprint vector and matches the topology fingerprint vector with a preset fault knowledge base to output the fault diagnosis conclusion of the energy storage box substation. Specifically, in this embodiment, the topology fingerprinting and fault diagnosis module receives the anomaly alarm signal output by the topology map differential and anomaly detection module, as well as the generated differential map. .

[0095] Disruption modes in difference maps Specific topological features will be left behind. Difference map It not only reflects the degree to which the system deviates from the healthy baseline, but more importantly, its topology (i.e. the distribution pattern of abnormal information flow) reveals the propagation path and nature of the fault.

[0096] Difference map The topological features may include: The complete failure of the sensor may occur in the differential spectrum. This manifests as follows: the weights of all causal edges pointing to the sensor node are significantly weakened or become negative (input information is lost), and the weights of all causal edges originating from the node are also significantly weakened or become negative (output information is invalid).

[0097] Thermal runaway of the battery cell may occur in the differential plot. In this context, it is represented by the nodes that indicate the temperature of the unit. , and the node representing the temperature of its adjacent unit Between them, a line appears on the theoretical health baseline map. Newly generated strong causal edges that do not exist in the middle (i.e.) Significantly positive).

[0098] A logic error in the controller may manifest as a reversal of cause and effect, for example, an error that should have occurred on the BMS master node. Affected contactor node Its difference spectrum The middle is manifested as Significantly negative, while Significantly positive.

[0099] Difference map The complete topology structure includes which nodes (sensors) are the source of abnormal information (weighted out-degree is significantly positive), which nodes are the convergence points of abnormal information (weighted in-degree is significantly positive), and which specific newly formed edges, disappearing edges, or edges with significantly deviated weights are activated. Together, these constitute a high-dimensional, highly discriminative fault mode representation, namely, topological fingerprint.

[0100] The function of the topology fingerprinting and fault diagnosis module is to analyze the differential map. The topological characteristics are matched or classified with a pre-established fault topology fingerprint database, thereby enabling the transition from detecting anomalies to diagnosing specific fault types, providing a basis for subsequent fault location and decision-making.

[0101] The differential map is received by the topology fingerprinting and fault diagnosis module. Subsequently, in order to use it in subsequent diagnostic models, it is necessary to extract it from the difference map. This involves extracting a set of quantitative features that can efficiently characterize the topological fingerprint properties of the graph. Topological feature extraction aims to transform high-dimensional, complex graph structure data into a fixed-dimensional topological feature vector, which can then be used for pattern matching or input into classification models.

[0102] The construction of topological eigenvectors is achieved by analyzing the difference map at different scales. This is achieved through the following levels: Node-level topological features are used to describe the difference graph. The local deviation properties of each independent node (i.e., sensor).

[0103] A key node-level feature is the weighted in / out degree deviation. For difference maps... any node in ; Its weighted in-degree deviation Calculated as all points pointing to this node The sum of the weights of the edges: ; In the formula, Represents a node The weighted in-degree deviation, Let represent the k-th node in the network, t represent the time variable used to characterize the state of the system or network topology as it changes over time, and N represent the total number of nodes in the network, i.e., the number of nodes contained in the topology graph. Representing the difference plot From the node Pointing to node The difference in edge weights.

[0104] Its weighted out-degree deviation Calculated as all from that node Sum of the weights of the starting edges: ; In the formula, Represents a node The weighted deviation of the degree. Represents a node In a graph or network, t represents the time variable and is used in this formula to describe the nodes. The state change at a certain moment, where N represents the total number of nodes in the graph, i.e., the number of nodes in the network. Representing the difference plot From the node Pointing to node The difference in edge weights.

[0105] All The weighted in-degree deviation and weighted out-degree deviation of each node can be combined to form a topological feature vector. Subvectors of dimension.

[0106] Graph-level topological features are used to describe changes in the overall structure of the system's information flow network.

[0107] The key graph-level feature is the change in network centrality of key nodes. Network centrality measures the importance or influence of a node in the graph. The topology fingerprinting and fault diagnosis module can calculate a real-time dynamic causal entropy graph. key nodes Central indicators and its role in the theoretical health baseline map The corresponding centrality index .

[0108] The centrality metric C can be PageRank centrality or betweenness centrality.

[0109] Topological features That is, it is calculated as the difference between the two: ; In the formula, Represents a node At any moment The deviation of the indicator. Represents a node The actual observed index at time t Represents a node The theoretical benchmark index at time t; Let t represent the k-th node in the network or system, and t represent the time variable.

[0110] Significant changes indicate key nodes. Its role and status in the system's information exchange network have undergone a fundamental transformation.

[0111] Subgraph-level topological features focus on identifying and counting differential maps. Local patterns that appear in the process and have specific fault diagnosis significance; Chain propagation motif: Identification in differential maps Does a path exist in the path consisting of multiple significantly positive weighted edges?

[0112] Fan-out phantom: Identification in difference maps Does there exist a node in it? It simultaneously sends to multiple other nodes Edges with significant positive weights. Fan-out phantoms typically indicate nodes. It is the source of a malfunction and simultaneously has an abnormal impact on multiple downstream components.

[0113] Anomaly feedback closed-loop phantom: identification in differential mapping Has a new closed loop been formed in the middle?

[0114] The topological fingerprinting and fault diagnosis module analyzes the differential map. A motif detection algorithm is used to count the occurrence frequency or total weight of each specific motif and incorporate it into the topological feature vector as a subgraph-level topological feature.

[0115] From the difference spectrum After extracting node-level, graph-level, and subgraph-level topological features, the topological fingerprint and fault diagnosis module combines these quantized features extracted at different scales to construct a topological fingerprint vector. .

[0116] In one embodiment, the combination process is accomplished through vector concatenation. The topology fingerprinting and fault diagnosis module concatenates sub-vectors composed of node-level topology features, graph-level topology features, and sub-graph-level topology features in a pre-defined fixed order.

[0117] Topological fingerprint vector This can be formally represented as: ; In the formula, In order to be in The topological fingerprint vector constructed at each moment, These are node-level topological feature vectors. These are graph-level topological feature vectors. These are sub-vectors representing topological features at the subgraph level. This indicates a vector concatenation operation.

[0118] Difference map The complex topological structure information is encoded into a fixed-dimensional topological fingerprint vector. Topological fingerprint vector This will then serve as a standardized input to the diagnostic unit in the topology fingerprinting and fault diagnosis module, used to perform subsequent fault mode matching or classification tasks.

[0119] In topological fingerprint vector After being constructed, the topology fingerprint and fault diagnosis module utilizes a pre-built fault knowledge base to analyze the topology fingerprint vector. Analysis is performed to match and determine the final fault diagnosis conclusion. This process is divided into two stages: offline construction of the fault knowledge base and online matching and diagnosis.

[0120] Offline Construction of the Fault Knowledge Base: The construction of the fault knowledge base is the foundation of the online diagnostic system. Its purpose is to establish a mapping relationship between the known typical fault modes of energy storage prefabricated substations and their corresponding typical topological fingerprint vectors. The data for constructing the fault knowledge base can be derived from historical fault data or fault injection simulation data.

[0121] For each known fault type Multiple sets of fault time series data are collected. Each set of fault time series data is input into the complete processing flow of this invention, ultimately generating a fault type-specific data. A set of topological fingerprint vectors. Used to establish fault types. By establishing the correspondence between vectors and typical topological fingerprints, the statistical representation of all vectors in the set can be calculated. For example, by calculating the mean vector, the fault type can be obtained. Typical topological fingerprint vector .

[0122] All known All fault types are handled using the methods described above, resulting in the final fault knowledge base. It can be represented as a containing A collection of entries; In the formula, This represents a fault knowledge base. Indicates the first Labels for different types of faults Indicates the first Topological reference feature matrix for each type of fault.

[0123] Fault Knowledge Base It is stored in the topology fingerprint and fault diagnosis module; Online matching and diagnosis; During the online operation phase of the energy storage box-type substation, once the topology differential and anomaly detection module detects an anomaly and triggers the topology fingerprint and fault diagnosis module to generate a real-time topology fingerprint vector. The online matching and diagnostic process is then executed.

[0124] In one embodiment, online matching diagnostics is achieved by calculating a real-time topological fingerprint vector. With fault knowledge base Each typical topological fingerprint vector stored in It is achieved through the similarity between them. Cosine similarity can be used for calculation: In the formula, Indicates at time System status and the Similarity of the various failure modes Represents the real-time topological fingerprint vector. Indicates the first Typical topological fingerprint vectors for various fault types Representing vectors The Euclidean norm, Represents the reference vector The Euclidean norm, Represents a time variable. This represents the fault category index. The topology fingerprint and fault diagnosis module calculates... With all The similarity of each reference vector is calculated, and the fault index with the highest similarity is identified. .

[0125] make The maximum similarity value. The maximum similarity value... With a preset diagnostic confidence threshold Comparison: like : Indicates the real-time topological fingerprint vector With the first in the fault knowledge base The topological fingerprint and the diagnostic conclusions output by the fault diagnosis module are highly matched to the typical fault modes, and these correspond to the fault labels. .

[0126] like This indicates that the topological fingerprint does not match any of the known typical failure modes. This situation is diagnosed as an unknown failure.

Claims

1. A fault prediction and early warning system for energy storage box-type substations, characterized in that, include; The multimodal data acquisition and preprocessing module is used to acquire raw time series data and process it to output a synchronized and normalized multivariate time series dataset. The macroscopic operation mode vector extraction module, which is connected to the multimodal data acquisition and preprocessing module, is used to generate macroscopic operation mode vectors that quantify the current physical operation conditions based on multivariate time series datasets. The real-time causal computation engine module, which is connected to the multimodal data acquisition and preprocessing module, is used to calculate and generate a real-time transfer entropy matrix based on a multivariate time series dataset. The real-time dynamic causal entropy graph construction module, which is connected to the real-time causal calculation engine module, is used to transform the real-time transfer entropy matrix into a real-time dynamic causal entropy graph. The dynamic mapping module for causal baselines of operating conditions is connected to the macroscopic operation mode vector extraction module. It is used to generate a theoretical health baseline map corresponding to the current operating conditions based on the macroscopic operation mode vector. The topology difference and anomaly detection module is connected to the real-time dynamic causal entropy map construction module and the working condition causal baseline dynamic mapping module. It is used to receive the real-time dynamic causal entropy map and the theoretical health baseline map, calculate the topological differences between them to generate a difference map, and detect the existence of topological anomalies based on the difference map. The topology fingerprint and fault diagnosis module, which is connected to the topology differential and anomaly detection module, is used to analyze the differential map to extract the topology fingerprint vector when the topology anomaly is detected, and match the topology fingerprint vector with a preset fault knowledge base to output the fault diagnosis conclusion of the energy storage box substation.

2. The energy storage box-type substation fault prediction and early warning system according to claim 1, characterized in that, The dynamic mapping module for causal baselines under operating conditions includes: Using a pre-trained nonlinear mapping model, the macroscopic operating mode vector is received as input, and the theoretical health transfer entropy matrix is ​​output. Based on the theoretical health transfer entropy matrix, noise thresholding is applied to generate a theoretical health baseline map.

3. The energy storage box-type substation fault prediction and early warning system according to claim 2, characterized in that, The nonlinear mapping model was obtained during the offline training phase by training a training dataset containing paired macroscopic operating mode vectors and a health transfer entropy matrix.

4. The energy storage box-type substation fault prediction and early warning system according to claim 1, characterized in that, The real-time dynamic causal entropy map construction module includes: A specific noise threshold is applied to the elements in the real-time transfer entropy matrix for comparison; When an element is greater than the specific noise threshold, a corresponding weighted directed edge is established in the real-time dynamic causal entropy graph.

5. The energy storage box-type substation fault prediction and early warning system according to claim 4, characterized in that, The specific noise threshold is determined individually for each pair of time series using a proxy data method.

6. The energy storage box-type substation fault prediction and early warning system according to claim 1, characterized in that, The map topology difference and anomaly detection module includes: The difference map is generated by calculating the difference between the weight matrix of the real-time dynamic causal entropy map and the theoretical health baseline map. The difference map is then examined according to the topological anomaly determination rules to detect topological anomalies.

7. The energy storage box-type substation fault prediction and early warning system according to claim 1, characterized in that, The topology fingerprinting and fault diagnosis module includes: When the topological anomaly is detected, a topological fingerprint vector is extracted from the differential map. The similarity between the topological fingerprint vector and the typical topological fingerprint vectors pre-stored in the fault knowledge base is calculated to match and output the fault diagnosis conclusion.

8. The energy storage box-type substation fault prediction and early warning system according to claim 7, characterized in that, The topological fingerprint and fault diagnosis module extracts the topological fingerprint vector by extracting and combining at least one of the following features: Node-level topological features, including weighted in-degree deviation or weighted out-degree deviation; Graph-level topological features, including network centrality changes of key nodes; Subgraph-level topological features, which include the count of specific network motifs.

9. The energy storage box-type substation fault prediction and early warning system according to claim 7, characterized in that, The topological fingerprint and the fault diagnosis module obtain the following after performing similarity calculation: The maximum similarity value is compared with a preset diagnostic confidence threshold; If the maximum similarity value is greater than the diagnostic confidence threshold, the corresponding fault label is output. If the maximum similarity value is less than or equal to the diagnostic confidence threshold, the output diagnostic conclusion is an unknown fault.

10. The energy storage box-type substation fault prediction and early warning system according to claim 1, characterized in that, The macroscopic operation mode vector extraction module includes: Normalization processing is performed on the continuous operating condition features in the preset operating condition features; One-hot encoding is performed on the discrete operating condition features in the preset operating condition features; The processed continuous operating condition features are then concatenated with the discrete operating condition features to construct the macroscopic operating mode vector.

Images