Fault prediction method and system of dedicated metering acquisition terminal fusing demand load characteristics, and medium

Abstract

The application discloses a fault prediction method and system of a dedicated terminal for collecting load characteristics, and a medium, and relates to the field of power system fault prediction; each physical unit and load fluctuation characteristics of a target dedicated terminal for collecting are taken as graph nodes to construct a causal graph, and a causal graph attention network is constructed based on the causal graph; basic operation data and prior causal strength are input into the causal graph attention network for message transmission to obtain fault probabilities and fault propagation paths of each physical unit; information is transmitted in the causal direction to improve the accuracy of causal inference and avoid learning false correlations; by introducing prior causal strength and counterfactual reasoning, direct load fluctuation interference and indirect load fluctuation interference can be distinguished, communication abnormalities caused by user-side impact loads are avoided from being misjudged as terminal faults, and invalid replacement is reduced.

Description

Technical Field

[0001] This invention relates to the field of power system fault prediction, specifically to a fault prediction method, system, and medium for dedicated transformer data acquisition terminals that integrate demand load characteristics. Background Technology

[0002] Dedicated transformer data acquisition terminals are core metering and monitoring devices specifically deployed on transformers of 10kV and above in power supply or distribution circuit devices or systems for large industrial users. They undertake multiple functions such as electricity consumption monitoring, remote fee control, demand metering, and power quality analysis. With the deepening of power market reform and the increasing demands for demand-side response, the precise sensing and real-time control capabilities of dedicated transformer data acquisition terminals have become increasingly important.

[0003] However, large industrial users often experience significant load fluctuations during production processes, such as the start-up and shutdown of large motors, electric furnace smelting, and the periodic loads of rolling mills. These operations cause a sharp increase in the rate of change of current within a demand billing cycle (typically 15 minutes), resulting in noticeable power ripple in the power module of the dedicated transformer data acquisition terminal. Simultaneously, this induces broadband electromagnetic interference through conduction and radiation paths. This unstable and highly interfering electromagnetic environment directly impacts the communication stability of the dedicated transformer data acquisition terminal (e.g., increased bit error rate in 4G and carrier communication) and the reliable execution of control commands (e.g., malfunctions or delayed responses in remote tripping and closing).

[0004] Currently, fault prediction for dedicated transformer data acquisition terminals mainly relies on traditional methods, such as status monitoring based on fixed thresholds, trend analysis of single electrical parameters, or simple historical fault statistical models. These methods typically assume a relatively stable power grid environment and predictable interference patterns. However, under complex operating conditions involving dynamic ripples caused by high current change rates and the superposition of random electromagnetic interference, traditional methods struggle to effectively distinguish interference signals from genuine early-stage fault characteristics. They are prone to misjudging transient interference as hardware degradation or permanent faults, leading to frequent false alarms and missed alarms. This not only reduces operation and maintenance efficiency but may also create unnecessary on-site inspections or risks of misoperation. Therefore, a fault prediction method for dedicated transformer data acquisition terminals that can adapt to highly dynamic and high-interference operating conditions is needed. Summary of the Invention

[0005] The technical problem this invention aims to solve is that, under complex operating conditions involving dynamic ripple and random electromagnetic interference superimposed by high current change rates in dedicated transformer data acquisition terminals, traditional fault prediction methods struggle to effectively distinguish interference signals from genuine prior fault characteristics. This often leads to misjudging transient interference as hardware degradation or permanent faults, resulting in frequent false alarms and missed alarms. The purpose of this invention is to provide a fault prediction method, system, and medium for dedicated transformer data acquisition terminals that integrates demand load characteristics. It constructs a causal graph using the physical units and load fluctuation characteristics of the target dedicated transformer data acquisition terminal as graph nodes. Based on this causal graph, a causal graph attention network is built. Basic operating data and prior causal strength are input into the causal graph attention network for message passing to obtain the fault probability and fault propagation path of each physical unit. Information is transmitted along the causal direction, improving the accuracy of causal inference and avoiding learning spurious correlations. By introducing prior causal strength and counterfactual reasoning, it can distinguish between direct and indirect load fluctuation interference paths, avoiding misjudging communication anomalies caused by user-side impulsive loads as terminal faults and reducing unnecessary replacements.

[0006] This invention is achieved through the following technical solution:

[0007] A fault prediction method for dedicated transformer data acquisition terminals that integrates demand and load characteristics, characterized in that the method includes:

[0008] The system collects load fluctuation data and basic operating data of each physical unit of the target transformer acquisition terminal during the demand metering cycle, and extracts original features and load fluctuation features from the basic operating data and load fluctuation data respectively; the physical unit includes: an uplink communication unit, a downlink communication unit, a load control loop unit, a power supply unit, and a clock battery unit;

[0009] A causal graph is constructed using each physical unit and load fluctuation characteristics as graph nodes. A causal relationship is defined for each edge and a priori causal strength is assigned.

[0010] A causal graph attention network is constructed based on a causal graph. Basic operational data and prior causal strength are input into the causal graph attention network for message passing to obtain the fault probability and fault propagation path of each physical unit. The causal graph attention network includes: a causal graph message passing module, a counterfactual reasoning module, and a fault probability output module. The causal graph message passing module generates node feature representations based on original features and attention coefficients fused with prior causal strength. The counterfactual reasoning module intervenes in load control loop unit nodes and load fluctuation feature nodes respectively to perform counterfactual reasoning to obtain differential features. The fault probability output module concatenates the differential features with the corresponding node feature representations to calculate the fault probability.

[0011] Operation and maintenance decisions are made based on the failure probability and failure propagation path of each physical unit, resulting in decisions such as replacing the SIM card or clock battery, maintaining user-side load, or replacing the dedicated transformer acquisition terminal.

[0012] A further optimized solution is as follows: when generating a strategy to replace the dedicated transformer acquisition terminal, the optimal replacement time is calculated and pushed to both the operation and maintenance side and the user side. After receiving confirmation of replacement from both the operation and maintenance side and the user side, a start command is issued to trigger the load control circuit anti-tripping device. The load control circuit anti-tripping device is installed between the target dedicated transformer acquisition terminal and the load circuit. Under normal conditions, the load control circuit unit connects the load circuit. After receiving the start command, the load control circuit anti-tripping device connects the load circuit.

[0013] A further optimized solution is that the method for extracting the original features includes:

[0014] Within the sampling period T, statistical characteristics are calculated for the operating indicators reflecting the quality of the uplink communication unit channel; where the sampling period T is an integer multiple of the demand period;

[0015] Within the sampling period T, statistical characteristics are calculated for the operating indicators that reflect the status of the downlink communication unit bus.

[0016] Within the sampling period T, statistical characteristics of the operating indicators of the hardware health of the reaction load control loop unit are calculated.

[0017] Within the sampling period T, statistical characteristics of the operating indicators of the aging degree of the reaction power supply unit are calculated.

[0018] Within the sampling period T, statistical characteristics of the operating indicators of the battery life state of the reaction clock battery cell are calculated.

[0019] The method for extracting load fluctuation characteristics includes: summarizing the load fluctuation data of all demand measurement cycles within the sampling period T to obtain load fluctuation characteristics.

[0020] A further optimized solution is that the method for defining causal influence relationships and assigning prior causal strength to each edge includes:

[0021] The directed edge L→C is defined as the state change of load fluctuation characteristics causing a causal effect on the load control loop unit. The causal relationship of L→C is that the load fluctuation characteristics lead to an increase or abnormality in the operation of the load control loop.

[0022] The directed edge L→P is defined as representing that the state changes of load fluctuation characteristics will causally affect the power supply unit, and the causal relationship of L→P is that load fluctuation characteristics accelerate the aging of the power supply unit.

[0023] A directed edge L→U is defined to represent that the state changes of load fluctuation characteristics will causally affect the uplink communication unit. The causal relationship of L→U is that load fluctuation characteristics directly interfere with uplink communication.

[0024] A directed edge P→U is defined to represent that the state change of the power supply unit will cause and affect the uplink communication unit. The causal relationship of P→U is that the power supply unit supplies power to the uplink communication unit.

[0025] A directed edge P→D is defined to represent that the state change of the power supply unit will causally affect the downlink communication unit. The causal relationship of P→D is that the power supply unit supplies power to the downlink communication unit.

[0026] A directed edge P→C is defined to represent that the state change of the power supply unit will cause and affect the load control loop unit. The causal relationship of P→C is that the power supply unit supplies power to the load control loop unit.

[0027] The directed edge C→U is defined as the state change of the load control loop causally affecting the uplink communication unit. The causal relationship of C→U is that the electromagnetic interference generated by the operation of the load control loop unit affects the uplink communication.

[0028] The directed edge C→D is defined as the state change of the load control loop causally affecting the downlink communication unit, and the causal relationship of C→D is that the electromagnetic interference generated by the operation of the load control loop affects the uplink communication.

[0029] Define a directed edge B→U to represent that the state of the clock battery cell will causally affect uplink communication. The causal relationship of B→U is that the clock battery affects the timestamp, which in turn affects uplink communication.

[0030] Based on domain expert knowledge, each directed edge i→j is assigned a prior causal strength c. ij ∈[0,1]; where i=L, P, C, B; j=C, P, U, D.

[0031] A further optimization is that the causal graph message passing module includes Q graph attention layers, and the l-th graph attention layer performs message passing according to the following formula:

[0032] ;

[0033] in, This represents the attention coefficient of the directed edge i→j in the m-th attention head of the l-th graph attention layer; Represent the set of incoming neighbors of node j; Let f(x) represent the learnable weight matrix of the m-th attention head in the l-th graph attention layer. This represents the input feature vector of node i in the l-th graph attention layer; σ represents the output feature vector of node j in the (l+1)th graph attention layer; σ() represents the activation function; M represents the number of attention heads in the graph attention layer.

[0034] A further optimized solution is that the method for obtaining the attention coefficients that fuse prior causal strength includes:

[0035] Attention coefficient α of directed edge i→j ij for:

[0036] ;

[0037] in, The attention coefficients represent the input features of nodes i and j in the current graph attention layer, and the attention coefficients incorporate the prior causal strength of the directed edge i→j. || represents the learnable weight matrix; || represents the vector concatenation operation; c represents the attention weight vector; ij c represents the prior causal strength of the directed edge i→j; kj β represents the prior causal strength of the directed edge k→j; β represents a learnable scalar parameter that controls the degree of influence of the prior causal strength on the final attention coefficient; LeakyReLU() represents the linear rectified activation function; exp() represents the exponential function. This represents the input features of node k in the current graph attention layer.

[0038] A further optimized solution is that the method for obtaining the differential features includes:

[0039] First-level counterfactual reasoning: Replace the characteristics of load fluctuation characteristic node L with stable load characteristics, and calculate the change in the failure probability of the uplink communication unit ΔpU(L);

[0040] Second-level counterfactual reasoning: The characteristic of the forced control loop unit node C is that it is inactive. Calculate the change in the uplink communication unit failure probability ΔpU(C);

[0041] A first threshold is set. If the change ΔpU(L) exceeds the first threshold and the change ΔpU(C) is less than half of the first threshold, it is determined to be a fault propagation path where load fluctuation characteristics directly interfere with uplink communication. If both the change ΔpU(L) and the change ΔpU(C) exceed the first threshold, it is determined to be a fault propagation path where load fluctuation characteristics indirectly interfere with uplink communication through the control loop unit.

[0042] A further optimized solution is that the method for generating the operation and maintenance decisions includes:

[0043] Preset a second threshold and two thresholds for binary classification: a high threshold and a low threshold.

[0044] If the clock battery cell failure probability P b If the fault probability of the clock battery is higher than the high threshold and the fault probability of the other units is lower than the low threshold, it is recommended to replace the clock battery.

[0045] If the uplink communication unit failure probability P u If the SIM card is replaced when the failure probability of the remaining units is less than the low threshold and the change ΔpU(L) is less than the second threshold;

[0046] If the uplink communication unit failure probability P u If the failure probability of the remaining units is less than the low threshold and the change ΔpU(L) is greater than or equal to the second threshold, then it is recommended to perform load maintenance on the user side.

[0047] Otherwise, it is recommended to replace the dedicated transformer data acquisition terminal.

[0048] This solution also provides a system for predicting faults in dedicated transformer data acquisition terminals by integrating demand and load characteristics. This system is used to implement the aforementioned method for predicting faults in dedicated transformer data acquisition terminals by integrating demand and load characteristics. The system includes:

[0049] The acquisition module is used to acquire load fluctuation data and basic operating data of each physical unit of the target transformer acquisition terminal during the demand metering cycle, and extract the original features and load fluctuation features from the basic operating data and load fluctuation data respectively; the physical unit includes: an uplink communication unit, a downlink communication unit, a load control loop unit, a power supply unit, and a clock battery unit.

[0050] The graph construction module uses each physical unit and load fluctuation characteristics as graph nodes to construct a causal graph, defines the causal influence relationship for each edge and assigns a priori causal strength;

[0051] The message passing module is used to construct a causal graph attention network based on the causal graph. Basic operational data and prior causal strength are input into the causal graph attention network for message passing to obtain the fault probability and fault propagation path of each physical unit. The causal graph attention network includes: a causal graph message passing module, a counterfactual reasoning module, and a fault probability output module. The causal graph message passing module generates node feature representations based on the original features and attention coefficients fused with prior causal strength. The counterfactual reasoning module intervenes in the load control loop unit nodes and load fluctuation feature nodes respectively to perform counterfactual reasoning to obtain differential features. The fault probability output module concatenates the differential features with the corresponding node feature representations to calculate the fault probability.

[0052] The decision module is used to make operation and maintenance decisions based on the failure probability and failure propagation path of each physical unit, and generate decisions such as replacing the SIM card or clock battery, user-side load operation and maintenance, or replacing the dedicated transformer acquisition terminal.

[0053] This solution also provides a computer-readable medium having a computer program stored thereon, which, when executed by a processor, enables the implementation of the above-described method for predicting faults in a dedicated transformer data acquisition terminal by integrating demand load characteristics.

[0054] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0055] 1. This invention provides a method, system, and medium for predicting faults in dedicated transformer data acquisition terminals by integrating demand and load characteristics. It constructs a causal graph knowledge graph using the physical units and load fluctuation characteristics of the target dedicated transformer data acquisition terminal as graph nodes. Based on the causal graph, a causal graph attention network is built. Basic operating data and prior causal strength are input into the causal graph attention network for message passing to obtain the fault probability and fault propagation path of each physical unit. Information is transmitted along the causal direction, improving the accuracy of causal inference and avoiding learning spurious correlations. By introducing prior causal strength and counterfactual reasoning, it can distinguish between direct and indirect load fluctuation interference paths, avoiding misjudging communication anomalies caused by user-side impact loads as terminal faults and reducing unnecessary replacements.

[0056] 2. This invention provides a method, system, and medium for predicting faults in dedicated transformer data acquisition terminals by integrating demand and load characteristics. Based on typical faults of the dedicated transformer data acquisition terminals, the terminals are divided into physical units that can characterize the faults, thereby enabling counterfactual intervention on different physical units. Furthermore, by integrating prior causal strength, the causal graph attention network prioritizes the transmission of information along the causal direction, while suppressing directed edges that are reversed or non-causal, thereby improving the accuracy of causal inference and avoiding learning spurious correlations.

[0057] 3. This invention provides a method, system, and medium for predicting faults in dedicated transformer data acquisition terminals by integrating demand load characteristics. It uses periodic statistics such as current change rate and number of step jumps as graph node features, enabling the model to predict faults from the load source, providing early warnings several days earlier than using only internal terminal indicators. By introducing prior causal strength and counterfactual reasoning, it can distinguish between two paths: direct interference from load fluctuations and indirect interference from load fluctuations. This avoids misjudging communication anomalies caused by user-side impulsive loads as terminal faults, reducing unnecessary replacements.

[0058] 4. This invention provides a method, system, and medium for predicting faults in dedicated transformer data acquisition terminals that integrate demand and load characteristics. It is designed to prevent tripping in normally closed load control circuits. Under normally closed load control circuit conditions, it can ensure timely replacement of terminals during data acquisition and maintenance, and also ensure that dedicated transformer customer equipment does not trip or lose power. This enables normal data acquisition and timely uploading of electricity to provide accurate data for meter reading and line loss assessment, thereby improving customer service satisfaction. Attached Figure Description

[0059] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0060] Figure 1 A schematic diagram of a fault prediction method for dedicated transformer data acquisition terminals that integrates demand load characteristics;

[0061] Figure 2 Structure diagram of a fault prediction system for dedicated transformer data acquisition terminals that integrates demand load characteristics;

[0062] Figure 3 This is a schematic diagram of the assembly of the anti-tripping device for the load control circuit.

[0063] The attached diagram shows the markings and corresponding component names:

[0064] 1-First crimping screw, 2-Second crimping screw, 3-Third crimping screw, 4-Fourth crimping screw, 5-Short connecting piece. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0066] Under the complex operating conditions of dynamic ripple and random electromagnetic interference caused by the high current change rate of the dedicated transformer acquisition terminal, traditional acquisition terminal fault prediction methods are difficult to effectively distinguish between interference signals and true early fault characteristics. They are prone to misjudging transient interference as hardware degradation or permanent faults, resulting in frequent false alarms and missed alarms. In view of this, this solution provides the following embodiments to solve the above technical problems.

[0067] Example 1

[0068] This embodiment provides a fault prediction method for dedicated transformer data acquisition terminals that integrates demand load characteristics, such as... Figure 1 As shown, the method includes:

[0069] Step 1: Collect load fluctuation data and basic operating data of each physical unit of the target transformer acquisition terminal within the demand metering cycle, and extract the original features and load fluctuation features from the basic operating data and load fluctuation data respectively.

[0070] In step one, the physical unit includes: an uplink communication unit, a downlink communication unit, a load control loop unit, a power supply unit, and a clock battery unit;

[0071] The maximum demand is calculated every 15 or 30 minutes using a dedicated transformer terminal. In this embodiment, the following load fluctuation data are recorded for each demand cycle: the maximum instantaneous current value is taken as the maximum current in the demand cycle, the minimum instantaneous current value is taken as the minimum current in the demand cycle, the maximum current change rate, the current fluctuation variance, the number of load steps, the time of exceeding the demand set value, the voltage sag depth, and the demand calculation value.

[0072] In step one, the methods for extracting the original features include:

[0073] Within the sampling period T, statistical characteristics are calculated for the operational indicators reflecting the quality of the uplink communication unit channel (specifically including: average RSSI, retransmission rate, connection success rate, average latency, and number of primary / backup handovers; these operational indicators are parameters built into the communication module and can be directly read from the communication management center); wherein, the sampling period T is an integer multiple of the demand period;

[0074] Within the sampling period T, statistical characteristics are calculated for the operating indicators reflecting the status of the downlink communication unit bus (specifically including: RS485 bus collision count, slave timeout rate, CRC error count, and bus voltage).

[0075] Within the sampling period T, statistical characteristics are calculated for the operating indicators of the hardware health of the response load control loop unit (specifically including: number of relay actions, average relay load current, and bit error rate increment after action).

[0076] Within the sampling period T, statistical characteristics are calculated for the operating indicators of the aging degree of the reaction power supply unit (specifically including: internal temperature, voltage ripple, number of resets, and input voltage).

[0077] Within the sampling period T, statistical characteristics are calculated for the operating indicators of the battery life state of the reaction clock battery cell (specifically including battery voltage and battery temperature).

[0078] Specifically, with a period T of one day, the statistical characteristics of each operating indicator are calculated, such as mean, maximum value, and variance.

[0079] In step one, the method for extracting load fluctuation characteristics includes: summarizing the load fluctuation data of all demand metering cycles within the sampling period T to obtain the load fluctuation characteristics. Specifically, each dedicated transformer acquisition terminal has 96 demand cycles (each demand cycle is 15 minutes) or 48 demand cycles (each demand cycle is 30 minutes) per day. This scheme statistically analyzes all demand cycles daily to obtain the daily load fluctuation characteristics (mean, maximum value, 95th percentile, trend slope, etc.).

[0080] Finally, the load fluctuation characteristics and the original characteristics are normalized using the Z-score method to eliminate the influence of dimensions.

[0081] Step two involves constructing a causal graph using each physical unit and load fluctuation characteristics as graph nodes. A causal relationship is defined for each edge, and a priori causal strength is assigned. Each node represents an observable or interveneable component. The nodes in the causal graph must be physical entities or external variables, not abstract faults, because faults are attributes of entities. This scheme divides the dedicated transformer data acquisition terminal into physical units that can characterize faults based on typical faults, thereby enabling counterfactual intervention on different physical units.

[0082] In step two, based on the physical connection and electromagnetic coupling relationship, directed edges are defined to represent possible causal influence directions. A causal influence relationship is defined for each edge, and a priori causal strength is assigned. Specifically, this includes:

[0083] The directed edge L→C is defined as the state change of load fluctuation characteristics that causally affects the load control loop unit (e.g., the probability of relay operation increases significantly after a load step). The causal relationship of L→C is that the load fluctuation characteristics lead to an increase or abnormality in the operation of the load control loop.

[0084] The directed edge L→P is defined as representing that the state changes of load fluctuation characteristics will cause and affect the power supply unit (such as sudden current changes causing fluctuations in the input voltage of the switching power supply, and stress on the capacitor). The causal relationship of L→P is that load fluctuation characteristics accelerate the aging of the power supply unit.

[0085] A directed edge L→U is defined to represent that the state changes of load fluctuation characteristics will causally affect the uplink communication unit (e.g., power quality monitoring shows that load fluctuations are related to the communication bit error rate). The causal relationship of L→U is that load fluctuation characteristics directly interfere with uplink communication.

[0086] A directed edge P→U is defined to represent that the state change of the power supply unit will cause a causal effect on the uplink communication unit (e.g., excessive ripple of the power supply unit will cause the uplink communication unit to malfunction). The causal relationship of P→U is that the power supply unit supplies power to the uplink communication unit.

[0087] A directed edge P→D is defined to represent that the state change of the power supply unit will cause a causal effect on the downlink communication unit (e.g., excessive ripple of the power supply unit will cause the downlink communication unit to malfunction). The causal relationship of P→D is that the power supply unit supplies power to the downlink communication unit.

[0088] A directed edge P→C is defined to represent that the state change of the power supply unit will cause and affect the load control circuit unit (such as insufficient drive voltage causing poor relay engagement). The causal relationship of P→C is that the power supply unit supplies power to the load control circuit unit.

[0089] The directed edge C→U is defined as follows: the state change of the load control loop will cause and affect the uplink communication unit (such as the transient pulse of hundreds of volts generated when the relay is activated, which interferes with communication through ground or space radiation). The causal relationship of C→U is that the electromagnetic interference generated by the operation of the load control loop unit affects the uplink communication.

[0090] The directed edge C→D is defined as the state change of the load control loop causally affecting the downlink communication unit, and the causal relationship of C→D is that the electromagnetic interference generated by the operation of the load control loop affects the uplink communication.

[0091] Define a directed edge B→U to represent that the state of the clock battery cell will causally affect uplink communication (such as time disorder causing the master station to reject data or encryption authentication failure). The causal relationship of B→U is that the clock battery affects the timestamp, which indirectly affects uplink communication.

[0092] Based on domain expert knowledge, each directed edge i→j is assigned a prior causal strength c. ij ∈[0,1]; where i=L, P, C, B; j=C, P, U, D.

[0093] Specifically, based on the expert's experience and knowledge, each directed edge i→j is assigned a value. For example, the prior causal strength of P→U is assigned a value of 0.9 (certain influence), and the prior causal strength of C→U is assigned a value of 0.7 (significant influence).

[0094] Step 3: Construct a causal graph attention network based on the causal graph. Input the basic operational data and prior causal strength into the causal graph attention network to obtain the failure probability and failure propagation path of each physical unit through message passing.

[0095] The specific causal graph attention network includes: a causal graph message passing module, a counterfactual reasoning module, and a fault probability output module;

[0096] The causal graph message passing module generates node feature representations for message passing based on the original features and the attention coefficients that fuse prior causal strength;

[0097] In step three, the causal graph message passing module includes Q graph attention layers, and the l-th graph attention layer performs message passing according to the following formula:

[0098] ;

[0099] in, The attention coefficient represents the directed edge i→j in the m-th attention head of the l-th graph attention layer, and the attention coefficient incorporates the prior causal strength of the directed edge i→j. Represent the set of incoming neighbors of node j; Let f(x) represent the learnable weight matrix of the m-th attention head in the l-th graph attention layer. This represents the input feature vector of node i in the l-th graph attention layer; σ represents the output feature vector of node j in the (l+1)th graph attention layer; σ() represents the activation function; M represents the number of attention heads in the graph attention layer.

[0100] Methods for obtaining attention coefficients that incorporate prior causal strength include:

[0101] Attention coefficient α of directed edge i→j ij for:

[0102] ;

[0103] in, This represents the input features of node i and node j in the current graph attention layer; || represents the learnable weight matrix; || represents the vector concatenation operation; c represents the attention weight vector; ij c represents the prior causal strength of the directed edge i→j; kj β represents the prior causal strength of the directed edge k→j; β represents a learnable scalar parameter that controls the degree of influence of the prior causal strength on the final attention coefficient; LeakyReLU() represents the linear rectified activation function; exp() represents the exponential function. This represents the input features of node k in the current graph attention layer.

[0104] Specifically, in this embodiment, the causal graph contains 6 nodes (L, U, D, C, P, B) and 9 directed edges, forming a small causal graph. The shortest path length between any two nodes does not exceed 3. Two graph attention layers are sufficient to propagate information from each node to its neighbors within a distance of 2, covering all causal paths while preserving the unique characteristics of each node, improving computational efficiency, and maintaining the model's lightweight and stability. Traditional graph attention networks rely solely on feature similarity, making them prone to learning spurious correlations (e.g., C and U increase simultaneously without direction). This scheme, by incorporating prior causal strength, prioritizes propagating information along causal directions (e.g., C→U) while suppressing reverse (U→C) or non-causal directed edges, thereby improving the accuracy of causal inference and avoiding learning spurious correlations. The graph attention layers in this scheme use a multi-head attention mechanism, allowing the model to learn multiple attention patterns from different feature subspaces, improving robustness and expressiveness.

[0105] The counterfactual reasoning module intervenes in the load control loop unit node and the load fluctuation characteristic node respectively to perform counterfactual reasoning to obtain the difference characteristics;

[0106] Specifically, methods for obtaining differential features include:

[0107] First-level counterfactual reasoning: Replace the characteristics of load fluctuation characteristic node L with stable load characteristics, and calculate the change in the failure probability of the uplink communication unit ΔpU(L);

[0108] Second-level counterfactual reasoning: The characteristic of the forced control loop unit node C is that it is inactive. Calculate the change in the uplink communication unit failure probability ΔpU(C);

[0109] A first threshold is set. If the change ΔpU(L) exceeds the first threshold and the change ΔpU(C) is less than half of the first threshold, it is determined to be a fault propagation path where load fluctuation characteristics directly interfere with uplink communication. If both the change ΔpU(L) and the change ΔpU(C) exceed the first threshold, it is determined to be a fault propagation path where load fluctuation characteristics indirectly interfere with uplink communication through the control loop unit.

[0110] Specifically, first-level and second-level counterfactual reasoning can be implemented by training a Conditional Variational Autoencoder (CVAE). The CVAE is fed the original features (factual features) and intervention features, and outputs counterfactual features corresponding to the intervention features. The counterfactual reasoning process infers causal relationships by asking "What would happen if a variable were intervened?", effectively distinguishing between direct and indirect paths. During training, the CVAE and the causal graph attention network can be jointly trained, with gradients simultaneously updating the parameters of both networks, optimizing the discriminative features to values ​​most conducive to the correct classification of node U. During inference, the discriminative features are concatenated with the corresponding node feature representations to obtain the final node feature representations used for probability calculation.

[0111] The fault probability output module concatenates the difference features with the corresponding node feature representations and then calculates the fault probability.

[0112] Specifically, after passing through two graph attention layers, each physical unit node obtains a node feature representation. The two differential feature change quantities ΔpU(L) and ΔpU(C) generated by the counterfactual reasoning module are calculated around the failure probability change of the uplink communication node U. Therefore, they are only concatenated with the final node representation of the uplink communication node U to correct the failure probability prediction of node U.

[0113] Then, each node is connected to an independent binary classification multilayer perceptron (MLP), ultimately obtaining the failure probability P of the five physical units. U P D P C P P P B .

[0114] If the user side experiences impact loads such as electric arc furnaces or rolling mills, harmonics and voltage dips can occur, directly coupling to the power supply or antenna of the terminal communication module. This can lead to increased bit error rate and connection interruptions. If misjudged as an indirect path (or directly treated as a terminal failure), maintenance personnel may repeatedly replace SIM cards or even the dedicated transformer data acquisition terminal. However, the impact on the user side persists, and the newly replaced dedicated transformer data acquisition terminal may quickly fail again. By learning differential characteristics through a counterfactual reasoning module, the direct path can be effectively identified, guiding the user side and implementing timely rectification measures such as installing filters to eliminate interference sources.

[0115] Step 4: Make operation and maintenance decisions based on the failure probability and failure propagation path of each physical unit, and generate options such as replacing the SIM card or clock battery, user-side load operation and maintenance, or replacing the dedicated transformer acquisition terminal.

[0116] The method for generating the operation and maintenance decisions includes:

[0117] Preset a second threshold and two thresholds for binary classification: a high threshold and a low threshold.

[0118] If the clock battery cell failure probability P B If the fault probability of the clock battery is higher than the high threshold and the fault probability of the other units is lower than the low threshold, it is recommended to replace the clock battery.

[0119] If the uplink communication unit failure probability P U If the SIM card is replaced when the failure probability of the remaining units is less than the low threshold and the change ΔpU(L) is less than the second threshold;

[0120] If the uplink communication unit failure probability P U If the failure probability of the remaining units is less than the low threshold and the change ΔpU(L) is greater than or equal to the second threshold, then it is recommended to perform load maintenance on the user side.

[0121] Otherwise, it is recommended to replace the dedicated transformer data acquisition terminal.

[0122] Specifically, in this embodiment, considering measurement errors and occasional voltage fluctuations, a high threshold of 0.6 is set to avoid false alarms for critical states. When recommending the replacement of the clock battery, the failure probability of the remaining units is required to be less than the low threshold of 0.4 to prevent misjudgment when battery undervoltage and other faults coexist. When recommending the replacement of the SIM card, the failure probability of the remaining units is required to be less than the low threshold of 0.4 to exclude the situation where other units fail at the same time.

[0123] Scenarios where it is recommended to replace the dedicated transformer data acquisition terminal include: downlink communication failure (the judgment rule can be set to failure probability > 0.5), control loop failure (the judgment rule can be set to failure probability > 0.5), power module failure (the judgment rule can be set to failure probability > 0.5), and multi-unit failure (failure probability > 0.5 for two or more communication units).

[0124] When generating a strategy to replace the dedicated transformer data acquisition terminal, the optimal replacement time is calculated and pushed to the operation and maintenance side and the user side. After receiving confirmation of replacement from the operation and maintenance side and the user side, a start command is issued to trigger the load control circuit anti-trip device. The load control circuit anti-trip device is installed between the target dedicated transformer data acquisition terminal and the load circuit. Under normal conditions, the load control circuit unit connects the load circuit. After receiving the start command, the load control circuit anti-trip device connects the load circuit.

[0125] Specifically, such as Figure 3 As shown, the load control circuit of the dedicated transformer data acquisition terminal includes normally open and normally closed types. When using a normally closed load control circuit, the load circuit is closed under normal user-side load conditions. When the user is in arrears, the internal circuits of the first port 1 and the second port 2 automatically disconnect, causing the user-side load switch to trip. If the dedicated transformer data acquisition terminal needs to be replaced under normal user-side load conditions, disconnecting the wiring of the first port 1 and the second port 2 will cause the load circuit to disconnect, leading to the tripping of the user's load switch. High-voltage dedicated transformer users, due to the nature of their loads, mostly cannot afford power outages; a sudden power outage would cause significant losses to customer production and even safety. Currently, due to the user-side... Due to factors such as equipment load, a relatively high proportion of normally closed load circuits are used. As the requirements for power collection become increasingly stringent, more and more data collection terminals need to be replaced and upgraded. To ensure that power collection is normal and timely when terminals are replaced while maintaining uninterrupted power supply for customers, this solution incorporates a load control circuit anti-tripping device and an interactive confirmation mechanism between the maintenance side and the user side. After the maintenance side and the user side have interactively confirmed the replacement time, a start command is sent to the load control circuit anti-tripping device to facilitate preparation by the user side. Upon receiving the start command, the load control circuit anti-tripping device is triggered and connects the load circuit.

[0126] The load control circuit anti-trip device includes a connecting piece 5, a first crimping screw 1, a second crimping screw 2, a third crimping screw 3, and a fourth crimping screw 4. The first crimping screw 1 is used to realize the electrical connection between the first end of the load circuit and the third crimping screw 3. The first crimping screw 2 is used to realize the electrical connection between the second end of the load circuit and the fourth crimping screw 4. The third crimping screw 3 is connected to the first end of the load control circuit. The fourth crimping screw 4 is connected to the second end of the load control circuit. The connecting piece 5 is movably connected and fixedly connected between the first crimping screw 1 and the second crimping screw 2. When the dedicated transformer data acquisition terminal is working normally, the movable connection between the connecting short piece and the first crimping screw 1 and the second crimping screw 2 is disconnected. The load control circuit of the target dedicated transformer data acquisition terminal is connected to the load circuit through the first crimping screw 1, the second crimping screw 2, the third crimping screw 3, and the fourth crimping screw 4. When the load control circuit anti-trip device receives the start command, it pushes the connecting short piece to connect the movable connection between the connecting short piece and the first crimping screw 1 and the second crimping screw 2. At this time, the load control circuit anti-trip device connects the load circuit. Removing the dedicated transformer data acquisition terminal will not cause a trip.

[0127] Example 2

[0128] This embodiment provides a fault prediction method system for dedicated transformer data acquisition terminals that integrates demand load characteristics, such as... Figure 2 As shown, the system for implementing the fault prediction method for dedicated transformer data acquisition terminals that integrates demand and load characteristics as described in Example 1 includes:

[0129] The acquisition module is used to acquire load fluctuation data and basic operating data of each physical unit of the target transformer acquisition terminal during the demand metering cycle, and to extract the original features and load fluctuation features from the basic operating data and load fluctuation data respectively; the physical unit includes: an uplink communication unit, a downlink communication unit, a load control loop unit, a power supply unit, and a clock battery unit.

[0130] The graph construction module uses each physical unit and load fluctuation characteristics as graph nodes to construct a causal graph, defines the causal influence relationship for each edge and assigns a priori causal strength;

[0131] The message passing module is used to construct a causal graph attention network based on the causal graph. Basic operational data and prior causal strength are input into the causal graph attention network for message passing to obtain the fault probability and fault propagation path of each physical unit. The causal graph attention network includes: a causal graph message passing module, a counterfactual reasoning module, and a fault probability output module. The causal graph message passing module generates node feature representations based on the original features and attention coefficients fused with prior causal strength. The counterfactual reasoning module intervenes in the load control loop unit nodes and load fluctuation feature nodes respectively to perform counterfactual reasoning to obtain differential features. The fault probability output module concatenates the differential features with the corresponding node feature representations to calculate the fault probability.

[0132] The decision module is used to make operation and maintenance decisions based on the failure probability and failure propagation path of each physical unit, and generate decisions such as replacing the SIM card or clock battery, user-side load operation and maintenance, or replacing the dedicated transformer acquisition terminal.

[0133] Example 3

[0134] This embodiment provides a computer-readable medium storing a computer program, characterized in that the computer program is executed by a processor to perform the fault prediction method for dedicated transformer data acquisition terminals that integrates demand load characteristics as described in Embodiment 1; specifically, it performs the following steps:

[0135] Step 1: Collect load fluctuation data and basic operating data of each physical unit of the target transformer acquisition terminal within the demand metering cycle, and extract the original features and load fluctuation features from the basic operating data and load fluctuation data respectively; the physical unit includes: uplink communication unit, downlink communication unit, load control loop unit, power supply unit and clock battery unit;

[0136] Step 2: Construct a causal graph using each physical unit and load fluctuation characteristics as graph nodes, define causal influence relationships for each edge and assign a prior causal strength;

[0137] Step 3: Construct a causal graph attention network based on the causal graph. Input the basic operating data and prior causal strength into the causal graph attention network for message passing to obtain the fault probability and fault propagation path of each physical unit. The causal graph attention network includes: a causal graph message passing module, a counterfactual reasoning module, and a fault probability output module. The causal graph message passing module generates node feature representations based on the original features and the attention coefficients fused with prior causal strength. The counterfactual reasoning module intervenes in the load control loop unit nodes and load fluctuation feature nodes respectively to perform counterfactual reasoning to obtain differential features. The fault probability output module concatenates the differential features with the corresponding node feature representations to calculate the fault probability.

[0138] Step 4: Make operation and maintenance decisions based on the failure probability and failure propagation path of each physical unit, and generate options such as replacing the SIM card or clock battery, user-side load operation and maintenance, or replacing the dedicated transformer acquisition terminal.

[0139] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A fault prediction method for dedicated transformer data acquisition terminals that integrates demand and load characteristics, characterized in that, The method includes: The system collects load fluctuation data and basic operating data of each physical unit of the target transformer acquisition terminal during the demand metering cycle, and extracts original features and load fluctuation features from the basic operating data and load fluctuation data respectively; the physical unit includes: an uplink communication unit, a downlink communication unit, a load control loop unit, a power supply unit, and a clock battery unit; A causal graph is constructed using each physical unit and load fluctuation characteristics as graph nodes. A causal relationship is defined for each edge and a priori causal strength is assigned. A causal graph attention network is constructed based on a causal graph. Basic operational data and prior causal strength are input into the causal graph attention network for message passing to obtain the fault probability and fault propagation path of each physical unit. The causal graph attention network includes: a causal graph message passing module, a counterfactual reasoning module, and a fault probability output module. The causal graph message passing module generates node feature representations based on original features and attention coefficients fused with prior causal strength. The counterfactual reasoning module intervenes in load control loop unit nodes and load fluctuation feature nodes respectively to perform counterfactual reasoning to obtain differential features. The fault probability output module concatenates the differential features with the corresponding node feature representations to calculate the fault probability. Operation and maintenance decisions are made based on the failure probability and failure propagation path of each physical unit, resulting in decisions such as replacing the SIM card or clock battery, maintaining user-side load, or replacing the dedicated transformer acquisition terminal.

2. The fault prediction method for dedicated transformer data acquisition terminals based on integrated demand and load characteristics according to claim 1, characterized in that, When generating a strategy to replace the dedicated transformer data acquisition terminal, the optimal replacement time is calculated and pushed to the operation and maintenance side and the user side. After receiving confirmation of replacement from the operation and maintenance side and the user side, a start command is issued to trigger the load control circuit anti-trip device. The load control circuit anti-trip device is installed between the target dedicated transformer data acquisition terminal and the load circuit. Under normal conditions, the load control circuit unit connects the load circuit. After receiving the start command, the load control circuit anti-trip device connects the load circuit.

3. The fault prediction method for dedicated transformer data acquisition terminals based on integrated demand and load characteristics according to claim 1, characterized in that, The method for extracting the original features includes: Within the sampling period T, statistical characteristics are calculated for the operating indicators reflecting the quality of the uplink communication unit channel; where the sampling period T is an integer multiple of the demand period; Within the sampling period T, statistical characteristics are calculated for the operating indicators that reflect the status of the downlink communication unit bus. Within the sampling period T, statistical characteristics of the operating indicators of the hardware health of the reaction load control loop unit are calculated. Within the sampling period T, statistical characteristics of the operating indicators of the aging degree of the reaction power supply unit are calculated. Within the sampling period T, statistical characteristics of the operating indicators of the battery life state of the reaction clock battery cell are calculated. The method for extracting load fluctuation characteristics includes: summarizing the load fluctuation data of all demand measurement cycles within the sampling period T to obtain load fluctuation characteristics.

4. The fault prediction method for dedicated transformer data acquisition terminals based on integrated demand and load characteristics according to claim 1, characterized in that, The method for defining causal influence relationships for each edge and assigning prior causal strength includes: The directed edge L→C is defined as the state change of load fluctuation characteristics causing a causal effect on the load control loop unit. The causal relationship of L→C is that the load fluctuation characteristics lead to an increase or abnormality in the operation of the load control loop. The directed edge L→P is defined as representing that the state changes of load fluctuation characteristics will causally affect the power supply unit, and the causal relationship of L→P is that load fluctuation characteristics accelerate the aging of the power supply unit. A directed edge L→U is defined to represent that the state changes of load fluctuation characteristics will causally affect the uplink communication unit. The causal relationship of L→U is that load fluctuation characteristics directly interfere with uplink communication. A directed edge P→U is defined to represent that the state change of the power supply unit will cause and affect the uplink communication unit. The causal relationship of P→U is that the power supply unit supplies power to the uplink communication unit. A directed edge P→D is defined to represent that the state change of the power supply unit will causally affect the downlink communication unit. The causal relationship of P→D is that the power supply unit supplies power to the downlink communication unit. A directed edge P→C is defined to represent that the state change of the power supply unit will cause and affect the load control loop unit. The causal relationship of P→C is that the power supply unit supplies power to the load control loop unit. The directed edge C→U is defined as the state change of the load control loop causally affecting the uplink communication unit. The causal relationship of C→U is that the electromagnetic interference generated by the operation of the load control loop unit affects the uplink communication. The directed edge C→D is defined as the state change of the load control loop causally affecting the downlink communication unit, and the causal relationship of C→D is that the electromagnetic interference generated by the operation of the load control loop affects the uplink communication. Define a directed edge B→U to represent that the state of the clock battery cell will causally affect uplink communication. The causal relationship of B→U is that the clock battery affects the timestamp, which in turn affects uplink communication. Based on domain expert knowledge, each directed edge i→j is assigned a prior causal strength c. ij ∈[0,1]; where i=L, P, C, B; j=C, P, U, D.

5. The fault prediction method for dedicated transformer data acquisition terminals based on integrated demand and load characteristics according to claim 4, characterized in that, The causal graph message passing module includes Q graph attention layers, and the l-th graph attention layer performs message passing according to the following formula: ； in, The attention coefficient represents the directed edge i→j in the m-th attention head of the l-th graph attention layer, and the attention coefficient incorporates the prior causal strength of the directed edge i→j. Represent the set of incoming neighbors of node j; Let f(x) represent the learnable weight matrix of the m-th attention head in the l-th graph attention layer. This represents the input feature vector of node i in the l-th graph attention layer; σ represents the output feature vector of node j in the (l+1)th graph attention layer; σ() represents the activation function; M represents the number of attention heads in the graph attention layer.

6. The fault prediction method for dedicated transformer data acquisition terminals based on integrated demand and load characteristics according to claim 5, characterized in that, The method for obtaining the attention coefficients that fuse prior causal strength includes: Attention coefficient α of directed edge i→j ij for: ； in, This represents the input features of node i and node j in the current graph attention layer; || represents the learnable weight matrix; || represents the vector concatenation operation; c represents the attention weight vector; ij c represents the prior causal strength of the directed edge i→j; kj β represents the prior causal strength of the directed edge k→j; β represents a learnable scalar parameter that controls the degree of influence of the prior causal strength on the final attention coefficient; LeakyReLU() represents the linear rectified activation function; exp() represents the exponential function. This represents the input features of node k in the current graph attention layer.

7. The fault prediction method for dedicated transformer data acquisition terminals based on integrated demand and load characteristics according to claim 1 or 4, characterized in that, The methods for obtaining the differential features include: First-level counterfactual reasoning: Replace the characteristics of load fluctuation characteristic node L with stable load characteristics, and calculate the change in the failure probability of the uplink communication unit ΔpU(L); Second-level counterfactual reasoning: The characteristic of the forced control loop unit node C is that it is inactive. Calculate the change in the uplink communication unit failure probability ΔpU(C); A first threshold is set. If the change ΔpU(L) exceeds the first threshold and the change ΔpU(C) is less than half of the first threshold, it is determined to be a fault propagation path where load fluctuation characteristics directly interfere with uplink communication. If both the change ΔpU(L) and the change ΔpU(C) exceed the first threshold, it is determined to be a fault propagation path where load fluctuation characteristics indirectly interfere with uplink communication through the control loop unit.

8. The fault prediction method for dedicated transformer data acquisition terminals based on integrated demand and load characteristics according to claim 7, characterized in that, The method for generating the operation and maintenance decisions includes: Preset a second threshold and two thresholds for binary classification: a high threshold and a low threshold. If the clock battery cell failure probability P b If the fault probability of the clock battery is higher than the high threshold and the fault probability of the other units is lower than the low threshold, it is recommended to replace the clock battery. If the uplink communication unit failure probability P u If the SIM card is replaced when the failure probability of the remaining units is less than the low threshold and the change ΔpU(L) is less than the second threshold; If the uplink communication unit failure probability P u If the failure probability of the remaining units is less than the low threshold and the change ΔpU(L) is greater than or equal to the second threshold, then it is recommended to perform load maintenance on the user side. Otherwise, it is recommended to replace the dedicated transformer data acquisition terminal.

9. A fault prediction method system for dedicated transformer data acquisition terminals that integrates demand and load characteristics, characterized in that: The system is used to implement the fault prediction method for a dedicated transformer data acquisition terminal that integrates demand and load characteristics as described in any one of claims 1-8, the system comprising: The acquisition module is used to acquire load fluctuation data and basic operating data of each physical unit of the target transformer acquisition terminal during the demand metering cycle, and extract the original features and load fluctuation features from the basic operating data and load fluctuation data respectively; the physical unit includes: an uplink communication unit, a downlink communication unit, a load control loop unit, a power supply unit, and a clock battery unit. The graph construction module uses each physical unit and load fluctuation characteristics as graph nodes to construct a causal graph, defines the causal influence relationship for each edge and assigns a priori causal strength; The message passing module is used to construct a causal graph attention network based on the causal graph. Basic operational data and prior causal strength are input into the causal graph attention network for message passing to obtain the fault probability and fault propagation path of each physical unit. The causal graph attention network includes: a causal graph message passing module, a counterfactual reasoning module, and a fault probability output module. The causal graph message passing module generates node feature representations based on the original features and attention coefficients fused with prior causal strength. The counterfactual reasoning module intervenes in the load control loop unit nodes and load fluctuation feature nodes respectively to perform counterfactual reasoning to obtain differential features. The fault probability output module concatenates the differential features with the corresponding node feature representations to calculate the fault probability. The decision module is used to make operation and maintenance decisions based on the failure probability and failure propagation path of each physical unit, and generate decisions such as replacing the SIM card or clock battery, user-side load operation and maintenance, or replacing the dedicated transformer acquisition terminal.

10. A computer-readable medium having a computer program stored thereon, characterized in that, The computer program, when executed by a processor, can implement the fault prediction method for dedicated transformer acquisition terminals that integrates demand load characteristics as described in any one of claims 1-8.

Images