Intelligent power distribution room ai decision analysis system based on edge computing

Abstract

The application discloses an intelligent power distribution room AI decision analysis system based on edge computing, which comprises a data acquisition preprocessing module, a friction state modeling module, a comprehensive state construction module, a fast and slow state evolution module, an operation decision determination module and a control instruction generation module.The data acquisition preprocessing module forms standardized multi-dimensional operation state data.The friction state modeling module constructs an improved GMS friction model and forms a friction state variable set.The comprehensive state construction module constructs a comprehensive operation state vector.The fast and slow state evolution module forms fast variable state evolution sequences and slow variable state evolution sequences.The operation decision determination module executes transient operation state determination and operation degradation trend determination, and generates edge side operation decision analysis information.The control instruction generation module generates operation control instructions.The application realizes stable operation decision analysis of the intelligent power distribution room on the nonlinear behavior of the actuator and the fast and slow evolution state by introducing the improved GMS friction model and the singular perturbation theory.

Description

Technical Field

[0001] This invention relates to the field of edge computing technology, and in particular to an AI decision analysis system for smart power distribution rooms based on edge computing. Background Technology

[0002] With the continuous expansion of urban power distribution systems, substations, as crucial infrastructure responsible for power distribution and operational assurance, directly impact the safety and stability of power supply. Traditional substation operation and management primarily rely on online monitoring of electrical quantities such as voltage, current, and power, combined with threshold judgments, status alarms, or human experience for operational decisions. Operational status analysis is typically based on data change characteristics at a single time scale, focusing on the identification and alarm of instantaneous anomalies, while paying insufficient attention to the mechanical wear, changes in frictional characteristics, and environmental disturbances that occur in actuators during long-term operation.

[0003] In the development of smart power distribution rooms, edge computing technology has been gradually introduced into the operation and management system of power distribution rooms for real-time processing and analysis of multi-source operational data on the field side. Basic solutions based on this technology typically integrate electrical operation data, actuator action data, and environmental status data to construct a comprehensive operational status description, and then execute status judgment and control decisions at the edge. However, existing basic solutions mostly use linear or simplified models to describe actuator behavior at the modeling level, making it difficult to characterize the nonlinear frictional characteristics inside the actuator and their evolution over time. For the simultaneous occurrence of fast-changing and slow-changing degradation processes in the operational state, a uniform time scale is often used for analysis, lacking a systematic approach to separate fast and slow state modeling. This results in room for improvement in the stability and interpretability of decision results under complex operating conditions.

[0004] Therefore, how to provide an AI decision analysis system for smart power distribution rooms based on edge computing is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an AI-based decision analysis system for smart power distribution rooms based on edge computing. This invention comprehensively utilizes edge computing, multi-source operational state fusion modeling, and fast / slow time-scale analysis methods. It addresses the operational scenarios of power distribution rooms by providing unified modeling and analysis of the electrical states, actuator action states, environmental disturbance states, and historical operational states generated during operation. By introducing an improved GMS friction model at the edge, the nonlinear friction and wear evolution of actuators in actual operation is explicitly characterized. Furthermore, singular perturbation theory is used to decompose the comprehensive operational state vector into fast and slow states, describing both transient changes and long-term degradation evolution processes during operation. This results in operational state analysis results with time-scale discrimination capabilities at the edge. The system completes operational state determination and control command generation at the edge computing node, achieving stable and continuous decision-making regarding the operational state of the power distribution room. This invention avoids the shortcomings of relying solely on electrical quantities or a single time scale for decision-making, improving the stability, interpretability, and timeliness of on-site response of operational decisions. It is suitable for smart power distribution room operation management scenarios under complex operating conditions.

[0006] The AI ​​decision analysis system for smart power distribution rooms based on edge computing according to an embodiment of the present invention includes: The data acquisition and preprocessing module collects multi-dimensional operational status data generated during the operation of the smart power distribution room at the edge computing node of the power distribution room, performs preprocessing on the multi-dimensional operational status data, and forms standardized multi-dimensional operational status data. The friction state modeling module constructs an improved GMS friction model based on standardized multidimensional operating state data, and performs parallel recursive calculations of friction state for each Maxwell-Slip friction unit to form a set of friction state variables. The integrated state construction module integrates the set of friction state variables with standardized multidimensional operating state data to construct an integrated operating state vector that includes electrical state variables, action response state variables, friction state variables, and environmental disturbance state variables. The fast and slow state evolution module is constructed based on the singular perturbation theory. It performs fast and slow state decomposition on the comprehensive operating state vector and performs state evolution calculations on the fast variable state subset and the slow variable state subset respectively under the event triggering condition, forming the fast variable state evolution sequence and the slow variable state evolution sequence. The operation decision-making module performs transient operation state determination based on the fast variable state evolution sequence and operation degradation trend determination based on the slow variable state evolution sequence. It then performs joint analysis of the transient operation state determination results and the operation degradation trend determination results to generate edge-side operation decision analysis information. The control command generation module generates corresponding operation control commands based on the edge-side operation decision analysis information.

[0007] Optionally, the multidimensional operating status data specifically includes electrical operating data, actuator action data, environmental status data, and equipment operating history data.

[0008] Optionally, the preprocessing of the multidimensional operational status data specifically includes time alignment, data cleaning, missing data completion, abnormal data removal, and unit unification.

[0009] Optionally, the friction state modeling module includes: Based on standardized multidimensional operational status data, displacement change, velocity change and action response are extracted from the actuator action data at each sampling time. These are then combined with environmental disturbance and historical action status data at the same sampling time to form an action input sequence. Unit-level mapping processing is performed on the action input sequence to generate the corresponding unit input sequence, thus obtaining the unit input set. An improved GMS friction model was constructed, which divided the Maxwell-Slip friction unit into state tracking unit and smoothing unit, and bound and registered each unit input sequence in the unit input set with the corresponding Maxwell-Slip friction unit. Based on the Maxwell-Slip friction unit and its historical internal state sequence, the historical internal state sequence of each Maxwell-Slip friction unit is divided into short-term historical state segments and long-term historical state segments according to the time span. These segments are written into the corresponding short-term cache partition and long-term cache partition, respectively, and a partition index table is constructed to form historical state cache data. Based on the unit input set and historical state cache data, the friction state recursive calculation is performed on each Maxwell-Slip friction unit in sequence. Based on the unit input sequence item and partition index table at the current sampling time, the internal state of each Maxwell-Slip friction unit is updated and the corresponding unit friction state quantity is generated, forming a set of unit friction state quantities. Based on the role labeling of Maxwell-Slip friction elements, role index tables and topology index tables are established for the set of element friction state variables according to the element role dimension and the topology connection dimension, respectively. Based on the role index table and topology index table, the set of unit friction state variables is indexed and summarized to obtain the comprehensive friction state variables at the current sampling time. The comprehensive friction state variables are then associated and encapsulated with the role index table and topology index table to form a set of friction state variables.

[0010] Optionally, the integrated state construction module includes: Based on standardized multidimensional operating status data and friction state variable set, the electrical state quantity in the electrical operating data, the action response state quantity in the actuator action data, the environmental disturbance state quantity in the environmental state data, and the comprehensive friction state quantity, unit role marking information, topology index information and historical state cache partition identification information in the friction state variable set are read respectively to form a comprehensive state fusion input dataset. Based on the comprehensive state fusion input dataset, the electrical state variables, action response state variables, environmental disturbance state variables, and comprehensive friction state variables are synchronized and aligned according to the sampling time, and structured mapping is performed to form a fusion state field sequence that corresponds one-to-one with each sampling time. Based on the fused state field sequence, vector concatenation processing is performed on the electrical state quantity, action response state quantity, comprehensive friction state quantity, environmental disturbance state quantity, and structured mapping unit role labeling information, topology index information, and historical state cache partition identification information corresponding to the same sampling time to form a single-time comprehensive state vector sequence corresponding to each sampling time. Based on the single-time comprehensive state vector sequence, sequence encapsulation processing is performed according to the sampling time order, and the single-time comprehensive state vectors corresponding to each sampling time are combined in sequence to form a comprehensive operating state vector.

[0011] Optionally, the fast and slow state evolution module includes: A fast and slow state decomposition module based on singular perturbation theory is constructed. The fast and slow state decomposition module includes a state basis transformation unit, a cascaded fast and slow state decomposition unit, a triggered fragmented evolution unit, and a fast and slow state output organization unit. The state basis transformation unit performs state basis mapping processing on the integrated operating state vector to obtain the basis space state vector, and performs component rearrangement processing on the basis space state vector according to the fast state basis component set and the slow state basis component set to form the basis space state sequence. In the cascaded fast and slow state decomposition unit, a bidirectional cascaded decomposition structure is introduced. Cascaded fast and slow state splitting is performed along the fast-to-slow decomposition path and the slow-to-fast decomposition path, respectively. In the fast-to-slow decomposition path, a first-level fast state subset is first generated and then the corresponding slow state subset is generated. In the slow-to-fast decomposition path, a first-level slow state subset is first generated and then the corresponding fast state subset is generated, forming bidirectional cascaded fast and slow state subsets, resulting in the fast state subset time series and the slow state subset time series. The triggering fragmented evolution unit divides the fast state subset time series and the slow state subset time series into segments based on the time series within the triggering window, and establishes segment graph connection relationships with each segment as a node, thus constructing a segment connection graph; Based on the fragment connection graph, state evolution calculations are performed on the fast state subset and slow state subset within each fragment according to the fragment connection order, forming a fragment-level fast state evolution sequence and a fragment-level slow state evolution sequence. The fast and slow state output organization unit constructs a dual index table for decomposition path segments based on the segment-level fast state evolution sequence, the segment-level slow state evolution sequence, the decomposition path identifier, and the segment node identifier. It then performs indexing and encapsulation processing on the fast and slow state evolution sequences according to the dual index table to form the fast variable state evolution sequence and the slow variable state evolution sequence.

[0012] Optionally, the operation decision-making module includes: The state fluctuation amplitude, state change rate, segment switching frequency and sequence extreme value distribution information of the fast variable state evolution sequence are extracted according to the sampling time, and combined in time order to form a transient operating state characteristic sequence. Based on the transient operating state feature sequence, the transient operating state features corresponding to each sampling time are matched with the transient state interval, and the operating state at each sampling time is classified and labeled according to the matching results to form the transient operating state judgment result. For the slow variable state evolution sequence, extract the state growth amount, state change slope, stage cumulative offset and adjacent time period difference information according to the sampling time, and combine them in time order to form the operation degradation trend feature sequence; Based on the operational degradation trend feature sequence, the operational degradation trend features corresponding to each sampling time are matched with the degradation stage interval to form the operational degradation trend judgment result. The transient operational status judgment result and the operational degradation trend judgment result are jointly mapped and processed according to the same sampling time to form edge-side operational decision analysis information.

[0013] Optionally, the control command generation module includes: Based on edge-side operation decision analysis information, the transient operation status judgment identifier, operation degradation trend judgment identifier and sampling time index information corresponding to each sampling time are read respectively, and then aligned and combined according to the sampling time order to form a joint judgment status input sequence. Based on the joint judgment state input sequence, the transient operation state judgment identifier and operation degradation trend judgment identifier corresponding to each sampling time are combined and mapped to form a control state coding sequence that corresponds one-to-one with each sampling time. Based on the control state coding sequence, the control state code corresponding to each sampling time is matched with the instruction coding table to generate the operation control instruction coding sequence corresponding to each sampling time. Based on the sequence of operation control instructions, each operation control instruction is encoded and encapsulated according to the sampling time order to form the corresponding operation control instruction.

[0014] The beneficial effects of this invention are: This invention proposes an AI decision analysis system for smart power distribution rooms based on edge computing. This system uses edge computing nodes in the power distribution room as its core carrier and constructs a state modeling and analysis method for real-time decision-making based on multi-dimensional operational state data generated during the operation of the power distribution room. At the edge, the system performs unified preprocessing and structured fusion of electrical operation data, actuator action data, environmental state data, and equipment operation history data. It also introduces an improved GMS friction model to explicitly model the nonlinear friction characteristics and internal state evolution of actuators during long-term operation, supplementing mechanical operation information that is difficult to reflect by traditional electrical quantity monitoring at the state level. Furthermore, based on singular perturbation theory, a fast-slow state decomposition method is constructed to distinguish and describe fast-changing and slow-changing states in the comprehensive operational state vector, enabling transient operational changes and long-term degradation evolution to be simultaneously characterized within the same analytical framework. By jointly judging the fast and slow state evolution results at the edge, operational decision analysis information is formed, and operational control commands are generated, achieving continuous perception and stable decision-making regarding the operational state of the power distribution room.

[0015] This invention introduces nonlinear friction modeling and fast / slow time-scale analysis into the edge-side decision-making process, avoiding the instability problems caused by relying solely on instantaneous electrical quantities for judgment. This improves the adaptability and interpretability of operational decisions to complex operating conditions, while reducing reliance on cloud computing and enhancing the real-time performance and reliability of on-site responses. It is suitable for the long-term safe operation and management of smart power distribution rooms. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the structure of the AI ​​decision analysis system for smart power distribution rooms based on edge computing proposed in this invention; Figure 2 This is a functional flowchart of the improved GMS friction model for the AI ​​decision analysis system for smart power distribution rooms based on edge computing proposed in this invention. Figure 3 This is a schematic diagram of the fast and slow state decomposition module of the AI ​​decision analysis system for smart power distribution rooms based on edge computing proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figure 1 , Figure 2 and Figure 3 The AI ​​decision analysis system for smart power distribution rooms based on edge computing includes: The data acquisition and preprocessing module collects multi-dimensional operational status data generated during the operation of the smart power distribution room at the edge computing node of the power distribution room, performs preprocessing on the multi-dimensional operational status data, and forms standardized multi-dimensional operational status data. The friction state modeling module constructs an improved GMS friction model based on standardized multidimensional operating state data, and performs parallel recursive calculations of friction state for each Maxwell-Slip friction unit to form a set of friction state variables. The integrated state construction module integrates the set of friction state variables with standardized multidimensional operating state data to construct an integrated operating state vector that includes electrical state variables, action response state variables, friction state variables, and environmental disturbance state variables. The fast and slow state evolution module is constructed based on the singular perturbation theory. It performs fast and slow state decomposition on the comprehensive operating state vector and performs state evolution calculations on the fast variable state subset and the slow variable state subset respectively under the event triggering condition, forming the fast variable state evolution sequence and the slow variable state evolution sequence. The operation decision-making module performs transient operation state determination based on the fast variable state evolution sequence and operation degradation trend determination based on the slow variable state evolution sequence. It then performs joint analysis of the transient operation state determination results and the operation degradation trend determination results to generate edge-side operation decision analysis information. The control command generation module generates corresponding operation control commands based on the edge-side operation decision analysis information.

[0019] In this embodiment, the multidimensional operating status data specifically includes electrical operating data, actuator action data, environmental status data, and equipment operating history data.

[0020] In this embodiment, the preprocessing of multidimensional operational status data specifically includes time alignment, data cleaning, missing data completion, abnormal data removal, and unit unification.

[0021] In this embodiment, the friction state modeling module includes: Based on standardized multidimensional operational status data, displacement change, velocity change, and action response are extracted from the actuator's motion data at each sampling time. These are then combined with environmental disturbances and historical motion state data at the same sampling time to form an action input sequence. Unit-level mapping processing is performed on the action input sequence to generate corresponding unit input sequences, resulting in a unit input set, where: Historical motion state data refers to the set of motion-related state data formed and recorded by the actuator at historical sampling times. It is stored in chronological order and reflects the motion change trajectory of the actuator's internal state data corresponding to the most recent five consecutive sampling times, including the actuator's displacement state, motion change direction, and motion continuity state corresponding to the historical sampling times. The action input sequence is processed using a unit-level mapping method, specifically: After obtaining the action input sequence arranged according to the sampling time, a correspondence table between the action input sequence and the friction element is established based on the numbering information of each Maxwell-Slip friction element in the improved GMS friction model. At each sampling time, according to the correspondence table, the displacement change, velocity change and action response in the action input sequence at the sampling time are copied and written into the input buffer corresponding to each Maxwell-Slip friction element, forming an independent element input sequence for each Maxwell-Slip friction element. All element input sequences are combined according to the friction element numbering order to form an element input set. An improved GMS friction model is constructed, dividing the Maxwell-Slip friction element into state tracking elements and smoothing elements. The input sequence of each element in the element input set is bound and registered with the corresponding Maxwell-Slip friction element, where: The Maxwell-Slip friction element is divided into a state-tracking element and a smoothing element, specifically: The Maxwell-Slip friction elements are numbered and registered. Based on the arrangement order of the friction elements in the model, some friction elements are designated as state tracking elements and others are designated as smoothing elements. The state tracking elements and smoothing elements maintain the same calculation form in the model structure, and only the different element role identifiers are marked in the numbering attribute. Based on the Maxwell-Slip friction cells and their historical internal state sequences, the historical internal state sequences of each Maxwell-Slip friction cell are divided into short-term and long-term historical state segments according to their time span. These segments are then written into the corresponding short-term and long-term cache partitions, respectively. A partition index table is constructed to form historical state cache data, where: The historical internal state sequence refers to the set of internal state data formed and recorded by each Maxwell-Slip friction unit at historical sampling time, reflecting the state evolution trajectory of the corresponding friction unit at multiple historical sampling times; The historical internal state sequence of each Maxwell–Slip friction element is divided into short-term historical state segments and long-term historical state segments according to the time span, as follows: After obtaining the historical internal state sequence arranged in chronological order, the internal state data corresponding to the five most recent sampling times are extracted from the historical internal state sequence to form a short-term historical state segment, and the internal state data corresponding to the remaining earlier sampling times are combined to form a long-term historical state segment. The partitioned index table is constructed as follows: After the short-term historical state segment and the long-term historical state segment are formed, index numbers are generated for the two types of state segments, and index entries are established. Each index entry records the corresponding friction unit number, state segment type identifier, and the start and end positions of the state segment in the storage space. All index entries are summarized according to the friction unit number to form a partitioned index table. The historical state cache data is generated as follows: Short-term historical state segments and long-term historical state segments are written into their respective cache partitions, and the partition index table is associated with the historical internal state data in the cache partition. The historical internal state data of each friction unit can be located in the corresponding cache partition through the partition index table, forming historical state cache data. Based on the unit input set and historical state cache data, the friction state recursive calculation is performed sequentially on each Maxwell-Slip friction unit. Based on the unit input sequence items and partition index table at the current sampling time, the internal state of each Maxwell-Slip friction unit is updated and the corresponding unit friction state quantity is generated, forming a set of unit friction state quantities, where: The friction state recursive calculation is performed sequentially for each Maxwell-Slip friction element, specifically as follows: According to the numbering order of the Maxwell-Slip friction elements, at each sampling time, the unit input sequence item of the corresponding friction element is read sequentially, and the corresponding historical internal state data is located and read through the partition index table. The unit input sequence item of the current sampling time and the read historical internal state data are used together as the input for recursion. After the internal state update of the Maxwell-Slip friction element at the current sampling time is completed, the same recursive process is performed on the next numbered Maxwell-Slip friction element. The set of unit friction state quantities is formed as follows: After updating the internal state of all Maxwell-Slip friction elements at the current sampling time, the updated internal state of each Maxwell-Slip friction element is read, and the corresponding element friction state variables are arranged and combined according to the number order to form a set of element friction state variables. Based on the role labeling of Maxwell-Slip friction elements, role index tables and topology index tables are established for the set of element friction state variables according to the element role dimension and topology connectivity dimension, respectively, where: For the set of unit friction state variables, establish role index tables and topology index tables according to the unit role dimension and topology connection dimension, respectively, as follows: After obtaining the set of unit friction state variables, the unit friction state variables are classified into state tracking units and smoothing units according to the role labeling information of Maxwell-Slip friction units. The state variable index position corresponding to each type of Maxwell-Slip friction unit is recorded to form a role index table. According to the parallel and serial connection relationship of Maxwell-Slip friction units in the improved GMS friction model, the set of unit friction state variables is grouped and registered, and the positional relationship of each group of friction unit state variables in the set is recorded to form a topology index table. Based on the role index table and topology index table, an indexed summary process is performed on the set of unit friction state variables to obtain the comprehensive friction state variables at the current sampling time. These comprehensive friction state variables are then associated and encapsulated with the role index table and topology index table to form a set of friction state variables, where: The set of element friction state variables is indexed and summarized, specifically as follows: After the role index table and topology index table are constructed, the unit friction state quantities under the corresponding role category and topology group are read according to the index position relationship recorded in the index table. The read unit friction state quantities are combined according to the index order to generate the comprehensive friction state quantity at the current sampling time. The comprehensive friction state quantity is encapsulated together with the corresponding role index table and topology index table to form a set of friction state variables.

[0022] In this embodiment, the comprehensive state construction module includes: Based on standardized multidimensional operating status data and friction state variable set, the electrical state quantity in the electrical operating data, the action response state quantity in the actuator action data, the environmental disturbance state quantity in the environmental state data, and the comprehensive friction state quantity, unit role marking information, topology index information and historical state cache partition identification information in the friction state variable set are read respectively to form a comprehensive state fusion input dataset. Based on the comprehensive state fusion input dataset, the electrical state variables, action response state variables, environmental disturbance state variables, and comprehensive friction state variables are synchronized and aligned according to the sampling time, and then structured mapping is performed to form a fusion state field sequence that corresponds one-to-one with each sampling time, where: Perform structured mapping processing, specifically: After aligning the sampling times of electrical state quantities, action response state quantities, environmental disturbance state quantities, and comprehensive friction state quantities, a fixed field identifier and field order are assigned to each type of state quantity, and a field position index table is predefined. The field position index table records the writing position of each type of state quantity in the fused state field. At each sampling time, according to the field order determined in the field position index table, the electrical state quantity of the corresponding sampling time is written to the electrical state field position, the action response state quantity is written to the action response field position, the environmental disturbance state quantity is written to the environmental disturbance field position, and the comprehensive friction state quantity is written to the friction state field position. The state variables written in each field position at the same sampling time are combined to form a complete field record. All field records corresponding to the sampling time are arranged in chronological order to form a fused state field sequence. Based on the fused state field sequence, vector concatenation processing is performed on the electrical state quantity, action response state quantity, comprehensive friction state quantity, environmental disturbance state quantity, and structured mapping unit role labeling information, topology index information, and historical state cache partition identification information corresponding to the same sampling time to form a single-time comprehensive state vector sequence corresponding to each sampling time. Based on the single-time comprehensive state vector sequence, sequence encapsulation processing is performed according to the sampling time order, and the single-time comprehensive state vectors corresponding to each sampling time are combined in sequence to form a comprehensive operating state vector.

[0023] In this embodiment, the fast and slow state evolution module includes: A fast-slow state decomposition module based on singular perturbation theory is constructed. This module includes a state basis transformation unit, a cascaded fast-slow state decomposition unit, a triggered fragmented evolution unit, and a fast-slow state output organization unit, wherein: A fast-slow state decomposition module based on singular perturbation theory is constructed, specifically as follows: The state basis transformation unit performs state basis mapping on the synthesized operating state vector to obtain the basis space state vector, and then performs component rearrangement on the basis space state vector according to the fast state basis component set and the slow state basis component set to form the basis space state sequence, where: The execution of state basis mapping processing is as follows: After obtaining the comprehensive operating state vector, position mapping is performed on each dimension of the comprehensive operating state vector. Each state component in the original state vector is written into the corresponding position index in the basis space according to the mapping relationship. This completes the vector rearrangement from the original state space to the basis space, resulting in a basis space state vector that is numerically consistent with the comprehensive operating state vector but completes the basis mapping in terms of structural arrangement. The component rearrangement process is executed as follows: After obtaining the basis space state vector, the state components in the basis space state vector are reordered according to the index order of the basis components recorded in the fast state basis component set and the slow state basis component set. The state components belonging to the fast state basis component set are arranged continuously, and the state components belonging to the slow state basis component set are arranged continuously. The rearranged basis space state vector is formed into a basis space state sequence according to the sampling time order. A bidirectional cascaded decomposition structure is introduced into the cascaded fast-slow state decomposition unit. Cascaded fast-slow state splitting is performed along both the fast-to-slow decomposition path and the slow-to-fast decomposition path. In the fast-to-slow decomposition path, a first-level fast state subset is first generated, followed by the corresponding slow state subset; conversely, in the slow-to-fast decomposition path, a first-level slow state subset is first generated, followed by the corresponding fast state subset, forming bidirectional cascaded fast and slow state subsets. This yields the fast state subset time series and the slow state subset time series, where: A bidirectional cascaded decomposition structure refers to establishing two decomposition paths on the same set of base space state sequences, one along the fast state priority direction and the other along the slow state priority direction, and performing fast and slow state splitting in the two paths according to different decomposition orders. In the fast-to-slow decomposition path, a first-level fast state subset is first generated, and then the corresponding slow state subset is generated. Specifically: Under the fast-to-slow decomposition path, the corresponding state components are extracted from the base space state sequence according to the fast state base component set, and combined at each sampling time to form a first-level fast state subset. The first-level fast state subset is removed from the base space state sequence, and the remaining state components are recombined according to the slow state base component set to form a slow state subset that corresponds one-to-one with the first-level fast state subset on the time axis. In the slow-to-fast decomposition path, a first-level slow state subset is first generated, and then the corresponding fast state subset is generated. Specifically: Under the slow-to-fast decomposition path, the corresponding state components are extracted from the base space state sequence according to the slow state base component set, and combined at each sampling time to form a first-level slow state subset. The first-level slow state subset is removed from the base space state sequence, and the remaining state components are recombined according to the fast state base component set to form a fast state subset that corresponds one-to-one with the first-level slow state subset on the time axis. The time series of the fast state subset and the slow state subset are obtained as follows: After completing the state splitting under the fast-to-slow decomposition path and the slow-to-fast decomposition path, the fast state subsets generated in the two decomposition paths are arranged and merged according to the sampling time order to form a fast state subset time series. At the same time, the slow state subsets generated in the two decomposition paths are arranged and merged according to the sampling time order to form a slow state subset time series. The triggering fragmentation evolution unit divides the fast-state subset time series and the slow-state subset time series into fragments based on the time series within the trigger window, and establishes fragment graph connections using each fragment as a node, constructing a fragment connection graph, where: Constructing a fragment connection graph, specifically: After obtaining the fast state subset time series and the slow state subset time series, the time series is divided into continuous segments according to the start and end positions of the trigger window on the time axis. Each segment is assigned a unique segment node identifier. According to the order of the segments on the time axis, the connection relationship between adjacent segment nodes is established. All segment nodes and connection relationships together constitute the segment connection graph. Based on the fragment connectivity graph, state evolution calculations are performed on the fast state subsets and slow state subsets within each fragment according to the fragment connectivity order, forming fragment-level fast state evolution sequences and fragment-level slow state evolution sequences, where: State evolution calculations are performed on the fast state subsets and slow state subsets within each segment according to the segment connection order, specifically as follows: After the fragment connection graph is constructed, the fast state subset and slow state subset corresponding to each fragment are read sequentially according to the fragment connection order recorded in the fragment connection graph. Within each fragment, the fast state subset and slow state subset are processed separately according to the sampling time order, forming a continuous fragment-level fast state evolution sequence and fragment-level slow state evolution sequence within each fragment. The fast and slow state output organization unit constructs a dual-index table for decomposition path fragments based on fragment-level fast state evolution sequences, fragment-level slow state evolution sequences, decomposition path identifiers, and fragment node identifiers. Then, it performs indexing and encapsulation processing on the fast and slow state evolution sequences according to the dual-index table, forming fast variable state evolution sequences and slow variable state evolution sequences, where: Construct a dual-index table for the decomposed path segments, specifically as follows: After obtaining the fragment-level fast state evolution sequence and the fragment-level slow state evolution sequence, an index entry is created for each decomposition path and each fragment node. The index entry records the decomposition path identifier, fragment node identifier, and the positional relationship of the corresponding state evolution sequence in the storage space. All index entries are combined and arranged according to the decomposition path identifier and fragment node identifier to form a decomposition path fragment dual index table. The fast variable state evolution sequence and the slow variable state evolution sequence are formed, specifically as follows: Based on the index relationship recorded in the decomposition path segment dual index table, the segment-level fast state evolution sequences under the corresponding decomposition path and segment nodes are read sequentially in chronological order and spliced ​​together to form a complete fast variable state evolution sequence. At the same time, the segment-level slow state evolution sequences are read and spliced ​​in the same way to form a slow variable state evolution sequence that corresponds one-to-one with the fast variable state evolution sequence on the time axis.

[0024] In this embodiment, the operation decision-making module includes: The state fluctuation amplitude, state change rate, segment switching frequency and sequence extreme value distribution information of the fast variable state evolution sequence are extracted according to the sampling time, and combined in time order to form a transient operating state characteristic sequence. Based on the transient operating state feature sequence, the transient operating state features corresponding to each sampling time are matched with the transient state interval, and the operating state at each sampling time is classified and labeled according to the matching results to form a transient operating state determination result, wherein: Transient state intervals refer to the set of state value segments defined by the distribution range of transient operating state characteristics in the numerical space. The upper and lower bounds of the transient operating state characteristics are used as boundary conditions to describe the value range of different transient operating states in the feature space. The transient operating state determination result is as follows: After obtaining the transient operating state feature sequence arranged according to the sampling time, for each sampling time, the transient operating state feature value corresponding to the sampling time is read and compared with the value boundary of each segment in the transient state interval one by one. When the transient operating state feature value falls within the value range of the transient state interval, the corresponding state identifier is recorded as the transient operating state mark of the sampling time. All the state marks of the sampling times are arranged in chronological order to form the transient operating state judgment result. For the slow variable state evolution sequence, extract the state growth amount, state change slope, stage cumulative offset and adjacent time period difference information according to the sampling time, and combine them in time order to form the operation degradation trend feature sequence; Based on the operational degradation trend feature sequence, the operational degradation trend features corresponding to each sampling time are matched with the degradation stage interval to form the operational degradation trend determination result. The transient operational state determination result and the operational degradation trend determination result are then jointly mapped at the same sampling time to form edge-side operational decision analysis information, wherein: The degradation stage interval refers to a set of continuous segments in the numerical space defined by the magnitude and cumulative range of changes in the characteristics of operational degradation trends in the time series. Each segment corresponds to a stage position in the operational degradation process. Each degradation stage interval is distinguished from each other on the numerical boundary, representing the characteristic value range of the operational state under different degradation stages. The results of the operational degradation trend determination are as follows: After forming the operational degradation trend feature sequence, for each sampling time, the operational degradation trend feature vector corresponding to that sampling time is read and compared with the numerical boundaries of each segment in the degradation stage interval. Based on the comparison result, the degradation stage interval identifier corresponding to that sampling time is determined. The degradation stage interval identifiers corresponding to all sampling times are arranged in chronological order to form the operational degradation trend determination result. The information generated for edge-side operational decision analysis includes: After obtaining the transient operating status determination results and the operating degradation trend determination results, the two types of determination results are time-aligned according to the sampling time. At each sampling time, the corresponding transient operating status mark and the operating degradation stage mark are combined and encoded to generate a joint status identifier sequence that corresponds one-to-one with the sampling time. The joint status identifier sequence, together with the sampling time index information, constitutes the edge-side operating decision analysis information.

[0025] In this embodiment, the control command generation module includes: Based on edge-side operation decision analysis information, the transient operation status judgment identifier, operation degradation trend judgment identifier and sampling time index information corresponding to each sampling time are read respectively, and then aligned and combined according to the sampling time order to form a joint judgment status input sequence. Based on the joint judgment state input sequence, the transient operating state judgment identifier and the operating degradation trend judgment identifier corresponding to each sampling time are combined and mapped to form a control state coding sequence that corresponds one-to-one with each sampling time, wherein: The transient operating state determination identifier and the operating degradation trend determination identifier corresponding to each sampling time are combined and mapped, specifically as follows: After obtaining the joint judgment state input sequence, for each sampling time, the transient operation state judgment identifier and the operation degradation trend judgment identifier corresponding to the sampling time are read respectively, and the two types of identifiers are combined into a joint state code input item. The joint state code input item is written into the code position of the corresponding sampling time. Each sampling time corresponds to a control state code composed of the transient operation state judgment identifier and the operation degradation trend judgment identifier, forming a control state code sequence arranged in the order of sampling time. Based on the control state coding sequence, the control state codes corresponding to each sampling time are matched with the instruction coding table to generate the operation control instruction coding sequence corresponding to each sampling time, wherein: The instruction encoding table is a structured data table stored in the edge computing nodes of the smart power distribution room. This data table uses control status codes as index keys and operation control instruction codes as corresponding values ​​to record the correspondence between various control status codes and operation control instruction codes. The instruction encoding table is constructed during the system deployment phase and remains unchanged during operation. The specific steps for generating the operation control instruction encoding sequence corresponding to each sampling time are as follows: After forming the control status code sequence, the control status codes corresponding to each sampling time are read sequentially according to the sampling time order. The control status codes are used as index keys to perform a lookup operation in the instruction code table to read the corresponding operation control instruction codes. The read operation control instruction codes are arranged and combined according to the sampling time order to form an operation control instruction code sequence that corresponds one-to-one with each sampling time. Based on the sequence of operation control instructions, each operation control instruction is encoded and encapsulated according to the sampling time order to form the corresponding operation control instruction.

[0026] Example 1: To verify the feasibility of this invention in practice, it was applied to a 110kV smart substation in a coastal city. This substation serves surrounding industrial parks and residential areas and is equipped with multiple circuit breakers, disconnect switches, and automatic actuators. It operates under high load and frequent start-stop conditions for extended periods. Due to the high humidity and large diurnal temperature range in the region, the actuators gradually exhibited problems such as increased mechanical friction and slow response during long-term operation. Traditional monitoring systems mainly rely on electrical quantities such as voltage and current for judgment. In multiple actual operations, there were instances where electrical parameters were normal but the actuators malfunctioned, making it difficult for maintenance personnel to identify potential risks in a timely manner.

[0027] The system of this invention is deployed in the power distribution room. First, at the edge computing node, it performs unified preprocessing and structured integration of collected electrical operation data, actuator action data, environmental status data, and equipment operation history data. The system continuously samples the displacement changes, operating speed, and response delay of the circuit breaker actuator. It also models the internal friction state of the actuator using an improved GMS friction model, obtaining friction state variables that reflect mechanical wear and friction evolution. These friction state variables are then fused with electrical state variables and environmental disturbance state variables to form a comprehensive operating state vector. Based on singular perturbation theory, the operating state is decomposed into fast and slow states, distinguishing between transient action anomalies over short periods and degradation trends accumulated over long periods.

[0028] In a comparative test running continuously for 60 days, the system monitored the actuators of three key circuit breakers with a sampling period of 1 second. Experimental data showed that from day 37 to day 45, the system detected that the cumulative offset of friction-related state quantities in the slow variable state increased by approximately 18% compared to the initial stage. Simultaneously, multiple short-term action response delays occurred in the fast variable state, but the corresponding current and voltage fluctuations remained within the rated range. Based on the joint determination results of the fast and slow states, the system generated operational decision analysis information at the edge and output operational control commands in advance to adjust the operational strategies of relevant equipment. Post-inspection results showed that the internal lubrication of one circuit breaker was significantly deteriorated, posing a risk of jamming if operation continued. Compared to historical data without the deployment of this invention's system, the average identification time of potential actuator anomalies in this embodiment was reduced by approximately 3.5 days, and the false positive rate was reduced from approximately 12% in the original system to below 4%, significantly improving the stability and reliability of operational decisions.

[0029] As can be seen from the embodiments, the present invention can perform detailed modeling and analysis of complex operating states at the edge of the smart power distribution room, effectively solving the problem that it is difficult to identify the hidden risks of the actuator in a timely manner by relying solely on electrical quantities, and has good practical value in actual engineering applications.

[0030] Table 1. Statistical Table of Comparative Monitoring and Judgment Results of the Operating Status of the Actuators in the Smart Power Distribution Room

[0031] As can be seen from the data in Table 1, the operating status of the actuator exhibited obvious phased changes during the 60-day continuous monitoring process. In the initial stage, the average action time remained between 128ms and 131ms, with relatively small fluctuations. The comprehensive friction state quantity and the cumulative offset rate of slow variables were both at low levels. The transient state judgment results and degradation trend judgment results both showed that the equipment was in a normal or initial stage, indicating that the equipment was operating stably at this time, and mechanical friction and wear had not yet accumulated significantly. The judgment results of the traditional judgment method based on electrical quantities were basically consistent with those of the system of this invention.

[0032] As operation progressed, around day 30, the average action time increased to 138ms, with significantly increased fluctuations. The comprehensive friction state quantity and the cumulative offset rate of slow variables rose simultaneously, indicating anomalies in the transient operational status assessment, and the degradation trend entered a development phase. At this time, the current deviation rate remained below 1.4%, and traditional electrical quantity assessment results still showed normality. However, the system of this invention had already issued a risk warning, demonstrating its ability to proactively detect changes in the mechanical state of the actuator. From day 37 to day 45, the cumulative offset rate of slow variables continued to rise to 21.4%, the comprehensive friction state quantity approached 0.70, the frequency of transient anomalies increased, and the degradation trend shifted from the development phase to the acceleration phase. The system of this invention continuously output operational intervention and enhanced control judgments, while traditional methods still did not trigger alarms.

[0033] In the later stages of operation, the average action time reached 156ms to 160ms, with a fluctuation range exceeding 22ms and a cumulative offset rate of nearly 30% for slow variables, indicating that the actuator was in a significant degradation state. Traditional electrical quantity judgments only begin to generate alarms at this stage, while the system of this invention has already completed load reduction and shutdown maintenance decisions in advance. Overall, the method of this invention can identify potential risks to the actuator in advance through friction state modeling and joint analysis of fast and slow states, even before electrical parameters show obvious abnormalities. This significantly improves the foresight, stability, and reliability of operational decisions, verifying its practical application value in the complex operating conditions of smart power distribution rooms.

[0034] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An AI decision analysis system for smart power distribution rooms based on edge computing, characterized in that, include: The data acquisition and preprocessing module collects multi-dimensional operational status data generated during the operation of the smart power distribution room at the edge computing node of the power distribution room, performs preprocessing on the multi-dimensional operational status data, and forms standardized multi-dimensional operational status data. The friction state modeling module constructs an improved GMS friction model based on standardized multidimensional operating state data, and performs parallel recursive calculations of friction state for each Maxwell-Slip friction unit to form a set of friction state variables. The integrated state construction module integrates the set of friction state variables with standardized multidimensional operating state data to construct an integrated operating state vector that includes electrical state variables, action response state variables, friction state variables, and environmental disturbance state variables. The fast and slow state evolution module is constructed based on the singular perturbation theory. It performs fast and slow state decomposition on the comprehensive operating state vector and performs state evolution calculations on the fast variable state subset and the slow variable state subset respectively under the event triggering condition, forming the fast variable state evolution sequence and the slow variable state evolution sequence. The operation decision-making module performs transient operation state determination based on the fast variable state evolution sequence and operation degradation trend determination based on the slow variable state evolution sequence. It then performs joint analysis of the transient operation state determination results and the operation degradation trend determination results to generate edge-side operation decision analysis information. The control command generation module generates corresponding operation control commands based on the edge-side operation decision analysis information.

2. The AI ​​decision analysis system for smart power distribution rooms based on edge computing according to claim 1, characterized in that, The multidimensional operating status data specifically includes electrical operating data, actuator action data, environmental status data, and equipment operating history data.

3. The AI ​​decision analysis system for smart power distribution rooms based on edge computing according to claim 1, characterized in that, The preprocessing of multidimensional operational status data specifically includes time alignment, data cleaning, missing data completion, abnormal data removal, and unit unification.

4. The AI ​​decision analysis system for smart power distribution rooms based on edge computing according to claim 1, characterized in that, The friction state modeling module includes: Based on standardized multidimensional operational status data, displacement change, velocity change and action response are extracted from the actuator action data at each sampling time. These are then combined with environmental disturbance and historical action status data at the same sampling time to form an action input sequence. Unit-level mapping processing is performed on the action input sequence to generate the corresponding unit input sequence, thus obtaining the unit input set. An improved GMS friction model was constructed, which divided the Maxwell-Slip friction unit into state tracking unit and smoothing unit, and bound and registered each unit input sequence in the unit input set with the corresponding Maxwell-Slip friction unit. Based on the Maxwell-Slip friction unit and its historical internal state sequence, the historical internal state sequence of each Maxwell-Slip friction unit is divided into short-term historical state segments and long-term historical state segments according to the time span. These segments are written into the corresponding short-term cache partition and long-term cache partition, respectively, and a partition index table is constructed to form historical state cache data. Based on the unit input set and historical state cache data, the friction state recursive calculation is performed on each Maxwell-Slip friction unit in sequence. Based on the unit input sequence item and partition index table at the current sampling time, the internal state of each Maxwell-Slip friction unit is updated and the corresponding unit friction state quantity is generated, forming a set of unit friction state quantities. Based on the role labeling of Maxwell-Slip friction elements, role index tables and topology index tables are established for the set of element friction state variables according to the element role dimension and the topology connection dimension, respectively. Based on the role index table and topology index table, the set of unit friction state variables is indexed and summarized to obtain the comprehensive friction state variables at the current sampling time. The comprehensive friction state variables are then associated and encapsulated with the role index table and topology index table to form a set of friction state variables.

5. The AI ​​decision analysis system for smart power distribution rooms based on edge computing according to claim 1, characterized in that, The comprehensive state construction module includes: Based on standardized multidimensional operating status data and friction state variable set, the electrical state quantity in the electrical operating data, the action response state quantity in the actuator action data, the environmental disturbance state quantity in the environmental state data, and the comprehensive friction state quantity, unit role marking information, topology index information and historical state cache partition identification information in the friction state variable set are read respectively to form a comprehensive state fusion input dataset. Based on the comprehensive state fusion input dataset, the electrical state variables, action response state variables, environmental disturbance state variables, and comprehensive friction state variables are synchronized and aligned according to the sampling time, and structured mapping is performed to form a fusion state field sequence that corresponds one-to-one with each sampling time. Based on the fused state field sequence, vector concatenation processing is performed on the electrical state quantity, action response state quantity, comprehensive friction state quantity, environmental disturbance state quantity, and structured mapping unit role labeling information, topology index information, and historical state cache partition identification information corresponding to the same sampling time to form a single-time comprehensive state vector sequence corresponding to each sampling time. Based on the single-time comprehensive state vector sequence, sequence encapsulation processing is performed according to the sampling time order, and the single-time comprehensive state vectors corresponding to each sampling time are combined in sequence to form a comprehensive operating state vector.

6. The AI ​​decision analysis system for smart power distribution rooms based on edge computing according to claim 1, characterized in that, The fast and slow state evolution module includes: A fast and slow state decomposition module based on singular perturbation theory is constructed. The fast and slow state decomposition module includes a state basis transformation unit, a cascaded fast and slow state decomposition unit, a triggered fragmented evolution unit, and a fast and slow state output organization unit. The state basis transformation unit performs state basis mapping processing on the integrated operating state vector to obtain the basis space state vector, and performs component rearrangement processing on the basis space state vector according to the fast state basis component set and the slow state basis component set to form the basis space state sequence. In the cascaded fast and slow state decomposition unit, a bidirectional cascaded decomposition structure is introduced. Cascaded fast and slow state splitting is performed along the fast-to-slow decomposition path and the slow-to-fast decomposition path, respectively. In the fast-to-slow decomposition path, a first-level fast state subset is first generated and then the corresponding slow state subset is generated. In the slow-to-fast decomposition path, a first-level slow state subset is first generated and then the corresponding fast state subset is generated, forming bidirectional cascaded fast and slow state subsets, resulting in the fast state subset time series and the slow state subset time series. The triggering fragmented evolution unit divides the fast state subset time series and the slow state subset time series into segments based on the time series within the triggering window, and establishes segment graph connection relationships with each segment as a node, thus constructing a segment connection graph; Based on the fragment connection graph, state evolution calculations are performed on the fast state subset and slow state subset within each fragment according to the fragment connection order, forming a fragment-level fast state evolution sequence and a fragment-level slow state evolution sequence. The fast and slow state output organization unit constructs a dual index table for decomposition path segments based on the segment-level fast state evolution sequence, the segment-level slow state evolution sequence, the decomposition path identifier, and the segment node identifier. It then performs indexing and encapsulation processing on the fast and slow state evolution sequences according to the dual index table to form the fast variable state evolution sequence and the slow variable state evolution sequence.

7. The AI ​​decision analysis system for smart power distribution rooms based on edge computing according to claim 1, characterized in that, The operation decision-making module includes: The state fluctuation amplitude, state change rate, segment switching frequency and sequence extreme value distribution information of the fast variable state evolution sequence are extracted according to the sampling time, and combined in time order to form a transient operating state characteristic sequence. Based on the transient operating state feature sequence, the transient operating state features corresponding to each sampling time are matched with the transient state interval, and the operating state at each sampling time is classified and labeled according to the matching results to form the transient operating state judgment result. For the slow variable state evolution sequence, extract the state growth amount, state change slope, stage cumulative offset and adjacent time period difference information according to the sampling time, and combine them in time order to form the operation degradation trend feature sequence; Based on the operational degradation trend feature sequence, the operational degradation trend features corresponding to each sampling time are matched with the degradation stage interval to form the operational degradation trend judgment result. The transient operational status judgment result and the operational degradation trend judgment result are jointly mapped and processed according to the same sampling time to form edge-side operational decision analysis information.

8. The AI ​​decision analysis system for smart power distribution rooms based on edge computing according to claim 1, characterized in that, The control command generation module includes: Based on edge-side operation decision analysis information, the transient operation status judgment identifier, operation degradation trend judgment identifier and sampling time index information corresponding to each sampling time are read respectively, and then aligned and combined according to the sampling time order to form a joint judgment status input sequence. Based on the joint judgment state input sequence, the transient operation state judgment identifier and operation degradation trend judgment identifier corresponding to each sampling time are combined and mapped to form a control state coding sequence that corresponds one-to-one with each sampling time. Based on the control state coding sequence, the control state code corresponding to each sampling time is matched with the instruction coding table to generate the operation control instruction coding sequence corresponding to each sampling time. Based on the sequence of operation control instructions, each operation control instruction is encoded and encapsulated according to the sampling time order to form the corresponding operation control instruction.

Images