Auxiliary decision-making method and device based on alarm signal of power system, electronic equipment and storage medium

Abstract

The invention discloses an auxiliary decision-making method and device based on an electric power system alarm signal, electronic equipment and a storage medium, and belongs to the technical field of electric power system automation, and the method comprises the steps: carrying out the feature space mapping processing of an alarm signal sequence, and generating an alarm signature; calculating the similarity between the alarm signature and a historical alarm signature of each historical case in a historical case knowledge base, and generating a similarity score of each historical case; selecting a preset number of historical cases of which the similarity scores are higher than a preset similarity threshold as target reference cases; performing aggregation processing on the historical processing information in each target reference case to generate background context information; according to background context information and the alarm signal sequence, constructing prompt information; the prompt information is input into a generative decision model, and auxiliary decision information for the alarm signal sequence is generated; according to the invention, the problem of low decision-making efficiency of the alarm signal in the prior art can be solved.

Description

Technical Field

[0001] This invention relates to the field of power system automation technology, and specifically to an auxiliary decision-making method, device, electronic device, and storage medium based on power system alarm signals. Background Technology

[0002] As the energy hub of modern society, the safe and stable operation of the power system is directly related to the national economy and people's livelihood. With the deepening of smart grid construction, the scale of the power system is growing larger and larger, especially the secondary professional systems involving multiple fields such as relay protection, automation, network security, communication, and security control, which have become a key line of defense for ensuring the reliable operation of the power grid. In the daily operation and maintenance and fault handling of the power system, the ability to identify anomalies from massive amounts of monitoring data in a timely and accurate manner and make correct dispatch decisions at the first moment is the core guarantee for preventing the escalation of accidents and shortening power outage time. Therefore, efficient auxiliary decision analysis based on alarm signals has become an important research direction for improving the intelligent operation and maintenance level of the power grid.

[0003] However, current auxiliary decision-making schemes for alarm signals in power systems mainly rely on human experience, resulting in low efficiency. Specifically, when the dispatch center or monitoring system receives an alarm signal, decisions are primarily made by maintenance personnel based on their personal experience. In this model, maintenance personnel need to spend a significant amount of time recalling similar historical fault scenarios or manually searching and comparing massive amounts of historical data to find a reference for handling the situation. This decision-making process, which heavily relies on individual proficiency and lacks automated assistance, leads to a long processing cycle for alarm signals, making it difficult to meet the high efficiency requirements of timely decision-making in power systems during fault emergency response, thus resulting in low efficiency in alarm signal decision-making. Summary of the Invention

[0004] This invention provides an auxiliary decision-making method, device, electronic device, and storage medium based on power system alarm signals, which can solve the problem of low decision-making efficiency for alarm signals in the prior art.

[0005] One embodiment of the present invention provides an auxiliary decision-making method based on power system alarm signals, comprising: Acquire the alarm signal sequence of the power system within a preset time period; The alarm signal sequence is subjected to feature space mapping processing to generate an alarm signature that characterizes the semantic features of the alarm signal sequence; Based on a pre-set historical case knowledge base, the similarity between the alarm signature and the historical alarm signatures of each historical case in the historical case knowledge base is calculated, and a similarity score for each historical case is generated. Select a preset number of historical cases with similarity scores higher than a preset similarity threshold as target reference cases; The historical handling information in each target reference case is aggregated to generate background context information; based on the background context information and the alarm signal sequence, a prompt message is constructed. The prompt information is input into a preset generative decision model, so that the generative decision model generates auxiliary decision information for the alarm signal sequence based on the prompt information; wherein, the generative decision model is trained by several training samples; each training sample includes historical prompt information and corresponding auxiliary decision labels.

[0006] Furthermore, the alarm signal sequence is subjected to feature space mapping processing to generate an alarm signature for characterizing the semantic features of the alarm signal sequence, including: The alarm signal sequence is sliced ​​to generate several alarm segments; Numerical mapping is performed on each alarm segment to generate the basic values ​​for each alarm segment; Based on the base values ​​of each alarm segment, construct a set of basic values; For each hash function in the preset set of hash functions, hash calculation is performed on each basic value in the basic value set based on the current hash function to generate the current hash value set; Select the smallest hash value from the current set of hash values, and use it as the minimum hash value corresponding to the current hash function; The minimum hash values ​​are combined to generate an alarm signature that represents the semantic features of the alarm signal sequence.

[0007] Furthermore, the alarm signal sequence is sliced ​​to generate several alarm segments, including: Based on a preset sliding window and a preset sliding step size, the alarm signal sequence is traversed by sliding to determine the alarm subsequence covered by each sliding step; wherein, the alarm signal sequence is composed of several alarm signals; The alarm signals of each alarm subsequence are spliced ​​together in chronological order to generate the corresponding alarm segments.

[0008] Furthermore, each alarm segment is numerically mapped to generate basic values ​​for each alarm segment, including: For each alarm segment, based on a preset numerical encoding algorithm, the current alarm segment is encoded to generate a current integer identifier value that uniquely identifies the current alarm segment; Use the current integer flag value as the base value for the current alarm segment.

[0009] Furthermore, based on a preset historical case knowledge base, the similarity between the alarm signature and the historical alarm signatures of each historical case in the historical case knowledge base is calculated, and a similarity score for each historical case is generated, including: For each historical case in the preset historical case knowledge base, the alarm signature is compared with the current historical alarm signature at the corresponding position to determine the number of elements whose alarm signature and current historical alarm signature values ​​are consistent. The number of elements whose alarm signature matches the current historical alarm signature value is taken as the number of matching features of the current historical alarm signature; Calculate the ratio of the number of matching features in the current historical alarm signature to the total number of elements contained in the alarm signature, and generate the similarity score of the current historical case.

[0010] Furthermore, the historical handling information in each target reference case is aggregated to generate background context information, including: For each target reference case, based on the historical handling information contained in the current target reference case and the preset case guidance words, a standard reference text for the current target reference case is generated. Based on the similarity scores of each target reference case arranged from high to low, and using a preset delimiter, the standard reference texts are sequentially concatenated to generate background context information.

[0011] Furthermore, the generative decision-making model is trained in the following way: Obtain several training samples; Each training sample is sequentially input into the generative decision model to train it until a preset number of training iterations are reached. Each time a training sample is received, the generative decision model outputs the corresponding predictive auxiliary decision information. A loss function value is calculated based on the predictive auxiliary decision information and the corresponding auxiliary decision label. The generative decision model is then updated based on the loss function value.

[0012] Based on the above method embodiments, the present invention provides corresponding apparatus embodiments.

[0013] An embodiment of the present invention provides an auxiliary decision-making device based on power system alarm signals, comprising: a data acquisition module, a target reference case determination module, a prompt information determination module, and an auxiliary decision-making module; The data acquisition module is used to acquire the alarm signal sequence of the power system within a preset time period; The target reference case determination module is used to perform feature space mapping processing on the alarm signal sequence to generate an alarm signature that represents the semantic features of the alarm signal sequence; based on a preset historical case knowledge base, it calculates the similarity between the alarm signature and the historical alarm signatures of each historical case in the historical case knowledge base to generate a similarity score for each historical case; and selects a preset number of historical cases with similarity scores higher than a preset similarity threshold as target reference cases. The prompt information determination module is used to aggregate and process the historical handling information in each target reference case to generate background context information; and to construct prompt information based on the background context information and the alarm signal sequence. The auxiliary decision-making module is used to input the prompt information into a preset generative decision-making model, so that the generative decision-making model generates auxiliary decision-making information for the alarm signal sequence based on the prompt information; wherein, the generative decision-making model is trained by several training samples; each training sample includes historical prompt information and corresponding auxiliary decision-making labels.

[0014] Based on the above method embodiments, the present invention provides corresponding electronic device embodiments.

[0015] An embodiment of the present invention provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the auxiliary decision-making method based on power system alarm signals as described in any of the above-described method embodiments.

[0016] Based on the above method embodiments, the present invention provides corresponding storage medium embodiments.

[0017] One embodiment of the present invention provides a storage medium storing a computer program thereon, wherein, when the computer program is running, it controls the device where the storage medium is located to execute any of the auxiliary decision-making methods based on power system alarm signals described in the above-described method embodiments.

[0018] Compared with the prior art, the present invention has the following beneficial effects: This invention provides an auxiliary decision-making method, apparatus, electronic device, and storage medium based on power system alarm signals. The method acquires an alarm signal sequence of the power system within a preset time period and maps it to a feature space to obtain an alarm signature representing semantic features. Based on a preset historical case knowledge base, it calculates the similarity between the alarm signature and each historical alarm signature, generating a similarity score. Several historical cases with similarity scores higher than a preset threshold are selected as target reference cases. Historical handling information of the target reference cases is aggregated to form background context information. Based on the background context information and the current alarm signal sequence, prompt information is constructed and input into a preset generative decision-making model to generate auxiliary decision-making information for the alarm signal sequence.

[0019] This invention generates alarm signatures by mapping alarm signal sequences to their feature space and automatically retrieves highly relevant target reference cases based on a historical case knowledge base, thus replacing the inefficient manual retrieval and comparison process. Simultaneously, to address the problem of low decision-making efficiency due to reliance on personal experience, this invention constructs a background context by aggregating historical handling information and inputting it into a trained generative decision-making model to directly generate auxiliary decision-making information, achieving automated reasoning based on historical experience. This approach effectively avoids the subjectivity and lag of manual decision-making, significantly shortens the fault response cycle, and improves the efficiency of power system alarm signal decision-making. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating an auxiliary decision-making method based on power system alarm signals provided in an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the structure of an auxiliary decision-making device based on power system alarm signals provided in an embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] like Figure 1 As shown, to address the problem of low decision-making efficiency for alarm signals in existing technologies, an embodiment of the present invention provides an auxiliary decision-making method based on power system alarm signals, comprising at least the following steps: Step S1: Obtain the alarm signal sequence of the power system within a preset time period.

[0024] Specifically, the first step is to establish a communication connection with the power system monitoring data source to receive various status data reported by underlying devices in real-time or near real-time. The power system is a complex network encompassing multiple stages, including power generation, transmission, substation, distribution, and consumption. It includes several secondary professional subsystems such as relay protection systems, integrated automation systems, network security monitoring systems, communication systems, and automatic safety devices. Acquiring alarm signal sequences is not simply data reading; it requires unified collection of multi-source heterogeneous data generated by these multiple secondary professional subsystems. The preset time period serves as a time window for data processing in support decision-making methods. This preset time period can be a fixed time interval or a dynamic window that slides forward over time. Within the preset time period, various sensors, protection devices, and monitoring backends in the power system will detect events such as voltage exceeding limits, sudden current changes, switch changes, communication interruptions, or device malfunctions, and define these events as alarm signals.

[0025] To facilitate subsequent feature space mapping processing, the collected discrete alarm data needs to be organized into an ordered set. Let the alarm signal sequence be denoted as... Then the alarm signal sequence It can be represented as containing An ordered set of alarm signal elements, i.e. ,in This indicates the total number of alarm signals collected within a preset time period. This indicates the alarm signal sequence arranged in chronological order. Each alarm signal. Each contains several key attribute fields, specifically represented as follows: ,in, Indicates the first The specific timestamp of each alarm signal is used to determine the temporal sequence of the alarm signals. Indicates the generation of the first Each alarm signal is a unique device identifier used to locate the source of the fault. Indicates the first The specific text description of each alarm signal, such as "Phase A current too high" or "Circuit breaker tripped." After obtaining the raw alarm data, the raw data is cleaned and standardized, invalid data missing key attributes is removed, and alarm description texts from different manufacturers and in different formats are converted into a unified encoding format. Finally, the data is sorted according to timestamps. Arrange all alarm signals in ascending order to form the alarm signal sequence described above. .

[0026] By acquiring the alarm signal sequence of the power system within a preset time period, fragmented power equipment status information can be transformed into a structured time-series data stream, thereby providing a comprehensive and accurate data foundation for subsequent extraction of alarm signatures and the use of historical cases for decision support.

[0027] It should be noted that, in response to the high-concurrency, instantaneous alarm storms that may occur in the power system under complex fault conditions such as bus faults, cascading trips, or grid oscillations, the steps for obtaining the alarm signal sequence of the power system within a preset time period include processing mechanisms for anti-interference and timing correction.

[0028] First, to address the interference caused by device contact jitter, the auxiliary decision-making method performs anti-jitter filtering operations for the same power equipment identifier. For state signals that frequently change within a preset, extremely short time threshold, the auxiliary decision-making method only extracts the identifier of the same power equipment. The final state after stabilization serves as a valid alarm signal. This eliminates false intermediate state signals.

[0029] Secondly, to address the timing disruptions caused by communication transmission under high concurrency, this invention does not rely on the reception time of alarm signals, but strictly adheres to the high-precision action time stamp carried in the original message. Reorder all concurrent signals.

[0030] If there are multiple alarm signals' action time markers Completely consistent, the action time stamp will be based on the preset device priority or channel logic. Multiple alarm signals are assigned micro-level indexes with completely consistent order, ensuring that the alarm signal sequence... In this context, even signals with the same physical time have a definite mathematical order.

[0031] Through the above processing, both sparse signals during stable operation and massive signals during fault outbreak are ultimately transformed into alarm signal sequences with uniform format, rigorous timing logic, and no ambiguity. This fully preserves the state snapshot of all power equipment during the fault, thus providing high-quality and interference-resistant underlying data support for subsequent alarm signature generation and large model inference.

[0032] Step S2: Perform feature space mapping processing on the alarm signal sequence to generate an alarm signature that represents the semantic features of the alarm signal sequence.

[0033] In a preferred embodiment, the alarm signal sequence is subjected to feature space mapping processing to generate an alarm signature for characterizing the semantic features of the alarm signal sequence, including: The alarm signal sequence is sliced ​​to generate several alarm segments; Numerical mapping is performed on each alarm segment to generate the basic values ​​for each alarm segment; Based on the base values ​​of each alarm segment, construct a set of basic values; For each hash function in the preset set of hash functions, hash calculation is performed on each basic value in the basic value set based on the current hash function to generate the current hash value set; Select the smallest hash value from the current set of hash values, and use it as the minimum hash value corresponding to the current hash function; The minimum hash values ​​are combined to generate an alarm signature that represents the semantic features of the alarm signal sequence.

[0034] In a preferred embodiment, the alarm signal sequence is sliced ​​to generate several alarm segments, including: Based on a preset sliding window and a preset sliding step size, the alarm signal sequence is traversed by sliding to determine the alarm subsequence covered by each sliding step; wherein, the alarm signal sequence is composed of several alarm signals; The alarm signals of each alarm subsequence are spliced ​​together in chronological order to generate the corresponding alarm segments.

[0035] In a preferred embodiment, each alarm segment is numerically mapped to generate a base value for each alarm segment, including: For each alarm segment, based on a preset numerical encoding algorithm, the current alarm segment is encoded to generate a current integer identifier value that uniquely identifies the current alarm segment; Use the current integer flag value as the base value for the current alarm segment.

[0036] Specifically, after obtaining the standardized alarm signal sequence, the variable-length, unstructured alarm signal sequence needs to be converted into a fixed-length mathematical vector suitable for fast computer comparison. Step S2 mainly adopts the MinHash technique in the Locality Sensitive Hash (LSH) algorithm to achieve dimensionality reduction and mapping of the feature space. The core objective of feature space mapping is to significantly reduce the data dimensionality while preserving the similarity features of the alarm signal sequence in terms of temporal structure and semantic content to the greatest extent. Step S2 first transforms the continuous alarm data stream into discrete fragments containing local contextual information through slicing, then transforms the text fragments into a set of numbers that can be computed by a computer through numerical mapping, and finally extracts a signature vector that can represent the features of the set through multi-hashing calculation.

[0037] In a preferred embodiment, the process of slicing the alarm signal sequence to generate several alarm segments employs an N-gram sliding window mechanism. The decision support method sets two key parameters: a preset sliding window length. and preset sliding step size Sliding window length This determines the number of alarm signals contained in each alarm segment, used to capture alarm concurrency or causal relationships within a local time range; sliding step size. This determines the magnitude by which the sliding window moves across the alarm signal sequence. The auxiliary decision-making method is based on the sliding window length. and sliding step size The alarm signal sequence is traversed sequentially by sliding window. During each slide, the sliding window covers a portion of the continuous signals in the alarm signal sequence, and the auxiliary decision-making method identifies the alarm subsequence covered by each slide as a processing unit. Subsequently, the text content or device identifiers of all alarm signals contained in the alarm subsequence are concatenated into strings strictly in chronological order to generate a corresponding alarm segment. In this way, the original alarm signal sequence is transformed into a set of several alarm segments, each containing the combined logic of device actions within a specific time window.

[0038] After generating several alarm segments, each segment is further mapped numerically to generate its base value. Since alarm segments are essentially string text, they cannot be directly hashed; therefore, they need to be mapped to an integer space. For each alarm segment, the auxiliary decision-making method calls a preset numerical encoding algorithm (such as CRC32, MurmurHash, or FNV) to encode the text content of the current alarm segment. The numerical encoding algorithm can convert a string input of arbitrary length into a fixed-digit integer output, which is the current integer identifier used to uniquely identify the current alarm segment. The auxiliary decision-making method directly uses the calculated current integer identifier as the base value of the current alarm segment. By performing the above operations on all alarm segments, the auxiliary decision-making method constructs a base value set containing several integer elements, denoted as set . .gather Each element in Each represents a semantic fingerprint of an alarm segment.

[0039] Based on the constructed base set of values, the core signature generation stage begins. To generate statistically significant alarm signatures, a pre-defined set of hash functions is used, which contains... Let be a set of distinct hash functions, denoted as . ,in The signature dimension is a preset value (e.g., 128 or 256). For each hash function in the preset set of hash functions... ( This represents the index of the hash function, with a value ranging from 1 to... The auxiliary decision-making method is based on the current hash function. For the basic set of values Each basic value in Perform hash calculations separately to obtain the result with the current hash function. The corresponding current set of hash values. Next, the auxiliary decision-making method iterates through the current set of hash values ​​and selects the hash value with the smallest value. This selected hash value with the smallest value is defined as the current hash function. The corresponding minimum hash value is denoted as The selection logic for the minimum hash value is based on the Jaccard Similarity theorem, which states that the probability that two sets have the same minimum hash value is equal to the Jaccard similarity between the two sets. The mathematical principle is shown in the following formula: in, Represents probability. This indicates the operation of selecting the minimum hash value. and These represent two different sets of basic values ​​(e.g., the current real-time alarm set and the historical alarm set). Represents a set With sets The similarity of Jaccard, Represents a set With sets The number of elements in the intersection. Represents a set With sets The number of elements in the union. The formula above shows that the similarity of the original sets can be estimated by comparing whether the minimum hash values ​​are equal.

[0040] Finally, the calculated The smallest hash values ​​are combined according to the preset order of the hash function to construct a hash function. A dimensional vector, which is the alarm signature used to characterize the semantic features of the alarm signal sequence, denoted as . .

[0041] Through the processing in step S2, massive and complex power alarm data is compressed into compact digital signatures. This alarm signature not only significantly reduces the overhead of data storage and transmission, but more importantly, it enables subsequent steps to complete similarity matching with a massive historical case database in a very short time, thereby significantly improving the real-time performance and accuracy of power system fault auxiliary decision-making.

[0042] For example, in a practical application scenario of power system alarm analysis, the preset sliding window length Preset sliding step size and the size of the preset hash function set (i.e., the dimension of the alarm signature) are three key hyperparameters that affect the quality of feature extraction.

[0043] In this embodiment, the preset sliding window length is set. The preferred setting is 3. Set the preset sliding window length. Setting the value to 3 is not an arbitrary choice, but rather based on an in-depth analysis of the operational behavior patterns of power system relay protection devices. In the power system's fault handling system, relay protection devices serve as the first line of defense for fault clearing, and their operational behavior forms the core framework of the alarm signal sequence. Typically, a complete fault clearing process presents a tightly linked "triple" characteristic at the alarm signal level. For example, in the event of a line short-circuit fault, the typical signal sequence is "protection device starting element operation" – "protection device output relay operation" – "circuit breaker position change"; or in the reclosing process, it presents a logical chain of "reclosing charging completed" – "reclosing started" – "reclosing output". Therefore, setting the sliding window length... Setting the window size to 3 ensures that each generated alarm segment mathematically covers exactly one complete physical action unit. This avoids both situations where the window is too small, causing logical chain breaks, and where the window is too large, allowing irrelevant noise signals to be mixed into the same feature segment. Simultaneously, to maximize the extraction of all potential feature patterns in the alarm signal sequence, the auxiliary decision-making method uses a preset sliding step size. Set to 1. This sets the preset sliding step size. Setting it to 1 means that the sliding window moves backward by only one alarm signal position at a time, thereby generating a set of alarm segments with extremely high overlap. This high-overlap slicing method ensures that even if there are tiny timing insertions or omissions in the alarm signal sequence, the generated feature set can still maintain high robustness, thus accurately mapping the discrete action behavior of the relay protection device into a digital signature with strong semantic features.

[0044] In the numerical mapping and hash calculation stages, to balance computational efficiency and the accuracy of similarity estimation, the auxiliary decision-making method sets the size of the pre-defined hash function set. Set to 200. Based on the mathematical properties of the MinHash algorithm, the expected error for similarity estimation is approximately... ,when When the value is 200, the theoretical estimation error can be controlled within 7%. For scenarios involving the retrieval of massive historical cases, 200 is an optimal balance point that balances retrieval speed and accuracy. The preset set of hash functions does not require writing 200 completely different algorithms; instead, it is constructed based on the common MurmurHash3 or CityHash algorithms, using 200 distinct prime numbers as random seeds. Specifically, the set of hash functions can be represented as... .

[0045] Based on the above parameter settings, the following specific example demonstrates the entire feature generation process. Assume a simplified alarm signal sequence is obtained. ,in to These represent "differential protection action", "outlet pressure plate engagement", "circuit breaker position change", and "reclosing action", respectively. The auxiliary decision-making method first uses a sliding window of length 3 and step size 1 to analyze the alarm signal sequence. Slice the data to generate the first alarm segment. (Include (a combination) and the second alarm segment (Include (A combination of [various methods]). Subsequently, the CRC32 algorithm is used to separate the alarm fragments. Mapping to basic numerical values (For example, the integer 3948201), will trigger an alarm segment. Mapping to basic numerical values (For example, the integer 5920183), thereby constructing a basic set of values. Next, using 200 pre-defined hash functions, the basic set of values ​​is processed respectively. Perform the calculation. For the first hash function... ,calculate and The smaller of the two values ​​is taken as the value of the first dimension of the signature vector; for the second hash function... ,calculate and The smaller of the two values ​​is taken as the value of the second dimension of the signature vector; this process is repeated until the 200th hash function is calculated. Finally, these 200 smallest hash values ​​are concatenated in order to generate the sequence used to represent the alarm signal. 200-dimensional alarm signature.

[0046] Through the specific parameter settings and implementation steps described above, the abstract power business logic can be transformed into a precise mathematical calculation process, ensuring that the generated alarm signature not only contains the physical laws of power faults but also has the computational characteristics of millisecond-level retrieval in a large-scale knowledge base.

[0047] In one specific embodiment, for each alarm segment, based on a preset numerical encoding algorithm, the current alarm segment is encoded to generate a current integer identifier value that uniquely identifies the current alarm segment: First, regarding the composition of alarm segments, in the slicing process mentioned in step S2, each generated alarm segment is not a simple stack of raw messages, but a string sequence defined by a standardized format. An alarm segment is composed of key attribute fields of several alarm signals covered by a sliding window, arranged in chronological order. To ensure that the alarm segment can represent the physical logic pattern of the fault without being affected by the absolute time of the fault occurrence (i.e., ensuring that a "trip" that occurred yesterday and a "trip" that occurred today can be identified as having the same characteristics), when constructing the alarm segment, the "device unique identifier" and "standard semantic text" of each alarm signal are extracted and connected using preset field separators (such as underscores "_" or vertical bars "|").

[0048] For example, suppose the sliding window length is 2, and the current sliding window covers two consecutive alarm signals: the first signal is a "differential protection action (Prot_Act)" issued by "Main Transformer No. 1 (Dev_001)," and the second signal is a "circuit breaker trip (Brk_Trip)" issued by "Main Transformer No. 1 (Dev_001)." The auxiliary decision-making method concatenates these two signals into a standardized text format string, that is, the content of the current alarm fragment is represented as: "Dev_001_Prot_Act|Dev_001_Brk_Trip." Through this concatenation method, the alarm fragment completely retains all semantic information of "which device" and "what action sequence occurred," while masking the interference of absolute timestamps, making the text strings generated by the same fault mode occurring at different times completely consistent.

[0049] After determining the text composition of the alarm fragment, a numerical mapping operation is performed. Since the MinHash algorithm and subsequent similarity calculations are based on integer mathematical operations, the variable-length text string needs to be converted into a fixed-length integer. In this embodiment, the preferred numerical encoding algorithm is the Cyclic Redundancy Check (CRC32) algorithm. The CRC32 algorithm is characterized by its fast computation speed, uniform hash distribution, and extremely low collision probability, making it very suitable for mapping short texts to integer indices.

[0050] In practice, the current alarm segment (e.g., the string "Dev_001_Prot_Act|Dev_001_Brk_Trip") is passed as input to the CRC32 function. The CRC32 function performs polynomial division modulo 2 on the byte stream of the input string, ultimately outputting a 32-bit unsigned integer (e.g., "35982174"). This 32-bit unsigned integer is the current integer identifier used to uniquely identify the current alarm segment. This current integer identifier is directly used as the base value for the current alarm segment.

[0051] If the alarm signal sequence contains multiple different alarm segments, the above process is repeated for each alarm segment. For example, for another alarm segment "Dev_002_OverCurrent|Dev_002_Warning", the integer identifier value "10293847" is generated after CRC32 operation. Through this deterministic numerical encoding algorithm, the originally complex and unstructured semantic description of power faults is accurately mapped to basic values ​​that occupy very little space in computer memory (usually only 4 bytes). This not only significantly reduces the memory overhead of subsequent storage and retrieval, but also ensures that the same fault mode has a unique and deterministic digital fingerprint in the feature space, thereby improving the accuracy of auxiliary decision-making.

[0052] In one specific embodiment, the preset set of hash functions is constructed in the following way: In the implementation of the MinHash algorithm, the preset set of hash functions is not composed of two hundred completely different algorithm logics (such as a simple stack of MD5, SHA-1 or SHA-256), but is a family of linear transformation functions generated by changing the internal operation coefficients of the same family of universal hash algorithms.

[0053] Specifically, a linear congruence formula is used to construct a predefined set of hash functions. Assume the size of the predefined set of hash functions (i.e., the signature dimension) is... (For example Then the set of hash functions is represented as Each hash function in the set ( From 1 to Each integer is composed of two unique random coefficients. and and a fixed maximum prime number Common definition. (The first...) The mathematical expression of a hash function is shown in the following formula: in, This represents the basic input value (i.e., the integer identifier value generated by the alarm segment in step S2). Indicates the first Each hash function is used to evaluate the underlying value. The hash value obtained after performing the calculation; Indicates the first The multiplication coefficients of a hash function It is in the interval A randomly selected integer; Indicates the first The addition coefficients of a hash function, It is in the interval A randomly selected integer; It represents a pre-defined maxima prime (Mersenne Prime). The value must be greater than the base value. The maximum possible value is determined to avoid hash collisions. In this embodiment, since the base value is a 32-bit unsigned integer, The preferred setting is a Mersenne prime. or greater than The smallest prime number (e.g., 4294967311).

[0054] During the stage of constructing the hash function set, it is pre-generated during system initialization. Groups of distinct coefficient pairs For example, for the first hash function Generation coefficient pair For the second hash function Generation coefficient pair And so on, until generated Group coefficients. This A fixed set of coefficient pairs constitutes the kernel of a pre-defined hash function set. Because of the coefficients... and It is random and independent, which statistically guarantees the validity of each hash function. The permutation rules for the input data are independent of each other, thus ensuring that each dimension of the MinHash signature can independently capture the characteristics of the set.

[0055] Based on the hash function set constructed above, specific hash calculation steps are performed. For each hash function in the preset hash function set... Based on the current hash function Corresponding coefficients and For the basic set of values (Include Basic values The hash function is performed on each of the basic values ​​in the dataset. That is, each value is hashed separately. , ,..., The calculated set of values This constitutes the current hash value set. Then, the minimum value is selected from the current hash value set and used as the alarm signature vector at the th position. Dimensional components.

[0056] Through this linear transformation construction method, the present invention can quickly generate high-dimensional independent hash maps with extremely low computational cost (involving only multiplication, addition and modulo operations), effectively solving the problem of excessive computation time of traditional cryptographic hash algorithms and significantly improving the generation efficiency of alarm signatures.

[0057] Step S3: Based on the preset historical case knowledge base, calculate the similarity between the alarm signature and the historical alarm signatures of each historical case in the historical case knowledge base, and generate a similarity score for each historical case.

[0058] In a preferred embodiment, based on a preset historical case knowledge base, the similarity between the alarm signature and the historical alarm signatures of each historical case in the historical case knowledge base is calculated, and a similarity score for each historical case is generated, including: For each historical case in the preset historical case knowledge base, the alarm signature is compared with the current historical alarm signature at the corresponding position to determine the number of elements whose alarm signature and current historical alarm signature values ​​are consistent. The number of elements whose alarm signature matches the current historical alarm signature value is taken as the number of matching features of the current historical alarm signature; Calculate the ratio of the number of matching features in the current historical alarm signature to the total number of elements contained in the alarm signature, and generate the similarity score of the current historical case.

[0059] Specifically, after generating the alarm signature, the core retrieval and matching stage begins. A pre-set historical case knowledge base forms the basis for the reasoning in the decision-making process. This knowledge base stores a large number of cleaned and labeled historical power fault cases. To ensure that the real-time generated alarm signature can be effectively compared with historical data, each historical case in the knowledge base has been preprocessed before being added to the database according to the same feature space mapping logic as step S2 (including the same slicing rules, the same numerical encoding algorithm, and the same set of pre-set hash functions), thus pre-generating and storing the corresponding historical alarm signature. The data structure of each historical case contains at least two parts: a historical alarm signature vector for retrieval and matching, and historical handling information (such as fault cause, operation records, and handling suggestions) for subsequent decision generation. Step S3 aims to quickly locate the past case most closely related to the current alarm scenario from the massive historical data by calculating the distance or similarity between mathematical vectors.

[0060] Based on a pre-defined historical case knowledge base, the similarity between the alarm signature and the historical alarm signatures of each historical case in the knowledge base is calculated, generating a similarity score for each historical case. An unbiased Jaccard similarity estimation method based on MinHash signatures is employed. First, the pre-defined historical case knowledge base is traversed, and the historical alarm signature corresponding to each historical case is obtained one by one. Assume the current alarm signature is denoted as a vector. The historical alarm signature of a given historical case currently being traversed is denoted as a vector. And both vectors have dimensions of 1. (That is, the size of the preset hash function set, such as 200). Sign the alarm signature. Current historical warning signature Perform a numerical comparison at the corresponding positions. Here, "corresponding positions" refers to the one-to-one correspondence between the hash function index order, i.e., the alarm signature. From the middle The minimum hash value generated by each hash function must be compared with the historical alarm signature. From the middle The comparison is performed using the smallest hash value generated by each hash function, but out-of-order comparisons are not allowed.

[0061] During the comparison process, an indicator function is introduced to determine whether two elements are identical, and the number of elements whose values ​​at the same dimension index are completely identical to the current historical alarm signature is counted. This number of elements with identical values ​​is defined as the matching feature count of the current historical alarm signature, denoted as . Subsequently, the auxiliary decision-making method calculates the number of matching features for the current historical alarm signature. The total number of elements contained in the alarm signature The ratio is calculated and directly used as the similarity score for the current historical case. The calculation process can be expressed by the following mathematical formula: in, This represents the similarity score of the current historical case, ranging from 0 to 1; This indicates the total number of elements contained in the alarm signature, which is also the total number of hash functions; This represents the dimension index of the current traversal, with values ​​ranging from 1 to... ; This indicates that the current alarm signature is at the [number]th [position]. The values ​​of each dimension; This indicates the current historical alert signature is at the [number]th [number]. The values ​​of each dimension; This indicates an indicator function; the function value is 1 when the condition within the parentheses is true (i.e., the two values ​​are equal), and 0 otherwise. Based on the theoretical foundation of the MinHash algorithm, the above formula calculates the similarity score. It is an unbiased estimate of the Jaccard similarity between the current real-time alarm signal sequence and the historical case alarm signal sequence, that is, it reflects the proportion of overlapping features in the two sets.

[0062] The above comparison and calculation process is repeated for each historical case in the preset historical case knowledge base, thereby generating a corresponding similarity score for each historical case.

[0063] Through the processing in step S3, the complex text semantic matching problem is transformed into a simple integer vector bit operation problem. This similarity calculation method based on MinHash signature does not require complex matrix multiplication or deep learning inference, but only linear numerical comparison. Therefore, it has extremely high computational efficiency and can accurately quantify the similarity to the current fault scenario from thousands of historical cases, providing a reliable quantitative basis for subsequent selection of high-quality reference cases.

[0064] It's important to note that the pre-defined historical case knowledge base is not a simple log file, but a cleaned, labeled, and vectorized structured database. Logically, the pre-defined historical case knowledge base is divided into a vector index area for fast similarity matching and a details storage area for storing detailed text content. The vector index area primarily stores mathematically reduced features, while the details storage area primarily stores human-readable natural language descriptions.

[0065] Specifically, each historical case in the pre-defined historical case knowledge base is defined as a structured data unit containing multi-dimensional attributes. Each historical case contains at least four core fields: a unique case identifier, a historical alarm signature vector, the original alarm signal sequence, and historical handling information.

[0066] First, the unique identifier for a case is the primary key used to uniquely index the current historical case in the pre-defined historical case knowledge base (e.g., "CASE_20241012_001").

[0067] Secondly, the historical alarm signature vector is the mathematical expression of each historical case in the feature space. The historical alarm signature vector is a MinHash signature pre-calculated and stored using the same feature space mapping method as step S2 (i.e., the same slicing rules, numerical encoding algorithm, and hash function set). Storing the pre-calculated historical alarm signature vector can avoid repeated hash operations during the retrieval stage, thereby significantly improving the retrieval speed.

[0068] Furthermore, the original alarm signal sequence records a snapshot of the actual underlying device messages when the historical case occurred, which can be used for backtracking verification in subsequent steps or as the original input for comparison in the large model prompt words.

[0069] Finally, historical handling information is the most valuable knowledge carrier in each historical case. This information specifically covers three dimensions: fault assessment conclusions, standard handling procedures, and on-site feedback records. Fault assessment conclusions are a qualitative description of the nature of the fault at the time (e.g., "The differential protection of main transformer No. 1 tripped, and it was found to be a short circuit between turns in phase B winding"); standard handling procedures record a series of standardized operating steps taken by the dispatcher or maintenance personnel at the time (e.g., "Disconnect the three switches on the main transformer," "Open the neutral grounding switch," "Activate the standby power supply automatic transfer device"); and on-site feedback records include equipment status confirmation information after fault isolation and subsequent maintenance suggestions.

[0070] Through this structured data organization method, the pre-set historical case knowledge base not only provides vector features to support retrieval in step S3, but also provides rich and logically rigorous background context materials for step S5, ensuring that the generated auxiliary decision-making information has both mathematical similarity basis and business processing logic support.

[0071] Step S4: Select a preset number of historical cases with similarity scores higher than a preset similarity threshold as target reference cases.

[0072] Specifically, after generating similarity scores for each historical case in the historical case knowledge base, it is necessary to select the most valuable subset of cases from the entire historical case library for use in subsequent large-scale model inference. The core logic of step S4 lies in constructing a high-quality set of target reference cases through a dual screening mechanism combining "quality threshold filtering" and "quantity truncation control." The preset similarity threshold is a criterion used to measure whether the correlation between historical cases and the current alarm signal sequence meets the standard. The preset similarity threshold is usually set based on the statistical characteristics of the MinHash algorithm and the need to balance recall and precision (e.g., set to 0.5 or 0.6). The preset quantity is an integer limit value used to control the length of the context input to the generative decision model (e.g., set to Top-3 or Top-5). The preset quantity aims to prevent the model's context window from being exceeded due to too many reference cases or the introduction of irrelevant noise information.

[0073] First, a threshold filtering operation is performed. All similarity scores calculated in step S3 are iterated through, and the similarity score of each historical case is compared with a preset similarity threshold. If the similarity score of a historical case is less than or equal to the preset similarity threshold, the auxiliary decision-making method considers the historical case insufficiently relevant to the current power failure and directly removes it; if the similarity score of a historical case is greater than the preset similarity threshold, the historical case is retained and added to a temporary candidate case set. This process can be described using mathematical set theory as follows: Assume the historical case knowledge base is a set. For the set of the first Historical Cases , No. The similarity score for each historical case is: The preset similarity threshold is Then the set of candidate cases after filtering Defined as: in, This represents the set of all candidate cases that satisfy the threshold condition. This indicates the first case in the historical case knowledge base. A historical case, Indicates the first The similarity score of each historical case This indicates the preset similarity threshold.

[0074] After completing the candidate case set Next, a sorting and truncation operation based on similarity scores is performed on the candidate case set. All historical cases are sorted in descending order of similarity score from highest to lowest. After sorting, the candidate case set is checked. The total number of elements in the candidate case set. The total number of elements in the array is less than the preset number (denoted as ). ), to set up candidate cases All historical cases in the candidate case set are directly identified as target reference cases; The total number of elements in the database is greater than or equal to the preset number. The decision support method only selects those ranked first. Historical cases were used as target reference cases, and those ranked in [position missing] were discarded. The historical cases following this position. The final selected target reference cases will collectively constitute the target reference case set.

[0075] The dual screening mechanism in step S4 effectively eliminates interference cases with low similarity and ensures that the reference information submitted to the generative decision model is a high-quality sample that is highly relevant to the current power system alarm scenario. This significantly reduces the computational load and contextual interference of the subsequent generative model while ensuring the accuracy of the auxiliary decision.

[0076] Step S5: Aggregate the historical handling information in each target reference case to generate background context information; construct prompt information based on the background context information and the alarm signal sequence.

[0077] In a preferred embodiment, historical handling information in each target reference case is aggregated to generate background context information, including: For each target reference case, based on the historical handling information contained in the current target reference case and the preset case guidance words, a standard reference text for the current target reference case is generated. Based on the similarity scores of each target reference case arranged from high to low, and using a preset delimiter, the standard reference texts are sequentially concatenated to generate background context information.

[0078] Specifically, after the target reference case is determined in step S4, the PromptEngineering construction phase begins. The core task of step S5 is to transform the discrete, structured database records into a natural language text stream that the generative decision-making model can understand and process.

[0079] In a preferred embodiment, knowledge integrity verification and completion operations are performed first. Although step S4 has eliminated cases with low similarity, this may result in too few target reference cases remaining, or even an empty target reference case set in extreme cases (e.g., when facing a completely new type of fault). To prevent the generative decision-making model from developing illusions or lacking sufficient reasoning basis due to sparse reference information, the auxiliary decision-making method counts the total number of elements in the target reference case set output by step S4, denoted as . The decision support method considers the total number of elements. The preset quantity mentioned in step S4 Compare. If the total number of elements... Less than the preset quantity It determines that the current reference information is insufficient and initiates a general knowledge completion strategy.

[0080] Specifically, no additional external retrieval is required; instead, pre-defined general handling principles stored locally are retrieved directly. These pre-defined general handling principles are standardized guidelines applicable to most fault scenarios in power systems (e.g., "1. Prioritize personal safety; 2. Isolate faulty equipment; 3. Keep fault-free parts operational"). The number of gaps that need to be filled is then calculated. ( Copy the preset general handling principles. Alternatively, the preset general handling principles can be used as a single supplementary reference information and combined with the target reference cases determined in step S4 to form the final set of reference information. This is for extreme scenarios where there are absolutely no historical reference cases (i.e.,...). This strategy ensures that the context input to the generative decision model includes at least the bottom-line safety rules for power dispatch, thereby guiding the generative decision model to generate conservative and safe disposal recommendations based on general logic.

[0081] After determining the reference information set containing the target reference case and supplementary reference information, standardization and aggregation operations are performed. For each piece of information in the reference information set, it is encapsulated based on a preset guide word template. If the current information originates from a historical case, a guide word such as "[Historical Reference]:" is used; if the current information originates from a preset general handling principle, a guide word such as "[General Principle]:" is used. The decision support method concatenates the guide word with the corresponding information content to generate standard reference text. Subsequently, following the arrangement logic of "prioritizing similar historical cases and supplementing with general principles," and combining the similarity scores, all standard reference texts are sorted. Based on preset delimiters (such as the newline character "\n" or a special delimiter Token), the sorted standard reference texts are concatenated sequentially to generate a long text string, i.e., the background context information.

[0082] Finally, based on the background context information and the alarm signal sequence obtained in step S1, the final prompt information input to the model is constructed. The process of constructing the prompt information can be described as a structured combination of role setting instructions, background context information, and current input data. To clearly express the composition logic of the prompt information, the decision support method adopts the following prompt information construction formula: in, This indicates the final build prompt; This represents a preset role setting instruction string, used to define the identity of the generative decision model (e.g., "You are a professional power dispatcher"). Indicates the preset section separator; This indicates the background context information generated above, which is a long text that aggregates multiple historical cases; A text description representing the current sequence of alarm signals to be processed; symbols This represents the ordered concatenation operation of strings.

[0083] Through the processing in step S5, unstructured historical experience and real-time alarm data are integrated into standardized instructions that conform to the input specifications of generative decision-making models. This not only provides the model with rich analogical reasoning materials, but also ensures that the generated auxiliary decision-making information is highly professional through a strict sorting and separation mechanism.

[0084] Step S6: Input the prompt information into a preset generative decision model, so that the generative decision model generates auxiliary decision information for the alarm signal sequence based on the prompt information; wherein, the generative decision model is trained by several training samples; each training sample includes historical prompt information and corresponding auxiliary decision labels.

[0085] In a preferred embodiment, the generative decision model is trained in the following manner: Obtain several training samples; Each training sample is sequentially input into the generative decision model to train it until a preset number of training iterations are reached. Each time a training sample is received, the generative decision model outputs the corresponding predictive auxiliary decision information. A loss function value is calculated based on the predictive auxiliary decision information and the corresponding auxiliary decision label. The generative decision model is then updated based on the loss function value.

[0086] Specifically, after constructing the prompt information containing background context information and the current alarm signal sequence, the system enters the final inference generation stage. The preset generative decision model is a large-scale pre-trained language model (LLM) based on the Transformer architecture, which has powerful natural language understanding and logical reasoning capabilities.

[0087] In step S6, the prompt information constructed in step S5 is input as a sequence into a preset generative decision-making model. The preset generative decision-making model is based on an auto-regressive generation mechanism, utilizing an attention mechanism to deeply capture the semantic relationships between historical cases and the current fault in the prompt information. Based on the learned probability distribution, the preset generative decision-making model predicts and generates tokens in the output sequence one by one until an end-of-sequence symbol is generated. The final generated text sequence is the auxiliary decision-making information for the alarm signal sequence. The auxiliary decision-making information includes a comprehensive assessment of the nature of the current fault, suggested troubleshooting steps (e.g., "check the pressure of the XX switch chamber"), and safety precautions for emergency handling.

[0088] In a preferred embodiment, to ensure that the pre-defined generative decision-making model can adapt to the professional context of power dispatching and follow specific instruction formats, it is trained using supervised fine-tuning (SFT). The training process first involves constructing training samples. The auxiliary decision-making method acquires several original historical fault records. For each historical fault record, the auxiliary decision-making method generates corresponding historical prompt information using logic completely consistent with steps S1 to S5 (i.e., feature mapping, similarity retrieval, context aggregation, and prompt word construction), and uses the correct handling solution given by the expert at that time as the corresponding auxiliary decision-making label. Each training sample constructed in this way maintains a high degree of consistency between its data distribution and the data distribution encountered by the model during actual inference, thereby ensuring the effectiveness of the training.

[0089] After acquiring several training samples, an iterative training process is performed. Each training sample is sequentially input into the generative decision model. During each forward propagation, the generative decision model receives historical prompts and outputs predictive auxiliary decision information for that sample (i.e., the probability distribution of the next word predicted by the model). Based on the predictive auxiliary decision information and the corresponding auxiliary decision label (i.e., the ground truth), the difference between the model output and the ground truth label is calculated using the cross-entropy loss function, i.e., the loss function value is calculated. To quantify this difference, the auxiliary decision method uses the following loss function formula: in, This represents the calculated loss function value; The set of model parameters representing a generative decision-making model; Indicates the total length of the auxiliary decision label sequence (i.e., the total number of tokens); Indicates the index of the currently calculated sequence position; The label indicates the decision support label in the first place. The real word element at each position; The label indicates the decision support label in the first place. All lexical units preceding each position; This indicates the input history prompts; This indicates the input history prompt information. and pre-order morphemes Under the condition of [condition], the generative decision model predicts the [condition]. Each lexical unit is a real lexical unit. The conditional probability.

[0090] After calculating the loss function value, perform the backpropagation operation to calculate the loss function value. Relative to model parameters The gradient is then calculated. Subsequently, a pre-defined optimization algorithm (such as the AdamW optimizer or stochastic gradient descent (SGD) is used to adjust the model parameters of the generative decision model based on the calculated gradient. Update the function to minimize the loss function value. Repeat the above input, calculation, and update steps until the preset number of training epochs is reached or the loss function value converges to a preset threshold range.

[0091] Through the above training and inference process, not only is the general large model endowed with professional decision-making capabilities in the vertical field of power, but it also ensures that the model can fully understand and utilize the background context information provided by the RAG retrieval enhancement framework, thereby generating logically rigorous, well-founded, and power safety-compliant auxiliary decision-making suggestions.

[0092] It should be noted that in this embodiment, the pre-defined generative decision model is built based on a Transformer decoder-only architecture. The generative decision model mainly consists of an input embedding layer, a positional encoding layer, several stacked decoder blocks, and an output linear layer. The input embedding layer converts text data (i.e., prompt information) from the power system domain into a distributed representation in a high-dimensional vector space. The stacked decoder blocks extract long-range semantic dependency features; each decoder block includes a multi-head self-attention sublayer and a feed-forward network sublayer, and each sublayer is followed by a layer normalization operation and a residual connection. The output linear layer maps the hidden layer features back to the vocabulary space to calculate the probability distribution of the next predicted word. Specifically, the parameter size of the generative decision model is preferably set between 7B and 13B, the hidden layer dimension is set to 4096, the number of attention heads is set to 32, and the number of layers in the decoder module is set to 32. This architecture configuration can meet the computing power constraints of local deployment in substations while ensuring a deep understanding of power industry expertise.

[0093] Regarding the integration of the model with specific domains and the input / output settings, a dedicated tokenization process is used to establish an intrinsic connection between alarm signals and model computations. The decision support method pre-defines a specialized vocabulary for the power dispatching vertical domain. This vocabulary, in addition to general natural language terms, expands to include professional terms such as primary power equipment names, secondary protection action terms, and dispatching operation instructions. In step S6, the decision support method first uses a byte-pair encoding (BPE) algorithm to segment the prompt information constructed in step S5 (including role settings, background context, and the current alarm sequence) into a token sequence. Subsequently, the token sequence is mapped to an input tensor. ,in For sequence length, For the embedding dimension. Generative decision models receive input tensors. The corresponding logits vector is output. A Softmax normalization operation is performed on the logits vector to obtain the generation probability distribution for power fault handling suggestions. A greedy search or nucleus sampling strategy is used to decode the auxiliary decision-making information text for the alarm signal sequence from the probability distribution.

[0094] Regarding the training steps and parameter settings of the model, a supervised fine-tuning (SFT) strategy was adopted to enable the generative decision-making model to possess professional decision-making capabilities in the power field. The specific training steps are as follows: First, a command fine-tuning dataset is constructed. Each sample in the dataset consists of an "input command (Prompt)" and an "expert response (Completion)." The "input command" is constructed strictly according to the structure described in step S5 (i.e., including the historical case context retrieved by RAG), and the "expert response" is derived from standard handling records in historical real accident reports. Second, training hyperparameters are set. The AdamW algorithm is selected as the optimizer, and the learning rate is set to [value missing]. The learning rate scheduling strategy employs cosine annealing, with a batch size of 64 and 3 epochs. During training, the auxiliary decision method uses the cross-entropy loss function to calculate the difference between the model's predicted output and the "expert answer," calculating only the loss value for the "expert answer" portion (loss masking) and omitting the loss value for the "input instruction" portion. Model parameters are updated using backpropagation until the loss function value converges to below 0.1.

[0095] Through the model architecture, domain data mapping logic, and training parameter configuration clearly described above, it is ensured that the generative decision model can accurately understand the complex alarm signal sequence input and generate auxiliary decision information that conforms to power safety regulations based on the retrieval-enhanced context.

[0096] It should be noted that this specification provides the following specific embodiments regarding the processing logic and output results when facing entirely new fault types that have never occurred in the power system. Assume that the power system experiences a rare fault caused by a novel network attack leading to logical chaos in secondary equipment, or a complex inter-regional oscillation fault caused by extreme weather. In this scenario, the alarm signal sequence obtained in step S1 contains a set of extremely unique signal combinations that have never been recorded in the preset historical case knowledge base. When executing steps S3 and S4, the similarity score between the alarm signature corresponding to this alarm signal sequence and all historical alarm signatures in the historical case knowledge base is calculated. Since this fault type is entirely new, the highest calculated similarity score (e.g., 0.15) is far below the preset similarity threshold (e.g., 0.6), resulting in an empty target reference case set output in step S4. In this situation, the general knowledge completion strategy is automatically triggered in step S5. After identifying that the number of target reference cases is insufficient, the preset general handling principles stored locally are directly called. These include standardized instructions such as prioritizing personal safety, immediately stopping related switching operations, isolating the fault area to prevent the accident from escalating, and strictly prohibiting blindly attempting to energize. The preset general handling principles are then used to construct background context information. Combined with the role setting instruction of "Current fault characteristics are unknown, please provide conservative suggestions based on general safety specifications", the final prompt message is generated.

[0097] Subsequently, in step S6, a prompt containing general handling principles is input into the preset generative decision-making model. Since the preset generative decision-making model has learned during the training phase how to follow general safety standards when specific reference cases are lacking, the generative decision-making model will output a conservative auxiliary decision-making information. Specifically, the content of this conservative auxiliary decision-making information will first clearly indicate in the fault analysis section that an abnormal signal combination pattern has been detected and no matching historical similar cases have been found, and it is determined to be an unknown type of complex fault or a potential atypical anomaly; secondly, in the risk warning section, it will emphasize that the current equipment status is unclear, and there is a risk of protection malfunction or failure to operate, and blind operation may lead to the expansion of the accident scope; finally, in the recommended handling strategy section, it will give suggestions on maintaining the status quo, fault isolation, and manual verification. For example, it is recommended to immediately suspend all ongoing automatic reclosing or remote control operations, switch the physical network elements involved in the alarm to cold standby status and disconnect the relevant protection output pressure plate, and at the same time assign maintenance personnel with professional testing instruments to go to the site to focus on checking the secondary circuit and communication module, clearly indicating that the fault has exceeded the confidence interval of the automatic decision-making system, and requesting the dispatcher to intervene immediately for manual analysis. As can be seen from the above embodiments, even when faced with a completely new fault that has never been seen before, the present invention will not output incorrect specific instructions due to the inability to find cases. Instead, it can output compliant, safe and conservative macro-level guidance suggestions based on preset general principles, ensuring the bottom-line safety of the auxiliary decision-making system under extreme conditions.

[0098] Based on the above method embodiments, the present invention provides corresponding apparatus embodiments.

[0099] like Figure 2 As shown, an embodiment of the present invention provides an auxiliary decision-making device based on power system alarm signals, including: a data acquisition module, a target reference case determination module, a prompt information determination module, and an auxiliary decision-making module; The data acquisition module is used to acquire the alarm signal sequence of the power system within a preset time period; The target reference case determination module is used to perform feature space mapping processing on the alarm signal sequence to generate an alarm signature that represents the semantic features of the alarm signal sequence; based on a preset historical case knowledge base, it calculates the similarity between the alarm signature and the historical alarm signatures of each historical case in the historical case knowledge base to generate a similarity score for each historical case; and selects a preset number of historical cases with similarity scores higher than a preset similarity threshold as target reference cases. The prompt information determination module is used to aggregate and process the historical handling information in each target reference case to generate background context information; and to construct prompt information based on the background context information and the alarm signal sequence. The auxiliary decision-making module is used to input the prompt information into a preset generative decision-making model, so that the generative decision-making model generates auxiliary decision-making information for the alarm signal sequence based on the prompt information; wherein, the generative decision-making model is trained by several training samples; each training sample includes historical prompt information and corresponding auxiliary decision-making labels.

[0100] It should be noted that the embodiments of the device described above correspond to the embodiments of the present invention described above, and can realize the auxiliary decision-making method based on power system alarm signals described in any one of the above embodiments of the present invention. Furthermore, the embodiments of the device described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the accompanying drawings of the device embodiments provided by the present invention, the connection relationship between modules indicates that they have a communication connection, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without creative effort.

[0101] Based on the above-described method embodiments of the present invention, a corresponding embodiment of an electronic device is provided.

[0102] An embodiment of the present invention provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the auxiliary decision-making method based on power system alarm signals as described in any one of the present invention, or the processor executes the computer program to implement the functions of each module in the above-described device embodiments.

[0103] For example, the computer program may be divided into one or more modules, which are stored in the memory and executed by the processor to perform the present invention. The one or more modules may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the terminal device.

[0104] The terminal device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0105] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.

[0106] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the terminal device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function, etc.; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital card (SD card), flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0107] Based on the above method embodiments, the present invention provides corresponding storage medium embodiments; Another embodiment of the present invention provides a storage medium including a stored computer program, wherein, when the computer program is executed, the device where the storage medium is located executes any of the above-described auxiliary decision-making methods based on power system alarm signals of the present invention.

[0108] The aforementioned storage medium is a computer-readable storage medium, and the computer program includes computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0109] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0110] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A decision-making auxiliary method based on power system alarm signals, characterized in that, include: Acquire the alarm signal sequence of the power system within a preset time period; The alarm signal sequence is subjected to feature space mapping processing to generate an alarm signature that characterizes the semantic features of the alarm signal sequence; Based on a pre-set historical case knowledge base, the similarity between the alarm signature and the historical alarm signatures of each historical case in the historical case knowledge base is calculated, and a similarity score for each historical case is generated. Select a preset number of historical cases with similarity scores higher than a preset similarity threshold as target reference cases; The historical handling information in each target reference case is aggregated and processed to generate background context information; Based on the background context information and the alarm signal sequence, a prompt message is constructed; The prompt information is input into a preset generative decision model, so that the generative decision model generates auxiliary decision information for the alarm signal sequence based on the prompt information; wherein, the generative decision model is trained by several training samples; each training sample includes historical prompt information and corresponding auxiliary decision labels.

2. The auxiliary decision-making method based on power system alarm signals as described in claim 1, characterized in that, The alarm signal sequence is subjected to feature space mapping processing to generate an alarm signature for characterizing the semantic features of the alarm signal sequence, including: The alarm signal sequence is sliced ​​to generate several alarm segments; Numerical mapping is performed on each alarm segment to generate the basic values ​​for each alarm segment; Based on the base values ​​of each alarm segment, construct a set of basic values; For each hash function in the preset set of hash functions, hash calculation is performed on each basic value in the basic value set based on the current hash function to generate the current hash value set; Select the smallest hash value from the current set of hash values, and use it as the minimum hash value corresponding to the current hash function; The minimum hash values ​​are combined to generate an alarm signature that represents the semantic features of the alarm signal sequence.

3. The auxiliary decision-making method based on power system alarm signals as described in claim 2, characterized in that, The alarm signal sequence is sliced ​​to generate several alarm segments, including: Based on a preset sliding window and a preset sliding step size, the alarm signal sequence is traversed by sliding to determine the alarm subsequence covered by each sliding step; wherein, the alarm signal sequence is composed of several alarm signals; The alarm signals of each alarm subsequence are spliced ​​together in chronological order to generate the corresponding alarm segments.

4. The auxiliary decision-making method based on power system alarm signals as described in claim 3, characterized in that, Numerical mapping is performed on each alarm segment to generate the basic values ​​for each alarm segment, including: For each alarm segment, based on a preset numerical encoding algorithm, the current alarm segment is encoded to generate a current integer identifier value that uniquely identifies the current alarm segment; Use the current integer flag value as the base value for the current alarm segment.

5. The auxiliary decision-making method based on power system alarm signals as described in claim 4, characterized in that, Based on a pre-defined historical case knowledge base, the similarity between the alarm signature and the historical alarm signatures of each historical case in the historical case knowledge base is calculated, and a similarity score for each historical case is generated, including: For each historical case in the preset historical case knowledge base, the alarm signature is compared with the current historical alarm signature at the corresponding position to determine the number of elements whose alarm signature and current historical alarm signature values ​​are consistent. The number of elements whose alarm signature matches the current historical alarm signature value is taken as the number of matching features of the current historical alarm signature; Calculate the ratio of the number of matching features in the current historical alarm signature to the total number of elements contained in the alarm signature, and generate the similarity score of the current historical case.

6. The auxiliary decision-making method based on power system alarm signals as described in claim 5, characterized in that, Historical handling information from each target reference case is aggregated to generate background context information, including: For each target reference case, based on the historical handling information contained in the current target reference case and the preset case guidance words, a standard reference text for the current target reference case is generated. Based on the similarity scores of each target reference case arranged from high to low, and using a preset delimiter, the standard reference texts are sequentially concatenated to generate background context information.

7. The auxiliary decision-making method based on power system alarm signals as described in claim 6, characterized in that, The generative decision-making model can be trained in the following way: Obtain several training samples; Each training sample is sequentially input into the generative decision model to train it until a preset number of training iterations are reached. Each time a training sample is received, the generative decision model outputs the corresponding predictive auxiliary decision information. A loss function value is calculated based on the predictive auxiliary decision information and the corresponding auxiliary decision label. The generative decision model is then updated based on the loss function value.

8. An auxiliary decision-making device based on power system alarm signals, characterized in that, include: The module includes a data acquisition module, a target reference case determination module, a prompt information determination module, and a decision support module. The data acquisition module is used to acquire the alarm signal sequence of the power system within a preset time period; The target reference case determination module is used to perform feature space mapping processing on the alarm signal sequence to generate an alarm signature that characterizes the semantic features of the alarm signal sequence. Based on a pre-set historical case knowledge base, the similarity between the alarm signature and the historical alarm signatures of each historical case in the historical case knowledge base is calculated, and a similarity score for each historical case is generated. Select a preset number of historical cases with similarity scores higher than a preset similarity threshold as target reference cases; The prompt information determination module is used to aggregate and process the historical handling information in each target reference case to generate background context information; Based on the background context information and the alarm signal sequence, a prompt message is constructed; The auxiliary decision-making module is used to input the prompt information into a preset generative decision-making model, so that the generative decision-making model generates auxiliary decision-making information for the alarm signal sequence based on the prompt information; wherein, the generative decision-making model is trained by several training samples; each training sample includes historical prompt information and corresponding auxiliary decision-making labels.

9. An electronic device, characterized in that, The system includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the auxiliary decision-making method based on power system alarm signals as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device where the storage medium is located to perform the auxiliary decision-making method based on power system alarm signals as described in any one of claims 1 to 7.

Images