Power distribution network reconstruction decision explainable optimization method, system, device and medium facing multi-scene knowledge base fusion

Abstract

The application discloses a power distribution network reconstruction decision-making explainable optimization method, system, equipment and medium for multi-scene knowledge base fusion, wherein the method comprises the following steps: using an optimized aggregated feature attribution method to extract features from historical power distribution network reconstruction decision-making and dispatcher evaluation data of a specific scene, calculating the importance score of a given evaluation rule index, and forming a historical evaluation system; optimizing and fusing the historical evaluation system and a general evaluation system according to the importance classification; using an enhanced thinking chain technology based on a reasoning mode, using the historical reconstruction scheme evaluation to extract operators and generate a demonstration set, training a large model to simulate the reasoning path of the demonstration set, optimizing the scheme of a power distribution network reconstruction decision-making candidate set, and showing the reasoning process. The application not only makes up for the defect that the current evaluation system cannot correctly optimize the network reconstruction decision-making in a specific scene, but also significantly improves the transparency and explainability of the scheme optimization.

Description

Technical Field

[0001] This invention relates to the field of power system distribution network reconfiguration and large-scale modeling, and in particular to an interpretable optimization method, system, device and medium for distribution network reconfiguration decision-making oriented towards multi-scenario knowledge base fusion. Background Technology

[0002] Distribution network reconfiguration is a key technology for optimizing power grid operation by adjusting switch combinations, aiming to minimize network losses, balance loads, and improve power supply reliability. Traditional methods rely on a pre-defined rule knowledge base (such as security constraints and economic indicators) to evaluate and select reconfiguration strategies, with rules typically ordered by a fixed priority. However, in the face of dynamic and changing operating scenarios (such as fault recovery and peak-valley load switching), fixed-priority rules are difficult to flexibly adapt to the core needs of different scenarios, resulting in the evaluation system failing to select the optimal strategy. To improve the adaptability of the knowledge base, existing research attempts to introduce feature attribution methods (such as SHAP and LIME), extracting feature importance by analyzing historical reconfiguration data and dispatcher evaluation records to construct historical evaluation indicators. However, traditional feature attribution methods suffer from unstable identification results: small perturbations to the input data can lead to significant shifts in feature importance ranking, resulting in insufficient credibility of the constructed historical evaluation indicators and thus affecting the accuracy of rule priorities. In the strategy selection stage, the Chain-of-Thought (CoT) technology of Large Language Models (LLM) needs to be used to guide the model to perform multi-indicator reasoning and evaluation of candidate strategies. Current mainstream unsupervised CoT methods rely on semantic similarity to select demonstration samples, but they face problems such as logical noise interference and lack of interpretability.

[0003] The aforementioned shortcomings limit the reliability of reconfiguration decision-making scheme selection: static rule bases cannot respond to dynamic needs of the scenario, unstable feature attribution reduces the effectiveness of historical evaluation systems, and noise-sensitive CoT mechanisms may output strategy evaluation results that violate the physical laws of the power grid. Therefore, there is an urgent need for an intelligent decision-making method that integrates dynamic rule optimization, robust feature analysis, and interpretable reasoning to achieve high-confidence selection of distribution network reconfiguration strategies. Summary of the Invention

[0004] To improve the credibility and reliability of reconfiguration decisions by providing interpretable optimization schemes for existing distribution network reconfiguration decisions, this invention provides an interpretable optimization method, system, device, and medium for distribution network reconfiguration decisions based on multi-scenario knowledge base fusion. It uses optimized aggregation feature attribution technology to extract feature importance scores of historical strategies in different scenarios, integrates existing knowledge bases to construct a strategy evaluation system, and adopts reasoning-based enhanced thinking chain technology to guide a large model to optimize reconfiguration strategies by providing interpretable optimization schemes.

[0005] Firstly, an interpretable optimization method for power distribution network reconfiguration decisions based on multi-scenario knowledge base fusion is provided, including the following steps:

[0006] S1: Based on the current distribution network reconfiguration scenario, recall historical reconfiguration decisions and corresponding dispatcher evaluation data under the same scenario; and use the optimized aggregation attribution method to calculate the importance scores of each evaluation rule indicator by performing feature attribution calculation on historical reconfiguration decisions according to the evaluation rule indicator library, and construct a historical evaluation system for distribution network reconfiguration decision schemes in specific scenarios.

[0007] S2: Obtain the dispatcher's ranking of the importance of each evaluation rule indicator in the evaluation rule indicator library, and establish a fixed general evaluation system; integrate the historical evaluation system with the general evaluation system, and adaptively adjust the priority order of the evaluation rule indicators in the general evaluation system for different scenarios, and establish a reconstructed decision evaluation system;

[0008] S3: Using the candidate set of distribution network reconfiguration decisions and the reconfiguration decision evaluation system in the current scenario as input, the large model is guided by the enhanced thinking chain technology based on the reasoning mode to comprehensively select the optimal reconfiguration decision. The evaluation of each evaluation rule indicator is used as an explanatory reason and output together with the optimal reconfiguration decision.

[0009] Furthermore, S1 specifically includes:

[0010] S11: Based on the current distribution network reconfiguration scenario, generate prompt words to guide the large model to recall historical reconfiguration decisions and corresponding dispatcher evaluation data under the same scenario;

[0011] S12: Using different feature attribution algorithms, calculate the importance score of each evaluation rule indicator according to the established evaluation rule indicator library, the historical reconstruction decision of the recall, and the corresponding scheduler evaluation data;

[0012] S13: Use the optimized aggregation attribution method to aggregate the importance scores of each evaluation rule indicator and construct a historical evaluation system for distribution network reconfiguration decision schemes in specific scenarios.

[0013] Furthermore, S13 specifically includes:

[0014] The attribution results of multiple feature attribution algorithms are aggregated by weighted summation;

[0015] Construct an objective function with the goal of minimizing the average sensitivity of aggregated attribution;

[0016] The optimal weights of the attribution results of each feature attribution algorithm are obtained by solving the problem, and the importance scores of each evaluation rule indicator after optimized aggregate attribution are calculated. A historical evaluation system is then constructed according to the importance scores of each evaluation rule indicator from high to low.

[0017] Furthermore, different feature attribution algorithms were used, including saliency maps, integral gradients, and SHAP feature attribution algorithms.

[0018] Furthermore, S2 specifically includes:

[0019] S21: Construct an evaluation rule indicator library and obtain the dispatcher's priority ranking of the importance of each evaluation rule indicator to form a fixed and universal evaluation system;

[0020] S22: The historical evaluation system and the general evaluation system are weighted and integrated. The weighted and integrated evaluation system is compared with the general evaluation system. The weighted and integrated evaluation system is integrated with the general evaluation system by combining the importance level judgment and the binary comparison weighting method to form a reconstructed decision evaluation system for multiple scenarios.

[0021] Furthermore, S22 specifically includes:

[0022] By weighted and fused together the historical evaluation system and the general evaluation system, an initial fused evaluation system is obtained;

[0023] The evaluation rules indicators are divided into important indicators and secondary indicators. If the priority order of important indicators in the initial fusion evaluation system is consistent with that in the general evaluation system, but the priority order of secondary indicators is inconsistent, the priority order of secondary indicators in the general evaluation system is changed to that in the initial fusion evaluation system to form a reconstructed decision evaluation system. If the priority order of important indicators in the initial fusion evaluation system is inconsistent with that in the general evaluation system, the scheduler is prompted to confirm whether the priority order of important indicators in the general evaluation system needs to be adjusted in the current scenario to form a reconstructed decision evaluation system after expert supervision and correction.

[0024] Furthermore, S3 specifically includes:

[0025] The complete evaluation process of historical reconfiguration decisions is obtained as a demonstration set. Each demonstration sample takes the candidate set of reconfiguration decisions for the distribution network and the corresponding reconfiguration decision evaluation system as the problem. The evaluation and screening of each evaluation rule indicator in the reconfiguration decision evaluation system is taken as the reasoning process. The optimal reconfiguration decision is finally selected as the answer. The reasoning pattern in the demonstration set is extracted with the help of the operator set.

[0026] The reasoning patterns are converted into vectors, and the reasoning pattern vectors are clustered to obtain... A cluster containing semantically similar reasoning pattern vectors;

[0027] A number of representative demonstration samples are selected from each cluster to form the final demonstration set;

[0028] The final demo set is used as contextual cue input to guide the large model using the reasoning pattern-based augmented mind chain technique for training.

[0029] The current distribution network reconfiguration decision candidate set and reconfiguration decision evaluation system are used as problem inputs. Based on the reasoning mode, the enhanced thinking chain technology guides the large model, triggers the chain reasoning ability of the large model, guides the large model to imitate the demonstration set to generate a complete evaluation and optimization process, obtains the optimal reconfiguration decision, and outputs the optimal reconfiguration decision and the complete evaluation and optimization reasoning process together.

[0030] Secondly, an interpretable optimization system for power distribution network reconfiguration decisions based on multi-scenario knowledge base fusion is provided, including:

[0031] The historical evaluation system construction module is used to recall historical reconstruction decisions and corresponding dispatcher evaluation data under the same scenario based on the current distribution network reconstruction scenario; and to use the optimized aggregation attribution method to calculate the importance scores of each evaluation rule indicator by performing feature attribution calculation on historical reconstruction decisions according to the evaluation rule indicator library, thereby constructing a historical evaluation system for distribution network reconstruction decision schemes in specific scenarios.

[0032] The module for reconstructing the decision evaluation system is used to obtain the priority ranking of various evaluation rule indicators in the evaluation rule indicator library by the dispatcher, and establish a fixed general evaluation system; the historical evaluation system is integrated with the general evaluation system, and the priority order of evaluation rule indicators in the general evaluation system is adaptively adjusted for different scenarios to establish a reconstructed decision evaluation system;

[0033] The optimal reconfiguration decision module takes the candidate set of reconfiguration decisions for the current distribution network and the reconfiguration decision evaluation system as input. Based on the reasoning mode, it guides the large model to comprehensively select the optimal reconfiguration decision and outputs the evaluation of each evaluation rule indicator as explanatory reasons along with the optimal reconfiguration decision.

[0034] Thirdly, an electronic device is provided, comprising:

[0035] A memory on which computer programs are stored;

[0036] The processor, when loading and executing the computer program, implements the previously described interpretable optimization method for power distribution network reconfiguration decisions based on multi-scenario knowledge base fusion.

[0037] Fourthly, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the previously described interpretable optimization method for power distribution network reconfiguration decision-making based on multi-scenario knowledge base fusion.

[0038] This invention proposes an interpretable optimization method, system, device, and medium for power distribution network reconfiguration decision-making based on multi-scenario knowledge base fusion, which has the following beneficial effects:

[0039] 1. Compared to traditional single-feature attribution methods for extracting historical decision indicator importance scores, this invention employs an optimized aggregate attribution method. This method integrates multiple feature attribution algorithms and aims to optimize the overall robustness and loyalty of feature extraction to improve the importance scores of evaluation rule indicators. It has been proven that its aggregated results are significantly superior to average or variance-weighted fusion feature attribution methods in terms of robustness and loyalty. This method significantly improves the accuracy and robustness of extracting the importance scores of evaluation rule indicators in historical decision-making schemes, thereby constructing a more reliable historical evaluation system.

[0040] 2. Compared with evaluation rule index libraries that simply rely on the priority of predetermined evaluation rules, this invention analyzes historical reconfiguration decision schemes in specific scenarios through feature attribution methods, adaptively integrates historical evaluation systems with existing general evaluation systems, and optimizes the priority order of rules in the evaluation system, thereby more reliably optimizing and interpreting distribution network reconfiguration decisions in specific scenarios.

[0041] 3. To address the problem that demonstration samples in traditional unsupervised thinking chain technology are prone to introducing noise, leading to logical errors and poor interpretability in large models, this invention, after generating a reconstruction decision evaluation system that integrates historical evaluation systems and existing general evaluation systems, adopts an enhanced thinking chain technology based on reasoning patterns. This technology provides a demonstration set with clear reasoning paths for large models by extracting operators and clustering reasoning patterns, and finally generates and outputs the optimal power distribution network reconstruction decision and its explanatory basis. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a flowchart of an interpretable and optimal method for power distribution network reconfiguration decision-making based on multi-scenario knowledge base fusion, provided in an embodiment of the present invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0045] The existing knowledge base (evaluation rule index base) contains several rules ordered by importance, with a relatively fixed priority order. However, facing different distribution network reconfiguration scenarios, the priority order of rules based on this knowledge base may not be optimal, making it impossible to select the most suitable reconfiguration strategy for the current scenario. To construct an optimal strategy evaluation system rule priority ranking for different reconfiguration scenarios, feature attribution is used to construct historical evaluation indicators from historical data and dispatcher historical evaluations. However, traditional feature attribution may lead to unstable or inconsistent feature recognition results, making the constructed historical evaluation system inaccurate. When using the evaluation system to optimize various distribution network reconfiguration decisions, it is necessary to use thought chain technology to guide the large model in reasoning and evaluation. Existing unsupervised thought chain methods select demonstration samples based on the semantic similarity of questions / answers. However, similar semantics may contain different computational logics, introducing irrelevant noise and causing the model to generate incorrect reasoning paths. Moreover, existing methods cannot explain why the demonstration samples are effective, lacking sufficient interpretability. Therefore, this invention proposes an interpretable optimization method for distribution network reconfiguration decisions that integrates dynamic rule optimization and robust feature analysis to achieve high-confidence optimization of distribution network reconfiguration strategies.

[0046] like Figure 1 As shown, this embodiment of the invention provides an interpretable optimization method for power distribution network reconfiguration decisions based on multi-scenario knowledge base fusion, comprising the following steps:

[0047] S1: Based on the current distribution network reconfiguration scenario, recall historical reconfiguration decisions and corresponding dispatcher evaluation data under the same scenario; and use the optimized aggregation attribution method to calculate the importance scores of each evaluation rule indicator by performing feature attribution calculation on historical reconfiguration decisions according to the evaluation rule indicator library, and construct a historical evaluation system for distribution network reconfiguration decision schemes in specific scenarios.

[0048] Specifically, S1 includes:

[0049] S11: Based on the current distribution network reconfiguration scenario, generate prompts to guide the large-scale model to recall historical reconfiguration decisions and corresponding dispatcher evaluation data under the same scenario. The data includes the impact of the reconfiguration decision on important users of the line, the current magnitude of the opposite line, the number of switching operations and the degree of automation, the load rates of the local and opposite lines and main transformers, the dispatcher's adoption or rejection of the decision, and the corresponding reasons. The data on the dispatcher's adoption or rejection of the decision and the corresponding reasons is used to subsequently build a demonstration set for training the large-scale model guided by the enhanced thinking chain technology based on inference patterns; the other data is used to calculate the scores of each evaluation rule indicator according to the evaluation rule indicator library.

[0050] For example, Table 1 shows a library of evaluation rule indicators constructed in one embodiment.

[0051] ;

[0052] In the calculation process of the above evaluation rule indicators, if the calculation result of a certain reconfiguration decision is less than or equal to 0, the reconfiguration decision is directly excluded. For these 9 evaluation rule indicators, their scores are summed according to certain weights to form the scoring function for distribution network reconfiguration decisions:

[0053] ;

[0054] in, This represents the total score of the power distribution network reconfiguration decision; This represents the score of the i-th evaluation rule indicator. This represents the weight of the i-th evaluation rule indicator.

[0055] S12: Using different feature attribution algorithms, calculate the importance score of each evaluation rule indicator according to the established evaluation rule indicator library, the historical reconstruction decision of recall, and the corresponding scheduler evaluation data.

[0056] For example, this embodiment uses saliency maps, integral gradients, and the SHAP feature attribution algorithm. The feature attribution algorithms used are shown in Table 2.

[0057] ;

[0058] in, This represents the feature input values ​​of each evaluation rule indicator (i.e., the corresponding values ​​from the previous section). ); This represents the output score of the corresponding indicator (i.e., the score corresponding to the previous one). ); This represents the importance score under different feature attribution methods, where l takes the values ​​of 1, 2, or 3 respectively. , or ; express For input The gradient; The baseline input is the feature input value that achieves a perfect score in all nine rules mentioned above. For integration variables; Indicates element-wise product; The set of features representing all evaluation rule indicators; A feature subset, representing from Excluded The set after; Representing a subset The number of features contained therein; The total number of features; It indicates the first Add the characteristics of the evaluation rules for distribution network reconfiguration to a subset. The marginal contribution of the reconstructed decision score.

[0059] S13: Use the optimized aggregation attribution method to aggregate the importance scores of each evaluation rule indicator and construct a historical evaluation system for distribution network reconfiguration decision schemes in specific scenarios.

[0060] First, we define the average sensitivity to measure the stability of the characteristic attribution results under small disturbances in the distribution network reconfiguration rule indicators, and express it as:

[0061] ;

[0062] in, Indicates average sensitivity. This represents the importance score corresponding to feature x under the feature attribution method. This represents a small perturbation to the network reconstruction rule index characteristics under a uniform distribution.

[0063] Then, the attribution results from multiple feature attribution algorithms are aggregated through weighted summation. :

[0064] ;

[0065] in, This is the weight vector between different feature attribution methods applied to the nine evaluation rule indicators, and the above formula satisfies... , ; express The matrix formed.

[0066] To minimize the average sensitivity of aggregate attribution Construct an objective function for the target:

[0067] ;

[0068] ;

[0069] Since the expected value is difficult to calculate directly, this embodiment uses Monte Carlo sampling to approximate the calculation:

[0070] Select One perturbation sample: For the first A perturbation sample, defined as:

[0071] ;

[0072] The original objective function can then be transformed into the following form:

[0073] ;

[0074] Further expansion and simplification by removing the constant term yields:

[0075] ;

[0076] make:

[0077] ;

[0078] ;

[0079] The objective function can then be simplified to a convex quadratic programming problem:

[0080] ;

[0081] Finally, by solving this convex programming problem, the optimal weights are obtained. Calculate the importance score of aggregate attribution and construct a historical evaluation system based on the index scores from high to low:

[0082] ;

[0083] ;

[0084] in, This refers to the importance score of the evaluation rule indicator after optimizing the aggregated attribution; This is a historical evaluation system based on historical power distribution network reconfiguration schemes.

[0085] S2: Construct an evaluation rule index library using the distribution network reconfiguration constraint rules, obtain the dispatcher's priority ranking of the importance of each evaluation rule index in the evaluation rule index library, and establish a fixed general evaluation system; integrate the historical evaluation system with the general evaluation system, and adaptively adjust the priority order of the evaluation rule indexes in the general evaluation system for different scenarios to establish a reconfiguration decision evaluation system.

[0086] Specifically, S2 includes:

[0087] S21: Construct an evaluation rule indicator library containing the 9 evaluation rule indicators from step S1, and obtain the scheduler's priority ranking of the importance of each evaluation rule indicator to form a fixed and universal evaluation system. , means as follows:

[0088] ;

[0089] in, Indicates that for the first Each evaluation rule indicator assigns an importance score based on the dispatcher's experience.

[0090] S22: The historical evaluation system and the general evaluation system are weighted and integrated. The weighted and integrated evaluation system is compared with the general evaluation system. The weighted and integrated evaluation system is integrated with the general evaluation system by combining the importance level judgment and the binary comparison weighting method to form a reconstructed decision evaluation system for multiple scenarios.

[0091] First, the historical evaluation system and the general evaluation system are weighted and integrated to obtain an initial integrated evaluation system. :

[0092] ;

[0093] in, These represent the general evaluation system and the historical evaluation system, respectively. The weights of each evaluation rule indicator are determined by a pre-set method.

[0094] Then, the initial fusion evaluation system is compared with the general evaluation system. The evaluation rule indicators are divided into important indicators and secondary indicators (e.g., the top five evaluation rule indicators are important indicators, and the last four are secondary indicators). If the priority order of important indicators in the initial fusion evaluation system and the general evaluation system is consistent, but the priority order of secondary indicators is inconsistent, the priority order of secondary indicators in the general evaluation system is changed to the priority order of secondary indicators in the initial fusion evaluation system to form a reconstruction decision evaluation system. If the priority order of important indicators in the initial fusion evaluation system and the general evaluation system is inconsistent, the scheduler is prompted to confirm whether the priority order of important indicators in the general evaluation system needs to be adjusted in the current scenario to form a reconstruction decision evaluation system after expert supervision and correction.

[0095] S3: The current scenario's distribution network reconfiguration decision candidate set and reconfiguration decision evaluation system are used as problem inputs. Based on the reasoning mode, the enhanced thinking chain technology guides the large model to rank, score, and record the performance of all reconfiguration decisions under each evaluation rule indicator. After all evaluation rule indicators have been screened and evaluated, the optimal reconfiguration decision is selected comprehensively. The evaluation of each evaluation rule indicator by the large model is used as an explanatory reason and output together with the optimal reconfiguration decision.

[0096] The core of this technology lies in the generation of a demonstration training set with inference patterns, as detailed below:

[0097] S31: Obtain the complete evaluation process of historical reconfiguration decisions as a demonstration set. Each demonstration sample uses the candidate set of distribution network reconfiguration decisions and the corresponding reconfiguration decision evaluation system as the problem; the evaluation and selection of each evaluation rule indicator in the reconfiguration decision evaluation system is the reasoning process; the optimal reconfiguration decision is finally selected as the answer; the reasoning pattern in the demonstration set is extracted using the operator set.

[0098] ;

[0099] ;

[0100] in, This represents a demonstration set of historical distribution network reconfiguration decisions. These represent the problem, reasoning, and answer in the demonstration set: the problem is how to evaluate and select the best candidate set for power distribution network reconfiguration under a given evaluation system; the reasoning is the complete evaluation process; and the answer is the optimal reconfiguration decision scheme in the current scenario. Indicates the first The reasoning process of a demonstration sample; This represents the set of operators related to reasoning, including arithmetic tasks such as operators and comparison tasks such as greater than and less than; For extraction functions, used to extract from Select from the middle to belong to Operators , used to form the first Reasoning patterns of a demonstration sample .

[0101] S32: Convert the inference patterns into vectors, and perform k-means clustering on the inference pattern vectors to obtain... A cluster containing semantic similarity reasoning pattern vectors:

[0102] ;

[0103] ;

[0104] in, Encoding Sentence-BERT to represent inference patterns Encoded as semantic vectors ; Indicates the first Each cluster contains semantically similar reasoning pattern vectors; The value is determined by the number of task operation types and the number of demonstration samples, and is calculated using an adaptive method.

[0105] ;

[0106] in, The number of operator types related to the task, such as an arithmetic task containing four operation types: addition, subtraction, multiplication, and division; This represents the total number of samples in the demonstration set; This represents the function for rounding up.

[0107] S33: Select several representative demonstration samples from each cluster to form the final demonstration set:

[0108] ;

[0109] in, They represent the first The representative questions, reasoning processes, and answers demonstrated in each cluster; This will be the final demo set used to improve the large model. It should be noted that one or more style samples will be selected from each cluster.

[0110] S34: Use the final demo set as contextual cue input to guide the large model with the enhanced thinking chain technology based on reasoning patterns for training the large model.

[0111] S35: The current distribution network reconfiguration decision candidate set (which can be generated by the dispatcher based on experience or by a known distribution network reconfiguration decision generation model) and the reconfiguration decision evaluation system are used as input to guide the large model based on the reasoning mode of the enhanced thinking chain technology. This triggers the chain reasoning ability of the large model, guides the large model to imitate the demonstration set to generate a complete evaluation and optimization process, obtains the optimal reconfiguration decision, and outputs the optimal reconfiguration decision and the complete evaluation and optimization reasoning process together.

[0112] The above embodiments provide an interpretable optimization method for power distribution network reconfiguration decisions based on multi-scenario knowledge base fusion, which has the following advantages:

[0113] 1. Compared to traditional single-feature attribution methods for extracting historical decision indicator importance scores, this invention employs an optimized aggregate attribution method. This method integrates multiple feature attribution algorithms and aims to optimize the overall robustness and loyalty of feature extraction to improve the importance scores of evaluation rule indicators. It has been proven that its aggregated results are significantly superior to average or variance-weighted fusion feature attribution methods in terms of robustness and loyalty. This method significantly improves the accuracy and robustness of extracting the importance scores of evaluation rule indicators in historical decision-making schemes, thereby constructing a more reliable historical evaluation system.

[0114] 2. Compared with evaluation rule index libraries that simply rely on the priority of predetermined evaluation rules, this invention analyzes historical reconfiguration decision schemes in specific scenarios through feature attribution methods, adaptively integrates historical evaluation systems with existing general evaluation systems, and optimizes the priority order of rules in the evaluation system, thereby more reliably optimizing and interpreting distribution network reconfiguration decisions in specific scenarios.

[0115] 3. To address the problem that demonstration samples in traditional unsupervised thinking chain technology are prone to introducing noise, leading to logical errors and poor interpretability in large models, this invention, after generating a reconstruction decision evaluation system that integrates historical evaluation systems and existing general evaluation systems, adopts an enhanced thinking chain technology based on reasoning patterns. This technology provides a demonstration set with clear reasoning paths for large models by extracting operators and clustering reasoning patterns, and finally generates and outputs the optimal power distribution network reconstruction decision and its explanatory basis.

[0116] This invention also provides an interpretable optimization system for power distribution network reconfiguration decisions based on multi-scenario knowledge base fusion, comprising:

[0117] The historical evaluation system construction module is used to recall historical reconstruction decisions and corresponding dispatcher evaluation data under the same scenario based on the current distribution network reconstruction scenario; and to use the optimized aggregation attribution method to calculate the importance scores of each evaluation rule indicator by performing feature attribution calculation on historical reconstruction decisions according to the evaluation rule indicator library, thereby constructing a historical evaluation system for distribution network reconstruction decision schemes in specific scenarios.

[0118] The module for reconstructing the decision evaluation system is used to obtain the priority ranking of various evaluation rule indicators in the evaluation rule indicator library by the dispatcher, and establish a fixed general evaluation system; the historical evaluation system is integrated with the general evaluation system, and the priority order of evaluation rule indicators in the general evaluation system is adaptively adjusted for different scenarios to establish a reconstructed decision evaluation system;

[0119] The optimal reconfiguration decision module takes the candidate set of reconfiguration decisions for the current distribution network and the reconfiguration decision evaluation system as input. Based on the reasoning mode, it guides the large model to comprehensively select the optimal reconfiguration decision and outputs the evaluation of each evaluation rule indicator as explanatory reasons along with the optimal reconfiguration decision.

[0120] It should be understood that the functional unit modules in the various embodiments of the present invention can be concentrated in one processing unit, or each unit module can exist physically separately, or two or more unit modules can be integrated into one unit module, and can be implemented in hardware or software.

[0121] Furthermore, embodiments of the present invention also provide an electronic device, comprising:

[0122] A memory on which computer programs are stored;

[0123] The processor, when loading and executing the computer program, implements the previously described interpretable optimization method for power distribution network reconfiguration decisions based on multi-scenario knowledge base fusion.

[0124] Furthermore, embodiments of the present invention also provide a computer-readable storage medium storing a computer program thereon, wherein when the computer program is executed by a processor, it implements the aforementioned interpretable optimization method for power distribution network reconfiguration decision-making oriented towards multi-scenario knowledge base fusion.

[0125] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0126] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0127] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0128] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0129] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0130] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A power distribution network reconstruction decision explainable optimization method for multi-scenario knowledge base fusion, characterized in that, Comprise the following steps: S1: Based on the current power distribution network reconstruction scene, recall the historical reconstruction decision and the corresponding dispatcher evaluation data under the same scene; and use the optimization aggregation attribution method to calculate the importance score of each evaluation rule index according to the evaluation rule index library, and construct the historical evaluation system of the power distribution network reconstruction decision scheme in the specific scene; The importance score of each evaluation rule index is aggregated by using the optimization aggregation attribution method, and the historical evaluation system of the power distribution network reconstruction decision scheme in the specific scene is constructed, which specifically comprises: Aggregating the attribution results of multiple feature attribution algorithms by weighted summation; A target function is constructed to minimize the average sensitivity of the aggregated attribution; The optimal weight of the attribution result of each feature attribution algorithm is obtained by solving, and the importance score of each evaluation rule index after optimization aggregation attribution is calculated, and the historical evaluation system is constructed according to the importance score of each evaluation rule index from high to low; S2: Obtain the importance priority order of each evaluation rule index in the evaluation rule index library, and establish a fixed general evaluation system; fuse the historical evaluation system with the general evaluation system, adjust the evaluation rule index priority order in the general evaluation system according to different scenes, and establish a reconstruction decision evaluation system; S3: The power distribution network reconstruction decision candidate set and the reconstruction decision evaluation system in the current scene are input into the enhanced thinking chain technology guided large model based on reasoning mode as a problem, and the optimal reconstruction decision is selected, and the evaluation of each evaluation rule index is taken as an explanatory reason, which is output together with the optimal reconstruction decision; The enhanced thinking chain technology guided large model based on reasoning mode is trained by the following method: Obtain the complete evaluation process of historical reconstruction decision as a demonstration set, and each demonstration sample takes the power distribution network reconstruction decision candidate set and the corresponding reconstruction decision evaluation system as a problem; the evaluation of each evaluation rule index in the reconstruction decision evaluation system is taken as a reasoning process; the finally selected optimal reconstruction decision is taken as an answer; and the reasoning mode in the demonstration set is extracted by means of operator set; The reasoning patterns are converted into vectors, and the reasoning pattern vectors are clustered to obtain clusters containing reasoning pattern vectors that are semantically similar. Select several representative demonstration samples from each cluster to form a final demonstration set; The final demonstration set is input into the enhanced thinking chain technology guided large model based on reasoning mode as a context prompt to train the enhanced thinking chain technology guided large model based on reasoning mode.

2. The power distribution network reconfiguration decision explainable preference method oriented to multi-scene knowledge base fusion according to claim 1, characterized in that, S1 specifically comprises: S11: According to the current power distribution network reconstruction scene, generate prompt words to guide the large model to recall the historical reconstruction decision and the corresponding dispatcher evaluation data under the same scene; S12: Use different feature attribution algorithms to calculate the importance score of each evaluation rule index according to the established evaluation rule index library and the recalled historical reconstruction decision and the corresponding dispatcher evaluation data; S13: Use the optimization aggregation attribution method to aggregate the importance score of each evaluation rule index, and construct the historical evaluation system of the power distribution network reconstruction decision scheme in the specific scene.

3. The power distribution network reconfiguration decision explainable preference method oriented to multi-scene knowledge base fusion according to claim 2, characterized in that, Using different feature attribution algorithms includes significance map, integral gradient and SHAP feature attribution algorithm.

4. The power distribution network reconfiguration decision explainable preference method oriented to multi-scene knowledge base fusion according to claim 1, characterized in that, S2 specifically comprises: S21: construct an evaluation rule index library, and obtain the importance priority order of dispatchers for each evaluation rule index, to form a fixed general evaluation system; S22: the historical evaluation system and the general evaluation system are weighted and fused, the fused evaluation system and the general evaluation system are compared, the fused evaluation system and the general evaluation system are fused by combining the importance level judgment and the binary comparison weighting method, and a multi-scene-oriented reconstructed decision evaluation system is formed.

5. The power distribution network reconfiguration decision explainable preference method oriented to multi-scene knowledge base fusion according to claim 4, characterized in that, S22 specifically includes: The historical evaluation system and the general evaluation system are weighted and fused to obtain an initial fused evaluation system; The evaluation rule indexes are divided into important indexes and secondary indexes, if the importance priority order of the important indexes of the initial fused evaluation system and the general evaluation system is consistent and the importance priority order of the secondary indexes is inconsistent, the importance priority order of the secondary indexes in the general evaluation system is changed to that in the initial fused evaluation system, and a reconstructed decision evaluation system is formed; if the importance priority order of the important indexes of the initial fused evaluation system and the general evaluation system is inconsistent, the dispatcher is prompted to confirm whether the importance priority order of the important indexes in the general evaluation system needs to be adjusted under the current scene, and a supervised modified reconstructed decision evaluation system is formed.

6. The power distribution network reconfiguration decision explainable preference method oriented to multi-scene knowledge base fusion according to claim 4, characterized in that, S3 specifically includes: The power distribution network reconstruction decision candidate set and the reconstruction decision evaluation system under the current scene are input into the enhanced thinking chain technology based on reasoning mode to guide the large model, the chain reasoning ability of the large model is triggered, the large model is guided to imitate the demonstration set to generate a complete evaluation and optimization process, the optimal reconstruction decision is obtained, and the optimal reconstruction decision and the complete evaluation and optimization reasoning process are output together.

7. A power distribution network reconfiguration decision explainable preference system oriented to multi-scenario knowledge base fusion, characterized in that, The system for implementing the multi-scene knowledge base fusion-oriented power distribution network reconstruction decision explainable optimization method according to any one of claims 1 to 6 includes: A historical evaluation system construction module is configured to recall historical reconstruction decisions and corresponding dispatcher evaluation data under the same scene based on the current power distribution network reconstruction scene, and to calculate the importance scores of each evaluation rule index by performing feature attribution calculation on the historical reconstruction decisions according to the evaluation rule index library by using an optimized aggregation attribution method, so as to construct a historical evaluation system of the power distribution network reconstruction decision scheme under a specific scene. A reconstruction decision evaluation system construction module is configured to obtain the importance priority order of dispatchers for each evaluation rule index in the evaluation rule index library, establish a fixed general evaluation system, fuse the historical evaluation system and the general evaluation system, and adaptively adjust the importance priority order of the evaluation rule indexes in the general evaluation system for different scenes and establish a reconstruction decision evaluation system. A reconstruction decision optimization module is configured to input the power distribution network reconstruction decision candidate set and the reconstruction decision evaluation system under the current scene into the enhanced thinking chain technology based on reasoning mode to guide the large model, comprehensively select the optimal reconstruction decision, and output the evaluation of each evaluation rule index as an explanatory reason together with the optimal reconstruction decision.

8. An electronic device, comprising: The system includes: A memory having a computer program stored thereon. A processor for loading and executing the computer program to implement the method for power distribution network reconfiguration decision explainable preferred method facing multi-scenario knowledge base fusion according to any one of claims 1 to 6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the method for power distribution network reconfiguration decision explainable preferred method facing multi-scenario knowledge base fusion according to any one of claims 1 to 6.

Images