A method and apparatus for responding to demand for electric power supplies

Abstract

The application discloses a kind of power material demand response method and device, method includes: obtaining the demand influence file of power material;According to the file type of demand influence file, different file processing strategies are matched, and the demand influence file is structured and processed to obtain material demand data;According to the material demand generation model constructed in advance, the material demand data is audited and handled by fusing power multi-element knowledge graph, to obtain the first material demand file;Combining material demand standard, the first material demand file is detected by difference, and the first material demand file is associated with the standard clause corresponding to material demand standard, and the result of responding power material demand is given.Through the detection of the first material demand file by combining material demand standard, the result of responding power material demand is obtained, the scientificity and rationality of power material demand are guaranteed, the accuracy of responding power material demand result is improved, and reference and help are provided for power material demand.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of electric power material management, and particularly relates to a method and device for responding to electric power material demand. BACKGROUND

[0002] Electric power materials refer to various materials required in the construction, operation and maintenance of an electric power system, and an electric power material plan responding to electric power material demand is a prior arrangement of the types, quantity, procurement time and supply channel of required electric power materials according to the development, operation demand and maintenance of the electric power system. A reasonable electric power material plan can ensure the smooth progress of the construction, operation and maintenance of the electric power system, avoid material shortage or overstock, reduce cost and improve the reliability and economy of the electric power system.

[0003] At present, in the process of constructing a power grid, the demand for electric power materials is large in quantity and various in type, and the electric power grid material plan is mainly generated by approving, summarizing and on-site surveying the information reported by lower-level institutions according to the investigation, statistics, estimation and report of the lower-level institutions by upper-level institutions. This process not only consumes a lot of manpower and resources, but also has a particularly large amount of work to review, and the generation of the demand cycle takes a particularly long time. Meanwhile, the filling is not standardized in this process, and the system input is prone to errors, which results in low accuracy of the material demand plan, is not conducive to the arrangement of material procurement plans, engineering construction and production plans, and also restricts the improvement of the lean management level and ability of electric power materials.

[0004] Patent application CN120409766A provides an electric power material demand prediction method, system, device and medium, which comprises: collecting historical electric power material demand data and influence factor data, and performing preprocessing of cleaning, transformation and feature extraction; building a deep learning model based on a TensorFlow deep learning framework; training the deep learning model using the historical data after preprocessing, adjusting the model parameters and evaluating the performance to obtain a final prediction model; collecting real-time electric power material demand data and influence factor data, preprocessing and inputting into the final prediction model for demand electric power material prediction. The deep learning model is built through the TensorFlow deep learning framework, and the prediction of electric power material demand data is realized by combining the processed historical electric power material demand data, thereby improving the prediction accuracy.

[0005] In the above related technology, a deep learning model is used to predict material demand, but responding to electric power material demand involves many aspects, and also needs to be modified and adjusted in time according to the actual development situation. How to ensure the scientificity and rationality of electric power material demand and improve the accuracy of electric power material demand is a problem to be solved at present. SUMMARY

[0006] In view of the defects in the prior art, the present application provides a power material demand response method and device, the method comprising: obtaining a demand influence file of power materials; according to the file type of the demand influence file, matching different file processing strategies, structuring the demand influence file to obtain material demand data; according to a pre-constructed material demand generation model, fusing a power multi-element knowledge graph to audit the material demand data, obtaining a first material demand file; combining a material demand standard, detecting the differences of the first material demand file, and associating the first material demand file with the standard clauses corresponding to the material demand standard to give the result of responding to the power material demand. By processing the demand influence file, combining the material demand generation model, obtaining the first material demand file and detecting it, the result of responding to the power material demand is obtained, and according to the actual material demand standard, the first material demand file is modified and adjusted in time to ensure the scientificity and rationality of the power material demand and improve the accuracy of the power material demand, providing reference and help for the power material demand.

[0007] In a first aspect, the present application provides a power material demand response method, specifically comprising the following steps:

[0008] Obtaining a demand influence file of power materials;

[0009] According to the file type of the demand influence file, matching different file processing strategies, structuring the demand influence file to obtain material demand data;

[0010] According to a pre-constructed material demand generation model, fusing a power multi-element knowledge graph to audit the material demand data, obtaining a first material demand file;

[0011] Combining a material demand standard, detecting the differences of the first material demand file, and associating the first material demand file with the standard clauses corresponding to the material demand standard to give the result of responding to the power material demand.

[0012] Further, the file type includes at least one of text, table and image;

[0013] According to the file type of the demand influence file, matching different file processing strategies, structuring the demand influence file to obtain material demand data, specifically comprising:

[0014] If the file type of the demand influence file is text, the demand influence file is segmented and tagged by word, and converted into structured data to obtain corresponding material demand data;

[0015] If the file type of the demand influence file is a table, the demand influence file is structurally parsed, the relationship between the content and position of each cell in the demand influence file is analyzed, and the content of the demand influence file is extracted to give corresponding material demand data.

[0016] If the file type of the demand influence file is an image, the text in the demand influence file is enhanced and recognized, and is converted into structured data to give corresponding material demand data.

[0017] Further, the relationship between the content and position of each cell in the demand influence file is represented by cell correlation, specifically as follows:

[0018]

[0019] Wherein, R(c i ,c j ) is the cell correlation of the i-th cell and the j-th cell, a is the correlation weight coefficient, PosSim(c i ,c j ) is the position similarity of the i-th cell and the j-th cell, ContSim(c i ,c j ) is the content similarity of the i-th cell and the j-th cell.

[0020] Further, according to the pre-constructed material demand generation model, the power multi-knowledge graph is fused to audit the material demand data to obtain a first material demand file, which is obtained by the following steps:

[0021] According to the sequence length of the material demand data, the rotating position coding extrapolation technology is used to expand the processing window of the material demand generation model;

[0022] The correlation similarity between each demand vector of the material demand data and each material vector in the power multi-knowledge graph is analyzed, and the attention weight in the attention mechanism in the material demand generation model is adjusted, wherein the power multi-knowledge graph includes material vectors of each power material, and each material vector of the power material includes technical parameters, coding rules and procurement specifications of the corresponding power material in the power industry standard;

[0023] The material demand data is substituted into the adjusted material demand generation model to obtain the first material demand file.

[0024] Further, the attention weight is specifically represented as follows:

[0025]

[0026] Wherein, A kis the attention weight corresponding to the kth material vector in the power multi-knowledge graph, S k is the semantic similarity between the kth material vector in the power multi-knowledge graph and the material demand data, λ k is the importance coefficient of the kth material vector in the power multi-knowledge graph, and M is the total number of material vectors in the power multi-knowledge graph.

[0027] Further, the demand reward function is specifically represented as:

[0028]

[0029] wherein β1 is the weight coefficient corresponding to the demand basic reward, β2 is the weight coefficient corresponding to the demand incremental reward, and β3 is the weight coefficient corresponding to the demand penalty term, R base is the demand basic reward, R inc is the demand incremental reward, P i is the demand penalty term of the ith type of violation.

[0030] Further, the material demand generation model includes a demand input layer, an attention mechanism, a demand reward layer, and a demand output layer.

[0031] The attention mechanism is connected with the demand input layer, and the material demand data enters the attention mechanism through the demand input layer to extract the core demand features in the material demand data.

[0032] The demand output layer is connected with the attention mechanism, and the intermediate demand file is generated based on the core demand features.

[0033] The demand reward layer is connected with the demand output layer, analyzes the compliance in the intermediate demand file, obtains the total demand reward, and feeds back to the attention mechanism and the demand output layer.

[0034] The demand reward layer includes a basic reward layer, an incremental reward layer, a violation penalty layer, and a reward fusion layer.

[0035] The basic reward layer is connected with the demand output layer, and is used to analyze the matching degree between the intermediate demand file and the corresponding procurement specification to obtain the intermediate demand basic reward.

[0036] The incremental reward layer is connected with the demand output layer, and is used to analyze the fluctuation difference between each value in the intermediate demand file and the corresponding standard value to obtain the intermediate demand incremental reward.

[0037] The violation penalty layer is connected with the demand output layer, and is used to analyze the violation between each data in the intermediate demand file to give the intermediate demand penalty term.

[0038] The reward fusion layer is connected with the basic reward layer, the incremental reward layer and the violation punishment layer respectively, and is used for fusing the intermediate demand basic reward, the intermediate demand incremental reward and the intermediate demand punishment term, obtaining the total demand reward and feeding back to the attention mechanism and the demand output layer.

[0039] Further, the construction of the material demand generation model is obtained through the following steps:

[0040] Obtain historical material demand data;

[0041] Analyze the matching degree of the historical material demand data and the corresponding procurement specification to obtain a demand basic reward;

[0042] Analyze the fluctuation difference between each value in the historical material demand data and the corresponding standard value to obtain a demand incremental reward;

[0043] Analyze the violation between each data in the historical material demand data to give a demand penalty term;

[0044] Fuse the demand basic reward, the demand incremental reward and the demand penalty term to train the initial neural network model until convergence to obtain a first demand generation model;

[0045] Insert a low-rank matrix into the attention mechanism of the first demand generation model, freeze the model parameters of the first demand generation model, adjust and optimize the parameters corresponding to the low-rank matrix, and obtain a material demand generation model.

[0046] Further, it further comprises:

[0047] A two-stage distillation framework is constructed, wherein the first stage is used to output the original score distribution of the demand teacher model, and the second stage is used to minimize the output difference between the demand student model and the demand teacher model in the power material demand generation task;

[0048] Analyze the influence of the model matrix in the demand teacher model on the output result of the demand teacher model to give the Hessian matrix of the demand teacher model;

[0049] Cholesky decomposition is performed on the Hessian matrix of the demand teacher model to obtain a lower triangular matrix;

[0050] The model matrix in the demand teacher model is blocked to obtain a plurality of parameter sub-matrices;

[0051] The model parameters in each parameter sub-matrix are quantized in turn, and the corresponding quantization error is given in combination with the lower triangular matrix;

[0052] The quantization process of the model parameters is repeated until all the model parameters are quantized and the quantization error converges to obtain a compressed model matrix;

[0053] The compressed model matrix is integrated into the demand student model to complete the construction of the demand student model.

[0054] The demand teacher model or the demand student model is used as a material demand generation model.

[0055] Further, the deployment and optimization of the demand teacher model and the demand student model are also included:

[0056] The demand teacher model is deployed on a cloud device, and the demand student model is deployed on an edge device.

[0057] The demand student model obtains a demand influence file on the edge device to give a corresponding response power material demand result and synchronizes the response power material demand result to the cloud device.

[0058] The demand teacher model audits the response power material demand result in response to the edge device, and gives an audit result.

[0059] Based on the audit result, the demand student model is optimized.

[0060] Further, the response power material demand result is synchronized to the cloud device, specifically including:

[0061] The response power material demand result is compressed to obtain a demand binary file.

[0062] Based on a pre-configured encryption algorithm, the demand binary file is encrypted to obtain a demand encrypted file.

[0063] Based on an encryption channel of the cloud device and the edge device, the demand encrypted file is transmitted to the cloud device.

[0064] The demand teacher model audits the response power material demand result in response to the edge device, and gives an audit result, specifically including:

[0065] The cloud device obtains the demand encrypted file from the corresponding encryption channel.

[0066] According to the device model of the edge device, a corresponding decryption key is searched from a key library to decrypt the demand encrypted file to obtain a compressed file.

[0067] The compressed file is decompressed, and the demand teacher model audits the decompressed compressed file to give an audit result.

[0068] Further, the material demand standard includes a basic material standard, a professional material standard, and a dynamic material standard.

[0069] The first material demand file is subjected to difference detection in combination with the material demand standard, and the first material demand file is associated with a standard clause corresponding to the material demand standard to give a result of responding to the power material demand, specifically including:

[0070] Based on the basic material standard, the uniqueness of various power materials in the first material demand file is checked, and a corresponding color is selected for marking;

[0071] Based on the professional material standard, the technical parameters and corresponding equipment demand of various power materials in the first material demand file are matched, and a corresponding color is selected for marking;

[0072] Based on the dynamic material standard, the quantity of various power materials in the first material demand file is associated with the dynamic inventory, and a corresponding color is selected for marking;

[0073] The first material demand file is traversed to give a result of responding to the power material demand.

[0074] In a second aspect, the present application also provides a power material demand response device, which adopts the power material demand response method of any one of the above, including:

[0075] A data acquisition module is configured to acquire a demand influence file of power materials;

[0076] A data processing module is configured to match different file processing strategies according to the file type of the demand influence file, and to structure-process the demand influence file to obtain material demand data;

[0077] A demand generation module is configured to fuse a power multi-element knowledge graph according to a pre-constructed material demand generation model, and to audit-process the material demand data to obtain a first material demand file;

[0078] A demand review module is configured to detect differences in the first material demand file in combination with the material demand standard, and to associate the first material demand file with a standard clause corresponding to the material demand standard to give a result of responding to the power material demand.

[0079] The power material demand response method and device provided by the present application at least have the following beneficial effects:

[0080] (1) By processing the demand influence file, combining the material demand generation model, obtaining the first material demand file and detecting it, obtaining the power material demand result, and according to the actual material demand standard, the first material demand file is modified and adjusted in time to ensure the scientificity and rationality of the power material demand, improve the accuracy of the power material demand, and provide reference and help for the power material demand.

[0081] (2) By building a demand reward layer including a basic reward layer, an incremental reward layer, a violation penalty layer and a reward fusion layer, a demand reward function is integrated into the material demand generation model, and based on the synergistic effect of the demand reward function and the low-rank matrix, the performance of the first demand generation model on the power material demand task is effectively optimized. The demand reward function provides feedback signals through a three-level reward mechanism to fine-tune the first demand generation model and ensure that the model can generate high-quality power material demand.

[0082] (3) In the process of training the material demand generation model, by freezing the model parameters of the first demand generation model, the parameters corresponding to the low-rank matrix are adjusted and optimized, which can improve the training efficiency of the material demand generation model and reduce the demand for computing resources. At the same time, only the low-rank matrix is adjusted, which can reduce the complexity of the material demand generation model, reduce the risk of overfitting, and enhance the generalization ability of the model. BRIEF DESCRIPTION OF DRAWINGS

[0083] Figure 1 The flowchart of the response method of the power material demand provided by the embodiment of the present application;

[0084] Figure 2 The flowchart of data processing provided by the embodiment of the present application;

[0085] Figure 3 The flowchart of obtaining the first material demand file provided by the embodiment of the present application;

[0086] Figure 4 The architecture diagram of the material demand generation model provided by the embodiment of the present application;

[0087] Figure 5 The training flowchart of the material demand generation model provided by the embodiment of the present application;

[0088] Figure 6 The structural block diagram of the response device of the power material demand provided by the embodiment of the present application.

[0089] Among them, 201, data acquisition module; 202, data processing module; 203, demand generation module; 204, demand review module. DETAILED DESCRIPTION

[0090] In order to better understand the above technical solutions, the above technical solutions will be described in detail in conjunction with the drawings and specific embodiments in the specification. Obviously, the described embodiments are only part of the embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0091] The terminology used in the description of the application herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. As used in the description of the application and the appended claims, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be understood that "a" or "an" or "the" or "one or more" or "at least one" or "one or more of the following items" can be construed to mean one or more, unless otherwise indicated herein.

[0092] It should also be noted that the terms "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The materials, methods, and examples provided herein are illustrative only and not intended to be limiting.

[0093] Electric power is a power production and consumption system composed of power generation, power transmission, power transformation, power distribution and power utilization. As a clean and efficient energy, electric power is an important foundation for modern social production and life, and is widely used in industrial manufacturing, residential life, transportation and other fields.

[0094] With the continuous development of the electric power industry, the scale of the electric power system is increasing, and the types and quantities of electric power materials are increasing, which puts forward higher requirements on the accuracy, timeliness and scientificity of the electric power material plan in response to the demand for electric power materials. How to use information technology and intelligent means to manage and support the generation, review and optimization of electric power material plan, improve the efficiency and level of plan management.

[0095] Electric power material plan is an important part of electric power system operation management, which is directly related to material procurement, inventory management and operation cost control of electric power enterprises. Electric power material demand is influenced by many factors, such as seasonal demand changes, periodic fluctuations, economic conditions, weather conditions, equipment status, etc. The complex relationship and nonlinear correlation between these factors make it difficult for traditional methods to meet the fine management needs of electric power enterprises. In the face of the above situation, it is necessary to use demand generation model, which can analyze the demand of various electric power materials, and provide support for the implementation of various procurement modes.

[0096] Meanwhile, the power material plan involves multi-source data including text, table, image and various forms, the existing technology mainly relies on manual or simple automatic tools for cleaning and parsing of the data, which is slow in processing speed and low in accuracy, difficult to meet the demand of large-scale data processing, the power field has the characteristics of strong professionalism and complex knowledge system, the existing system is difficult to fully integrate the professional knowledge of the power field in the plan generation and review process of responding to material demand, resulting in deviation between the plan and the actual demand, the existing large model has large parameter size and high computing resource consumption, which is difficult to deploy on edge devices and cannot meet the real-time processing demand in the edge scene of the power system, resulting in insufficient adaptability and practicality of the system, affecting the use effect of the power material plan.

[0097] Therefore, the present application provides a power material demand response method, which comprises the following steps:

[0098] By processing the demand influence file, combining the material demand generation model, obtaining the first material demand file and detecting it, the result of responding to the power material demand is obtained, the first material demand file is modified and adjusted in time according to the actual material demand standard, the scientificity and rationality of the power material demand are ensured, the accuracy of the power material demand is improved, and reference and help are provided for the power material demand.

[0099] As shown in Figure 1 The present application provides a power material demand response method, which comprises the following steps:

[0100] S101: Obtain the demand influence file of the power material.

[0101] Specifically, the demand impact file is a file that has an impact on the demand of the power material in the planning period, such as maintenance records, replacement records, and information or files of the use environment of each power material.

[0102] In a specific example, the annual power material procurement of a power company needs to be planned, involving multiple types of power materials such as power generation equipment, power transmission lines, and power transformation facilities. The demand impact files include unstructured data such as handwritten inspection records, paper report images, and equipment nameplate photos.

[0103] S102: According to the file type of the demand impact file, match different file processing strategies, and structure the demand impact file to obtain material demand data.

[0104] Specifically, the file type includes at least one of text, table, and image.

[0105] Further, according to the file type of the demand impact file, different file processing strategies are matched to structure the demand impact file to obtain material demand data, specifically including: Figure 2 If the file type of the demand impact file is text, the demand impact file is segmented and tagged with parts of speech, and is converted into structured data to obtain corresponding material demand data;

[0106]

[0107] If the file type of the demand impact file is a table, the demand impact file is structure parsed, the relationship between the content and the position of each cell in the demand impact file is analyzed, and the content of the demand impact file is extracted to give corresponding material demand data;

[0108] If the file type of the demand impact file is an image, the text in the demand impact file is enhanced and recognized, and is converted into structured data to give corresponding material demand data.

[0109] Wherein, the relationship between the content and the position of each cell in the demand impact file is represented by cell correlation degree, specifically:

[0110]

[0111] Wherein, R(c i ,c j ) is the cell correlation degree of the i-th cell and the j-th cell, a is the correlation weight coefficient, PosSim(c i ,c j ) is the position similarity of the i-th cell and the j-th cell, ContSim(c i ,c j ​) is the content similarity of the i-th cell and the j-th cell.

[0112] In a specific embodiment, when the file type of the demand impact file is text, a BERT model variant that integrates BiLSTM (Bidirectional Long Short-Term Memory Network) + CRF (Conditional Random Field) is used to perform word segmentation, part-of-speech tagging, and entity recognition on the demand impact file, extract key information such as "transformer model specifications" and "cable quantity", and convert it into structured data to obtain the corresponding material demand data.

[0113] In the BERT model variant that integrates BiLSTM + CRF, the demand impact file is input into BERT to obtain the context representation of each word, and a fixed-dimensional vector is output to represent the semantic information of each word in the demand impact file. Then, the output of BERT is taken as the input of BiLSTM to further capture the forward and backward dependencies in the sequence, and BiLSTM outputs a vector containing forward and backward information. Finally, the output of BiLSTM is taken as the input of the CRF layer to model the dependencies between label sequences through the CRF layer, and the CRF layer outputs the final label sequence to complete the processing of the demand impact file.

[0114] When the file type of the demand impact file is a table, a layout language model (LayoutLMv3) is used to parse the structure of the demand impact file, analyze the relationship between the content and location of each cell in the demand impact file, and extract the content of the demand impact file to obtain the corresponding material demand data. In a specific example, there is an image of a power material inventory report that contains a diagonal table header. Optical Character Recognition (OCR) technology is used to recognize the text content and its location information from the image of the power material inventory report. The extracted text content and its corresponding bounding box are converted into a format that can be processed by the LayoutLMv3 model and input into the LayoutLMv3 model to obtain an understanding of the table structure and the content and location information of each cell. By calculating the correlation between cells, the row and column relationships of the table are established. According to the correlation of the cells, the row and column structure of the table is determined and the content of each cell is extracted, and organized according to the row and column structure to obtain the corresponding material demand data.

[0115] When the file type of the demand impact file is an image, a multi-scale OCR engine is used to enhance and recognize the text in the demand impact file, and convert it into structured data, giving the corresponding material demand data. In a specific example, the demand impact file is an equipment nameplate image. After enhancing the blurred text in the equipment nameplate image, the text is recognized, and the result is converted into structured data in JSON format, obtaining the corresponding material demand data, for example, the parameter information of "10kV circuit breaker" is standardized and stored.

[0116] S103: According to the pre-constructed material demand generation model, the power multi-knowledge graph is fused, and the material demand data is audited to obtain a first material demand file.

[0117] Referring to Figure 3 , specifically comprising:

[0118] According to the sequence length of the material demand data, a rotating position encoding extrapolation technique is used to expand the processing window of the material demand generation model;

[0119] The correlation similarity between each demand vector of the material demand data and each material vector in the power multi-knowledge graph is analyzed, and the attention weight in the attention mechanism in the material demand generation model is adjusted, wherein the power multi-knowledge graph includes material vectors of each power material, and each material vector of the power material includes technical parameters, coding rules and procurement specifications of the corresponding power material in the power industry standard;

[0120] The material demand data is substituted into the adjusted material demand generation model to obtain a first material demand file.

[0121] Wherein, the attention weight is specifically represented as:

[0122]

[0123] Wherein, A k is the attention weight corresponding to the kth material vector in the power multi-knowledge graph, S k is the semantic similarity between the kth material vector in the power multi-knowledge graph and the material demand data, λ k is the important coefficient of the kth material vector in the power multi-knowledge graph, and M is the total number of material vectors in the power multi-knowledge graph.

[0124] It can be understood that the RoPE extrapolation technique is directly applied to the model based on the Transformer architecture. By adjusting the RoPE parameters of the model (such as the rotation angle, the scaling factor, etc.), longer sequences can be processed at model runtime without retraining. The RoPE extrapolation technique can be implemented in different ways, for example, Position Interpolation (PI) maps positions beyond the training length back into the training range by reducing the rotation radian. NTK-Aware interpolation distinguishes between high-frequency (low-dimensional) and low-frequency (high-dimensional) components and extrapolates the high-frequency part (small scaling amplitude) and interpolates the low-frequency part (large scaling amplitude). Dynamic NTK adjusts the scaling factor dynamically to gradually reduce the rotation radian according to the current sequence length, avoiding performance discontinuity. In other examples, other ways can be used to implement the expansion of the processing window of the material demand generation model, which are not limited herein.

[0125] In a specific embodiment, NTK-Aware interpolation is used, which is specifically represented as:

[0126]

[0127] where β' is the adjusted RoPE parameter, β is the original RoPE parameter, L extra is the sequence length of the material demand data, L train is the maximum sequence length of the material demand generation model during training (i.e., the processing window of the material demand generation model), d is the dimension of the material demand generation model, and i is the current dimension index.

[0128] The adjusted RoPE parameter is substituted into the material demand generation model to expand the processing window of the material demand generation model.

[0129] The correlation similarity of each demand vector of the material demand data and each material vector in the power multi-element knowledge graph is analyzed, and the attention weight in the attention mechanism in the material demand generation model is adjusted to strengthen the correlation weight between the power material field knowledge and the general semantics.

[0130] The power multi-knowledge graph includes a material vector of each power material, and the material vector of each power material includes technical parameters, coding rules and procurement specifications of the corresponding power material in the power industry standard. Specifically, the power industry standard is obtained, and technical parameters, coding rules and procurement specifications of various power materials including power generation, power transmission and power transformation equipment in the power industry standard are collected to construct a triple knowledge graph including entities, attributes and relationships, such as "lightning arrester - applicable voltage - 110 kV". In this example, the power multi-knowledge graph is a triple knowledge graph, and in other examples, the content in the knowledge graph is increased or decreased according to actual conditions, and this is not limited.

[0131] Finally, the material demand data is substituted into the adjusted material demand generation model to obtain a first material demand file.

[0132] Further, with reference to Figure 4 , the material demand generation model includes a demand input layer, an attention mechanism, a demand reward layer and a demand output layer;

[0133] The attention mechanism is connected with the demand input layer, and the material demand data enters the attention mechanism through the demand input layer to extract core demand features in the material demand data;

[0134] The demand output layer is connected with the attention mechanism, and an intermediate demand file is generated based on the core demand features;

[0135] The demand reward layer is connected with the demand output layer, and the compliance of the intermediate demand file is analyzed to obtain a total demand reward and feedback to the attention mechanism and the demand output layer;

[0136] The demand reward layer includes a basic reward layer, an incremental reward layer, a violation penalty layer and a reward fusion layer;

[0137] The basic reward layer is connected with the demand output layer, and is used to analyze the matching degree of the intermediate demand file and the corresponding procurement specification to obtain an intermediate demand basic reward;

[0138] The incremental reward layer is connected with the demand output layer, and is used to analyze the fluctuation difference between each value in the intermediate demand file and the corresponding standard value to obtain an intermediate demand incremental reward;

[0139] The violation penalty layer is connected with the demand output layer, and is used to analyze the violation between each data in the intermediate demand file to give an intermediate demand penalty term;

[0140] The reward fusion layer is connected with the basic reward layer, the incremental reward layer and the violation penalty layer respectively, and is used to fuse the intermediate demand basic reward, the intermediate demand incremental reward and the intermediate demand penalty term to obtain the total demand reward and feedback to the attention mechanism and the demand output layer.

[0141] Further, referring to Figure 5 , the construction of the material demand generation model is obtained by the following steps:

[0142] Obtain historical material demand data;

[0143] Analyze the matching degree of the historical material demand data and the corresponding procurement specification to obtain a demand base reward;

[0144] Analyze the fluctuation difference between each value in the historical material demand data and the corresponding standard value to obtain a demand increment reward;

[0145] Analyze the violation between each data in the historical material demand data to give a demand penalty term;

[0146] Fuse the demand base reward, the demand increment reward, and the demand penalty term to train the initial neural network model until convergence to obtain a first demand generation model;

[0147] Insert a low-rank matrix into the attention mechanism of the first demand generation model, freeze the model parameters of the first demand generation model, and adjust and optimize the parameters corresponding to the low-rank matrix to obtain a material demand generation model.

[0148] Wherein, the demand reward function is obtained by fusing the demand base reward, the demand increment reward, and the demand penalty term, and is specifically expressed as:

[0149]

[0150] Wherein, β1 is the weight coefficient corresponding to the demand base reward, β2 is the weight coefficient corresponding to the demand increment reward, and β3 is the weight coefficient corresponding to the demand penalty term. R base is the demand base reward, R inc is the demand increment reward, and P i is the demand penalty term of the i-th violation.

[0151] Based on the demand reward function, after the first demand generation model is obtained by training the initial neural network model, a low-rank matrix is inserted into the attention mechanism of the first demand generation model to adapt to the power material demand task.

[0152] In a specific example, the weight matrix of the first demand generation model attention mechanism is W, the inserted low-rank matrices are A and B, and the new weight matrix W' is:

[0153]

[0154] Wherein, the dimension of the low-rank matrix A is d x r, the dimension of the low-rank matrix B is r x d, and r is the rank of the low-rank matrix.

[0155] Then, the model parameters of the first demand generation model are frozen, and only the parameters corresponding to the newly added low-rank matrix are fine-tuned. This can significantly reduce the consumption of computing resources while maintaining the performance of the model. In this example, the AdamW optimizer is used for fine-tuning. Using historical material demand data, the low-rank matrix parameters are adjusted based on the demand reward function to optimize the model performance, and the material demand generation model is obtained.

[0156] In one specific example, there is a power material demand task that needs to generate a power material procurement plan. Based on the completion of the first demand generation model, low-rank matrices A and B are inserted into the attention mechanism of the first demand generation model. The model parameters of the first demand generation model are frozen, and only the parameters corresponding to the low-rank matrix are fine-tuned. Historical material demand data is prepared, including historical plans, historical procurement specifications, historical material quantities, and other data.

[0157] The historical material demand data is input into the first demand generation model to generate a plan P. The cosine similarity between the generated plan P and the standard procurement specification S is calculated to obtain a demand base reward:

[0158]

[0159] The pre-set reasonable quantity range is [Q min ,Q max ], the generated material quantity is Q, and the demand increment reward is calculated:

[0160]

[0161] Set: the penalty corresponding to the model specification conflict is P spec , the penalty corresponding to the delivery period conflict is P delivery , and the demand penalty term P penalty is calculated:

[0162]

[0163] The total reward is obtained by summing the demand base reward and the demand increment reward and subtracting the demand penalty term. The loss function is calculated, and the parameters corresponding to the low-rank matrix are updated through the AdamW optimizer. Through multiple iterations of training, the low-rank matrix parameters are optimized, and the model performance is improved.

[0164] By building a demand reward layer including a base reward layer, an increment reward layer, a violation penalty layer, and a reward fusion layer, the demand reward function is integrated into the material demand generation model. Based on the synergistic effect of the demand reward function and the low-rank matrix, the performance of the first demand generation model in the power material demand task is effectively optimized. The demand reward function provides feedback signals through a three-level reward mechanism to fine-tune the first demand generation model, ensuring that the model can generate high-quality power material demand.

[0165] Further comprising:

[0166] A two-stage distillation framework is constructed, wherein the first stage is used to output the raw score distribution of the demand teacher model, and the second stage is used to minimize the output difference between the demand student model and the demand teacher model in the power material demand generation task;

[0167] The influence of the model matrix in the demand teacher model on the output result of the demand teacher model is analyzed, and the Hessian matrix of the demand teacher model is given;

[0168] The Hessian matrix of the demand teacher model is subjected to Cholesky decomposition to obtain a lower triangular matrix;

[0169] The model matrix in the demand teacher model is blocked to obtain a plurality of parameter sub-matrices;

[0170] The model parameters in each parameter sub-matrix are quantized in turn, and the corresponding quantization error is given in combination with the lower triangular matrix;

[0171] The process of quantizing the model parameters is repeated until all the model parameters are quantized and the quantization error converges, and a compressed model matrix is obtained;

[0172] The compressed model matrix is integrated into the demand student model to complete the construction of the demand student model;

[0173] The demand teacher model or the demand student model is used as a material demand generation model.

[0174] In a specific embodiment, a demand teacher model and a demand student model are prepared, wherein the demand teacher model is a large language model that has been trained and has high performance and large parameter quantity, and the demand student model is a lightweight model. Through the construction of a two-stage distillation framework, through the first stage and the second stage, the output performance of the demand student model is ensured while the parameter quantity and the calculation cost are reduced.

[0175] The raw score distribution of the demand teacher model obtained through the first stage is combined with the model matrix of the demand teacher model to analyze the influence of the model matrix in the demand teacher model on the output result of the demand teacher model. For each weight parameter in the model matrix, the second derivative thereof is calculated to give the Hessian matrix of the demand teacher model. Then, the Hessian matrix is subjected to Cholesky decomposition to obtain a lower triangular matrix L, wherein H=L·L T , H is the Hessian matrix, and L T is the transpose matrix of the lower triangular matrix L.

[0176] The weight matrix is divided into multiple small blocks, i.e., parameter sub-matrices, and the model parameters in each block are quantized one by one. The quantization formula is:

[0177]

[0178] wherein W quant is the quantized model parameter, W float is the floating-point parameter of the model parameter, and scale is a quantization scale factor for mapping the floating-point parameter to the quantization range.

[0179] The error between the demand teacher model output before and after quantization is analyzed and calculated, and the corresponding quantization error is given. According to the quantization error, the quantized model parameters are adjusted to improve the accuracy of the model. The process of quantization and adjustment is repeated until the quantization of all model parameters is completed. The compressed model matrix is integrated into the demand student model to complete the construction of the demand student model. Through quantization and adjustment, the demand student model reduces the parameter quantity while maintaining high performance.

[0180] According to different actual situations, the demand teacher model or the demand student model can be used as a material demand generation model. For example, when the equipment can meet the operation of the demand teacher model, the demand teacher model can be used as the material demand generation model. When a quick result is needed, the demand student model is used as the material demand generation model, which is not limited. In this example, the trained material demand generation model is used as the demand teacher model, and the demand student model is obtained by completing model compression.

[0181] Further, the deployment and optimization of the demand teacher model and the demand student model are also included:

[0182] The demand teacher model is deployed on a cloud device, and the demand student model is deployed on an edge device;

[0183] The demand student model obtains the demand influence file on the edge device to give the corresponding response power material demand result and synchronizes the response power material demand result to the cloud device;

[0184] The cloud device responds to the edge device, and the demand teacher model audits the response power material demand result to give an audit result;

[0185] Based on the audit result, the demand student model is optimized.

[0186] In a specific implementation, the demand student model is deployed on the edge device, i.e., runs locally, to generate a corresponding first material demand file. To ensure the consistency of the output result of the edge device with that of the cloud device, the edge device will periodically upload the output result to the cloud device. After receiving the output result uploaded by the edge device, the cloud device will audit the output result.

[0187] The cloud device compares the output result uploaded by the edge device with the expected result of the cloud device. It checks whether the output result is consistent. If inconsistency is found, further analysis is performed to find the cause. If an error is found, i.e., the audit result is failed, the cloud device sends feedback to the edge device, requiring the edge device to re-execute the task or update the corresponding demand student model. Otherwise, if no error is found, i.e., the audit result is passed, the cloud device does not send feedback information to the edge device.

[0188] In order to ensure the security of information transmission between the cloud device and the edge device and improve the efficiency of data transmission, the result of responding to the demand for power materials is compressed and encrypted for transmission.

[0189] Further, the result of responding to the demand for power materials is synchronized to the cloud device, specifically including:

[0190] The result of responding to the demand for power materials is compressed to obtain a demand binary file;

[0191] Based on a pre-configured encryption algorithm, the demand binary file is encrypted to obtain a demand encrypted file;

[0192] Based on an encryption channel between the cloud device and the edge device, the demand encrypted file is transmitted to the cloud device;

[0193] The cloud device responds to the edge device, and the demand teacher model audits the result of responding to the demand for power materials to give an audit result, specifically including:

[0194] The cloud device obtains the demand encrypted file from the corresponding encryption channel;

[0195] According to the device model of the edge device, a corresponding decryption key is searched from a key library to decrypt the demand encrypted file to obtain a compressed file;

[0196] The compressed file is decompressed, and the demand teacher model audits the decompressed compressed file to give an audit result.

[0197] In a specific embodiment, after the edge device obtains the result of responding to the power material demand, the result of responding to the power material demand is compressed to obtain a demand binary file. In the present example, the compression of the result of responding to the power material demand can adopt Huffman coding, LZW coding, Gzip, Bzip2 or LZMA, and in other embodiments, other file compression methods can be adopted, which are not limited. After completing the file compression, the demand binary file is encrypted based on the encryption algorithm pre-configured in the edge device to obtain a demand encrypted file. A corresponding encryption algorithm is pre-configured in each edge device of the present application, and different encryption algorithms are configured according to the actual use scene of the edge device to adapt to the encryption demand of different edge devices. The encryption algorithm can be AES, DES, 3DES and RC4, etc. Based on the encryption channel of the cloud device and the edge device, the demand encrypted file is transmitted to the cloud device, for example, a secure communication channel is established using TLS (Transport Layer Security).

[0198] The cloud device obtains the demand encrypted file of the edge device from the corresponding encryption channel. According to the device model of the edge device, the corresponding decryption key is searched from the key library to decrypt the demand encrypted file to obtain a compressed file. The device model of the edge device is a unique identifier of each edge device, and the corresponding decryption key is searched from the key library through the device model of the edge device. The key library is a database used by the cloud device to store encryption and decryption keys of each edge device, and is updated according to the communication with the edge device. The compressed file is decompressed, and the decompressed compressed file is audited by the demand teacher model to give an audit result.

[0199] In the interaction process between the cloud device and the edge device, the transmitted file is compressed and then encrypted, which can improve the transmission efficiency and reduce the data amount in the encryption process to improve the efficiency of data encryption. The file encryption and the encryption channel are combined to further ensure the interaction security of the cloud device and the edge device.

[0200] S104: The first material demand file is detected in combination with the material demand standard, and the first material demand file is associated with the standard clauses corresponding to the material demand standard to give a result of responding to the power material demand.

[0201] In particular, the material demand standard includes a basic material standard, a professional material standard, and a dynamic material standard. The basic material standard is used to verify the uniqueness of the power material code, ensuring that the material code of each power material is unique, preventing duplicate or incorrect material codes. The professional material standard includes technical parameters of various power materials and corresponding equipment requirements, such as matching requirements of insulation level and voltage level, ensuring that the insulation level of the equipment meets the requirements of its voltage level, ensuring the safety and reliability of the power material. The dynamic material standard is associated with real-time inventory and procurement cycle data, ensuring the rationality and timeliness of procurement demand, reducing inventory backlog and shortage risk.

[0202] Further, in combination with the material demand standard, the first material demand file is subjected to difference detection, and the first material demand file is associated with the standard clauses corresponding to the material demand standard to give the result of responding to the power material demand, specifically including:

[0203] Based on the basic material standard, the uniqueness of various power materials in the first material demand file is verified, and the corresponding color is selected for marking;

[0204] Based on the professional material standard, the technical parameters and corresponding equipment requirements of various power materials in the first material demand file are matched, and the corresponding color is selected for marking;

[0205] Based on the dynamic material standard, the quantity of various power materials in the first material demand file is associated with the dynamic inventory, and the corresponding color is selected for marking;

[0206] Traverse the first material demand file to give the result of responding to the power material demand.

[0207] In a specific embodiment, it is determined whether the material code of each power material in the first material demand file exists uniquely in the existing power material code library. If yes, no marking is needed, and if not, the corresponding material code in the first material demand file is highlighted with the color corresponding to the basic material standard.

[0208] It is determined whether the technical parameters and corresponding equipment requirements of various power materials are matched, and whether the voltage level of various power materials matches its insulation level. If the voltage level meets the corresponding insulation level, it meets the standard and no marking is needed, otherwise, the relevant technical parameters and equipment requirements of the corresponding power material in the first material demand file are highlighted with the color corresponding to the professional material standard.

[0209] The relationship between the quantity of each power material in the first material demand file and the quantity in the dynamic inventory can be understood as that the quantity in the dynamic inventory is real-time changing, and there is a certain difference between the dynamic inventory when the first material demand file is given and the dynamic inventory in the first material demand file. The rationality of the first material demand file needs to be checked according to the real-time inventory and the procurement cycle. When the difference is within the corresponding range, no marking is made, otherwise, when the difference exceeds the corresponding range, the corresponding power material in the first material demand file is highlighted and marked with the color corresponding to the dynamic material standard.

[0210] In a specific example, the first material demand file is marked, first, the material code is judged based on the basic material standard, for example, the material code is 00001, the material code exists only in the power material code library, and no marking is needed. Then, the judgment is made based on the professional material standard, the power material with the material code 00001 has a device voltage level of 10kV and a device insulation level of 30kV, and the device voltage level is lower than the device insulation level, which meets the requirements, and no marking is needed. Based on the dynamic material standard, the quantity of the material with the material code 00001 in the real-time inventory is 50, the planned procurement quantity in the first material demand file is 30, the demand quantity is 70, 50+30>70, which meets the demand, but (50+30-70) / 70=1 / 7∈[0,0.2], which is within the unsafe quantity range, and the corresponding warning color of the dynamic material standard is used for highlighting and marking. For example, the quantity of the material with the material code 00001 in the real-time inventory is 75, the planned procurement quantity in the first material demand file is 50, the demand quantity is 100, 75+50>100, which meets the demand, and (75+50-100) / 100=0.25∈[0.2,0.4], which is within the safe quantity range, and no marking is needed. When the planned procurement quantity is too much, it reaches the range of overstock, and the corresponding color marking is also made.

[0211] Through comprehensive checking and marking of the first material demand file, it is ensured that the power material demand meets the basic, professional and dynamic material standards, the accuracy and rationality of the power material demand are improved, and it is also convenient to quickly identify and handle key differences.

[0212] The present application combines the power industry standards to ensure the compliance of the power material demand, and also deploys demand teacher models and demand student models on the cloud and the edge to meet the use demand of edge sites (such as substations), and maximizes the use efficiency of the material demand generation model.

[0213] Reference Figure 6 The embodiment of the present application provides a power material demand response device, which comprises:

[0214] The data acquisition module 201 is configured to acquire a demand influence file of the electric power material;

[0215] The data processing module 202 is configured to perform structural processing on the demand influence file according to different file processing strategies matched according to the file type of the demand influence file, to obtain material demand data.

[0216] The demand generation module 203 is configured to perform auditing processing on the material demand data according to a pre-constructed material demand generation model and by fusing an electric power multi-element knowledge graph, to obtain a first material demand file.

[0217] The demand review module 204 is configured to perform difference detection on the first material demand file in combination with a material demand standard, and to establish association between the first material demand file and a standard clause corresponding to the material demand standard, to give a result of responding to the electric power material demand.

[0218] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the described modules can refer to the corresponding process in the foregoing method embodiments, which will not be described herein.

[0219] Although the preferred embodiments of the present application have been described, those skilled in the art can make further changes and modifications to these embodiments once they understand the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application. Obviously, those skilled in the art can make various modifications and changes to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and changes of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these modifications and changes.

Claims

1. A method for responding to the demand for electrical materials, characterized in that, The method comprises the following steps: obtaining a demand impact file of power materials; according to the file type of the demand impact file, matching different file processing strategies to perform structured processing on the demand impact file to obtain material demand data; according to the sequence length of the material demand data, using a rotating position encoding extrapolation technology to expand the processing window of the material demand generation model; analyzing the correlation similarity between each demand vector of the material demand data and each material vector in the power multi-knowledge graph, and adjusting the attention weight in the attention mechanism in the material demand generation model, wherein the power multi-knowledge graph comprises material vectors of each power material, and each material vector of the power material comprises technical parameters, coding rules and procurement specifications of the corresponding power material in the power industry standard; substituting the material demand data into the adjusted material demand generation model to obtain a first material demand file; the material demand generation model comprises a demand input layer, an attention mechanism, a demand reward layer and a demand output layer; the attention mechanism is connected with the demand input layer, the material demand data enters the attention mechanism through the demand input layer, and the core demand features in the material demand data are extracted; the demand output layer is connected with the attention mechanism, and an intermediate demand file is generated based on the core demand features; the demand reward layer is connected with the demand output layer, and the compliance of the intermediate demand file is analyzed to obtain a total demand reward and feedback to the attention mechanism and the demand output layer; wherein the demand reward layer comprises a basic reward layer, an incremental reward layer, a violation penalty layer and a reward fusion layer; the basic reward layer is connected with the demand output layer, and is used to analyze the matching degree between the intermediate demand file and the corresponding procurement specification to obtain an intermediate demand basic reward; the incremental reward layer is connected with the demand output layer, and is used to analyze the fluctuation difference between each value in the intermediate demand file and the corresponding standard value to obtain an intermediate demand incremental reward; the violation penalty layer is connected with the demand output layer, and is used to analyze the violation between each data in the intermediate demand file to give an intermediate demand penalty term; the reward fusion layer is connected with the basic reward layer, the incremental reward layer and the violation penalty layer respectively, and is used to fuse the intermediate demand basic reward, the intermediate demand incremental reward and the intermediate demand penalty term to obtain the total demand reward and feedback to the attention mechanism and the demand output layer; combining the material demand standard, performing difference detection on the first material demand file, and establishing an association between the first material demand file and the standard clauses corresponding to the material demand standard to give a result of responding to the power material demand.

2. The method of claim 1, wherein, The file type comprises at least one of a text, a table and an image; according to the file type of the demand impact file, matching different file processing strategies to perform structured processing on the demand impact file to obtain material demand data, specifically comprising: if the file type of the demand impact file is a text, performing word segmentation and part-of-speech tagging on the demand impact file, and converting the demand impact file into structured data to obtain corresponding material demand data; if the file type of the demand impact file is a table, performing structural analysis on the demand impact file, analyzing the relationship between the content and position of each cell in the demand impact file, extracting the content of the demand impact file and giving corresponding material demand data; If the file type of the demand influence file is an image, the text in the demand influence file is enhanced and recognized, and is converted into structured data, and corresponding material demand data is given.

3. The method of claim 1, wherein the power asset demand response is performed in response to a request from a power grid operator. The construction of the material demand generation model is obtained through the following steps: Obtain historical material demand data; Analyze the matching degree of the historical material demand data and the corresponding procurement specification to obtain a demand base reward; Analyze the fluctuation difference between each value in the historical material demand data and the corresponding standard value to obtain a demand increment reward; Analyze the violation between each data in the historical material demand data to give a demand penalty term; Fuse the demand base reward, the demand increment reward and the demand penalty term to train the initial neural network model until convergence to obtain a first demand generation model; Insert a low-rank matrix into the attention mechanism of the first demand generation model, freeze the model parameters of the first demand generation model, and adjust and optimize the parameters corresponding to the low-rank matrix to obtain a material demand generation model.

4. The method of claim 1, wherein the power asset demand response is performed in response to a request from a power grid operator. Also includes: A two-stage distillation framework is constructed, wherein the first stage is used to output the original score distribution of the demand teacher model, and the second stage is used to minimize the output difference between the demand student model and the demand teacher model in the electric power material demand generation task; Analyze the influence of the model matrix in the demand teacher model on the output result of the demand teacher model to give the Hessian matrix of the demand teacher model; Cholesky decomposition is performed on the Hessian matrix of the demand teacher model to obtain a lower triangular matrix; The model matrix in the demand teacher model is blocked to obtain a plurality of parameter sub-matrices; The model parameters in each parameter sub-matrix are quantized in turn, and the corresponding quantization error is given in combination with the lower triangular matrix; Repeat the quantization process of the model parameters until all the model parameters are quantized and the quantization error converges to obtain a compressed model matrix; The compressed model matrix is integrated into the demand student model to complete the construction of the demand student model; The demand teacher model or the demand student model is used as a material demand generation model.

5. The method of claim 4, wherein the power asset demand response is performed in response to a request from a power grid operator. Also includes the deployment and optimization of the demand teacher model and the demand student model: The demand teacher model is deployed on a cloud device, and the demand student model is deployed on an edge device; The demand student model obtains the demand influence file on the edge device to give the corresponding response electric power material demand result and synchronizes the response electric power material demand result to the cloud device; The cloud device responds to the edge device, and the demand teacher model audits the response electric power material demand result to give an audit result; Based on the audit result, the demand student model is optimized.

6. The method of claim 5, wherein the power asset demand response is performed in response to a request from a power grid operator. Synchronize the response electric power material demand result to the cloud device, specifically including: Compress the response electric power material demand result to obtain a demand binary file; Based on a pre-configured encryption algorithm, encrypt the demand binary file to obtain a demand encrypted file; Based on the encryption channel of the cloud device and the edge device, transmit the demand encrypted file to the cloud device; The cloud device responds to the edge device, and the demand teacher model audits the response electric power material demand result to give an audit result, specifically including: The cloud device obtains the demand encrypted file from the corresponding encryption channel; According to the device model of the edge device, a corresponding decryption key is searched from a key library, a demand encrypted file is decrypted to obtain a compressed file; The compressed file is decompressed, and the demand teacher model audits the decompressed compressed file to give an audit result.

7. The method of claim 1, wherein the power asset demand response is performed in response to a request from a power grid operator. The material demand standard includes a basic material standard, a professional material standard, and a dynamic material standard; In combination with the material demand standard, a first material demand file is subjected to difference detection, and the first material demand file is associated with a standard clause corresponding to the material demand standard to give a result of responding to the demand for electric power materials, specifically including: Based on the basic material standard, the uniqueness of various electric power materials in the first material demand file is checked, and a corresponding color is selected for marking; Based on the professional material standard, the technical parameters of various electric power materials in the first material demand file and the corresponding equipment demand are matched, and a corresponding color is selected for marking; Based on the dynamic material standard, the quantity of various electric power materials in the first material demand file is associated with the dynamic inventory, and a corresponding color is selected for marking; The first material demand file is traversed to give a result of responding to the demand for electric power materials.

8. An electric power asset demand response apparatus, characterized by, The method for responding to the demand for electric power materials as claimed in any one of claims 1-7, comprising: a data acquisition module configured to acquire a demand influence file of electric power materials; a data processing module configured to match different file processing strategies according to the file type of the demand influence file, and to perform structured processing on the demand influence file to obtain material demand data; The demand generation module is configured to expand the processing window of the material demand generation model by using a rotating position encoding extrapolation technique according to the sequence length of the material demand data; analyze the correlation similarity between each demand vector of the material demand data and each material vector in the power multi-knowledge graph, and adjust the attention weight in the attention mechanism in the material demand generation model, wherein the power multi-knowledge graph includes material vectors of each power material, and each material vector of the power material includes technical parameters, coding rules and procurement specifications of the corresponding power material in the power industry standard; and substitute the material demand data into the adjusted material demand generation model to obtain a first material demand file; the material demand generation model includes a demand input layer, an attention mechanism, a demand reward layer and a demand output layer; the attention mechanism is connected with the demand input layer, the material demand data enters the attention mechanism through the demand input layer, and core demand features in the material demand data are extracted; the demand output layer is connected with the attention mechanism, and an intermediate demand file is generated based on the core demand features; the demand reward layer is connected with the demand output layer, analyzes compliance in the intermediate demand file, obtains a total demand reward and feeds back to the attention mechanism and the demand output layer; wherein the demand reward layer includes a basic reward layer, an incremental reward layer, a violation penalty layer and a reward fusion layer; the basic reward layer is connected with the demand output layer, and is configured to analyze the matching degree between the intermediate demand file and the corresponding procurement specification to obtain an intermediate demand basic reward; the incremental reward layer is connected with the demand output layer, and is configured to analyze the fluctuation difference between each value in the intermediate demand file and the corresponding standard value to obtain an intermediate demand incremental reward; the violation penalty layer is connected with the demand output layer, and is configured to analyze the violation between each data in the intermediate demand file to give an intermediate demand penalty term; and the reward fusion layer is connected with the basic reward layer, the incremental reward layer and the violation penalty layer, respectively, and is configured to fuse the intermediate demand basic reward, the intermediate demand incremental reward and the intermediate demand penalty term to obtain the total demand reward and feed back to the attention mechanism and the demand output layer; The demand review module is configured to detect differences in the first material demand file in combination with the material demand standard, associate the first material demand file with the standard clauses corresponding to the material demand standard, and give a result of responding to the power material demand.

Images