Knowledge graph retrieval intelligent electric storage system based on feedforward neural network prediction

Abstract

The invention discloses a knowledge graph retrieval intelligent power storage system based on a feedforward neural network, and relates to the field of power equipment accessory management. The method comprises the following steps: constructing a power equipment resource classification model and a meteorological disaster risk periodic evaluation model in a coverage storage range; obtaining static data and dynamic data of different power devices; predicting the replacement demand level of the equipment through the feedforward neural network model; a model prediction evaluation result is used as a data source of the knowledge graph warehousing system, and terminal real-time fault prediction evaluation query, periodic equipment purchase guidance and backlog warning are realized. According to the invention, the storage demand can be intelligently predicted according to the actual state of the power equipment, the accuracy and response speed of storage management are improved, and an efficient spare part support scheme is provided for power network maintenance and equipment storage.

Description

Technical Field

[0001] The present invention belongs to the technical field of warehouse management, and in particular relates to a knowledge graph retrieval intelligent power storage system based on feedforward neural network prediction. Background Art

[0002] Smart supply chain refers to the realization of intelligent supply chain, network collaboration, and intelligent decision-making through cloud computing, big data, and machine learning, combined with current supply chain management theories and methods. The power grid is highly dependent on the accurate and orderly supply of power grid materials in terms of power transmission and transformation project construction and the safe and stable operation of the power grid. Therefore, the material supply chain is a very important supply chain in the power industry. In actual work, since power grid projects have strict time limits for material supply, it is difficult to accurately predict the supply cycle. Pre-purchase is likely to cause inventory backlogs, and simultaneous purchases will easily cause delayed purchases. Ultimately, the imbalance between supply and demand of materials will lead to project construction progress, and the losses caused will have a huge impact on power grid construction and the national economy.

[0003] In the field of power warehouse management, the relatively traditional supply chain management method is still adopted, which is mostly through advance forecasting or simultaneous procurement with projects. The procurement plan is mostly based on the past experience of procurement personnel. With the continuous deepening of the development of the power industry, it is impossible to achieve lean management and flexible supply requirements by relying solely on manual and experience-based methods to manage supply chain procurement plans. In the field of smart supply chain, there have been studies and applications using technologies such as big data, the Internet of Things, edge computing, and artificial intelligence to explore high business continuity, low shortages, and low inventory. A relatively mature theoretical system and implementation plan have been established. Its prediction principle for market demand can be used as a reference theory for the construction of smart power storage systems. Summary of the invention

[0004] The present invention aims to solve at least one of the technical problems existing in the prior art.

[0005] The technical solution of the present invention is: a knowledge graph retrieval intelligent power storage method based on a feedforward neural network, comprising the following steps:

[0006] S1: Construct a classification model for power equipment resources within the storage area and a periodic assessment model for meteorological disaster risks;

[0007] S2: Obtain static data, dynamic data of different power equipment and the model results in S1 as training sets;

[0008] S3: Predict and evaluate the replacement demand level of equipment through a feedforward neural network model;

[0009] S4: The model prediction evaluation results serve as the data source of the knowledge graph warehousing system to realize real-time terminal status query, periodic equipment procurement guidance and backlog warning.

[0010] Furthermore, the classification model of power equipment resources in step S1 divides the power equipment in the lines within the coverage of the warehouse into primary equipment and secondary equipment: primary equipment is equipment that directly produces, transmits and distributes electric energy, including generators, transformers, switchgear, power lines, mutual inductors, lightning arresters, etc.; secondary equipment is equipment that controls, regulates, protects and monitors primary equipment, including measuring meters, relay protection and automatic devices, operating electrical appliances and DC power supply equipment, etc. Neural network fault assessment is performed on the different operating characteristics of the above-mentioned classified equipment.

[0011] Furthermore, the meteorological disaster risk periodic assessment model in step S1 includes:

[0012]

[0013] Where: WREM(t) is the evaluation parameter, H i is the severity of historical disasters, ranging from 1 to 5, H max is the maximum severity of historical disasters, F i is the failure frequency (times / month), F max is the maximum fault frequency, λ is the time attenuation coefficient, λ=0.15, t i is a historical time point, t is the current time node, and T is the total monthly data.

[0014] Furthermore, the training set in step S2 will include static data and dynamic data, wherein the static data is the number of uses and connection relationships of the power equipment on the lines within the coverage of the warehouse, and the dynamic data is the operating years, performance degradation degree, and carried business type of the power equipment after numerical processing;

[0015] Furthermore, the power equipment fault assessment algorithm of the feedforward neural network model in step S3 is specifically:

[0016] A1: The dynamic and static indicators of the power equipment are numerically processed and used as the sample data training set;

[0017] A2: The replacement demand level of the equipment to be repaired is obtained as the verification result through historical verification data and manual analysis of A1 sample data;

[0018] A3: Take the sample data training set as input, set the initial weights of the feedforward neural network model, perform learning and training in the feedforward neural network model, obtain the prediction results, compare the prediction results with the verification results, perform error backpropagation, update the neural network weight parameters, repeat the training process until the model accuracy meets the expected requirements, and obtain the feedforward neural network model for the knowledge graph data source.

[0019] Furthermore, the input layer P and output layer Q of the sample data training set in step A1 in the neural network are:

[0020] P=[P 11 ,P 12 …P ij ,Z,C,] (2)

[0021] Q=[Q1,Q2,…Q i ] (3)

[0022] Among them, P ij Q is the dynamic data of the jth numerical quantization processing in the i-th power equipment, Z is the predicted month information, and C is the climate information of the predicted time period after quantification. i The feedforward neural network model predicts the replacement demand level of the i-th power equipment.

[0023] Furthermore, in step A2, the replacement demand level of the equipment to be repaired is obtained by manually analyzing and judging the A1 sample data through historical verification data, specifically:

[0024] Q c =[Q c1 ,Q c2 ,…Q ci ] (4)

[0025] Among them, Q ci It is the verification result after manual analysis and judgment.

[0026] Furthermore, in step A3, the AdamW algorithm is used to optimize the weight parameters of the feedforward neural network to improve the accuracy of power equipment fault assessment, specifically including:

[0027] A3-1: The replacement demand levels of power equipment are divided into "urgent", "suspected" and "normal", and they are mapped numerically to facilitate the output and error backpropagation of the neural network.

[0028] A3-2: The loss function is defined as MAE, which is used to calculate the error between the prediction evaluation result and the verification result:

[0029]

[0030] Where N is the total number of samples of power equipment, and L(θ) is the loss function of the neural network model.

[0031] A3-3: Use a feedforward neural network for probability prediction evaluation. The core prediction probability calculation formula is:

[0032]

[0033] σ(x)=max(0,x) (7)

[0034] Among them, σ is the activation function, and the multi-layer feedforward neural network adopts a multi-layer ReLU activation structure; W i is the weight matrix; b i is the bias term; L is the number of neural network layers.

[0035] A3-4: Use the AdamW algorithm to minimize the loss function L(θ) and update the weight parameters of the neural network. The optimization process of AdamW is as follows:

[0036] Momentum Update:

[0037]

[0038] Among them, m t is the momentum term of the tth iteration, recording the first-order momentum of the gradient, v t It is the accumulation of the square of the gradient at the tth iteration, recording the second-order momentum of the gradient. t-1 is the momentum term of the t-1th iteration, v t-1 is the accumulation of the squared gradient of the t-1th iteration. β1 and β2 are momentum coefficients used to control the decay rate of the first-order and second-order momentum.

[0039] Bias correction:

[0040]

[0041] in and They are the momentum term and the squared gradient accumulation after bias correction, respectively.

[0042] Weight update (introducing weight decay):

[0043]

[0044] Among them, θ t is the weight parameter of the tth iteration, θ t+1 is the updated value of the weight parameter for the t+1th iteration, η is the learning rate, λ θ is the weight decay coefficient, ∈ is the smoothing term.

[0045] Furthermore, the knowledge graph storage system in step S4 includes:

[0046] B1: Build a link portrait of power equipment, receive the results of the feedforward neural network model prediction and evaluation model, and organize and build a fault description knowledge graph;

[0047] B2: Generate knowledge graphs for power equipment warehouse space allocation module, order preprocessing module and backlog warning module based on the prediction and evaluation model;

[0048] B3: Provides a breadth-first search strategy based on graphics processor unit (GPU) programming to achieve intelligent question-answering and fast retrieval, improve the usability of the graph, and maximize the value of data.

[0049] Furthermore, the power equipment link portrait in step B1 includes a three-layer data preprocessing framework based on data input, data preprocessing, and data analysis.

[0050] Furthermore, the fault description knowledge graph constructed in step B1 includes a triple expression form, which is a common representation form of the knowledge graph, namely

[0051] G=(E,R,S) (12)

[0052] Among them, E is the entity set in the knowledge base, R is the relationship set, and S is the triple set. By modeling the power system and equipment components covered by the warehouse and combining them with the prediction and evaluation results of the feedforward neural network, the knowledge graph is organized and constructed to realize the structured data storage of fault information triples.

[0053] Furthermore, the power equipment warehouse space allocation module in step B2 includes:

[0054] Analyze the properties of power equipment to form a basic property profile of each device;

[0055] Allocate warehouse storage space for incoming goods based on incoming order forecasts and basic attribute portraits;

[0056] Allocate storage space in the warehouse based on the status of the stored power equipment, the remaining storage space, and the power equipment order forecast.

[0057] Furthermore, the order pre-processing module in step B2 includes:

[0058] The degree of equipment replacement demand based on feedforward neural network prediction and evaluation model;

[0059] Consider the dynamic adjustment mechanism of equipment replacement cycle;

[0060] Emergency reserve combined with meteorological disaster risk assessment

[0061] Consider the provisioning redundancy quota during the equipment application circulation cycle.

[0062] Furthermore, the dynamic adjustment mechanism also includes:

[0063] Regularly update the forecast optimal inventory of each type of equipment;

[0064] Adjust the warning threshold according to the actual inventory consumption rate;

[0065] The warning level is revised based on seasonal factors.

[0066] Furthermore, the backlog warning module in step B2 proposes a method for evaluating inventory backlog risk through the backlog warning index OWI. The specific calculation formula is:

[0067]

[0068] Where: S i is the current inventory, S i * is the predicted optimal inventory, B i is the unit price and M is the number of inventory types.

[0069] Furthermore, the backlog warning module also includes a warning level classification mechanism:

[0070] When OWI ≥ 30%, the system issues a red warning, indicating a serious backlog risk;

[0071] When 20% ≤ OWI < 30%, the system issues a yellow warning, indicating that there is a medium backlog risk;

[0072] When 10%≤OWI<20%, the system issues a blue warning, indicating a slight backlog risk;

[0073] When OWI < 10%, the system shows that the inventory level is normal.

[0074] Furthermore, the breadth-first search strategy based on graphics processor unit (GPU) programming in step B3 is a knowledge graph path search method based on GPU parallel computing. The method adopts a breadth-first search strategy, and the specific execution process is as follows: First, the method receives the knowledge graph database D i As the data input source, it also determines the starting vertex V of the search. s and the target vertex V i The search process starts from the starting vertex V s Initially, in each iteration, the algorithm searches in parallel for all adjacent vertices V of the currently visited vertex i For each adjacent vertex V that is found i, the system will dynamically allocate an independent GPU computing unit and the corresponding distributed memory unit to achieve parallel processing of search tasks. i With the target vertex V t If there is a match, the algorithm immediately returns the currently found path as the result; otherwise, the search loop continues to search for the next layer of adjacent nodes until the termination condition is met.

[0075] In the work, the present invention integrates the innovative solution of feedforward neural network prediction and knowledge graph, which can effectively realize the lean management and flexible supply of power material storage, intelligently predict and reasonably allocate material demand, and reduce the economic losses caused by the imbalance of material supply and demand. The present invention promotes the innovative development of the smart supply chain in the power industry, reduces out-of-stock and inventory costs, and improves the construction efficiency of power engineering projects and the safety and stability of the power grid. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 It is a schematic diagram of the process of the present invention;

[0077] Figure 2 This is a schematic diagram of the feedforward neural network prediction and evaluation model;

[0078] Figure 3 A schematic diagram of a three-layer data preprocessing framework for power equipment link profiling;

[0079] Figure 4 A schematic diagram of the structure for generating a knowledge graph;

[0080] Figure 5 Schematic diagram of a breadth-first search strategy for programming on a graphics processor unit (GPU). DETAILED DESCRIPTION

[0081] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0082] like Figure 1 As shown, the present invention provides a knowledge graph retrieval intelligent power storage method based on feedforward neural network prediction, and the method is implemented for a power equipment storage warehouse in a certain area of ​​Jiangsu and the power network within its storage coverage, including the following steps:

[0083] S1: Construct a classification model for power equipment resources within the storage area and a periodic assessment model for meteorological disaster risks;

[0084] S2: Obtain static data, dynamic data of different power equipment and the model results in S1 as training sets;

[0085] S3: Predict and evaluate the replacement demand level of equipment through a feedforward neural network model;

[0086] S4: The model prediction evaluation results serve as the data source of the knowledge graph warehousing system to realize real-time terminal status query, periodic equipment procurement guidance and backlog warning.

[0087] The requirements for constructing the classification model of power equipment resources within the storage range and the meteorological disaster risk periodic assessment model in step S1 of this embodiment are as follows:

[0088] For the classification model of power equipment resources, the power equipment in the lines within the coverage of the warehouse is divided into primary equipment and secondary equipment: primary equipment is equipment that directly produces, transmits and distributes electric energy, including generators, transformers, switchgear, power lines, mutual inductors, lightning arresters, etc.; secondary equipment is equipment that controls, regulates, protects and monitors primary equipment, including measuring meters, relay protection and automatic devices, operating electrical appliances and DC power supply equipment, etc. According to the different working scenarios of primary and secondary equipment and the basic parameters of their power equipment, different feedforward neural network prediction and evaluation models and retrograde targeted training will be used to avoid mutual influence between power equipment with significant attribute differences, which will hinder the convergence of the neural network prediction and evaluation model. Finally, the classification model of power equipment resources becomes one of the input parameters of the feedforward neural network prediction model.

[0089] In this embodiment, the devices within the warehouse coverage area are divided into two categories:

[0090] (1) Primary equipment: including 86 types of 35kV~220kV transformers, 53 types of circuit breakers, 31 types of lightning arresters, etc.;

[0091] (2) Secondary equipment: including 128 types of relay protection devices, 45 types of automatic devices, etc.

[0092] For the meteorological disaster risk periodic assessment model, big data meteorological warning information will be used to analyze the impact of different meteorological disasters on power equipment. According to the nature and characteristics of meteorological disasters, they will be divided into different types such as rainstorms, blizzards and typhoons, and key factors affecting the safe operation of power equipment, such as rainfall, wind speed and temperature, will be extracted to understand and identify the impact of periodic climate risks on the power system and power equipment storage warehouses; a meteorological disaster risk assessment model will be established to classify different types of meteorological disasters and quantify the output, which will become one of the input parameters of the feedforward neural network prediction model.

[0093] The meteorological disaster risk periodic assessment model includes:

[0094]

[0095] Where: WREM(t) is the evaluation parameter, H i is the severity of historical disasters, ranging from 1 to 5, H max is the maximum severity of historical disasters, F i is the failure frequency (times / month), F max is the maximum fault frequency, λ is the time attenuation coefficient, and λ=0.15; t i is a historical time point, t is the current time node, T is the total amount of monthly data, and T=60.

[0096] The acquisition of the data set in step S2 of this embodiment is an important prerequisite, and the related operations are related to the accuracy and authenticity of the information source obtained by the knowledge graph. For data set acquisition, it is necessary to provide static indicators of power equipment, that is, the physical connection relationship between power equipment to establish node connections and fill in triple information for the knowledge graph; it is necessary to provide dynamic indicators of power equipment, that is, the operating years, performance degradation degree, and business types of power equipment after numerical processing. For the training of feedforward neural network prediction and evaluation models, it is necessary to consider the impact of differences between different power equipment on the generalization and convergence ability of the neural network model. While improving the accuracy of the training set, it is also necessary to ensure that the model has good generalization ability and can realize the prediction of unidentified power equipment samples.

[0097] Based on the above content, step S2 is specifically as follows: In this embodiment, the connection status of the power equipment storage warehouse in the area and the power network it covers is obtained. For the topological connection structure and equipment usage on the line, the static data and dynamic data in the input parameters of the step S2 training set are obtained, wherein the static data is the number of power equipment used and the connection relationship on the line within the coverage of the warehouse, and the dynamic data is the operating years, performance degradation degree, and business type of the power equipment in the power system within the coverage of the warehouse after numerical processing for one year; the other input parameters of the training set have been obtained in S1. In step S3 of this embodiment, the feedforward neural network model is used to predict and evaluate the replacement demand level of the equipment:

[0098] A1: The dynamic and static indicators of the power equipment are numerically processed and used as the sample data training set;

[0099] A2: The replacement demand level of the equipment to be repaired is obtained as the verification result through historical verification data and manual analysis of A1 sample data;

[0100] A3: Take the sample data training set as input, set the initial weights of the feedforward neural network model, perform learning and training in the feedforward neural network model, obtain the prediction results, compare the prediction results with the verification results, perform error backpropagation, update the neural network weight parameters, repeat the training process until the model accuracy meets the expected requirements, and obtain the feedforward neural network model for the knowledge graph data source.

[0101] In step A1, the dynamic and static indicators of the power equipment are numerically processed as the sample data training set, and the training set data is divided into a training set, a validation set, and a test set according to the ratio of 80%, 10%, and 10%, respectively. The input layer P and output layer Q of the sample data training set in S1 in the neural network are:

[0102] P=[P 11 ,P 12 …P ij ,Z,C,] (2)

[0103] Q=[Q1,Q2,…Q i ] (3)

[0104] Among them, P ij Q is the dynamic data of the jth numerical quantization processing in the i-th power equipment, Z is the predicted month information, and C is the climate information of the predicted time period after quantification. i The feedforward neural network model predicts the replacement demand level of the i-th power equipment.

[0105] In step A2, the replacement demand level of the equipment to be repaired is obtained by using historical verification data and manual analysis and judgment of A1 sample data as the verification result. The verification result and the training set data correspond to form an important part of the prediction error of the neural network model, which provides conditions for error back propagation and gradient descent of the neural network model. Finally, the replacement demand level of the equipment to be repaired is obtained by using historical verification data and manual analysis and judgment of A1 sample data in step A2, which is specifically:

[0106] Q c =[Q c1 ,Q c2 ,…Q ci ] (4)

[0107] Among them, Q ci It is the verification result after manual analysis and judgment.

[0108] Step A3 is an important step in building a feedforward neural network prediction and evaluation model and training. Figure 2The feedforward neural network model shown. Take the sample data training set as input, set the initial weights of the feedforward neural network model, conduct learning and training in the feedforward neural network model, obtain the prediction results, compare the prediction results with the verification results, perform error back propagation, update the neural network weight parameters, repeat the training process until the model accuracy meets the expected requirements, and obtain the feedforward neural network model for the knowledge graph data source. At the same time, step A3 also has the following requirements:

[0109] A3-1: The replacement demand levels of power equipment are divided into "urgent", "suspected" and "normal", and they are mapped numerically to facilitate the output and error backpropagation of the neural network.

[0110] A3-2: The loss function is defined as MAE, which is used to calculate the error between the prediction evaluation result and the verification result:

[0111]

[0112] Where N is the total number of samples of power equipment, and L(θ) is the loss function of the neural network model.

[0113] A3-3: Use a feedforward neural network for probability prediction evaluation. The core prediction probability calculation formula is:

[0114]

[0115] σ(x)=max(0,x) (7)

[0116] Among them, σ is the activation function, and the multi-layer feedforward neural network adopts a multi-layer ReLU activation structure; W i is the weight matrix; b i is the bias term; L is the number of neural network layers.

[0117] A3-4: Use the AdamW algorithm to minimize the loss function L(θ) and update the weight parameters of the neural network. The optimization process of AdamW is as follows:

[0118] Momentum Update:

[0119]

[0120] Among them, m t is the momentum term of the tth iteration, recording the first-order momentum of the gradient, v t It is the accumulation of the square of the gradient at the tth iteration, recording the second-order momentum of the gradient. t-1 is the momentum term of the t-1th iteration, v t-1is the accumulation of the squared gradient of the t-1th iteration. β1 and β2 are momentum coefficients used to control the decay rate of the first-order and second-order momentum.

[0121] Bias correction:

[0122]

[0123] in and They are the momentum term and the squared gradient accumulation after bias correction, respectively.

[0124] Weight update (introducing weight decay):

[0125]

[0126] Among them, θ t is the weight parameter of the tth iteration, θ t+1 is the updated value of the weight parameter for the t+1th iteration, η is the learning rate, λ θ is the weight decay coefficient, ∈ is the smoothing term.

[0127] Based on the above content, step A3 is specifically as follows:

[0128] 1) Model training: Establish a feedforward neural network prediction and evaluation model, use the AdamW algorithm to minimize the loss function L(θ) to update the weight parameters of the neural network, and use small batch random training for a total of 200 epochs, where the bitch size is 64 and α=0.001. That is, the entire training data set was trained for 200 rounds, 64 samples were randomly selected for each training, and the initial learning rate was 0.001. After repeated training and tuning, the final model achieved a high accuracy on the test set, with a prediction accuracy of 87.3% for the first device and 91.2% for the second device. The accuracy directly affects the actual application effect and reliability of the model. Therefore, it is necessary to continuously optimize the model during the training process to improve the classification accuracy to meet the needs of actual applications.

[0129] 2) Model testing: After the model training is completed, the model needs to be tested to evaluate its performance. The test set is a data set that has not been touched during the model training process and can be used to verify the generalization ability of the model. Some indicators are usually used to evaluate the performance of the model on the test set. In this embodiment, after the test set test in the total sample division, the prediction accuracy of the primary device is 82.1%, and the prediction accuracy of the secondary device is 85.9%.

[0130] Table 1 Feedforward neural network prediction evaluation model training and test accuracy

[0131]

[0132] Step S4 of this embodiment is an important step in creating a knowledge graph storage system. Establishing a knowledge graph storage system includes:

[0133] B1: Build a link portrait of power equipment, receive the results of the feedforward neural network model prediction and evaluation model, and organize and build a fault description knowledge graph;

[0134] B2: Generate knowledge graphs for power equipment warehouse space allocation module, order preprocessing module and backlog warning module based on the prediction and evaluation model;

[0135] B3: Provides a breadth-first search strategy based on graphics processor unit (GPU) programming to achieve intelligent question-answering and fast retrieval, improve the usability of the graph, and maximize the value of data.

[0136] For building the power equipment link portrait in step B1, receiving the prediction and evaluation model results of the feedforward neural network model, and organizing the construction of the fault description knowledge graph, the following requirements are required: first, according to Figure 3 The structure is used to implement a three-layer data preprocessing framework based on data input, data preprocessing, and data analysis for the power equipment link portrait. Then, according to the power system network topology link structure within the storage range of the warehouse, a Figure 4 Finally, according to the fault description knowledge graph, which includes a triple expression form, a general representation form of the knowledge graph is constructed, namely

[0137] G=(E,R,S) (12)

[0138] Among them, E is the entity set in the knowledge base, R is the relationship set, and S is the triple set. By modeling the power system and equipment components covered by the warehouse and combining them with the prediction and evaluation results of the feedforward neural network, the knowledge graph is organized and constructed to realize the structured data storage of fault information triples.

[0139] In step B2, the knowledge graphs of the power equipment warehouse space allocation module, order preprocessing module, and backlog warning module are generated according to the prediction and evaluation model, with the following requirements:

[0140] For the power equipment warehouse space allocation module, the incoming order forecast and power equipment order forecast included in the results of the feedforward neural network model prediction and evaluation model are further processed. According to the incoming order forecast and basic attribute portrait, the warehouse storage space is allocated for the incoming goods. At the same time, the storage space of the warehouse is allocated according to the storage power equipment, the remaining storage space, and the power equipment order forecast, which ultimately increases the storage space utilization rate of the warehouse by 22.7%.

[0141] For the order preprocessing module, the knowledge graph retrieval intelligent power storage method will process the following data: the output results of the neural network evaluation model, including but not limited to the overall output trend and the specific output results of the previous outputs; the results of the meteorological disaster risk assessment model; the static image of electrical equipment; the equipment application circulation cycle of the redundant quota, and provide purchase order suggestions for the next cycle. At the same time, the knowledge graph retrieval intelligent power storage method will regularly update the predicted optimal inventory of various types of equipment according to the neural network prediction and evaluation model, adjust the warning threshold, and revise the warning level, ultimately reducing the shortage rate of power equipment in the warehouse by 5.6%.

[0142] For the backlog warning module, a method of evaluating inventory backlog risk through the backlog warning index OWI is proposed. The specific calculation formula is:

[0143]

[0144] Where: S i is the current inventory, S i * is the predicted optimal inventory, B i is the unit price, and M is the number of inventory types. Combining the comparison between the warehouse before and after the use of the present invention, the OWI index dropped from 18.7% before use to 9.6% after use, and the overall blue warning inventory level of the warehouse was reduced to normal inventory. Finally, the total inventory amount of the warehouse was reduced by 18.3%.

[0145] In step B3, a breadth-first search strategy based on graphics processor unit (GPU) programming is provided to achieve intelligent question-answering and fast retrieval, improve the usability of the graph, and maximize the value of data. The specific steps are as follows:

[0146] The breadth-first search strategy based on graphics processor unit (GPU) programming is a knowledge graph path search method based on GPU parallel computing, which can effectively improve the speed of users asking questions and querying data. This method adopts a breadth-first search strategy, such as Figure 5 As shown, the specific execution process is as follows: First, the method receives the knowledge graph database D i As the data input source, it also determines the starting vertex V of the search. s and the target vertex V i The search process starts from the starting vertex V s Initially, in each iteration, the algorithm searches in parallel for all adjacent vertices V of the currently visited vertex i For each adjacent vertex V that is found i , the system will dynamically allocate an independent GPU computing unit and the corresponding distributed memory unit to achieve parallel processing of search tasks.i With the target vertex V t If the algorithm matches, the algorithm immediately returns the currently found path as the result; otherwise, the search loop continues to be executed to find the next layer of adjacent nodes until the termination condition is met. By completing the above steps, the breadth-first search strategy based on graphics processor unit (GPU) programming can speed up the average time required for manual information query of data storage devices by 35.3%, and reduce the data query and verification work time of warehouse managers by an average of 17.5%.

[0147] This example uses the following three indicators to measure warehousing efficiency, namely inventory turnover rate, project delay rate, and material supply and demand balance.

[0148] Inventory turnover rate refers to the number of times a company's inventory is sold and replaced within a certain period of time, and is usually used to measure the efficiency of inventory management and the liquidity of goods. The specific calculation method is to divide the sales cost of a single year by the average inventory of that year; the project delay rate refers to the difference between the actual completion time and the planned completion time of the power project, which is usually used to evaluate the effectiveness of project management. The specific calculation method is to subtract the planned completion time from the actual completion time, and then divide the result by the planned completion time. The balance between supply and demand of materials refers to the degree of match between the supply capacity and demand capacity of materials within a certain period of time. It is evaluated by comparing the actual demand with the actual supply. The specific power storage effect is shown in Table 2:

[0149] Table 2 Comparison of power storage effects

[0150]

[0151] As can be seen from Table 2, the present invention has shown significant effects in inventory management. Compared with traditional power warehousing, the inventory turnover rate of intelligent warehousing has been significantly improved, inventory liquidity has been enhanced, and capital turnover has been faster. The project delay rate continues to remain at a low level. The balance between supply and demand of materials has been significantly improved, reflecting the effectiveness of supply and demand matching and the optimization of resource utilization. Overall, intelligent power warehousing is superior to traditional models in terms of efficiency and management level, and improves operational benefits.

[0152] It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0153] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A knowledge graph retrieval intelligent power storage method based on feedforward neural network, characterized in that: The steps include: S1: Construct a classification model for power equipment resources within the storage area and a periodic assessment model for meteorological disaster risks; S2: Obtain static data, dynamic data of different power equipment and the model results in S1 as training sets; S3: Predict and evaluate the replacement demand level of equipment through a feedforward neural network model; S4: The model prediction evaluation results serve as the data source of the knowledge graph warehousing system to realize real-time terminal status query, periodic equipment procurement guidance and backlog warning.

2. According to claim 1, a knowledge graph retrieval intelligent power storage method based on feedforward neural network is characterized in that: The classification model of power equipment resources in step S1 includes: primary equipment and secondary equipment, wherein the primary equipment includes generators, transformers, switchgear, power lines, mutual inductors, and lightning arresters; the secondary equipment includes measuring meters, relay protection and automatic devices, operating electrical appliances, and DC power supply equipment.

3. According to claim 1, a knowledge graph retrieval intelligent power storage method based on feedforward neural network is characterized in that: The meteorological disaster risk periodic assessment model in step S1 includes: Where: WREM(t) is the evaluation parameter, H i is the severity of historical disasters, H max is the maximum severity of historical disasters, F i is the fault frequency, F max is the maximum fault frequency, λ is the time attenuation coefficient, t i is a historical time point, t is the current time node, and T is the total monthly data.

4. According to claim 1, a knowledge graph retrieval intelligent power storage method based on feedforward neural network is characterized in that: The training set in step S2 will include static data and dynamic data, wherein the static data is the usage quantity and connection relationship of the power equipment on the lines within the coverage of the warehouse, and the dynamic data is the operating years of the power equipment, the degree of performance degradation, and the type of business carried.

5. According to claim 1, a knowledge graph retrieval intelligent power storage method based on feedforward neural network is characterized in that: The power equipment fault assessment algorithm of the feedforward neural network model in step S3 specifically includes: A1: The dynamic and static indicators of the power equipment are numerically processed and used as the sample data training set; A2: The replacement demand level of the equipment to be repaired is obtained as the verification result through historical verification data and manual analysis of A1 sample data; A3: Take the sample data training set as input, set the initial weights of the feedforward neural network model, perform learning and training in the feedforward neural network model, obtain the prediction results, compare the prediction results with the verification results, perform error backpropagation, update the neural network weight parameters, repeat the training process until the model accuracy meets the expected requirements, and obtain the feedforward neural network model for the knowledge graph data source.

6. The method of intelligent power storage based on knowledge graph retrieval based on feedforward neural network according to claim 5 is characterized in that: The input layer P and output layer Q of the sample data training set in step A1 in the neural network are: P=[P11,P12…P i j,Z,C,] (2) Q=[Q1,Q2,…Q i ](3) Among them, P ij is the dynamic data of the ith power equipment after numerical quantization, Z is the forecast month information, C is the climate information of the forecast time period after quantification, Q i The feedforward neural network model predicts the replacement demand level of the i-th power equipment.

7. The method of intelligent power storage based on knowledge graph retrieval of feedforward neural network according to claim 5 is characterized in that: In step A2, the replacement demand level of the equipment to be repaired is obtained by historical verification data and manual analysis of the A1 sample data, which is specifically: Q c =[Q c1 ,Q c2 ,…Q ci ] (4) Among them, Q ci It is the verification result after manual analysis and judgment.

8. The method of intelligent power storage based on knowledge graph retrieval based on feedforward neural network according to claim 5 is characterized in that: In step A3, the AdamW algorithm is used to optimize the weight parameters of the feedforward neural network, including: A3-1: The replacement demand levels of power equipment are divided into "urgent", "suspected" and "normal", and numerically mapped; A3-2: Define the loss function according to the absolute value error to calculate the error between the prediction evaluation result and the verification result: Where N is the total number of samples of power equipment, and L(θ) is the loss function of the neural network model; A3-3: Use a feedforward neural network for probability prediction evaluation. The prediction probability calculation formula is: σ(x)=max(0,x) (7) Among them, σ is the activation function; W i is the weight matrix; b i is the bias term; L is the number of neural network layers; A3_4: Use the AdamW algorithm to minimize the loss function L(θ) and update the weight parameters of the neural network; the optimization process of AdamW is as follows: Momentum Update: Among them, m t is the momentum term of the tth iteration, v t is the accumulation of the squared gradient of the tth iteration, m t-1 is the momentum term of the t-1th iteration, v t-1 is the accumulation of the squared gradient of the t-1th iteration, β1 and β2 are momentum coefficients; Bias correction: in, and They are the momentum term and the squared gradient accumulation after bias correction respectively; Weight update: Among them, θ t is the weight parameter of the tth iteration, θ t+1 is the updated value of the weight parameter for the t+1th iteration, η is the learning rate, and λ θ is the weight decay coefficient, ∈ is the smoothing term.

9. The method of intelligent power storage based on knowledge graph retrieval of feedforward neural network according to claim 1 is characterized in that: The knowledge graph storage system in step S4 includes: B1: Build a link portrait of power equipment, receive the results of the feedforward neural network model prediction and evaluation model, and organize and build a fault description knowledge graph; B2: Generate knowledge graphs for power equipment warehouse space allocation module, order preprocessing module and backlog warning module based on the prediction and evaluation model; B3: Provides a breadth-first search strategy based on graphics processor unit programming to achieve intelligent question answering and fast retrieval.

10. The method of intelligent power storage based on knowledge graph retrieval of feedforward neural network according to claim 9, characterized in that: The power equipment warehouse space allocation module in step B2 includes: Analyze the properties of power equipment to form a basic property profile of each device; Allocate warehouse storage space for incoming goods based on incoming order forecasts and basic attribute portraits; Allocate storage space in the warehouse based on the status of the stored power equipment, the remaining storage space, and the power equipment order forecast.

11. The method of intelligent power storage based on knowledge graph retrieval of feedforward neural network according to claim 9, characterized in that: In step B2, the backlog warning module uses the backlog warning index OWI to evaluate the inventory backlog risk. The specific calculation formula is: Where: S i is the current inventory, S i * is the predicted optimal inventory, B i is the unit price and M is the number of inventory types.

12. The method of intelligent power storage based on knowledge graph retrieval of feedforward neural network according to claim 11, characterized in that: The backlog warning module also includes a warning level classification mechanism: When OWI ≥ 30%, the system issues a red warning, indicating a serious backlog risk; When 20% ≤ OWI < 30%, the system issues a yellow warning, indicating that there is a medium backlog risk; When 10%≤OWI<20%, the system issues a blue warning, indicating a slight backlog risk; When OWI < 10%, the system shows that the inventory level is normal.

13. The method of intelligent power storage based on knowledge graph retrieval of feedforward neural network according to claim 9 is characterized in that: The breadth-first search strategy based on graphics processor unit programming in step B3 includes: first, receiving the knowledge graph database D i As the data input source, it also determines the starting vertex V of the search. s and the target vertex V i ; The search process starts from the starting vertex V s Initially, in each iteration, all adjacent vertices V of the currently visited vertex are searched in parallel i ; For each adjacent vertex V found i , the system will dynamically allocate an independent GPU computing unit and the corresponding distributed memory unit to achieve parallel processing of search tasks; During the search process, if an adjacent vertex V i With the target vertex V t If there is a match, the currently found path is immediately returned as the result; otherwise, the search loop continues to search for the next layer of adjacent nodes until the termination condition is met.

Images