CN111401808A

CN111401808A - Material agreement inventory demand prediction method based on hybrid model

Info

Publication number: CN111401808A
Application number: CN202010169815.8A
Authority: CN
Inventors: 马新强; 黄羿; 刘友缘
Original assignee: Chongqing University of Arts and Sciences
Current assignee: Chongqing University of Arts and Sciences
Priority date: 2020-03-12
Filing date: 2020-03-12
Publication date: 2020-07-10

Abstract

The invention provides a material agreement inventory demand prediction method based on a hybrid model, and belongs to the technical field of material demand prediction. The material agreement inventory demand forecasting method based on the hybrid model comprises the following steps: s1: acquiring historical data corresponding to all materials; s2: classifying the historical data to obtain material subclass data with different classification characteristics; s3: establishing a mixed model library containing a plurality of prediction models; s4: and selecting a preset model from the mixed model library according to different classification characteristics, and inputting the material subclass data acquired in the step S2 into the preset model to predict the preset material requirements. According to the invention, the preset model is selected from the mixed model library, and the material subclass data is input into the preset model to realize prediction of the preset material requirements, so that the classification of the materials is more accurate, and the prediction effect is better.

Description

Material agreement inventory demand prediction method based on hybrid model

Technical Field

The invention belongs to the technical field of material demand prediction, and relates to a material agreement inventory demand prediction method based on a hybrid model.

Background

The agreement inventory purchasing mode classifies the required materials according to the power grid construction requirements in a period of time in the future, determines the unified technical requirements and service standards, and purchases various materials required by the power grid construction in the period of time in the future at one time. The agreement inventory purchasing mode can effectively reduce the repeated bidding times of similar products, reduce the purchasing cost, and effectively integrate resources so as to improve the power grid material bidding purchasing efficiency and the power grid investment efficiency. The agreement inventory purchasing is a main bidding purchasing mode for the power grid materials with uniform specifications and large use amount at present, and is a key point for researching power grid material management. The demand management of the agreement inventory mainly comprises three links of determining an agreement inventory catalog, predicting an agreement inventory purchasing plan, tendering and purchasing and the like.

The protocol inventory catalog is determined by a centralized purchasing catalog issued by a national power grid company aiming at the characteristics of protocol purchasing and an independent bidding catalog actually supplemented by each provincial power grid company according to self project construction, and the national power grid company and each provincial power grid company regularly adjust and supplement the protocol inventory purchasing catalog according to construction requirements and market supply and demand environment changes; the forecast of the agreement stock purchasing demand is generally a process from the project department to the national network material department from bottom to top, firstly, an annual material demand plan is submitted to the provincial company material department by each construction project department, the provincial company material department calculates the annual actual purchasing demand according to the current entity material residual quantity and the agreement stock material residual quantity on the basis, and the calculated actual purchasing demand is compared, adjusted and analyzed with the historical demand data to form the final purchasing demand which is informed to the national network company material department or is prepared for autonomous purchasing; centralized purchasing of catalog materials in the agreement inventory is realized by unified bidding purchasing of national power grid company organizations, autonomous bidding purchasing of the materials is realized by provincial power companies, and the agreement inventory purchasing is generally realized in a centralized manner in april and october each year in a half-year period.

With the enhancement of the construction strength of the power distribution network and the improvement of the technical standard of the power grid construction, the following problems may exist in the forecast of the protocol inventory requirement:

1) due to the fact that protocol inventory prediction is conducted by a material department, due to the fact that detailed project construction basic information is not available, historical protocol purchasing data can only be referred to, and appropriate adjustment is conducted on the basis, so that the protocol inventory material demand prediction is inaccurate;

2) the project material demand is reported by the project department and the demand is summarized and readjusted by the material department to predict the demand, because the ambiguity of the project scale and progress can easily cause the standard and quantity of the tender purchase material to have certain deviation with the actual construction demand;

3) the uncertainty of the use period of the project materials, the time period of material purchase and the uncertainty of the order fulfillment condition ensure that the agreement inventory material requirement is met with certain risk, and the project construction progress is influenced or the material backlog is caused.

4) The variety of project materials is various, and the service life/demand of different materials is different, so the correct classification of the materials is often the basis of accurate prediction, but the traditional classification according to the function may cause adverse effect on the prediction.

Chinese patent CN201910574743, a distribution network material demand prediction method based on region and project clustering and linear regression, which is characterized in that the distribution network material is classified, the project is preprocessed and then clustered, the region clustering is performed, the money amount of any type of project in any type of region in the next year is predicted, the money amount involved by each standard packet in each type of project and the money amount of each standard packet in all project types and region types are calculated, and the calculated money amounts are distributed to the predicted 12 months according to the proportion to serve as the distribution network material demand in the next year. The method realizes the high-characteristic classification of the materials, analyzes the material demand rule, establishes the prediction model, predicts the amount of money of each type of items in each type of area to realize the accurate prediction of the material demand, improves the material demand prediction capability, performs auxiliary support on the material purchasing related work of a material department, ensures the normal production of power enterprises, greatly saves the purchasing cost and the inventory cost, and improves the enterprise competitiveness. In the above patent, materials are classified according to projects and regions, the classification is inaccurate, the materials with the same functions and the requirements are not necessarily the same, the historical data of part of the materials are sparse, meanwhile, a linear regression analysis method is used, the fitting may be poor, the prediction effect is poor, and therefore different inventory demand prediction algorithms for different materials need to be considered again.

Disclosure of Invention

Aiming at the problems in the prior art, the invention provides a material agreement inventory demand prediction method based on a mixed model, and the technical problems to be solved by the invention are as follows: how to provide a material agreement inventory demand prediction method based on a mixed model, which classifies historical data corresponding to materials and predicts different materials by using a plurality of prediction models.

The purpose of the invention can be realized by the following technical scheme:

a material agreement inventory demand forecasting method based on a hybrid model comprises the following steps:

s1: acquiring historical data corresponding to all materials;

s2: classifying the historical data to obtain material subclass data with different classification characteristics;

s3: establishing a mixed model library containing a plurality of prediction models;

s4: and selecting a preset model from the mixed model library according to different classification characteristics, and inputting the material subclass data acquired in the step S2 into the preset model to predict the preset material requirements.

Preferably, step S1 includes: acquiring historical project information and historical material availability data of the power distribution network from a power grid planning plan management system and an ERP system; constructing a historical sample library containing distribution network project attributes of project types, project investment sums and voltage levels according to historical project information of the power grid and historical material utilization data; and acquiring historical data corresponding to all materials from a historical sample library.

Preferably, step S1 further includes preprocessing the history data.

Preferably, the preprocessing of the history data in step S1 specifically includes: selecting data; integrating data; clearing data; data specification; and (5) data transformation.

Preferably, step S2 specifically includes:

s21: removing attributes irrelevant to the predicted value in the historical data by evaluating the importance of the attributes of the historical data to obtain historical subclass data;

s22: and clustering the historical subclass data to aggregate the historical subclass data with similar types into one type of characteristics so as to obtain the material subclass data with a plurality of different classification characteristics.

Preferably, the models in the mixed model library include a classical prediction algorithm model and a machine learning algorithm model, the classical prediction algorithm model includes a time series prediction model and a regression analysis prediction model, and the machine learning algorithm model includes a decision tree algorithm model, a bayesian algorithm model, a support vector machine algorithm model, an artificial neural network algorithm model, an integrated learning algorithm model, an association rule algorithm model, an EM algorithm model and a deep learning algorithm model.

Preferably, the artificial neural network algorithm model comprises a BP neural network prediction model, the BP neural network prediction model is selected as a preset model in the step S3, the hyper-parameters of the BP neural network are optimized through a particle swarm algorithm, and the material subclass data acquired in the step S2 is input into the BP neural network prediction model in the step S4 to realize prediction of preset material requirements.

Preferably, in step S3, the material subclass data with different classification features obtained in step S2 is divided into a training set for inputting a preset model to train the preset model and a test set for inputting a preset model to predict the preset material requirement.

Preferably, step S4 is followed by performing model evaluation on each model in the mixed model library, wherein the model evaluation obtains the inventory occupancy rate of each model by inputting the demand data into the mixed model library to test each model.

Preferably, the hyper-parameters include the number of network layers, the learning rate, the initial weight and the threshold value of the BP neural network prediction model.

According to the invention, historical data corresponding to all materials is firstly obtained, then the historical data is classified to obtain material subclass data with different classification characteristics, then a mixed model library comprising multiple prediction models is established, finally, according to the difference of the classification characteristics, different material subclasses are obtained, a preset model is selected from the mixed model library, the material subclass data obtained in the step S2 is input into the preset model to realize prediction of the preset material requirements, the materials are classified according to the material characteristics found in the data, the materials are classified according to the classification characteristics of the different material subclasses, the classification is more accurate, models suitable for the different material subclasses are selected to predict the material requirements with different classification characteristics, and the prediction effect is better.

Drawings

FIG. 1 is a schematic flow diagram of the present invention;

FIG. 2 is a flow chart of a particle swarm optimized BP neural network prediction model in the present invention;

FIG. 3 is a diagram of a BP neural network architecture in accordance with the present invention;

FIG. 4 is a diagram showing the result of clustering 6 material small groups into 3 classes by the clustering algorithm in the present invention;

FIG. 5 is a schematic diagram of the basic prediction flow of machine learning in the present invention;

FIG. 6 is a schematic diagram of a machine learning algorithm in the present invention;

FIG. 7 is a flow chart of the prediction of the inventory requirements of the distribution network material agreement in the present invention;

FIG. 8 is a schematic diagram of a sample library of protocol inventory demand forecast histories in accordance with the present invention;

FIG. 9 is a flow diagram of protocol inventory requirements management in the present invention.

Detailed Description

The following are specific embodiments of the present invention and are further described with reference to the drawings, but the present invention is not limited to these embodiments.

Referring to fig. 1-9, the method for forecasting inventory requirements of material agreement based on a hybrid model in the present embodiment includes the following steps:

s1: acquiring historical data corresponding to all materials;

The method comprises the steps of firstly obtaining historical data corresponding to all materials, classifying the historical data to obtain material subclass data with different classification characteristics, then establishing a mixed model library containing multiple prediction models, and finally selecting preset models from the mixed model library and inputting the material subclass data obtained in the step S2 into the preset models to predict preset material requirements according to different classification characteristics of different material subclasses, classifying the materials according to the material characteristics found in the data, classifying the materials accurately, selecting models suitable for the different material subclasses to predict the material requirements with different classification characteristics, and achieving a good prediction effect.

Step S1 includes: acquiring historical project information and historical material availability data of the power distribution network from a power grid planning plan management system and an ERP system; constructing a historical sample library containing distribution network project attributes of project types, project investment sums and voltage levels according to historical project information of the power grid and historical material utilization data; and acquiring historical data corresponding to all materials from a historical sample library. The materials in the invention include, but are not limited to, power grid materials.

Referring to fig. 7-9, the agreement inventory demand forecast key factor:

(1) critical factors. Factors which have great influence on annual material demand, such as annual investment amount, annual engineering construction amount and the like, and play a key role in prediction work;

(2) uncertainty factors. Relevant policies, work requirement changes (budget management, settlement requirements, etc.), technical requirement changes, design changes, etc., are major factors that cause demand fluctuations, but with great uncertainty. The forecast of the material demand of the protocol inventory is urgently needed to be changed from the past plan management mode of mainly passive demand management and secondarily artificial experience forecast into an active plan management mode of scientifically forecasting the material demand and differentially adjusting the plan based on the change data of the material demand. The method is characterized in that scientific material demand prediction is carried out by adopting a machine learning algorithm based on relevant data of project construction and relevant data of material demand change of a power grid ERP and ECP system and assisted by other peripheral data, and is a key for solving the problem of accuracy of protocol inventory demand prediction.

Analyzing data of each link of reporting and bidding purchasing of the agreement inventory requirement in the ECP system to obtain a plurality of key data nodes in the management of the agreement inventory material requirement: the agreement stock tenders the purchase amount, the agreement stock matching order amount, the agreement stock order fulfillment arrival amount and the material receiving amount, and the like. The agreement inventory material tendering purchase amount refers to the one-time tendering purchase amount of various materials according to the demand purchase suggestion in a purchase period; the agreement stock matching order amount is a result of matching bid amount in a purchasing period; the order fulfillment arrival amount refers to the agreement inventory fulfillment arrival amount in a purchasing period; the material receiving amount refers to the amount of materials of each item receiving agreement inventory and historical inventory in a purchasing period according to requirements. Under the integrated and intensive trend of material demand and purchase, data of links such as demand submission, bid order, purchase order and reception of the agreed inventory materials are recorded in ERP and ECP systems, and specific numerical values of the indexes can be obtained through correlation and comparison of related data. Obtaining historical project information and historical material availability data of the power distribution network from a power grid planning plan management system and an ERP system, constructing a historical sample library containing distribution network project attributes such as project types, project investment amounts, voltage levels and the like, establishing a machine learning model on the basis, and predicting the demand of materials.

Step S1 also includes preprocessing the history data.

The preprocessing of the history data in step S1 specifically includes: selecting data; integrating data; clearing data; data specification; and (5) data transformation. The data selection can be to select the power grid material data, select data of which attributes or data of which materials in which period; the data integration can be power grid material data integration, and is to integrate data of all sources and arrange the data according to material-time period-attribute; the data cleaning can be used for cleaning materials of the power grid and removing dirty data; the data protocol can be a power grid material data protocol, is a sample protocol, and is convenient for subsequent operation when sampling some mass data; the data transformation can be power grid material data transformation, and is to normalize, remove abnormal values, fill missing values and the like for data of some dimensions.

Step S2 specifically includes:

s22: and clustering the historical subclass data to aggregate the historical subclass data with similar types into one type of characteristics so as to obtain the material subclass data with a plurality of different classification characteristics. Because some attributes of data selection may be redundant and attributes irrelevant to predicted values, some unimportant attributes can be removed through some correlation analysis and variance analysis, and dimensionality reduction can improve the efficiency of model accuracy.

Proper classification of materials is often the basis for accurate predictions, but traditional classification by function may adversely affect predictions. Because of the same functional supplies, the requirements are not necessarily the same. For example, by utilizing the homology and inheritance of spare parts, a multiple linear regression method combined with a pre-classification technology is provided for predicting the demand of the spare parts. Firstly, a principal component analysis pedigree clustering method is used for classifying spare parts which have the same working environment and basically consistent requirements into one class. And performing regression analysis prediction by taking the classification result as the basis of the multiple linear regression prediction model, and then calculating the economic ordering batch of the spare parts according to the prediction result. For materials with obvious seasonal change characteristics, the periodicity can be captured by using a time series prediction model. According to Poisson distribution, the concept of cluster analysis is utilized to classify the prediction contents, and spare parts of different types have different distribution curves (considering factors such as service life, whether maintenance can be carried out and the like) so as to construct inventory demand prediction models of different devices. Some material subclasses have sparse historical data, insufficient data are used as training data, the data demand with the least gray theoretical prediction is four samples, and the material subclasses are suitable for being predicted by the gray theory. Therefore, necessary features are extracted for clustering according to historical information of different materials, similar materials are clustered together in small categories, efficiency can be improved, and prediction errors can be reduced.

The power grid materials can be classified into ABC three types according to the control degree:

A) the material is also called batch material, generally comprises five hundred fifty kilovolts of high-voltage cable line cabinet material and more than five hundred fifty kilovolts of high-voltage cable line cabinet material, is used for main line transmission and high-voltage transformer substations, and is high in price, high in quality requirement and relatively large in purchase quantity, so that the material is intensively purchased and controlled by national power grid companies;

B) the materials generally comprise high-voltage cable cabinet materials of one hundred and ten kilovolts and more than five hundred and fifty kilovolts, office supplies with higher unit price and the like, and management and control catalogues of the materials are set by various network provinces companies;

C) the material generally comprises office supplies and other materials required by daily office work, has the characteristics of large purchase quantity, low price and short updating period, and belongs to the material managed by non-national grid companies and the material managed by each provincial and rural company.

According to the classification of demand characteristics, the power grid materials can be divided into capital construction materials, operation and maintenance materials, first-aid repair materials and office supplies. According to the purchasing classification standard, the power grid materials are classified into primary equipment (main transformers, alternating current transformers, voltage transformers, direct current transformers, and the like), secondary equipment (relay protection and automation devices, measurement and control and online monitoring systems, and the like), device materials (pole towers, ground wires, hardware fittings, optical cables, insulators, and the like), communication equipment (power line carrier equipment, access equipment, optical fiber communication equipment, and the like), and other materials (office supplies, and the like).

Referring to fig. 4, the characteristics of the materials (such as the periodicity of the materials, the device lifetime, the size of the historical data, and the like) found in the data based on the three material classification methods can be combined together for cluster analysis, so that all the material subclasses can be divided into material subclass groups with different characteristics.

The model in the mixed model library comprises a classic prediction algorithm model and a machine learning algorithm model, the classic prediction algorithm model comprises a time sequence prediction model and a regression analysis prediction model, and the machine learning algorithm model comprises a decision tree algorithm model, a Bayesian algorithm model, a support vector machine algorithm model, an artificial neural network algorithm model, an integrated learning algorithm model, an association rule algorithm model, an EM algorithm model and a deep learning algorithm model. The models in the mixed model library in the present invention may include, but are not limited to, classical prediction algorithm models and machine learning algorithm models, the classical prediction algorithm models may also include, but are not limited to, the 2 models mentioned above, and the machine learning algorithm models may also include, but are not limited to, the above lists.

Methods in classical prediction are mainly classified into two types: time series prediction method and regression analysis prediction method. The classical prediction algorithm has the advantages that the classical prediction method is based on the asymptotic theory, and the theoretical basis is mature. Meanwhile, the research searches for a data statistical rule when the samples tend to be infinite, and the data statistical processing process is not complex and is convenient to understand and apply. The disadvantages are as follows: the classical prediction method is to look for an explicit rule presented by a sample according to historical data and construct a prediction model, and the data in the future time is considered to be generated according to an established forecast model rule. However, in reality, many data are generated not according to the expected model rule. Meanwhile, in the case of enough sample data, it is very difficult to find out the rule of the sample data. Therefore, when the sample data amount is limited and the sample data rule is not obvious, the classic prediction algorithm is used for predicting the future situation to have unreliability.

1) The time series prediction method is to arrange the data of the prediction variables in time sequence, such as: y1, y2, y3, … and yt-1, a mathematical model of a symbol is selected according to the numerical value change type of the mathematical model to describe the change situation of the numerical value, and then a new demand is predicted by a method for extending the new demand to the future according to the past demand change rule on the basis of the mathematical model. The time series prediction method mainly comprises the following steps: exponential smoothing, moving average, etc.

2) Regression analysis prediction rules are a common method used to determine the interdependence between two or more variables when they are encountered. The method is characterized in that a regression equation is established by finding out the correlation between a dependent variable (a prediction object) and an independent variable (various factors influencing the prediction object), and then the numerical value of the independent variable is introduced to calculate the numerical value of the dependent variable. Regression analysis prediction methods include univariate regression analysis, multivariate regression analysis, linear regression analysis, nonlinear regression analysis, and the like.

Referring to fig. 5-6, the machine learning algorithm is classified into classification, regression, clustering and dimensionality reduction according to task types, and has the advantage that the machine learning method is similar to data mining and finds rules contained in existing data, so as to realize construction of a prediction model. Compared with the traditional classical prediction algorithm, the machine learning method can be used for discovering the data rule according to the existing data and constructing the prediction model on the basis of not needing to know the structure of the function or rule in advance. The method has the disadvantages that the machine learning method constructs a prediction model according to the 'approximation' idea, the data statistics processing process is relatively complex and difficult to understand, and a model with a good effect usually needs a large amount of independent and equally distributed historical data and has high requirements on data.

Referring to fig. 2-3, the artificial neural network algorithm model includes a BP neural network prediction model, the BP neural network prediction model is selected as a preset model in step S3, the hyper-parameters of the BP neural network are optimized through a particle swarm algorithm, and the material subclass data acquired in step S2 is input into the BP neural network prediction model in step S4 to predict the preset material demand.

After the materials are classified, the BP + particle swarm optimization algorithm prediction model can predict the material requirements of one material class, the data volume is rich, the material attributes are multiple, and the BP + particle swarm optimization algorithm prediction model can be adopted for constructing high-dimensional material historical data. Firstly, determining the structure of a BP neural network, such as an input layer, an output layer and a hidden layer of the BP neural network, then obtaining a particle fitness function by the BP neural network, wherein the fitness function can be the predicted mean square error of the BP neural network, the fitness of a certain particle = the predicted mean square error after the BP neural network is trained by using the parameter of the particle, and the particle fitness calculation is the process of training the BP neural network by using the parameter of the particle and then calculating the predicted error; secondly, encoding the particles of the particle swarm, and initializing the particle swarm so as to calculate the particle fitness; if the particle fitness meets the condition that the error is within a preset value or reaches a preset iteration number, jumping out of a particle swarm algorithm cycle, and if the particle fitness does not meet the condition that the error is within the preset value or reaches the preset iteration number, sequentially calculating the individual optimal value of the particles, calculating the global optimal value of the particle swarm, evolving the particle speed and evolving the particle position through the particle swarm algorithm; and if the particle fitness meets the condition that the error is within a preset value or reaches a preset iteration number, assigning values to an initial value and a threshold value of the BP neural network, and when the error between the actual output and the expected output of an output layer of the BP neural network is calculated to be larger, continuously adjusting the weight and the threshold value of the BP neural network until the error is a preset value, and ending the prediction.

The BP neural network has the following two advantages: 1) and the nonlinear mapping relation is relatively discrete in input samples, and the mapping relation can be set by using the weight and the threshold between the neurons only by providing enough training samples, so that the nonlinear mapping relation has stronger nonlinear mapping capability. (2) Fault tolerance capability. In a large number of input samples, even errors in the individual input data do not affect the entire training process. The BP neural network can transmit and adjust the weight through the reverse error, and the correct rule of the training sample is counted.

The BP neural network has many specific advantages in solving the machine learning problems of small samples, nonlinearity and high dimension, but the small sample data learning simply relying on the BP neural network technology still hardly obtains stable and good learning effect, the hyperparameter is optimized through the particle swarm optimization on the basis of the BP neural network, and a BP neural network protocol inventory demand prediction model based on the particle swarm optimization is established. The BP neural network is a neural network which propagates reversely and is composed of more than three layers of neuron cells, wherein the upper layer of neuron cells and the lower layer of neuron cells are mutually connected, and the same layer of neuron cells are mutually independent. The BP neural network signal processing flow comprises the following steps: 1) the input layer acquires input information from the outside and transmits the information to the output layer through the middle layer to obtain an output result of the output layer; 2) comparing the direction of the expected output layer result with the direction of the network output result, reversely adjusting the connection weight between the neurons layer by layer according to the result, and transmitting the result to the input layer; 3) and repeatedly correcting and adjusting the connection weight between the neurons, and enhancing the response of the network to the correct mode until the output error is acceptable. The BP neural network model is realized mainly by a BP algorithm, which is a supervised learning algorithm and consists of an input layer-output layer learning process and an output layer-input layer error adjusting process. The specific implementation process of the BP algorithm comprises the following steps:

1) forward propagation: training sample data is input through an input layer, data processing is carried out through a hidden layer, and data output is carried out on an output layer;

2) and (3) error back propagation: when the error between the actual output of the output layer and the expected output is larger, entering a back transmission process from the output layer to the input layer, wherein the back transmission process is carried out layer by layer from the output layer to the input layer through each neuron of the hidden layer, and in the error back transmission process, each neuron cell shares the error;

3) weight adjustment: continuously adjusting the weight of the neuron cells in the forward propagation and error reverse transmission correction process, namely the learning process of the neural network;

4) threshold determination: there are generally two criteria for the stopping of the neural network learning process: one is that the error in the output of the network reaches a desired threshold, and the other is that the training process reaches a predetermined number of times.

In the particle swarm optimization, each particle f is an individual in the bird population, namely a feasible solution of the solution space , the fitness of each particle is set according to an objective function, and the passing speed of the particles determines the moving direction and distance of the particles so as to carry out continuous optimization. In the particle swarm optimization process, firstly, a particle swarm is initialized, namely a solution space, and then, the particles are optimized by continuously tracking two extreme values through iteration for one time: the particles track individual extreme values obtained in the self-optimizing process on one hand, and track global extreme values obtained by iteration of the population on the other hand.

In step S3, the material subclass data with different classification characteristics obtained in step S2 are divided into a training set for inputting a preset model for training a preset model and a test set for inputting a preset model for predicting the preset material requirement.

Step S4 is followed by model evaluation of the models in the hybrid model library, the model evaluation obtaining inventory occupancy rates of the models by inputting the demand data into the hybrid model library to test the models.

The hyper-parameters comprise the network layer number, the learning rate, the initial weight and the threshold value of the BP neural network prediction model.

CPFR, a collaborative supply chain inventory management technique, can reduce inventory while increasing sales. Its four main characteristics: and processing modes of cooperation, planning, prediction and replenishment. Problems possibly encountered in the management and control of the materials of the current power grid company through the CPFR cooperative management and control are analyzed for project characteristics, material current situation, purchasing mode and plan execution, related data including internal material circulation, plan making, production operation, financial funds and the like are communicated through the data of the current ERP and ECP systems, information sharing between provincial material purchasing and city and county material purchasing is achieved, cooperative sales prediction of purchasing parties is facilitated, the common prediction of the information can greatly reduce the low efficiency and dead stock of the whole value chain system, better product sales is promoted, and resources of the whole supply chain are saved and used.

The problems and influence factors in the distribution network material agreement inventory demand prediction are analyzed, the fact that the quantity and the variety of the material demands are large and the correlation between the investment quantity and historical data is large is found, meanwhile, the difficulty lies in that the variation of power grid distribution network items is large, the types are large, the related materials are small and various, the characteristic difference of different materials is large, a historical sample library is established according to the characteristics, the materials are classified around the historical sample library, and then models suitable for different materials are selected from the model library to predict the different material inventory demands.

The specific embodiments described herein are merely illustrative of the spirit of the invention. Various modifications or additions may be made to the described embodiments or alternatives may be employed by those skilled in the art without departing from the spirit or ambit of the invention as defined in the appended claims.

Claims

1. A material agreement inventory demand forecasting method based on a hybrid model is characterized by comprising the following steps:

s1: acquiring historical data corresponding to all materials;

2. The method as claimed in claim 1, wherein the step S1 includes: acquiring historical project information and historical material availability data of the power distribution network from a power grid planning plan management system and an ERP system; constructing a historical sample library containing distribution network project attributes of project types, project investment sums and voltage levels according to historical project information of the power grid and historical material utilization data; and acquiring historical data corresponding to all materials from a historical sample library.

3. The material agreement inventory demand forecasting method based on the hybrid model as set forth in claim 1 or 2, characterized in that: step S1 also includes preprocessing the history data.

4. The method for forecasting the inventory requirement of the material agreement based on the hybrid model as claimed in claim 3, wherein the preprocessing of the historical data in the step S1 specifically includes: selecting data; integrating data; clearing data; data specification; and (5) data transformation.

5. The method for forecasting the inventory requirement of the material agreement based on the hybrid model as claimed in claim 4, wherein the step S2 specifically includes:

6. The material agreement inventory demand forecasting method based on the hybrid model as claimed in claim 3, characterized in that: the model in the mixed model library comprises a classical prediction algorithm model and a machine learning algorithm model, the classical prediction algorithm model comprises a time sequence prediction model and a regression analysis prediction model, and the machine learning algorithm model comprises a decision tree algorithm model, a Bayesian algorithm model, a support vector machine algorithm model, an artificial neural network algorithm model, an integrated learning algorithm model, an association rule algorithm model, an EM algorithm model and a deep learning algorithm model.

7. The material agreement inventory demand forecasting method based on the hybrid model as claimed in claim 6, characterized in that: the artificial neural network algorithm model comprises a BP neural network prediction model, the BP neural network prediction model is selected as a preset model in the step S3, the hyper-parameters of the BP neural network are optimized through a particle swarm algorithm, and in the step S4, the material subclass data obtained in the step S2 are input into the BP neural network prediction model to realize prediction of preset material requirements.

8. The material agreement inventory demand forecasting method based on the hybrid model as claimed in claim 4, characterized in that: in step S3, the material subclass data with different classification characteristics obtained in step S2 are divided into a training set for inputting a preset model for training a preset model and a test set for inputting a preset model for predicting the preset material requirement.

9. The material agreement inventory demand forecasting method based on the hybrid model as claimed in claim 7, characterized in that: step S4 is followed by model evaluation of the models in the hybrid model library, the model evaluation obtaining inventory occupancy rates of the models by inputting the demand data into the hybrid model library to test the models.

10. The material agreement inventory demand forecasting method based on the hybrid model as claimed in claim 7, characterized in that: the hyper-parameters comprise the number of network layers, the learning rate, the initial weight and the threshold value of the BP neural network prediction model.