Digital twin supply chain logistics demand forecasting method based on hybrid model

Abstract

A method for forecasting logistics demand for a digital twin supply chain based on a hybrid model includes the following steps: constructing an index system of influencing factors with the local industrial structure as the core. Constructing the YWP‑Elman model and the FWOA‑Elman model and making predictions. A hybrid prediction model with adaptive weights is established for prediction, in which the adaptive weights take into account the contribution of the prediction model and the error of the prediction results. Based on big data analysis and with the local industrial structure as the core, the present invention studies and constructs a prediction model around the correlation between material demand and supply, considers a variety of value data, conducts data training, takes into account economies of scale, carries out accurate prediction work, and forms a complete logistics demand forecasting system. This prediction system can be flexibly adjusted according to the contribution of the prediction model and the error of the prediction results, which can effectively improve the prediction accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of forecasting models, and specifically relates to a digital twin supply chain logistics demand forecasting method based on a hybrid model. Background Art

[0002] As a key technology for realizing information-physical interaction and integration, digital twin mainly studies how to obtain a virtual image of all elements of the system to meet the monitoring and simulation requirements of customized logistics systems for unknown needs and environments. With the in-depth development of IoT technology, digital twin technology can provide all operating status information in IoT devices and provide end-to-end process visualization tools for supply chain networks, improve production efficiency, and monitor and warn equipment operating status. Summary of the invention

[0003] The technical problem to be solved by the present invention is to overcome the shortcomings of the existing technology and provide a digital twin supply chain logistics demand forecasting method based on a hybrid model. It is based on big data analysis, takes the local industrial structure as the core, and focuses on the correlation between material demand and supply. It studies and constructs a forecasting model, considers a variety of valuable data, conducts data training, takes into account economies of scale, carries out accurate forecasting work, and forms a complete logistics demand forecasting system.

[0004] The present invention provides a method for predicting logistics demand of digital twin supply chain based on hybrid model, comprising the following steps:

[0005] Step S1. Establish an index system of influencing factors with the local industrial structure as the core, including economic industrial structure, economic development level, geographical location, logistics factors, consumer market factors, alternative factors, and force majeure factors;

[0006] Step S2. Construct the YWP-Elman model and the FWOA-Elman model and make predictions, and record the prediction data and errors;

[0007] Step S3. Based on the contribution of the prediction model and the error of the prediction result, a hybrid prediction model with adaptive weights is established for prediction.

[0008] As a further technical solution of the present invention, in step S1, the secondary indicator of economic development level is the GDP level of the region and the total retail sales of consumer goods; the secondary indicator of geographical location is the regional storage capacity and warehouse distance; the secondary indicator of logistics factors is the transportation network, freight volume and cargo turnover; the secondary indicator of consumer market factors is the permanent population, total product output, social network trends and residents' consumption level; the secondary indicator of substitutable factors is the out-of-stock rate and the output of substitutable products; the secondary indicator of force majeure factors is weather and natural disasters; the secondary indicator of economic industrial structure is the output value of the primary industry, the output value of the secondary industry and the output value of the tertiary industry.

[0009] Furthermore, the secondary indicators of the primary indicator force majeure are abstractly defined, so as to abstract the force majeure factors under the influence of different weather conditions;

[0010] The impact of weather and natural disasters on logistics and transportation is defined as

[0011]

[0012] The definition of weather impact on road transport is

[0013]

[0014] The definition of weather impact on railway transportation is

[0015]

[0016] The definition of weather impact on air transport is

[0017]

[0018] The definition of weather impact on water transport is

[0019]

[0020] The impact of extreme weather and natural disasters on transportation is defined as

[0021]

[0022] In summary, force majeure factors under different weather conditions are defined as:

[0023]

[0024] Furthermore, we use historical data analysis to determine product attribute weights and establish a product attribute value ranking based on historical data prediction to estimate product out-of-stock replacement rates. Specifically,

[0025] Conduct index evaluation on product attribute values;

[0026] Normalize the data to obtain the normalized matrix

[0027] Calculate the attribute weight and get the weight of the jth item as

[0028]

[0029] Furthermore, in step S1, the social network trend influence is defined as:

[0030]

[0031] Among them, q is the social network trend impact factor of the product, R is the boundary length of the market, and R 3 is the boundary of the market and K is the diffusion constant.

[0032] Furthermore, in step S3, the weights are set by adaptive combination, and the weights are adaptively changed in combination with the contribution of the prediction model and the error of the prediction result. Specifically,

[0033] According to the marginal benefit created by a single model for the overall model, the mean value is used as the basis for division. The benefit that model i gets from the overall benefit is

[0034]

[0035] Among them, |S| is the number of single models, is the number of model i excluded from the combination, and n is the total number of prediction models used.

[0036] The mean absolute error of the i-th prediction model during the prediction process:

[0037]

[0038] Among them, E i is the average error of the i-th model in prediction, m is the number of samples, e ij is the prediction error of the jth data in the prediction process of the i-th prediction model;

[0039] Therefore, the mean error when the combined model makes predictions is

[0040]

[0041] According to the calculation results of the above steps, the weights in the combined prediction model can be obtained:

[0042]

[0043] Furthermore, in step S3, a hybrid prediction model with adaptive weights is established, specifically,

[0044] The combined model is

[0045]

[0046]

[0047] is the combined prediction result, α1 is the proportion of the YWP-Elman neural network model, α2 is the proportion of the FWOA-Elman neural network prediction model, is the predicted value of the YWP-Elman neural network model; is the predicted value of FWOA-Elman neural network model.

[0048] Furthermore, new parameters M and N are introduced into the YWP-Elman model for correction:

[0049]

[0050]

[0051] Therefore, the predicted value is

[0052]

[0053] Further, in step S2, a YWP-Elman neural network model is constructed, including the following steps:

[0054] (1) Let X 0 (k) As the original data series, perform quadratic root processing on each data and obtain the sequence

[0055] (2) Construct the YWP model to complete the logarithmic series Prediction analysis to obtain the prediction sequence

[0056] (3) Change the sequence Perform a quadratic operation, i.e. the predicted sequence of the actual sample;

[0057] (4) Construct an Elman neural network, set relevant parameters, use the original data as training samples, and perform model training;

[0058] (5) The values ​​of various factors predicted by the YWP model are used as neural network input values ​​and input into the trained Elman neural network to obtain the final prediction value.

[0059] Furthermore, in step S2, a convergence factor is introduced into the FWOA-Elman neural network model. Its equation is:

[0060]

[0061] Coefficient vector Used to represent the ability of local search and global search, according to the original convergence factor

[0062] It can be seen that it decreases monotonically at (0,1), so the convergence factor should maintain a decreasing trend as the number of iterations increases;

[0063] The FWOA-Elman neural network model is constructed as follows:

[0064] (1) Input influencing factor data and target output data, divide the ELMAN neural network into training set and test set, and normalize the data;

[0065] (2) Construct the ELMAN neural network and initialize the network structure;

[0066] (3) Initialization of algorithm parameters: The independent variable of the objective function is taken as the position information X of the individual whale. In the solution space, the population position is randomly initialized. At the same time, the parameters are initialized, including the population number N, the logarithmic spiral shape constant b, the random number l, the number of iterations t, and the maximum number of iterations T. max ;

[0067] (4) Calculate the number of variable elements that need to be optimized based on the step network structure;

[0068] (5) Set the FWOA fitness function to the mean square error of the ELMAN prediction; execute the FWOA loop process, i.e., velocity update and position update, until the maximum number of iterations is reached, and terminate FWOA;

[0069] (6) Assign the optimal weight and threshold parameters obtained by FWOA to the neural network;

[0070] (7) The ELMAN neural network after FWOA will be used for training and prediction, and the errors will be recorded and analyzed.

[0071] The advantages of the present invention are that it is based on big data analysis, takes the local industrial structure as the core, focuses on the correlation between material demand and supply, studies and constructs a forecasting model, considers multiple value data, conducts data training, takes into account scale effects, carries out accurate forecasting work, and forms a complete logistics demand forecasting system. This forecasting system can be flexibly adjusted according to the contribution of the forecasting model and the error of the forecasting results, which can effectively improve the forecasting accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 It is the Elman network diagram of the present invention;

[0073] Figure 2 It is a flow chart of the YWP-Elman neural network model of the present invention;

[0074] Figure 3 It is a flow chart of the FWOA-Elman neural network model of the present invention. DETAILED DESCRIPTION

[0075] See also Figure 1This embodiment provides a digital twin supply chain logistics demand forecasting method based on a hybrid model, comprising the following steps:

[0076] Step S1. Establish an index system of influencing factors with the local industrial structure as the core, including economic industrial structure, economic development level, geographical location, logistics factors, consumer market factors, alternative factors, and force majeure factors;

[0077] Step S2. Construct the YWP-Elman model and the FWOA-Elman model and make predictions, and record the prediction data and errors;

[0078] Step S3. Based on the contribution of the prediction model and the error of the prediction result, a hybrid prediction model with adaptive weights is established for prediction.

[0079] In step S1, the secondary indicators of economic development level are the GDP level of the region and the total retail sales of consumer goods; the secondary indicators of geographical location are regional storage volume and warehouse distance; the secondary indicators of logistics factors are transportation network, freight volume and cargo turnover; the secondary indicators of consumer market factors are permanent population, total product output, social network trends and residents' consumption level; the secondary indicators of substitutable factors are out-of-stock rate and substitutable output; the secondary indicators of force majeure factors are weather and natural disasters; the secondary indicators of economic industrial structure are the output value of the primary industry, the output value of the secondary industry and the output value of the tertiary industry. As shown in the following table

[0080]

[0081] Abstractly define the secondary indicators of the primary indicator force majeure, so as to abstract the force majeure factors under different weather conditions;

[0082] The impact of weather and natural disasters on logistics and transportation is defined as

[0083]

[0084] The definition of weather impact on road transport is

[0085]

[0086] The definition of weather impact on railway transportation is

[0087]

[0088] The definition of weather impact on air transport is

[0089]

[0090] The definition of weather impact on water transport is

[0091]

[0092] The impact of extreme weather and natural disasters on transportation is defined as

[0093]

[0094] In summary, force majeure factors under different weather conditions are defined as:

[0095]

[0096] By analyzing historical data to determine the weights of product attributes, we established a product attribute value ranking based on historical data prediction to estimate the product out-of-stock replacement rate. Specifically,

[0097] Conduct index evaluation on product attribute values;

[0098] Normalize the data to obtain the normalized matrix

[0099] Calculate the attribute weight and get the weight of the jth item as

[0100]

[0101] In step S1, the social network trend influence is defined:

[0102]

[0103] Among them, q is the social network trend impact factor of the product, R is the boundary length of the market, and R 3 is the boundary of the market and K is the diffusion constant.

[0104] In step S3, the weights are set by adaptive combination, and the weights are adaptively changed in combination with the contribution of the prediction model and the error of the prediction result. Specifically,

[0105] According to the marginal benefit created by a single model for the overall model, the mean value is used as the basis for division. The benefit that model i gets from the overall benefit is

[0106]

[0107] Among them, |S| is the number of single models, is the number of model i excluded from the combination, and n is the total number of prediction models used.

[0108] The mean absolute error of the i-th prediction model during the prediction process:

[0109]

[0110] Among them, Ei is the average error of the i-th model in prediction, m is the number of samples, e ij is the prediction error of the jth data in the prediction process of the i-th prediction model;

[0111] Therefore, the mean error when the combined model makes predictions is

[0112]

[0113] According to the calculation results of the above steps, the weights in the combined prediction model can be obtained:

[0114]

[0115] In step S3, a hybrid prediction model with adaptive weights is established, specifically,

[0116] The combined model is

[0117]

[0118]

[0119] is the combined prediction result, α1 is the proportion of the YWP-Elman neural network model, α2 is the proportion of the FWOA-Elman neural network prediction model, is the predicted value of the YWP-Elman neural network model; is the predicted value of FWOA-Elman neural network model.

[0120] New parameters M and N are introduced into the YWP-Elman model for correction:

[0121]

[0122]

[0123] Therefore, the predicted value is

[0124]

[0125] In step S2, a YWP-Elman neural network model is constructed, including the following steps:

[0126] (1) Let X 0 (k) As the original data series, perform quadratic root processing on each data and obtain the sequence

[0127] (2) Construct the YWP model to complete the logarithmic series Prediction analysis to obtain the prediction sequence

[0128] (3) Change the sequence Perform a quadratic operation, i.e. the predicted sequence of the actual sample;

[0129] (4) Construct an Elman neural network, set relevant parameters, use the original data as training samples, and perform model training;

[0130] (5) The values ​​of various factors predicted by the YWP model are used as neural network input values ​​and input into the trained Elman neural network to obtain the final prediction value.

[0131] In step S2, the convergence factor is introduced into the FWOA-Elman neural network model. Its equation is:

[0132]

[0133] Coefficient vector Used to represent the ability of local search and global search, according to the original convergence factor

[0134] It can be seen that it decreases monotonically at (0,1), so the convergence factor should maintain a decreasing trend as the number of iterations increases;

[0135] The FWOA-Elman neural network model is constructed as follows:

[0136] (1) Input influencing factor data and target output data, divide the ELMAN neural network into training set and test set, and normalize the data;

[0137] (2) Construct the ELMAN neural network and initialize the network structure;

[0138] (3) Initialization of algorithm parameters: The independent variable of the objective function is taken as the position information X of the individual whale. In the solution space, the population position is randomly initialized. At the same time, the parameters are initialized, including the population number N, the logarithmic spiral shape constant b, the random number l, the number of iterations t, and the maximum number of iterations T. max ;

[0139] (4) Calculate the number of variable elements that need to be optimized based on the step network structure;

[0140] (5) Set the FWOA fitness function to the mean square error of the ELMAN prediction; execute the FWOA loop process, i.e., velocity update and position update, until the maximum number of iterations is reached, and terminate FWOA;

[0141] (6) Assign the optimal weight and threshold parameters obtained by FWOA to the neural network;

[0142] (7) The ELMAN neural network after FWOA will be used for training and prediction, and the errors will be recorded and analyzed.

[0143] In this embodiment, firstly, the factors affecting the logistics demand of the supply chain are analyzed and collected based on the influencing factor index system with the local industrial structure as the core. In terms of the measurement index of logistics demand, the logistics volume, that is, the logistics freight volume, is generally used as a measurement index to measure the scale of logistics demand, and the freight volume is used as an analogy index for logistics volume analysis to predict and analyze logistics demand. The factors affecting the logistics demand of the supply chain are mainly the level of economic development and the economic industrial structure at a macro level. Secondly, the consumer market factors and the factors of substitutable substances also have a great impact on the logistics demand; finally, the geographical location and logistics factors.

[0144] Force majeure factors will affect logistics demand. Through analysis and research, we select regional GDP X1, total retail sales of consumer goods X2, regional storage X3, warehouse distance X4, transportation network X5, freight volume X6, cargo turnover X7, permanent population X8, total product output X9, social network trend X10, resident consumption level X11, out-of-stock rate X12, substitute output X13, force majeure X14, primary industry output X15, secondary industry output X16, tertiary industry output X17 as characteristic values ​​for analysis and prediction of logistics demand Y.

[0145] The acquired data is transformed and integrated. The data features that affect the supply chain logistics demand are numerous and the data structure is complex. Before data processing, data transformation and integration processing is required. The data is normalized. The normalized data is conducive to the establishment of subsequent algorithm models. The data normalized in the range of 0 to 1 is relatively stable and reliable, and the model has good robust performance, which can avoid many modeling problems. Then the transformed and integrated data is used to establish a database. The characteristic database that affects the supply chain logistics forecast includes the regional GDP X1, the total retail sales of social consumer goods X2, the regional storage volume X3, the warehouse distance X4, the transportation network X5, the freight volume X6, the cargo turnover X7, the permanent population X8, the total product output X9, the social network trend X10, the residents' consumption level X11, the out-of-stock rate X12, the output of the substitute product X13, the force majeure X14, the output value of the primary industry X15, the output value of the secondary industry X16, and the output value of the tertiary industry X17, which are used as characteristic values ​​for analysis to predict the logistics demand Y.

[0146] The data in the database is divided into training samples and test samples, a small part of which is used as test samples to test the effectiveness of the model, and the rest of the data samples are used as training samples for training the model.

[0147] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above specific embodiments. The above specific embodiments and the description in the specification are only for further illustrating the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of the present invention to be protected is defined by the claims and their equivalents.

Claims

1. A digital twin supply chain logistics demand forecasting method based on a hybrid model, characterized by: The following steps are included: Step S1. Establish an index system of influencing factors with the local industrial structure as the core, including economic industrial structure, economic development level, geographical location, logistics factors, consumer market factors, alternative factors, and force majeure factors; Step S2. Construct the YWP-Elman model and the FWOA-Elman model and make predictions, and record the prediction data and errors; Step S3. According to the contribution of the prediction model and the error of the prediction result, a hybrid prediction model with adaptive weights is established for prediction; In step S1, the secondary indicators of economic development level are the GDP level of the region and the total retail sales of consumer goods; the secondary indicators of geographical location are regional storage volume and warehouse distance; the secondary indicators of logistics factors are transportation network, freight volume and cargo turnover; the secondary indicators of consumer market factors are permanent population, total product output, social network trends and residents' consumption level; the secondary indicators of substitute factors are out-of-stock rate and substitute output; the secondary indicators of force majeure factors are weather and natural disasters; the secondary indicators of economic industrial structure are the output value of the primary industry, the output value of the secondary industry and the output value of the tertiary industry; In step S3, a hybrid prediction model with adaptive weights is established, specifically, The combined model is is the combined prediction result, α1 is the proportion of the YWP-Elman neural network model, α2 is the proportion of the FWOA-Elman neural network prediction model, is the predicted value of the YWP-Elman neural network model; is the predicted value of FWOA-Elman neural network model; New parameters M and N are introduced into the YWP-Elman model for correction: Therefore, the predicted value is In step S2, a YWP-Elman neural network model is constructed, comprising the following steps: (1) Let X 0 (k) As the original data series, perform quadratic root processing on each data and obtain the sequence (2) Construct the YWP model to complete the logarithmic series Prediction analysis to obtain the prediction sequence (3) Change the sequence Perform a quadratic operation, i.e. the predicted sequence of the actual sample; (4) Construct an Elman neural network, set relevant parameters, use the original data as training samples, and perform model training; (5) The values ​​of various factors predicted by the YWP model are used as neural network input values ​​and input into the trained Elman neural network to obtain the final prediction value; In step S2, a convergence factor is introduced into the FWOA-Elman neural network model. Its equation is: Coefficient vector Used to represent the ability of local search and global search, according to the original convergence factor It can be seen that it decreases monotonically at (0,1), so the convergence factor should maintain a decreasing trend as the number of iterations increases; The FWOA-Elman neural network model is constructed as follows: (1) Input influencing factor data and target output data, divide the ELMAN neural network into training set and test set, and normalize the data; (2) Construct the ELMAN neural network and initialize the network structure; (3) Initialization of algorithm parameters: The independent variable of the objective function is taken as the position information X of the individual whale. In the solution space, the population position is randomly initialized. At the same time, the parameters are initialized, including the population number N, the logarithmic spiral shape constant b, the random number l, the number of iterations t, and the maximum number of iterations T. max ; (4) Calculate the number of variable elements that need to be optimized based on the step network structure; (5) Set the FWOA fitness function to the mean square error of the ELMAN prediction; execute the FWOA loop process, i.e., velocity update and position update, until the maximum number of iterations is reached, and terminate FWOA; (6) Assign the optimal weight and threshold parameters obtained by FWOA to the neural network; (7) The ELMAN neural network after FWOA will be used for training and prediction, and the errors will be recorded and analyzed.

2. The method for predicting logistics demand of digital twin supply chain based on hybrid model according to claim 1 is characterized in that: Abstractly define the secondary indicators of the primary indicator force majeure, so as to abstract the force majeure factors under different weather conditions; The impact of weather and natural disasters on logistics and transportation is defined as The definition of weather impact on road transport is The definition of weather impact on railway transportation is The definition of weather impact on air transport is The definition of weather impact on water transport is The impact of extreme weather and natural disasters on transportation is defined as Force majeure factors under different weather conditions are defined as: 。 3. The method for predicting logistics demand of digital twin supply chain based on hybrid model according to claim 2 is characterized in that: By analyzing historical data to determine the weights of product attributes, we established a product attribute value ranking based on historical data prediction to estimate the product out-of-stock replacement rate. Specifically, Conduct index evaluation on product attribute values; Normalize the data to obtain the normalized matrix Calculate the attribute weight and get the weight of the jth item as 4. The method for predicting logistics demand of digital twin supply chain based on hybrid model according to claim 1 is characterized in that: In step S1, the social network trend influence is defined: Among them, q is the social network trend impact factor of the product, R is the boundary length of the market, and R 3 is the boundary of the market and K is the diffusion constant.

5. The method for predicting logistics demand of digital twin supply chain based on hybrid model according to claim 1 is characterized in that: In step S3, the weights are set by adaptive combination, and the weights are adaptively changed in combination with the contribution of the prediction model and the error of the prediction result. Specifically, According to the marginal benefit created by a single model to the overall model, the mean value is used as the basis for division. The benefit that model i gets from the overall benefit is Among them, |S| is the number of single models, is the exclusion model i in the combination, n is the total number of prediction models used; the mean absolute error of the i-th prediction model in the prediction process is: Among them, E i is the average error of the i-th model in prediction, m is the number of samples, e ij is the i-th prediction model The prediction error of the jth data in the prediction process; The mean error when the combined model is used for prediction is According to the calculation results of the above steps, the weights in the combined prediction model can be obtained:

Images