Hami melon supply chain demand prediction system based on artificial intelligence

Abstract

The invention discloses a Hami melon supply chain demand prediction system based on artificial intelligence, which belongs to the technical field of intelligent agriculture and comprises a multi-source data preparation module, a supply and demand feature modeling module, a bidirectional linkage prediction module and a supply and demand prediction report generation module. According to the method, a heterogeneous association embedding modeling method is adopted to carry out supply and demand feature modeling, a multi-dimensional supply and demand heterogeneous graph covering a planting land parcel and a sales area is constructed, spatial-temporal change and supply and demand linkage characteristics are dynamically sensed, comprehensive features with higher expressivity are generated, and prediction accuracy can be improved; the supply and demand prediction method based on bidirectional coordination is adopted to perform bidirectional linkage prediction, combine modeling of planting yield and market sales volume, dynamic capture of mutual influence between the planting yield and the market sales volume, and timely discovery and correction of supply and demand difference, so that prediction accuracy and supply chain response capability are improved, and supply and demand balance and operation efficiency of the Hami melon market are guaranteed.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent agriculture, and particularly relates to a Hami melon supply chain demand prediction system based on artificial intelligence. BACKGROUND

[0002] The Hami melon supply chain demand prediction based on artificial intelligence utilizes artificial intelligence technology, fuses multi-source data such as planting and sales, jointly models supply and demand change characteristics, realizes bidirectional prediction of planting yield and market sales, aims to assist supply chain decision-making, guarantee market supply and demand balance, improve operation efficiency and response capability of the supply chain, reduce resource waste and risks, and promote intelligent development of the Hami melon industry.

[0003] However, in the existing Hami melon supply chain demand prediction process, there are technical problems that planting information and sales information are scattered, supply and demand correlation is weak, and actual supply and demand linkage cannot be effectively reflected, thereby affecting prediction accuracy; and there are technical problems that traditional demand prediction usually relies on single modeling of planting information or sales information, ignores the supply and demand linkage relationship, and leads to unbalanced supply and demand in the prediction result, and it is difficult to adjust in time to cope with the unbalanced supply and demand. SUMMARY

[0004] In view of the above problems, in order to overcome the defects of the prior art, the application provides a Hami melon supply chain demand prediction system based on artificial intelligence, which creatively adopts a heterogeneous correlation embedded modeling method to model supply and demand characteristics, constructs a multi-dimensional supply and demand heterogeneous graph covering planting plots and sales areas, dynamically perceives spatio-temporal changes and supply and demand linkage characteristics, generates more expressive comprehensive features, and helps to improve prediction accuracy; and creatively adopts a bidirectional collaborative supply and demand prediction method to perform bidirectional linkage prediction, jointly models planting yield and market sales, dynamically captures mutual influence between the two, discovers and corrects supply and demand differences in time, thereby improving prediction accuracy and supply chain response capability, guaranteeing supply and demand balance and operation efficiency of the Hami melon market.

[0005] The technical scheme adopted by the application is as follows: The Hami melon supply chain demand prediction system based on artificial intelligence provided by the application comprises a multi-source data preparation module, a supply and demand feature modeling module, a bidirectional linkage prediction module, and a supply and demand prediction report generation module.

[0006] The multi-source data preparation module is configured to prepare multi-source data, obtain melon supply chain multi-source data through multi-source data preparation, and send the melon supply chain multi-source data to the supply-demand feature modeling module;

[0007] The supply-demand feature modeling module is configured to model supply-demand features, model supply-demand features according to melon supply chain multi-source data by using a heterogeneous correlation embedding modeling method, obtain a supply-demand fusion feature vector, and send the supply-demand fusion feature vector to the bidirectional linkage prediction module;

[0008] The bidirectional linkage prediction module is configured to predict bidirectional linkage, predict bidirectional linkage according to the supply-demand fusion feature vector by using a supply-demand prediction method based on bidirectional collaboration, obtain melon supply-demand prediction data, and send the melon supply-demand prediction data to the supply-demand prediction report generation module;

[0009] The supply-demand prediction report generation module is configured to generate a supply-demand prediction report, and obtain a melon supply chain demand prediction report.

[0010] Further, the multi-source data preparation specifically includes collecting melon historical planting information, melon historical sales records, and melon supply and marketing information, and performing data cleaning and structured processing to obtain melon supply chain multi-source data;

[0011] The melon historical planting information includes plot number, planting area, planting variety, and melon yield;

[0012] The melon historical sales records include sales region number, melon sales volume, and promotion activity intensity;

[0013] The melon supply and marketing information includes spatial distance between sales regions, spatial distance between planting plots and sales regions, transportation cost, and transportation cargo volume.

[0014] Further, the supply-demand feature modeling specifically includes modeling supply-demand features according to melon supply chain multi-source data by using a heterogeneous correlation embedding modeling method to obtain a supply-demand fusion feature vector, including the following steps: supply-demand heterogeneous graph construction, dynamic embedding of time and space weights, and supply-demand feature fusion;

[0015] The supply-demand heterogeneous graph construction is configured to construct a heterogeneous graph containing multiple relationship structures between the planting end and the sales end, and specifically includes constructing the supply-demand heterogeneous graph through node setting, edge setting, and edge weight setting;

[0016] The node setting specifically includes setting nodes of the supply-demand heterogeneous graph by constructing planting nodes and constructing sales nodes;

[0017] The construction planting node, specifically, each planting plot is taken as a planting node, and the plot number, planting area, planting variety and Hami melon yield are taken as planting node attributes; the construction sales node, specifically, each sales area is taken as a sales node, and the sales area number, Hami melon sales volume and promotion activity intensity are taken as sales node attributes;

[0018] The edge setting, specifically, the supply edge and the sales history edge are constructed to set the edges of the supply-demand heterogeneous graph;

[0019] The construction of the supply edge, specifically, when the planting plot and the sales area have a direct supply relationship, the supply edge is constructed between the corresponding planting node and the sales node, and the spatial distance, transportation cost and transportation cargo volume are taken as the supply edge attributes;

[0020] The construction of the sales history edge, specifically, the sales volume similarity between two sales areas is calculated through the Pearson correlation coefficient, and when the sales volume similarity exceeds the sales volume similarity threshold, the sales history edge is constructed between the corresponding two sales nodes, and the sales volume similarity and the spatial distance are taken as the sales history edge attributes;

[0021] The edge weight setting, specifically, the standard multi-layer perception machine is constructed to combine the node attributes and the edge attributes to calculate the edge weight;

[0022] The spatio-temporal weight dynamic embedding, specifically, the node attribute feature vector is taken as the initial node embedding vector, the adjacent nodes of each node are obtained, the attention score is calculated through the design of the Fourier spatio-temporal periodic function, the spatio-temporal perception weight is generated through the graph attention mechanism, then the node embedding vector is updated by weighted aggregation of adjacent node information;

[0023] The supply-demand feature fusion, specifically, in the supply-demand heterogeneous graph, for each supply edge, the corresponding planting node embedding vector and the sales node embedding vector are extracted, and the feature fusion is performed to obtain the supply-demand fusion feature vector;

[0024] The feature fusion, specifically, the planting node embedding vector and the sales node embedding vector are first spliced to obtain a supply-demand preliminary splicing feature vector, and the planting node embedding vector and the sales node embedding vector are multiplied element by element to obtain a supply-demand feature interaction term; then the supply-demand preliminary splicing feature vector and the supply-demand feature interaction term are spliced to obtain a supply-demand splicing feature vector, and finally the supply-demand splicing feature vector is nonlinearly transformed by a tanh activation function to generate a supply-demand fusion feature vector.

[0025] Further, the bidirectional linkage prediction, specifically according to the supply-demand fusion feature vector, adopts a supply-demand prediction method based on bidirectional collaboration to perform bidirectional linkage prediction to obtain the Hami melon supply-demand prediction data, including the following steps: bidirectional prediction backbone construction, collaborative attention interaction, supply-demand residual correction, and supply-demand prediction model construction.

[0026] The bidirectional prediction backbone construction, specifically, through constructing a time series encoder, models the time series features of the supply-demand fusion feature vector to obtain supply-demand time series features, and then constructs a parallel decoding branch to decode the supply-demand time series features respectively to generate the yield prediction sequence and the sales prediction sequence.

[0027] The collaborative attention interaction is used to model the mutual influence relationship between supply and demand, specifically by introducing a bidirectional collaborative attention mechanism between the parallel decoding branches, extracting the sales hidden state features and the yield hidden state features from the hidden layers in the parallel decoding branches respectively, calculating the sales influence factor and the yield influence factor through the bidirectional collaborative attention mechanism, and feeding back to the parallel decoding branches to adjust the sales hidden state features and the yield hidden state features in the decoding process.

[0028] The supply-demand residual correction, specifically, through the difference between the sales prediction sequence and the yield prediction sequence, obtains the supply-demand residual, when the absolute value of the supply-demand residual is greater than the residual threshold, the supply-demand residual is dynamically smoothed by a gated recurrent unit to obtain a feedback adjustment amount, and the sales prediction sequence and the yield prediction sequence are corrected according to the feedback adjustment amount.

[0029] The supply-demand prediction model construction, specifically, by executing the bidirectional prediction backbone construction, the collaborative attention interaction, the supply-demand residual correction, and model training, constructs a supply-demand prediction model, and then performs bidirectional linkage prediction through the supply-demand prediction model to obtain the Hami melon supply-demand prediction data, including the yield prediction sequence and the sales prediction sequence.

[0030] Further, the supply-demand prediction report generation, specifically, according to the Hami melon supply-demand prediction data, extracts the supply-demand change trend by a time series trend analysis method to generate a Hami melon supply chain demand prediction report; the Hami melon supply chain demand prediction report includes the Hami melon supply-demand prediction data, the Hami melon supply-demand change trend chart, and the planting adjustment suggestion.

[0031] The above-mentioned scheme has the following beneficial effects:

[0032] (1) In view of the technical problem that in the existing Hami melon supply chain demand prediction process, planting information and sales information are scattered, the correlation between supply and demand is weak, and it is difficult to effectively reflect the actual supply and demand linkage, thereby affecting the prediction accuracy, the scheme creatively uses a heterogeneous correlation embedded modeling method to model the supply and demand characteristics, constructs a multi-dimensional supply and demand heterogeneous graph covering planting plots and sales areas, dynamically perceives the spatio-temporal changes and supply and demand linkage characteristics, generates more expressive comprehensive features, and helps to improve the prediction accuracy;

[0033] (2) In view of the technical problem that in the existing Hami melon supply chain demand prediction process, traditional demand prediction usually relies on single modeling of planting information or sales information, ignoring the supply and demand linkage relationship, resulting in imbalance between supply and demand in the prediction result, and it is difficult to adjust in time to cope with the imbalance between supply and demand, the scheme creatively uses a two-way collaborative supply and demand prediction method for two-way linkage prediction, jointly models planting yield and market sales, dynamically captures the mutual influence between the two, and timely discovers and corrects the supply and demand difference, thereby improving the prediction accuracy and supply chain response ability, and guaranteeing the supply and demand balance and operation efficiency of the Hami melon market. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 The structure diagram of the Hami melon supply chain demand prediction system based on artificial intelligence provided by the present application is provided.

[0035] Figure 2 The flowchart of the supply and demand feature modeling module is provided.

[0036] Figure 3 The flowchart of the two-way linkage prediction module is provided.

[0037] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation on the present application. DETAILED DESCRIPTION

[0038] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

[0039] In the description of the present application, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0040] Embodiment one, refer to Figure 1 The present application provides a hamigua supply chain demand prediction system based on artificial intelligence, comprising: a multi-source data preparation module, a supply and demand feature modeling module, a two-way linkage prediction module and a supply and demand prediction report generation module;

[0041] The multi-source data preparation module is used for multi-source data preparation, and through multi-source data preparation, hamigua supply chain multi-source data is obtained, and the hamigua supply chain multi-source data is sent to the supply and demand feature modeling module;

[0042] The supply and demand feature modeling module is used for supply and demand feature modeling, and according to the hamigua supply chain multi-source data, a heterogeneous correlation embedding modeling method is used for supply and demand feature modeling to obtain a supply and demand fusion feature vector, and the supply and demand fusion feature vector is sent to the two-way linkage prediction module;

[0043] The two-way linkage prediction module is used for two-way linkage prediction, and according to the supply and demand fusion feature vector, a supply and demand prediction method based on two-way collaboration is used for two-way linkage prediction to obtain hamigua supply and demand prediction data, and the hamigua supply and demand prediction data is sent to the supply and demand prediction report generation module;

[0044] The supply and demand prediction report generation module is used for supply and demand prediction report generation to obtain a hamigua supply chain demand prediction report.

[0045] Embodiment two, refer to Figure 1 This embodiment is based on the above embodiment, and the multi-source data preparation specifically collects hamigua historical planting information, hamigua historical sales records and hamigua supply and marketing information, and performs data cleaning and structured processing to obtain hamigua supply chain multi-source data;

[0046] The hamigua historical planting information includes plot number, planting area, planting variety and hamigua yield;

[0047] The hamigua historical sales records include sales area number, hamigua sales volume and promotion activity intensity;

[0048] The hamigua supply and marketing information includes the spatial distance between sales regions, the spatial distance between planting plots and sales regions, transportation cost and transportation volume;

[0049] The data cleaning includes outlier removal, duplicate value filtering and missing value filling;

[0050] The structured processing includes timestamp alignment, spatial annotation and semantic labeling.

[0051] Embodiment three, refer to Figure 1 and Figure 2 , based on the above embodiment, the supply-demand feature modeling, specifically according to the heterogeneous correlation embedding modeling method is used to model the supply-demand feature, and the supply-demand fusion feature vector is obtained, including the following steps: supply-demand heterogeneous graph construction, dynamic embedding of time and space weight and supply-demand feature fusion;

[0052] The supply-demand heterogeneous graph construction is used to construct a heterogeneous graph containing multiple relationship structures between the planting end and the sales end, specifically by node setting, edge setting and edge weight setting to construct the supply-demand heterogeneous graph;

[0053] The node setting specifically sets the nodes of the supply-demand heterogeneous graph by constructing planting nodes and constructing sales nodes;

[0054] The construction of planting nodes specifically takes each planting plot as a planting node, and takes plot number, planting area, planting variety and Hami melon yield as planting node attributes; the construction of sales nodes specifically takes each sales area as a sales node, and takes sales area number, Hami melon sales volume and promotion activity intensity as sales node attributes;

[0055] The edge setting specifically sets the edges of the supply-demand heterogeneous graph by constructing supply edges and constructing sales history edges;

[0056] The construction of supply edges specifically constructs a supply edge between the corresponding planting node and the sales node when the planting plot and the sales area have a direct supply relationship, and takes spatial distance, transportation cost and transportation volume as supply edge attributes;

[0057] The construction of sales history edges specifically calculates the sales similarity between two sales areas by Pearson correlation coefficient, and constructs a sales history edge between the corresponding two sales nodes when the sales similarity exceeds the sales similarity threshold, and takes the sales similarity and the spatial distance as the sales history edge attributes;

[0058] The edge weight setting specifically calculates the edge weight by constructing a standard multi-layer perception machine combined with node attributes and edge attributes, and the calculation formula is:

[0059] ;

[0060] In the formula, is an edge weight between the ith node and the jth node under the rth edge type, the edge types include supply edges and sales history edges, i is a first index of a node of a supply-demand heterogeneous graph, j is a second index of a node of the supply-demand heterogeneous graph, i≠j, and the node of the supply-demand heterogeneous graph and the first index of the node of the supply-demand heterogeneous graph are adjacent nodes, is a standard multi-layer perception function, f i (t) is a node attribute feature vector of the ith node at the tth time step, f j (t) is a node attribute feature vector of the jth node at the tth time step, is an edge attribute feature vector between the ith node and the jth node under the rth edge type;

[0061] The spatio-temporal weight dynamic embedding, specifically, takes the node attribute feature vector as an initial node embedding vector, obtains adjacent nodes of each node, calculates an attention score through a designed Fourier spatio-temporal periodic function, generates a spatio-temporal perception weight through a graph attention mechanism, and then updates the node embedding vector through weighted aggregation of adjacent node information;

[0062] The calculation formula of the attention score is:

[0063] ;

[0064] In the formula, sc ij (t) is an attention score of the ith node to the jth node at the tth time point, LeakyReLU(·) is a ReLU activation function with leakage, a is a learnable weight vector, T is a transpose operation, concat(·) is a concatenation operation, W is a shared weight matrix, h i (t) is an embedding vector of the ith node at the tth time point, h j (t) is an embedding vector of the jth node at the tth time point, is a Fourier spatio-temporal periodic function, dis ij is a spatial distance between the ith node and the jth node;

[0065] The calculation formula of the Fourier spatio-temporal periodic function is:

[0066] ;

[0067] In the formula, sin(·) is a sine function, is a circular constant, Time is a period length, e is a base number of a natural logarithm, is a spatial decay coefficient;

[0068] The calculation formula of the spatio-temporal perception weight generated through the graph attention mechanism is:

[0069] ;

[0070] wherein, is the spatio-temporal aware weight of the ith node to the jth node at the tth time point, exp(·) is the exponential function, N(j) is the set of adjacent nodes of the jth node, k is the third index of the supply-demand heterogeneous graph, sc kj (t) is the attention score of the kth node to the jth node at the tth time point;

[0071] The calculation formula for updating the node embedding vector by weighted aggregation of adjacent node information is:

[0072] ;

[0073] wherein, h j (t+1) is the jth node embedding vector at the t+1 time point, ReLU(·) is the ReLU activation function, W r is the rth edge type weight matrix;

[0074] The supply-demand feature fusion is specifically that, in the supply-demand heterogeneous graph, for each supply edge, the corresponding planting node embedding vector and the sales node embedding vector are extracted and feature fusion is performed to obtain a supply-demand fusion feature vector;

[0075] The feature fusion is specifically that the planting node embedding vector and the sales node embedding vector are first spliced to obtain a supply-demand preliminary splicing feature vector, and the planting node embedding vector and the sales node embedding vector are multiplied element by element to obtain a supply-demand feature interaction term; then the supply-demand preliminary splicing feature vector and the supply-demand feature interaction term are spliced to obtain a supply-demand splicing feature vector, and finally the supply-demand splicing feature vector is nonlinearly transformed by a tanh activation function to generate a supply-demand fusion feature vector, and the calculation formula is:

[0076] ;

[0077] ;

[0078] ;

[0079] wherein, z v (t) is the supply-demand fusion feature vector of the vth supply edge, v is the supply edge index, tanh(·) is the tanh activation function, W c is the compression weight matrix, h one (t) is the supply-demand preliminary splicing feature vector, h two (·) is the supply-demand feature interaction term, b is the bias term, h s (t) is the planting node embedding vector, h g(t) is a sales node embedding vector, is an element-wise multiplication symbol.

[0080] By performing the above operations, in view of the technical problem that in the existing Hami melon supply chain demand prediction process, the planting information and sales information are scattered, the supply and demand correlation is weak, and it is difficult to effectively reflect the actual supply and demand linkage, thereby affecting the prediction accuracy, the scheme creatively uses a heterogeneous correlation embedding modeling method to model the supply and demand characteristics, constructs a multi-dimensional supply and demand heterogeneous graph covering planting plots and sales areas, dynamically perceives the spatio-temporal changes and supply and demand linkage characteristics, and generates more expressive comprehensive features, which helps to improve the prediction accuracy.

[0081] Embodiment four, see Figure 1 and Figure 3 This embodiment is based on the above-mentioned embodiment. The bidirectional linkage prediction is specifically a bidirectional linkage prediction using a bidirectional collaborative supply and demand prediction method based on the supply and demand fusion feature vector, to obtain Hami melon supply and demand prediction data, including the following steps: bidirectional prediction backbone construction, collaborative attention interaction, supply and demand residual correction, and supply and demand prediction model construction.

[0082] The bidirectional prediction backbone construction is specifically constructing a time series encoder to model the time series characteristics of the supply and demand fusion feature vector, to obtain supply and demand time series characteristics, and then constructing a parallel decoding branch to decode the supply and demand time series characteristics respectively, to generate yield prediction sequences and sales volume prediction sequences.

[0083] The parallel decoding branch includes a yield decoding branch and a sales volume decoding branch.

[0084] Preferably, the Transformer encoder is used as the time series encoder, and two Transformer decoders are used as the yield decoding branch and the sales volume decoding branch respectively.

[0085] The collaborative attention interaction is used to model the mutual influence relationship between supply and demand. Specifically, a bidirectional collaborative attention mechanism is introduced between the parallel decoding branches to extract sales volume hidden state features and yield hidden state features from the hidden layers in the parallel decoding branches, calculate sales volume influence factors and yield influence factors through the bidirectional collaborative attention mechanism, and feed back to the parallel decoding branches to adjust the sales volume hidden state features and yield hidden state features in the decoding process.

[0086] The calculation formula of the sales volume influence factor and the yield influence factor is:

[0087] ;

[0088] ;

[0089] In the formula, CoAttsg is a sales influence factor, used to represent the influence of sales on production, softmax(·) is a softmax normalization function, H s is a sales hidden state feature, W q is a query weight matrix, H g is a production hidden state feature, W k is a key weight matrix, d is a scaling factor, W v is a value weight matrix, CoAtt gs is a production influence factor, used to represent the influence of production on sales;

[0090] The feedback is to the parallel decoding branch, and the calculation formula of the sales hidden state feature and the production hidden state feature is adjusted during the decoding process:

[0091] ;

[0092] ;

[0093] In the formula, is the adjusted sales hidden state feature, is the adjusted production hidden state feature;

[0094] The supply and demand residual correction is specifically: through the difference between the sales prediction sequence and the production prediction sequence, the supply and demand residual is obtained, when the absolute value of the supply and demand residual is greater than the residual threshold, the supply and demand residual is dynamically smoothed by the gated recurrent unit to obtain the feedback adjustment amount, and the sales prediction sequence and the production prediction sequence are corrected according to the feedback adjustment amount;

[0095] The calculation formula of the supply and demand residual is:

[0096] ;

[0097] In the formula, R(t+∆) is the supply and demand residual, ∆ is the prediction step, is the sales prediction sequence, is the production prediction sequence;

[0098] The calculation formula for correcting the sales prediction sequence and the production prediction sequence according to the feedback adjustment amount is:

[0099] ;

[0100] ;

[0101] In the formula, is the corrected sales prediction sequence, is the sales correction coefficient with a value range of [0.05, 0.5], which is used to control the correction strength. GRU(·) is the gated recurrent unit. is the feedback regulation quantity, is the revised production forecast series, is the yield correction factor with a value range of [0.05,0.5];

[0102] The calculation formula of the residual threshold is:

[0103] ;

[0104] Where, is the residual threshold, is the elastic coefficient with a value range of [0.05,0.2], is the historical sales average, is the historical production average;

[0105] The supply and demand forecasting model is constructed by executing the bidirectional forecasting backbone construction, the collaborative attention interaction, the supply and demand residual correction, and performing model training to construct a supply and demand forecasting model, and then performing bidirectional linkage forecasting through the supply and demand forecasting model to obtain Hami melon supply and demand forecasting data, which includes a production forecasting sequence and a sales forecasting sequence.

[0106] By performing the above operations, in order to address the technical problem that in the existing Hami melon supply chain demand forecasting process, traditional demand forecasting usually relies on a single modeling of planting information or sales information, ignoring the supply and demand linkage relationship, resulting in the forecast results being prone to supply and demand imbalance and difficult to adjust in time to cope with the supply and demand imbalance, this solution creatively adopts a supply and demand forecasting method based on two-way collaboration to perform two-way linkage forecasting, jointly modeling planting output and market sales, dynamically capturing the mutual influence between the two, and timely discovering and correcting supply and demand differences, thereby improving forecast accuracy and supply chain responsiveness, and ensuring the supply and demand balance and operational efficiency of the Hami melon market.

[0107] Example 5, see Figure 1 This embodiment is based on the above embodiment. The supply and demand forecast report is generated by extracting the supply and demand change trend based on the Hami melon supply and demand forecast data through a time series trend analysis method to generate a Hami melon supply chain demand forecast report; the Hami melon supply chain demand forecast report includes Hami melon supply and demand forecast data, a Hami melon supply and demand change trend chart, and planting adjustment suggestions.

[0108] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; it is not intended to exclude myriad other embodiments of the present application that other presenters can develop. It is also possible, however, that only a single element can be present. Terms such as "a", "an", and "the" are not intended to refer to only a singular entity but include the general class of which a single element is only one species, unless otherwise indicated. Furthermore, the use of the terms "primary" and "secondary", "first" and "second", etc., designate different Stages in the process, and are not intended to otherwise limit the number of stages which can be employed. The terminology includes the words specifically noted above, derivatives thereof, and words of similar import. The designation of a process, method, article, or apparatus as including a list of elements is not a limitation as to additional elements that can be present or additional elements that can be utilized. Nor is it implied that any or all of the claimed elements are required for the optimal performance or success of the process, method, article, or apparatus.

[0109] While the embodiments of the application have been shown and described, it is to be understood that for the purpose of the present application, the change, modification, substitution and variation can be made by those skilled in the art without departing from the spirit and principle of the present application.

[0110] The above description of the application and its embodiments is not restrictive, and the embodiments shown in the drawings are only one of the embodiments of the present application, and the actual structure is not limited thereto. In general, if a person skilled in the art is inspired by it, without departing from the purpose of the present application, without creative design, similar structure and embodiments of the technical solution can be developed, which should belong to the protection scope of the present application.

Claims

1. The Hami melon supply chain demand forecasting system based on artificial intelligence is characterized by: include: Multi-source data preparation module, supply and demand feature modeling module, two-way linkage forecasting module and supply and demand forecast report generation module; The multi-source data preparation module is used for multi-source data preparation, obtains Hami melon supply chain multi-source data through multi-source data preparation, and sends the Hami melon supply chain multi-source data to the supply and demand feature modeling module; The supply and demand feature modeling module is used for supply and demand feature modeling. Based on the multi-source data of the Hami melon supply chain, a heterogeneous association embedding modeling method is used to perform supply and demand feature modeling, obtain a supply and demand fusion feature vector, and send the supply and demand fusion feature vector to the bidirectional linkage prediction module; The two-way linkage prediction module is used for two-way linkage prediction, and adopts a supply and demand prediction method based on two-way collaboration to perform two-way linkage prediction based on the supply and demand fusion feature vector to obtain Hami melon supply and demand prediction data, and sends the Hami melon supply and demand prediction data to the supply and demand forecast report generation module; The supply and demand forecast report generation module is used to generate a supply and demand forecast report to obtain a Hami melon supply chain demand forecast report.

2. The Hami melon supply chain demand forecasting system based on artificial intelligence according to claim 1, characterized in that: The supply and demand feature modeling is specifically based on the multi-source data of the Hami melon supply chain, using a heterogeneous association embedding modeling method to model the supply and demand features, and obtain a supply and demand fusion feature vector, including the following steps: supply and demand heterogeneous graph construction, spatiotemporal weight dynamic embedding and supply and demand feature fusion; The supply and demand heterogeneous graph construction is used to construct a heterogeneous graph containing multiple relationship structures between the planting end and the sales end, specifically by setting nodes, edges and edge weights to construct the supply and demand heterogeneous graph; The spatiotemporal weighted dynamic embedding specifically uses the node attribute feature vector as the initial node embedding vector, obtains the adjacent nodes of each node, calculates the attention score by designing the Fourier spatiotemporal periodic function, generates the spatiotemporal perception weight through the graph attention mechanism, and then updates the node embedding vector by weighted aggregation of adjacent node information; The supply and demand feature fusion is specifically to extract the corresponding planting node embedding vector and sales node embedding vector for each supply edge in the supply and demand heterogeneous graph, and perform feature fusion to obtain the supply and demand fusion feature vector; The feature fusion is specifically as follows: first, the planting node embedding vector and the sales node embedding vector are spliced ​​together to obtain the preliminary supply and demand splicing feature vector; then, the supply and demand feature interaction term is obtained by element-by-element multiplication of the planting node embedding vector and the sales node embedding vector; then, the preliminary supply and demand splicing feature vector and the supply and demand feature interaction term are spliced ​​together to obtain the supply and demand splicing feature vector; finally, the supply and demand splicing feature vector is nonlinearly transformed through the tanh activation function to generate the supply and demand fusion feature vector.

3. The Hami melon supply chain demand forecasting system based on artificial intelligence according to claim 2, characterized in that: In the construction of the heterogeneous supply and demand graph, the node setting is specifically to set the nodes of the heterogeneous supply and demand graph by constructing a planting node and constructing a sales node; The construction of the planting node specifically involves taking each planting plot as a planting node, and taking the plot number, planting area, planting variety and Hami melon yield as planting node attributes; the construction of the sales node specifically involves taking each sales area as a sales node, and taking the sales area number, Hami melon sales volume and promotion intensity as sales node attributes.

4. The Hami melon supply chain demand forecasting system based on artificial intelligence according to claim 3, characterized in that: In the supply and demand heterogeneous graph construction, the edge setting is specifically setting the edge of the supply and demand heterogeneous graph by constructing a supply edge and constructing a sales history edge; The supply edge is constructed by constructing a supply edge between the corresponding planting node and the sales node when there is a direct supply relationship between the planting site and the sales area, and using the spatial distance, transportation cost and transportation volume as the supply edge attributes; The sales history edge is constructed by calculating the sales similarity between two sales regions using the Pearson correlation coefficient. When the sales similarity exceeds a sales similarity threshold, a sales history edge is constructed between the corresponding two sales nodes, and the sales similarity and spatial distance are used as sales history edge attributes. In the construction of the heterogeneous supply and demand graph, the edge weight is set by constructing a standard multi-layer perceptron and combining node attributes and edge attributes to calculate the edge weight.

5. The Hami melon supply chain demand forecasting system based on artificial intelligence according to claim 4, characterized in that: The two-way linkage prediction is specifically based on the supply and demand fusion feature vector and a supply and demand prediction method based on two-way collaboration to perform two-way linkage prediction to obtain Hami melon supply and demand forecast data, including the following steps: two-way prediction backbone construction, collaborative attention interaction, supply and demand residual correction and supply and demand prediction model construction; The bidirectional prediction backbone is constructed by constructing a time series encoder to model the supply and demand fusion feature vector to obtain the supply and demand time series features. Then, by constructing a parallel decoding branch, the supply and demand time series features are decoded respectively to generate the production forecast sequence and the sales forecast sequence. The collaborative attention interaction is used to model the mutual influence relationship between supply and demand. Specifically, a bidirectional collaborative attention mechanism is introduced between the parallel decoding branches. The sales hidden state features and the output hidden state features are extracted from the hidden layers of the parallel decoding branches respectively. The sales impact factors and the output impact factors are calculated through the bidirectional collaborative attention mechanism and fed back to the parallel decoding branches. The sales hidden state features and the output hidden state features are adjusted during the decoding process. The supply and demand residual correction is specifically to obtain the supply and demand residual by taking the difference between the sales forecast sequence and the output forecast sequence. When the absolute value of the supply and demand residual is greater than the residual threshold, the supply and demand residual is dynamically smoothed by a gated recurrent unit to obtain a feedback adjustment amount, and the sales forecast sequence and the output forecast sequence are corrected according to the feedback adjustment amount. The supply and demand forecasting model is constructed by executing the bidirectional forecasting backbone construction, the collaborative attention interaction, the supply and demand residual correction, and performing model training to construct a supply and demand forecasting model, and then performing bidirectional linkage forecasting through the supply and demand forecasting model to obtain Hami melon supply and demand forecasting data, which includes a production forecasting sequence and a sales forecasting sequence.

6. The Hami melon supply chain demand forecasting system based on artificial intelligence according to claim 5, characterized in that: The supply and demand forecast report is generated by extracting the supply and demand change trend based on the Hami melon supply and demand forecast data through a time series trend analysis method to generate a Hami melon supply chain demand forecast report; the Hami melon supply chain demand forecast report includes Hami melon supply and demand forecast data, a Hami melon supply and demand change trend chart and planting adjustment suggestions.

7. The Hami melon supply chain demand forecasting system based on artificial intelligence according to claim 6, characterized in that: The multi-source data preparation specifically involves collecting historical Hami melon planting information, Hami melon historical sales records, and Hami melon supply and marketing information, and performing data cleaning and structured processing to obtain Hami melon supply chain multi-source data; The Hami melon historical planting information includes plot number, planting area, planting variety and Hami melon yield; The historical sales records of Hami melons include sales region numbers, Hami melon sales volume, and intensity of promotional activities; The Hami melon supply and marketing information includes the spatial distance between sales areas, the spatial distance between planting plots and sales areas, transportation costs and transportation volume.

Images