A bulk agricultural product price trend prediction method based on multi-source data

Abstract

This invention discloses a method for predicting the price trend of bulk agricultural products based on multi-source data, belonging to the field of agricultural product price prediction technology. The method includes: using a multi-source heterogeneous dataset, performing an inversion using the output-flow-accumulation mass conservation equation to obtain the daily dynamic implicit inventory sequence and daily dynamic implicit inventory change rate of each related commodity; constructing a multi-commodity weighted directed graph of production and consumption substitution relationships, and embedding the daily dynamic implicit inventory sequence and daily dynamic implicit inventory change rate into the feature vectors of the corresponding agricultural product nodes to generate multi-commodity enhanced graph data; using a graph attention network to aggregate cross-node association information of the multi-commodity enhanced graph data to form a cross-node association information aggregation representation, and extracting the dynamic change features of each agricultural product node by combining historical time-series evolution patterns to obtain a multi-commodity joint price prediction sequence. This invention achieves continuous characterization of the implicit inventory status and inventory change trend of each related commodity.

Description

Technical Field

[0001] This invention relates to the field of agricultural product price forecasting technology, and in particular to a method for forecasting the price trends of major agricultural products based on multi-source data. Background Technology

[0002] With the development of remote sensing, digital supply chain data collection, market transaction tracking, and industrial cloud computing, price forecasting for bulk agricultural products has gradually shifted from statistical analysis based on single price series to multi-source collaborative forecasting that integrates information on production, logistics, inventory, transactions, and external disturbances. Industrial cloud computing provides fundamental support for the aggregation, storage, scheduling, and parallel analysis of data from multiple regions, multiple links, and multiple varieties, making cross-temporal and multi-granular data-driven price forecasting possible.

[0003] However, existing related technologies still have the following two main problems: First, existing schemes mostly use observable variables such as explicit prices, transaction volume, and weather as core inputs, which do not make sufficient use of the implicit balance relationship between output, flow, and accumulation. They are difficult to continuously characterize the quality in transit, implicit inventory, and the direction of inventory changes, resulting in limited ability to identify supply and demand mismatches and trend inflection points. Second, existing multi-product forecasting schemes still have a relatively rough characterization of substitution relationships, production competition relationships, and consumption migration relationships. They often fail to reflect cross-price elasticity and multi-product linkage processes within a unified framework, thus affecting forecast accuracy and interpretability. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for predicting the price trend of bulk agricultural products based on multi-source data to solve the problems of insufficient characterization of implicit inventory and inadequate reflection of the linkage between multiple varieties in the existing technology.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a method for predicting the price trend of bulk agricultural products based on multi-source data, comprising: collecting multi-source heterogeneous basic data and performing spatiotemporal benchmark unification processing to generate a multi-source heterogeneous dataset; based on the multi-source heterogeneous dataset, using the output-flow-accumulation mass conservation equation for inversion to obtain the daily dynamic implicit inventory sequence and daily dynamic implicit inventory change rate of each related variety; constructing a multi-variety weighted directed graph of production and consumption substitution relationship, and embedding the daily dynamic implicit inventory sequence and daily dynamic implicit inventory change rate into the feature vector of the corresponding agricultural product node after encoding, to generate multi-variety enhanced graph data; using a graph attention network to aggregate cross-node association information of the multi-variety enhanced graph data to form a cross-node association information aggregation representation, and extracting the dynamic change characteristics of each agricultural product node in combination with historical time series evolution patterns to obtain a multi-variety joint price prediction sequence; based on the multi-variety joint price prediction sequence, combining the market equilibrium constraint loss function reflecting the cross-price elasticity of substitute crops for parameter backpropagation optimization to obtain the price prediction value, and generating a bulk agricultural product price trend prediction report based on the final price prediction value.

[0007] As a preferred embodiment of the method for predicting the price trend of bulk agricultural products based on multi-source data described in this invention, the multi-source heterogeneous basic data includes remote sensing image data of the main producing areas of the target bulk agricultural products and alternative crops, supply chain logistics trajectory data, and historical market transaction data.

[0008] As a preferred embodiment of the bulk agricultural product price trend prediction method based on multi-source data described in this invention, the specific steps for generating the multi-source heterogeneous dataset are as follows: Collect multi-source heterogeneous basic data and encapsulate them in a structured manner according to source identifier, variety code, spatial identifier, observation time, unit of measurement and original observation value to obtain the original record set; The variety codes and spatial identifiers in the records from various sources are uniformly mapped using the original record set. The original observations are then subjected to dimensional unification, time alignment, anomaly correction, fusion completion, and structured organization to generate a multi-source heterogeneous dataset.

[0009] As a preferred embodiment of the method for predicting the price trend of bulk agricultural products based on multi-source data described in this invention, the specific steps for obtaining the daily dynamic implicit inventory sequence and the daily dynamic implicit inventory change rate of each related commodity are as follows: The multi-source heterogeneous dataset is used to collect the daily output records, flow records, accumulation records and processing and consumption records of related varieties in each regional node, and to construct a two-domain inversion map of edge points. By using the edge point dual-domain inversion map, continuous event-driven reconstruction is performed on the discrete observation records of each related variety at each regional node, and the transport characteristics of the in-transit quality at the regional edge are performed to obtain a continuous transport field. Based on the continuous transport field, the output, inflow, outflow, accumulation and processing consumption of each node are substituted into the output-flow-accumulation mass conservation equation for joint inversion to obtain the daily dynamic implicit inventory sequence of each related commodity at each regional node. Based on the daily dynamic implicit inventory sequence, the daily dynamic implicit inventory change rate of each related product is calculated according to the inventory increase and decrease relationship between adjacent dates.

[0010] As a preferred embodiment of the method for predicting the price trend of bulk agricultural products based on multi-source data described in this invention, the specific steps for constructing a multi-variety weighted directed graph of production and consumption substitution relationships are as follows: Based on the daily dynamic implicit inventory sequence and the daily dynamic implicit inventory change rate, combined with the competition relationship of planting resources, the substitution relationship of processing formulas and the substitution relationship of consumption, candidate substitution relationships among varieties are screened to form candidate substitution units; Extract production substitution events that characterize the transfer of resources on the production side and consumption substitution events that characterize the migration of demand on the consumption side from the candidate substitution units to obtain a set of substitution events; Based on the set of substitution events, the daily directed edge weights between varieties are calculated according to the substitution direction, substitution intensity, and duration, forming a set of daily directed edges; The daily directed edge set is used to associate nodes and attach edge weights to each related variety, and a multi-variety weighted directed graph of production and consumption substitution relationships is constructed.

[0011] As a preferred embodiment of the method for predicting the price trend of bulk agricultural products based on multi-source data according to the present invention, the specific steps for generating multi-variety enhanced map data are as follows: The daily dynamic implicit inventory sequence and daily dynamic implicit inventory change rate are extracted from the multi-variety weighted directed graph, and the node time sequence is organized by combining the node association edge information to obtain the node event package. The daily dynamic implicit inventory sequence and the daily dynamic implicit inventory change rate are jointly encoded using node event packages, and the node joint encoding vector is generated by combining the directional features and edge weight features of the node associated edges. The joint encoding vector of nodes is embedded into the feature vector of the corresponding agricultural product node, and then associated with the multi-variety weighted directed graph to generate multi-variety enhanced graph data.

[0012] As a preferred embodiment of the bulk agricultural product price trend prediction method based on multi-source data described in this invention, the specific steps for forming a cross-node associated information aggregation representation are as follows: Extract the node feature vectors of each agricultural product node and the edge weights and directions of directed edges between nodes from the multi-variety enhanced graph data, and construct the node directional neighborhood relationship. By utilizing the directional neighborhood relationship of nodes, the node feature vectors of each agricultural product node and the edge weights and directions of the corresponding directed edges are input into the graph attention network to calculate the attention weights between the node and its neighboring nodes. Attention weights are used to associate and aggregate the incoming and outgoing neighborhood information and node feature vectors of each agricultural product node, forming a cross-node association information aggregation representation.

[0013] As a preferred embodiment of the method for predicting the price trend of bulk agricultural products based on multi-source data according to the present invention, the specific steps for obtaining the multi-commodity joint price prediction sequence are as follows: Extract the node feature vectors and node association edge information corresponding to each agricultural product node from the multi-variety enhanced graph data in historical time sequence, and construct the node historical evolution input sequence; The historical evolution input sequence of nodes is used to recursively calculate the historical time-series state of each agricultural product node, and to form dynamic change characteristics that characterize the state level, rate of change and turning trend of each agricultural product node. Based on dynamic changes, and combined with the correlation between various agricultural product nodes, a joint recursive decoding is performed to obtain a multi-variety joint price prediction sequence.

[0014] As a preferred embodiment of the method for predicting the price trend of bulk agricultural products based on multi-source data according to the present invention, the specific steps for obtaining the price prediction value are as follows: Based on the multi-variety joint price forecast series, and combined with the substitution relationship edge weights of each agricultural product, the daily dynamic implicit inventory series, and the daily dynamic implicit inventory change rate, a market equilibrium constraint loss function reflecting the cross price elasticity of substitute crops is constructed. The market equilibrium constraint loss function is used to map the multi-product joint price forecast series to a constrained price series that satisfies the supply and demand balance constraint and the substitution relationship constraint. By backpropagating the market equilibrium constraint loss function using the constrained price series, we obtain the gradient information corresponding to each parameter. Based on the gradient information, we iteratively update each parameter to form the price prediction value.

[0015] As a preferred embodiment of the bulk agricultural product price trend prediction method based on multi-source data described in this invention, the specific steps for generating the bulk agricultural product price trend prediction report are as follows: Based on price forecasts, the magnitude, rate of change, and degree of fluctuation of prices for each agricultural product within the forecast period are calculated to form a price trend evolution description sequence. Based on the price trend evolution description sequence, the trend status and key turning points of each agricultural product are identified. Combined with the daily dynamic implicit inventory change rate, substitution relationship margin weight and market equilibrium constraint deviation, trend interpretation information is formed, and a price trend forecast report for bulk agricultural products is generated according to the preset report structure.

[0016] The beneficial effects of this invention are as follows: By daily aggregating the output, flow, accumulation, and processing / consumption information of related varieties at various regional nodes, and combining it with edge point dual-domain inversion maps, the output-flow-accumulation mass conservation equation, and data collaborative processing supported by industrial cloud computing, the invention achieves continuous characterization of the implicit inventory status and inventory change trends of each related variety, thereby improving the ability to identify supply and demand imbalances and enhancing the accuracy of trend inflection point judgment. By constructing a market equilibrium constraint loss function that reflects the cross-price elasticity of substitute crops, the invention optimizes the multi-variety joint price prediction sequence through backpropagation to obtain price prediction values, thereby achieving coordinated correction of substitution relationship transmission and supply and demand balance constraints, and improving the accuracy of price prediction, consistency of linkage, and interpretability of results. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a method for predicting price trends of bulk agricultural products based on multi-source data.

[0019] Figure 2 A flowchart for generating multi-source heterogeneous datasets.

[0020] Figure 3 A flowchart for generating enhanced spectral data for multiple varieties.

[0021] Figure 4 This is a heatmap of the daily directed edge weight matrix.

[0022] Figure 5 A line chart comparing the forecasts for the normalized price index of corn. Detailed Implementation

[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0024] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0025] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0026] Reference Figures 1-5 This is one embodiment of the present invention, which provides a method for predicting the price trend of bulk agricultural products based on multi-source data, including the following steps: S1. Collect multi-source heterogeneous basic data and perform spatiotemporal benchmark unification processing to generate multi-source heterogeneous datasets.

[0027] Collect multi-source heterogeneous basic data and encapsulate them in a structured manner according to source identifier, variety code, spatial identifier, observation time, unit of measurement and original observation value to obtain the original record set.

[0028] It should be noted that remote sensing image data of the main producing areas of the target bulk agricultural products and alternative crops are obtained from publicly available satellite data sources, supply chain logistics trajectory data are obtained from logistics information platforms, and historical market transaction data are obtained from exchanges or financial data terminals.

[0029] Deep analysis of multi-source heterogeneous basic data is performed to accurately extract six key fields, including data source identifier, agricultural product variety code, spatial identifier of geographical location or trading venue, specific observation time, unit of measurement, and core original observation value. These fields are then packaged and organized according to a pre-set unified standard to form standardized structured records, which are then incorporated into the original record set.

[0030] The unified standards are pre-set based on the core business attributes and data characteristics of bulk agricultural products and their substitute crops in supply chain logistics and market transactions.

[0031] The variety codes and spatial identifiers in the records from various sources are uniformly mapped using the original record set. The original observations are then subjected to dimensional unification, time alignment, anomaly correction, fusion completion, and structured organization to generate a multi-source heterogeneous dataset.

[0032] The specific process includes: using the original record set, according to the preset variety and spatial mapping dictionary, converting the variety codes in the records from each source into unified variety identifiers, and converting the spatial identifiers into unified geographic coordinates; based on the records after unified mapping, converting the original observation values ​​into standard observation values ​​under the standard unit of measurement according to the unit of measurement conversion rules, and at the same time normalizing the observation time according to the preset time window to achieve time alignment and form a unified index record set.

[0033] Outlier detection is performed on the unified index record set based on statistics. Abnormal standard observations that exceed the statistical judgment boundary are corrected by interpolation of neighboring data. For records under the same unified variety identifier and the same unified geographic coordinates, standard observations are weighted and fused according to the priority strategy of source identifier, and missing time node data is filled in. The data is arranged and encapsulated in an orderly manner according to the preset data structure to generate a multi-source heterogeneous dataset.

[0034] It should be noted that the variety and spatial mapping dictionary is preset based on the industry standard classification system of major agricultural products and their substitute crops and the standard geocoding system in the geographic information system; the time window is preset based on the business cycle characteristics and data update frequency of major agricultural products and their substitute crops in supply chain logistics and market transactions; the statistical judgment boundary is preset based on the outlier cutoff point rule in the interquartile range algorithm; and the data structure is preset based on the business cycle characteristics and data features of major agricultural products and their substitute crops in supply chain logistics and market transactions.

[0035] S2. Based on multi-source heterogeneous datasets, the output-flow-stacking mass conservation equation is used for inversion to obtain the daily dynamic implicit inventory sequence and daily dynamic implicit inventory change rate of each related commodity.

[0036] Using multi-source heterogeneous datasets, the production, flow, accumulation, and processing / consumption records of related varieties in each regional node are collected daily, and a two-domain inversion graph of edge points is constructed.

[0037] The specific process includes using multi-source heterogeneous datasets to aggregate and statistically analyze the production records, flow records, accumulation records, and processing and consumption records of related varieties in each regional node according to daily time windows, generating a daily aggregated record set.

[0038] The output, accumulation, and processing / consumption quantities in the daily aggregated record set are used as attribute information and mapped to the corresponding node entities. The flow records in the daily aggregated record set are used as connecting edge entities, and the flow quantity is used as the weight of the connecting edge entities. A linked list structure is established using the node entities as indexes, and connecting edge entities originating from nodes in the same region are linked by pointers to form an outgoing edge list, thus constructing an adjacency list storage structure. Based on the outgoing edge list pointing relationship in the adjacency list storage structure, the state change logic of the node entities is deduced in reverse, generating a two-domain inversion graph of edge points.

[0039] It should be noted that the edge-point dual-domain inversion graph is a graph structure data that connects different node entities through connecting edge entities and can infer the state changes of node entities through the weights of the connecting edge entities; the associated varieties include target bulk agricultural products and their substitute crops.

[0040] By using the edge-point dual-domain inversion map, continuous event-driven reconstruction is performed on the discrete observation records of each related variety at each regional node, and the transport characteristics of the in-transit quality at the regional edge are performed to obtain a continuous transport field.

[0041] The specific process includes arranging the discrete observation records on the node entities in the edge-point dual-domain inversion graph in chronological order, using the occurrence time of the driving event as the state transition node, using the daily dynamic implicit inventory change rate to obtain the state increment between two adjacent discrete observation records, and filling the time gap between discrete observation records through a linear interpolation algorithm, thereby transforming the discrete output, accumulation, and processing consumption data into a continuously changing time series curve, forming a continuous event-driven reconstruction record.

[0042] The continuous event-driven reconstruction record decomposes the flow record corresponding to the connected edge entity into five elements: starting region node, ending region node, departure time, arrival time, and flow volume. The spatial coordinates of the starting region node and departure time are used to construct the spatiotemporal starting point, and the spatial coordinates of the ending region node and arrival time are used to construct the spatiotemporal ending point. The flow volume is linearly distributed between the spatiotemporal starting point and the spatiotemporal ending point as the transport flux. The in-transit mass distribution density at any time on the transport path is obtained, and a continuous transport field covering the entire transport path and transport time period is generated.

[0043] Based on the continuous transport field, the output, inflow, outflow, accumulation and processing consumption of each node are substituted into the output-flow-accumulation mass conservation equation for joint inversion to obtain the daily dynamic implicit inventory sequence of each related commodity at each regional node.

[0044] The specific process includes: analyzing the flow status of regional nodes within a daily time window using a continuous transport field; accumulating the flow of entities connecting the regional nodes as terminating regional nodes to obtain the node inflow; accumulating the flow of entities connecting the regional nodes as starting regional nodes to obtain the node outflow; reading the daily node output, node accumulation, and processing / consumption from a multi-source heterogeneous dataset; adding the node output to the node inflow to obtain the daily total supply; adding the node outflow, node accumulation, and processing / consumption to obtain the daily total consumption; calculating the difference between the daily total supply and total consumption using the output-flow-accumulation mass conservation equation to obtain the daily net inventory change; using the previous day's daily net inventory change as the initial benchmark for the current day, superimposing the current day's daily net inventory change and accumulating it day by day to generate a daily dynamic implicit inventory sequence for each related commodity at each regional node.

[0045] It should be noted that the output-flow-accumulation mass conservation equation is based on the law of conservation of mass. It determines the expression for the net change in inventory by balancing the quantitative relationships between node output, node inflow and outflow, node accumulation and processing consumption within a specific time window for each regional node.

[0046] Based on the daily dynamic implicit inventory sequence, the daily dynamic implicit inventory change rate of each related product is calculated according to the inventory increase / decrease relationship between adjacent days. The expression is as follows: ; in, Indicates related varieties On the date The daily dynamic rate of change of implicit inventory; Indicates the related product index; Indicates date index; Indicates related varieties On the date Daily dynamic hidden inventory levels; Indicates related varieties On the date Daily dynamic hidden inventory levels; This indicates the time interval between two adjacent dates.

[0047] The specific process includes extracting the daily dynamic implicit inventory of each related product on two adjacent dates at each regional node from the daily dynamic implicit inventory sequence, and calculating the difference between the daily dynamic implicit inventory of two adjacent dates to obtain the inventory increase or decrease; using the arithmetic mean of the daily dynamic implicit inventory of two adjacent dates as the benchmark inventory, the inventory increase or decrease is compared with the benchmark inventory, and the result is applied to the reciprocal of the time interval between two adjacent dates to obtain the daily dynamic implicit inventory change rate of each related product.

[0048] It should be noted that the daily dynamic hidden inventory change rate reflects the dynamic rate and direction of the hidden inventory level of each related commodity at each regional node within two consecutive daily time windows. It is a key indicator for measuring the degree of short-term imbalance in supply and demand and potential market pressure.

[0049] S3. Construct a multi-variety weighted directed graph of the production and consumption substitution relationship, and embed the daily dynamic implicit inventory sequence and the daily dynamic implicit inventory change rate into the feature vector of the corresponding agricultural product node after encoding, to generate multi-variety enhanced graph data.

[0050] Based on the daily dynamic implicit inventory sequence and the daily dynamic implicit inventory change rate, and combined with the competition relationship of planting resources, the substitution relationship of processing formulas and the substitution relationship of consumption, candidate substitution relationships among varieties are screened to form candidate substitution units.

[0051] The specific process includes: using daily dynamic implicit inventory sequences and daily dynamic implicit inventory change rates to analyze the correlation between the inventory fluctuation trends and change rates of related varieties at nodes in the same region, and identifying variety combinations with synchronous or inverse inventory changes; combining planting resource competition relationships to screen related variety pairs with overlapping planting seasons and conflicting arable land resources; combining processing formula substitution relationships to screen related variety pairs with interchangeable raw materials in the processing stage; combining consumption substitution relationships to screen related variety pairs with similar terminal market functions and price linkages; and integrating related variety pairs that satisfy at least one of the following: inventory fluctuation correlation, planting resource competition relationship, processing formula substitution relationship, and consumption substitution relationship into candidate substitution relationships between varieties, and encapsulating them into candidate substitution units.

[0052] Extract production substitution events that characterize the transfer of resources on the production side and consumption substitution events that characterize the migration of demand on the consumption side from the candidate substitution units to obtain a set of substitution events.

[0053] The specific process includes: traversing candidate substitution units and screening candidate substitution units whose relationship type is identified as a competition relationship for planting resources; combining the daily dynamic implicit inventory change rate, identifying time segments where the inventory changes of two related varieties on the same regional node are in opposite directions and the change rates are lagging, defining the related varieties whose inventory decreases in the time segment as resource transferors and the related varieties whose inventory increases as resource transferees; encapsulating the resource transferor, resource transferee, regional node identifier, start and end time of the time segment, and inventory change rate ratio into a production substitution event and merging it into the production substitution event subset.

[0054] Based on the set of substitution events, the daily directed edge weights between varieties are calculated according to the substitution direction, substitution intensity, and duration, forming a set of daily directed edges.

[0055] The specific process includes: traversing the set of substitution events, defining the resource transferor or demand migration origin as the source product index, defining the resource transferee or demand migration origin as the target product index, and defining the event time segment as the duration of the substitution relationship; calculating the substitution strength of the source product index to the target product index by combining the inventory change rate ratio, inventory fluctuation correlation coefficient, and daily dynamic implicit inventory change rate; calculating the daily directed edge weights using the substitution strength, the duration of the substitution relationship, and the corresponding standardized benchmark value through an exponential decay function; and encapsulating the source product index, the target product index, the daily directed edge weights, and the date into a daily directed edge set.

[0056] The expression for calculating the daily directed edge weights between varieties is: ; in, Indicates on date seasonal varieties Target varieties The daily degree has a directional weight; Indicates the source variety index; Indicates the target variety index; Represents the natural exponential function; Indicates on date seasonal varieties For varieties The strength of substitution; This represents the standardized benchmark value for alternative strength. Indicates the duration adjustment factor; Indicates on date seasonal varieties Target varieties The duration of the substitution relationship; A standardized baseline value representing the duration.

[0057] It should be noted that, It is obtained by statistically analyzing the maximum value or high quantile reference value of the substitution intensity samples of each variety within a historical period; This was obtained through fitting analysis of the duration of substitution relationships and the degree of price linkage in historical samples; It is obtained by statistically analyzing the mean, median, or quantile reference values ​​of the duration of each substitution relationship in historical samples.

[0058] The daily directed edge set is used to associate nodes and attach edge weights to each related variety, and a multi-variety weighted directed graph of production and consumption substitution relationships is constructed.

[0059] The specific process includes: extracting the source variety index and target variety index from the daily directed edge set; identifying all unique variety indexes as graph node entities; assigning an independent node identifier to each graph node entity to complete the node association of each related variety; establishing a directed connection edge between the corresponding two graph node entities based on the pointing relationship between the source variety index and the target variety index of each daily directed edge record in the daily directed edge set; and attaching the daily directed edge weight encapsulated in the daily directed edge record to the directed connection edge as the edge weight attribute to complete the edge weight attachment.

[0060] By utilizing graph node entities with completed node associations and edge weights, and organizing them according to graph data structure standards, a multi-variety weighted directed graph of production and consumption substitution relationships is constructed, with all graph node entities as the vertex set and all directed edges as the edge set.

[0061] It should be noted that the multi-variety weighted directed graph visually presents the substitution relationship and its strength between production and consumption sides of various varieties through nodes and weighted directed edges, providing structured data support for analyzing the transmission path of variety substitution and predicting market fluctuations.

[0062] like Figure 4 The daily directed edge weight matrix heatmap illustrates the strength and direction of substitution relationships among related commodities. The horizontal axis represents the target commodity index, and the vertical axis represents the source commodity index. The darker the color, the higher the daily directed edge weight of the corresponding source commodity on the target commodity. This clearly shows that within a specific daily time window, the production and consumption substitution relationships among different related commodities are not fixed but dynamically adjusted with changes in the daily dynamic implicit inventory sequence, the daily dynamic implicit inventory change rate, and the set of substitution events. Therefore, this graph demonstrates that the present invention has a good ability to characterize the substitution relationship transmission path and intensity.

[0063] The daily dynamic implicit inventory sequence and daily dynamic implicit inventory change rate are extracted from the multi-variety weighted directed graph, and the node event package is obtained by combining the node association edge information to organize the node time sequence.

[0064] The specific process includes: traversing each graph node entity in the multi-variety weighted directed graph; matching the corresponding associated varieties based on the node identifier of the graph node entity; extracting the inventory data of the associated varieties in each regional node from the daily dynamic implicit inventory sequence; extracting the corresponding change rate data from the daily dynamic implicit inventory change rate; arranging the inventory data and change rate data in chronological order to form the node time series data of the graph node entity; extracting all directed edges connected to the graph node entity from the multi-variety weighted directed graph; obtaining the source variety index, target variety index, and daily directed edge weight of each directed edge; and encapsulating the node time series data and directed edge information into a node event package.

[0065] The daily dynamic implicit inventory sequence and the daily dynamic implicit inventory change rate are jointly encoded using node event packages, and the node joint encoding vector is generated by combining the directional features and edge weight features of the node associated edges.

[0066] The specific process includes extracting node time-series data from the node event package, and vectorizing the numerical distribution characteristics of the daily dynamic implicit inventory sequence and the fluctuation trend characteristics of the daily dynamic implicit inventory change rate to generate an inventory status encoding vector; at the same time, parsing the directed connection edge information in the node event package, determining the edge direction features based on the pointing relationship between the source variety index and the target variety index, extracting the numerical magnitude of the daily directed edge weights to determine the edge weight features, and fusing the edge direction features and edge weight features to generate an association structure encoding vector.

[0067] By using inventory status encoding vectors and associated structure encoding vectors, vectors are concatenated according to a preset feature concatenation dimension. This integrates the daily dynamic implicit inventory sequence, the daily dynamic implicit inventory change rate, and the directional and weight features of the node-associated edges into a vector representation with a unified dimension, generating a node joint encoding vector.

[0068] It should be noted that the feature concatenation dimension is a pre-defined dimension parameter used to integrate the features of the two types of vectors, based on the vector lengths of the inventory status encoding vector and the associated structure encoding vector, as well as the feature fusion strategy.

[0069] The joint encoding vector of nodes is embedded into the feature vector of the corresponding agricultural product node, and then associated with the multi-variety weighted directed graph to generate multi-variety enhanced graph data.

[0070] The specific process includes: traversing each graph node entity in the multi-variety weighted directed graph, matching the node joint encoding vector according to the node identifier of the graph node entity, and writing the node joint encoding vector into the feature vector storage space of the corresponding graph node entity; associating all graph node entities with embedded node joint encoding vectors with the directed connection edges in the multi-variety weighted directed graph according to the graph data structure, so that the feature vector of each graph node entity contains the daily dynamic implicit inventory sequence, the daily dynamic implicit inventory change rate, and the directional features and edge weight features of the node association edges, thereby generating multi-variety enhanced graph data.

[0071] It should be noted that the multi-variety enhanced atlas data is constructed based on multi-source heterogeneous datasets, covering core data such as daily dynamic implicit inventory sequences, daily dynamic implicit inventory change rates, daily directed edge sets, daily directed edge weights, candidate substitution units, subsets of production substitution events, and candidate substitution relationships between varieties. It is used to characterize the complex relationships and dynamic evolution characteristics of major agricultural products and their substitute crops in the supply chain, production end, and consumption end.

[0072] S4. Graph attention network is used to aggregate cross-node association information of multi-variety enhanced graph data to form a cross-node association information aggregation representation. Combined with historical time series evolution patterns, dynamic change features of each agricultural product node are extracted to obtain a multi-variety joint price prediction sequence.

[0073] The node feature vectors of each agricultural product node and the edge weights and directions of directed edges between nodes are extracted from the multi-variety enhanced graph data, and the node directional neighborhood relationships are constructed.

[0074] The specific process includes reading the joint encoding vector of each graph node entity from the multi-variety enhanced graph data as the node feature vector, and obtaining the daily directed edge weight of the directed connection edge and the pointing relationship between the source variety index and the target variety index.

[0075] Based on the extracted node feature vectors, edge weights, and direction information, for each graph node entity, all directed connecting edges with the source variety index of that graph node entity are selected, and the target variety index corresponding to the selected directed connecting edges is determined as the outgoing neighbor node of that graph node entity; all directed connecting edges with the target variety index of that graph node entity are selected, and the source variety index corresponding to the selected directed connecting edges is determined as the incoming neighbor node of that graph node entity; all outgoing and incoming neighbor nodes are classified and organized according to edge direction information to construct node directional neighborhood relationships.

[0076] By utilizing the directional neighborhood relationship of nodes, the node feature vectors of each agricultural product node and the edge weights and directions of the corresponding directed edges are input into the graph attention network, and the attention weights between the node and its neighboring nodes are output.

[0077] The specific process includes: traversing each graph node entity in the directed neighborhood relationship; concatenating the node feature vector of the graph node entity with the node feature vectors of its outgoing and incoming neighboring nodes; and using the daily directed edge weights on the directed connections as additional features of the concatenated vector to form a neighborhood feature set containing information about the center node, neighboring nodes, and edge weights. The neighborhood feature set is then fed into the attention mechanism layer of the graph attention network. The shared attention parameters in the attention mechanism layer are used to perform correlation analysis on the feature vectors of the center node and neighboring nodes in the neighborhood feature set, and a weighted mapping is performed in combination with the daily directed edge weights to obtain the attention weights between the graph node entity and each outgoing and incoming neighboring node.

[0078] Furthermore, the training process of the graph attention network involves inputting node features and inter-node relationships into the graph attention network, obtaining a unified representation through feature mapping, acquiring the attention scores between the target node and its neighboring nodes, and normalizing these scores to form attention weights. The attention weights are then used to weight and aggregate the features of neighboring nodes to obtain updated node representations, which are then input into the prediction layer to form corresponding prediction outputs. A loss function is constructed based on the error between the predicted output and the true value. The gradients corresponding to each parameter are obtained through backpropagation, and the weight parameters are iteratively updated until the loss converges, thus completing the training of the graph attention network.

[0079] Attention weights are used to associate and aggregate the incoming and outgoing neighborhood information and node feature vectors of each agricultural product node, forming a cross-node association information aggregation representation.

[0080] The specific process includes: traversing each graph node entity in the node-oriented neighborhood relationship; using the attention weights between the graph node entity and each outgoing and incoming neighboring node, performing a weighted summation operation on the node feature vectors of the outgoing neighboring nodes of the graph node entity to obtain the outgoing aggregated feature of the graph node entity; and performing a weighted summation operation on the node feature vectors of the incoming neighboring nodes of the graph node entity to obtain the incoming aggregated feature of the graph node entity; and concatenating the node feature vectors, outgoing aggregated features, and incoming aggregated features of the graph node entity to form a cross-node association information aggregation representation.

[0081] Extract the node feature vectors and node association edge information corresponding to each agricultural product node from the multi-variety enhanced graph data in historical time sequence, and construct the node historical evolution input sequence.

[0082] The specific process includes: reading the joint encoding vectors of each graph node entity within a preset historical time window from the multi-variety enhanced graph data according to historical time sequence, and arranging the joint encoding vectors in chronological order to obtain the node feature vector sequence corresponding to each agricultural product node; extracting the directed connection edges of each graph node entity within a preset historical time window from the multi-variety enhanced graph data, and arranging the source variety index, target variety index, and daily directed edge weights corresponding to the directed connection edges in chronological order to obtain the node association edge information sequence corresponding to each agricultural product node.

[0083] Align the node feature vector sequence and the node association edge information sequence corresponding to each agricultural product node according to the time index, encapsulate the node feature vector and the node association edge information under the same time index to obtain the node historical state package corresponding to each time index; concatenate all the node historical state packages according to the chronological order of the time index to construct the node historical evolution input sequence.

[0084] It should be noted that the historical time window is preset based on the business cycle characteristics and data update frequency of bulk agricultural products and their substitute crops in supply chain logistics and market transactions.

[0085] The historical evolution input sequence of nodes is used to recursively calculate the historical time-series state of each agricultural product node, and to form dynamic change characteristics that characterize the state level, rate of change and turning trend of each agricultural product node.

[0086] The specific process includes: processing the time-series information corresponding to each agricultural product node in the historical evolution input sequence in chronological order, continuously updating the historical time-series state of the previous moment with the current time-series information to form the historical time-series state corresponding to each agricultural product node; extracting state increase / decrease information based on the changes between adjacent historical time-series states, extracting change rate information based on the rate of change between multiple consecutive historical time-series states, and extracting turning trend information based on the changing relationship of historical time-series states from rising to slowing down, from falling to rising, or from stable to fluctuating; and merging and organizing the state increase / decrease information, change rate information, and turning trend information to form dynamic change characteristics.

[0087] Based on dynamic changes, and combined with the correlation between various agricultural product nodes, a joint recursive decoding is performed to obtain a multi-variety joint price prediction sequence.

[0088] The specific process includes: aggregating dynamic change features and cross-node correlation information into vector concatenation to form a joint state representation that includes the evolutionary laws of each agricultural product node and the substitution relationship between varieties; passing the joint state representation into a fully connected layer for linear transformation to obtain the initial prediction state of each agricultural product node at the start of the prediction; using the initial prediction state as a benchmark, recursively iterating according to the prediction time step, superimposing the prediction state of the previous prediction time step with the turning trend information in the dynamic change features to obtain the prediction value of the agricultural product node at the current prediction time step; and traversing all prediction time steps, arranging the prediction values ​​of the agricultural product nodes corresponding to each prediction time step in chronological order to form a multi-variety joint price prediction sequence.

[0089] S5. Based on the multi-variety joint price forecast sequence, the parameters are optimized by backpropagation using the market equilibrium constraint loss function that reflects the cross-price elasticity of substitute crops to obtain the price forecast value. Based on the final price forecast value, a price trend forecast report for major agricultural products is generated.

[0090] Based on the multi-variety joint price forecast series, and combined with the substitution relationship edge weights of each agricultural product, the daily dynamic implicit inventory series, and the daily dynamic implicit inventory change rate, a market equilibrium constraint loss function reflecting the cross-price elasticity of substitute crops is constructed.

[0091] The specific process includes: extracting the predicted price sequence corresponding to each agricultural product node from the multi-variety joint price prediction sequence; combining the daily dynamic implicit inventory sequence and the daily dynamic implicit inventory change rate to obtain the price change rate and inventory change rate of each agricultural product node within the prediction time step; based on the inter-variety correlation direction determined by the substitution relationship edge weights; calculating the ratio between the price change rate corresponding to the source variety index and the inventory change rate corresponding to the target variety index to obtain the elasticity coefficient reflecting the cross-price elasticity of substitute crops; weighting and fusing the elasticity coefficient with the substitution relationship edge weights to construct a constraint term characterizing the market substitution transmission strength; obtaining the prediction deviation between each predicted price sequence and the historical real price sequence in the multi-variety joint price prediction sequence; and superimposing the prediction deviation and the constraint term to construct a market equilibrium constraint loss function reflecting the cross-price elasticity of substitute crops.

[0092] It should be noted that the market equilibrium constraint loss function is a mathematical function used to quantify the degree to which the actual market state deviates from the ideal supply and demand equilibrium state. Its magnitude reflects the level of market distortion or efficiency loss.

[0093] The market equilibrium constraint loss function is used to map the multi-product joint price forecast series to a constrained price series that satisfies the supply and demand balance constraint and the substitution relationship constraint.

[0094] The specific process includes taking the multi-variety joint price prediction sequence as the initial solution, obtaining the gradient direction of the market equilibrium constraint loss function reflecting the cross price elasticity of substitute crops relative to each predicted price sequence in the multi-variety joint price prediction sequence, and using the gradient descent algorithm to iteratively correct each predicted price sequence along the reverse gradient direction to obtain a constraint price sequence that satisfies the supply and demand balance constraint and the substitution relationship constraint.

[0095] It should be noted that the supply and demand balance constraint refers to ensuring that the market price adjusts to a state of equilibrium where the total supply and total demand are equal or tend to be consistent by constraining the difference between the supply and demand. The substitution relationship constraint refers to using the cross-price elasticity characteristics between substitute crops to limit the correlation strength between the price change of one commodity and the corresponding change in the demand of another commodity, so as to reflect the substitution transmission law between varieties.

[0096] By backpropagating the market equilibrium constraint loss function using the constrained price series, we obtain the gradient information corresponding to each parameter. Based on the gradient information, we iteratively update each parameter to form the price prediction value.

[0097] The specific process includes using the constrained price series to obtain the value of the market equilibrium constrained loss function that reflects the cross-price elasticity of alternative crops, and using an automatic differentiation mechanism to backpropagate and differentiate the constrained price series to obtain gradient information that reflects the sensitivity of the loss function to changes in the constrained price series.

[0098] Based on the descent direction and learning rate determined by the gradient information, each predicted price sequence in the constrained price sequence is iteratively updated until the market equilibrium constrained loss function, which reflects the cross-price elasticity of alternative crops, converges to the loss convergence threshold, thus forming the price prediction value.

[0099] It should be noted that the automatic differentiation mechanism refers to the method of using the backpropagation algorithm to obtain the gradient information of the market equilibrium constraint loss function, which reflects the cross-price elasticity of substitute crops, relative to each predicted price sequence in the constraint price sequence, so as to guide the iterative update of parameters. Each parameter refers to all learnable variables used in the multi-variety enhancement graph to construct and calculate the market equilibrium constraint loss function, specifically including node features (such as the implicit inventory of each variety in each region, supply and demand status, etc.), edge features (such as the substitution relationship weight between varieties, logistics cost coefficient, etc.), and model weights (such as the internal parameters used by the neural network layer, fully connected layer, etc. when predicting the price sequence).

[0100] like Figure 5The graph shows a comparison of the normalized price index (NPI) predictions for corn. Control group A is a scheme that relies solely on historical market transaction data for trend judgment, without inverting the daily dynamic implicit inventory sequence, constructing a multi-variety weighted directed graph of production and consumption substitution relationships, and without constructing a market equilibrium constraint loss function. Control group B is a scheme that has inverted the daily dynamic implicit inventory sequence and the daily dynamic implicit inventory change rate based on multi-source heterogeneous datasets to form a multi-variety joint price prediction sequence, but has not used the market equilibrium constraint loss function for constraint optimization. As can be seen from the graph, the experimental group tracks the true normalized price index more closely to the actual change trajectory within the key turning point interval. In particular, the fitting of the turning point position, change magnitude, and fluctuation convergence process is more stable in the locally magnified region. This indicates that the present invention, by combining the characterization of substitution relationship transmission and the synergistic correction of supply and demand balance constraints, can effectively improve the price prediction accuracy, correlation consistency, and result interpretability.

[0101] Based on price forecasts, the magnitude, rate of change, and degree of fluctuation of prices for each agricultural product within the forecast period are calculated to form a price trend evolution description sequence.

[0102] The specific process includes: extracting the price forecast values ​​of agricultural products for two consecutive days within the forecast period from the price forecast values; calculating the ratio of the difference to the previous day's forecast value to obtain the price change range; dividing the price change range by the time interval to obtain the price change rate; calculating the arithmetic mean of the price change ranges for each day within the forecast period to obtain the mean; then taking the square root of the squared average of the differences between the daily ranges and the mean to obtain the price volatility; and arranging the daily price change ranges, rates, and volatility in chronological order to form a price trend evolution description sequence.

[0103] The expressions for calculating the magnitude, rate of change, and degree of fluctuation of prices for each agricultural product within the forecast period are as follows: ; ; ; ; in, Indicates agricultural product varieties On the date The range of price changes; Indicates an index of agricultural product varieties; Indicates the date index within the prediction period; Indicates agricultural product varieties On the date Price forecasts on the chart; Indicates agricultural product varieties On the date Price forecasts on the chart; Indicates agricultural product varieties On the date The rate of price change; Indicates agricultural product varieties On the date The degree of price fluctuation; Indicates the length of the statistical window for the degree of fluctuation; Indicates the backtracking step size index within the window; Indicates agricultural product varieties In the date For the current time point, looking back The price change range on the corresponding date for each time interval; Indicates agricultural product varieties In This represents the average price change within the statistical window at the current point in time.

[0104] It should be noted that the dimensions of this expression have been unified using price normalization before calculation; specifically, the magnitude of price change. In the calculation formula, the numerator is the difference between the price forecast at the current time and the price forecast at the previous time, and the denominator is the price forecast at the previous time. This method of calculating the relative rate of change is used to differentiate between different agricultural product varieties. The absolute price (usually in yuan / jin or yuan / ton) is converted into a dimensionless percentage or proportion value, thereby eliminating the dimensional influence caused by the difference in price base between different varieties, making the price change range, change speed and fluctuation degree of different varieties comparable.

[0105] It is formed by combining the dynamic changes of each agricultural product node with the relationships between nodes through joint recursive decoding.

[0106] Based on the price trend evolution description sequence, the trend status and key turning points of each agricultural product are identified. Combined with the daily dynamic implicit inventory change rate, substitution relationship margin weight and market equilibrium constraint deviation, trend interpretation information is formed, and a price trend forecast report for bulk agricultural products is generated according to the preset report structure.

[0107] The specific process includes: traversing the price trend evolution description sequence, identifying the dates when the rate of price change changes from positive to negative or vice versa, and marking them as key turning points from rising to falling and falling to rising, respectively; defining three trend states—rapid rise, rapid fall, and stable fluctuation—based on the numerical range of the rate of price change; extracting the daily dynamic implicit inventory change rate, comparing the direction of inventory changes before and after the key turning points, and determining the supporting role of inventory in price turning points; reading the substitution relationship edge weights of corresponding substitute products, and determining the impact of substitution transmission on price turning points; obtaining the market equilibrium constraint deviation, and determining the driving role of the easing of supply and demand imbalances in price turning points; and integrating the trend state, key turning points, and various influencing factors to form trend explanation information.

[0108] Furthermore, based on the pre-set report structure, the report begins with the forecast period, integrates price trends and status into the price operation analysis chapter, marks key turning points, and categorizes the daily dynamic implicit inventory change rate, substitution relationship weights, and market equilibrium constraint deviation analysis results into the driving factors chapter. The report also embeds trend explanation information as the core conclusion, generating a price trend forecast report for bulk agricultural products.

[0109] It should be noted that the report structure is pre-set in a logical order of forecast cycle, price movement analysis, key turning points, driving factors, and core conclusions.

[0110] In summary, this invention achieves continuous characterization of the implicit inventory status and inventory change trends of various related commodities by: daily collection of output, flow, accumulation, and processing / consumption information of related commodities at various regional nodes; and combined with edge point dual-domain inversion maps, the output-flow-accumulation mass conservation equation, and data collaborative processing supported by industrial cloud computing. This enhances the ability to identify supply-demand imbalances and improves the accuracy of trend inflection point judgment. Furthermore, by constructing a market equilibrium constraint loss function reflecting the cross-price elasticity of substitute crops, backpropagation optimization is performed on the multi-commodity joint price prediction sequence to obtain price prediction values. This achieves coordinated correction of substitution relationship transmission and supply-demand balance constraints, thereby improving price prediction accuracy, consistency, and interpretability.

[0111] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting the price trend of bulk agricultural products based on multi-source data, characterized in that, include: Collect multi-source heterogeneous basic data and perform spatiotemporal benchmark unification processing to generate multi-source heterogeneous datasets; Based on multi-source heterogeneous datasets, the output-flow-stacking mass conservation equation is used for inversion to obtain the daily dynamic implicit inventory sequence and daily dynamic implicit inventory change rate of each related commodity. Construct a multi-variety weighted directed graph of the production and consumption substitution relationship, and embed the daily dynamic implicit inventory sequence and the daily dynamic implicit inventory change rate into the feature vector of the corresponding agricultural product node after encoding, to generate multi-variety enhanced graph data; By using graph attention networks to aggregate cross-node association information of multi-variety enhanced graph data, a cross-node association information aggregation representation is formed. Combined with historical time series evolution patterns, the dynamic change characteristics of each agricultural product node are extracted to obtain a multi-variety joint price prediction sequence. Based on the multi-variety joint price forecast sequence, the parameters are optimized by backpropagation using the market equilibrium constraint loss function that reflects the cross price elasticity of substitute crops to obtain the price forecast value. Based on the final price forecast value, a price trend forecast report for major agricultural products is generated.

2. The method for predicting the price trend of bulk agricultural products based on multi-source data as described in claim 1, characterized in that, The multi-source heterogeneous basic data includes remote sensing image data of the main producing areas of the target bulk agricultural products and alternative crops, supply chain logistics trajectory data, and historical market transaction data.

3. The method for predicting the price trend of bulk agricultural products based on multi-source data as described in claim 1, characterized in that, The specific steps for generating the multi-source heterogeneous dataset are as follows: Collect multi-source heterogeneous basic data and encapsulate them in a structured manner according to source identifier, variety code, spatial identifier, observation time, unit of measurement and original observation value to obtain the original record set; The variety codes and spatial identifiers in the records from various sources are uniformly mapped using the original record set. The original observations are then subjected to dimensional unification, time alignment, anomaly correction, fusion completion, and structured organization to generate a multi-source heterogeneous dataset.

4. The method for predicting the price trend of bulk agricultural products based on multi-source data as described in claim 1, characterized in that, The specific steps for obtaining the daily dynamic implicit inventory sequence and daily dynamic implicit inventory change rate for each related commodity are as follows: The multi-source heterogeneous dataset is used to collect the daily output records, flow records, accumulation records and processing and consumption records of related varieties in each regional node, and to construct a two-domain inversion map of edge points. By using the edge point dual-domain inversion map, continuous event-driven reconstruction is performed on the discrete observation records of each related variety at each regional node, and the transport characteristics of the in-transit quality at the regional edge are performed to obtain a continuous transport field. Based on the continuous transport field, the output, inflow, outflow, accumulation and processing consumption of each node are substituted into the output-flow-accumulation mass conservation equation for joint inversion to obtain the daily dynamic implicit inventory sequence of each related commodity at each regional node. Based on the daily dynamic implicit inventory sequence, the daily dynamic implicit inventory change rate of each related product is calculated according to the inventory increase and decrease relationship between adjacent dates.

5. The method for predicting the price trend of bulk agricultural products based on multi-source data as described in claim 1, characterized in that, The specific steps for constructing the multi-variety weighted directed graph of production and consumption substitution relationships are as follows: Based on the daily dynamic implicit inventory sequence and the daily dynamic implicit inventory change rate, combined with the competition relationship of planting resources, the substitution relationship of processing formulas and the substitution relationship of consumption, candidate substitution relationships among varieties are screened to form candidate substitution units; Extract production substitution events that characterize the transfer of resources on the production side and consumption substitution events that characterize the migration of demand on the consumption side from the candidate substitution units to obtain a set of substitution events; Based on the set of substitution events, the daily directed edge weights between varieties are calculated according to the substitution direction, substitution intensity, and duration, forming a set of daily directed edges; The daily directed edge set is used to associate nodes and attach edge weights to each related variety, and a multi-variety weighted directed graph of production and consumption substitution relationships is constructed.

6. The method for predicting the price trend of bulk agricultural products based on multi-source data as described in claim 1, characterized in that, The specific steps for generating multi-variety enhanced atlas data are as follows: The daily dynamic implicit inventory sequence and daily dynamic implicit inventory change rate are extracted from the multi-variety weighted directed graph, and the node time sequence is organized by combining the node association edge information to obtain the node event package. The daily dynamic implicit inventory sequence and the daily dynamic implicit inventory change rate are jointly encoded using node event packages, and the node joint encoding vector is generated by combining the directional features and edge weight features of the node associated edges. The joint encoding vector of nodes is embedded into the feature vector of the corresponding agricultural product node, and then associated with the multi-variety weighted directed graph to generate multi-variety enhanced graph data.

7. The method for predicting the price trend of bulk agricultural products based on multi-source data as described in claim 1 or 6, characterized in that, The specific steps for forming the cross-node association information aggregation representation are as follows: Extract the node feature vectors of each agricultural product node and the edge weights and directions of directed edges between nodes from the multi-variety enhanced graph data, and construct the node directional neighborhood relationship. By utilizing the directional neighborhood relationship of nodes, the node feature vectors of each agricultural product node and the edge weights and directions of the corresponding directed edges are input into the graph attention network to calculate the attention weights between the node and its neighboring nodes. Attention weights are used to associate and aggregate the incoming and outgoing neighborhood information and node feature vectors of each agricultural product node, forming a cross-node association information aggregation representation.

8. The method for predicting the price trend of bulk agricultural products based on multi-source data as described in claim 1, characterized in that, The specific steps for obtaining the multi-commodity joint price prediction sequence are as follows: Extract the node feature vectors and node association edge information corresponding to each agricultural product node from the multi-variety enhanced graph data in historical time sequence, and construct the node historical evolution input sequence; The historical evolution input sequence of nodes is used to recursively calculate the historical time-series state of each agricultural product node, and to form dynamic change characteristics that characterize the state level, rate of change and turning trend of each agricultural product node. Based on dynamic changes, and combined with the correlation between various agricultural product nodes, a joint recursive decoding is performed to obtain a multi-variety joint price prediction sequence.

9. The method for predicting the price trend of bulk agricultural products based on multi-source data as described in claim 1 or 8, characterized in that, The specific steps to obtain the price prediction value are as follows: Based on the multi-variety joint price forecast series, and combined with the substitution relationship edge weights of each agricultural product, the daily dynamic implicit inventory series, and the daily dynamic implicit inventory change rate, a market equilibrium constraint loss function reflecting the cross price elasticity of substitute crops is constructed. The market equilibrium constraint loss function is used to map the multi-product joint price forecast series to a constrained price series that satisfies the supply and demand balance constraint and the substitution relationship constraint. By backpropagating the market equilibrium constraint loss function using the constrained price series, we obtain the gradient information corresponding to each parameter. Based on the gradient information, we iteratively update each parameter to form the price prediction value.

10. The method for predicting the price trend of bulk agricultural products based on multi-source data as described in claim 1, characterized in that, The specific steps for generating a price trend forecast report for major agricultural products are as follows: Based on price forecasts, the magnitude, rate of change, and degree of fluctuation of prices for each agricultural product within the forecast period are calculated to form a price trend evolution description sequence. Based on the price trend evolution description sequence, the trend status and key turning points of each agricultural product are identified. Combined with the daily dynamic implicit inventory change rate, substitution relationship margin weight and market equilibrium constraint deviation, trend interpretation information is formed, and a price trend forecast report for bulk agricultural products is generated according to the preset report structure.

Images