Real-time commodity sales prediction system and method based on edge calculation

Abstract

The invention relates to the technical field of big data, in particular to a real-time commodity sales prediction system and method based on edge calculation. The system comprises a data acquisition module, an edge calculation module, a commodity sales prediction module, an optimization feedback module and an intelligent scheduling module. The system collects sales, inventory, customer behaviors and environment data in real time through the Internet of Things, and utilizes a graph embedding technology to construct a dynamic relation graph and enhance features. Modeling is carried out based on a space-time diagram network algorithm, and accurate prediction of commodity sales volume is realized. Model parameters are optimized through reinforcement learning, inventory replenishment, a distribution path, a production schedule and a promotion strategy are dynamically adjusted in combination with a discrete differential evolution algorithm, and a real-time resource scheduling scheme is generated. According to the invention, the accuracy and real-time performance of commodity sales volume prediction are improved, optimal configuration of resources is realized through intelligent scheduling, and the operation efficiency and market competitiveness of enterprises are effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of big data technology, and in particular to a real-time commodity sales forecasting system and method based on edge computing. Background Art

[0002] With the rapid development of e-commerce, logistics distribution and retail industries, commodity sales forecasting has become one of the key technologies to improve business efficiency and resource utilization. Traditional commodity sales forecasting methods usually rely on large-scale cloud computing, and obtain forecast results by modeling and analyzing historical data, market trends, seasonal changes and other factors. However, there are still the following problems: the existing technology cannot fully capture the complex relationship between commodity sales and time, space and multiple factors, resulting in inaccurate forecast results; the existing system is difficult to dynamically adjust model parameters according to actual sales data, and the forecast accuracy decreases with environmental changes; the existing technology does not respond in time in inventory replenishment, distribution routes, production scheduling and other links, and the ability to respond to dynamic market demand is insufficient. The scheduling algorithm is inefficient, resulting in unreasonable resource allocation, high operating costs and missed market opportunities. Summary of the invention

[0003] To solve the above problems, the present invention provides a real-time commodity sales forecasting system and method based on edge computing, which solves the problem of how to fully capture the complex correlation between time, space and multiple factors in commodity sales forecasting, and dynamically adjust model parameters, while achieving efficient inventory replenishment, distribution route optimization and production scheduling, thereby improving forecasting accuracy and resource utilization.

[0004] To achieve the above object, the technical solution adopted by the present invention is: On the one hand, a real-time commodity sales prediction system based on edge computing includes a data acquisition module, an edge computing module, a commodity sales prediction module, an optimization feedback module and an intelligent scheduling module which are sequentially communicatively connected; The data collection module is used to collect commodity data in real time through IoT devices; the commodity data includes sales data, inventory data, customer behavior data and environmental data; The edge computing module is used to construct the commodity data into a dynamic relationship graph based on graph embedding technology, and perform feature enhancement processing on the vertices and edges of the graph using a dynamic weighting mechanism based on structural adaptation; The commodity sales forecasting module is used to extract and model the spatiotemporal features of the dynamic relationship graph based on the spatiotemporal graph network algorithm, generate a commodity sales forecasting model, and generate commodity sales forecasting results by capturing the complex dynamic associations between commodity sales and time, space and related factors; The optimization feedback module is used to dynamically adjust the parameters of the commodity sales forecasting model using a reinforcement learning algorithm according to the difference between the commodity sales forecast result and the actual sales data, optimize the forecast accuracy, and generate a commodity sales forecast analysis report; The intelligent scheduling module is used to optimize commodity inventory replenishment, distribution route optimization, production schedule adjustment and promotion activity strategy in real time according to the commodity sales forecast analysis report, and generate a resource scheduling report by using a dynamic resource scheduling method based on a discrete differential evolution algorithm.

[0005] Furthermore, the operation process of the edge computing module includes the following steps: Based on the commodity data, a dynamic relationship graph is constructed using graph embedding technology, where the nodes of the graph represent commodities, customers or environmental factors, and the edges represent the interaction relationship between entities; Through the structural adaptive dynamic weighting mechanism, the weights of the edges in the graph are dynamically adjusted by combining the topological characteristics of the nodes, the changing trends of real-time data, and the time dimension information; By using the dynamically weighted relationship graph, the high-order features of nodes and edges are integrated, and multi-dimensional features are extracted through the spatiotemporal attention mechanism to optimize the spatiotemporal expression ability of the graph. The dynamic relationship graph after feature enhancement is output to the product sales prediction module in real time.

[0006] Furthermore, the dynamic weighting mechanism specifically adopts commodity data and calculates the initial edge weight in combination with the attributes of the nodes. At the same time, based on the sales change trend, customer behavior temporal pattern and the dynamic influence of environmental factors, it introduces the time decay function and context-aware model to adaptively adjust the edge weight and generate a real-time optimized dynamic relationship graph.

[0007] Furthermore, the operation process of the commodity sales forecasting module includes the following steps: Receiving the dynamic relationship graph, and performing normalization and noise reduction processing on node features and edge features in the graph; Based on the spatiotemporal graph network algorithm, a multi-layer graph network model with time and space dimensions is constructed. The changing characteristics of nodes in different time periods are extracted through the temporal attention mechanism, and the interaction pattern of nodes in the graph structure is extracted through the spatial attention mechanism. Perform feature fusion processing on the extracted spatiotemporal features, construct a high-dimensional feature representation that includes the complex relationship between product sales and time, space and related factors, and generate a preliminary product sales prediction model; Use the rolling window mechanism to input real-time updated data, combine historical sales data and external environment data, and dynamically adjust model parameters; Based on the optimized product sales forecasting model, the input dynamic relationship graph is inferred and calculated to generate multi-dimensional product sales forecasting results including time series, regional distribution and product category dimensions, and the results are output to the optimization feedback module in real time.

[0008] Furthermore, the formula of the commodity sales prediction model is as follows: in, Indicates the product sales forecast result; The feature vector representing the node v at time t, i.e., the commodity data; Represents the set of neighbor nodes of node v; represents the time attention weight, which measures the impact of historical time t on current sales; represents the spatial attention weight, which measures the influence of neighbor node u on node v; The weight matrix representing the temporal features; The weight matrix representing the spatial features; Represents the final fused output weight matrix; Represents the feature vector of neighbor node u; Represents the output bias term of the final fusion; represents a nonlinear activation function; T represents the number of historical time points to be predicted.

[0009] Furthermore, the optimization feedback module adopts a dynamic signal sampling and weight updating mechanism to adjust the target weight of the reinforcement learning reward function according to different time periods, sales cycles and external environmental changes, and dynamically optimize the model parameters.

[0010] Furthermore, the operation process of the intelligent scheduling module includes the following steps: Based on the sales forecast analysis report output by the optimization feedback module, identify the resource scheduling requirements for inventory replenishment, delivery routes, production scheduling, and promotional activities; A dynamic resource scheduling optimization model is built based on discrete differential evolution algorithm, and a multi-objective optimization function is designed, including inventory minimization, distribution path optimization, production efficiency maximization, and promotion effect improvement. Use real-time environment data and resource status to generate the optimal resource scheduling strategy through iterative optimization; The optimized resource scheduling solution is generated into a report output, and a real-time dynamic resource scheduling optimization model is created based on execution feedback.

[0011] Furthermore, the formula of the discrete differential evolution algorithm is as follows: in, Represents the optimal resource scheduling strategy; represents the current resource scheduling scheme of the i-th individual in the t-th generation; and They represent the reference resource scheduling scheme individuals randomly selected from the population; F represents the differential variation factor; Represents the crossover factor, which is used to control the crossover probability of the newly generated scheme; represents the random disturbance.

[0012] On the other hand, a real-time commodity sales prediction method based on edge computing includes the following steps: Collect commodity data in real time through IoT devices; Based on graph embedding technology, the commodity data is constructed into a dynamic relationship graph, and a dynamic weighting mechanism based on structural adaptation is used to perform feature enhancement processing on the vertices and edges of the graph; Based on the spatiotemporal graph network algorithm, the spatiotemporal features of the dynamic relationship graph are extracted and modeled to generate a commodity sales forecasting model, and the commodity sales forecasting results are generated by capturing the complex dynamic relationship between commodity sales and time, space and related factors; According to the difference between the commodity sales forecast result and the actual sales data, a reinforcement learning algorithm is used to dynamically adjust the parameters of the commodity sales forecast model, optimize the forecast accuracy, and generate a commodity sales forecast analysis report; According to the commodity sales forecast analysis report, a dynamic resource scheduling method based on discrete differential evolution algorithm is adopted to optimize commodity inventory replenishment, distribution route optimization, production schedule adjustment and promotion activity strategy in real time, and generate a resource scheduling report.

[0013] The beneficial effects of the present invention are: The data acquisition module of the present invention collects multi-dimensional commodity data in real time through the Internet of Things devices, including sales data, inventory data, customer behavior data and environmental data, ensuring the real-time and comprehensiveness of the data, and providing accurate basic data support for subsequent prediction and optimization. The edge computing module uses graph embedding technology to construct the collected data into a dynamic relationship graph, and enhances the feature representation ability of graph vertices and edges through a structurally adaptive dynamic weighting mechanism, thereby improving the accuracy and effectiveness of data modeling. The commodity sales forecasting module is based on the spatiotemporal graph network algorithm, which fully explores the complex dynamic relationship between commodity sales and time, space and related factors, realizes accurate prediction of commodity sales, and helps enterprises better grasp market trends and customer needs. The optimization feedback module dynamically adjusts the model parameters based on the difference between actual sales data and prediction results through a reinforcement learning algorithm, continuously optimizes the prediction accuracy, thereby improving the stability and prediction reliability of the system, and generating an in-depth commodity sales forecast analysis report. The intelligent scheduling module is based on the discrete differential evolution algorithm to realize dynamic resource scheduling for commodity inventory replenishment, distribution path optimization, production schedule adjustment and promotion activity strategy. By optimizing resource allocation and operation strategies, the supply chain efficiency of the enterprise is significantly improved and the operating costs are reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a module schematic diagram of a real-time commodity sales prediction system based on edge computing of the present invention.

[0015] Figure 2 It is a flowchart of the operation process of the commodity sales forecasting module provided by one embodiment of the present invention.

[0016] Figure 3 It is a flow chart of a real-time commodity sales prediction method based on edge computing of the present invention. DETAILED DESCRIPTION

[0017] See also Figure 1-3 As shown, the present invention relates to a real-time commodity sales prediction system and method based on edge computing.

[0018] Example 1 A real-time commodity sales prediction system based on edge computing, comprising a data acquisition module, an edge computing module, a commodity sales prediction module, an optimization feedback module and an intelligent scheduling module which are sequentially connected in communication; The data collection module is used to collect commodity data in real time through IoT devices; the commodity data includes sales data, inventory data, customer behavior data and environmental data; It should be noted that IoT devices collect multi-dimensional commodity data in real time, including: Sales data: By connecting to the cash register system, POS machine and e-commerce platform interface, real-time commodity sales records are obtained, including sales volume, sales time, price, etc.

[0019] Inventory data: Through RFID tags, barcode scanners and warehouse management systems, the inventory quantity, in and out time and inventory turnover rate of goods can be obtained in real time.

[0020] Customer behavior data: Use cameras, heat map analysis tools and customer shopping trajectory analysis systems to collect customer behavior data in the store, such as dwell time, number of product views, purchase conversion rate, etc.

[0021] Environmental data: Deployed environmental sensors collect real-time temperature and humidity, light intensity, passenger flow density and other external environmental information.

[0022] Perform preliminary data filtering and formatting through edge devices, such as deleting abnormal data and standardizing data formats, and then upload the cleaned data to the edge computing module.

[0023] The edge computing module is used to construct the commodity data into a dynamic relationship graph based on graph embedding technology, and perform feature enhancement processing on the vertices and edges of the graph using a dynamic weighting mechanism based on structural adaptation; The operation process of the edge computing module includes the following steps: Based on the commodity data, a dynamic relationship graph is constructed using graph embedding technology, where the nodes of the graph represent commodities, customers or environmental factors, and the edges represent the interaction relationship between entities; It should be noted that the definitions of nodes and edges of the dynamic relationship graph are as follows: Node: Map products, customers, and environmental factors (such as weather, holidays, etc.) to nodes in the graph. Each node contains attribute information, such as: Product nodes: price, inventory level, historical sales.

[0024] Customer nodes: purchase preferences, behavioral data (such as touch frequency and dwell time).

[0025] Environmental factor nodes: weather conditions, promotional activities.

[0026] Edge: represents the interaction between entities, such as: Item-item edge: joint sales relationship (such as the frequency with which item A and item B are purchased together).

[0027] Product-customer edge: the frequency of interaction between a customer and a product.

[0028] Product-environmental factor edge: the impact of certain weather or activities on product sales.

[0029] Using graph embedding technology (such as Node2Vec, DeepWalk), the high-dimensional attribute information of the node is mapped to a low-dimensional vector to maintain the structure and semantic information of the node. Based on the product sales data, the initial weight of the edge is calculated, such as the statistical value of joint sales or the strength of the association between customer behavior and products.

[0030] Through the structural adaptive dynamic weighting mechanism, the weights of the edges in the graph are dynamically adjusted by combining the topological characteristics of the nodes, the changing trends of real-time data, and the time dimension information; Specifically, we calculate the topological characteristic indicators of the nodes, such as degree centrality, betweenness centrality, and clustering coefficient. We adjust the edge weights according to the local structure of the graph (such as triangle relationships). For example, if products A and B frequently interact with the same customer, the edge weights will increase.

[0031] According to the changing trend of real-time data, the edge weights are adjusted: Sales change trend: When product sales increase suddenly, increase the weight of the product-product edge to strengthen the impact of joint sales.

[0032] Customer behavior temporal pattern: adjust the weight of the product-customer edge according to the customer's behavior (such as visit frequency).

[0033] Introduce time decay function: By setting the decay parameter, the impact of historical data on edge weights gradually weakens over time, ensuring that the model pays more attention to recent data. Assign weight proportions to short-term changes and long-term trends respectively. Build a context-aware model to adjust edge weights based on node attributes and external environmental factors. For example: during holidays, increase the product-customer edge weight to reflect the impact of promotional activities; when the weather changes, strengthen the connection strength between the weather node and the related product nodes.

[0034] By using the dynamically weighted relationship graph, the high-order features of nodes and edges are integrated, and multi-dimensional features are extracted through the spatiotemporal attention mechanism to optimize the spatiotemporal expression ability of the graph. It should be noted that the spatiotemporal attention mechanism is used to capture the multi-dimensional interaction characteristics of the time dimension and the space dimension: Temporal attention: weights the data of historical time steps to highlight the features of key time points.

[0035] Spatial attention: Dynamically adjust the interaction weights between different nodes to highlight the products, customers or environmental factors that have the greatest impact on sales.

[0036] The dynamic relationship graph after feature enhancement is output to the product sales prediction module in real time.

[0037] The dynamic weighting mechanism specifically uses commodity data and calculates the initial edge weight in combination with the attributes of the nodes. At the same time, based on the sales change trend, customer behavior temporal pattern and the dynamic influence of environmental factors, it introduces the time decay function and the context perception model to adaptively adjust the edge weight and generate a real-time optimized dynamic relationship graph.

[0038] Furthermore, the initial edge weight calculation formula is as follows: in, represents edge weight; represents the joint sales of products i and j; Indicates the frequency of customer behavior towards a product.

[0039] The commodity sales forecasting module is used to extract and model the spatiotemporal features of the dynamic relationship graph based on the spatiotemporal graph network algorithm, generate a commodity sales forecasting model, and generate commodity sales forecasting results by capturing the complex dynamic associations between commodity sales and time, space and related factors; The operation process of the commodity sales forecasting module includes the following steps: Receiving the dynamic relationship graph, and performing normalization and noise reduction processing on node features and edge features in the graph; Based on the spatiotemporal graph network algorithm, a multi-layer graph network model with time and space dimensions is constructed. The changing characteristics of nodes in different time periods are extracted through the temporal attention mechanism, and the interaction pattern of nodes in the graph structure is extracted through the spatial attention mechanism. Specifically, the time dimension and the space dimension are combined to construct a graph network containing multiple levels; the time dimension is used to capture the pattern of changes in product sales over time; the space dimension is used to analyze the association between products and the interaction between customers and products.

[0040] Use time series modeling techniques to analyze the changing patterns of product sales and identify key time points or periodic patterns (such as peak and trough periods). Divide time periods (such as hours, days, or weeks) to refine time analysis and highlight the differences in characteristics of different time periods. Use graph neural networks to extract the interactive features of products, customers, and environmental factors in the relationship graph. Analyze the joint sales relationship between products or the commonality of customer behavior in a region to explore potential association patterns.

[0041] Among them, the temporal attention mechanism focuses on the time points in historical data that have a significant impact on sales, such as sales changes at the beginning of a promotion.

[0042] Spatial attention mechanism: dynamically adjust the weights between nodes and edges to strengthen the influence of key products and customer nodes.

[0043] By setting up a multi-layer structure, higher-order spatiotemporal features are extracted layer by layer to capture complex correlation patterns. The output of each layer is used as the input of the next layer, gradually optimizing the expressive power of the graph.

[0044] Perform feature fusion processing on the extracted spatiotemporal features, construct a high-dimensional feature representation that includes the complex relationship between product sales and time, space and related factors, and generate a preliminary product sales prediction model; Use the rolling window mechanism to input real-time updated data, combine historical sales data and external environment data, and dynamically adjust model parameters; Specifically, continuously receive real-time updated data, including the latest sales data, inventory data, customer behavior data, and environmental factor data. Set up a rolling window mechanism to dynamically input new time period data into the model, while retaining certain historical data for contextual analysis. Retrain some model parameters according to changes in real-time data, such as updating weights or adjusting the length of the time window. During peak hours or emergencies (such as weather changes), the model will give priority to short-term real-time data. Use incremental learning methods to quickly adjust some parameters without retraining the entire model. Regularly calibrate the model to ensure the accuracy and real-time nature of the prediction results.

[0045] Based on the optimized product sales forecasting model, the input dynamic relationship graph is inferred and calculated to generate multi-dimensional product sales forecasting results including time series, regional distribution and product category dimensions, and the results are output to the optimization feedback module in real time.

[0046] Furthermore, the formula of the commodity sales prediction model is as follows: in, Represents the sales forecast results of products, that is, the sales values ​​in terms of time series, regional distribution and product category; The feature vector representing the node v at time t, i.e., the commodity data; Represents the set of neighbor nodes of node v, such as other products, customers, or environmental factors associated with product v; represents the time attention weight, which measures the impact of historical time t on current sales; represents the spatial attention weight, which measures the influence of neighbor node u on node v; The weight matrix representing the temporal features; The weight matrix representing the spatial features; Represents the final fused output weight matrix, which is used to map temporal and spatial features into sales forecast results; Represents the feature vector of neighbor node u; Represents the output bias term of the final fusion; represents a nonlinear activation function; T represents the number of historical time points to be predicted.

[0047] The optimization feedback module is used to dynamically adjust the parameters of the commodity sales prediction model using a reinforcement learning algorithm according to the difference between the commodity sales prediction results and the actual sales data, optimize the prediction accuracy, and generate a commodity sales prediction analysis report; the optimization feedback module adopts a dynamic signal sampling and weight update mechanism to adjust the target weight of the reinforcement learning reward function according to different time periods, sales cycles and external environment changes, and dynamically optimize the model parameters.

[0048] Specifically, the error is calculated using the historical actual sales data (obtained from the data collection module) and the current product sales forecast results. A variety of error evaluation indicators are used: Mean Squared Error (MSE): Measures the average of the squared errors between the predicted values ​​and the actual values, suitable for overall evaluation.

[0049] Mean absolute error (MAE): avoids over-amplifying extreme value errors and is closer to actual business needs.

[0050] Weighted error evaluation: Set different weights for different time periods (such as peak sales period, off-season) or specific products to highlight the impact of errors in key scenarios on model optimization.

[0051] The error is used as the environmental feedback signal of reinforcement learning to provide input basis for model optimization.

[0052] Combined with the error sources (such as model underfitting, overfitting or external factors), a multi-dimensional optimization feedback report is generated.

[0053] It should be noted that the method based on deep reinforcement learning, such as Deep Q Network (DQN), Policy Gradient or Proximal Policy Optimization (PPO), is selected to optimize the accuracy of sales forecast results by adjusting the prediction model parameters (such as embedding dimension, learning rate, regularization weight).

[0054] Adjust the weight of the reward function for different time periods (such as holidays and promotional periods) to enhance the accuracy of sales forecasts for key time periods. Add weights of environmental factors (such as weather and policy changes) to the reward function to ensure the robustness of the model to environmental fluctuations. Combine the accuracy of short-term forecasts and the effects of long-term trend forecasts to dynamically adjust the reward target to avoid the model focusing too much on a certain time scale.

[0055] Strategy update and parameter optimization: After each round of prediction error feedback, reinforcement learning is used to update the strategy.

[0056] Strategy update process: Use the error feedback collected in real time as the current state; use the reinforcement learning algorithm to generate a new model parameter adjustment strategy; update the parameters of the prediction model by executing the strategy; verify the performance of the updated model again and record the improvement effect.

[0057] The dynamic signal sampling mechanism classifies external signals (such as weather, holidays, social activities) and internal signals (such as sales fluctuations, insufficient inventory). The impact of the signals is graded, and the signal weights are dynamically adjusted according to their historical relevance. A dynamic sampling mechanism is implemented to increase the sampling frequency when the signal fluctuates greatly or the prediction error increases significantly. Combined with an event-driven sampling mechanism, for example, when abnormal sales data (such as a sudden increase or decrease) is detected, new data is immediately collected and the optimization process is triggered.

[0058] The product sales forecast analysis report includes multi-dimensional error analysis results, model optimization history records, reinforcement learning reward function change trends, and final optimized parameter adjustment details. Dynamically generate visual reports, such as error change trend charts, reward trend charts during optimization, and sales forecast comparison heat maps. Provide interactive dashboards to support managers to select specific time periods or regions to view optimization effects.

[0059] The intelligent scheduling module is used to optimize commodity inventory replenishment, distribution route optimization, production schedule adjustment and promotion activity strategy in real time according to the commodity sales forecast analysis report, and generate a resource scheduling report by using a dynamic resource scheduling method based on a discrete differential evolution algorithm.

[0060] The operation process of the intelligent scheduling module includes the following steps: Based on the sales forecast analysis report output by the optimization feedback module, identify the resource scheduling requirements for inventory replenishment, delivery routes, production scheduling, and promotional activities; It should be noted that the product sales forecast analysis report output by the receiving optimization feedback module includes: product sales forecast data for different time periods and regions, error analysis between actual sales data and forecast data, and forecast of the impact of environmental factors (such as weather and traffic) on sales and logistics.

[0061] According to the analysis report, the following four types of resource scheduling requirements are extracted: Inventory replenishment needs: Analyze whether each inventory node (store or warehouse) has insufficient or excessive inventory.

[0062] Delivery route needs: Identify whether the delivery route needs to be optimized, such as a surge in regional sales or the risk of delivery delays.

[0063] Production scheduling requirements: Adjust production plans based on predicted demand peaks to avoid overcapacity or undercapacity.

[0064] Promotional needs: Identify which products require additional promotional support to increase sales.

[0065] Prioritize scheduling tasks based on the urgency and impact of demand. For example, insufficient inventory takes precedence over promotion strategy adjustments; delivery optimization in high-sales areas takes precedence over low-sales areas.

[0066] A dynamic resource scheduling optimization model is built based on discrete differential evolution algorithm, and a multi-objective optimization function is designed, including inventory minimization, distribution path optimization, production efficiency maximization, and promotion effect improvement. Specifically, an initial population is generated, and each individual represents a resource scheduling scheme, including inventory replenishment plan, distribution route, production schedule and promotion strategy.

[0067] Represent the scenarios as discrete variables (e.g., replenishment quantities, delivery sequences, production time periods, etc.).

[0068] The objective optimization function is designed as follows: Inventory minimization goal: reduce inventory holding costs while ensuring that inventory meets sales demand.

[0069] The goal of distribution route optimization is to shorten the total distribution distance and time and reduce transportation costs.

[0070] The goal of maximizing production efficiency is to maximize equipment utilization and reduce production plan adjustment costs.

[0071] Promotion effect improvement goal: Increase sales of low-selling products and increase profit margins through promotions.

[0072] The above objective functions are combined into a weighted multi-objective optimization model, and the weights are adjusted dynamically: when inventory is insufficient, the weight of the inventory replenishment goal is increased; when the distribution pressure is high, the weight of the route optimization is increased.

[0073] Use real-time environment data and resource status to generate the optimal resource scheduling strategy through iterative optimization; The optimized resource scheduling solution is generated into a report output, and a real-time dynamic resource scheduling optimization model is created based on execution feedback.

[0074] Furthermore, the formula of the discrete differential evolution algorithm is as follows: in, Represents the optimal resource scheduling strategy; represents the current resource scheduling scheme of the i-th individual in the t-th generation; and They represent the reference resource scheduling scheme individuals randomly selected from the population; F represents the differential variation factor, which is used to control the differential amplitude and adjust the variation intensity; Represents the crossover factor, which is used to control the crossover probability of the newly generated scheme; represents the random disturbance.

[0075] Example 2 A real-time commodity sales prediction method based on edge computing includes the following steps: Collect commodity data in real time through IoT devices; Based on graph embedding technology, the commodity data is constructed into a dynamic relationship graph, and a dynamic weighting mechanism based on structural adaptation is used to perform feature enhancement processing on the vertices and edges of the graph; Based on the spatiotemporal graph network algorithm, the spatiotemporal features of the dynamic relationship graph are extracted and modeled to generate a commodity sales forecasting model, and the commodity sales forecasting results are generated by capturing the complex dynamic relationship between commodity sales and time, space and related factors; According to the difference between the commodity sales forecast result and the actual sales data, a reinforcement learning algorithm is used to dynamically adjust the parameters of the commodity sales forecast model, optimize the forecast accuracy, and generate a commodity sales forecast analysis report; According to the commodity sales forecast analysis report, a dynamic resource scheduling method based on discrete differential evolution algorithm is adopted to optimize commodity inventory replenishment, distribution route optimization, production schedule adjustment and promotion activity strategy in real time, and generate a resource scheduling report.

[0076] In this embodiment, a real-time commodity sales prediction method based on edge computing is applied to a real-time commodity sales prediction system based on edge computing described in Example 1, which will not be repeated here.

[0077] In summary, the present invention collects multi-dimensional data such as sales, inventory, customer behavior and environment in real time through IoT devices, and uses edge computing modules to preprocess and analyze data to ensure the timeliness and efficiency of data, realize real-time prediction and dynamic optimization of commodity sales, and significantly improve the response speed of enterprises to market changes. The system uses graph embedding technology to construct a dynamic relationship graph, fully integrates multi-dimensional data such as commodities, customers and environment, dynamically models the complex association relationship between nodes, and uses the space-time graph network algorithm to further extract the interactive features of time and space dimensions, thereby capturing the complex dynamic association between commodity sales and related factors, and improving the accuracy and robustness of the prediction model.

[0078] The present invention introduces a structurally adaptive dynamic weighting mechanism and a time decay function, and dynamically optimizes the edge weights in the graph according to the changing trend of real-time data and the weight adjustment of historical data. The model parameters are iteratively optimized through the reinforcement learning algorithm, which effectively improves the accuracy of the prediction results while maintaining the ability to balance short-term changes and long-term trends. The system combines the sales forecast results and adopts a dynamic resource scheduling method based on a discrete differential evolution algorithm to perform multi-objective optimization for inventory replenishment, distribution routes, production scheduling, and promotion strategies. While meeting market demand, it minimizes inventory costs, transportation costs, and waste of production resources, and achieves efficient use of resources throughout the entire process.

[0079] The present invention continuously receives and dynamically adjusts model parameters through a rolling window mechanism and an incremental learning method, and has the ability to flexibly respond to emergencies (such as promotional activities, weather changes, etc.) and long-term trends (such as holiday sales peaks). The optimization feedback module can adjust the model strategy in real time according to environmental changes to ensure that the system always maintains good performance in a complex and changing business environment. By generating commodity sales forecast analysis reports and resource scheduling reports, it provides a visual display of error analysis, model optimization history, and resource scheduling strategies, providing managers with a clear and intuitive reference for decision-making. At the same time, the interactive dashboard function supports key data analysis for specific time periods or regions, enhancing the business application value of the system.

[0080] The above implementation modes are merely descriptions of the preferred implementation modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary engineering and technical personnel in the field shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A real-time commodity sales prediction system based on edge computing, characterized in that: It includes a data acquisition module, an edge computing module, a commodity sales forecasting module, an optimization feedback module and an intelligent scheduling module which are sequentially connected in communication; The data collection module is used to collect commodity data in real time through IoT devices; the commodity data includes sales data, inventory data, customer behavior data and environmental data; The edge computing module is used to construct the commodity data into a dynamic relationship graph based on graph embedding technology, and perform feature enhancement processing on the vertices and edges of the graph using a dynamic weighting mechanism based on structural adaptation; The commodity sales forecasting module is used to extract and model the spatiotemporal features of the dynamic relationship graph based on the spatiotemporal graph network algorithm, generate a commodity sales forecasting model, and generate commodity sales forecasting results by capturing the complex dynamic associations between commodity sales and time, space and related factors; The optimization feedback module is used to dynamically adjust the parameters of the commodity sales forecasting model using a reinforcement learning algorithm according to the difference between the commodity sales forecast result and the actual sales data, optimize the forecast accuracy, and generate a commodity sales forecast analysis report; The intelligent scheduling module is used to optimize commodity inventory replenishment, distribution route optimization, production schedule adjustment and promotion activity strategy in real time according to the commodity sales forecast analysis report, and generate a resource scheduling report by using a dynamic resource scheduling method based on a discrete differential evolution algorithm.

2. A real-time commodity sales forecasting system based on edge computing according to claim 1, characterized in that: The operation process of the edge computing module includes the following steps: Based on the commodity data, a dynamic relationship graph is constructed using graph embedding technology, where the nodes of the graph represent commodities, customers or environmental factors, and the edges represent the interaction relationship between entities; Through the structural adaptive dynamic weighting mechanism, the weights of the edges in the graph are dynamically adjusted by combining the topological characteristics of the nodes, the changing trends of real-time data, and the time dimension information; By using the dynamically weighted relationship graph, the high-order features of nodes and edges are integrated, and multi-dimensional features are extracted through the spatiotemporal attention mechanism to optimize the spatiotemporal expression ability of the graph. The dynamic relationship graph after feature enhancement is output to the product sales prediction module in real time.

3. A real-time commodity sales forecasting system based on edge computing according to claim 2, characterized in that: The dynamic weighting mechanism specifically uses commodity data and calculates the initial edge weight in combination with the attributes of the nodes. At the same time, based on the sales change trend, customer behavior temporal pattern and the dynamic influence of environmental factors, it introduces the time decay function and the context perception model to adaptively adjust the edge weight and generate a real-time optimized dynamic relationship graph.

4. The real-time commodity sales forecasting system based on edge computing according to claim 1 is characterized in that: The operation process of the commodity sales forecasting module includes the following steps: Receiving the dynamic relationship graph, and performing normalization and noise reduction processing on node features and edge features in the graph; Based on the spatiotemporal graph network algorithm, a multi-layer graph network model with time and space dimensions is constructed. The changing characteristics of nodes in different time periods are extracted through the temporal attention mechanism, and the interaction pattern of nodes in the graph structure is extracted through the spatial attention mechanism. Perform feature fusion processing on the extracted spatiotemporal features, construct a high-dimensional feature representation that includes the complex relationship between product sales and time, space and related factors, and generate a preliminary product sales prediction model; Use the rolling window mechanism to input real-time updated data, combine historical sales data and external environment data, and dynamically adjust model parameters; Based on the optimized product sales forecasting model, the input dynamic relationship graph is inferred and calculated to generate multi-dimensional product sales forecasting results including time series, regional distribution and product category dimensions, and the results are output to the optimization feedback module in real time.

5. A real-time commodity sales forecasting system based on edge computing according to claim 4, characterized in that: The formula of the commodity sales forecasting model is as follows: in, Indicates the product sales forecast result; The feature vector representing the node v at time t, i.e., the commodity data; Represents the set of neighbor nodes of node v; represents the time attention weight, which measures the impact of historical time t on current sales; represents the spatial attention weight, which measures the influence of neighbor node u on node v; The weight matrix representing the temporal features; The weight matrix representing the spatial features; Represents the final fused output weight matrix; Represents the feature vector of neighbor node u; Represents the output bias term of the final fusion; represents a nonlinear activation function; T represents the number of historical time points to be predicted.

6. A real-time commodity sales forecasting system based on edge computing according to claim 1, characterized in that: The optimization feedback module adopts a dynamic signal sampling and weight updating mechanism to adjust the target weight of the reinforcement learning reward function according to different time periods, sales cycles and external environmental changes, and dynamically optimize the model parameters.

7. The real-time commodity sales forecasting system based on edge computing according to claim 1 is characterized in that: The operation process of the intelligent scheduling module includes the following steps: Based on the sales forecast analysis report output by the optimization feedback module, identify the resource scheduling requirements for inventory replenishment, delivery routes, production scheduling, and promotional activities; A dynamic resource scheduling optimization model is built based on discrete differential evolution algorithm, and a multi-objective optimization function is designed, including inventory minimization, distribution path optimization, production efficiency maximization, and promotion effect improvement. Use real-time environment data and resource status to generate the optimal resource scheduling strategy through iterative optimization; The optimized resource scheduling solution is generated into a report output, and a real-time dynamic resource scheduling optimization model is created based on execution feedback.

8. The real-time commodity sales forecasting system based on edge computing according to claim 7 is characterized in that: The formula of the discrete differential evolution algorithm is as follows: in, Represents the optimal resource scheduling strategy; represents the current resource scheduling scheme of the i-th individual in the t-th generation; and They represent the reference resource scheduling scheme individuals randomly selected from the population; F represents the differential variation factor; Represents the crossover factor, which is used to control the crossover probability of the newly generated scheme; represents the random disturbance.

9. A real-time commodity sales prediction method based on edge computing, characterized in that: The method is applied to a real-time commodity sales forecasting system based on edge computing as described in any one of claims 1 to 8, comprising the following steps: Collect commodity data in real time through IoT devices; Based on graph embedding technology, the commodity data is constructed into a dynamic relationship graph, and a dynamic weighting mechanism based on structural adaptation is used to perform feature enhancement processing on the vertices and edges of the graph; Based on the spatiotemporal graph network algorithm, the spatiotemporal features of the dynamic relationship graph are extracted and modeled to generate a commodity sales forecasting model, and the commodity sales forecasting results are generated by capturing the complex dynamic relationship between commodity sales and time, space and related factors; According to the difference between the commodity sales forecast result and the actual sales data, a reinforcement learning algorithm is used to dynamically adjust the parameters of the commodity sales forecast model, optimize the forecast accuracy, and generate a commodity sales forecast analysis report; According to the commodity sales forecast analysis report, a dynamic resource scheduling method based on discrete differential evolution algorithm is adopted to optimize commodity inventory replenishment, distribution route optimization, production schedule adjustment and promotion activity strategy in real time, and generate a resource scheduling report.

Images