Comprehensive energy cost management method and system for green low-carbon supply chain

Abstract

The invention relates to the field of data processing, in particular to a comprehensive energy cost management method and system for a green low-carbon supply chain. According to the technical scheme provided by the invention, aiming at the characteristics of a complex chain structure and energy flow and fund flow of a supply chain, association of nodes and edges is modeled by using a graph neural network, accurate accounting and management of full-chain energy cost are realized, energy market real-time data is introduced, an original data set is dynamically optimized, the difference between historical data and a current market environment is eliminated, and the real-time calculation of the energy market is realized. And the decision timeliness and accuracy are ensured. The invention provides a dynamic and multi-dimensional supply chain energy cost management scheme integrating a graph neural network and energy field specific constraint.

Description

Technical Field

[0001] This invention relates to the field of data processing, and more specifically, to a comprehensive energy cost management method and system for a green and low-carbon supply chain. Background Technology

[0002] Driven by global "dual carbon" goals, green and low-carbon supply chains have become a core competitive advantage for enterprises' sustainable development. With tightening policies such as carbon tariffs and ESG disclosure requirements, energy efficiency and carbon footprint management within the supply chain are becoming increasingly critical. Currently, the low-carbon transformation of the entire chain, from raw material procurement to end-user distribution, is evolving from a voluntary corporate behavior to an industry-wide mandatory requirement, driving energy cost management towards greater precision and dynamism.

[0003] However, enterprises face a dual challenge in green supply chain management: on the one hand, the supply chain has many nodes and complex links, and energy consumption is significantly affected by factors such as market price fluctuations and seasonal supply and demand changes. Traditional static accounting methods struggle to capture dynamic correlations, leading to significant deviations in energy cost calculations. On the other hand, the comprehensive decision-making regarding energy consumption and cost expenditures lacks a systematic approach, and carbon costs and energy costs are often managed separately, failing to support enterprises in balancing low-carbon transformation and cost optimization. Therefore, a dynamic, multi-dimensional, and intelligent supply chain energy cost management system is needed to provide a systematic solution to these problems. Summary of the Invention

[0004] The purpose of this invention is to provide a dynamic, multi-dimensional supply chain energy cost management solution that integrates graph neural networks with energy-specific constraints.

[0005] According to a first aspect of the present invention, a comprehensive energy cost management method for a green and low-carbon supply chain is proposed, comprising the following steps: S1. Obtain a comprehensive dataset D1 of the green and low-carbon supply chain. The comprehensive dataset D1 includes basic information of entities in each link of the supply chain, energy consumption data, expenditure data, and transportation characteristic data. S2. Dynamically optimize the comprehensive dataset D1 to generate the first optimized dataset D2; S3. Construct an energy graph neural network model based on the first optimized dataset D2; S4. Train the energy graph neural network model based on the first optimized dataset D2; S5. Output energy cost prediction indicators for entities in the green and low-carbon supply chain based on the trained energy graph neural network model; S6. Generate energy structure adjustment plans and energy procurement combination plans based on energy cost forecast indicators.

[0006] According to some embodiments, in the method of the first aspect of the present invention, step S2. dynamically optimizing the comprehensive dataset D1 to generate a first optimized dataset D2 includes: Access the real-time energy system market and obtain real-time energy market data; The comprehensive dataset D1 is corrected based on real-time energy market data, following the formula below: in, Market sensitivity coefficient To synthesize the data in dataset D1, For real-time energy market data and historical market data The difference, This refers to the data in the first optimized dataset D2.

[0007] According to some embodiments, in the method of the first aspect of the present invention, step S3. constructing an energy graph neural network model based on a first optimized dataset D1 includes: Data modeling and graph structure construction are performed based on the first optimized dataset D1, where entities are defined as nodes and the energy flow relationship between nodes is defined as edges. Based on the constructed graph structure, an energy neural network model is built that includes a physical constraint embedding layer, a multi-energy coupled attention layer, and an output layer.

[0008] According to some embodiments, in the method of the first aspect of the present invention, the physical constraint embedding layer includes graph structure constraints generated based on the physical rules of the energy system, and the implementation process includes: Obtain the physical formulas corresponding to the energy type; Obtain the physical parameters of the transmission medium used for energy transfer between nodes; The energy flow relationship between nodes is corrected based on physical formulas and parameters to ensure that the constructed energy neural network model conforms to physical laws.

[0009] According to some embodiments, in the method of the first aspect of the present invention, in the step of data modeling and graph structure construction based on the first optimized dataset D2, an energy cost transmission mechanism is introduced to enable the transmission of energy costs between the nodes, wherein the update formula of the cost transmission function is: in, Let N(i) be the energy cost of node i at the t-th iteration, including energy consumption and energy expenditure; N(i) is the set of adjacent nodes of node i. It is the cost propagation weight from node j to node i.

[0010] According to some embodiments, in the method of the first aspect of the present invention, the implementation of the multi-energy coupled attention layer includes: Construct heterogeneous graph structures based on the interactions between various types of energy; Dynamic weights are assigned based on attention mechanisms across energy types to quantify the cost transmission differences of different energy conversion pathways.

[0011] According to some embodiments, in the method of the first aspect of the present invention, step S4. training the energy graph neural network model based on the first optimized dataset D2 includes: A spatiotemporal dynamic data augmentation mechanism is used to process the first optimized dataset D2 to obtain the second optimized dataset D3. The spatiotemporal dynamic data augmentation mechanism combines the historical energy consumption time series of nodes with their geographical location to capture the seasonal and regional differences in energy consumption. Based on the second optimized dataset D3, the energy graph neural network model was trained by combining the loss function and optimization algorithm.

[0012] According to some embodiments, in the method of the first aspect of the present invention, the spatiotemporal dynamic data enhancement mechanism is implemented through the following steps: LSTM is used to process the energy consumption data sequence of node v over T historical time periods. Perform temporal feature encoding to obtain the encoded temporal features. ; The geographic distance dvu between node v and its neighbor u is obtained based on a geographic information system, and the distance decay coefficient is calculated. Follow the formula below: Where γ is the decay factor; the time series characteristics and distance attenuation coefficient The node features are updated by incorporating edge weights, following the formula: Among them, W and This is the weight matrix. This is the activation function.

[0013] According to some embodiments, in the method of the first aspect of the present invention, the energy cost prediction indicators include node-level indicators, edge-level indicators, and graph-level indicators; the node-level indicators include node energy consumption, node energy expenditure, and node carbon emission value; the edge-level indicators include the cost transmission coefficient value of energy transmission path and the path congestion risk index; the graph-level indicators include the total energy cost prediction value of the entire supply chain, the prediction error confidence interval, and the carbon emission reduction target achievement rate.

[0014] According to a second aspect of the present invention, a comprehensive energy cost management system for a green and low-carbon supply chain is proposed for implementing the method described in the first aspect of the present invention, comprising: The data acquisition module is used to acquire the comprehensive dataset D1 of the green and low-carbon supply chain. The comprehensive dataset D1 includes basic information of entities in each link of the supply chain, energy consumption data, expenditure data, and transportation characteristic data. The data optimization module is used to dynamically optimize the comprehensive dataset D1 to generate the first optimized dataset D2; The model building module is used to build an energy graph neural network model based on the first optimized dataset D2; The model training module is used to train the energy graph neural network model based on the first optimized dataset D2; The indicator generation module is used to output energy cost prediction indicators for entities in the green and low-carbon supply chain based on the trained energy graph neural network model. The decision generation module is used to generate energy structure adjustment plans and energy procurement combination plans based on energy cost forecast indicators.

[0015] The solution proposed in this invention has the following beneficial effects: 1. In response to the complex chain structure in the green and low-carbon supply chain, as well as the characteristics of energy flow and capital flow, graph neural networks are used to model the relationship between nodes and edges to achieve accurate accounting and management of energy costs across the entire chain.

[0016] 2. Given the high volatility of the energy market, real-time energy market data is introduced to dynamically optimize the original dataset, eliminate the discrepancy between historical data and the current market environment, and ensure the timeliness and accuracy of decision-making.

[0017] 3. Incorporate physical rules to construct a constraint layer to ensure that energy transmission and cost transmission conform to actual physical laws and improve the physical feasibility of model predictions.

[0018] 4. The multi-energy coupling attention layer accurately associates electricity, gas, heat and other types, dynamically quantifies the cost transmission of conversion paths, and optimizes comprehensive energy cost decisions.

[0019] 5. The spatiotemporal dynamic data enhancement mechanism integrates temporal and geographical features to capture seasonal and regional differences in energy consumption, thereby improving the model's adaptability to dynamic changes. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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 exceeding the scope of protection claimed by the present invention.

[0021] Figure 1 This is a flowchart illustrating an embodiment 1000 of the integrated energy cost management method for a green and low-carbon supply chain according to the present invention. Figure 2 for Figure 1 A flowchart illustrating the detailed steps S2 of Example 1000; Figure 3 for Figure 1 A flowchart illustrating the detailed steps S3 of Example 1000; Figure 4 for Figure 3 A flowchart illustrating the implementation process S32A of the physical constraint embedding layer in step S32; Figure 5 for Figure 3 A flowchart illustrating the implementation process of the multi-energy coupled attention layer in step S32B; Figure 6 for Figure 1 A flowchart illustrating the detailed steps S4 of Example 1000; Figure 7 for Figure 6 A flowchart illustrating the implementation process of the spatiotemporal dynamic data augmentation mechanism S42A in step S4; Figure 8 This is a schematic diagram of the results of an embodiment 2000 of the integrated energy cost management system for a green and low-carbon supply chain according to the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] See Figure 1 , Figure 1 This is a flowchart illustrating an embodiment 1000 of the integrated energy cost management method for a green and low-carbon supply chain according to the present invention. Figure 1 As shown, Example 1000 includes steps S1-S6.

[0024] In some specific embodiments, in step S1, a comprehensive energy cost management system (such as a processor, hereinafter referred to as a processor) for a green and low-carbon supply chain acquires a comprehensive dataset D1 of the green and low-carbon supply chain. The comprehensive dataset D1 includes basic information of entities in each link of the supply chain, energy consumption data, expenditure data, and transportation characteristic data.

[0025] In some specific embodiments, in step S2, the processor dynamically optimizes the comprehensive dataset D1 to generate a first optimized dataset D2.

[0026] In some specific embodiments, in step S2, the processor accesses the real-time energy system market and obtains real-time energy market data. Then, it corrects the comprehensive dataset D1 based on the real-time energy market data to make the dataset more closely reflect the impact of market price fluctuations on actual energy consumption, thus obtaining the optimized first optimized dataset D2.

[0027] In some specific embodiments, in step S3, an energy graph neural network model is constructed based on the first optimized dataset D2.

[0028] In some specific embodiments, in step S3, the processor performs data modeling and graph structure construction based on the first optimized dataset D1, defining entities as nodes and energy flow relationships between nodes as edges. Optionally, in the data modeling and graph structure construction based on the first optimized dataset D2, the processor introduces an energy cost transmission mechanism, enabling energy costs to be transferred between nodes.

[0029] In some specific embodiments, in step S3, the processor constructs an energy neural network model based on the constructed graph structure, including a physical constraint embedding layer, a multi-energy coupling attention layer, and an output layer. The physical constraint embedding layer ensures that the constructed energy neural network model conforms to physical laws, and the multi-energy coupling attention layer ensures that in a complex supply chain containing multiple energy types, the model can accurately capture the cost transmission differences of multi-energy coupling.

[0030] In some specific embodiments, in step S4, the energy graph neural network model is trained based on the first optimized dataset D2.

[0031] In some specific embodiments, in step S4, the processor uses a spatiotemporal dynamic data augmentation mechanism to process the first optimized dataset D2 to obtain the second optimized dataset D3; wherein, the spatiotemporal dynamic data augmentation mechanism combines the historical energy consumption time series of nodes with geographical location to capture the seasonal and regional differences in energy consumption.

[0032] In some specific embodiments, in step S4, the processor trains the energy graph neural network model based on the second optimization dataset D3, combining a loss function and an optimization algorithm. The process includes: designing a loss function; using a combined loss function to balance multiple task objectives; and iteratively updating parameters using an optimization algorithm. Optionally, the optimization algorithm includes the Adam optimizer and an early stopping strategy.

[0033] In some specific embodiments, in step S5, the energy cost prediction index of entities in the green and low-carbon supply chain is output based on the trained energy graph neural network model.

[0034] In some specific embodiments, in step S5, the energy cost prediction indicators include node-level indicators, edge-level indicators, and graph-level indicators. Optionally, node-level indicators include node energy consumption, node energy expenditure, and node carbon emissions, wherein the node-level indicators are derived with a preset time window as the granularity. For example, the output node-level indicators are derived with daily, weekly, and monthly granularities, respectively.

[0035] Optionally, edge-level indicators include the cost transmission coefficient of energy transmission paths and the path congestion risk index. Optionally, graph-level indicators include the predicted total energy cost of the entire supply chain, the prediction error confidence interval, and the carbon emission reduction target achievement rate. In step S5, all energy cost prediction indicators are calculated through the forward propagation of the model and validated using the historical data optimized in step S2.

[0036] In some specific embodiments, in step S6, an energy structure adjustment plan and an energy procurement combination plan are generated based on energy cost forecast indicators.

[0037] In some specific embodiments, in step S6, the processor, based on a long-term forecast objective (such as annual energy cost trends), solves for the energy structure adjustment scheme using a multi-objective optimization algorithm to minimize the total cost function, as shown in the formula: Where x represents the energy structure adjustment vector (such as the proportion of photovoltaic installed capacity). This is the carbon price coefficient. Let R(x) be the risk preference coefficient, and R(x) be the investment risk function.

[0038] In some specific embodiments, in step S6, the processor optimizes the energy procurement mix based on medium-term forecasting indicators (such as monthly procurement cycles) to minimize procurement costs, using the following formula: in, Let i be the quantity of energy purchased. For real-time prices, To predict demand, This is the penalty coefficient for supply and demand.

[0039] According to such Figure 1 The embodiments shown in this invention propose a technical solution based on the specific field of supply chain energy cost accounting. Targeting the complex chain structure in green and low-carbon supply chains, as well as the characteristics of energy and capital flows, the technical solution uses the association relationships between nodes and edges in a graph neural network to accurately model the energy flow relationships between various entities in the supply chain and between entities, thereby achieving accurate energy cost accounting and management across the entire chain.

[0040] Figure 2 for Figure 1 A flowchart illustrating the detailed steps S2 of Example 1000. (See attached diagram.) Figure 2As shown, step S2 includes steps S21-S22.

[0041] In some specific embodiments, in step S21, the processor accesses the real-time energy system market and acquires real-time energy market data. In some specific embodiments, the real-time energy market data includes spot electricity prices (such as the regional marginal price (LMP) updated every 15 minutes), instant natural gas trading prices, coal port settlement prices, and real-time renewable energy subsidy amounts, collected at minute / hourly frequencies via real-time energy system APIs (such as the European Powernext and the Guangzhou Power Exchange Center interface). Optionally, real-time price data is bound to supply chain nodes; for example, a manufacturer node is bound to the real-time electricity price of its region, and a logistics company node is bound to the real-time retail price of diesel in the area covered by its transportation routes.

[0042] In some specific embodiments, in step S22, the processor corrects the comprehensive dataset D1 based on real-time energy market data, following the formula: in, Market sensitivity coefficient To synthesize the data in dataset D1, For real-time energy market data and historical market data The difference, The data in the first optimized dataset D2 is used for optimization. In step S2, real-time energy market data is introduced to make the first optimized dataset D2 more closely reflect the impact of market price fluctuations on actual energy consumption.

[0043] According to such Figure 2 The implementation method shown in this invention introduces real-time energy market data, upgrading the original dataset from static historical data to dynamic market-aware data. This enables the energy graph neural network model to learn long-term patterns and respond to short-term market changes, ultimately improving the flexibility and economy of energy cost management in the supply chain.

[0044] Figure 3 for Figure 1 A flowchart illustrating the detailed step S3 of Example 1000. (See attached diagram.) Figure 3 As shown, step S3 includes steps S31-S32.

[0045] In some specific embodiments, in step S31, the processor performs data modeling and graph structure construction based on the first optimized dataset D1, defining entities as nodes and energy flow relationships between nodes as edges.

[0046] In some specific embodiments, the entities in the supply chain are the actual enterprises participating in each link of the supply chain, such as raw material suppliers, manufacturers, logistics companies, and distributors. The node attributes include basic enterprise information, energy consumption data, cost data, and technical parameters. Basic enterprise information includes enterprise production capacity and size; energy consumption data includes the consumption of different types of energy, including electricity, gas, and oil; technical parameters include equipment energy efficiency and production processes. In some specific embodiments, each node includes a unique ID used for identification throughout the entire process.

[0047] In some specific embodiments, edges represent energy flow relationships between nodes. For example, an edge is defined as a raw material transportation link from a supplier to a manufacturer, or an edge is defined as a finished product logistics link from a manufacturer to a warehouse. Edge characteristics include transportation distance, energy type, transportation volume, and unit energy consumption coefficient. In some specific embodiments, edges represent energy procurement relationships between nodes. For example, an edge is defined as the purchase price from a raw material supplier to a manufacturer.

[0048] In step S31, during data modeling and graph structure construction based on the first optimized dataset D2, an energy cost transmission mechanism is introduced to allow energy costs to be transferred between nodes. The update formula for the cost transmission function is as follows: in, Let N(i) be the energy cost of node i at the t-th iteration, including energy consumption and energy expenditure; N(i) is the set of adjacent nodes of node i. It is the cost propagation weight from node j to node i.

[0049] In some specific embodiments, in step S32, the processor constructs an energy neural network model based on the constructed graph structure, which includes a physical constraint embedding layer, a multi-energy coupled attention layer, and an output layer.

[0050] In some specific embodiments, in step S32, the processor obtains the physical formula corresponding to the energy type, obtains the physical parameters of the transmission medium for energy transmission between nodes, and corrects the energy flow relationship between nodes according to the physical formula and physical parameters to ensure that the constructed energy neural network model conforms to the physical laws.

[0051] In some specific embodiments, in step S32, the processor constructs a heterogeneous graph structure based on the interaction between multiple different types of energy, and then assigns dynamic weights according to the attention mechanism across energy types to quantify the cost transmission differences of different energy conversion paths.

[0052] Figure 4 for Figure 3A flowchart illustrating the implementation process S32A of the physical constraint embedding layer in step S32. (See attached diagram.) Figure 4 As shown, the implementation process S32A includes steps S321-S323.

[0053] In some specific embodiments, the physical constraint embedding layer in step S32 includes graph structure constraints generated based on the physical rules of the energy system, and its implementation process S32A includes: In step S321, the processor obtains the physical formula corresponding to the energy type. In some specific embodiments, in step S321, the processor selects applicable physical laws and further determines the physical formula for the specific scenario of energy transmission and conversion in the supply chain. For example, optionally, the physical laws include the law of conservation of energy and the transmission limits of different energy types (such as the maximum transmission capacity of power lines).

[0054] In step S322, the processor acquires the physical parameters of the transmission medium for energy transfer between nodes. In some specific embodiments, in step S322, the processor converts physical laws into computable parameters and associates these parameters with the edge or node attributes of the graph structure. In step S323, the processor corrects the energy flow relationship between nodes based on the physical formulas and physical parameters to ensure that the constructed energy neural network model conforms to the physical laws.

[0055] The following is an example of implementing power transmission between power grid nodes in process S32A: In Example 1, the physical rule is that the transmission power of the line must not exceed its rated capacity. The processor obtains the rated transmission capacity of each transmission line. (Unit: MW), used as a characteristic parameter of the edge. For example, the characteristic parameter of a certain line. =500MW. When node v receives a power message from its neighbor node u. At that time, messages exceeding the rated transmission capacity are truncated using the following formula: For example, if the transmission power is 600MW, exceeding If the capacity is 500MW, the actual message transmission is corrected to 500MW to avoid physically infeasible overcapacity transmission.

[0056] The following is a second embodiment of natural gas pipeline transportation that implements process S32A: In Example 2, the physical rule is Poiseuille's law of fluid dynamics, specifically including that the pipe flow rate is directly proportional to the pressure difference and inversely proportional to the pipe length. The processor calculates the maximum allowable flow rate of the pipe. Where k is the pipeline coefficient. Let L be the pressure difference between the two ends, and L be the pipe length. For example, a certain pipe... In the natural gas cost transmission message, the cost coefficient λ related to the limited flow rate satisfies... ,in, The unit price of natural gas is set to ensure that the cost calculated based on flow does not exceed the maximum value corresponding to the physical limit.

[0057] The third embodiment of energy conversion within the factory that achieves process S32A is as follows: In Example 3, the physical rule is the first law of thermodynamics, i.e., energy conversion efficiency <= 100%. The processor determines the maximum efficiency of the electrothermal conversion device. This is used as an attribute parameter of the node. For example, the electric boiler's... =90%. When calculating the energy conversion message from the power node to the heat node, the converted heat message Q is constrained to satisfy Q<= ,in For input electrical energy. For example, when 100 kWh of electrical energy is input, the maximum heat information is 90 kWh, to avoid unreasonable predictions of an increase in energy value out of thin air.

[0058] According to such Figure 4 The implementation method shown in this invention incorporates physical rule constraints to ensure that energy transmission and cost transmission conform to actual laws, avoid predictions beyond physical limitations, and improve model feasibility.

[0059] Figure 5 for Figure 3 A flowchart illustrating the implementation process of the multi-energy coupled attention layer in step S32, S32B. (See attached flowchart for example.) Figure 5 As shown, the implementation process S32B includes steps S324-S325.

[0060] In some specific embodiments, in step S324, the processor constructs a heterogeneous graph structure based on the interaction between multiple different types of energy.

[0061] For example, in some specific embodiments, the energy types include electrical energy, gas energy, and thermal energy, and constructing the heterogeneous graph structure includes: the processor dividing nodes of different energy types into sets of child nodes. , , The energy conversion and transmission relationships between nodes are defined as heterogeneous edges, and edge characteristics include energy conversion efficiency and unit cost; the processor performs a unified dimensional mapping on the characteristics of various types of nodes, including electrical node characteristics. through Mapping, gas node features through Mapping, hot node features through Mapping yields feature vectors of the same dimension.

[0062] In some specific embodiments, in step S325, the processor assigns dynamic weights according to an attention mechanism across energy types to quantify the cost transmission differences of different energy conversion paths.

[0063] For example, in some specific embodiments, the processor computes the attention weights between node v and its neighbors u of different types. The formula is: In step S325, the processor aggregates neighbor features according to weights to update node features, enabling the model to accurately capture the cost transmission differences of multi-energy coupling. The update formula for node features is: According to such Figure 5 The implementation method shown in this invention enables the model to accurately capture and dynamically quantify the cost transmission differences of multiple energy couplings by coupling multiple energy types, accurately associate different energy types such as electricity, gas, and heat, and optimize comprehensive energy cost decision-making.

[0064] Figure 6 for Figure 1 A flowchart illustrating the detailed step S4 of Example 1000. (See attached diagram.) Figure 6 As shown, step S4 includes steps S41-S42.

[0065] In some specific embodiments, in step S41, the processor uses a spatiotemporal dynamic data augmentation mechanism to process the first optimized dataset D2 to obtain the second optimized dataset D3; wherein, the spatiotemporal dynamic data augmentation mechanism combines the historical energy consumption time series of nodes with geographical location to capture the seasonal and regional differences in energy consumption.

[0066] In some specific embodiments, in step S41, the processor uses LSTM to process the energy consumption data sequence of node v over T historical time periods. Perform temporal feature encoding to obtain the encoded temporal features. In some specific embodiments, in step S41, the processor obtains the geographical distance dvu between node v and its neighbor node u based on the geographic information system, and calculates the distance attenuation coefficient. In some specific embodiments, in step S41, the processor will determine the timing characteristics. and distance attenuation coefficient The node features are updated by incorporating edge weights.

[0067] In some specific embodiments, in step S42, the processor trains the energy graph neural network model based on the second optimized dataset D3, combining a loss function and an optimization algorithm. The specific process includes: S421, Design the loss function, using a combined loss function to balance multiple task objectives. In step S42, considering both energy consumption and carbon emission indicators, the designed loss function is as follows: Where MSE is the mean square error between the predicted value y' and the actual value y, in order to optimize energy cost prediction; CE is the carbon emission classification loss, in order to optimize low-carbon targets; For model parameters The regularization term is used to prevent overfitting; α, β, These are the weighting coefficients.

[0068] S422, the parameters are iteratively updated using an optimization algorithm. Optionally, the optimization algorithm includes the Adam optimizer and an early stopping strategy. In some specific embodiments, the Adam optimizer is used to iteratively update the parameters: an initial learning rate is set, which decays by 10% every 5 epochs; the number of training epochs is controlled by an early stopping strategy, for example, if the validation set loss does not decrease for 10 consecutive epochs, the iteration stops; projected gradient descent is used for the parameters of the physically constrained embedding layer to ensure that hard constraints such as physical laws are met, and finally, converged model parameters are obtained. '.

[0069] S423, Model Performance Verification. In some specific embodiments, 30% of the data is allocated as a test set. The node-level prediction accuracy and graph-level cost error rate are calculated and compared with preset target values ​​to ensure that the model meets the accuracy requirements of supply chain energy cost management. For example, in some specific embodiments, if the node-level prediction accuracy is >= 85% and the graph-level cost error rate is <= 5%, the model is considered to meet the requirements.

[0070] Figure 7 for Figure 6 A flowchart illustrating the implementation process of the spatiotemporal dynamic data augmentation mechanism in step S4, S41A. (See attached flowchart.) Figure 7 As shown, the implementation process S41A includes steps S411-S413.

[0071] In some specific embodiments, in step S411, the processor uses LSTM to process the energy consumption data sequence of node v over T historical time periods. Perform temporal feature encoding to obtain the encoded temporal features. Among them, LSTM (Long Short-Term Memory) is a type of recurrent neural network suitable for processing and predicting events with long intervals and delays in time series. In some specific embodiments, the input data in the time series feature encoding is the energy consumption data sequence of each node within a certain continuous historical period (e.g., 3-5 years). The time granularity is accurate to the hour, and T is the total time count (e.g., if the historical period is 3 years, T is 26280 hours). In step S411, the input gate, forget gate, cell state, output gate, and hidden state in LSTM are used for encoding, and the temporal features of the node energy consumption sequence are extracted layer by layer to capture the intraday peaks and troughs, seasonal fluctuations, and other patterns in the historical data. In step S411, the temporal features... The edge weights in the energy graph neural network model are modified to reflect the impact of time factors on energy transmission efficiency. For example, cargo transportation and storage during extreme weather (rain, snow, or temperatures exceeding a certain range) require more energy and incur higher costs.

[0072] In some specific embodiments, in step S412, the processor obtains the geographical distance dvu between node v and its neighbor node u based on the geographic information system, and calculates the distance attenuation coefficient. Follow the formula below: Where γ is the attenuation factor. Optionally, γ varies depending on the industry to which the node belongs; for example, γ=0.01 for the logistics industry and γ=0.005 for the power transmission industry. In step S412, the distance attenuation coefficient... The edge weights in the energy graph neural network model are modified to reflect the impact of geographical distance on energy transmission efficiency. For example, the energy consumed in long-distance transportation increases with distance.

[0073] In some specific embodiments, in step S412, the processor obtains the latitude and longitude information of node v and neighboring node u based on the geographic information system, and then calculates the straight-line distance dvu between the two based on the latitude and longitude information, in kilometers.

[0074] In some specific embodiments, in step S413, the processor will determine the timing characteristics. and distance attenuation coefficient The node features are updated by incorporating edge weights, following the formula: Among them, W and This is the weight matrix. This is the activation function. In some specific embodiments, in step S413, the processor fuses temporal trends and spatial correlations to improve prediction accuracy for different scenarios.

[0075] According to such Figure 7 The embodiments shown in this invention demonstrate that the technical solution proposed in this invention integrates temporal and geographical features through a spatiotemporal dynamic data enhancement mechanism to capture seasonal and regional differences in energy consumption and improve the model's adaptability to dynamic changes.

[0076] Figure 8 This is a schematic diagram of an embodiment 2000 of a comprehensive energy cost management system for a green and low-carbon supply chain according to the present invention. Figure 8As shown, embodiment 2000 includes a data acquisition module 201, a data optimization module 202, a model building module 203, a model training module 204, an indicator generation module 205, and a decision generation module 206.

[0077] In some specific embodiments, the data acquisition module 201 acquires a comprehensive dataset D1 of the green and low-carbon supply chain. The comprehensive dataset D1 includes basic information of entities in each link of the supply chain, energy consumption data, expenditure data, and transportation characteristic data.

[0078] In some specific embodiments, the data optimization module 202 dynamically optimizes the comprehensive dataset D1 to generate a first optimized dataset D2.

[0079] In some specific embodiments, the data optimization module 202 accesses the real-time energy system market and obtains real-time energy market data. Then, it corrects the comprehensive dataset D1 based on the real-time energy market data, making the dataset more consistent with the impact of market price fluctuations on actual energy consumption, thus obtaining the optimized first optimized dataset D2.

[0080] In some specific embodiments, the model building module 203 builds an energy graph neural network model based on the first optimized dataset D2.

[0081] In some specific embodiments, the model building module 203 performs data modeling and graph structure construction based on the first optimized dataset D1, defining entities as nodes and energy flow relationships between nodes as edges. Optionally, in the data modeling and graph structure construction based on the first optimized dataset D2, the model building module 203 introduces an energy cost transmission mechanism, enabling energy costs to be transferred between nodes.

[0082] In some specific embodiments, the model building module 203 constructs an energy neural network model based on the constructed graph structure, including a physical constraint embedding layer, a multi-energy coupling attention layer, and an output layer. The physical constraint embedding layer ensures that the constructed energy neural network model conforms to physical laws, while the multi-energy coupling attention layer ensures that in a complex supply chain containing multiple energy types, the model can accurately capture the cost transmission differences of multi-energy coupling.

[0083] In some specific embodiments, the model training module 204 trains the energy graph neural network model based on the first optimized dataset D2.

[0084] In some specific embodiments, the model training module 204 uses a spatiotemporal dynamic data augmentation mechanism to process the first optimized dataset D2 to obtain the second optimized dataset D3; wherein, the spatiotemporal dynamic data augmentation mechanism combines the historical energy consumption time series of nodes with their geographical location to capture the seasonal and regional differences in energy consumption.

[0085] In some specific embodiments, the model training module 204 trains the energy graph neural network model based on the second optimization dataset D3, combining a loss function and an optimization algorithm. The process includes: designing a loss function, using a combined loss function to balance multiple task objectives, and iteratively updating parameters using an optimization algorithm. Optionally, the optimization algorithm includes the Adam optimizer and an early stopping strategy.

[0086] In some specific embodiments, the indicator generation module 205 outputs energy cost prediction indicators for entities in the green and low-carbon supply chain based on the trained energy graph neural network model.

[0087] In some specific embodiments, energy cost prediction indicators include node-level indicators, edge-level indicators, and graph-level indicators. Optionally, node-level indicators include node energy consumption, node energy expenditure, and node carbon emissions, wherein the node-level indicators are derived with a preset time window as the granularity. For example, the node-level indicators output by the indicator generation module 205 are derived with daily, weekly, and monthly granularities, respectively.

[0088] Optionally, edge-level indicators include the cost transmission coefficient of energy transmission paths and the path congestion risk index. Optionally, graph-level indicators include the predicted total energy cost of the entire supply chain, the prediction error confidence interval, and the carbon emission reduction target achievement rate. All energy cost prediction indicators are calculated through the forward propagation of the model and validated using historical data optimized in the data optimization module 202.

[0089] In some specific embodiments, the decision generation module 206 generates energy structure adjustment schemes and energy procurement combination schemes based on energy cost forecast indicators.

[0090] In some specific embodiments, the decision generation module 206, based on long-term forecast objectives (such as annual energy cost trends), solves for energy structure adjustment schemes through a multi-objective optimization algorithm to minimize the total cost function, as shown in the formula: Where x represents the energy structure adjustment vector (such as the proportion of photovoltaic installed capacity). This is the carbon price coefficient. Let R(x) be the risk preference coefficient, and R(x) be the investment risk function.

[0091] In some specific embodiments, the decision generation module 206 optimizes the energy procurement combination scheme and minimizes procurement costs based on medium-term forecast indicators (such as monthly procurement cycles), using the following formula: in, Let i be the quantity of energy purchased. For real-time prices, To predict demand, This is the penalty coefficient for supply and demand.

[0092] The embodiments of the present invention have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, any changes or modifications made by those skilled in the art based on the ideas of the present invention, its specific implementation methods, and its application scope, are all within the scope of protection of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A comprehensive energy cost management method for a green and low-carbon supply chain, characterized in that, include: S1. Obtain a comprehensive dataset D1 of the green and low-carbon supply chain, wherein the comprehensive dataset D1 includes basic information of entities in each link of the supply chain, energy consumption data, expenditure data, and transportation characteristic data; S2. Dynamically optimize the comprehensive dataset D1 to generate a first optimized dataset D2; S3. Construct an energy graph neural network model based on the first optimized dataset D2; S4. Train the energy graph neural network model based on the first optimized dataset D2; S5. Output the energy cost prediction index of the entity in the green and low-carbon supply chain based on the trained energy graph neural network model; S6. Generate an energy structure adjustment plan and an energy procurement combination plan based on the energy cost forecast indicators.

2. The method according to claim 1, characterized in that, Step S2. Dynamically optimize the comprehensive dataset D1 to generate a first optimized dataset D2, including: Access the real-time energy system market and obtain real-time energy market data; The comprehensive dataset D1 is corrected based on the real-time energy market data, following the formula: ； in, Market sensitivity coefficient The data in the comprehensive dataset D1, For the aforementioned real-time energy market data and historical market data The difference, The data is from the first optimized dataset D2 after optimization.

3. The integrated energy cost management method for a green and low-carbon supply chain according to claim 1, characterized in that, Step S3, constructing an energy graph neural network model based on the first optimized dataset D2, includes: Data modeling and graph structure construction are performed based on the first optimized dataset D2, where the entities are defined as nodes and the energy flow relationships between the nodes are defined as edges. Based on the constructed graph structure, an energy neural network model is built that includes a physical constraint embedding layer, a multi-energy coupled attention layer, and an output layer.

4. The integrated energy cost management method for a green and low-carbon supply chain according to claim 3, characterized in that, The physical constraint embedding layer includes graph structure constraints generated based on the physical rules of the energy system, and the implementation process includes: Obtain the physical formulas corresponding to the energy type; Obtain the physical parameters of the transmission medium used for energy transfer between the nodes; The energy flow relationship between the nodes is corrected according to the physical formula and the physical parameters to ensure that the constructed energy neural network model conforms to the physical laws.

5. The method according to claim 3, characterized in that, In the step of data modeling and graph structure construction based on the first optimized dataset D2, an energy cost transmission mechanism is introduced to allow energy costs to be transferred between the nodes. The update formula for the cost transmission function is as follows: ； in, Let N(i) be the energy cost of node i at the t-th iteration, including energy consumption and energy expenditure; N(i) is the set of adjacent nodes of node i. It is the cost propagation weight from node j to node i.

6. The method according to claim 5, characterized in that, The implementation of the multi-energy coupled attention layer includes: Construct heterogeneous graph structures based on the interactions between various types of energy; Dynamic weights are assigned based on attention mechanisms across energy types to quantify the cost transmission differences of different energy conversion pathways.

7. The method according to claim 1, characterized in that, Step S4. Training the energy graph neural network model based on the first optimized dataset D2 includes: The first optimized dataset D2 is processed using a spatiotemporal dynamic data augmentation mechanism to obtain the second optimized dataset D3; wherein, the spatiotemporal dynamic data augmentation mechanism combines the historical energy consumption time series and geographical location of the nodes to capture the seasonal and regional differences in energy consumption; Based on the second optimized dataset D3, the energy graph neural network model is trained using a loss function and an optimization algorithm.

8. The method according to claim 7, characterized in that, The spatiotemporal dynamic data enhancement mechanism is implemented through the following steps: LSTM is used to process the energy consumption data sequence of node v over T historical time periods. Perform temporal feature encoding to obtain the encoded temporal features. ; The geographic distance dvu between node v and its neighbor node u is obtained based on a geographic information system, and the distance attenuation coefficient is calculated. Follow the formula below: ; Where γ is the attenuation factor; The time series features and the distance attenuation coefficient The features of the nodes are updated by incorporating the edge weights, following the formula: ; Among them, W and This is the weight matrix. This is the activation function.

9. The method according to claim 1, characterized in that, The energy cost forecasting indicators include node-level indicators, edge-level indicators, and graph-level indicators. The node-level indicators include node energy consumption, node energy expenditure, and node carbon emissions. The edge-level indicators include the cost transmission coefficient of energy transmission paths and the path congestion risk index. The graph-level indicators include the total energy cost forecast of the entire supply chain, the forecast error confidence interval, and the carbon emission reduction target achievement rate.

10. A comprehensive energy cost management system for a green and low-carbon supply chain, used to implement the methods of claims 1-9 above, characterized in that, include: The data acquisition module is used to acquire a comprehensive dataset D1 of the green and low-carbon supply chain. The comprehensive dataset D1 includes basic information of entities in each link of the supply chain, energy consumption data, expenditure data, and transportation characteristic data. The data optimization module is used to dynamically optimize the comprehensive dataset D1 to generate a first optimized dataset D2; The model building module is used to build an energy graph neural network model based on the first optimized dataset D2; The model training module is used to train the energy graph neural network model based on the first optimized dataset D2; The indicator generation module is used to output the energy cost prediction indicators of the entities in the green and low-carbon supply chain based on the trained energy graph neural network model. The decision generation module is used to generate energy structure adjustment schemes and energy procurement combination schemes based on the energy cost forecast indicators.

Images