A four-stream coupling full life cycle evaluation method for resource processing of crop and livestock waste based on a graph neural network

Abstract

The present application relates to livestock manure and crop straw resource processing technology field, specifically provides a kind of four-flow coupling full life cycle evaluation method of species cultivation waste resource processing based on graph neural network, comprising: obtaining four-flow real-time working condition data of system state, construct heterogeneous process atlas;Topology generation network is constructed, and a group of candidate process topology adjacency matrix is generated;The topological total expected value, path robustness score and reconfiguration conversion cost corresponding to each candidate process topology adjacency matrix are calculated;Obtain final decision score, select the highest final decision score candidate process topology adjacency matrix as optimal process topology, output the adaptive execution strategy corresponding to the optimal process topology.This method breaks through the limitation that traditional fixed process flow can only carry out parameter optimization, realizes the dynamic reconfiguration of whole process topology structure.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of waste resource processing, and particularly discloses a four-stream coupling full life cycle evaluation method for seed and livestock waste resource processing based on a graph neural network. BACKGROUND

[0002] In the current field of seed and livestock waste resource processing, the core goal is to convert organic waste such as livestock manure and crop straw into energy or high-value chemicals. To achieve this goal, multiple processing units such as anaerobic digestion, biological desulfurization, methane reforming, and methanol synthesis are usually combined to form a complete process flow. The performance of this flow needs to be evaluated comprehensively from four dimensions: material flow, energy flow, environmental flow, and economic flow, involving multiple complex indicators such as material conversion efficiency, energy transfer efficiency, environmental load, and operating benefits. Traditional process design relies on single-dimensional, fixed, and static flow topology, i.e., once the combination and connection mode of processing units are determined, they rarely change.

[0003] In existing technologies, optimization methods mainly focus on parameter tuning based on a given process flow, such as adjusting internal operating parameters such as temperature, pressure, or flow rate of a certain reaction unit. However, this approach has fundamental limitations, as it cannot reconstruct the process topology itself in response to changes in external conditions. In real production, system input conditions are dynamic and variable, including fluctuations in raw material properties, changes in market signals, and shifts in equipment internal conditions. When these conditions change significantly, the original fixed flow may no longer be the optimal choice, and may even become inefficient or uneconomical.

[0004] Existing technologies generally lack an intelligent mechanism that can respond to real-time data-driven, autonomously generate and evaluate new process paths. They cannot deeply correlate and uniformly model multi-dimensional input data such as raw materials, markets, and operating conditions with the topology of the process flow, which limits the system's ability to fundamentally change the flow path to avoid negative impacts or seize market opportunities, such as activating an ammonia capture unit to deal with high-nitrogen raw materials or switching to a biomass gasification path to quickly respond to methanol market demand. This parameter optimization approach greatly limits the system's potential to maximize full life cycle value in complex and dynamic environments.

[0005] Therefore, how to provide a method that can generate, evaluate, and select the topology of the processing technology based on real-time, multi-dimensional input data, achieve a leap from parameter optimization to process reconstruction, and enable the system to dynamically adapt and always approach the globally optimal operating state, is a technical problem that needs to be solved by those skilled in the art.

[0006] The above information disclosed in the BACKGROUND section is only for enhancing the understanding of the background of the present disclosure, and therefore it can include information that does not constitute the prior art known to those of ordinary skill in the art. SUMMARY

[0007] The purpose of the present application is to provide a kind of four-flow coupling full life cycle evaluation method for resource processing of seed-raising waste based on graph neural network, to solve the problems raised in the above background.

[0008] The technical scheme of the present application is as follows:

[0009] A kind of four-flow coupling full life cycle evaluation method for resource processing of seed-raising waste based on graph neural network, comprising:

[0010] Step one, obtain the multi-dimensional real-time working condition data representing system state, the multi-dimensional real-time working condition data includes raw material characteristic vector, market signal vector and internal working condition vector, and construct the heterogeneous process graph including the preset processing unit as node and potential process path as edge ; the quantitative value of full life cycle material flow, energy flow, environmental flow and economic flow calculated based on the multi-dimensional real-time working condition data is set as the four-dimensional weight tensor of each edge in the heterogeneous process graph ;

[0011] Step two, construct topological generation network, input the heterogeneous process graph And the multi-dimensional real-time working condition data to the topological generation network, generate a group of candidate process topology adjacency matrix ;

[0012] Step three, construct multi-agent reinforcement learning system, each candidate process topology adjacency matrix Instance is used to learn optimal operation parameter combination, and the optimal operation parameter combination and preset 4F-LCA function are used to calculate the corresponding topology total expected value Of each candidate process topology adjacency matrix ;

[0013] Step four, calculate the path robustness score And reconstruction conversion cost Of each candidate process topology adjacency matrix ; combine the topology total expected value , path robustness score And reconstruction conversion cost , obtain final decision score , and select the candidate process topology adjacency matrix With the highest final decision score ​As the optimal process topology , an adaptive execution strategy corresponding to the optimal process topology is outputted.

[0014] Preferably, in the step one, the step of acquiring the multi-dimensional real-time working condition data comprises:

[0015] the raw material property vector is obtained by measuring through an online sensor, the raw material property vector comprising carbon-nitrogen ratio, water content and lignin content;

[0016] the market signal vector is acquired through an external interface, the market signal vector comprising methanol market price, carbon credit price and power grid price;

[0017] the internal working condition vector is obtained by sensor collection of each processing unit, the internal working condition vector comprising temperature, pressure and flow.

[0018] Preferably, in the step one, the step of setting the four-dimensional weight tensor comprises:

[0019] a transmission coefficient representing the conversion efficiency of a substance between two nodes is set as the substance flow weight in the four-dimensional weight tensor ;

[0020] a value representing the energy transfer efficiency or loss on the path is set as the energy flow weight in the four-dimensional weight tensor ;

[0021] an equivalent value representing the environmental load generated by the path is set as the environmental flow weight in the four-dimensional weight tensor ;

[0022] a value representing the operation benefit associated with the path is set as the economic flow weight in the four-dimensional weight tensor .

[0023] Preferably, in the step two, the topology generation network comprises a graph variational autoencoder and a generative decoder;

[0024] the graph variational autoencoder is used to encode the heterogeneous process graph and the multi-dimensional real-time working condition data into a topology distribution latent space ;

[0025] the generative decoder is used to sample from the topology distribution latent space and generate the set of candidate process topology adjacency matrices according to the multi-dimensional real-time working condition data.

[0026] Preferably, the step two is performed by generating a set of candidate process topology adjacency matrices Then, the method further comprises:

[0027] For each candidate process topology adjacency matrix a structure and physical constraint check is performed, and the candidate process topology adjacency matrix that fails the check is discarded;

[0028] a value agent model is constructed, and the value agent model is used to calculate a topology total expected value of the candidate process topology adjacency matrix that passes the check;

[0029] a heuristic value score threshold is set , and the candidate process topology adjacency matrix whose heuristic value score HVS is higher than the heuristic value score threshold is selected for subsequent processing.

[0030] Preferably, in the step three, the 4F-LCA function is set to be composed of a weighted sum of four sub-functions of material flow, energy flow, environmental flow and economic flow;

[0031] The weight factors corresponding to the four sub-functions are introduced to balance economic benefit, environmental impact, energy self-sufficiency rate and material conversion rate;

[0032] The weight factors are adjusted in real time according to the market signal vector to realize dynamic calculation of the topology total expected value .

[0033] Preferably, in the step three, the step of instantiating each candidate process topology adjacency matrix comprises:

[0034] In the multi-agent reinforcement learning system, each functional node in the candidate process topology adjacency matrix is instantiated as an independent agent;

[0035] The goal of the independent agent is set to find the optimal operation parameters of each agent through collaborative learning;

[0036] The optimal operation parameters are used to maximize the topology total expected value of the candidate process topology adjacency matrix calculated based on the 4F-LCA function .

[0037] Preferably, in the step four, the path robustness score and the reconstruction conversion cost are calculatedThe step four includes:

[0038] The step of constructing a Monte Carlo simulation method, by forward propagation of input uncertainty in the multi-dimensional real-time operating data, to calculate the path robustness score ;

[0039] The step of setting an integrated loss assessment model, by assessing energy, material and time losses generated when switching from the current operating topology to the candidate process topology adjacency matrix , to calculate the reconstruction conversion cost .

[0040] The step of calculating the final decision score in the step four includes:

[0041] Setting a small regularization constant ;

[0042] After the reconstruction conversion cost is non-dimensionalized by a cost normalization factor , and added to the small regularization constant , a denominator item is obtained;

[0043] The numerator item is obtained by multiplying the topology total expected value and the path robustness score ;

[0044] The final decision score is calculated by dividing the numerator item by the denominator item.

[0045] The adaptive execution strategy in the step four is preferably a hierarchical dynamic reconstruction strategy, and the execution of the strategy includes:

[0046] Setting a low-cost reconstruction threshold ;

[0047] Comparing the reconstruction conversion cost corresponding to the optimal process topology with the low-cost reconstruction threshold ;

[0048] When the reconstruction conversion cost is lower than the low-cost reconstruction threshold , a first-level reconstruction is triggered to perform a smooth switching;

[0049] When the reconstruction conversion cost is higher than the low-cost reconstruction threshold , a second-level reconstruction is triggered to start a planned reconstruction, and an optimal transition operation sequence is generated.

[0050] Compared with the prior art, the present application has at least the following beneficial effects:

[0051] 1. The present method breaks through the limitation of traditional fixed process flow that can only optimize parameters, and realizes dynamic reconstruction of the entire process topology. By autonomously generating a new process path, the system can fundamentally change the combination and connection mode of the processing units to adapt to real-time changes in raw material characteristics, market signals and internal working conditions, significantly improving the environmental adaptability and overall optimization potential of the entire system.

[0052] 2. The present method constructs a scientific and comprehensive intelligent decision-making framework that unifies and quantitatively models material flow, energy flow, environmental flow and economic flow. The final decision not only depends on the comprehensive expected value of the path, but also weighs the stability risk in actual operation and the conversion cost of switching to a new process, ensuring that the selected process path balances between theoretical optimality, practical feasibility and economic rationality.

[0053] 3. The present method realizes efficient target focusing among a large number of possibilities through generative network and phased screening mechanism. The system first creatively generates multiple candidate process topologies, and then quickly eliminates a large number of infeasible or low-potential schemes using physical constraint verification and value agent models, significantly reducing the computational load of subsequent refined evaluation and improving the efficiency and speed of decision-making.

[0054] 4. The present method establishes a safe and efficient bridge from optimal topology calculation to actual physical execution, and differentially processes process switching of different costs through hierarchical dynamic reconstruction strategy. For low-cost changes, smooth switching is performed, and for high-cost changes, planned reconstruction is started and an optimal transition scheme is generated to minimize losses, ensuring that the optimal solution at the calculation level can be safely and economically applied to actual production, forming a complete closed-loop adaptive control. BRIEF DESCRIPTION OF DRAWINGS

[0055] The present application will be further explained below in conjunction with the accompanying drawings and examples:

[0056] Figure 1 is a flowchart of a kind of four-flow coupling whole life cycle evaluation method for breeding waste resource processing based on graph neural network of the present application. DETAILED DESCRIPTION

[0057] To make the purpose, technical scheme and advantages of the present application more clear and explicit, the present application will be further described in detail below in conjunction with specific examples.

[0058] Example 1

[0059] A kind of four-flow coupling full life cycle evaluation method for kind of waste resource processing based on graph neural network, comprising the following steps:

[0060] Step one, obtain multi-dimensional real-time working condition data representing system state, multi-dimensional real-time working condition data includes raw material characteristic vector, market signal vector and internal working condition vector, and construct heterogeneous process graph including preset processing unit as node and potential process path as edge ;The quantitative value of material flow, energy flow, environmental flow and economic flow calculated based on multi-dimensional real-time working condition data is set as the four-dimensional weight tensor of each edge in the heterogeneous process graph ;

[0061] Step two, construct topology generation network, input the heterogeneous process graph And multi-dimensional real-time working condition data into topology generation network, generate a group of candidate process topology adjacency matrix ;

[0062] Step three, construct multi-agent reinforcement learning system, each candidate process topology adjacency matrix Instantiation is used to learn optimal operation parameter combination, and the optimal operation parameter combination and the preset 4F-LCA function are used to calculate the corresponding topology total expected value Of each candidate process topology adjacency matrix ;

[0063] Step four, calculate the path robustness score And reconstruction conversion cost Of each candidate process topology adjacency matrix ;Combine topology total expected value , path robustness score And reconstruction conversion cost , calculate the final decision score , and select the candidate process topology adjacency matrix With the highest final decision score As optimal process topology , to output the adaptive execution strategy corresponding to the optimal process topology .

[0064] The application provides a kind of four-flow coupling full life cycle evaluation method for kind of waste resource processing based on graph neural network;Multi-dimensional real-time working condition data is obtained to represent system state, wherein raw material characteristic vector Characterize the physical and chemical properties of input material, market signal vector Reflect external economic and environmental information, internal working condition vector Describe system internal operating state parameters;Heterogeneous process graph​ is constructed, which includes nodes representing processing units and edges representing potential process paths; a four-dimensional weight tensor is set as a heterogeneous process map attributes of each edge, which unifies the transmission characteristics of material flow, energy flow, environmental flow and economic flow, where the material flow is material conversion efficiency, the energy flow is energy consumption and energy recovery, the environmental flow is environmental load impact, and the economic flow is cost-benefit index; a topology generation network is constructed and used to input the heterogeneous process map and multi-dimensional real-time operating data to generate a set of candidate process topology adjacency matrices ;

[0065] A multi-agent reinforcement learning system is constructed, each candidate process topology adjacency matrix is instantiated to learn optimal operating parameters cooperatively; using the optimal operating parameters and a preset 4F-LCA function, each candidate process topology adjacency matrix corresponding topology total expected value is calculated; path robustness score and reconstruction conversion cost are also calculated for each candidate process topology adjacency matrix ; the final decision score is obtained by combining the topology total expected value , the path robustness score and the reconstruction conversion cost ; the candidate process topology adjacency matrix with the highest final decision score is selected as the optimal process topology , and an adaptive execution strategy is output;

[0066] This method realizes the transition from parameter optimization to topology reconstruction by constructing a complete decision space driven by real-time multi-dimensional data, enabling the system to autonomously create and evaluate new process flows according to dynamic changes in external environment and internal state, ultimately forming a closed-loop adaptive control system with evaluation, generation, decision-making and execution.

[0067] Embodiment 2

[0068] In step one, the step of obtaining the multi-dimensional real-time operating data includes:

[0069] The raw material property vector is obtained by measuring with online sensors, including carbon-nitrogen ratio, moisture content and volatile solid content;

[0070] The market signal vector is obtained through external interfaces, including methanol market price, carbon credit price and power grid price;

[0071] The internal working condition vector is obtained by sensor acquisition of each processing unit, and includes temperature, pressure, and flow.

[0072] In step one, the four-dimensional weight tensor is set The steps include:

[0073] The transmission coefficient representing the conversion efficiency of the substance between two nodes is set as the substance flow weight in the four-dimensional weight tensor

[0074] The value representing the energy transfer efficiency or loss on the path is set as the energy flow weight in the four-dimensional weight tensor

[0075] The equivalent value representing the environmental load generated by the path is set as the environmental flow weight in the four-dimensional weight tensor

[0076] The value representing the operation benefit associated with the path is set as the economic flow weight in the four-dimensional weight tensor

[0077] This embodiment is a further description of embodiment 1; the acquisition of multi-dimensional real-time working condition data is refined; the raw material characteristic vector is obtained by online sensor measurement, which specifically includes carbon-nitrogen ratio , water content , and lignin content ; the market signal vector is obtained through external interface, which includes methanol market price , carbon credit price , and power grid price ; the internal working condition vector is obtained by sensor acquisition of each processing unit, which includes temperature , pressure , and flow ; the setting steps of the four-dimensional weight tensor are also specified; the substance flow weight is set as the transmission coefficient representing the conversion efficiency of the substance between two nodes; the energy flow weight is set as the value representing the energy transfer efficiency or loss on the path; the environmental flow weight is set as the equivalent value representing the environmental load generated by the path; the economic flow weight is set as the value representing the operation benefit associated with the path;

[0078] For example, for an anaerobic fermentation unit, the substance flow weight can be quantified by the following empirical formula: ​​​​

[0079] ;

[0080] wherein: is a dimensionless process intrinsic coefficient calibrated by historical data fitting or experiments, is the raw material moisture content, is the carbon to nitrogen ratio, is the fermentation temperature; the exponential term in the formula is used to describe the effect of temperature on fermentation efficiency, where the constant 35 is the optimal fermentation temperature, and the constant 10 is the temperature sensitivity coefficient, which has the same dimension as temperature to ensure that the exponential is dimensionless;

[0081] The value representing the energy transfer efficiency or loss on a path is set as the energy flow weight in the four-dimensional weight tensor, for example, the energy flow weight of a certain path can represent the net energy consumption per unit of material processing, and its calculation formula is:

[0082] ;

[0083] wherein: and are the input power and recovered power of the process, respectively, is the material flow rate. The energy value can be further quantified economically in combination with the grid electricity price;

[0084] The equivalent value representing the environmental load generated by a path is set as the environmental flow weight in the four-dimensional weight tensor; for example, the environmental flow weight of a certain path can focus on greenhouse gas emissions, which is estimated by the following formula:

[0085] ;

[0086] wherein: , is the gas emission amount, and GWP is the global warming potential, is the carbon credit price, and is introduced as a unit conversion factor to convert the emission amount from kilograms to tons to match the unit of the carbon credit price, ensuring dimensional consistency;

[0087] The value representing the operational benefit associated with a path is set as the economic flow weight in the four-dimensional weight tensor; for example, for a process path that produces methanol, the economic flow weight can be represented as:

[0088] ;

[0089] wherein: is the methanol mass yield per unit mass of raw material, which is dimensionless, is the material flow rate, is the market price of methanol, is the comprehensive operating cost per unit time, including energy consumption, labor, etc.

[0090] By accurately and multi-dimensionally defining the input data and mapping it into the four-dimensional weight tensor of the graph, a quantitative and comprehensive decision basis is provided for the subsequent model, ensuring that the evaluation of the process path can dynamically reflect the real-time changes of raw materials, market and equipment working conditions.

[0091] Embodiment 3

[0092] In step two, the topology generation network includes a graph variational autoencoder and a generative decoder;

[0093] The graph variational autoencoder is used to encode the heterogeneous process graph and multi-dimensional real-time working condition data into a topological distribution latent space ;

[0094] The generative decoder is used to sample from the topological distribution latent space and generate a set of candidate process topology adjacency matrices according to the multi-dimensional real-time working condition data ;

[0095] After step two generates a set of candidate process topology adjacency matrices , it further includes:

[0096] For each candidate process topology adjacency matrix , structural and physical constraint checking is performed, and candidate process topology adjacency matrices that fail the checking are discarded;

[0097] A value agent model is constructed, and the value agent model is used to calculate the heuristic value score HVS of the candidate process topology adjacency matrices that pass the checking;

[0098] A heuristic value score threshold is set, and candidate process topology adjacency matrices with a heuristic value score HVS higher than the heuristic value score threshold are selected for subsequent processing.

[0099] This embodiment is a detailed description of the topology generation and screening link in embodiment 1; the topology generation network is designed as a structure including a graph variational autoencoder and a generative decoder; the graph variational autoencoder encodes the heterogeneous process graph and multi-dimensional real-time working condition data into a low-dimensional, continuous topological distribution latent space ; wherein the graph variational autoencoder includes an encoder network and a decoder network, the encoder network adopts a 3-layer graph convolutional neural network, each layer contains 128 hidden units, and the activation function is selected as the ReLU function, and the final output is the mean of the latent space vector ​​and variance ; latent space dimension is set to 1 / 4 to 1 / 2 of the original graph node number, the specific value is determined according to the size of the graph, less than 50 nodes take 16 dimensions, 50-200 nodes take 32 dimensions, more than 200 nodes take 64 dimensions; the variational inference adopts the reparameterization technique, that is:

[0100] ;

[0101] wherein is the standard normal distribution sampling value, represents element-wise multiplication;

[0102] The generative decoder samples from the latent space and conditionally generates a set of candidate process topology adjacency matrices according to specific working condition data; the generation process is represented by the following function:

[0103] ;

[0104] wherein, is the candidate process topology adjacency matrix; is the latent space vector; is the raw material characteristic vector; is the market signal vector; is the internal working condition vector; is the generative decoder function, which is based on a graph neural network architecture, and maps the latent space vector and the condition vector to the adjacency matrix through multi-layer graph convolution operation, and the specific calculation process is:

[0105] The latent space vector and the condition vector are fused and a feature embedding matrix of all nodes is generated:

[0106] ;

[0107] wherein, is the total number of nodes in the graph, is the node embedding dimension, MLP is a multi-layer perceptron, represents vector splicing.

[0108] By performing inner product calculation on the node embedding matrix and applying the Sigmoid activation function, each element (i.e. edge existence probability) in the adjacency matrix is decoded:

[0109] ;

[0110] wherein: Sigmoid activation function is used to ensure the value of each element in the adjacency matrix is between 0 and 1, representing the probability of edge generation; finally, the binary adjacency matrix is determined by threshold judgment or Bernoulli sampling; The intermediate node embedding matrix generated by the decoder;

[0111] The training of the graph variational autoencoder uses the negative variational lower bound as the loss function, and the goal is to minimize the loss, which is expressed as:

[0112]

[0113] The first term is the reconstruction loss, which measures the similarity between the generated graph and the original graph; the second term is the KL divergence regular term, which constrains the latent space distribution to be close to the prior distribution The reconstruction loss uses binary cross-entropy to calculate the element-wise error of the adjacency matrix, and the analytical expression of the KL divergence is:

[0114]

[0115] where, is the dimension of the latent space; and represent the j-th component of the mean vector and variance vector output by the encoder; the training uses the Adam optimizer, the learning rate is set to 0.001, the batch size is 32, and the training rounds are 1000 rounds;

[0116] After generating the candidate process topology adjacency matrix , a two-stage screening process is performed; first, each is checked for structural and physical constraints to discard infeasible solutions that have isolated nodes or violate physical and chemical laws; then, the checked topology will have its heuristic value score HVS calculated by a value agent model, whose calculation formula is:

[0117]

[0118] where, is the heuristic value score of the i-th candidate topology, is the value agent model's fast estimate of the potential four-flow value of the topology, which uses a deep neural network structure, with the input being the topology adjacency matrix and the corresponding node feature matrix, and the output being a scalar value estimate, whose calculation formula is:

[0119]

[0120] where: MLP represents a multi-layer perceptron, ​​​​This represents the graph pooling operation. This is the node feature matrix, with dimensions N (where N is the number of nodes). (This is the node feature dimension), each row represents the feature of a processing unit, which is usually composed of the components of the internal working condition vector of that unit;

[0121] Only one heuristic value score (HVS) is above the preset threshold. Candidate process topology adjacency matrix Only then is it sent to the next evaluation stage; among which, the heuristic value score threshold The setting method is as follows:

[0122] First, calculate the mean heuristic value score of all successful topologies in the historical running data. and standard deviation Then set Where k is the screening strictness coefficient, which is taken when a large number of candidate solutions are needed. When it is necessary to select a few options, choose When the system has extremely high quality requirements, take If historical data is lacking, then... The initial value is set to 0.6, and it is dynamically adjusted based on subsequent running results. The adjustment formula is as follows:

[0123] ;

[0124] in: The heuristic value score for the best-performing topology in the current period;

[0125] This design enables the system to autonomously create entirely new process flows, and through an efficient phased screening mechanism, it can quickly focus on a few high-potential candidate solutions from a vast range of possibilities, laying a reliable foundation for subsequent accurate decision-making.

[0126] Example 4

[0127] In step three, the 4F-LCA function is set to be composed of a weighted sum of four sub-functions: material flow, energy flow, environmental flow, and economic flow.

[0128] The weighting factors corresponding to the four sub-functions are introduced to balance economic benefits, environmental impact, energy self-sufficiency rate and material conversion rate;

[0129] The weighting factor is based on the market signal vector. Make real-time adjustments to achieve the total expected value of the topology. Dynamic calculation;

[0130] In step three, the adjacency matrix of each candidate process topology is... The instantiation process includes:

[0131] In the multi-agent reinforcement learning system, each functional node in the candidate process topology adjacency matrix is instantiated as an independent agent;

[0132] The goal of the independent agent is to find the optimal operating parameters of each respective node through collaborative learning;

[0133] The optimal operating parameters are used to maximize the total expected value of the candidate process topology adjacency matrix based on the 4F-LCA function calculation .

[0134] This embodiment is a detailed description of the path value evaluation link in Embodiment 1, and it also clarifies the relationship between the value agent model in Embodiment 3 and the 4F-LCA function in this embodiment: the value agent model is used for rapid pre-screening of candidate topologies, and the heuristic value score HVS output by the value agent model is an approximate estimate of the true 4F-LCA function calculation result. By learning the mapping relationship between the topology structure and the final value in the historical data through a neural network, an efficient candidate scheme screening is provided before the complete multi-agent optimization. The 4F-LCA function is then calculated based on the candidate topologies screened by the value agent model, and the accurate value evaluation is obtained through the complete multi-agent reinforcement learning process.

[0135] The 4F-LCA function is set as the weighted sum of four sub-functions, which is used to calculate the total expected value of each topology ; The specific function expression is:

[0136] ;

[0137] Among them: is the total expected value of the topology, is each sub-flow, where each sub-function adopts the min-max normalization method to unify the dimension, and the normalization formula is , so that all the normalized sub-functions take values in the range of [0, 1] dimensionless values; is the weight factor, and the subscripts mat, eng, env, and eco represent material, energy, environment, and economic flow, respectively. The corresponding weight factors are , , , , each weight factor is a dimensionless parameter and satisfies ;

[0138] The weight factors of the four sub-functions are introduced to balance the economic benefits, environmental impact, energy self-sufficiency rate, and material conversion rate; these weight factors are determined according to the market signal vector Real-time adjustment is performed to realize dynamic calculation of the total expected value of the topology ;

[0139] For each filtered candidate process topology adjacency matrix , the multi-agent reinforcement learning system instantiates each of its functional nodes as an independent agent; each agent uses a deep Q network architecture, the state space includes the operating parameters of the node itself and the state information of the adjacent nodes, and the action space is the discretized value of the adjustable operating parameters of the node; the communication mechanism between agents is based on the sharing of neighborhood state information, and each agent broadcasts an information vector containing its state, reward and action to its adjacent nodes at each time step, where is the state of agent i, is the reward value, is the action, and the adjacent agent j uses it as part of its observation state; the convergence criterion for collaborative learning is that the average Q value of all agents in the last 50 training episodes changes by less than 0.001, or the maximum number of training rounds is reached 5000 rounds; when the actions of multiple agents conflict, a priority scheduling mechanism is used, and the priority is determined according to the criticality of the node in the process flow, and the criticality is calculated by the weighted average of the node degree centrality and the intermediate centrality;

[0140] The goal of these independent agents is to find their optimal operating parameters, such as temperature, pressure, etc., through collaborative learning; this optimal combination of operating parameters is ultimately used to maximize the total expected value of the candidate process topology adjacency matrix based on the 4F-LCA function ;

[0141] This evaluation method ensures that each candidate process not only has a reasonable structure, but also has optimized operating parameters to the best state under current conditions, providing a detailed and quantitative value basis for the final decision.

[0142] Example 5

[0143] In step four, the path robustness score and the reconstruction conversion cost comprise:

[0144] A Monte Carlo simulation method is constructed to calculate the path robustness score by forward propagation of input uncertainty in multi-dimensional real-time operating data;

[0145] An integrated loss evaluation model is set up to evaluate the switching from the current operating topology to the candidate process topology adjacency matrix The energy, material and time loss generated during reconstruction, and the reconstruction conversion cost .

[0146] In step four, the final decision score is calculated The steps include:

[0147] A small regularization constant is set ;

[0148] The reconstruction conversion cost is dimensionless processed by the cost standardization factor , and then added to the small regularization constant to obtain a denominator term;

[0149] The topology total expected value is multiplied by the path robustness score to obtain a numerator term;

[0150] The numerator term is divided by the denominator term to calculate the final decision score .

[0151] This embodiment takes over the calculation of the topology total expected value in Example 4, further supplements the calculation method of the path robustness score and the reconstruction conversion cost required for optimal topology selection, and gives the specific calculation formula of the final decision score , so as to complete the complete decision chain from candidate topology generation, value evaluation to optimal selection;

[0152] A Monte Carlo simulation method is constructed, which calculates the path robustness score by forward propagation of input uncertainty in multi-dimensional real-time operating data, and the specific calculation method is:

[0153] The uncertainty distribution of the input parameters is determined based on historical data analysis, and each component of the raw material characteristic vector is assumed to follow a truncated normal distribution , each component of the market signal vector is assumed to follow a lognormal distribution , and each component of the internal operating condition vector is assumed to follow a uniform distribution , wherein the distribution parameters are estimated according to the operating data of the past 6 months; N Monte Carlo samples (N≥1000) are performed, and the Latin hypercube sampling (LHS) method is used for each sampling to improve sampling efficiency and representativeness; the system output variance is calculated for each sampling result, and the output indicators include material conversion rate, energy efficiency, environmental impact value and economic benefit, and the variance calculation formula is:

[0154] ;

[0155] wherein: is the dimensionless comprehensive output value of the i-th sampling, is the output mean value;

[0156] then , wherein is the dimensionless sensitivity coefficient, the specific value is determined according to the requirement of the system stability, when the system requires high stability , when the system requires medium stability , and when the system requires low stability ; the convergence judgment criterion is that the change of is less than 1% of the current value after continuously increasing 100 times of sampling.

[0157] An integrated loss evaluation model is set, which evaluates the energy, material and time loss generated when switching from the current running topology to the candidate process topology adjacency matrix , to calculate the reconstruction conversion cost ; then, the calculation steps of the final decision score are defined; a regularization small constant is set; the reconstruction conversion cost is added to the regularization small constant to obtain the denominator item; the total expected value of the topology is multiplied by the path robustness score to obtain the numerator item; finally, the denominator item is divided by the numerator item to calculate the final decision score ; the calculation formula is:

[0158] ;

[0159] wherein: is the final decision score; is the total expected value; is the reconstruction conversion cost; is the cost standardization factor, the unit is monetary unit (yuan), the value is equal to the annual average operating cost of the system, so that is a dimensionless value, is a regularization small constant to prevent the denominator from being zero, the value range is to , which is a dimensionless value; is the path robustness score, which is a dimensionless value with a value range of 0 to 1, representing the resistance of the system to disturbance;

[0160] The final decision score The calculation not only considers the theoretical optimal value of the topology, but also balances the stability risk and switching cost in actual operation, ensuring that the finally selected process topology is optimal in theory, practice and economy.

[0161] Embodiment 6

[0162] In step four, the adaptive execution strategy is a hierarchical dynamic reconstruction strategy, and the execution of the strategy includes:

[0163] A low-cost reconstruction threshold is set ;

[0164] The optimal process topology The corresponding reconstruction conversion cost is compared with the low-cost reconstruction threshold ;

[0165] When the reconstruction conversion cost is lower than the low-cost reconstruction threshold , a first-level reconstruction is triggered to perform a smooth switching;

[0166] When the reconstruction conversion cost is higher than the low-cost reconstruction threshold , a second-level reconstruction is triggered to start a planned reconstruction, and an optimal transition operation sequence is generated.

[0167] This embodiment is a specific description of the adaptive execution strategy in Embodiment 1; the strategy is designed as a hierarchical dynamic reconstruction strategy; a low-cost reconstruction threshold is set; the optimal process topology The corresponding reconstruction conversion cost is compared with the threshold ; wherein The superscript of indicates the optimal topology selected by optimization, The subscript C1 in indicates the first-level cost threshold, which is set to 5-15% of the system's daily operating cost; when the reconstruction conversion cost is lower than the low-cost reconstruction threshold , a first-level reconstruction is triggered to perform a smooth switching, which usually involves online adjustment of valves, activation or dormancy of a small number of processing units, without interrupting the main process; when the reconstruction conversion cost is higher than the low-cost reconstruction threshold , a second-level reconstruction is triggered to start a planned reconstruction; this process may require a short shutdown for maintenance to achieve more fundamental process changes, and the system will generate an optimal transition operation sequence to minimize downtime loss, which is calculated by a dynamic programming algorithm, and the objective function is: ​

[0168] ;

[0169] wherein: , , respectively represent the production loss, energy loss and material loss at the tth moment, T is the total reconstruction time; in order to ensure the dimensional uniformity, each loss term is converted into equivalent monetary units during calculation;

[0170] The state of dynamic programming is defined as:

[0171] ;

[0172] wherein: represents the device configuration state, material inventory state and energy reserve state at the tth moment; the state transition equation is , wherein is the operation action at the tth moment, is the external disturbance; the boundary condition is set as the initial state is the current running state, the target state is the optimal process topology corresponding to the steady-state running state;

[0173] The recursive relationship of dynamic programming is:

[0174] ;

[0175] wherein: is the minimum total loss starting from the tth moment state S, is the instantaneous loss function; the algorithm time complexity is , wherein and are the sizes of the state space and the action space respectively, and the space complexity is ;

[0176] The construction of this hierarchical strategy safely and efficiently converts the "optimal topology" at the calculation level into the actual production process in the physical world, and uses continuous feedback for self-optimization, ultimately realizing the dynamic process reconstruction of the entire system level, reflecting the comprehensive consideration of cost-effectiveness and operational feasibility.

[0177] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit it. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. A four-flow coupled full life-cycle assessment method for the resource treatment of agricultural and livestock waste based on graph neural networks, wherein the four flows include material flow, energy flow, environmental flow, and economic flow, characterized in that, include: Step 1: Obtain multi-dimensional real-time operating condition data characterizing the system state. The multi-dimensional real-time operating condition data includes raw material characteristic vectors, market signal vectors, and internal operating condition vectors. Construct a heterogeneous process graph encompassing preset processing units as nodes and potential process paths as edges. ; The quantified values ​​of the material flow, energy flow, environmental flow, and economic flow throughout the entire life cycle, calculated based on the multidimensional real-time operating data, are set as the heterogeneous process map. The four-dimensional weight tensor of each edge ; Step 2: Construct a topology generation network to generate the heterogeneous process map. The multidimensional real-time operating condition data is input into the topology generation network to generate a set of candidate process topology adjacency matrices. ; Step 3: Construct a multi-agent reinforcement learning system, and process each candidate process topology adjacency matrix. Instantiation is performed to collaboratively learn the optimal combination of operating parameters, and the optimal combination of operating parameters and a preset 4F-LCA function are used to calculate the adjacency matrix of each candidate process topology. Corresponding topological total expected value ; The 4F-LCA function is configured to be a weighted sum of four sub-functions: material flow, energy flow, environmental flow, and economic flow. Weighting factors corresponding to these four sub-functions are introduced to balance economic benefits, environmental impact, energy self-sufficiency rate, and material conversion rate. These weighting factors are adjusted in real-time based on the market signal vector to achieve the expected total value of the topology. Dynamic calculation; Step 4: Calculate the topological adjacency matrix for each candidate process. Path robustness score Reconstruction and conversion costs Combined with the total expected value of the topology Path robustness scoring Reconstruction and conversion costs Calculate the final decision score And select the final decision score. The highest candidate process topology adjacency matrix As the optimal process topology To output the optimal process topology The corresponding adaptive execution strategy.

2. The method according to claim 1, characterized in that, Step one, the step of obtaining the multidimensional real-time operating condition data, includes: The raw material characteristic vector is obtained by measuring online sensors. The raw material characteristic vector includes carbon-nitrogen ratio, moisture content and lignin content. The market signal vector is obtained through an external interface, and the market signal vector includes the methanol market price, carbon credit price, and grid electricity price. The internal operating condition vector is obtained by the sensors of each processing unit, and the internal operating condition vector includes temperature, pressure and flow rate.

3. The method according to claim 1, characterized in that, In step one, the four-dimensional weight tensor is set. The steps include: The transport coefficient, which characterizes the efficiency of material conversion between two nodes, is set as the four-dimensional weight tensor. Material flow weights in the data; The numerical value characterizing the energy transfer efficiency or loss along the path is set as the four-dimensional weight tensor. Energy flow weights in the data; The equivalent value representing the environmental load generated by the path is set as the four-dimensional weight tensor. Environmental flow weights in the context; The numerical value representing the operational benefits of the path association is set as the four-dimensional weight tensor. The weight of economic flows in the data.

4. The method according to claim 1, characterized in that, In step two, the topology generation network includes a graph variational autoencoder and a generative decoder. The graph variational autoencoder is used to process the heterogeneous process map. The multidimensional real-time operating condition data is encoded into a topological distribution latent space. middle; The generative decoder is used to extract data from the topological distribution latent space. Sampling is performed during the process, and a set of candidate process topology adjacency matrices is generated based on the multidimensional real-time operating condition data. .

5. The method according to claim 4, characterized in that, Step two involves generating the set of candidate process topology adjacency matrices. Following that, it also includes: For each candidate process topology adjacency matrix Perform structural and physical constraint verification and discard candidate process topology adjacency matrices that fail the verification; Construct a value proxy model, and use the value proxy model to analyze the topological adjacency matrix of the candidate processes that have passed the verification. Calculate its heuristic value score (HVS); Set a heuristic value score threshold And select those whose heuristic value score (HVS) is higher than the heuristic value score threshold. Candidate process topology adjacency matrix This is used for processing in subsequent steps.

6. The method according to claim 1, characterized in that, In step three, the adjacency matrix of each candidate process topology is... The instantiation process includes: In the multi-agent reinforcement learning system, the candidate process topology adjacency matrix is... Each functional node in the process is instantiated as an independent intelligent agent; The goal of the independent agents is to find their respective optimal operating parameters through collaborative learning. The optimal operating parameters are used to maximize the candidate process topology adjacency matrix. The topological total expected value calculated based on the 4F-LCA function .

7. The method according to claim 1, characterized in that, In step four, the path robustness score is calculated. With the reconstruction and transformation cost The steps include: A Monte Carlo simulation method is constructed to calculate the path robustness score by forward propagating the input uncertainties in the multidimensional real-time operating data. ; A comprehensive loss assessment model is established, which evaluates the adjacency matrix of switching from the current operating topology to the candidate process topology. The reconstruction and conversion cost is calculated based on the energy, material, and time losses incurred during the process. .

8. The method according to claim 7, characterized in that, In step four, the final decision score is calculated. The steps include: Set a small regularization constant. ; Reconstructing conversion costs Through cost standardization factors After dimensionless processing, it is compared with the regularization small constant. Add them together to get a denominator term; The expected value of the topology With the path robustness score Multiply them to get a numerator; The final decision score is obtained by dividing the numerator by the denominator. .

9. The method according to claim 1, characterized in that, In step four, the adaptive execution strategy is a hierarchical dynamic reconstruction strategy, and the execution of this strategy includes: Set a low-cost reconstruction threshold ; The optimal process topology Corresponding reconstruction and transformation costs With the aforementioned low-cost reconstruction threshold Compare; When the reconstruction conversion cost Below the low-cost reconstruction threshold At that time, a first-level refactoring is triggered to perform a smooth switch; When the reconstruction conversion cost Above the low-cost reconstruction threshold At that time, a secondary refactoring is triggered to initiate a planned refactoring and generate an optimal sequence of transition operations.

Images