An integrated energy system scheduling method based on intelligent compact constraint learning

Abstract

The application mainly relates to a comprehensive energy system scheduling method based on intelligent compact constraint learning, comprising the following steps: acquiring system historical operation data, containing multi-energy state parameters and labels; based on physical topology and rules, a physical adjacency mask matrix for initializing a sparse neural network (SNN) is analyzed and generated; taking the historical operation data as samples and the physical adjacency mask matrix as constraints, a sparse training algorithm guided by physical rules is adopted to train the SNN, so that a compact physical parameter mapping model is obtained; the model is converted into a constraint of an embeddable optimization problem; a scheduling model facing an optimization target is constructed, the constraint is embedded, and an enhanced model is formed; and an optimal scheduling scheme is solved. Through the cooperation of physical mechanism and data driving, the compact constraint which is efficient in calculation and can be embedded in optimization is constructed, and the difficulty in solving caused by model complexity and nonlinearity in the existing carbon-aware scheduling is effectively overcome.

Description

Technical Field

[0001] This invention relates to the field of integrated energy system optimization scheduling technology, and in particular to an integrated energy system scheduling method based on intelligent compact constraint learning. Background Technology

[0002] An integrated energy system (IES) is a system that operates in coordination with multiple energy sources. To meet the functional requirements of achieving low-carbon operation of this system, existing technologies have proposed carbon sensing scheduling as a key technical means. Its core lies in accurately depicting the generation, transmission, and distribution mechanisms of carbon emissions in the flow of multiple energy sources.

[0003] To achieve this function, existing research typically employs scheduling methods based on the Carbon Emission Flow (CEF) model. This model quantifies the propagation process of carbon emissions by tracking energy flow paths, thereby providing a carbon constraint basis for scheduling decisions. The CEF model can achieve full-process carbon emission analysis from primary energy conversion to end-use consumption, laying the theoretical foundation for carbon emission responsibility sharing.

[0004] However, because actual scheduling requires simultaneous optimization of carbon emissions and multiple energy flows as decision variables, the CEF model introduces a large number of bilinear and higher-order nonlinear terms caused by multi-energy coupled devices (such as cogeneration, heat pumps, etc.). These complex terms are difficult to transform into constraints that can be embedded in the optimization problem, making the model difficult to solve and severely limiting the practical application of CEF in large-scale IES scheduling.

[0005] In recent years, to overcome the challenges of nonlinear modeling, scholars have introduced machine learning methods, particularly deep neural networks (DNNs), aiming to approximate the complex relationship between carbon emissions and multi-energy scheduling through data-driven approaches. Some neural networks employing the ReLU activation function can also be equivalently represented as mixed-integer linear programming (MILP) constraints, thus embedding them into the optimal scheduling model. However, most existing methods rely on fully connected neural network structures with large parameter scales, resulting in excessively high optimization model complexity and low solution efficiency, making it difficult to meet the computational efficiency and stability requirements of IES day-ahead scheduling.

[0006] While sparse neural networks (SNNs) alleviate model complexity to some extent by pruning redundant connections, they still have significant limitations: on the one hand, CEF (Carbon Awareness and Fulfillment) is highly dependent on physical topology and energy conservation relationships, making it difficult for purely data-driven sparse training to stably identify key carbon emission pathways; on the other hand, traditional pruning strategies rely on weight or gradient information, are sensitive to data noise, and suffer from problems such as cold start difficulties, slow convergence speed, weak generalization ability, and poor interpretability. Therefore, current carbon-aware scheduling models still cannot effectively characterize the coupling relationship between carbon emissions and multi-energy scheduling while ensuring physical consistency and computational efficiency. There is an urgent need to develop new modeling methods to achieve compact, efficient, and easily embeddable carbon constraint expressions in optimization models.

[0007] It should be noted that the information disclosed in the background section above is only for understanding the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0008] The technical problem this application aims to solve is "how to construct a computationally efficient and compact carbon constraint that can be embedded in an optimization model through a collaborative intelligent method that combines physical mechanisms and data-driven approaches, so as to overcome the solution dilemma caused by the complexity and nonlinearity of existing integrated energy system carbon sensing scheduling".

[0009] The technical solution adopted in this application to solve the above-mentioned technical problems is as follows.

[0010] This application provides a comprehensive energy system scheduling method based on intelligent compact constraint learning, comprising the following steps:

[0011] S1: Obtain the historical operation dataset of the integrated energy system. The historical operation dataset includes the multi-energy operation status parameters of the integrated energy system and label data related to the system status.

[0012] S2: Based on physical rules, the physical topology of the integrated energy system is analyzed to generate a physical adjacency mask matrix for initializing the sparse neural network SNN;

[0013] S3: Using historical running datasets as training samples and physical adjacency mask matrix as initial structural constraints, a sparse training algorithm guided by physical rules is used to train the SNN to obtain a compact system physical parameter mapping model.

[0014] S4: Convert the trained system physical parameter mapping model into computational constraints that can be embedded into an optimization problem;

[0015] S5: Construct a comprehensive energy system scheduling model to achieve the predetermined optimization objectives, and embed computational constraints into the comprehensive energy system scheduling model to form an enhanced scheduling model;

[0016] S6: Solve the enhanced scheduling model to obtain the optimal scheduling scheme for the integrated energy system.

[0017] In some embodiments, the tag data in step S1 includes the node carbon emission rate of the integrated energy system, and the node carbon emission rate includes at least one of the grid node carbon emission rate, the heating network node carbon emission rate, and the gas network node carbon emission rate.

[0018] The multi-energy operating status parameters in step S1 include at least one of the following: active power injection of the grid node, heat injection power of the heating network node, gas injection of the gas network node, generator carbon intensity parameters, electrical load demand forecast, and heat load demand forecast.

[0019] In some embodiments, in step S2, the physical topology of the integrated energy system is analyzed based on physical rules, which is achieved by using a large language model (LLM) based on domain knowledge enhancement. The LLM is a specialized model obtained by fine-tuning a pre-trained model on professional literature and normative materials in the energy field, or a model that uses retrieval enhancement to generate a RAG architecture and connects it to the physical rule base of the energy system.

[0020] In some embodiments, step S2 generates a physical adjacency mask matrix for initializing a sparse neural network (SNN), specifically including: inputting a user instruction template containing a system topology description into the LLM; based on the user instruction template, the LLM performs decoupling analysis, coupling identification, key filtering, and redundancy removal according to a chain inference template; and outputting the physical adjacency mask matrix.

[0021] The physical rule-guided sparse training algorithm in step S3 is a dynamic sparse training algorithm. During the training process, the dynamic sparse training algorithm performs semantic scoring and guidance on the branching and pruning operations of network connections based on physical rules.

[0022] Semantic scoring and guidance for branching and pruning operations on network connections based on physical rules are implemented using a large-scale language model (LLM). The loss function of the dynamic sparse training algorithm is:

[0023]

[0024] in, This is the data fitting error term. For sparse regularization terms, An adaptive regularization term is constructed based on the semantic score of the LLM output. and These are the weighting coefficients.

[0025] In some embodiments, the construction rule for the adaptive regularization term is as follows: if the LLM determines that the physical correlation of a connection is low, then the corresponding penalty coefficient is increased; if the LLM determines that the physical correlation of a connection is high, then the corresponding penalty coefficient is decreased or waived.

[0026] In some embodiments, the pruning operation is triggered when the weight magnitude of the network connection is less than a first threshold and the semantic confidence score given by the LLM is lower than a second threshold.

[0027] The triggering conditions for branching operations are: the gradient magnitude of a network connection is greater than the third threshold but it has already been pruned, or LLM indicates based on physical reasoning that a certain path must exist but its corresponding weight is zero.

[0028] In some embodiments, the predetermined optimization objective in step S5 includes at least one of minimizing total operating cost and minimizing total carbon emissions, and the computational constraints of the embeddable optimization problem in step S4 include mixed integer linear programming (MILP) constraints. The transformation of MILP constraints is achieved by introducing auxiliary binary variables into the ReLU activation function in the system physical parameter mapping model and linearizing it.

[0029] In some embodiments, an intelligent compact constraint learning system is also provided for generating a compact system physical parameter mapping model in the integrated energy system scheduling method of this application. The system includes:

[0030] The data acquisition module is configured to acquire the historical operation dataset of the integrated energy system. The historical operation dataset includes multi-energy operation status parameters of the integrated energy system and label data related to the system status.

[0031] The topology parsing module is configured to parse the physical topology of the integrated energy system based on physical rules and generate a physical adjacency mask matrix for initializing the sparse neural network (SNN).

[0032] The model training module is configured to use historical running datasets as training samples and physical adjacency mask matrix as initial structural constraints, and to train the SNN using a sparse training algorithm guided by physical rules to obtain a compact system physical parameter mapping model.

[0033] The constraint transformation module is configured to convert the trained system physical parameter mapping model into computational constraints that can be embedded in the optimization problem.

[0034] In some embodiments, a carbon sensing and dispatching device for an integrated energy system is also provided, comprising:

[0035] The intelligent compact constraint learning system of this application is configured to generate a compact system physical parameter mapping model and convert it into computational constraints.

[0036] The scheduling modeling module is configured to construct a comprehensive energy system scheduling model to achieve a predetermined optimization objective, and to embed computational constraints into the comprehensive energy system scheduling model to form an enhanced scheduling model.

[0037] The solution module is optimized and configured to solve the enhanced scheduling model, thereby obtaining the optimal scheduling scheme for the integrated energy system.

[0038] In some embodiments, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the integrated energy system scheduling method of this application.

[0039] The present invention has the following beneficial effects:

[0040] This application constructs a computationally efficient and embeddable compact carbon constraint model by introducing a collaborative intelligent method combining physical mechanisms and data-driven approaches. This significantly improves the solution efficiency and practicality of integrated energy system scheduling models, effectively solving the solution dilemma caused by the complexity and nonlinearity of existing integrated energy system carbon-aware scheduling models. Specifically:

[0041] This application analyzes the physical topology of an integrated energy system and generates a physical adjacency mask matrix by referencing physical rules, such as using Large Language Modeling (LLM). This matrix serves as the initial structural constraint, and a sparse neural network (SNN) is trained using an LLM-guided sparse training algorithm. Finally, the resulting compact model is converted into computational constraints that can be embedded into the optimization problem. This technical solution effectively combines the accuracy of physical mechanisms with the efficiency of data-driven approaches. It significantly reduces model complexity and improves computational efficiency. Furthermore, the generated compact constraints can be directly embedded into the scheduling optimization model, avoiding the solution difficulties caused by the high nonlinearity and nonconvexity of traditional carbon flow models, thus achieving efficient and reliable carbon-aware scheduling.

[0042] Furthermore, by explicitly labeling the data as node carbon emission rates (including the carbon emission rates of power grid, heating network, and gas network nodes), this application ensures that the constructed mapping model can accurately reflect the carbon emission distribution of the multi-energy system, providing a data foundation for the precise embedding of carbon constraints and enhancing the environmental benefits of the scheduling scheme.

[0043] Furthermore, this application specifically defines the "multi-energy operating state parameters" as key physical quantities such as "active power injection at grid nodes, heat injection at heating network nodes, gas injection at gas network nodes, generator carbon intensity parameters, and predicted electricity / heat load demand," ensuring that the model input dataset fully covers the core elements of sources, grids, and loads affecting the system's carbon emission flow. This explicit input limitation avoids the arbitrariness of feature selection, enabling the trained mapping model to more accurately capture the inherent physical relationships throughout the entire process of energy production, transmission, and consumption, thereby significantly improving the accuracy of the model output and its practicality in real-world scheduling applications.

[0044] Furthermore, this application defines the construction method of "Large Language Model (LLM) based on domain knowledge enhancement" (such as fine-tuning based on professional literature and normative materials in the energy field, or using a RAG architecture to access the physical rule base), and clarifies that when generating the physical adjacency mask matrix, it is necessary to "perform decoupling analysis, coupling identification, key screening, and redundancy removal according to the chain-like reasoning template," transforming the abstract intelligent parsing process into a structured and supervised physical rule injection process. This limitation ensures that the parsing of the physical topology does not rely on the fuzzy reasoning or implicit knowledge of a general LLM, but is controlled by a clear domain knowledge base and a rigorous logical rule chain, thereby outputting an initial network mask matrix that is physically reliable and sparse in terms of connectivity (i.e., retaining only strong physical associations). This matrix provides an initialization structure that highly conforms to the actual physical architecture of the system for the subsequent training of the sparse neural network, fundamentally avoiding the training bias and slow convergence problems that may be caused by random initialization or pure data-driven initialization, and laying a reasonable foundation for quickly obtaining a high-precision and physically consistent compact model.

[0045] Furthermore, this application introduces an LLM-guided dynamic sparse training algorithm and corresponding loss function and operation rules, so that the network training process not only depends on the statistical features of the data, but is also guided by the semantics of physical knowledge. This significantly improves the physical consistency and generalization ability of the model while ensuring the sparsity of the model, avoids overfitting, and enhances the robustness of the constraints.

[0046] Furthermore, this application explicitly defines the system physical parameter mapping model as a carbon emission flow (CEF) mapping model, directly targeting the core objective of carbon constraint generation and ensuring a close correlation between model output and scheduling carbon sensing requirements.

[0047] Furthermore, this application ensures that the generated constraint form is compatible with widely used optimization solvers by converting the model into mixed integer linear programming (MILP) constraints, which greatly reduces the difficulty of solving the problem and improves the engineering practicality of the scheduling model.

[0048] Furthermore, this application demonstrates the advantages of this method in achieving both economic and environmental goals by defining optimization objectives (such as minimizing total operating costs or total carbon emissions).

[0049] In summary, the various technical features of this application are synergistic and organically combined: LLM's domain knowledge enhancement ensures a deep integration of physical mechanisms; the sparse neural network and LLM-guided training mechanism guarantee the model's efficiency and accuracy; and the conversion to MILP constraints achieves smooth integration with the optimization model. This integrated framework of "physical analysis-intelligent training-constraint transformation-optimization embedding" fundamentally overcomes the limitations of traditional methods, ultimately achieving the core objective of constructing computationally efficient, embeddable, and physically meaningful compact carbon constraints, providing strong technical support for the low-carbon, economical, and reliable scheduling of integrated energy systems.

[0050] Other beneficial effects of the present invention will be further described below. Attached Figure Description

[0051] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0052] Figure 1 This is a flowchart of one embodiment of this application. Detailed Implementation

[0053] The embodiments of the present invention will be described in detail below. It should be emphasized that the following description is merely exemplary and not intended to limit the scope and application of the present invention.

[0054] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of the present invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0055] This invention proposes a comprehensive energy system scheduling method based on intelligent compact constraint learning. In some embodiments, this invention integrates a large language model (LLM) with knowledge reasoning capabilities into the structure learning of a sparse neural network (SNN) for low-carbon scheduling of an integrated energy system (IES). First, this invention utilizes the LLM as a physical topology and constraint interpreter to guide the physically consistent sparsification of the SNN. Then, an LLM-assisted dynamic branching and pruning strategy is introduced. This strategy combines gradient information with the importance of physical constraints, which can alleviate cold start and noise sensitivity problems while compactly and accurately modeling the CEF relationship. Finally, by utilizing the equivalence between the trained SNN and the mixed integer linear programming (MILP) formula, the learned compact CEF constraints are embedded into the IES scheduling model, achieving co-optimization of carbon and energy.

[0056] In some embodiments, such as Figure 1 As shown, the specific implementation steps are as follows:

[0057] (1) Constructing the IES low-carbon scheduling model

[0058] This invention first constructs an IES low-carbon scheduling model to minimize the total cost of the IES, including operating costs and carbon emission costs;

[0059] 1) Low-carbon dispatching objectives based on load-side carbon emission calculation: The IES considered in this invention includes gas-fired combined heat and power (CHP) units, gas-fired generators (GFG), coal-fired generators (CFG), wind turbines (WT), photovoltaic (PV) units, electric boilers (EB), energy storage systems (ESS), thermostatically controlled loads (TCL), and gas wells (GW). Based on this system configuration, a carbon-sensing dispatching framework was developed to minimize the total cost of the integrated energy system, taking into account both operating costs and carbon emission costs.

[0060]

[0061] In the above model, These are the operating costs of CHP units, GFG units, CFG units, wind and solar curtailment costs, EB units, ESS units, GW gas purchase costs, grid node carbon emission costs, heating network node carbon emission costs, and gas network node carbon emission costs. For the operating costs of each part, and For the power generation and heat production of CHP, This refers to the power generation capacity of GFG. For the power generation capacity of CFG, and Forecast values ​​for wind and solar power output. and Actual output value of wind and solar power, and For the charging and discharging power of ESS, Enter the kilometers for the EB point. For GW of gas purchases, , and These represent the nodal carbon emissions of the power grid, heating network, and gas network, respectively. , and These represent the number of nodes in the power grid, heating network, and gas network, respectively.

[0062] 2) IES Operating Model:

[0063]

[0064]

[0065] in, , and These are the gas consumption, gas-to-electricity coefficient, and gas-to-heat coefficient of the CHP unit, respectively. This refers to the calorific value of natural gas. and These are the gas consumption and gas-to-electricity coefficient of the GFG unit, respectively. and This refers to the reactive power output and power factor of the CFG unit. and These represent the normal distribution and confidence level in the chance constraint, respectively. and These are the thermal power and efficiency of the EB unit, respectively. , , , , These are the upper limit of charging power, upper limit of discharging power, charging efficiency, discharging efficiency, and charging status of the ESS unit. and It is the square of the voltage amplitude and the current. and That is its upper limit. and It refers to the resistance and reactance of the branch circuit. and These represent the active and reactive power of the nodes. and They represent the nodes respectively. Injected power and slave node A set of circuits that output power.

[0066] In this study, using and These represent the outlet temperature and location temperature of the pipeline, respectively. The generation and consumption of heat are determined by the temperature difference.

[0067]

[0068] in For specific heat capacity, For flow through nodes Traffic, and For the supply water node temperature and the return water node temperature, For TCL power consumption, and These are the lower and upper limits of the pipeline temperature. For TCL temperature, and The initial temperature of the TCL circuit and the outdoor temperature are given. , and For TCL parameters, and These are the lower and upper limits of the TCL temperature.

[0069] The output of the gas flow rate (GW) is limited by its operating limits. The Weymouth equation is used to describe the relationship between gas flow rate and pressure in the pipeline, while the sign function indicates the direction of gas flow.

[0070]

[0071] in and The lower and upper limits for natural gas production per GW. For air pressure, and These are the lower and upper limits of air pressure. For pipeline gas flow rate, For pipeline parameters, This is the direction function.

[0072] 3) IES carbon emission flow model:

[0073] The carbon emission flow (CEF) theory consists of power generation carbon intensity (GCI), branch carbon intensity (BCI), and nodal carbon intensity (NCI). For example, in a power system, branch carbon intensity... Equal to the nodal carbon intensity of the inflow node This value is determined by the active power flow injection from the branch and generator connections:

[0074]

[0075] in The corresponding unit's carbon intensity for power generation.

[0076] Similarly, carbon potential in a thermodynamic system and and the carbon potential of natural gas systems It can be calculated using the same method.

[0077] This invention explores the mapping relationship between power generation carbon intensity and load-side node carbon intensity. Taking a power system as an example, it investigates the relationship between the power injection vector and its carbon potential and the nodal carbon flux. The mapping relationship.

[0078]

[0079] in and It is highly non-convex. For power grid flow information, Power injection matrix of the node The impact of this. Therefore, composite mapping. It cannot be directly embedded into the optimization model as a tractable constraint. This invention proposes learning a synthetic mapping. .

[0080] Similarly, in thermal and natural gas systems, the following mapping relationship can be derived:

[0081]

[0082] in Heat flow power injection matrix The impact, Gas-receiving power injection matrix The impact, and The final mapping is also highly non-convex.

[0083] When subsystems learn independently, cross-system coupling variables are treated as exogenous variables. Once the learned model is embedded into a unified scheduling framework, bilinear terms reappear. To eliminate the structural nonconvexity caused by modeling decoupled subsystems, this paper embeds all power-heating-gas coupling variables into a unified learning framework and jointly parameterizes the resulting composite mapping. Specifically, the proposed constrained learning method uses a... A parameterized representation is used to approximate the mapping. ,get and .

[0084] (2) Constructing an intelligent and compact constraint learning method

[0085] This section introduces a framework for intelligent compact constraint learning and integrates it into a carbon-aware scheduling model.

[0086] 1) Physical topology analysis:

[0087] To achieve controllable integration of LLM in this method, this invention positions the language model as a "physical topology parser + structural sparsity judge," with its interaction process strictly occurring outside the numerical training loop. Specifically, the scheduling modeling module first generates a structured system description (node / branch / device list, cross-system energy conversion relationships, carbon flow propagation rules, and learning input / output dimensions), which, together with the user instruction template, constitutes the Prompt. Subsequently, under the constraints of the chained inference template, LLM completes "decoupling analysis - coupling identification - key filtering - redundancy removal," outputting a physical adjacency mask matrix in standard JSON format. And necessary metadata (such as carbon flow direction, list of cross-system coupling paths). The system performs syntax and consistency checks on the JSON result (dimension matching, diagonal / boundary validity, direction constraints, etc.), and after passing the checks, it will... Structural initialization of the sparse neural network: The connection weights are structurally zeroed and frozen using the Hadamard product method, thereby completing a "constraint pre-pruning" before training begins.

[0088] In the subsequent dynamic sparse training phase, LLM does not participate in every gradient update, but only provides semantic-physical consistency scores periodically or when triggered by events to distinguish between "spuriously relevant connections that can be safely pruned" and "critical physical paths that should be retained / branched". To ensure engineering deployability, this invention encapsulates the LLM call interface as an independent evaluation service: the input is statistical information of candidate connections (weight magnitude, gradient magnitude, path length, evidence of violation of physical constraints, etc.) and retrieved domain rule fragments, and the output is a confidence score of [0, 1] and suggested actions. This output is only used to update the network structure mask and regularization strength parameters, without directly modifying the training data or numerical gradients, thereby avoiding embedding uncertain generation processes into the real-time control closed loop while maintaining interpretable physical priors.

[0089] User instruction templates use natural language to precisely describe learning objectives, optimization context, and the topology of the IES. Their standard format is as follows:

[0090] [Description] This system comprises a 30-node power system, a 6-node thermal system, and a 6-node natural gas system. The integrated energy system includes the following physical structure:

[0091] The power subsystem comprises 30 grid nodes (numbered 1–30), 34 distribution lines, 3 gas-fired combined heating and power generating units (CHP) connected to nodes 5, 12, and 21 respectively, 2 coal-fired generators (CFG) connected to nodes 3 and 18, 4 wind turbines (WT) connected to nodes 7, 14, 22, and 27, 3 photovoltaic units (PV) connected to nodes 9, 16, and 25, and 2 energy storage systems (ESS) connected to nodes 10 and 19.

[0092] The thermal subsystem includes 12 heat network nodes (numbered H1–H12), 15 heating pipelines, CHP units supplying heat to nodes H3 and H8, electric boilers (EB) connected to node H6, and 6 temperature control loads (TCL) distributed at H2, H4, H7, H9, H10, and H12.

[0093] Gas subsystem: It includes 8 gas nodes (numbered G1–G8), 10 gas pipelines, gas wells (GW) connected to G1 node, and CHP and GFG draw gas from G3 and G5 nodes respectively.

[0094] Cross-system coupling relationship: The CHP unit realizes the conversion of gas to electricity and heat, the GFG realizes the conversion of gas to electricity, and the EB realizes the conversion of electricity to heat. The time resolution is 1 hour and the scheduling cycle is 24 hours.

[0095] Instructions: Systematically analyze the physical connectivity between nodes, construct a complete adjacency matrix, and clearly identify which nodes' energy injection can be transferred to other nodes in the physical topology. Construct adjacency matrices for the three subsystems of electricity, heat, and gas; analyze the flow direction of energy within each subsystem; identify all cross-system coupling paths; determine the propagation direction of carbon emissions in the following cases: carbon intensity from power generation propagates through active power flow; carbon potential from heat sources propagates through heat flow; carbon potential from gas flows propagates through airflow.

[0096] Clearly identify which nodes do not have a physical direct or indirect carbon flow transfer path;

[0097] The output is a physical adjacency mask matrix that can be used for initializing sparse neural network structures. Note: Carbon flow can only propagate along the energy flow direction; please accurately determine its propagation direction based on the source-charge relationship.

[0098] Statistics: Power systems, heating systems, and gas systems are each modeled as sets. , , .

[0099] The input variables for this learning task include: active power injection at power grid nodes (30 dimensions), heat injection at heating network nodes (12 dimensions), gas injection at gas grid nodes (8 dimensions), generator carbon intensity parameters (9 dimensions), and load demand forecasts (30 dimensions of electrical load and 12 dimensions of thermal load), for a total of 91 dimensions.

[0100] The output variables are: carbon emission rate of power grid nodes (30 dimensions), carbon emission rate of heating network nodes (12 dimensions), and carbon emission rate of gas network nodes (8 dimensions), with a total output dimension of 50 dimensions.

[0101] Output: Output the physical adjacency mask matrix in standard JSON format. .

[0102] To guide LLM in its step-by-step analysis of the computational logic, propagation paths, and node correlations in IES carbon emission streams, this paper further designs a structured CoT inference template:

[0103] Decoupling analysis: distinguishing between basic constraints and complex coupled constraints;

[0104] Coupling identification: The system analyzes the physical coupling relationships between different power generation and energy conversion units;

[0105] Key screening: Identifying the key constraints necessary for learning carbon emission flows;

[0106] Redundancy removal: Prune redundant constraints and assign higher priority weights to critical constraints.

[0107] Obtaining the physical adjacency mask matrix Subsequently, this invention introduces a constrained pre-pruning mechanism during the SNN initialization stage to achieve structural sparsity through the following operations:

[0108]

[0109] in, This represents the Hadamard Product. For those that satisfy... The connections are forced to have their weights set to zero, and gradient updates are frozen in the early stages of training. This operation performs large-scale structural pruning before model training begins, effectively eliminating physically infeasible connections and ensuring that the model focuses on legitimate physical manifolds from the outset, thereby significantly improving training efficiency and physical consistency.

[0110] 2) LLM-guided dynamic sparse training algorithm:

[0111] Simply relying on physical priors may introduce redundant paths, while purely data-driven methods are prone to overfitting noisy data. Therefore, this paper further introduces LLM (Liquidity Matrix) to participate in dynamic branching and pruning discrimination during the training process, achieving deep co-optimization of physical knowledge and data-driven approaches. For a given set of physical priors... In a layered SNN, let the first layer be... The number of neurons in the layer is ,in Indicates the input layer. This represents the output layer. Its forward propagation process can be described as follows.

[0112] Input layer :

[0113]

[0114] in, The input vector represents .

[0115] Hidden layer :

[0116] For the The first in one neuron The calculation process is as follows:

[0117]

[0118] in, Indicates the first Layer The output of each neuron after ReLU activation; Indicates the connection of the first Layer The first neuron and the second Layer The weights of each neuron; For the first Layer Bias terms for each neuron.

[0119] Output layer :

[0120] For the first in the output layer The calculation process for each neuron is as follows:

[0121]

[0122] in, Let be the final output vector of the network, representing the vector generated by the function . Predicted carbon emission rates of each subsystem node The sparsity of this network is reflected in its weight parameters. and It contains a large number of zero elements. The neural connections corresponding to these zero elements are gradually pruned during training, resulting in a network with fewer parameters and a more compact structure.

[0123] For each potential connection in the SNN This article defines its importance as a weighted combination of data-driven and knowledge-driven items, specifically including the following two core components:

[0124] a. Semantic scoring :

[0125] When a connection weight is in an edge state (small but unstable amplitude), the LLM semantic query mechanism is triggered. For example:

[0126] [Query]: "Sparse neural network training shows a weak correlation (weight magnitude 0.03) between the gas pressure at node 5 and the electricity price at node 12, and the physical path between them passes through 3 pressurization stations. Based on IES physical modeling and engineering knowledge, please determine whether this correlation is more likely to stem from real physical coupling or data randomness, and provide a confidence score in the range of 0-1."

[0127] LLM outputs semantic scores based on its internalized knowledge of physics and engineering. ,in This indicates a high level of confidence in physical correlation. This indicates a possible spurious correlation.

[0128] b. Hybrid Loss Function:

[0129] Taking into account data fitting accuracy, structural sparsity, and semantic prior knowledge, this paper constructs the following multi-objective optimization function:

[0130]

[0131] in, Mean squared error loss is used to measure the quality of data fitting. For the traditional Sparse regularization terms improve the overall sparsity of the network. This is an adaptive regularization term guided by LLM semantics. The core mechanism of this design is: if the LLM determines low physical relevance ( ), then the penalty coefficient ( ), forcing the weight to decay rapidly to zero; conversely, if LLM determines that it is physically critical ( If the penalty tends towards zero, it effectively protects the weight from being... The regularization was mistakenly deleted.

[0132] During the training iteration, the SNN executes the following structure update logic driven by both LLM and gradient information:

[0133] a. Pruning:

[0134] Triggering condition: When and When the semantic confidence threshold is reached, LLM acts as a "validator," using physical reasoning to confirm that the connection indeed has no actual physical meaning, thus safely pruning it and effectively avoiding the accidental deletion of important connections with weak signals but reasonable physical meaning.

[0135] b. Growth:

[0136] Triggering condition: When gradient Significant, but currently (States that have been pruned); or when LLM, based on physical reasoning (such as the law of conservation of energy or the principle of thermal inertia), indicates that a certain path must exist, but the corresponding weight in the data fitting is zero. LLM acts as a "Discoverer".

[0137] In this invention, the LLM used is not an unconstrained general-purpose basic model, but a specialized inference model enhanced with knowledge from the energy domain to ensure the accuracy, stability, and repeatability of the physical inference results. Specifically, in one embodiment, the LLM is fine-tuned based on a publicly available pre-trained language model. The fine-tuning data comes from professional literature and standards related to energy systems, including but not limited to: textbooks on power system analysis and operation, textbooks on integrated energy system modeling, standard textbooks on thermodynamics and heat transfer, natural gas pipeline operation specifications, carbon emission accounting standards (such as relevant ISO standards), power grid dispatching procedures, and publicly available energy system operation case data.

[0138] In another implementation, LLM employs a Retrieval-Augmented Generation (RAG) architecture, which introduces a domain knowledge base as an external constraint information source during model reasoning. The knowledge base includes a structured energy system physics rule base, an electro-thermal-gas coupling relationship graph, a database of equipment operation constraint parameters, and a carbon emission calculation rule base. When performing physical structure analysis or semantic scoring tasks, the system first performs knowledge retrieval based on the topology and parameter information of the current integrated energy system. The matched rules and standard text are then input into the language model as context, and the model generates structured reasoning results within the defined knowledge scope. In this way, the model generation process is constrained within the energy physics rule space, thereby significantly reducing the risk of illusion and improving the consistency between the reasoning results and engineering principles.

[0139] 3) SNN embedded CEF constraint model:

[0140] After training, the SNN can be precisely converted into a set of equivalent mixed-integer linear programming (MILP) constraints. This conversion mainly relies on the linearization of the ReLU activation function, which is achieved by introducing auxiliary binary variables and employing a Big-M expression. Based on this conversion process and the resulting sparse network structure, Unified as input vector and carbon emission rate Represented as network output In its general form, the following linear inequality constraints can be obtained:

[0141]

[0142]

[0143] in, and They represent the first Layer Pre-activation and post-activation values ​​of each neuron; and These are the corresponding upper and lower bound constants; Let be the binary variable representing the ReLU activation state. Therefore, the CEF model, which is complex and difficult to handle directly in terms of operational constraints, can be approximated by the trained SNN and reformulated as a completely equivalent MILP constraint within the proposed framework.

[0144] In this invention, the input vector of the neural network An extended state vector containing information about the system's current operating status is used to characterize the dynamic operating features of the integrated energy system, including: system power load, CHP unit power generation, CFG unit power generation, carbon potential of each generator unit, TCL load demand, heat production capacity of CHP and EB units, natural gas load demand, and GW of gas purchases.

[0145] Assuming a neural network contains L layers, with an average of S neurons per layer, the total number of neurons in the network is approximately L×S. Since each ReLU activation function can be equivalently linearized by introducing a binary variable and several linear constraints, the number of newly introduced binary variables after conversion to a MILP model is approximately L×S, and the corresponding number of newly introduced linear constraints is also approximately L×S. If we consider a network sparsity rate of ρ (0 < ρ ≤ 1), meaning each layer retains only ρ×S effective neurons, the actual scale of the introduced binary variables is approximately L×ρ×S, which is significantly lower than that of a fully connected network.

[0146] To further illustrate the model scale and solution efficiency, this invention uses a small-scale test system as an example. This test system comprises a 6-node power grid, a 3-node heating network, and a 3-node gas network, totaling 12 energy nodes. A three-layer sparse neural network is used to approximate the carbon emission flow mapping relationship, with each layer containing 10 neurons and a sparsity rate of 0.6, resulting in approximately 18 effective neurons. After ReLU linearization, 18 binary variables and approximately 72 linear inequality constraints are introduced (each ReLU corresponds to 4 standard linearization constraints). Based on this, a complete scheduling MILP model is constructed, with a total number of integer variables less than 30 and a continuous variable count of approximately 120.

[0147] Under the same test system conditions, when using the conventional iterative adjustment method, the average solution time increases significantly with the refinement of the time scale (the solution time for some systems exceeds 1 hour), and there is even a risk of local optima. However, when using the SNN to linearize the model into a MILP model, and solving it with a modern commercial solver, a globally optimal solution can be obtained (average 25 seconds). Therefore, the linearized MILP model has an order-of-magnitude advantage in solution efficiency compared to the original method.

[0148] Furthermore, the MILP problem derived from this invention can be solved using modern high-performance optimization solvers, such as commercial mixed-integer programming solvers like Gurobi or CPLEX. This invention constructs a sparse neural network structure through a physical masking mechanism, resulting in a highly sparse constraint matrix after linearization and low constraint coupling, thereby further improving solution efficiency.

[0149] (3) Case analysis;

[0150] Taking an IES comprising a 30-node IEEE power system, a 6-node thermal system, and a 6-node natural gas system as an example, the learning and scheduling time resolution is 15 minutes. Randomly sampled power and natural gas load data are used to generate 6000 scenarios within ±20% of the nominal level. For each scenario, economic scheduling is performed without considering carbon emissions, and the resulting scheduling variables and carbon emissions are used as input and output vectors, respectively. All input data are normalized to the [0, 1] interval to address unit heterogeneity and improve approximation accuracy. To evaluate the advantages of the proposed carbon-sensing model, the following comparison method is used:

[0151] Method 1 (M1): Carbon-sensing scheduling driven by the proposed intelligent compact constraint learning.

[0152] Method 2 (M2): Carbon-sensing scheduling based on dense neural networks.

[0153] Method 3 (M3): Carbon-sensing scheduling based on sparse neural networks.

[0154] Method 4 (M4): Carbon-sensing scheduling based on iterative adjustment of load carbon potential.

[0155] Method 5 (M5): Economic dispatch model that does not consider the impact of carbon emissions.

[0156] The only difference between M2 and M3 and M1 is the different neural networks. M2 and M3 do not undergo any special processing, while M4 and M5 are both traditional scheduling methods.

[0157] Based on the system topology, the LLM resolves a 133x96 physical adjacency mask matrix with 8023 non-zero elements. We construct an SNN with 133 input nodes, 256 hidden nodes, and 96 output nodes. After training, the network weight sparsity reaches 36%. In the LLM auxiliary module, a pre-trained Qwen-7B model is used for policy generation. Qwen-7B is a Transformer-based large-scale language model (LLM) with 7 billion parameters and a vocabulary of 153,000 words.

[0158] Table 1 Comparison of Approximation Errors

[0159]

[0160] Table 1 compares the mean square error (MSE), root mean square error (RMSE), and R0 of M1-M3. 2 In terms of performance, M1 has the best overall performance among all models. Its prediction error is consistently low and R0 is high. 2 A value close to 1 indicates that the proposed method can accurately approximate the nonlinear carbon emission flow mapping. In contrast, M2 performs the worst, suggesting that ignoring physical constraints and structural regularization leads to redundancy and poor generalization ability. M3 provides a moderate improvement by partially introducing structural information and slightly outperforms in global trend fitting, revealing the trade-off between accuracy and consistency. The performance gap between M1 and M3 further confirms the effectiveness of the proposed method's LLM-assisted constraint learning and dynamic pruning. By embedding prior knowledge of physical topology and semantic guidance, the proposed method preserves key physical paths while eliminating redundant connections, thus generating a compact and expressive model.

[0161] To verify the training efficiency advantage of the proposed method in the cold start phase, we compared the number of convergence iterations and wall-bell time required for each method to reach the same accuracy threshold under the same training data scale, network width, optimizer settings (learning rate, batch size, etc.), and hardware environment. In this paper, "convergence" is defined as the validation set metric first satisfying RMSE ≤ 0.012 and not bouncing more than 2% in the subsequent 500 iterations. Experimental results show that M1 guided by physical prior masks and LLM semantic consistency achieves a convergence time of approximately 3.8 × 10⁻⁶. 4 Convergence can be achieved in just a few iterations; in contrast, the baseline M2 of a dense network requires approximately 8.9 × 10⁻⁶ iterations. 4 In the next iteration, the static sparse network M3 requires approximately 6.6 × 10⁻⁶. 4 The number of convergence iterations for M1 is reduced by approximately 57% and 42% compared to M2 and M3, respectively, and the convergence time is shortened by approximately 59% and 44%, respectively. These results demonstrate that M1 can significantly compress the invalid search space in the early stages of training through structural pre-pruning and suppress invalid exploration of spurious connections through semantic-physical consistency constraints provided by LLM, thereby achieving faster and more stable cold-start convergence performance.

[0162] In the compact constraint learning training process of this invention, LLM does not participate in every gradient iteration update, but rather participates in the physical consistency evaluation and adjustment of the network structure in a periodic, event-triggered manner. Specifically, in a typical implementation, the total number of training iterations of the neural network is set to 50,000. After every 1,000 parameter updates, a semantic physical consistency scoring and structural optimization evaluation based on the domain knowledge-enhanced language model is triggered. Therefore, the number of times LLM is called during the complete training cycle is approximately 50, accounting for 0.1% of the total number of training iterations. In another implementation, LLM is called using an event-driven mechanism, triggered when any of the following conditions are met: first, the network completes a pruning or pluralization operation; second, the convergence rate of the validation set loss function is lower than a preset threshold within several consecutive iterations; third, the physical consistency constraint violation rate exceeds a set proportion. Thus, LLM only participates in decision-making during the critical structural adjustment stage and does not participate in the conventional numerical gradient calculation process.

[0163] From a computational overhead perspective, under the condition of using a locally deployed domain-enhanced language model with 7B parameters, a single inference takes approximately 1 to 2 seconds, while a single network gradient update takes approximately 10 to 50 milliseconds. In the aforementioned scenario of 50 calls, the additional time overhead introduced by LLM is approximately 50 to 100 seconds, accounting for approximately 2% to 5% of the overall training time (typically 30 to 60 minutes). Therefore, the introduction of LLM does not significantly increase the computational burden during the training phase. Furthermore, this invention adopts an offline training and online deployment operation mode. The language model is only used during the model structure learning phase. After the structure is determined and parameter training is completed, LLM is no longer called during the actual integrated energy system scheduling operation phase; instead, only the linearized MILP solution process is executed. Since the typical day-ahead scheduling cycle of an integrated energy system is once every 24 hours, model training is only re-executed when the system topology or equipment parameters change. Therefore, the LLM interaction mechanism does not affect the real-time performance of scheduling and the stability of system operation.

[0164] In summary, by limiting the frequency of language model calls and adopting a mechanism that combines periodicity and event triggering, this invention effectively avoids the uncertainties and computational delays that may result from embedding the language model into the real-time control loop. While ensuring the optimization effect of structural physical consistency, it maintains the computational controllability and industrial feasibility required for engineering deployment.

[0165] Table 2 Comparison of Scheduling Results

[0166]

[0167] Table 2 summarizes the system's scheduling results. After embedding the constraints into the scheduling model, the solution time was reduced from over 1 hour in the traditional nonlinear model to 37.19 seconds, with a carbon emission calculation error of less than 0.5%. Compared to M5, M1–M4 all achieved emission reduction and cost optimization. Among them, M1 performed best, reducing carbon emissions by 9.55% and operating costs by 10.05%, while maintaining satisfactory computational efficiency. Furthermore, compared to M2–M4, M1 achieved additional reductions of 2.56% in emissions and 2.00% in costs, further validating the effectiveness of the proposed framework.

[0168] In some embodiments, Physically Guided Neural Architecture Search (NAS) can be used instead of LLM direct guidance. That is, instead of directly using LLM to score connections, a restricted search space is defined using hard constraints extracted from the physical topology (such as figure rules and differential equation constraints). Then, efficient Neural Architecture Search (NAS) algorithms (such as DARTS and SNAS) are used to find high-performance compact network structures within this restricted space.

[0169] Specifically, NAS based on graph rules can be used. Graph features such as node degree, centrality, and community structure are extracted from the IES topology graph and transformed into constraints on the connection probabilities of the SNN. For example, the initial probability of a connection between the input and output neurons of physically directly connected nodes is higher. Differentiable NAS methods are then used to search on this biased probability distribution.

[0170] Specifically, NAS can be based on differential equation constraints. The partial differential equations (PDEs) describing carbon flow propagation are discretized, and their discretization format naturally defines a sparse convolution kernel or graph convolution operation. This PDE discrete kernel is used as a fixed sparse connection pattern (not trainable) for some layers in the SNN, and then NAS is used to search for the structure of other layers.

[0171] In some embodiments, LLM can also be used to generate training data or synthesize features. LLM does not directly intervene in the training process, but is used in the earlier data augmentation or feature engineering stages. For example, the generative capabilities of LLM can be used to simulate and generate more training data that conforms to physical laws, or to interpret and synthesize new, higher-order features that are more effective in predicting carbon flow mapping from the original data.

[0172] Specifically, LLM can be used to assist in data generation. The system topology and some operating parameters are input into the LLM, which is then asked to infer and generate other possible operating states and their corresponding carbon emission rate labels based on physical laws, thereby expanding the training dataset. A regular SNN (with sparse regularization) is then trained using the expanded dataset.

[0173] Specifically, feature engineering can be assisted by LLM. Raw operational data is input into the LLM, and by designing specific instructions (such as "From the following data, find the derived index that best reflects the degree to which the carbon potential of node A is affected by the power generation of node B"), the LLM outputs a series of suggested new features. These LLM-generated features, along with the original features, are then input into the SNN for training.

[0174] In some embodiments, an intelligent compact constraint learning system is also provided for generating a compact system physical parameter mapping model in the integrated energy system scheduling method of this application. The system includes:

[0175] The data acquisition module is configured to acquire the historical operation dataset of the integrated energy system. The historical operation dataset includes multi-energy operation status parameters of the integrated energy system and label data related to the system status.

[0176] The topology parsing module is configured to parse the physical topology of the integrated energy system based on physical rules and generate a physical adjacency mask matrix for initializing the sparse neural network (SNN).

[0177] The model training module is configured to use historical running datasets as training samples and physical adjacency mask matrix as initial structural constraints, and to train the SNN using a sparse training algorithm guided by physical rules to obtain a compact system physical parameter mapping model.

[0178] The constraint transformation module is configured to convert the trained system physical parameter mapping model into computational constraints that can be embedded in the optimization problem.

[0179] In some embodiments, the topology parsing module (based on knowledge graph query) may not directly invoke generative LLM, but instead incorporate a pre-built integrated energy system knowledge graph. This graph stores the attributes of various devices, nodes, and branches, as well as their physical connections, energy conversion rules, and carbon flow propagation rules in a graph structure. Upon receiving a system topology description, the module parses the description text, extracts key entities and relationships, and then converts it into query statements for the knowledge graph (such as Cypher queries). It then generates a physical adjacency mask matrix through graph querying and reasoning. The knowledge graph can be built based on industry standards (such as the CIM standard) and stored and queried using a graph database (such as Neo4j). The parsed text can employ a lightweight NLP model or rule templates.

[0180] In some embodiments, the model training module (in a reinforcement learning-based environment) can construct the semantic scoring mechanism of LLM as a reward function and integrate it into the reinforcement learning (RL) environment. The structure search process of the SNN is modeled as a Markov decision process (MDP), where the agent's actions are adding or deleting connections, and the state is the current network structure and performance metrics. LLM does not directly score individual connections but provides a physically consistent reward signal for the overall network structure change proposed by the agent. The model training module then runs RL algorithms (such as PPO) to explore network structures that yield high rewards (high physical consistency + high accuracy).

[0181] In some embodiments, a carbon sensing and dispatching device for an integrated energy system is also provided, comprising:

[0182] The intelligent compact constraint learning system of this application is configured to generate a compact system physical parameter mapping model and convert it into computational constraints.

[0183] The scheduling modeling module is configured to construct a comprehensive energy system scheduling model to achieve a predetermined optimization objective, and to embed computational constraints into the comprehensive energy system scheduling model to form an enhanced scheduling model.

[0184] The solution module is optimized and configured to solve the enhanced scheduling model, thereby obtaining the optimal scheduling scheme for the integrated energy system.

[0185] In some embodiments, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the integrated energy system scheduling method of this application.

[0186] In summary, this application has the following advantages over the prior art:

[0187] This invention proposes a carbon-aware scheduling method for integrated energy systems driven by intelligent compact constraint learning. It introduces the knowledge reasoning capability of Large Language Models (LLMs) into the learning process of Sparse Neural Networks (SNNs) for carbon-aware scheduling of integrated energy systems (IES). Unlike existing research that uses SNNs to post-process the pruning results driven by LLMs, this invention uses LLMs as interpreters of physical topology and constraints, guiding the network to achieve physical consistency sparsity. It proposes an LLM-assisted dynamic branching and pruning strategy that combines gradient information with constraint importance assessment based on physical knowledge. This strategy alleviates the cold-start problem and noise sensitivity while achieving compact and high-precision modeling of carbon emission flows (CEFs). By leveraging the equivalence between the trained sparse neural network and mixed-integer linear programming (MILP) models, this invention embeds the learned compact CEF constraints into the IES scheduling model, achieving co-optimization of carbon and energy.

[0188] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0189] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0190] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0191] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0192] The background section of this invention may include background information about the problems or environment in which the invention is being developed, and is not necessarily a description of prior art. Therefore, the content included in the background section does not constitute an admission of prior art by the applicant.

[0193] The above description provides a further detailed explanation of the present invention in conjunction with specific / preferred embodiments, and it should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various substitutions or modifications can be made to these described embodiments without departing from the concept of the present invention, and all such substitutions or modifications should be considered within the scope of protection of the present invention. In the description of this specification, the reference to terms such as "an embodiment," "some embodiments," "preferred embodiment," "example," "specific example," or "some examples," etc., indicates that the specific features, structures, materials, or characteristics described in connection with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any suitable manner in one or more embodiments or examples. Without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification and the features of different embodiments or examples. Although the embodiments of the present invention and their advantages have been described in detail, it should be understood that various changes, substitutions, and modifications can be made herein without departing from the scope of protection of the patent application.

Claims

1. A comprehensive energy system scheduling method based on intelligent compact constraint learning, characterized in that, Includes the following steps: S1: Obtain the historical operation dataset of the integrated energy system, which includes multi-energy operation status parameters of the integrated energy system and label data related to the system status; S2: Based on physical rules, the physical topology of the integrated energy system is analyzed to generate a physical adjacency mask matrix for initializing the sparse neural network SNN; S3: Using the historical running dataset as training samples and the physical adjacency mask matrix as initial structural constraints, the SNN is trained using a sparse training algorithm guided by physical rules to obtain a compact system physical parameter mapping model. S4: Convert the trained system physical parameter mapping model into computational constraints that can be embedded in an optimization problem; S5: Construct a comprehensive energy system scheduling model to achieve the predetermined optimization objective, and embed the computational constraints into the comprehensive energy system scheduling model to form an enhanced scheduling model; S6: Solve the enhanced scheduling model to obtain the optimal scheduling scheme for the integrated energy system; In step S2, the physical topology of the integrated energy system is analyzed based on physical rules, which is achieved through a large-scale language model (LLM) based on domain knowledge enhancement; a physical adjacency mask matrix is ​​generated, specifically including: inputting a user instruction template containing a system topology description into the LLM; based on the user instruction template, the LLM performs decoupling analysis, coupling identification, key filtering, and redundancy removal according to a chain-like reasoning template; and outputting the physical adjacency mask matrix. The physical rule-guided sparse training algorithm in step S3 is a dynamic sparse training algorithm. During the training process, the dynamic sparse training algorithm performs semantic scoring and guidance on the branching and pruning operations of network connections based on physical rules. The computational constraints of the embeddable optimization problem in step S4 include mixed integer linear programming (MILP) constraints. The conversion to MILP constraints is achieved by introducing auxiliary binary variables into the activation function of the system physical parameter mapping model and then linearizing it.

2. The integrated energy system scheduling method based on intelligent compact constraint learning according to claim 1, characterized in that, The tag data in step S1 includes the node carbon emission rate of the integrated energy system, which includes at least one of the grid node carbon emission rate, heating network node carbon emission rate, and gas network node carbon emission rate. The multi-energy operating status parameters in step S1 include at least one of the following: active power injection of the power grid node, heat injection power of the heating network node, gas injection of the gas network node, generator carbon intensity parameter, predicted electrical load demand, and predicted heat load demand.

3. The integrated energy system scheduling method based on intelligent compact constraint learning according to claim 1, characterized in that, The LLM is a specialized model obtained by fine-tuning a pre-trained model based on professional literature and normative materials in the energy field, or a model that uses retrieval enhancement to generate a RAG architecture and connects to the physical rule base of the energy system.

4. The integrated energy system scheduling method based on intelligent compact constraint learning according to claim 3, characterized in that, The semantic scoring and guidance of network connection pruning operations based on physical rules is implemented through a large-scale language model (LLM), and the loss function of the dynamic sparse training algorithm is: in, This is the data fitting error term. For sparse regularization terms, An adaptive regularization term is constructed based on the semantic score of the LLM output. and These are the weighting coefficients. Score the semantics.

5. The integrated energy system scheduling method based on intelligent compact constraint learning according to claim 4, characterized in that, The construction rule for the adaptive regularization term is as follows: if the LLM determines that the physical correlation of a connection is low, then the corresponding penalty coefficient is increased; if the LLM determines that the physical correlation of a connection is high, then the corresponding penalty coefficient is decreased or waived.

6. The integrated energy system scheduling method based on intelligent compact constraint learning according to claim 4, characterized in that, The pruning operation is triggered when the weight magnitude of the network connection is less than the first threshold and the semantic confidence score given by LLM is lower than the second threshold. The triggering conditions for the branching operation are: the gradient magnitude of the network connection is greater than the third threshold but it has been pruned, or LLM indicates based on physical reasoning that a certain path must exist but its corresponding weight is zero.

7. The integrated energy system scheduling method based on intelligent compact constraint learning according to claim 1, characterized in that, The predetermined optimization objective in step S5 includes at least one of minimizing total operating cost and minimizing total carbon emissions, and the activation function is the ReLU activation function.

8. A smart compact constraint learning system for generating a compact system physical parameter mapping model in the integrated energy system scheduling method based on smart compact constraint learning according to any one of claims 1-7, characterized in that, The system includes: The data acquisition module is configured to acquire the historical operation dataset of the integrated energy system, which includes multi-energy operation status parameters of the integrated energy system and label data related to the system status. The topology parsing module is configured to parse the physical topology of the integrated energy system based on physical rules and generate a physical adjacency mask matrix for initializing the sparse neural network (SNN). The model training module is configured to use the historical running dataset as training samples and the physical adjacency mask matrix as initial structural constraints, and to train the SNN using a sparse training algorithm guided by physical rules to obtain a compact system physical parameter mapping model. The constraint transformation module is configured to convert the trained system physical parameter mapping model into computational constraints that can be embedded in the optimization problem.

9. A carbon sensing and dispatching device for an integrated energy system, characterized in that, include: The intelligent compact constraint learning system as described in claim 8 is configured to generate a compact system physical parameter mapping model and convert it into computational constraints; The scheduling modeling module is configured to construct a comprehensive energy system scheduling model for achieving a predetermined optimization objective, and to embed the computational constraints into the comprehensive energy system scheduling model to form an enhanced scheduling model; The optimization solution module is configured to solve the enhanced scheduling model to obtain the optimal scheduling scheme for the integrated energy system.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the integrated energy system scheduling method based on intelligent compact constraint learning as described in any one of claims 1-7.

Images