A multi-stage heating collaborative regulation method based on a physical perception large model and a double-flow decoupling LoRA, an electronic device, and a storage medium

Abstract

The application belongs to the technical field of energy control, and provides a multi-stage heat supply collaborative regulation method based on a physical perception large model and a double-flow decoupling LoRA, an electronic device and a storage medium, the method comprising: multi-stage heat supply collaborative regulation model regulation; the model training process comprises: graph coding, time series coding, multi-source data fusion, first constraint embedding, second constraint embedding, iterative training, prototype library construction, parameter matching and assembly; through an incremental fine-tuning mechanism, the fine-tuning parameter matrix is decomposed into two orthogonal low-rank matrix branches of static physical flow and dynamic environment flow, and an adaptive gating coefficient based on the system state context is introduced, so that decoupling learning and dynamic fusion of physical properties and environmental strategies are realized; through the first constraint and the second constraint, the model output is constrained to conform to the control instruction of the physical law; through the LoRA prototype library, rapid deployment under the condition of zero samples or few samples is realized.

Description

Technical Field

[0001] This invention relates to the field of energy control technology, and in particular to a multi-level heating coordinated control method, electronic device and storage medium based on a physical sensing large model and dual-flow decoupling LoRA. Background Technology

[0002] Urban centralized heating systems are typical complex physical systems characterized by large time delays, strong coupling, and nonlinearity. They typically consist of one or more energy supply units, N energy conversion / distribution units, and M energy consumption units. The energy supply unit, as the system's energy starting point, has the core task of adjusting the outlet water temperature and flow rate according to the total heat load demand of the entire network, ensuring a balance between total heat supply and demand. The energy conversion / distribution unit, as the hub connecting the primary and secondary networks, is tasked with distributing heat on demand according to the specific building characteristics of the areas under its jurisdiction, resolving the hydraulic imbalance problem in the last mile.

[0003] Although industrial internet technology has been widely applied in the heating sector, traditional heating systems still heavily rely on manual experience or mechanism-based PID control to achieve intelligent collaborative control across the entire source-network-station chain. This presents the following core challenges: 1) Multi-source heterogeneous data, making unified modeling difficult: Heating systems have extremely complex data sources, including high-frequency sampled meteorological, supply and return water temperatures, pump frequencies, and other time-series data; complex network topology and hydraulic balance states, resulting in graph-structured data with spatial attributes; and discrete static attributes such as building age, insulation materials, and user complaint records. Traditional control methods, such as PID or simple neural networks, struggle to process this heterogeneous data simultaneously within a unified semantic space, often leading to the neglect of spatial lag in the network or personalized user needs. 2) Weak model generalization ability and high migration costs: Heating systems exhibit strong non-standard attributes. The network structure and building thermal inertia vary significantly between different residential areas. Existing models are typically designed for a single region, meaning a high-precision model trained in one region often fails when directly applied to another due to feature distribution shifts. This necessitates retraining from scratch with a large amount of new data, resulting in long deployment cycles and high implementation costs. 3) Direct application of large models poses security risks: While general-purpose large language models possess powerful reasoning capabilities, they are essentially probabilistic models lacking prior knowledge of the physical world. Directly using them for industrial control can easily lead to misinterpretations, outputting control commands that violate thermodynamic laws or exceed equipment safety boundaries, posing security risks. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a multi-level heating coordinated control method, electronic device and storage medium based on a physical sensing large model and dual-flow decoupling LoRA, which solves the problems of existing methods being difficult to unify modeling, having high migration costs and posing safety hazards when directly applied.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A multi-stage coordinated heating regulation method based on a physical sensing large model and dual-flow decoupled LoRA includes:

[0007] Multi-source operational data of the target community is collected, and the multi-source operational data is input into a pre-trained multi-level heating coordinated control model to generate control commands. Control commands are then generated, and the heating system of the target community is controlled according to the control commands. The training process of the multi-level heating coordinated control model includes:

[0008] The graph neural network is used to encode the topological data of the thermal pipeline network to obtain the graph structure embedding;

[0009] Time-series slicing encoding and one-dimensional convolution are performed on high-frequency time-series signals to obtain time-series token embeddings;

[0010] Linear projection and concatenation operations are performed on the graph structure embedding, the temporal token embedding, and the discrete static attribute embedding to obtain the input sequence;

[0011] Embed the first layer of physical causality-based constraints in the self-attention computation layer of the Transformer model;

[0012] A second constraint based on the law of energy conservation and the mean square error of the control command is embedded at the output of the Transformer model.

[0013] Based on the input sequence, the Transformer model is iteratively trained using an incremental fine-tuning mechanism based on two-stream decoupling LoRA according to the first and second constraints, to obtain the original collaborative regulation model; the expression of the incremental fine-tuning mechanism includes:

[0014] ,

[0015] , ;in, To fine-tune the updated model weight matrix; This is the pre-trained weight matrix; This is the scaling factor for the static physical flow; This is the low-rank increment matrix of the static physical flow; This is the upgraded matrix of the static physical flow; This is the dimension reduction matrix of the static physical flow; This is the scaling factor for the dynamic environment flow; This is a low-rank increment matrix for dynamic environmental flow; This is an upgraded matrix representing the dynamic environment flow; This is a dimension-reduced matrix for the dynamic environment flow; For output features; For input features; For adaptive gating coefficients; For activation functions; Here is the weight matrix of the gated network; For system state context;

[0016] Based on pre-collected community data from the same climate zone, the original collaborative control model is transferred and fine-tuned using the incremental fine-tuning mechanism to obtain a LoRA prototype library; the LoRA prototype library includes: several static LoRA prototypes and dynamic LoRA prototypes;

[0017] Based on the static feature vector and climate feature vector of the target community, the prototypes in the LoRA prototype library are calculated and screened for cosine similarity to obtain initialization parameters. The initialization parameters are then assembled in a modular fashion to obtain the multi-level heating coordinated control model.

[0018] Preferably, the feature representation of the graph structure embedding is as follows: ;in, For nodes No. Layer feature representation; It is a non-linear activation function; To learn the weight matrix; This is a concatenation function; It is an aggregate function; for Nodes within; For nodes The neighboring nodes.

[0019] Preferably, the expression for the first constraint is:

[0020] ;in, ; Attention value; , , These are the query matrix, key matrix, and value matrix, respectively. For dimensions; This is a thermodynamic attention mask matrix; These are elements in the thermodynamic attention mask matrix; and They represent the first and second elements in the input sequence, respectively. The and the first A point in time; This refers to the physical transmission delay.

[0021] Preferably, the expression for the second constraint is:

[0022] ;in, This is the total loss function; To control the mean square error of the command; This is the penalty coefficient for physical constraints; Specific heat capacity; This refers to the water flow rate in the pipe network. For water supply temperature; The return water temperature; For the building's heat load; This is an estimated value for heat loss in the pipeline network.

[0023] Preferably, the expression for the cosine similarity includes:

[0024] ;in, For optimal static matching index; This represents the static feature vector of the newly accessed cell; This is the feature vector of the kth static prototype in the prototype library; For optimal dynamic matching index; This represents the climate feature vector of the newly accessed location; For the prototype library The feature vector of a dynamic prototype.

[0025] Preferably, an electronic device includes: at least one processor and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to execute the aforementioned multi-level heating coordinated control method based on a physical sensing large model and dual-flow decoupling LoRA.

[0026] Preferably, a non-transient computer-readable storage medium stores computer instructions for causing a computer to execute the aforementioned multi-level heating coordinated control method based on a physical sensing large model and dual-flow decoupling LoRA.

[0027] The present invention discloses the following technical effects:

[0028] This invention provides a multi-level heating coordinated control method, electronic device, and storage medium based on a physical sensing large model and dual-flow decoupled LoRA. By using an incremental fine-tuning mechanism based on dual-flow decoupled LoRA and establishing a LoRA prototype library based on the dual-flow decoupled architecture, it solves the problems of weak model generalization ability and high transfer cost of existing methods. It realizes decoupled learning and dynamic fusion of physical attributes and environmental strategies, as well as rapid deployment under zero-sample or few-sample conditions. Through the first and second constraints, it solves the problem of safety hazards in the direct application of existing methods and realizes that the constraint model outputs control commands that conform to physical laws. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 A schematic diagram of a multi-level heating coordinated control process based on a physical sensing large model and dual-flow decoupling LoRA provided in an embodiment of the present invention;

[0031] Figure 2 This is a schematic diagram of the overall technical architecture provided for an embodiment of the present invention;

[0032] Figure 3 This is a schematic diagram of the dual-stream decoupling LoRA incremental fine-tuning mechanism provided in an embodiment of the present invention.

[0033] Figure 4 This is a schematic diagram of the meta-learning fast transfer method provided in an embodiment of the present invention. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] The purpose of this invention is to provide a multi-level heating coordinated control method, electronic device and storage medium based on a physical sensing large model and dual-flow decoupled LoRA, which solves the problems of existing methods such as difficulty in unified modeling, high migration cost and safety hazards in direct application.

[0036] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0037] Figure 1 This is a schematic diagram of a multi-level heating coordinated control process based on a large physical sensing model and dual-flow decoupled LoRA provided in an embodiment of the present invention, as shown below. Figure 1 As shown, this invention provides a multi-level coordinated heating control method based on a large physical sensing model and dual-flow decoupled LoRA, including:

[0038] Step 100: Collect multi-source operational data of the target community, input the multi-source operational data into a pre-trained multi-level heating coordinated control model to generate control commands, and regulate the heating system of the target community according to the control commands; the training process of the multi-level heating coordinated control model includes:

[0039] Step 101: Encode the topology data of the thermal pipeline network using a graph neural network to obtain the graph structure embedding;

[0040] Step 102: Perform time-series slice encoding and one-dimensional convolution on the high-frequency time-series signal to obtain the time-series token embedding;

[0041] Step 103: Perform linear projection and splicing operations on the graph structure embedding, the temporal token embedding, and the discrete static attribute embedding to obtain the input sequence;

[0042] Step 104: Embed the first layer of physical causality-based constraints in the self-attention computation layer of the Transformer model;

[0043] Step 105: Embed a second constraint based on the law of energy conservation and the mean square error of the control command at the output of the Transformer model;

[0044] Step 106: Based on the input sequence, the Transformer model is iteratively trained using an incremental fine-tuning mechanism based on two-stream decoupling LoRA according to the first constraint and the second constraint to obtain the original collaborative control model; the expression of the incremental fine-tuning mechanism includes:

[0045] ,

[0046] , ;in, To fine-tune the updated model weight matrix; This is the pre-trained weight matrix; This is the scaling factor for the static physical flow; This is the low-rank increment matrix of the static physical flow; This is the upgraded matrix of the static physical flow; This is the dimension reduction matrix of the static physical flow; This is the scaling factor for the dynamic environment flow; This is a low-rank increment matrix for dynamic environmental flow; This is an upgraded matrix representing the dynamic environment flow; This is a dimension-reduced matrix for the dynamic environment flow; For output features; For input features; For adaptive gating coefficients; For activation functions; Here is the weight matrix of the gated network; For system state context;

[0047] Step 107: Based on the pre-collected cell data of the same climate zone, the original collaborative control model is transferred and fine-tuned using the incremental fine-tuning mechanism to obtain the LoRA prototype library; the LoRA prototype library includes: several static LoRA prototypes and dynamic LoRA prototypes;

[0048] Step 108: Based on the static feature vector and climate feature vector of the target community, perform cosine similarity calculation and screening on the prototypes in the LoRA prototype library to obtain initialization parameters, and assemble the initialization parameters in a modular manner to obtain the multi-level heating coordinated control model.

[0049] Specifically, in combination Figure 2 This document describes the implementation process of the overall technical solution. This embodiment proposes an end-to-end control architecture based on global topology awareness, dual-stream decoupled reasoning, and physical boundary constraints. The core idea is to transform the heating control problem into a physically constrained multimodal sequence generation problem, utilizing the semantic understanding and reasoning capabilities of a large model to solve decision-making challenges under complex operating conditions. The specific implementation process includes the following key steps:

[0050] 1) Graph-temporal semantic alignment of multi-source heterogeneous data:

[0051] To address the dimensional misalignment issue between time-series data and pipeline topology in heating systems, this embodiment designs a multimodal alignment encoder to map heterogeneous signals from the physical world to the high-dimensional semantic space of a large model. First, for the topological data of the heating pipeline network, the system constructs a graph neural network (GNN) encoding module. Heat exchange stations, valves, and user terminals are defined as nodes in the graph. Heating pipes are defined as side Construct a topology graph To capture the spatial transfer and hysteresis characteristics of heat in the pipe network, a message passing mechanism is used to aggregate node features. For any node... , its first Layer feature representation By aggregating its neighbor nodes The information is updated, and the calculation formula is as follows:

[0052]

[0053] in, It is a non-linear activation function. For learnable weight matrix, For aggregation functions (such as mean or max pooling). After... After layer aggregation, the generated node embedding vectors can explicitly encode the thermal inertia and resistance distribution characteristics of the pipeline network.

[0054] For high-frequency time-series signals such as weather forecasts and historical operating data of heating systems, a time-series patching encoding mechanism is adopted. This involves encoding continuous time series data... The time slice is divided into multiple non-overlapping time slices and mapped to serialized token embeddings through a one-dimensional convolutional layer.

[0055] Through a linear projection layer, the aforementioned graph structure embedding, temporal token embedding, and discrete static attribute embedding are uniformly mapped to the input dimension of the large model backbone network. The concatenation operation forms an input sequence containing complete spatiotemporal physical semantics. .

[0056] 2) Incremental fine-tuning mechanism based on two-stream decoupled LoRA:

[0057] refer to Figure 3 To address the challenge of migrating the model across different neighborhoods and climate zones, this embodiment proposes a dual-flow decoupled low-rank adaptation mechanism. This mechanism abandons the traditional LoRA approach of mixing and updating all knowledge, instead basing its implementation on the physical characteristics of the heating system and adjusting the large model weight matrix... Update increment Mathematically, this is decomposed into two orthogonal low-rank subspaces: the static physical flow and the dynamic environment flow. Specifically, the weight update process is formalized as follows:

[0058]

[0059] Static physical flow This branch is specifically designed for learning time-independent intrinsic physical properties, such as the heat loss coefficient of pipe networks and the heat transfer coefficient of building envelopes. Parameter updates in this section are triggered only by static property inputs. Dynamic environmental flow. This branch is specifically designed to learn the system's dynamic response strategies to changes in the external environment, such as the increase in water supply temperature when the outdoor temperature drops by 1°C. Parameter updates in this section are primarily driven by time-series meteorological data.

[0060] In the inference phase, an adaptive gating coefficient is introduced to address the differences in policy dependence under different operating conditions. This coefficient is determined by the current system state context. Real-time calculation Used to dynamically adjust the fusion weights of the two data streams:

[0061]

[0062] With this decoupled design, when the model is migrated to a new cell in the same climate zone, only the dynamic environmental flow parameters need to be retained, and the static physical flow parameters need to be retrained, thereby greatly reducing the dependence on training data.

[0063] 3) A physical sensing backbone network embedded with the laws of thermodynamics:

[0064] To prevent large models from generating illusory instructions that violate physical laws, this embodiment introduces a dual physical constraint mechanism both inside the Transformer backbone network and at the output:

[0065] The first constraint: Thermodynamic attention mask. In the Transformer's self-attention computation layer, a mask matrix based on physical causality is constructed. The shortest transmission time of the heat medium from the heat source to each user node is calculated based on the principles of fluid mechanics. In calculating attention scores If two time steps and The interval is less than the physical transmission delay Then its attention weight is reset to negative infinity:

[0066]

[0067]

[0068] in, Attention value; , , These are the query matrix, key matrix, and value matrix, respectively. For dimensions; This is a thermodynamic attention mask matrix; These are elements in the thermodynamic attention mask matrix; and They represent the first and second elements in the input sequence, respectively. The and the first A point in time; This refers to the physical transmission delay.

[0069] This ensures, at the network structure level, that the model will only focus on historical information that is physically causally related.

[0070] The second constraint: the physical consistency loss function, in the objective function of model training, besides the usual mean squared error of control commands ( In addition to the above, a physical residual term based on the law of conservation of energy has been added. The system calculates the residual based on the predicted water supply temperature. Return water temperature and traffic Calculate the total heat supply and require it to be compared with the building's heat load. Maintain balance. The total loss function is defined as:

[0071]

[0072] in, This is the total loss function; To control the mean square error of the command; This is the penalty coefficient for physical constraints; Specific heat capacity; This refers to the water flow rate in the pipe network. For water supply temperature; The return water temperature; For the building's heat load; This is an estimated value for heat loss in the pipeline network.

[0073] 4) Meta-learning transfer strategy based on decoupled prototypes:

[0074] refer to Figure 4 To address the cold start challenge of newly accessed cells, this embodiment utilizes the aforementioned decoupled architecture to establish a rapid migration process based on meta-learning. The system pre-builds a LoRA prototype library, containing... A static LoRA prototype with typical building physics characteristics and Dynamic LoRA prototypes with typical climatic characteristics When a new cell is connected, the system first extracts its static feature vector. and the local climate feature vector The best matching prototype is retrieved by calculating cosine similarity.

[0075]

[0076] in, For optimal static matching index; This represents the static feature vector of the newly accessed cell; This is the feature vector of the kth static prototype in the prototype library; For optimal dynamic matching index; This represents the climate feature vector of the newly accessed location; For the prototype library The feature vector of a dynamic prototype.

[0077] Then, the static model parameters are loaded directly. and dynamic model parameters This serves as the initialization parameter for the new community. This modular assembly strategy enables the model to have basic control capabilities even without historical operating data, and then only requires fine-tuning the gating network to achieve full adaptation.

[0078] As an optional implementation, this embodiment also provides an electronic device, including: at least one processor, and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to execute the aforementioned multi-level heating coordinated control method based on a physical sensing large model and dual-flow decoupling LoRA.

[0079] As an optional implementation, this embodiment also provides a non-transient computer-readable storage medium storing computer instructions for causing a computer to execute the aforementioned multi-level heating coordinated control method based on a physical sensing large model and dual-flow decoupling LoRA.

[0080] The beneficial effects of this invention are as follows:

[0081] This invention utilizes an incremental fine-tuning mechanism based on dual-flow decoupling LoRA to decompose the fine-tuning parameter matrix of a large model into two orthogonal low-rank matrix branches: a static physical flow and a dynamic environmental flow. It also introduces adaptive gating coefficients based on the system state context, achieving decoupled learning and dynamic fusion of physical attributes and environmental policies. Through first and second constraints, the model outputs control commands that conform to physical laws. Furthermore, the LoRA prototype library enables rapid deployment under zero-sample or few-sample conditions.

[0082] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0083] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A multi-stage coordinated heating control method based on a large-scale physical sensing model and dual-flow decoupled LoRA, characterized in that, include: Multi-source operation data of the target community is collected, and the multi-source operation data is input into a pre-trained multi-level heating coordinated control model to generate instructions, thereby obtaining control instructions, and the heating system of the target community is controlled according to the control instructions. The training process of the multi-level heating coordinated control model includes: The graph neural network is used to encode the topological data of the thermal pipeline network to obtain the graph structure embedding; Time-series slicing encoding and one-dimensional convolution are performed on high-frequency time-series signals to obtain time-series token embeddings; Linear projection and concatenation operations are performed on the graph structure embedding, the temporal token embedding, and the discrete static attribute embedding to obtain the input sequence; Embed the first layer of physical causality-based constraints in the self-attention computation layer of the Transformer model; A second constraint based on the law of energy conservation and the mean square error of the control command is embedded at the output of the Transformer model. Based on the input sequence, the Transformer model is iteratively trained using an incremental fine-tuning mechanism based on two-stream decoupling LoRA according to the first and second constraints, to obtain the original collaborative regulation model; the expression of the incremental fine-tuning mechanism includes: 、 , ;in, To fine-tune the updated model weight matrix; This is the pre-trained weight matrix; This is the scaling factor for the static physical flow; This is the low-rank increment matrix of the static physical flow; This is the upgraded matrix of the static physical flow; This is the dimension reduction matrix of the static physical flow; This is the scaling factor for the dynamic environment flow; This is a low-rank increment matrix for dynamic environmental flow; This is an upgraded matrix representing the dynamic environment flow; This is a dimension-reduced matrix for the dynamic environment flow; For output features; Input features; For adaptive gating coefficients; For activation functions; Here is the weight matrix of the gated network; For system state context; Based on pre-collected community data from the same climate zone, the original collaborative control model is transferred and fine-tuned using the incremental fine-tuning mechanism to obtain a LoRA prototype library; the LoRA prototype library includes: several static LoRA prototypes and dynamic LoRA prototypes; Based on the static feature vector and climate feature vector of the target community, the prototypes in the LoRA prototype library are calculated and screened for cosine similarity to obtain initialization parameters. The initialization parameters are then assembled in a modular fashion to obtain the multi-level heating coordinated control model.

2. The multi-stage coordinated heating control method based on a large-scale physical sensing model and dual-flow decoupling LoRA as described in claim 1, characterized in that, The feature representation of the graph structure embedding is as follows: ;in, For nodes No. Layer feature representation; It is a non-linear activation function; To learn the weight matrix; This is a concatenation function; It is an aggregate function; for Nodes within; For nodes The neighboring nodes.

3. The multi-stage coordinated heating control method based on a large-scale physical sensing model and dual-flow decoupling LoRA as described in claim 1, characterized in that, The expression for the first level of constraint is: ;in, ; Attention value; , , These are the query matrix, key matrix, and value matrix, respectively. For dimensions; This is a thermodynamic attention mask matrix; These are elements in the thermodynamic attention mask matrix; and They represent the first and second digits of the input sequence, respectively. The and the first A point in time; This refers to the physical transmission delay.

4. The multi-stage coordinated heating control method based on a large physical sensing model and dual-flow decoupled LoRA as described in claim 1, characterized in that, The expression for the second constraint is: ;in, This is the total loss function; To control the mean square error of the command; This is the penalty coefficient for physical constraints; Specific heat capacity; This refers to the water flow rate in the pipe network. For water supply temperature; The return water temperature; For the building's heat load; This is an estimated value for heat loss in the pipeline network.

5. The multi-stage coordinated heating control method based on a large physical sensing model and dual-flow decoupled LoRA as described in claim 1, characterized in that, The expression for the cosine similarity includes: ;in, For optimal static matching index; This represents the static feature vector of the newly accessed cell; This is the feature vector of the kth static prototype in the prototype library; For optimal dynamic matching index; This represents the climate feature vector of the newly accessed location; For the prototype library The feature vector of a dynamic prototype.

6. An electronic device, characterized in that, include: At least one processor, and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to perform a multi-level heating coordinated control method based on a physical sensing large model and dual-flow decoupled LoRA, as described in any one of claims 1 to 5.

7. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to execute the multi-level heating coordinated control method based on a physical sensing large model and dual-flow decoupled LoRA as described in any one of claims 1 to 5.

Images