Terminal air conditioner regulation and control method based on cross-modal large model

Abstract

The invention provides a terminal air conditioner regulation and control method based on a cross-modal large model, and the method comprises the steps: constructing an environment data set, an expert experience data set and an air conditioner strategy data set, and carrying out the correlation modeling, and obtaining multi-modal data containing a physical parameter vector, a text vector and a knowledge graph relation; expanding the open source cross-modal model to a third modal, mapping numerical parameters to a vector space through a full connection layer, and sharing a cross-modal attention mechanism with a text; the method comprises the following steps: training multi-modal data input through a cross-modal large model, introducing a thermodynamic loss function, limiting a fresh air opening value, and carrying out dynamic fine tuning by adopting a low-rank adapter technology to obtain a trained model; and real-time operation data and text logs of the air conditioning system are input into the trained model for multi-modal reasoning, and a real-time control strategy giving consideration to energy efficiency and equipment health is generated based on model output. Through a cross-modal fusion method, physical constraint enhancement AI and a dynamic parameter fine tuning mechanism, deep fusion of numerical values and texts is realized, domain knowledge is embedded into the model through loss function design, the reasonability of model output is ensured, energy consumption control extends to equipment health management, and a sustainable air conditioning system optimization framework is constructed.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent building energy conservation, and in particular to a terminal air-conditioning control method based on a cross-modal large model. Background Art

[0002] As the core unit of building energy consumption, the control strategy of modern air conditioning system directly affects energy efficiency and indoor environmental quality. The typical system realizes temperature and humidity control through fresh air-return air mixing regulation, where the fresh air enthalpy (h n ), return air enthalpy (h r ) and air supply enthalpy (Δh) are key parameters for measuring the thermodynamic energy of air. However, multivariable coupling, time-varying interference, and nonlinear heat transfer processes make it difficult for traditional control methods to achieve dynamic optimal regulation.

[0003] The existing technology has the following defects: overfitting and experience dependence of data-driven methods. Current data-driven algorithms (such as SVM and random forest) are easily affected by high-dimensional sparse data in terminal air conditioning energy consumption prediction, resulting in overfitting; and the control strategy that relies on artificial experience values ​​is prone to unreasonable lower control values. The nonlinear coupling defect of traditional PID control. The standard PID controller cannot decouple the fresh air enthalpy value (h n ), return air enthalpy (h r ) and the dynamic relationship between the air supply enthalpy difference (Δh). The spatiotemporal modeling limitations of AI models. Although the existing LSTM model can process numerical sensor data, it cannot integrate text-based maintenance logs for joint optimization. The strategy based on the rule engine lacks the correlation modeling of long-term factors such as equipment aging rate and spatial load distribution, resulting in rising energy consumption under delayed maintenance. The dynamic conflict between energy saving and comfort. The preset strategy for a fixed time period cannot adapt to the two scenarios of transition season and high-density crowd scene. When the outdoor enthalpy value in the transition season is appropriate, it still operates in mechanical circulation mode, resulting in energy redundancy; the high-density crowd scene makes CO 2 Short-term concentration exceeding the standard forces the fan to run at overclocked frequency, causing noise complaints and shortening of its life. Summary of the invention

[0004] The purpose of the present invention is to provide a terminal air conditioning control method based on a cross-modal large model to solve the deficiencies of the above-mentioned prior art.

[0005] To achieve the above object, the present invention adopts the following technical solution:

[0006] A terminal air conditioning control method based on a cross-modal large model includes the following steps:

[0007] S1. Obtain the operating parameters, equipment logs and expert strategies of the air-conditioning system, and construct an environmental data set, an expert experience data set and an air-conditioning strategy data set;

[0008] S2, the environment data set, expert experience data set and air conditioning strategy data set are analyzed through the semantic engine for unstructured data, and the domain knowledge is aligned using the knowledge graph embedding method to obtain multimodal data containing physical parameter vectors, text vectors and knowledge graph relationships;

[0009] S3, based on the cross-modal large model deepseekJanus-Pro-7B, expands the third modality of numerical data on the basis of its image-text dual modality, and adds a fully connected layer to map the running parameters to the vector space, sharing the same cross-modal attention layer with the text features;

[0010] S4. Input multimodal data into the cross-modal large model for training, learn the association between modalities through the cross-modal attention mechanism, introduce the thermodynamic loss function and limit the fresh air opening value during the training process, and use the low-rank adapter technology for dynamic fine-tuning to obtain a trained model;

[0011] S5. Input the real-time operation data and text logs of the air-conditioning system into the trained model for multimodal reasoning, and generate a real-time control strategy that takes into account both energy efficiency and equipment health based on the model output.

[0012] Furthermore, the environmental data set is specifically constructed by the following method: obtaining real-time sensor data in the air conditioning system, including but not limited to the temperature and humidity of the fresh air outlet, return air outlet, and supply air outlet, CO 2 concentration, valve opening and fan frequency; calculate the fresh air enthalpy value h through the enthalpy formula n , return air enthalpy value h r and the supply air enthalpy Δh;

[0013] Data cleaning is performed using methods including but not limited to threshold filtering, outlier removal, and missing value filling; an environmental data set is constructed.

[0014] Furthermore, the expert experience dataset is specifically constructed by the following method: obtaining text data of equipment operation logs, repair reports, inspections, maintenance work orders, alarm records and expert strategy libraries; using natural language processing tools NLP to encode text into semantic vectors and extract key features; and constructing an expert experience dataset.

[0015] Furthermore, the air conditioning strategy data set is specifically constructed by the following method: the air conditioning strategy data set is obtained by combining manual adjustment records of operation and maintenance personnel with expert strategy labels.

[0016] Furthermore, the step S2 specifically includes the following method implementation:

[0017] S21, semantic engine processing, using NLP technology to extract key entities and relationships based on the data set;

[0018] S22. Build a domain knowledge graph, with key entities as nodes and relationships as edges;

[0019] S23, embedding alignment, maps numerical parameters into vectors, aligns them with text entity vectors in a shared semantic space, and makes related nodes close in distance in the vector space through graph neural networks or TransE algorithms.

[0020] Furthermore, step S3 is specifically implemented by the following method:

[0021] S31. Based on the existing image-text dual modality of the base model DeepSeekJanus-Pro-7B, a new numerical sensor data modality is added;

[0022] S32, add a fully connected layer, map the 10-dimensional sensor data to a 768-dimensional vector space through the fully connected layer, and obtain a numerical vector;

[0023] This step is expressed by the following formula (1):

[0024] v num =ReLU(W·x sensor +b) (1);

[0025] In formula (1): x sensor represents the normalized sensor data of the input; W represents the weight matrix, which is a trainable parameter and is used to linearly transform the 10-dimensional input into a 768-dimensional vector, which is consistent with the text embedding dimension; b represents the bias vector, which is a trainable parameter; ReLU represents the activation function; v num a numeric vector representing the output;

[0026] S33, inputting the numerical vector and the text vector into the attention mechanism together, and sharing the same cross-modal attention layer with the text vector;

[0027] This step is expressed by the following formula (2):

[0028]

[0029] In formula (2), Q represents the query matrix, which is composed of the numerical mode vector v num Generate; K represents the key matrix, which is generated from the text vector; V represents the value matrix, which is homologous to K and undergoes independent linear transformation; d k Indicates the dimension of the key vector, usually d k =768; Softmax represents the normalization function, which normalizes the attention weights by row; Attention represents the cross-modal attention layer.

[0030] Furthermore, the step S4 specifically includes the following method implementation:

[0031] S41. During the training process, the thermodynamic loss function is introduced to penalize the prediction that deviates from the enthalpy equation, and the Sigmoid function is applied in the output layer to limit the fresh air opening β∈[0,1] to ensure that the model output conforms to the laws of thermodynamics.

[0032] The thermodynamic loss function is expressed by the following formula (3):

[0033] L=MSE(Δh 预测 ,Δh 实际 )+α·Penalty(β 预测 ) (3);

[0034] In formula (3): Δh 预测 Indicates the actual air supply enthalpy value; Δh 预测 represents the air supply enthalpy value predicted by the model; MSE represents the mean square error; β 实际 It represents the fresh air opening ratio output by the model; Penalty represents the physical constraint penalty term to prevent non-physical solution, Penalty=max(0,-β)+max(0,β-1); α represents the penalty term weight, which is a hyperparameter and can be adjusted to meet the needs of different scenarios; L represents the total loss;

[0035] S42. Use low-rank adapter technology for dynamic fine-tuning, freeze the original weight matrix, insert trainable low-rank matrices A and B for incremental training; when the error (Δh, β) ≤ 2%, merge the adapter parameters and release resources;

[0036] S43. Get the trained model.

[0037] Furthermore, the real-time control strategies based on the model output that take into account both energy efficiency and equipment health include:

[0038] Multimodal reasoning: Based on the input real-time data and text logs, the model matches key features through a cross-modal attention mechanism and outputs the fresh air opening degree β and control strategy;

[0039] Correction of physical rules: check whether the fresh air opening β conforms to the enthalpy equation. If the predicted air supply enthalpy Δh deviates from the actual value, adjust the fresh air opening β in the opposite direction to a reasonable range.

[0040] Equipment health management, monitoring equipment status and dynamically adjusting strategies, as well as generating maintenance reminders;

[0041] Energy-saving strategies are applied to identify high outdoor enthalpy values ​​in summer, close the fresh air to the minimum opening, and give priority to the use of return air; in the transition season, low outdoor enthalpy air is used to maximize the proportion of fresh air.

[0042] It can be seen from the above technical solutions that the present invention has the following advantages compared with the prior art:

[0043] (1) The present invention expands the third modality based on the cross-modal large model, realizes the deep association between physical parameters and text semantics, and improves the generalization ability of complex scenes;

[0044] (2) The present invention introduces a thermodynamic loss function during model training to force the model output to conform to the enthalpy equation and physical boundaries, thereby ensuring the feasibility of the strategy;

[0045] (3) The present invention adopts the dynamic adaptation mechanism and flexible freezing strategy of LoRA dynamic adaptation technology during the model training process to reduce computing power requirements and response delays, and realize rapid adaptation of the model in scenarios such as device degradation and environmental mutation;

[0046] (4) The present invention can dynamically adjust strategies and generate decline warnings by analyzing the decline trend of fan efficiency and the semantic features of maintenance records, thereby balancing energy-saving goals and equipment health and extending equipment life;

[0047] (5) The model and method described in the present invention can dynamically identify outdoor enthalpy values, minimize fresh air load in summer, maximize free cooling in transition seasons, and achieve systematic energy saving. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a schematic diagram of the steps of the terminal air conditioning control method based on the cross-modal large model of the present invention. DETAILED DESCRIPTION

[0049] A preferred embodiment of the present invention is described in detail below with reference to the accompanying drawings.

[0050] like Figure 1 The terminal air conditioning control method based on the cross-modal large model shown includes the following steps:

[0051] S1. Obtain the operating parameters, equipment logs and expert strategies of the air-conditioning system, and build an environmental data set, an expert experience data set and an air-conditioning strategy data set.

[0052] In the specific operation, the environmental data set (value) described in this step is specifically constructed by the following method: obtaining real-time sensor data in the air conditioning system, including but not limited to the temperature and humidity of the fresh air outlet, return air outlet, and supply air outlet, CO 2 concentration, valve opening and fan frequency; calculate the fresh air enthalpy value h through the enthalpy formula n , return air enthalpy value h r And the supply air enthalpy value Δh; use methods including but not limited to threshold filtering, outlier removal, and missing value filling to clean the data; and construct an environmental data set.

[0053] The expert experience data set (text) described in this step is specifically constructed by the following method: obtaining text data of equipment operation logs, repair reports, inspections, maintenance work orders (such as "fresh air valve stuck"), alarm records (such as "low fan efficiency"), and expert strategy libraries; using natural language processing tools NLP to encode text into semantic vectors and extract key features; and constructing an expert experience data set.

[0054] The air conditioning strategy data set described in this step is specifically constructed by the following method: combining the manual adjustment records of operation and maintenance personnel with expert strategy labels (such as "transition season fresh air ratio ≥ 60%") to construct an air conditioning strategy data set.

[0055] S2. The environmental data set, expert experience data set, and air conditioning strategy data set are parsed through a semantic engine for unstructured data, and the domain knowledge is aligned using the knowledge graph embedding method to obtain multimodal data containing physical parameter vectors, text vectors, and knowledge graph relationships.

[0056] Specifically, this step includes the following methods:

[0057] S21, semantic engine processing, based on the data set, uses NLP technology (such as BERT, knowledge graph entity recognition) to extract key entities and relationships, such as from the log "Afternoon meeting room CO 2 Exceeding the standard, insufficient opening of the fresh air valve" extracts the entity [Conference room, CO 2 , fresh air valve] and the relationship [exceeding the standard → insufficient opening];

[0058] S22. Build a domain knowledge graph, with key entities as nodes, such as physical parameters (such as Δh), equipment components (such as fresh air valves), events (such as "maintenance"), and relationships as edges (such as "influence", "cause", "need adjustment");

[0059] S23, embedding alignment, the numerical parameters (such as CO 2 = 1200ppm) is mapped to a vector and compared with the text entity vector (such as "CO 2 Exceeding the standard”) are aligned in the shared semantic space, and the graph neural network or TransE algorithm is used to make the related nodes close in the vector space (such as “CO 2 ” and “crowd density”).

[0060] This step S2 realizes the two-way translation of data and knowledge through physical-semantic joint representation, so that the output of the model conforms to the physical laws and operation and maintenance experience, avoiding absurd decisions; through the association rules in the knowledge graph, the model can handle edge scenarios that have not been seen in training; the decision-making process can be traced back to the entities and relationships in the knowledge graph, meeting the needs of industrial scenarios for AI transparency.

[0061] S3. Based on the cross-modal large model deepseekJanus-Pro-7B, the third modality of numerical data is expanded on the basis of its image-text dual modality, and a fully connected layer is added to map the running parameters to the vector space, sharing the same cross-modal attention layer with the text features.

[0062] Specifically, this step is implemented by the following method:

[0063] S31. Based on the existing image-text dual modality of the base model DeepSeekJanus-Pro-7B, a new numerical sensor data modality is added;

[0064] S32, add a fully connected layer (FCN), map the 10-dimensional sensor data to a 768-dimensional vector space through the fully connected layer, and obtain a numerical vector;

[0065] Numerical data (such as temperature = 30°C, CO 2 =1200ppm) is low-dimensional structured data, which needs to be mapped to a high-dimensional vector space to match the semantic dimension of the text modality; this step is represented by the following formula (1):

[0066] v num =ReLU(W·x sensor +b) (1);

[0067] In formula (1): x sensor represents the normalized sensor data of the input; W represents the weight matrix, which is a trainable parameter and is used to linearly transform the 10-dimensional input into a 768-dimensional vector, which is consistent with the text embedding dimension; b represents the bias vector, which is a trainable parameter; ReLU represents the activation function; v num a numeric vector representing the output;

[0068] S33. Input the numerical vector and the text vector into the attention mechanism together, and share the same cross-modal attention layer with the text vector.

[0069] The original model only supports image-text bimodal attention, and needs to be expanded to trimodal interaction: retain the original text-image attention head, add value-text and value-image attention heads, and calculate the attention weights for the Q of the value modality and the K / V of the text modality to achieve semantic alignment. This step is expressed by the following formula (2):

[0070]

[0071] In formula (2), Q represents the query matrix, which is composed of the numerical mode vector v num Generate; K represents the key matrix, which is generated from the text vector; V represents the value matrix, which is homologous to K and undergoes independent linear transformation; d kIndicates the dimension of the key vector, usually d k =768; Softmax represents the normalization function, which normalizes the attention weights by row; Attention represents the cross-modal attention layer.

[0072] For example: Enter the value CO 2 =1200ppm→Q vector; input text "crowd gathering in conference room"→K / V vector; then high attention weight associates the two, triggering the output strategy "increase the fresh air opening to 50%".

[0073] S4. Input multimodal data into the cross-modal large model for training, learn the association between modalities through the cross-modal attention mechanism, introduce the thermodynamic loss function and limit the fresh air opening value during the training process, and use the low-rank adapter technology for dynamic fine-tuning to obtain a trained model.

[0074] This preferred embodiment aims to enable the model to learn the semantic association between numerical data and text while satisfying physical constraints; the step S4 specifically includes the following method implementation:

[0075] S41. During the training process, the thermodynamic loss function is introduced to penalize the prediction that deviates from the enthalpy equation, and the Sigmoid function is applied in the output layer to limit the fresh air opening β∈[0,1] to ensure that the model output conforms to the laws of thermodynamics.

[0076] The thermodynamic loss function is expressed by the following formula (3):

[0077] L=MSE(Δh 预测 ,Δh 实际 )+α·Penalty(β 预测 ) (3);

[0078] In formula (3): Δh 预测 Indicates the actual air supply enthalpy value; Δh 预测 represents the air supply enthalpy value predicted by the model; MSE represents the mean square error; β 实际 It represents the fresh air opening ratio output by the model; Penalty represents the physical constraint penalty term to prevent non-physical solution, Penalty=max(0,-β)+max(0,β-1); α represents the penalty term weight, which is a hyperparameter and can be adjusted to meet the needs of different scenarios; L represents the total loss;

[0079] S42. Use low-rank adapter technology for dynamic fine-tuning, freeze the original weight matrix, insert trainable low-rank matrices A and B for incremental training; when the error (Δh, β) ≤ 2%, merge the adapter parameters and release resources;

[0080] S43. Get the trained model.

[0081] This step shortens the distance between related modal vectors (such as the "high temperature" text and the 35℃ numerical vector) through cross-modal contrast loss, and combines the soft constraint of the thermodynamic loss function with the hard constraint on the value of the fresh air opening β, while performing efficient parameter fine-tuning. The air conditioning control strategy has both data-driven flexibility and physical rule reliability.

[0082] S5. Input the real-time operation data and text logs of the air-conditioning system into the trained model for multimodal reasoning, and generate a real-time control strategy that takes into account both energy efficiency and equipment health based on the model output.

[0083] Specifically, the real-time control strategies based on model output that take into account both energy efficiency and equipment health include:

[0084] Multimodal reasoning, based on real-time input data such as CO 2 =1200ppm,h n =47kj / kg) and text logs (such as "crowds gather in the conference room"). The model matches key features through the cross-modal attention mechanism and outputs the fresh air opening β and control strategy (such as "opening increased to 50%");

[0085] Correction of physical rules: check whether the fresh air opening β conforms to the enthalpy equation. If the predicted air supply enthalpy Δh deviates from the actual value, adjust the fresh air opening β in the opposite direction to a reasonable range.

[0086] Equipment health management: monitor equipment status (e.g., fan efficiency drops by 10%) and dynamically adjust strategies (e.g., limit the upper limit of opening to 8%), as well as generate maintenance reminders (e.g., "valve is stuck, need to shut down for maintenance");

[0087] Energy-saving strategies are applied to identify high outdoor enthalpy values ​​in summer, close the fresh air to the minimum opening, and give priority to the use of return air; in the transition season, low outdoor enthalpy air is used to maximize the proportion of fresh air.

[0088] The above-described embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A terminal air conditioning control method based on a cross-modal large model, characterized in that: The following steps are involved: S1. Obtain the operating parameters, equipment logs and expert strategies of the air-conditioning system, and construct an environmental data set, an expert experience data set and an air-conditioning strategy data set; S2, the environment dataset, expert experience dataset and air conditioning strategy dataset are parsed through the semantic engine for unstructured data, and the domain knowledge is aligned using the knowledge graph embedding method to obtain multimodal data containing physical parameter vectors, text vectors and knowledge graph relationships; S3, based on the cross-modal large model deepseekJanus-Pro-7B, expands the third modality of numerical data on the basis of its image-text dual modality, and adds a fully connected layer to map the running parameters to the vector space, sharing the same cross-modal attention layer with the text features; S4. Input multimodal data into the cross-modal large model for training, learn the association between modalities through the cross-modal attention mechanism, introduce the thermodynamic loss function and limit the fresh air opening value during the training process, and use the low-rank adapter technology for dynamic fine-tuning to obtain a trained model; S5. Input the real-time operation data and text logs of the air-conditioning system into the trained model for multimodal reasoning, and generate a real-time control strategy that takes into account both energy efficiency and equipment health based on the model output.

2. According to claim 1, a terminal air conditioning control method based on a cross-modal large model is characterized in that: The environmental data set is specifically constructed by the following method: obtaining real-time sensor data in the air conditioning system, including but not limited to the temperature and humidity, CO2 concentration, valve opening and fan frequency of the fresh air inlet, return air inlet and supply air inlet; calculating the fresh air enthalpy value h by the enthalpy formula n , return air enthalpy value h r And the supply air enthalpy value Δh; use methods including but not limited to threshold filtering, outlier removal, and missing value filling to clean the data; and construct an environmental data set.

3. The terminal air conditioning control method based on a cross-modal large model according to claim 1 is characterized in that: The expert experience dataset is specifically constructed by the following method: obtaining text data of equipment operation logs, repair reports, inspections, maintenance work orders, alarm records and expert strategy libraries; using natural language processing tools NLP to encode text into semantic vectors and extract key features; and constructing an expert experience dataset.

4. The terminal air conditioning control method based on a cross-modal large model according to claim 1 is characterized in that: The air conditioning strategy data set is specifically constructed by the following method: combining manual adjustment records of operation and maintenance personnel with expert strategy labels to construct the air conditioning strategy data set.

5. The terminal air conditioning control method based on a cross-modal large model according to claim 1 is characterized in that: The step S2 specifically includes the following method implementation: S21, semantic engine processing, using NLP technology to extract key entities and relationships based on the data set; S22. Build a domain knowledge graph, with key entities as nodes and relationships as edges; S23, embedding alignment, maps numerical parameters into vectors, aligns them with text entity vectors in a shared semantic space, and makes related nodes close in distance in the vector space through graph neural networks or TransE algorithms.

6. The terminal air conditioning control method based on a cross-modal large model according to claim 1 is characterized in that: The step S3 is specifically implemented by the following method: S31. Based on the existing image-text dual modality of the base model DeepSeekJanus-Pro-7B, a new numerical sensor data modality is added; S32, add a fully connected layer, map the 10-dimensional sensor data to a 768-dimensional vector space through the fully connected layer, and obtain a numerical vector; This step is expressed by the following formula (1): v num =ReLU(W·x sensor +b) (1)? In formula (1): x sensor represents the normalized sensor data of the input; W represents the weight matrix, which is a trainable parameter and is used to linearly transform the 10-dimensional input into a 768-dimensional vector, which is consistent with the text embedding dimension; b represents the bias vector, which is a trainable parameter; ReLU represents the activation function; v num a numeric vector representing the output; S33, inputting the numerical vector and the text vector into the attention mechanism together, and sharing the same cross-modal attention layer with the text vector; This step is expressed by the following formula (2): In formula (2), Q represents the query matrix, which is composed of the numerical mode vector v num Generate; K represents the key matrix, which is generated from the text vector; V represents the value matrix, which is homologous to K and undergoes independent linear transformation; d k Indicates the dimension of the key vector, usually d k =768; Softmax represents the normalization function, which normalizes the attention weights by row; Attention represents the cross-modal attention layer.

7. The terminal air conditioning control method based on a cross-modal large model according to claim 1 is characterized in that: The step S4 specifically includes the following method implementation: S41. During the training process, the thermodynamic loss function is introduced to penalize the prediction that deviates from the enthalpy equation, and the Sigmoid function is applied in the output layer to limit the fresh air opening β∈[0,1] to ensure that the model output conforms to the laws of thermodynamics. The thermodynamic loss function is expressed by the following formula (3): L=MSE(Δh 预测 ,Dh 实际 )+α·Penalty(β 预测 ) (3); In formula (3): Δh 预测 Indicates the actual air supply enthalpy value; Δh 预测 represents the air supply enthalpy value predicted by the model; MSE represents the mean square error; β 实际 It represents the fresh air opening ratio output by the model; Penalty represents the physical constraint penalty term to prevent non-physical solution, Penalty=max(0,-β)+max(0,β-1); α represents the penalty term weight, which is a hyperparameter and can be adjusted to meet the needs of different scenarios; L represents the total loss; S42, use low-rank adapter technology for dynamic fine-tuning, freeze the original weight matrix, and insert trainable low-rank matrices A and B for incremental training; When the error (Δh, β) ≤ 2%, merge adapter parameters and release resources; S43. Get the trained model.

8. The terminal air conditioning control method based on a cross-modal large model according to claim 1 is characterized in that: Generating real-time control strategies that take into account both energy efficiency and equipment health based on model outputs includes: Multimodal reasoning: Based on the input real-time data and text logs, the model matches key features through a cross-modal attention mechanism and outputs the fresh air opening degree β and control strategy; Correction of physical rules: check whether the fresh air opening β conforms to the enthalpy equation. If the predicted air supply enthalpy Δh deviates from the actual value, adjust the fresh air opening β in the opposite direction to a reasonable range. Equipment health management, monitoring equipment status and dynamically adjusting strategies, as well as generating maintenance reminders; Energy-saving strategies are applied to identify high outdoor enthalpy values ​​in summer, close the fresh air to the minimum opening, and give priority to the use of return air; in the transition season, low outdoor enthalpy air is used to maximize the proportion of fresh air.

Images