A Central Air Conditioning Energy-Saving Control Method Based on AI Adaptive Adjustment

Abstract

This invention discloses an AI-based adaptive adjustment method for energy-saving control of central air conditioning systems, belonging to the field of intelligent control technology. The method includes: real-time acquisition of environmental data, equipment operation data, and energy consumption data from the central air conditioning system; preprocessing of multimodal data; constructing physical constraint equations based on thermodynamic laws to generate a multimodal dataset; generating multi-objective optimization control commands based on optimization weights; solving for the optimal combination of equipment parameters using the Pareto front algorithm; and transmitting the results to the central air conditioning actuators; collecting actual data after command execution; calculating the deviation between energy efficiency and comfort levels; and updating the physical information neural network parameters and adjusting the coefficients of the thermodynamic partial differential equations using a meta-learning framework. This invention improves the comprehensive control capability of central air conditioning systems in complex building environments by constructing a multimodal physical information fusion framework and a dynamic closed-loop optimization system.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and in particular to an energy-saving control method for central air conditioning based on AI adaptive adjustment. Background Technology

[0002] In recent years, the intelligent optimization control technology of central air conditioning has gradually developed from a single data-driven model to a multi-modal fusion approach. Traditional methods mainly rely on statistical learning of historical data or physical modeling based on thermodynamic equations, but both have significant limitations. Data-driven models lack explicit embedding of physical laws, causing prediction results to easily deviate from thermodynamic constraints under complex operating conditions; while pure physical models are difficult to adapt to dynamic environmental changes and have high computational complexity. With the rise of physical information neural networks, some studies have attempted to integrate partial differential equations as soft constraints into deep learning frameworks, but they still suffer from problems such as rigid weight allocation and insufficient causal reasoning in dynamic multi-objective optimization scenarios, failing to effectively balance the real-time conflict between energy saving and comfort.

[0003] The core deficiency of existing technologies lies in the insufficient integration of data-driven approaches and physical laws, which limits the effectiveness of multi-objective collaborative optimization. Physical constraints are superimposed posteriorly in the form of loss functions, making it difficult to guarantee that the prediction results strictly conform to thermodynamic equations; the lack of causal inference mechanisms means that the joint effects of key variables on energy consumption and comfort cannot be quantified, and dynamic weight allocation relies on empirical rules; the optimization model is updated lagging behind, making it unable to adapt online to dynamic disturbances such as equipment aging and sudden environmental changes. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an AI-based adaptive adjustment method for central air conditioning energy-saving control, which solves the problems of prediction bias and poor dynamic adaptability caused by insufficient integration of data-driven models and physical laws in the multi-objective optimization control of central air conditioning.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an AI-based adaptive adjustment method for energy-saving control of central air conditioning, which includes real-time collection of environmental data, equipment operation data and energy consumption data of central air conditioning, preprocessing of multimodal data, constructing physical constraint equations in combination with thermodynamic laws, and generating multimodal datasets.

[0008] By inputting a multimodal dataset into a physical information neural network, and using a deep learning model that embeds constraints from thermodynamic partial differential equations, the thermodynamic state and cooling load demand of buildings in future time periods are predicted, and the prediction results are output.

[0009] Based on the prediction results, a causal graph model is constructed to identify key variables affecting energy consumption and comfort, quantify the causal effects of key variables, and dynamically allocate optimization weights for energy saving and comfort.

[0010] Multi-objective optimization control commands are generated based on the optimization weights, the optimal combination of equipment parameters is solved using the Pareto front algorithm, and then transmitted to the central air conditioning actuator.

[0011] The system collects actual data after the command is executed, calculates the deviation between energy efficiency and comfort, updates the parameters of the physical information neural network through the meta-learning framework, and adjusts the coefficients of the thermodynamic partial differential equation.

[0012] As a preferred embodiment of the AI-based adaptive adjustment-based central air conditioning energy-saving control method of the present invention, the specific steps for generating the multimodal dataset are as follows:

[0013] The multimodal data is timestamped and the sampling frequency is normalized. A generative adversarial network is used to complete the missing data and output the preprocessed multimodal data.

[0014] Based on the equipment characteristics and thermodynamic laws of central air conditioning, physical constraint equations are constructed.

[0015] The preprocessed multimodal data is fused with physical constraint equations to generate a multimodal dataset containing physical laws.

[0016] As a preferred embodiment of the AI-based adaptive adjustment-based central air conditioning energy-saving control method of the present invention, the physical constraint equations include the heat conduction equation and the energy conservation equation.

[0017] As a preferred embodiment of the AI-based adaptive adjustment-based central air conditioning energy-saving control method of the present invention, the specific steps for outputting the prediction result are as follows:

[0018] The multimodal dataset is input into the spatiotemporal encoder and physical encoder of the physical information neural network to extract spatiotemporal features and generate implicit physical features;

[0019] The fusion layer of the input graph neural network integrates spatiotemporal features and implicit physical features, and uses a dynamic adaptive weight function to perform integral residual projection on the thermodynamic partial differential equations in the spatiotemporal domain to generate physically consistent spatiotemporal joint features.

[0020] Based on the spatiotemporal joint characteristics, the overall building cooling load demand forecast and the spatial distribution of the building's internal thermodynamic state are predicted at multiple scales. The overall building cooling load demand forecast is output through the fully connected layer, and the spatial distribution of the building's internal thermodynamic state is output through the U-Net structure.

[0021] The spatial distribution of predicted cooling load demand and thermodynamic state is input into the normalization mapping layer to generate prediction results.

[0022] As a preferred embodiment of the AI-based adaptive adjustment-based central air conditioning energy-saving control method of the present invention, the specific steps of dynamically allocating the optimization weights for energy saving and comfort are as follows:

[0023] Causal candidate variables are extracted from the prediction results, and a causal graph model is constructed using an improved PC algorithm;

[0024] Based on the causal graph model, key variables affecting energy consumption and comfort are identified. A dual machine learning method is used to calculate the joint causal effect coefficient of key variables on energy consumption and comfort. The distribution difference between historical data and real-time data is quantified by the KL divergence function.

[0025] Based on the joint causal effect coefficient and distribution difference, an information entropy-driven dynamic weight function is constructed to dynamically allocate the optimization weights for energy saving and comfort.

[0026] As a preferred embodiment of the AI-based adaptive adjustment-based central air conditioning energy-saving control method of the present invention, the key variable refers to the variable that has a causal impact on the energy consumption and comfort of the central air conditioning identified by the causal graph model;

[0027] The key variables include energy consumption-related variables, comfort-related variables, and environmental covariates.

[0028] As a preferred embodiment of the AI-based adaptive adjustment-based central air conditioning energy-saving control method of the present invention, the specific steps of transmitting data to the central air conditioning actuator are as follows:

[0029] The multi-objective optimization control command is decomposed into multiple sub-optimization problems, and each sub-problem is assigned dynamic weights and the device parameter search space is defined.

[0030] The Pareto front algorithm is used to solve the sub-optimization problem in parallel, and the non-dominated solution set that satisfies all constraints is selected to generate candidate equipment parameter combinations.

[0031] Based on the real-time load rate, select the optimal combination of device parameters from the candidate combinations of device parameters;

[0032] The optimal combination of equipment parameters is encoded into control commands compatible with the central air conditioning actuator protocol and transmitted to the central air conditioning actuator.

[0033] As a preferred embodiment of the AI-based adaptive adjustment-based central air conditioning energy-saving control method of the present invention, the specific steps of updating the physical information neural network parameters and adjusting the coefficients of the thermodynamic partial differential equation are as follows:

[0034] The system collects actual energy consumption data, temperature field distribution data, and equipment status data from central air conditioning actuators and sensors, and eliminates sensor noise through adaptive Kalman filtering.

[0035] Based on the filtered actual energy consumption data and the predicted cooling load demand, the dynamic energy-saving efficiency index is calculated, and the spatial weighted comfort deviation is calculated by the deviation between the temperature field distribution data and the preset value of the target temperature field distribution.

[0036] The dynamic energy efficiency index and comfort deviation are input into the meta-learning framework, and the parameters of the physical information neural network are updated through cross-task gradient descent, thereby simultaneously optimizing the coefficients of the thermodynamic partial differential equation.

[0037] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the AI-based adaptive adjustment central air conditioning energy-saving control method described in the first aspect of the present invention.

[0038] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the AI-based adaptive adjustment central air conditioning energy-saving control method as described in the first aspect of the present invention.

[0039] The beneficial effects of this invention are as follows: By constructing a multimodal physical information fusion framework and a dynamic closed-loop optimization system, the comprehensive control capability of central air conditioning in complex building environments is improved. Based on real-time acquisition of multi-source data and embedding of thermodynamic constraints, it can accurately predict the building's thermodynamic state and cooling load demand, ensuring that the prediction results strictly follow the laws of energy conservation and heat conduction. Through causal reasoning, the interaction effects between key variables are identified, the coupling relationship between energy consumption and comfort is dynamically quantified, and optimization weights are adaptively allocated to fully tap the energy-saving potential. Combined with the Pareto frontier algorithm, a globally optimal combination of equipment parameters is generated, improving the collaborative control efficiency of actuators such as compressors and dampers, and avoiding resource waste or performance imbalance caused by local optimization. Through a meta-learning-driven closed-loop update mechanism, it can perceive the deviation between actual operating data and prediction results in real time, autonomously adjust neural network parameters and thermodynamic equation coefficients, and continuously adapt to dynamic disturbances such as seasonal changes and equipment aging. Attached Figure Description

[0040] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0041] Figure 1 This is a flowchart of an AI-based adaptive adjustment method for central air conditioning energy-saving control.

[0042] Figure 2 This is a flowchart for multimodal data processing.

[0043] Figure 3 This is a flowchart of dynamic weight allocation and optimization control.

[0044] Figure 4 A flowchart for adaptive optimal device parameter updates. Detailed Implementation

[0045] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0046] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0047] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0048] Reference Figures 1-4 This is one embodiment of the present invention, which provides an AI-based adaptive adjustment method for central air conditioning energy-saving control, comprising the following steps:

[0049] S1. Real-time collection of environmental data, equipment operation data, and energy consumption data of the central air conditioning system; preprocessing of multimodal data; construction of physical constraint equations in conjunction with thermodynamic laws; and generation of multimodal datasets.

[0050] Furthermore, the multimodal data is timestamped and the sampling frequency is normalized. A generative adversarial network is used to complete the missing data and output the preprocessed multimodal data.

[0051] Specifically, this is achieved through a unified reference clock. After the timestamps of the central air conditioning's environmental data, equipment operation data, and energy consumption data are converted to standard UTC format, a linear interpolation method is used to synchronize data from different sampling frequencies to the same time interval. For example, if the environmental data sampling frequency is 1Hz, the equipment operation data is 0.5Hz, and the energy consumption data is 0.2Hz, all data are interpolated to a 1Hz frequency. Sampling frequency normalization is accomplished through resampling technology; high-frequency data is downsampled using mean pooling, and low-frequency data is upsampled using cubic spline interpolation, ensuring that all modal data have the same time resolution.

[0052] It should be noted that a generative adversarial network consists of a generator and a discriminator. The generator receives multimodal data containing missing values ​​and outputs completed data; the discriminator distinguishes between completed data and real complete data; the training data of the generative adversarial network must include artificially simulated random missing patterns to cover real-world scenarios.

[0053] The preprocessed multimodal data is stored as a structured array in time series format, containing three fields: environmental data, equipment operation data, and energy consumption data. Each field has consistent data dimensions and strictly aligned timestamps. The data is saved in HDF5 or NPZ format. For example, in the output file, the temperature, humidity, and wind speed of the environmental data are arranged at a 1Hz frequency; the compressor power and damper opening of the equipment operation data are aligned with the instantaneous power of the energy consumption data along the same time axis; missing values ​​are flagged as 0, and complete values ​​are flagged as 1.

[0054] Based on the equipment characteristics and thermodynamic laws of central air conditioning, physical constraint equations are constructed.

[0055] It should be noted that equipment characteristics refer to the inherent physical parameters and operating rules of various components of the central air conditioning system, such as compressors, air valves, and heat exchangers, as well as the nonlinear relationship between the compressor's power-cooling capacity curve and the air valve opening degree and air volume.

[0056] The laws of thermodynamics refer to the principles of energy conservation and heat transfer laws upon which physical constraint equations are constructed. For example, the input electrical energy is equal to the sum of the cooling capacity and the heat loss, and the heat flux density is proportional to the temperature gradient.

[0057] Specifically, the physical constraint equations are constructed based on the actual operating characteristics of the central air conditioning system and classical thermodynamic principles. Control conditions are formed by analyzing the correlation between pre-processed environmental data, equipment operating data, and energy consumption data. The temperature field distribution in the environmental data and parameters such as compressor power and valve opening in the equipment operating data establish energy transfer relationships, while energy consumption data is used to verify energy conservation. The first law of thermodynamics ensures total energy balance, and the principle of heat conduction reflects the dynamic changes in the temperature field. The heat transfer efficiency of each zone is dynamically adjusted according to real-time wind speed and valve status, and temperature gradient changes are constrained and optimized using historical operating data.

[0058] Construct the physical constraint equations, the expressions of which are:

[0059]

[0060] In the formula, T represents the temperature field, and t represents the current time. This represents the partial derivative of the temperature field T with respect to the current time t. Let κ represent the dynamic thermal conductivity, x represent the building's spatial location, and κ(t,x) represent the dynamic thermal conductivity at the building's spatial location x and the current time t. Let represent the gradient vector of the temperature field, where i represents the index of the discretized region, and N represents the total number of discretized regions. This represents the summation of N discretized regions, h i (t) represents the dynamic heat transfer coefficient of the i-th discretized region, A i Let T' represent the heat transfer area of ​​the i-th discretized region, T′ represent the external ambient temperature, and β represent the gradient suppression coefficient.

[0061] It should be noted that the dynamic heat transfer coefficient is calibrated using both convective heat transfer theory and equipment operating data to achieve real-time mapping of heat transfer efficiency. For example, when natural convection is 1–10 W / (m²·K), and the opening degree is 0%–50%, h i (t)∈[0,20], with the damper closed or at low wind speed, forced convection: at 10~100W / (m2·K), the opening degree is 50%-100%h. i (t)∈[20,40], the air valve is fully open; the gradient suppression coefficient is determined by historical data optimization and numerical stability analysis, for example, the optimal β = 1.2 for summer operation and the optimal β = 0.8 for winter operation.

[0062] The preprocessed multimodal data is fused with physical constraint equations to generate a multimodal dataset containing physical laws.

[0063] Specifically, the preprocessed multimodal data and physical constraint equations are fused using a data-physics coupling method. Temperature field distribution from environmental data and parameters such as compressor power and damper opening from equipment operation data are substituted into the physical constraint equations to verify the relationship between energy conservation and heat conduction. Energy consumption data is used to calculate residuals and correct equation coefficients. The generated multimodal dataset contains both the original observations and the physical equation verification results at each time step. For example, when the damper opening is 80% and the compressor power is 15kW, a residual of less than 5% between the calculated temperature field gradient of 0.5℃ / m and the measured value of 0.48℃ / m is marked as physically consistent data.

[0064] It should be noted that raw observations refer to the original data records that are directly collected from sensors or devices and have undergone preprocessing but have not been corrected by physical equations. These include environmental data observations, equipment operation data observations, and energy consumption data observations.

[0065] The physical constraint equations include the heat conduction equation and the energy conservation equation.

[0066] It should be noted that the thermal diffusivity in the heat conduction equation is decomposed into a composite parameter of thermal conductivity, density, and specific heat capacity. The thermal conductivity is dynamically adjusted according to the compressor efficiency and ambient humidity. The thermal resistance in the energy conservation equation is determined by the damper opening and the air velocity, while the specific heat capacity depends on the heat exchanger material properties. The heat transfer coefficient has a non-linear relationship with the damper opening and the air velocity.

[0067] S2. Input the multimodal dataset into the physical information neural network, and predict the thermodynamic state and cooling load demand of the building in the future period by embedding the thermodynamic partial differential equation constraint in the deep learning model, and output the prediction results.

[0068] Furthermore, the multimodal dataset and the parameterized physical constraint equations are input into the spatiotemporal encoder and physical encoder of the physical information neural network to extract spatiotemporal features and generate implicit physical features.

[0069] Specifically, the spatiotemporal encoder processes environmental data, equipment operation data, and energy consumption data from multimodal data through a three-dimensional convolutional network, extracting spatiotemporal features with temporal and spatial correlations; the physical encoder inputs parameterized physical constraint equation parameters such as thermal resistance, thermal capacity, and heat transfer coefficient into a fully connected network to generate implicit physical features characterizing thermodynamic laws.

[0070] The spatiotemporal features and implicit physical features are input into the fusion layer of the graph neural network. The thermodynamic partial differential equations in the spatiotemporal domain are integrally projected through a dynamic adaptive weight function to generate physically consistent spatiotemporal joint features.

[0071] Specifically, the fusion layer of the graph neural network receives spatiotemporal features extracted by the spatiotemporal encoder and implicit physical features generated by the physical encoder. The dynamic adaptive weight function assigns weight coefficients according to the spatiotemporal features and implicit physical features. The weighted features are projected onto the integral residual term of the thermodynamic partial differential equation. After the spatiotemporal features and physical features interact through the attention mechanism, physically consistent spatiotemporal joint features are generated.

[0072] The dynamic adaptive weight function is expressed as follows:

[0073]

[0074] In the formula, λ(t) represents the dynamic weight, exp represents the exponential function, η represents the decay rate coefficient, t0 represents the start time, and Ψ represents the deviation between the temperature field and the external ambient temperature. This is expressed as the gradient magnitude of the deviation field. This indicates that the integral is performed from the starting time t0 to the current time t, and dt represents the time integration variable.

[0075] It should be noted that the decay rate coefficient is obtained by fusing historical temperature gradient data with multimodal data of control commands, and is dynamically adjusted through offline optimization and online adaptive adjustment, for example in a shopping mall (sudden change in pedestrian flow): η∈[0.1,5].

[0076] The integral residual projection function is expressed as:

[0077]

[0078] In the formula, Represented as physical constraint residuals, j represents the sensor index, and S represents the total number of sensors. This represents the cumulative total number of sensors, where t′ represents a future time (e.g., 5 seconds in the future). This is expressed as the deviation between the temperature field and the external ambient temperature. For the current time The partial derivative, where α represents the thermal diffusivity. Let γ represent the Laplace operator of the bias field, where γ represents the KL divergence regularization coefficient, and KL(p||q) represents the difference between the probability distribution p of the temperature field output by the predictive physical information neural network and the statistical distribution q of the temperature field measured by the sensor.

[0079] It should be noted that the thermal diffusivity is obtained through material parameters. For example, when it is concrete, α = 1.0. The KL divergence regularization coefficient is calibrated by backpropagation of the error between the predicted distribution and the actual distribution. For example, in an office building (stable condition): γ = 0.5.

[0080] Based on the spatiotemporal joint characteristics, the overall building cooling load demand forecast and the spatial distribution of the building's internal thermodynamic state are predicted at multiple scales. The overall building cooling load demand forecast is output through the fully connected layer, and the spatial distribution of the building's internal thermodynamic state is output through the U-Net structure.

[0081] Specifically, the physical consistency spatiotemporal joint feature input multi-scale prediction structure adopts a three-layer fully connected network structure. The fully connected layer maps global features to the predicted value of the building's overall cooling load demand. The spatiotemporal joint feature input U-Net structure gradually restores spatial resolution through encoder-decoder architecture and skip connections, and outputs the spatial distribution of thermodynamic states such as temperature field and humidity field inside the building.

[0082] The spatial distribution of predicted cooling load demand and thermodynamic state is input into the normalization mapping layer to generate prediction results.

[0083] Specifically, the predicted cooling load demand is scaled to the [0,1] interval by Min-Max normalization, the thermodynamic state distribution is standardized by Z-score, and the normalized values ​​are used to generate the final prediction result through linear combination.

[0084] S3. Based on the prediction results, construct a causal graph model to identify key variables affecting energy consumption and comfort, quantify the causal effects of key variables, and dynamically allocate optimization weights for energy saving and comfort.

[0085] Furthermore, causal candidate variables are extracted from the prediction results, and a causal graph model is constructed using an improved PC algorithm;

[0086] Specifically, energy consumption-related variables, comfort-related variables, and environmental covariates from the prediction results are used as causal candidate variables input into the improved PC algorithm. A causal graph model is constructed through conditional independence tests and causal direction determination. The improved PC algorithm uses an information entropy-based conditional independence test method to screen for dependencies between variables and uses the time priority criterion to determine the causal direction.

[0087] Based on the causal graph model, key variables affecting energy consumption and comfort are identified. A dual machine learning method is used to calculate the joint causal effect coefficient of key variables on energy consumption and comfort. The distribution difference between historical data and real-time data is quantified by the KL divergence function.

[0088] Specifically, a causal graph model is used to identify key variables affecting energy consumption and comfort. Variables with a direct causal relationship to the target variable, including compressor power, damper opening, and outdoor temperature, are selected. A dual machine learning approach is employed to calculate the joint causal effect coefficients of the key variables. The first machine learning step fits the conditional expectation of the cooling load forecast and temperature field distribution with respect to the key variables. The second machine learning step calculates the partial derivatives based on the residuals and then seeks the expectation to obtain the joint causal effect coefficient of compressor power on energy consumption and comfort. The difference between historical and real-time data distributions is quantified using the KL divergence function.

[0089] It should be noted that the target variable is a predefined optimization objective, obtained by directly collecting data from sensors or calculating indicators; variables with direct causal edges refer to variables that are directly connected to the target variable through directed edges in the causal graph model, including equipment operation data and environmental data; the identified key variables are high-influence variables selected from the direct causal variables, such as those with an in-degree > 2.

[0090] The causal effect coefficient of the key variable is calculated using dual machine learning, expressed as follows:

[0091]

[0092] V∈{P,f,T out};

[0093] In the formula, This represents the joint causal effect coefficient of key variables on energy consumption and comfort, where Q represents the predicted cooling load and V represents the set of key variables. This is expressed as the mean global comfort deviation. This represents the partial derivative of the predicted cooling load Q with respect to the set of key variables V. Represented as the mean of global comfort deviation The partial derivatives with respect to the set of key variables V, where P represents compressor power, f represents frequency, and T... out Represented as external temperature, It is represented as the mathematical expectation operator.

[0094] Based on the joint causal effect coefficient and distribution difference, an information entropy-driven dynamic weight function is constructed to dynamically allocate the optimization weights for energy saving and comfort.

[0095] Specifically, based on the quantitative results of the joint causal effect coefficient and KL divergence, where the joint causal effect coefficient measures the strength of the causal effect of key variables on energy consumption and comfort (e.g., for every 1kW increase in compressor power, energy consumption increases by 5% and comfort decreases by 0.3%), and the KL divergence quantifies the degree of difference between historical and real-time data distributions (e.g., the KL divergence between real-time data and historical benchmarks under summer conditions is 0.8), a dynamic weighting function is constructed by combining variable information entropy. The dynamic weighting function uses the Softmax function to normalize the product of the joint causal effect coefficient and information entropy, generating dynamically optimized weights for energy-saving and comfort targets.

[0096] Construct an information entropy-driven dynamic weight function, expressed as:

[0097]

[0098] In the formula, λ′(t) represents the dynamic optimization weight of the key variable set V, H represents the information entropy of the key variable set V, k represents the key variable index, and K represents the total number of key variables. This indicates that the totals of the K key variables are summed. V is represented as the coefficient of the joint causal effect of the k-th key variable on energy consumption and comfort. k This is represented as the k-th key variable.

[0099] Key variables refer to those variables that have a causal impact on the energy consumption and comfort of central air conditioning, as identified through causal graph models.

[0100] It should be noted that the causal graphical model uses an improved PC algorithm to screen out key variables that have a causal relationship with the target variable. Causal effects are verified through conditional independence tests and causal direction determination. For example, the causal effect coefficient between compressor power and cooling load is 0.6, and the causal effect coefficient between damper opening and temperature field gradient is 0.3. The strength of the causal edge is quantified using partial correlation coefficients or mutual information, and only variables with an effect strength exceeding 0.2 are retained as key variables.

[0101] Key variables include energy consumption-related variables, comfort-related variables, and environmental covariates.

[0102] It should be noted that energy consumption-related variables include equipment operating parameters and energy efficiency indicators; comfort-related variables include thermodynamic state and uniformity indicators; and environmental covariates include external environment and building characteristics.

[0103] Equipment operating parameters include compressor power, refrigerant flow rate, fan frequency, and water pump power;

[0104] Energy efficiency indicators include instantaneous energy efficiency ratio, cooling capacity, and condenser heat load;

[0105] Thermodynamic state includes spatial distribution of temperature field, spatial distribution of humidity field, temperature gradient, and local wind speed;

[0106] Uniformity indicators include the standard deviation of temperature difference in each region, PMV index, and PPD.

[0107] The external environment includes outdoor temperature, outdoor humidity, solar radiation intensity, and wind speed;

[0108] Building characteristics include thermal resistance of the building envelope, window area ratio, and population density.

[0109] S4. Generate multi-objective optimization control commands based on the optimization weights, solve for the optimal combination of equipment parameters using the Pareto front algorithm, and transmit them to the central air conditioning actuator.

[0110] Furthermore, the multi-objective optimization control command is decomposed into multiple sub-optimization problems, each of which is assigned dynamic weights and the device parameter search space is defined;

[0111] Specifically, the multi-objective optimization control command is decomposed into energy-saving optimization sub-problems and comfort optimization sub-problems.

[0112] It should be noted that the energy-saving optimization subproblem takes the predicted value of cooling load demand as the objective function, and the equipment parameter search space is limited to the operating range of compressor power, air valve opening, etc. based on the statistical range of historical operating data.

[0113] The comfort optimization subproblem uses the temperature field gradient as the objective function, and the equipment parameter search space limits the adjustment range of fan frequency, refrigerant flow, etc.

[0114] The dynamic weighting function allocates energy-saving and comfort weights based on an information entropy-driven dynamic weighting function. The overall optimization objective is the sum of the weighted cooling load error and temperature gradient error. The upper and lower bounds of the equipment parameter search space are determined through quantile statistics of historical operating data to eliminate interference from extreme operating conditions and ensure that optimization instructions are generated within the feasible region.

[0115] The Pareto front algorithm is used to solve the sub-optimization problem in parallel, and the non-dominated solution set that satisfies all constraints is selected to generate candidate equipment parameter combinations.

[0116] Specifically, the Pareto front algorithm solves the two subproblems in parallel, generating a non-dominated solution set using the NSGA-II algorithm. The selection criteria for the non-dominated solution set are: it is not dominated by other solutions in terms of both energy saving and comfort objectives, and it satisfies the equipment parameter search space constraints. Candidate equipment parameter combinations are extracted from the non-dominated solution set, and a crowding ranking and uniform sampling strategy is used to select representative parameter combinations.

[0117] Based on the real-time load rate, select the optimal combination of device parameters from the candidate combinations of device parameters;

[0118] Specifically, the real-time load rate is determined by the ratio of the current predicted cooling load demand to the rated capacity of the equipment. Candidate equipment parameter combinations are assigned energy-saving and comfort optimization weights based on the range of the real-time load rate using a dynamic weighting function. A weighted summation method is used to comprehensively score and rank the energy-saving and comfort target achievement rates of each candidate combination. During the scoring process, it is simultaneously verified whether the candidate combinations meet the temperature field gradient residual requirements of the physical constraint equations and the equipment parameter search space limitations. The parameter combination with the highest score and satisfying all constraints is selected as the optimal solution. The output results ensure that the selected equipment parameter combination simultaneously achieves a balance between minimizing energy consumption and minimizing comfort deviation under real-time load rate conditions.

[0119] The optimal combination of equipment parameters is encoded into control commands compatible with the central air conditioning actuator protocol and transmitted to the central air conditioning actuator.

[0120] Specifically, parameters such as compressor power, damper opening, and fan frequency in the optimal equipment parameter combination are converted into binary instruction codes according to the register addresses and data formats defined in the central air conditioning actuator protocol. These instruction codes are then encapsulated into communication protocol frames after adding a checksum. The encoded control commands are transmitted to the central air conditioning actuator via an industrial communication protocol, using standard communication parameters. Before execution, the compatibility with the physical constraint equations is verified. If the verification fails, an instruction rollback mechanism is triggered, and a suboptimal parameter combination is selected. The issued control commands strictly match the function codes and address mapping table defined in the actuator protocol to ensure that the actuator accurately parses and executes the operation.

[0121] S5. Collect actual data after command execution, calculate the deviation between energy efficiency and comfort, update physical information neural network parameters and adjust thermodynamic partial differential equation coefficients through meta-learning framework.

[0122] Furthermore, actual energy consumption data, temperature field distribution data, and equipment status data are collected from the central air conditioning actuators and sensors, and sensor noise is eliminated through adaptive Kalman filtering.

[0123] Specifically, data is transmitted to the data processing unit via Modbus RTU or MQTT protocols. Actual energy consumption data includes instantaneous and cumulative compressor power values; temperature field distribution data covers temperature, humidity, and wind speed in each area; and equipment status data records valve opening, fan frequency, and refrigerant flow. Adaptive Kalman filtering dynamically adjusts the process noise covariance matrix and observation noise covariance matrix based on the sensor noise statistical characteristics to smooth abnormal fluctuations in the temperature field distribution data.

[0124] Based on the filtered actual energy consumption data and the predicted cooling load demand, the dynamic energy-saving efficiency index is calculated, and the spatial weighted comfort deviation is calculated by the deviation between the temperature field distribution data and the preset value of the target temperature field distribution.

[0125] It should be noted that the dynamic energy efficiency index is calculated by the integral ratio of the filtered actual energy consumption data and the predicted cooling load demand. The spatial weighted comfort deviation is calculated by taking the square root of the weighted sum of the squared deviations between the temperature of each area and the preset target temperature in the temperature field distribution data. The weighting coefficients are allocated according to the area and population density.

[0126] Specifically, the dynamic energy efficiency index is calculated using the following expression:

[0127]

[0128] In the formula, ξ(t) represents the dynamic energy-saving efficiency index, P′ represents the actual compressor power, and σ represents the normalization coefficient. It is expressed as the squared L2 norm of the temperature field gradient after execution.

[0129] It should be noted that the normalization coefficient is obtained based on the statistical distribution of the historical temperature field gradient, σ∈[0.1,1.0].

[0130] The spatially weighted comfort deviation is calculated using the following expression:

[0131]

[0132] In the formula, D(t) represents the spatially weighted comfort deviation, s represents the region index, and S represents the total number of regions. This indicates that the totals of the S regions are summed, w s Let the weight of the s-th region be denoted as . Let Ψ′ represent the temperature field gradient after execution in region s, and let Ψ′ represent the deviation of the preset value of the target temperature field distribution.

[0133] The dynamic energy efficiency index and comfort deviation are input into the meta-learning framework, and the parameters of the physical information neural network are updated through cross-task gradient descent, thereby simultaneously optimizing the coefficients of the thermodynamic partial differential equation.

[0134] Specifically, the meta-learning framework jointly updates the weights of the physical information neural network and the dynamic thermal conductivity and heat transfer coefficient in the thermodynamic partial differential equations through a cross-task optimization mechanism. The dynamic energy-saving efficiency index serves as the energy consumption optimization objective, and the spatially weighted comfort deviation serves as the comfort optimization objective. The weights of both, allocated through a dynamic weighting function, form the multi-task joint loss function. During cross-task gradient descent, the spatiotemporal encoder parameters of the physical information neural network are adjusted synchronously with the coefficients of the thermodynamic partial differential equations. The updates to thermal conductivity and heat transfer coefficients are calculated based on the temperature field residuals and energy conservation constraints, ensuring that the equation parameters dynamically adapt to real-time operating conditions.

[0135] This embodiment also provides a computer device applicable to the AI-based adaptive adjustment central air conditioning energy-saving control method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the AI-based adaptive adjustment central air conditioning energy-saving control method proposed in the above embodiment.

[0136] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0137] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the AI-based adaptive adjustment method for central air conditioning energy saving control as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0138] In summary, this invention enhances the comprehensive control capabilities of central air conditioning systems in complex building environments by constructing a multimodal physical information fusion framework and a dynamic closed-loop optimization system. Based on real-time acquisition of multi-source data and embedding thermodynamic constraints, it can accurately predict the building's thermodynamic state and cooling load demand, ensuring that the prediction results strictly adhere to the laws of energy conservation and heat conduction. Through causal reasoning, it identifies the interaction effects between key variables, dynamically quantifies the coupling relationship between energy consumption and comfort, and adaptively allocates optimization weights to fully tap energy-saving potential. Combined with Pareto frontier algorithms, it generates globally optimal equipment parameter combinations, improving the collaborative control efficiency of actuators such as compressors and dampers, avoiding resource waste or performance imbalances caused by local optimization. Through a meta-learning-driven closed-loop update mechanism, it can perceive the deviation between actual operating data and prediction results in real time, autonomously adjust neural network parameters and thermodynamic equation coefficients, and continuously adapt to dynamic disturbances such as seasonal changes and equipment aging.

[0139] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A central air conditioning energy-saving control method based on AI adaptive adjustment, characterized in that: include, Real-time collection of environmental data, equipment operation data, and energy consumption data of the central air conditioning system; preprocessing of multimodal data; construction of physical constraint equations in conjunction with thermodynamic laws; and generation of multimodal datasets. The multimodal dataset is input into a physical information neural network. A deep learning model, constrained by thermodynamic partial differential equations, is used to predict the building's thermodynamic state and cooling load demand over future periods, outputting the prediction results. The specific steps are as follows. The multimodal dataset is input into the spatiotemporal encoder and physical encoder of the physical information neural network to extract spatiotemporal features and generate implicit physical features; The spatiotemporal features and implicit physical features are input into the fusion layer of the graph neural network. The thermodynamic partial differential equations in the spatiotemporal domain are integrally projected through a dynamic adaptive weight function to generate physically consistent spatiotemporal joint features. Based on the spatiotemporal joint characteristics, the overall building cooling load demand forecast and the spatial distribution of the building's internal thermodynamic state are predicted at multiple scales. The overall building cooling load demand forecast is output through the fully connected layer, and the spatial distribution of the building's internal thermodynamic state is output through the U-Net structure. The spatial distribution of the predicted cooling load demand and the thermodynamic state is input into the normalization mapping layer to generate the prediction results. Based on the prediction results, a causal graph model is constructed to identify key variables affecting energy consumption and comfort, quantify the causal effects of key variables, and dynamically allocate optimization weights for energy saving and comfort. Multi-objective optimization control commands are generated based on the optimization weights, the optimal combination of equipment parameters is solved using the Pareto front algorithm, and then transmitted to the central air conditioning actuator. The system collects actual data after the command is executed, calculates the deviation between energy efficiency and comfort, updates the parameters of the physical information neural network through the meta-learning framework, and adjusts the coefficients of the thermodynamic partial differential equation.

2. The central air conditioning energy-saving control method based on AI adaptive adjustment as described in claim 1, characterized in that: The specific steps for generating the multimodal dataset are as follows: The multimodal data is timestamped and the sampling frequency is normalized. A generative adversarial network is used to complete the missing data and output the preprocessed multimodal data. Based on the equipment characteristics and thermodynamic laws of central air conditioning, physical constraint equations are constructed. The preprocessed multimodal data is fused with physical constraint equations to generate a multimodal dataset containing physical laws.

3. The central air conditioning energy-saving control method based on AI adaptive adjustment as described in claim 2, characterized in that: The physical constraint equations include the heat conduction equation and the energy conservation equation.

4. The central air conditioning energy-saving control method based on AI adaptive adjustment as described in claim 1, characterized in that: The specific steps for dynamically allocating the optimization weights for energy saving and comfort are as follows: Causal candidate variables are extracted from the prediction results, and a causal graph model is constructed using an improved PC algorithm; Based on the causal graph model, key variables affecting energy consumption and comfort are identified. A dual machine learning method is used to calculate the joint causal effect coefficient of key variables on energy consumption and comfort. The distribution difference between historical data and real-time data is quantified by the KL divergence function. Based on the joint causal effect coefficient and distribution difference, an information entropy-driven dynamic weight function is constructed to dynamically allocate the optimization weights for energy saving and comfort.

5. The central air conditioning energy-saving control method based on AI adaptive adjustment as described in claim 4, characterized in that: The key variables refer to those variables that have a causal impact on the energy consumption and comfort of central air conditioning, as identified by the causal graph model. The key variables include energy consumption-related variables, comfort-related variables, and environmental covariates.

6. The central air conditioning energy-saving control method based on AI adaptive adjustment as described in claim 1, characterized in that: The transmission to the central air conditioning actuator involves the following specific steps. The multi-objective optimization control command is decomposed into multiple sub-optimization problems, and each sub-problem is assigned dynamic weights and the device parameter search space is defined. The Pareto front algorithm is used to solve the sub-optimization problem in parallel, and the non-dominated solution set that satisfies all constraints is selected to generate candidate equipment parameter combinations. Based on the real-time load rate, select the optimal combination of device parameters from the candidate combinations of device parameters; The optimal combination of equipment parameters is encoded into control commands compatible with the central air conditioning actuator protocol and transmitted to the central air conditioning actuator.

7. The central air conditioning energy-saving control method based on AI adaptive adjustment as described in claim 1, characterized in that: The specific steps for updating the physical information neural network parameters and adjusting the coefficients of the thermodynamic partial differential equations are as follows. The system collects actual energy consumption data, temperature field distribution data, and equipment status data from central air conditioning actuators and sensors, and eliminates sensor noise through adaptive Kalman filtering. Based on the filtered actual energy consumption data and the predicted cooling load demand, the dynamic energy-saving efficiency index is calculated, and the spatial weighted comfort deviation is calculated by the deviation between the temperature field distribution data and the preset value of the target temperature field distribution. The dynamic energy efficiency index and comfort deviation are input into the meta-learning framework, and the parameters of the physical information neural network are updated through cross-task gradient descent, thereby simultaneously optimizing the coefficients of the thermodynamic partial differential equation.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the central air conditioning energy-saving control method based on AI adaptive adjustment as described in any one of claims 1 to 7.

9. 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 central air conditioning energy-saving control method based on AI adaptive adjustment as described in any one of claims 1 to 7.

Images