A two-stage optimization control method and system for a user-centered heating, ventilation, and air conditioning system

Abstract

The application discloses a kind of two-stage optimization control method and system for user-centered heating ventilation air conditioning system, it is related to building energy consumption control field, including the following steps: step 1, establish second-order equivalent thermal parameter model, obtain the online identification result of model initial parameter by genetic algorithm iteration, obtain the initial prediction value result of indoor temperature;Step 2, establish TimesNet model, construct the input and output feature set of model, obtain the correction value result of indoor temperature;Step 3, establish fusion model and complete accurate prediction to indoor temperature;Step 4, realize group-level regulation and control of indoor temperature based on model predictive control, obtain temperature setting interval;Step 5, realize individualized regulation and control of indoor temperature based on deep Q network, obtain temperature setting value.The application has significant technical advantages in modeling accuracy, control performance and response efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of building energy consumption control, and particularly relates to a two-stage optimization control method and system for a user-centered heating, ventilation and air conditioning (HVAC) system. BACKGROUND

[0002] Under the global carbon neutralization goal, building energy consumption optimization has become a core task for the development of smart grids and green buildings. As the largest energy consumption unit in buildings, the intelligent control of heating, ventilation and air conditioning (HVAC) systems is of great significance for improving energy efficiency, reducing operating costs and achieving coordinated regulation of buildings and power grids. Traditional HVAC control strategies are mostly based on physical thermal models, which have good interpretability and stability, but have limitations in dealing with changes in user behavior.

[0003] In recent years, control methods based on artificial intelligence, such as deep reinforcement learning (DRL) and data-driven prediction models, have been gradually introduced into HVAC system control. Such methods have good adaptive ability and non-linear modeling ability, and are particularly advantageous in handling multi-variable dynamic environments. However, relying solely on data-driven methods also has problems such as insufficient model stability, high learning cost and difficulty in constraining strategies. SUMMARY

[0004] The present application relates to the field of building energy consumption control, and particularly relates to a two-stage optimization control method and system for a user-centered heating, ventilation and air conditioning (HVAC) system.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0006] A two-stage optimization control method for user-centered HVAC systems, first, the equivalent thermal parameter (ETP) model is combined with the TimesNet to realize the complementary coupling of building thermal inertia modeling and dynamic human behavior disturbance modeling, effectively improving the prediction accuracy; second, a double-layer optimization control architecture based on model predictive control (MPC) and deep Q-network (DQN) cooperation is constructed to realize group-level optimization of temperature setting interval and fine-grained adaptive control of individual temperature set point, and the robustness and individual response ability of the system are considered; finally, a personalized modeling mechanism driven by user temperature preference historical data is introduced, the user's expected temperature is estimated by extracting historical room temperature data through a sliding window, and the individual comfort preference is integrated into the state space and reward function of the reinforcement learning controller to realize dynamic preference learning and adaptive comfort control.

[0007] Further, the method of the present application comprises the following steps:

[0008] Step 1, a second-order equivalent thermal parameter model is established, and the online identification result of the initial parameters of the model is obtained by genetic algorithm iteration to obtain the initial prediction value result of the indoor temperature;

[0009] Step 2, a TimesNet model is established, and the input and output feature sets of the model are constructed to obtain the correction value result of the indoor temperature;

[0010] Step 3, a fusion model is established and the accurate prediction of the indoor temperature is completed;

[0011] Step 4, group-level regulation and control of the indoor temperature is realized based on model predictive control to obtain the temperature setting interval;

[0012] Step 5, individualized regulation and control of the indoor temperature is realized based on deep Q-network to obtain the temperature setting value.

[0013] Further, the step 1 specifically comprises:

[0014] 1.1) A second-order ETP model is established to describe the heat exchange process inside the building, which contains four basic parameters: gas thermal resistance, gas heat capacity, solid thermal resistance and solid heat capacity, considering the heat exchange between indoor air, building structure and external environment, the thermal dynamic behavior of the system is described by a set of ordinary differential equations:

[0015]

[0016] wherein, represents the current time, represents the indoor temperature, represents the gas thermal resistance, represents the gas thermal capacity, represents the solid thermal resistance, represents the solid thermal capacity, is the temperature of the indoor solid, represents the outdoor temperature, represents the on-off state of the air conditioner, represents the thermal power of the air conditioner;

[0017] 1.2) Estimate using genetic algorithm GA , , and Optimal parameters: first, the parameter set is encoded into the form of a chromosome, and each parameter combination is represented as an individual. Through random initialization, an initial population of a certain size is generated, denoted as , where P is the population size; for each individual , its fitness is calculated (Fitness), and the initial predicted value of the indoor temperature is obtained by simulating the ETP model , and compared with the real temperature , the root mean square error (Root Mean Square Error, RMSE) is used as the evaluation index:

[0018]

[0019] where, represents the initial predicted value of the indoor temperature output by the ETP model, is the mean square error between the ETP model predicted temperature and the real indoor temperature; in each iteration process, new population is generated through crossover and mutation operations, and the above process is repeated, and the whole genetic process continues until the termination condition is met, and the final output optimal individual parameter set makes the ETP model prediction error minimum.

[0020] Further, the step 2 specifically comprises:

[0021] 2.1) Establish TimesNet model to capture the influence of user dynamic behavior on indoor temperature, TimesNet model divides the original input feature sequence into k frequency components through frequency decomposition, and applies a two-dimensional convolution structure to each component for feature extraction, thereby modeling local and global dynamic features at the same time;

[0022] Assuming that the first i processing module of the TimesNet model receives a one-dimensional sequence To perform two-dimensional processing on the sequence, first, it is expanded to a suitable length by padding with zeros, and then reshaped into a two-dimensional tensor. , The data is further fed into an improved feature extraction module to extract local and periodic features. The extracted two-dimensional features are then transformed back into one-dimensional space and aggregated based on frequency importance.

[0023] 2.2) The input to the TimesNet model is physical characteristics reflecting the user's dynamic behavior, and the output of the TimesNet model is... This is a temperature correction term used to compensate for disturbances that are not modeled in the traditional ETP model. The input-output relationship is expressed as follows:

[0024]

[0025] in, Indicates the electrical power of indoor equipment. Indicates indoor humidity and illuminance. It is the fitting function for the TimesNet model.

[0026] Furthermore, step 3 specifically includes:

[0027] 3.1) In the proposed hybrid modeling framework, the ETP model first generates a baseline value for indoor temperature prediction based on static physical parameters; simultaneously, TimesNet generates a temperature correction term, capturing residual changes caused by dynamic occupant behavior and environmental disturbances. The ETP output and the TimesNet correction are directly superimposed to obtain the final indoor temperature prediction, expressed as:

[0028] .

[0029] Furthermore, step 4 specifically includes:

[0030] 4.1) At each control moment, MPC dynamically adjusts the upper and lower limits of the temperature setpoint by solving an optimization problem to achieve a trade-off between energy saving and comfort. The decision variable of the optimization problem is the upper bound of the setpoint temperature. and the lower realm To determine the final temperature setpoint used for control The system is in the interval [ , Generate m random samples within the specified range and take their average:

[0031]

[0032] in, Let q be the q-th randomly generated sample. m Indicates the number of samples;

[0033] 4.2) Setpoint temperature used to determine the required heating power of the HVAC system at the current time instant to achieve the target temperature regulation, the state evolution of the control system is described by the following dynamics equation:

[0034]

[0035] 4.3) The objective function of the MPC is to minimize the total cost function within the prediction horizon n which is composed of two parts: energy cost and thermal comfort cost, weighted by energy weight coefficient and thermal comfort weight coefficient respectively:

[0036]

[0037] where, represents the energy cost, represents the thermal discomfort cost, represents the energy weight coefficient, represents the thermal comfort weight coefficient;

[0038] Energy cost is calculated based on the power consumption of the HVAC and the real-time electricity price:

[0039]

[0040] where, represents the electric-to-thermal power conversion efficiency of the HVAC system, represents the time step, represents the real-time electricity price at the current time instant;

[0041] Thermal discomfort cost is estimated by a simplified PMV model:

[0042]

[0043] where, represents the current relative humidity, represents the metabolic rate of human body, represents the clothing thermal resistance;

[0044] 4.4) To ensure the safe operation of the system and the comfort of the user, the constraint conditions of the MPC optimization problem are:

[0045] 1). Power constraint: ;

[0046] 2). Temperature setpoint range constraint: ;

[0047] ​3) Comfort constraint: ;

[0048] 4) Temperature setpoint interval width constraint: ;

[0049] where, denotes the maximum thermal power of the air conditioner, denotes the lower bound of the temperature setpoint interval, denotes the upper bound of the temperature setpoint interval, denotes the width of the temperature setpoint interval.

[0050] Further, the step 5 specifically comprises:

[0051] 5.1) The deep Q-network dynamically selects the optimal temperature setpoint within the temperature setpoint interval defined by the first-stage MPC, , to achieve a coordinated response of the environment state and the individualized demand;

[0052] 5.2) The indoor temperature of the same time period in the past seven working days is weighted and averaged by the moving average method to construct the time-varying individual preference temperature :

[0053]

[0054] where, denotes the indoor temperature value at time d in the t th past working day;

[0055] 5.3) The control process is modeled as a Markov decision process, which includes the following elements:

[0056] 1) State : composed of the current indoor temperature , electricity price , outdoor temperature , indoor relative humidity , and user preference temperature ;

[0057] 2) Action : a value in the discretized temperature setpoint set, satisfying ;

[0058] 3) Reward function : used to balance energy consumption and user comfort, defined as follows:

[0059]

[0060] where, representing energy cost a weight coefficient representing user preference cost, a weight coefficient representing user preference cost, respectively used to control the trade-off relationship between energy cost and preference temperature difference;

[0061] 5.4) At each time step, the agent selects an action according to the current state and - the greedy policy selects an action with probability randomly selects an action with probability 1- performs the current optimal action, and then interacts with the environment to obtain a new state and a reward , and stores this experience into the experience replay buffer.

[0062] 5.5) The current state is input into the deep Q network to predict the current Q value , while the target Q network is used to estimate the maximum Q value corresponding to the optimal action at the next time, and the target value is calculated as:

[0063]

[0064] wherein, is a discount factor used to balance short-term and long-term returns, is the action at the next time, and the following mean square error loss function is minimized:

[0065]

[0066] The loss is optimized by the backpropagation algorithm to optimize the parameters of the Q network. The target Q network synchronizes the parameters of the main Q network every few steps through soft update or copy strategy.

[0067] Further, the application also provides a two-stage optimization control system for a user-centered HVAC system, comprising a hybrid modeling module, a double-layer optimization module, and a personalized learning module. The hybrid modeling module inputs building physical structure parameters, historical sensor data, and user behavior data, and outputs future temperature prediction values via an ETP-TimesNet model. The double-layer optimization module solves the rolling control according to the predicted temperature and electricity price prediction, and generates temperature set upper and lower limits and final control actions. The personalized module estimates the user target temperature based on 7-day historical temperature preferences, which is used as a control reference for the DQN.

[0068] The mixed modeling module fuses the modeling advantages of the physical mechanism model and the data-driven model, describes the thermal inertia and heat transfer characteristics of the building structure based on the ETP model, and generates a baseline value of the room temperature; TimesNet takes the energy consumption data and environmental parameters of HVAC, lighting, and socket loads as inputs, captures the nonlinear disturbance characteristics caused by user behavior, and outputs a correction value. The final prediction result is the weighted superposition of the two, thereby realizing the joint modeling of static building physical properties and dynamic human behavior disturbance, and effectively improving the prediction accuracy.

[0069] The double-layer optimization module consists of two layers: the upper layer is based on the MPC method, which optimizes the upper and lower limits of the indoor temperature set under the consideration of electricity price, environmental parameters and comfort constraints, and provides a robust regulation boundary for system operation; the lower layer is based on constrained DQN, which further optimizes the actual set point in the feasible solution space output by MPC, and realizes the dynamic response to individual preferences. This structure has both the stability of MPC and the adaptive ability of DQN, effectively solving the problem that traditional controllers are difficult to balance system energy efficiency and individual comfort at the same time.

[0070] The individualized learning module extracts the preference value through a sliding window, the preference value is the room temperature data of the user in the historical 7-day working period, calculates the individual expected temperature at each time, and introduces the preference value in the DQN state space and the "temperature set point and preference temperature deviation square" term in the reward function, so that the controller can constrain the deviation degree of the set value while optimizing the energy consumption cost. Thus, dynamic learning of preferences and comfort maintenance are realized, and the individualized control capability and user acceptance of the system are improved.

[0071] Compared with the prior art, the present application has the advantages of: the present application has significant technical advantages in modeling accuracy, control performance and response efficiency. First, the physical model and deep time sequence neural network are fused, breaking through the problem that the prediction accuracy of traditional models is limited, and the indoor thermal environment changing with time can be described more comprehensively. Secondly, the double-layer optimization control architecture effectively combines the stability of model predictive control and the adaptive ability of deep reinforcement learning, which can guarantee the robustness of system operation and realize individual-oriented differentiated response, and improve user satisfaction. Finally, by introducing the constrained reinforcement learning mechanism, the explicit feasibility constraint of the control strategy is realized, the learning process is guided to be more efficient and stable, and the deployment feasibility and control safety of the algorithm in the actual intelligent building system are improved. BRIEF DESCRIPTION OF DRAWINGS

[0072] Figure 1 It is a two-stage optimization control system framework of the user-centered heating, ventilation and air conditioning system of the present application;

[0073] Figure 2 It is a second-order equivalent thermal parameter model diagram.

[0074] Figure 3 For the data-physical fusion modeling framework of the application;

[0075] Figure 4 For the TimesNet network architecture diagram of the application;

[0076] Figure 5 For the two-stage optimization control strategy diagram of the application;

[0077] Figure 6 For the precision improvement effect diagram of the fusion model of the application;

[0078] Figure 7 For the two-stage optimization control strategy effect diagram of the application. DETAILED DESCRIPTION

[0079] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the application.

[0080] Please refer to Figure 1 In the embodiments of the application, a two-stage optimization control method for a user-centered heating, ventilation and air conditioning system includes:

[0081] Step 1, a second-order equivalent thermal parameter model is established, an online identification result of initial parameters of the model is obtained through genetic algorithm iteration, an initial prediction value result of indoor temperature is obtained, and specifically includes:

[0082] 1.1) A second-order ETP model is established to describe the heat exchange process inside the building, as shown in Figure 2 The model includes four basic parameters: gas thermal resistance, gas heat capacity, solid thermal resistance and solid heat capacity. The heat exchange between indoor air, building structure and external environment is considered. The thermal dynamic behavior of the system is described by a set of ordinary differential equations:

[0083]

[0084] wherein, represents the current time, represents the indoor temperature, represents the gas thermal resistance, represents the gas heat capacity, represents the solid thermal resistance, represents the solid heat capacity, is the temperature of the indoor solid, represents the outdoor temperature, indicates the on-off state of the air conditioner, indicates the thermal power of the air conditioner.

[0085] 1.2) Since , , and cannot be directly measured in practice, a genetic algorithm (GA) is used to estimate these optimal parameters. First, the parameter set is encoded into the form of a chromosome, and each parameter combination is represented as an individual. By random initialization, an initial population of a certain size is generated, denoted as , where P is the population size. Next, the fitness of each individual is calculated by simulating the initial predicted value of the indoor temperature using the ETP model, and comparing it with the real temperature , using the root mean square error (RMSE) as the evaluation metric:

[0086]

[0087] where represents the initial predicted value of the indoor temperature output by the ETP model, is the mean square error between the predicted temperature of the ETP model and the real indoor temperature.

[0088] In each iteration process, crossover and mutation operations are performed to generate a new generation of population, and the above process is repeated. The entire genetic process continues until the termination condition is met, such as the maximum number of iterations is reached, or the RMSE converges to a set threshold. The final output of the optimal individual parameter set makes the ETP model prediction error minimum.

[0089] The optimal parameters estimated by the genetic algorithm not only improve the model prediction accuracy, but also maintain the physical reasonableness and interpretability of the model structure.

[0090] Step 2, establish the TimesNet model, construct the input and output feature set of the model, and obtain the corrected value result of the indoor temperature, which includes:

[0091] 2.1) Establish the TimesNet model to capture the influence of user dynamic behavior on indoor temperature. The TimesNet model divides the original input feature sequence into k frequency components through frequency decomposition, and applies a two-dimensional convolution structure to each component for feature extraction, thereby modeling both local (hourly, daily) and global (weekly) dynamic features.

[0092] Assume that the first i processing module of TimesNet receives the input as a one-dimensional sequence , which is processed in two dimensions by first extending it to the appropriate length through a padding operation and reshaping it into a two-dimensional tensor , is further input to an improved feature extraction module that extracts its local and periodic features. The extracted two-dimensional features are then converted back to one-dimensional space and aggregated according to the frequency importance weighting.

[0093] 2.2) The input of TimesNet model is physical features reflecting user dynamic behavior, including the electrical power of indoor devices such as HVAC system, lighting devices, and socket loads, as well as the humidity and illuminance of the indoor. The output of TimesNet model is a temperature correction term to compensate for the disturbance term that is not modeled in the traditional ETP model. The input-output relationship can be represented as:

[0094]

[0095] where represents the electrical power of indoor devices, represents the humidity and illuminance of the indoor, is the fitting function of TimesNet model.

[0096] Step 3, establish a fusion model and complete accurate prediction of indoor temperature, specifically including:

[0097] 3.1) As shown in Figure 3 , in the proposed hybrid modeling framework, the ETP model first generates a baseline value of indoor temperature prediction based on static physical parameters; at the same time, TimesNet generates a temperature correction term to capture the residual changes caused by dynamic occupant behavior and environmental disturbances. The ETP output and TimesNet correction are directly superimposed to obtain the final indoor temperature prediction, represented as:

[0098]

[0099] Step 4, realize group-level regulation of indoor temperature based on model predictive control, and obtain the temperature setting interval, specifically including:

[0100] 4.1) At each control time, MPC solves an optimization problem to dynamically adjust the upper and lower limits of the temperature set point to achieve a trade-off between energy saving and comfort. The decision variables of this optimization problem are the upper limit and the lower limit of the set temperature. To determine the final temperature set point ​, the system generates m random samples within the interval [ , ] and takes their average:

[0101]

[0102] where, is the qth randomly generated sample, m is the number of samples.

[0103] 4.2) Set temperature is used to determine the required thermal power of the HVAC system at the current time to achieve the target temperature regulation. The state evolution of the control system is described by the following dynamics equation:

[0104]

[0105] 4.3) The objective function of the MPC is to minimize the total cost function n within the prediction horizon , which is composed of two parts: energy cost and thermal comfort cost, weighted by energy weight coefficient and thermal comfort weight coefficient:

[0106]

[0107] where, is the energy cost, is the thermal discomfort cost, is the energy weight coefficient, is the thermal comfort weight coefficient.

[0108] Energy cost is calculated based on the power consumption of the HVAC and the real-time electricity price:

[0109]

[0110] where, is the electric-thermal power conversion efficiency of the HVAC system, is the time step, is the real-time electricity price at the current time.

[0111] Thermal discomfort cost is estimated by a simplified PMV (Predicted Mean Vote) model:

[0112]

[0113] where, is the current relative humidity, is the metabolic rate of the human body, represents the clothing thermal resistance.

[0114] 4.4) To ensure the safe operation of the system and the comfort of the user, the MPC optimization problem needs to satisfy the following set of constraints:

[0115] 1). Power constraint:

[0116] 2). Temperature set range constraint:

[0117] 3). Comfort constraint:

[0118] 4). Temperature set interval width constraint:

[0119] where, represents the maximum thermal power of the air conditioner, represents the lower bound of the temperature set interval, represents the upper bound of the temperature set interval, represents the width of the temperature set interval.

[0120] Step 5, based on the deep Q network, the individualized regulation of indoor temperature is realized, and the temperature setting value is obtained, specifically including:

[0121] 5.1) The deep Q network dynamically selects the optimal temperature set point within the temperature set interval defined by the first stage MPC , to realize the coordinated response of the environment state and individualized demand.

[0122] 5.2) In practical applications, the user's thermal comfort preference usually has periodic variation characteristics. In order to dynamically obtain the user's preferred temperature, the present application constructs a time-varying individual preference temperature :

[0123]

[0124] where, represents the indoor temperature value at time d in the t th past working day. With the passage of time, the average window moves gradually, so that the preference temperature can reflect the dynamic adjustment of seasonal changes and user behavior habits.

[0125] 5.3) The control process is modeled as a Markov Decision Process (MDP), which includes the following elements:

[0126] 1). State : composed of current indoor temperature , electricity price , outdoor temperature , indoor relative humidity and user preferred temperature ;

[0127] 2). Action : set one value in the discretized temperature setpoint collection, satisfying ;

[0128] 3). Reward function : used to balance energy consumption and user comfort, defined as follows:

[0129]

[0130] where represents the weight coefficient of energy consumption cost , and represents the weight coefficient of user preference cost, respectively used to control the trade-off relationship between energy consumption cost and preference temperature difference.

[0131] 5.4) At each time step, the agent selects an action according to the current state and - the greedy strategy selects an action , randomly selects an action with a probability , and performs the current optimal action with a probability of 1- . The agent then interacts with the environment to obtain a new state and a reward , and stores this experience in the experience replay buffer. The experience replay mechanism can break the time correlation, improve the sample efficiency and training stability.

[0132] 5.5) The current state is input into the deep Q network to predict the current Q value , while the target Q network is used to estimate the maximum Q value corresponding to the optimal action at the next time, and the target value is calculated as:

[0133]

[0134] where is the discount factor used to balance short-term and long-term returns, is the action at the next time. The following mean square error loss function is minimized:

[0135]

[0136] The loss is used to optimize the parameters of the Q network through a back propagation algorithm. To ensure the stability of the training, the target Q network synchronizes the parameters of the main Q network every few steps through a soft update or copy strategy to avoid model oscillation or divergence.

[0137] Further, the application also provides a user-centered HVAC system two-stage optimization control system, comprising a hybrid modeling module, a double-layer optimization module and a personalized learning module. The hybrid modeling module inputs building physical structure parameters, historical sensor data and user behavior data, and outputs future temperature prediction values via an ETP-TimesNet model. The double-layer optimization module solves the rolling control according to the predicted temperature and electricity price prediction, and generates temperature setting upper and lower limits and final control actions. The personalized module estimates the user target temperature based on 7-day historical temperature preferences as the control reference of DQN.

[0138] The hybrid modeling module fuses the modeling advantages of physical mechanism models and data-driven models, describes the thermal inertia and heat transfer characteristics of the building structure based on the ETP model, and generates the baseline value of the room temperature. The TimesNet takes the energy consumption data of HVAC, lighting, socket load and environmental parameters as inputs, captures the nonlinear disturbance characteristics caused by user behavior, and outputs the correction value. The final prediction result is the weighted superposition of the two, thereby realizing the joint modeling of static building physical properties and dynamic human behavior disturbance, and effectively improving the prediction accuracy.

[0139] The double-layer optimization module consists of two layers: the upper layer is based on the MPC method, which optimizes the indoor temperature setting upper and lower limits under the consideration of electricity price, environmental parameters and comfort constraints, and provides robust regulation boundaries for system operation; the lower layer is based on constrained DQN, which further optimizes the actual set point within the feasible solution space output by MPC, and realizes the dynamic response to individual preferences. This structure combines the stability of MPC and the adaptive ability of DQN, effectively solving the problem that traditional controllers are difficult to simultaneously consider system energy efficiency and individual comfort.

[0140] The personalized learning module extracts the preference value through a sliding window, the preference value being the room temperature data of the user in the historical 7-day working hours, and calculates the individual expected temperature at each time; and introduces the preference value in the DQN state space, and introduces the "square of the deviation between the temperature set point and the preference temperature" term in the reward function, so that the controller can constrain the deviation of the set value while optimizing the energy consumption cost, thereby realizing the dynamic learning of preferences and the maintenance of comfort, and improving the individualized control ability of the system and the user acceptance.

[0141] It will be obvious to a person skilled in the art that the application is not limited to the details of the foregoing exemplary embodiments and can be implemented in other concrete forms without departing from the spirit or essential characteristics of the application. The embodiments are therefore to be considered in all respects as illustrative and not restrictive, the scope of the application being indicated by the appended claims rather than by the foregoing description, and all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein. No reference signs in the claims should be considered as limiting the scope of the claims to the identity of the reference signs therein.

[0142] Furthermore, it should be understood that although the description is made on the basis of the embodiments, not every embodiment contains only one independent technical solution, and the description of the specification is only for the sake of clarity, and those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that those skilled in the art can understand.

Claims

1. A user-centric two-stage optimization control method for HVAC systems, characterized in that, Includes the following steps: Step 1: Establish a second-order equivalent thermal parameter model, obtain the online identification results of the initial parameters of the model through genetic algorithm iteration, and obtain the initial predicted value of indoor temperature. Step 2: Build a TimesNet model, construct the model's input and output feature sets, and obtain the corrected indoor temperature values. Step 3: Establish a fusion model and complete accurate prediction of indoor temperature; Step 4: Implement group-level control of indoor temperature based on model predictive control to obtain the temperature setpoint range; Step 5: Implement personalized indoor temperature control based on deep Q-network to obtain the temperature setpoint; Step 1 specifically includes: 1.1) A second-order ETP model is established to describe the heat exchange process inside the building. The second-order ETP model includes four basic parameters: gas thermal resistance, gas heat capacity, solid thermal resistance, and solid heat capacity. It considers the heat exchange between indoor air, building structure, and external environment. The thermal dynamic behavior of the system is described by a set of ordinary differential equations: in, Indicates the current moment. Indicates indoor temperature. Indicates the gas thermal resistance. Indicates the heat capacity of a gas. Indicates the thermal resistance of a solid. Indicates the heat capacity of a solid. It is the temperature of the solid inside the room. Indicates the outdoor temperature. Indicates the on / off status of the air conditioner. Indicates the thermal power of the air conditioner; 1.2) Estimation using a Genetic Algorithm (GA) , , and Optimal parameters: First, the parameter set Encoded in chromosome form, each parameter combination represents an individual. Through random initialization, an initial population of a certain size is generated, denoted as . Where P is the population size; for each individual Calculate its fitness level (Fitness) and obtain the initial predicted value of indoor temperature through ETP model simulation. and the actual temperature For comparison, the root mean square error (RMSE) is used as the evaluation metric: in, This represents the initial predicted indoor temperature output by the ETP model. This is the mean squared error between the temperature predicted by the ETP model and the actual indoor temperature. In each iteration, a crossover and mutation operation is performed to generate a new generation of the population, and this process is repeated until the termination condition is met, ultimately outputting the optimal set of individual parameters. This minimizes the prediction error of the ETP model; Step 2 specifically includes: 2.1) A TimesNet model is established to capture the impact of user dynamic behavior on indoor temperature. The TimesNet model uses frequency decomposition to divide the original input feature sequence into... k Each frequency component is identified, and a two-dimensional convolutional structure is applied to each component for feature extraction, thereby simultaneously modeling local and global dynamic features; Assuming the TimesNet model's first... i Each processing module receives a one-dimensional sequence as input. To perform two-dimensional processing on the sequence, first, it is expanded to a suitable length by padding with zeros, and then reshaped into a two-dimensional tensor. , The data is further fed into an improved feature extraction module to extract local and periodic features. The extracted two-dimensional features are then transformed back into one-dimensional space and aggregated based on frequency importance. 2.2) The input to the TimesNet model is physical characteristics reflecting the user's dynamic behavior, and the output of the TimesNet model is... This is a temperature correction term used to compensate for disturbances that are not modeled in the traditional ETP model. The input-output relationship is expressed as follows: in, Indicates the electrical power of indoor equipment. Indicates indoor humidity and illuminance. It is the fitting function for the TimesNet model; Step 3 specifically includes: 3.1) In the proposed hybrid modeling framework, the ETP model first generates a baseline value for indoor temperature prediction based on static physical parameters; simultaneously, TimesNet generates a temperature correction term, capturing residual changes caused by dynamic occupant behavior and environmental disturbances. The ETP output and the TimesNet correction are directly superimposed to obtain the final indoor temperature prediction, expressed as: ； Step 4 specifically includes: 4.1) At each control moment, MPC dynamically adjusts the upper and lower limits of the temperature setpoint by solving an optimization problem to achieve a trade-off between energy saving and comfort. The decision variable of the optimization problem is the upper bound of the setpoint temperature. and the lower realm To determine the final temperature setpoint used for control The system is in the interval [ , Generate m random samples within the specified range and take their average: in, Let q be the q-th randomly generated sample. m Indicates the number of samples; 4.2) Set temperature Used to determine the thermal power required by the HVAC system at the current moment. To achieve the target temperature regulation, the state evolution of the control system is described by the following dynamic equation: 4.3) The objective function of MPC is to minimize the prediction time domain. n Total cost function within It is composed of two weighted components: energy consumption cost and thermal comfort cost. in, Indicates energy consumption cost, Indicates the cost of thermal discomfort. This represents the energy consumption weighting coefficient. This represents the weighting coefficient for thermal discomfort; Energy consumption cost HVAC power consumption and real-time electricity price calculations: in, This indicates the electrothermal power conversion efficiency ratio of the HVAC system. Indicates the time step. This indicates the real-time electricity price at the current moment; Thermal discomfort cost Estimation using a simplified PMV model: in, Indicates the current relative humidity. Indicates the human metabolic rate. Indicates the thermal resistance of the clothing; 4.4) To ensure system security and user comfort, the constraints of the MPC optimization problem are as follows: 1) Power constraints: ; 2) Temperature set range constraints: ; 3) Comfort constraints: ; 4) Temperature setpoint range width constraint: ; in, This indicates the maximum thermal power of the air conditioner. This indicates the lower bound of the temperature setting range. This indicates the upper limit of the temperature setting range. This indicates the width of the temperature setting range.

2. The user-centered two-stage optimization control method for HVAC systems according to claim 1, characterized in that, Step 5 specifically includes: 5.1) The temperature setpoint range defined by the deep Q-network in the first stage MPC [ , Within the system, the optimal temperature setpoint is dynamically selected to achieve a coordinated response between environmental conditions and individual needs. 5.2) A time-varying individual preferred temperature is constructed by weighting indoor temperatures for the same time period over the past seven working days using a moving average method. : in, Indicates the first d A moment in the past workday t Indoor temperature value; 5.3) Model the control process as a Markov decision process, which includes the following elements: 1) Status Based on the current indoor temperature Electricity price Outdoor temperature Indoor relative humidity and user preferred temperature composition; 2) Actions : is a value in the discretized set of temperature setpoints that satisfies ; 3) Reward Function Used to balance energy consumption and user comfort, defined as follows: in, Indicates energy consumption cost The weighting coefficients, The weighting coefficients representing user preference costs are used to control the trade-off between energy consumption costs and preferred temperature differences. 5.4) At each time step, the agent adjusts its actions based on the current state. and Greedy strategy selects actions With probability Randomly select an action with a probability of 1- After performing the current optimal action, the agent then interacts with the environment to obtain a new state. and rewards and this experience Stored in the experience replay buffer; 5.5) Current Status It is fed into a deep Q-network to predict the current Q value. Meanwhile, the target Q network Used to estimate the maximum Q value corresponding to the optimal action at the next moment, and to calculate the target value: in, This is a discount factor used to balance short-term and long-term returns. The action at the next moment is to minimize the following mean squared error loss function: The loss is used to optimize the parameters of the Q-network using the backpropagation algorithm.

3. The user-centered two-stage optimization control method for HVAC systems according to claim 2, characterized in that, The target Q network synchronizes the parameters of the master Q network every few steps using a soft update or replication strategy.

4. A system employing the user-centered two-stage optimization control method for HVAC systems as described in any one of claims 1-3, characterized in that, include: The system comprises a hybrid modeling module, a two-layer optimization module, and a personalized learning module. The hybrid modeling module takes into account building physical structural parameters, historical sensor data, and user behavior data, and outputs future temperature predictions via the ETP-TimesNet model. The two-layer optimization module performs rolling control solutions based on predicted temperature and electricity price forecasts, generating upper and lower limits for temperature settings and the final control action. The personalized module estimates the user's target temperature based on 7-day historical temperature preferences, serving as a control reference for DQN.

5. The user-centric two-stage optimization control system for HVAC systems according to claim 4, characterized in that, The dual-layer optimization module consists of two layers: the upper layer is based on the MPC method, which optimizes the upper and lower limits of the indoor temperature setting under the conditions of electricity price, environmental parameters and comfort constraints, providing a robust control boundary for system operation; the lower layer is based on constrained DQN, which further optimizes the actual set point within the feasible solution space output by MPC, realizing a dynamic response to individual preferences.

6. The user-centric two-stage optimization control system for HVAC systems according to claim 5, characterized in that, The personalized learning module extracts preference values ​​through a sliding window. The preference values ​​are the room temperature data of the user during the historical 7-day working period, and calculate the individual expected temperature at each moment. The preference values ​​are introduced into the DQN state space, and the reward function introduces the term "squared deviation between the temperature setpoint and the preferred temperature", so that the controller can constrain the deviation of the setpoint while optimizing energy consumption costs.

Images