A HVAC optimization control method based on scene recognition and demand prediction

Through the combination of the LSTM model and PSO optimization algorithm, multi-source data is collected in real time for scene identification and demand prediction, and the optimal control strategy is generated, which solves the problem of low control delay and accuracy in HVAC systems, and achieves high efficiency and energy saving and Reduce emissions.

CN118729501BActive Publication Date: 2024-11-12ASIAINFO WANLIAN TECHNOLOGY (NANJING) CO LTD
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202411217593.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-02
Publication Date
2024-11-12
Estimated Expiration
2044-09-02

AI Technical Summary

Technical Problem

The existing HVAC system control technology has problems such as control delay, inaccurate load prediction, low control accuracy and high system complexity, resulting in waste of energy and inefficiency.

Method used

Using scene recognition and demand prediction methods based on LSTM model, multi-source data is collected in real time, and the optimal control strategy is generated to dynamically adjust the operating parameters of the HVAC system. The PSO optimization algorithm is used to achieve the trade-off between user comfort and energy consumption, and to generate specific HVACs. System control commands.

Benefits of technology

It improves system efficiency, reduces energy consumption, achieves the goal of energy conservation and emission reduction, improves the response speed and prediction accuracy to sudden load changes, and solves the problem of low control delay and accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118729501B_ABST
    Figure CN118729501B_ABST
Patent Text Reader

Abstract

The present invention provides an HVAC optimization control method based on scene recognition and demand prediction, including the following steps: collecting key variable data from multiple data sources in an independent space; performing data cleaning on the collected data; extracting useful features from the preprocessed data to construct a feature vector; inputting the feature vector into an LSTM model to obtain scene classification probabilities and demand prediction values; dynamically adjusting the operating parameters of the HVAC system according to the scene recognition and demand prediction results; and generating specific HVAC system control instructions according to the optimal control strategy. The present invention proposes an innovative HVAC optimization control method through multi-source data collection, the application of LSTM models, and the use of multi-objective optimization models. Compared with the prior art, the present invention has higher creativity and novelty, can effectively solve the problems of control delay and inaccurate prediction in the prior art, and improve the overall efficiency of the system and user comfort.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of optimization control of a heating, ventilation and air conditioning (HVAC) system, and in particular to an HVAC optimization control method based on scene recognition and demand prediction. Background Art

[0002] Currently popular HVAC optimization control technologies include:

[0003] Adopting variable frequency regulation technology, the operation mode can be flexibly changed according to external conditions during use, and the basic elements can be maintained even in energy-saving operation;

[0004] Introducing the spring / autumn transition mode, when the system enters the transition season mode, the system will automatically adjust the fresh air volume of the air conditioning unit according to the schedule to ensure indoor comfort. In the transition season, fresh air will be used as much as possible. When the temperature fluctuates, the system will only maintain the minimum fresh air supply because it has no cooling or heating capabilities;

[0005] Refrigeration station group control: Each control quantity in the cold source system will affect the characteristics of the refrigeration machine. Relying solely on the "selfish" control of each of them cannot achieve the goal of overall energy saving. The refrigeration station group control provides the peripheral conditions for the efficient operation of the refrigeration machine to the maximum extent, thereby optimizing the operation of the chilled water pump, cooling water pump, and cooling tower fan, so that the entire refrigeration room can achieve the maximum energy saving effect.

[0006] With the increase in building energy consumption and the improvement of environmental protection requirements, the optimization control of HVAC systems has become an important means of energy conservation and emission reduction. The existing HVAC system control technologies can be summarized as feedback control and feedforward control, but these methods have control delays, resulting in system inefficiency and energy waste.

[0007] Patent CN116147128A discloses a HVAC energy-saving control method and system, which mentions that by collecting system characteristic data, weather forecasts and historical data, a combined prediction algorithm combining a quadratic exponential smoothing algorithm with a neural network model is used to perform real-time high-precision load prediction, generate an optimized control strategy, and combine it with the basic control strategy to improve the operating efficiency and energy efficiency of the HVAC system while ensuring the safe and stable operation of the system. However, it relies on the collection and processing of a large amount of data, resulting in high system complexity and cost. At the same time, since the training of the neural network model requires a lot of time and data, and is sensitive to the initial parameters, it needs to be carefully adjusted to avoid overfitting or underfitting; this results in the technology being slow to respond to sudden load changes, poor prediction accuracy, and unable to combine seasonal time characteristics, resulting in control delays and low system efficiency.

[0008] Patent CN117287802A discloses a central air conditioning system energy-saving logic control method, which mentions that the dynamic and intelligent management of the central air conditioning system is achieved by combining CO2 sensor monitoring, big data analysis, artificial intelligence optimization and machine learning models. Although this technology provides innovative solutions for energy-saving control of central air conditioning systems, it is too dependent on sensor accuracy and data quality. In other words, when CO2 sensors and big data analysis are inaccurate, it will directly affect the control effect. There is a lack of comprehensive utilization of multi-source data, and the accuracy and adaptability are insufficient.

[0009] Therefore, in order to solve the above-mentioned drawbacks, the present invention proposes an HVAC optimization control method based on scene recognition and demand prediction to provide an optimal solution to this industry defect. Summary of the invention

[0010] In view of the shortcomings of the prior art, the purpose of the present invention is to provide an HVAC optimization control method based on scene recognition and demand prediction, in order to solve the problems of control delay, inaccurate load prediction, low control accuracy and high system complexity in the prior art. The present invention collects multi-source data in real time, uses the LSTM model for scene recognition and demand prediction, generates an optimal control strategy to dynamically adjust the operating parameters of the HVAC system, thereby improving system efficiency, reducing energy consumption, and achieving the goal of energy conservation and emission reduction. In order to achieve the above purpose, the present invention is implemented through the following technical solutions: A HVAC optimization control method based on scene recognition and demand prediction, comprising:

[0011] S1. Data collection and preprocessing

[0012] Collect key variable data representing the building energy consumption at the current time step based on multiple data sources in independent spaces, remove abnormal key variable data by setting expected ranges to ensure data accuracy and reliability, remove noise using smoothing filters, and fill in missing data through linear and / or polynomial interpolation to ensure data integrity;

[0013] S2. Feature extraction and normalization

[0014] Perform feature extraction on the processed key variable data to ensure that all variable data are aligned at the same time step t and then construct the feature vector F i,t ;

[0015] S3, scene recognition and demand prediction

[0016] The feature vector F i,t Input into the LSTM model for scene recognition and demand forecasting, and obtain the scene classification probability P(C i | F i,t) and the demand forecast value D at the future time step t+k i,t+k , according to the demand forecast value D i,t+k Generate a dynamic demand plan, indicating the demand in the future t+k time;

[0017] S4. Generation of optimal control strategy

[0018] Based on the scene recognition and demand prediction results provided by the LSTM model, the multi-objective optimization model is applied to generate the optimal control strategy to efficiently and comfortably manage the energy consumption demand and environmental control of the independent space. The process is as follows:

[0019] S41, through the user comfort function U i,t And the energy consumption function E i,t By weighted sum method, a comprehensive objective function J is constructed to quantitatively evaluate the performance of the HVAC system and quantify the multi-objective optimization problem: ,

[0020] In the formula, α and β represent the weight coefficients of user comfort and energy consumption; U i,t E refers to the user comfort level of independent space i at time step t; i,t refers to the energy consumption of independent space i at time step t; where, , ,

[0021] Where, T target Expressed as temperature target value; H target Expressed as humidity target value; C target is represented by the target value of CO2 concentration; WT, WH and WC are the weight parameters of temperature, humidity and CO2 concentration respectively; P i,t It is represented as the power of the HVAC system of independent space i at time step t; △t is represented as the time step;

[0022] S42, using the PSO optimization algorithm to solve the objective function J, obtain the control variables that make the objective function J optimal, and achieve the goals of maximizing environmental control and minimizing energy consumption demand;

[0023] S5. HVAC system control command generation

[0024] The optimal control variables obtained according to the PSO optimization algorithm and the dynamic demand plan are converted into a specific HVAC system control instruction operation plan within the future t+k time, and instructions are sent according to the instruction plan to achieve environmental control.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] In order to solve the problems in the prior art that the feedback control and feedforward control methods have lags in response to load changes, resulting in low system efficiency, the traditional load forecasting methods are slow to respond to sudden load changes, have poor forecasting accuracy, and fail to combine seasonal and operational planning and other human influences, the existing HVAC control methods are not accurate enough when adjusting the operating status of air-conditioning equipment, resulting in poor energy-saving effects; and the existing intelligent group control system is based on DDC (direct digital control system), which is complicated to program and debug, has high costs, and the solidified group control logic is prone to failure. The present invention collects multi-source data in real time, uses LSTM models for scene recognition and demand forecasting, generates optimal control strategies to dynamically adjust the operating parameters of the HVAC system, thereby improving system efficiency, reducing energy consumption, and achieving the goal of energy conservation and emission reduction. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The disclosure of the present invention is described with reference to the accompanying drawings. It should be understood that the drawings are only for illustrative purposes and are not intended to limit the scope of protection of the present invention. In the drawings, the same reference numerals are used to refer to the same components. Among them:

[0028] Figure 1 It is a flowchart of an HVAC optimization control method based on scene recognition and demand prediction proposed in one embodiment of the present invention. DETAILED DESCRIPTION

[0029] It is easy to understand that according to the technical solution of the present invention, without changing the essential spirit of the present invention, a person skilled in the art can propose a variety of interchangeable structural modes and implementation modes. Therefore, the following specific implementation modes and drawings are only exemplary descriptions of the technical solution of the present invention, and should not be regarded as the whole of the present invention or as a limitation or restriction to the technical solution of the present invention.

[0030] The present invention is further described in detail below with reference to the accompanying drawings, but is not intended to limit the present invention.

[0031] As an understanding of the technical concept and implementation principle of the present invention, the main purpose of the HVAC optimization control method based on scene recognition and demand prediction proposed in the present invention is to solve the problems of control delay, inaccurate load prediction, low control accuracy and high system complexity in the prior art. It is proposed to collect multi-source data in real time, use LSTM model for scene recognition and demand prediction, generate optimal control strategy to dynamically adjust the operating parameters of the HVAC system, thereby improving system efficiency, reducing energy consumption, and achieving the goal of energy saving and emission reduction.

[0032] like Figure 1 As shown, as an embodiment of the present invention, a HVAC optimization control method based on scene recognition and demand prediction is proposed, comprising the steps of:

[0033] S1. Data collection and preprocessing

[0034] Collect key variable data representing the building energy consumption at the current time step based on multiple data sources in independent spaces, and remove abnormal key variable data by setting expected ranges to ensure data accuracy and reliability;

[0035] Smoothing filters are used to remove noise and missing data are filled in through linear and / or polynomial interpolation to ensure data integrity.

[0036] Based on the above technical concept, it can be understood that the comprehensiveness and accuracy of data are improved through multi-source data collection in step S1, ensuring real-time control and optimization of the HVAC system. The specific implementation steps include:

[0037] S11. Collect environmental data, personnel activity data, equipment status data, operation plan data and external information data from multiple data sources in the independent space, wherein:

[0038] Environmental data include temperature, humidity, CO2 concentration, light intensity, and noise level. i,t The temperature of independent space i at time t is collected in real time through temperature sensors; humidity H i,t The purpose is to collect the humidity of independent space i at time t in real time through the humidity sensor; CO2 concentration C i,t The purpose is to collect the concentration of independent space i at time t in real time through the CO2 sensor; the light intensity L i,t The purpose is to measure the intensity of independent space i at time t, and collect the data in real time through the light sensor; the noise level N i,t The aim is to measure the noise level of independent space i at time t and collect the data in real time through noise sensors.

[0039] Personnel activity data includes the change in the number of people entering and leaving △P i,t and the total flow of people ∑P i,t . Change in number of people △P i,t The purpose is to collect the data of the number of people entering and leaving independent space i at time t in real time through access control system or people counting equipment; the total flow of people ∑P i,t , aims to collect the total flow of people in independent space i at time t in real time through people counting equipment.

[0040] Device status data includes device switch status S i,t and operating mode M i,t . Switching state S i,t The switch status of the equipment in independent space i at time t is collected in real time through the equipment management system. The status value can be on 1 or off 0; operation mode Mi,t , aims to collect the data of the equipment operation mode of independent space i at time t in real time through the equipment management system. The mode can be cooling, heating or ventilation.

[0041] Operation plan data includes activity type, activity time, equipment maintenance type, and maintenance time. Activity type A i,type , is the type of activity planned in the independent space i, such as meeting, training, etc., and this data is obtained through the schedule management system; activity time A i, start , A i, end , are the start and end time of the planned activities in the independent space i, respectively. The data is obtained through the schedule management system; equipment maintenance type M i,type , Maintenance time M i, start 、M i, end The equipment maintenance types planned in the independent space i are respectively, such as regular maintenance, emergency maintenance, etc. The data is obtained through the equipment management system; the start time and end time of the equipment maintenance are obtained through the equipment management system.

[0042] External information data includes outdoor temperature T outdoor,t , Outdoor humidity H outdoor,t and weather forecast information forecast,t . Outdoor temperature T outdoor,t , Outdoor humidity H outdoor,t They are the outdoor temperature and humidity of the area where the independent space is located at time t, and the data is obtained in real time through the meteorological data interface; the weather forecast information W forecast,t The weather forecast information of the area where the independent space is located at time t is obtained in real time through the meteorological data interface.

[0043] S121. Clean the collected data to remove outliers and noise to ensure the accuracy and reliability of the data. In specific implementation, the following operations are performed:

[0044] Example: For T i,t The expected range is set to -10°C to 50°C, and data outside this range are considered outliers and are recorded as ;H i,t The expected range of is set to 0% to 100%, denoted as For C i,t The expected range is set to 0ppm to 5000ppm, denoted as For L i,t The expected range is set to 0 to 100000 lux, denoted by For N i,t The expected range is set to 0 to 130 dB, denoted as .

[0045] S122. The calculation method for removing the noise of key variable data by using a smoothing filter is: (1)

[0046] In the formula, Represented as smoothed data; is represented as the original data; k is represented as half of the window size.

[0047] S123. After removing the noise of the key variable data by using a smoothing filter, the key variable data needs to be divided into time series data and non-time series data according to the time characteristics of the data. It should be noted that among the aforementioned collected data, time series data refers to data that changes with time, and each time step has a data value, while non-time series data does not change with time and is usually static data. Therefore,

[0048] For time series data, use linear interpolation to fill in missing values: (2)

[0049] In the formula, Represented as interpolated data; , Respectively represent the known data points before and after the missing point; t prev ,t next They represent the time steps before and after the missing point respectively;

[0050] For nonlinearly changing data, polynomial interpolation is used to fill in missing values: (3)

[0051] Where P(x) is the interpolation polynomial; yj is the value of the known data point; xj and xk are the time steps of the known data points.

[0052] In one embodiment of the present invention, the linear change data and the nonlinear change data are shown in the following table: .

[0053] S2. Feature extraction and normalization

[0054] Perform feature extraction on the processed key variable data to ensure that all variable data are aligned at the same time step t and then construct the feature vector F i,t The specific implementation steps include:

[0055] S21. Ensure that the data of all variables are aligned at the same time step t. All features should be collected and processed at the same time step t. The following operations are performed:

[0056] First, all collected data are sorted by timestamp, and then a uniform time step is set, such as 1 minute, to create a time index sequence covering the entire time range of the data set as a benchmark; then, all data are re-indexed according to this time index to ensure that the data is aligned with the uniform time step. For missing data points, they are processed through interpolation or filling methods to complete data alignment.

[0057] S22, concatenate all the feature data at each time step t after alignment to complete the feature vector F i,t Build:

[0058] F i,t =[T i,t , H i,t , C i,t , L i,t , N i,t , △P i,t ,∑P i,t , S i,t , M i,t , A i,type , A i, start , A i, end , M i,type , M i, start , M i, end , T outdoor,t , H outdoor,t , W forecast,t ] (4)

[0059] Based on the above technical concept, it can be understood that the feature vector F i,t The dimension is equal to the number of all feature data types 18. It should be noted that in the embodiments of the present invention, through flexible feature extraction and vector construction mechanisms, the effectiveness and adaptability of the system in the case of changes in data dimensions are guaranteed, that is, the feature vector can be increased or decreased for different usage environments and scenarios, and it does not affect the use of this method. For example, if there are no carbon dioxide collection sensors in some places, then the dimension of this data can be reduced or set to 0. In addition, if some environments have environmental data not listed in this example due to more advanced equipment, the dimension of the vector can also be increased to include the data, thereby enriching the scope of regulation.

[0060] Example: Assume that at a certain time step t, the collected feature data are as follows:

[0061] Temperature T i,t =25°C

[0062] Humidity i,t =60\%

[0063] CO2 concentration C i,t =400ppm

[0064] Light intensity L i,t =5000lux

[0065] Noise level N i,t =50dB

[0066] Change in the number of people entering and leaving DeltaP i,t =5

[0067] Total flow of people sumP i,t =20

[0068] Device switch status S i,t =1 (on)

[0069] Operation Mode M i,t =1 (cooling)

[0070] Activity Type A i,type =1 (meeting)

[0071] Event start time i, start =9:00

[0072] Event end time i, end =10:00

[0073] Equipment maintenance type M i,type =0 (no maintenance)

[0074] Equipment maintenance start time M i, start =None

[0075] Equipment maintenance end time M i, end =None

[0076] Outdoor temperature outdoor,t =30°C

[0077] Outdoor humidity H outdoor,t =70\%

[0078] Weather forecast information forecast,t =Sunny day

[0079] Then, the eigenvector F i,t =[25,60,400,5000,50,5,20,1,1,1,9:00,10:00,0,none,none,30,70,sunny].

[0080] S23, constructing the feature vector F i,t After that, and after the eigenvector F i,t Before inputting into the LSTM model, the feature vector Fi,t Each data point in is normalized: (5)

[0081] Calculate its mean and standard deviation, convert the data to zero mean and unit variance to prepare standardized data for subsequent model input, (6) (7)

[0082] In the formula, X represents the original data; Represented as normalized data; Expressed as mean; It is expressed as standard deviation; N is the total number of samples.

[0083] S3, scene recognition and demand prediction

[0084] The feature vector F i,t Input into the LSTM model for scene recognition and demand forecasting, and obtain the scene classification probability P(C i | F i,t ) and the demand forecast value D at the future time step t+k i,t+k , according to the demand forecast value D i,t+k Generate a dynamic demand plan, which represents the demand in the future t+k time.

[0085] Based on the above technical concept, it can be understood that the use of LSTM model for scene recognition and demand forecasting realizes dynamic control based on real-time data, improving the response speed and prediction accuracy of the system. In the HVAC optimization control method, scene recognition refers to identifying the usage or activity type of independent space at the current time step. The results of scene recognition are usually used to guide the generation of demand forecasting and control strategies. The following are several commonly used scene categories and their explanations:

[0086] Idle scenario: there is no activity in the independent space, with little or no personnel flow, and the characteristics are: stable temperature and humidity, low CO2 concentration, low light and noise levels; meeting scenario: a meeting or training is underway in the independent space with a large number of people, and the characteristics are: temperature and humidity may fluctuate, CO2 concentration is high, and light and noise levels are high; work scenario: daily work is being carried out in the independent space with frequent personnel activities, and the characteristics are: slight fluctuations in temperature and humidity, medium CO2 concentration, and medium light and noise levels; rest scenario: the independent space is in rest or lunch break time, with few personnel activities; the characteristics are: relatively stable temperature and humidity, low CO2 concentration, and low light and noise levels; equipment maintenance scenario: equipment maintenance is being carried out in the independent space, and there may be no personnel activities or only a small number of people; the characteristics are: the equipment switch status and operation mode will change, and the environmental parameters may be affected.

[0087] In one embodiment of the present invention, the LSTM model includes an input gate, a forget gate, a candidate memory and an output gate, and the construction steps include:

[0088] S3-1. Input feature vector F through the input gate i,t and the hidden state h at the previous moment t1 , calculate the input activation value i through the input gate t : (8)

[0089] Where W i Represented as the weight matrix of the input gate, with dimension (h+d)×h; b i It is represented as the bias term of the input gate, with dimension h=18, which is determined by the dimension of the input vector;

[0090] S3-2. Calculate the forget activation value f through the forget gate t : (9)

[0091] Where W f Represented as the weight matrix of the forget gate, with dimension (h+d)×h; b f It is represented as the bias term of the forget gate, with dimension h=18;

[0092] S3-3. Calculate candidate memory activation value through candidate memory : (10)

[0093] Where W C Represented as the weight matrix of candidate memory, with dimension (h+d)×h; b C It is represented as the bias term of the candidate memory, with dimension h=18;

[0094] S3-4. Update unit status C t : (11)

[0095] S3-5, calculate the output activation value O through the output gate t :

[0096] (12)

[0097] Where W O Represented as the weight matrix of the output gate, with dimension (h+d)×h; b O It is represented as the bias term of the output gate, with dimension h=18;

[0098] S3-6. Update hidden state h t : (13)

[0099] S3-7, output the scene classification probability P(C i | F i,t ): (14)

[0100] Where W sofmax is the weight matrix of the softmax layer, with a dimension of h×k; b sofmax is the bias term of the softmax layer, with a dimension of k=5, which is determined by the number of scene categories;

[0101] S3-8. Output the demand forecast value D at the future time step t+k through the fully connected layer i,t+k : (15)

[0102] Where W fc Represented as the weight matrix of the fully connected layer, with a dimension of h×1; b fc Represented as the bias term of the fully connected layer, with a dimension of 1.

[0103] Example: LSTM model calculates hidden state h through forward propagation t and the memory cell state C t , the scene classification probability P(C i | F i,t ), assuming that the output scene classification probability is:

[0104] P(C i | F i,t)=[0.1,0.2,0.4,0.2,0.2], we get that the category with the highest probability is 0.4, and the corresponding scene category is 3 (assuming it is a work scene).

[0105] At the same time, after the LSTM model is built, it needs to be trained based on historical data. The training steps include:

[0106] First, set the objective function of the LSTM model, where

[0107] For the task of scene recognition, the cross entropy loss function L is used CE : (16)

[0108] In the formula, yi represents the real scene category; Expressed as the predicted scene category probability;

[0109] For the demand forecasting task, the mean square error loss function L is used. MSE : (17)

[0110] In the formula, yi represents the real scene category; Expressed as the predicted scene category probability;

[0111] Secondly, the back propagation algorithm is used to calculate the gradient of the loss function to the LSTM model parameters; and the gradient descent method is used to update the model parameters, where the learning rate η is 0.001, the time step T is 20, the batch size B is 64, the iteration E is 100, and the weight matrix W i , W f , W C , W O and the bias term b i 、b f 、b C 、b O The settings are optimized through initialization (can be set to random values ​​or experience values) and training processes, thus ensuring that the LSTM model continues to join the training after obtaining new data during use, continuously updating the various parameters of the model, and having adaptability and scalability.

[0112] S4. Generation of optimal control strategy

[0113] According to the scene recognition and demand prediction results provided by the LSTM model, the multi-objective optimization model is applied to generate the optimal control strategy to efficiently and comfortably manage the energy consumption demand and environmental control of the independent space. It should be noted that in S3, the demand D at the future time step t+k is predicted using the LSTM model. i,t+k, the demand prediction value represents the energy consumption demand or environmental control demand of independent space i at the future time step t+k, and the multi-objective optimization model proposed in S4 is used to balance multiple objectives to find the optimal solution, with the aim of being able to balance user comfort, energy consumption and equipment life, thereby improving the overall efficiency of the system.

[0114] It is understandable that in HVAC system optimization, the usual goals include: 1. Maximizing user comfort: ensuring that the indoor environment (temperature, humidity, CO2 concentration, etc.) meets the user's comfort needs; 2. Minimizing energy consumption: reducing the energy consumption of the HVAC system and reducing operating costs; 3. Optimizing equipment life: extending the service life of the equipment and reducing maintenance costs.

[0115] In one embodiment of the present invention, when the multi-objective optimization model is used to generate the optimal control strategy, it is also necessary to:

[0116] S41, through the user comfort function U i,t And the energy consumption function E i,t The weighted sum method is used to construct the comprehensive objective function J and quantify the multi-objective optimization problem: (18)

[0117] In the formula, α and β represent the weight coefficients of user comfort and energy consumption; U i,t E refers to the user comfort level of independent space i at time step t; i,t refers to the energy consumption of independent space i at time step t; where, (19) (20)

[0118] Where, T target It is expressed as the temperature target value, which is 20°C to 25°C; H target Expressed as humidity target value, ranging from 40% to 60%; C target It is represented by the CO2 concentration target value, ranging from 400ppm to 800ppm; WT, WH and WC are the weight parameters of temperature, humidity and CO2 concentration, respectively, ranging from 0.1 to 1; P i,t It is represented as the power of the HVAC system of independent space i at time step t; △t is represented as the time step, which can be 5 minutes to 1 hour;

[0119] S42, using the PSO optimization algorithm to solve the objective function J, obtain the control variables that make the objective function J optimal, and achieve the goals of maximizing environmental control and minimizing energy consumption demand;

[0120] It should be noted that the control variables for achieving the goals of maximizing environmental control and minimizing energy consumption requirements include at least the temperature setting points T and T that optimize the objective function J. i,opt , humidity set point H i,opt and CO2 concentration set point C i,opt , whose expression is:

[0121] T i,opt =arg min T i,t J(T i,t , H i,t , C i,t , P i,t )、

[0122] H i,opt =arg min H i,t J(T i,t , H i,t , C i,t , P i,t )、

[0123] C i,opt =arg min C i,t J(T i,t , H i,t , C i,t , P i,t ),

[0124] Where, T i,t , H i,t , C i,t and P i,t They are respectively represented as the temperature, relative humidity, carbon dioxide concentration and human flow of independent space i at time t.

[0125] As an embodiment of the present invention, in S42, the control variable may also include the optimal ventilation setting point V i,opt , , it is understandable that V i,opt Represents the ventilation set point that optimizes the objective function J. If the HVAC system supports it, other control items can be added.

[0126] S5. HVAC system control command generation

[0127] The optimal control variables obtained according to the PSO optimization algorithm and dynamic demand plan are converted into a specific HVAC system control instruction operation plan within the future t+k time. Instructions are sent according to the instruction plan to achieve environmental control.

[0128] It should be noted that the control instructions include: adjusting the temperature set point, humidity set point and ventilation rate set point. In specific implementation, the system control instructions can only be issued by the system control, and users are prohibited from setting them by themselves. The generated control instructions include the temperature set point: TsetT_{\text{set}}Tset; the ventilation volume set point: VsetV_{\text{set}}Vset; and other equipment parameter set points: EsetE_{\text{set}}Eset.

[0129] As an embodiment of the present invention, in S5, the control instruction also includes a temperature adjustment factor ΔT and a ventilation volume adjustment factor ΔV, which are dynamically adjusted according to external environmental parameters: , where the ventilation adjustment factor △V is adjusted according to the CO2 concentration: .

[0130] Based on the above technical concept, it can be understood that in the optimization control of the HVAC system, in order to better adapt to the real-time changing environment and needs, the control instructions include not only the basic temperature, humidity and ventilation set points, but also some adaptive adjustment parameters, such as advance time parameters, temperature adjustment factors and ventilation adjustment factors. These adaptive adjustment parameters can be dynamically adjusted according to external environmental parameters, thereby improving the system's responsiveness and control effect. In specific implementation:

[0131] The principle of setting the advance time parameters is as follows: the advance time parameters are used to start or adjust the HVAC system in advance to ensure that the desired environmental conditions are achieved before a specific event occurs. For example, the cooling system is started 30 minutes before the meeting to ensure that the conference room reaches a comfortable temperature when the meeting starts. The purpose is to ensure that the environmental conditions (such as temperature and humidity) in the independent space reach the preset comfort standards when important events occur. Improve user satisfaction and avoid a decrease in comfort due to unsuitable environmental conditions. The implementation process is as follows:

[0132] a1. Determine the event time: Get the start time of the planned event, such as the start time of a meeting;

[0133] a2. Calculate the advance start time: Calculate the time when the system needs to start based on the default advance time parameter (such as 30 minutes);

[0134] a3. Adjust system start time: Start or adjust the HVAC system at the calculated advance time point to ensure that environmental conditions meet preset standards when the event begins.

[0135] The principle of the temperature adjustment factor is: the temperature adjustment factor dynamically adjusts the indoor temperature set point according to the change of the external temperature. For example, when the external temperature is high, the indoor temperature set point can be appropriately lowered to improve the indoor comfort; vice versa, its purpose is to dynamically adapt to the external temperature changes, ensure that the indoor environment is always within the comfortable range, improve the responsiveness and flexibility of the system, and enhance the user's comfort experience. Its implementation process is:

[0136] b1. Monitor external temperature: obtain external temperature in real time through external temperature sensor or meteorological data interface;

[0137] b2. Calculate the temperature adjustment factor: Calculate the temperature adjustment factor according to a preset adjustment rule or algorithm;

[0138] b3. Adjust the temperature set point: Dynamically adjust the indoor temperature set point based on the calculated temperature adjustment factor.

[0139] The principle of ventilation adjustment factor is: ventilation adjustment factor dynamically adjusts ventilation set point according to the change of indoor CO2 concentration. For example, when indoor CO2 concentration is high, increase ventilation to improve air quality; vice versa, its purpose is to dynamically adapt to the change of indoor CO2 concentration, ensure that air quality is always within the healthy range, improve the responsiveness and flexibility of the system, and enhance the user's comfort experience. Its implementation process is:

[0140] c1. Monitor indoor CO2 concentration: obtain indoor CO2 concentration in real time through CO2 sensor;

[0141] c2. Calculate ventilation volume adjustment factor: Calculate ventilation volume adjustment factor according to preset adjustment rules or algorithms.

[0142] As an embodiment of the present invention, the step of using the PSO optimization algorithm to solve the objective function J in S42 includes:

[0143] S421, particle initialization: randomly initialize the position and velocity of particles in the search space, each particle represents a control strategy, including a temperature set point, a humidity set point, and a ventilation set point;

[0144] S422, velocity initialization: The velocity of the particle is initialized to a small value, usually randomly generated in the range of [-V max, V max];

[0145] S423, parameter settings: group size (number of particles): for example, 30 particles; maximum number of iterations: for example, 100 times; inertia weight w: for example, 0.7; acceleration constants c_1 and c_2: for example, 1.5 and 1.5;

[0146] S424, fitness calculation: calculate the value of the comprehensive objective function J for each particle to evaluate its fitness;

[0147] S425, position and speed update, the process is as follows:

[0148] Establish the velocity update formula: , where v i, t+1 represents the velocity of particle i at time t+1; W represents the inertia weight; c1 and c2 represent acceleration constants; r1 and r2 represent random numbers between [0,1]; P i,best It is represented as the best historical position of particle i; g best Denoted as the global optimal position of the group;

[0149] Establish the position update formula: , where x i, t+1 It is represented by the position of particle i at time t+1; x i, t Expressed as the position of particle i at time t; v i, t+1 It is represented as the velocity of particle i at time t+1;

[0150] S426, Initialize the group: Initialize the position and velocity of the particles, set the initial global optimal position g best and the individual best position P i,best ;

[0151] S427, iterative process: calculate the fitness value of each particle; update the individual best position P i,best and the global optimal position g best ; Update the particle's velocity and position; Iterate until the maximum number of iterations is reached or the convergence condition is met

[0152] S428, optimal control strategy generation:

[0153] The optimal control strategy is obtained from the PSO algorithm, including the temperature set point, humidity set point, and ventilation set point. Based on the optimal control strategy, specific HVAC system control instructions are generated:

[0154] Example: Assume that the PSO algorithm is used and the following parameters are set:

[0155] Temperature target value T target =22°C

[0156] Humidity target value H target =50%

[0157] CO2 concentration target value C target =600ppm

[0158] User comfort weight α=0.7

[0159] Energy consumption weight β=0.3

[0160] Temperature weight WT=0.5.

[0161] Then, after the optimization algorithm runs, the optimal control strategy is: the optimal temperature setting point T i,opt =22.5°C; optimal humidity setting point H i,opt =49%; optimal CO2 concentration setting point C i,opt =590ppm; optimal ventilation setting point V i,opt =2000m³ / h.

[0162] Based on this, it can be understood that by solving the objective function J through the PSO optimization algorithm, a control strategy that optimizes the comprehensive objective function J can be found, so that user comfort and energy consumption are balanced between the optimal points.

[0163] The technical scope of the present invention is not limited to the contents in the above description. Those skilled in the art can make various deformations and modifications to the above embodiments without departing from the technical idea of ​​the present invention, and these deformations and modifications should all fall within the protection scope of the present invention.

Claims

1. A HVAC optimization control method based on scene recognition and demand prediction, characterized in that: Includes steps: S1. Data collection and preprocessing Collect key variable data representing the building energy consumption at the current time step based on multiple data sources in independent spaces, remove abnormal key variable data by setting expected ranges to ensure data accuracy and reliability, remove noise using smoothing filters, and fill in missing data through linear and / or polynomial interpolation to ensure data integrity; S2. Feature extraction and normalization Perform feature extraction on the processed key variable data to ensure that all variable data are aligned at the same time step t and then construct the feature vector F i,t ; S3, scene recognition and demand prediction The feature vector F i,t Input into the LSTM model for scene recognition and demand forecasting, and obtain the scene classification probability P(C i | F i,t ) and the demand forecast value D at the future time step t+k i,t+k , according to the demand forecast value D i,t+k Generate a dynamic demand plan, indicating the demand in the future t+k time; S4. Generation of optimal control strategy Based on the scene recognition and demand prediction results provided by the LSTM model, the multi-objective optimization model is applied to generate the optimal control strategy to efficiently and comfortably manage the energy consumption demand and environmental control of the independent space. The process is as follows: S41, through the user comfort function U i,t And the energy consumption function E i,t By weighted sum method, a comprehensive objective function J is constructed to quantitatively evaluate the performance of the HVAC system and quantify the multi-objective optimization problem: , In the formula, α and β represent the weight coefficients of user comfort and energy consumption; U i,t E refers to the user comfort level of independent space i at time step t; i,t refers to the energy consumption of independent space i at time step t; where, , , Where, T target Expressed as temperature target value; H target Expressed as humidity target value; C target is represented by the target value of CO2 concentration; WT, WH and WC are the weight parameters of temperature, humidity and CO2 concentration respectively; P i,t It is represented as the power of the HVAC system of independent space i at time step t; △t is represented as the time step; S42, using the PSO optimization algorithm to solve the objective function J, obtain the control variables that make the objective function J optimal, and achieve the goals of maximizing environmental control and minimizing energy consumption demand; S5. HVAC system control command generation The optimal control variables obtained according to the PSO optimization algorithm and the dynamic demand plan are converted into a HVAC system control instruction operation plan within a specific future time t+k, and instructions are sent according to the instruction operation plan to achieve environmental control.

2. The HVAC optimization control method based on scene recognition and demand prediction according to claim 1 is characterized in that: In S4, The control variables for achieving the goals of maximizing environmental control and minimizing energy consumption requirements include at least the temperature setting points T that optimize the objective function J. i,opt , humidity set point H i,opt and CO2 concentration set point C i,opt , whose expression is: T i,opt =arg min T i,t J ( T i,t ,H i,t ,C i,t ,P i,t )、 H i,opt =arg min H i,t J ( T i,t ,H i,t ,C i,t ,P i,t )、 C i,opt =arg min C i,t J ( T i,t ,H i,t ,C i,t ,P i,t ?, Where, T i,t , H i,t , C i,t and P i,t They are respectively represented as the temperature, relative humidity, carbon dioxide concentration and human flow of independent space i at time t.

3. The HVAC optimization control method based on scene recognition and demand prediction according to claim 1, characterized in that: In S1, The key variable data at least includes environmental data, personnel activity data, equipment status data, operation plan data and external information data; The calculation method for removing the noise of key variable data using smoothing filter is: , In the formula, Represented as smoothed data; is represented as the original data; k is represented as half of the window size.

4. The HVAC optimization control method based on scene recognition and demand prediction according to claim 1 or 3, characterized in that: After removing the noise of the key variable data by using a smoothing filter, and before supplementing the missing key variable data by linear and / or polynomial interpolation, the key variable data also needs to be divided into time series data and non-time series data according to the time characteristics of the data; Then, for time series data, linear interpolation is used to fill in missing values: , In the formula, Represented as interpolated data; , Respectively represent the known data points before and after the missing point; t prev ,t next They represent the time steps before and after the missing point respectively; For nonlinearly changing data, polynomial interpolation is used to fill in missing values: , Where P(x) is the interpolation polynomial; yj is the value of the known data point; xj and xk are the time steps of the known data points.

5. The HVAC optimization control method based on scene recognition and demand prediction according to claim 1, characterized in that: In S2, After all variable data are aligned at the same time step t, it is necessary to concatenate all feature data at each aligned time step t to complete the feature vector F i,t The construction of In constructing the feature vector F i,t After that, the feature vector F i,t Before inputting into the LSTM model, the feature vector F i,t Each data point in is normalized: , Calculate its mean and standard deviation, convert the data to zero mean and unit variance to prepare standardized data for subsequent model input, , , In the formula, X represents the original data; Represented as normalized data; Expressed as mean; It is expressed as standard deviation; N is the total number of samples.

6. The HVAC optimization control method based on scene recognition and demand prediction according to claim 1, characterized in that: In S3, The LSTM model includes an input gate, a forget gate, a candidate memory, and an output gate. Its construction steps include: S3-1. Input feature vector F through the input gate i,t and the hidden state h at the previous moment t1 , calculate the input activation value i through the input gate t : , Where W i Represented as the weight matrix of the input gate, with dimension (h+d)×h; b i It is represented as the bias term of the input gate, with dimension h=18, which is determined by the dimension of the input vector; S3-2. Calculate the forget activation value f through the forget gate t : , Where W f Represented as the weight matrix of the forget gate, with dimension (h+d)×h; b f It is represented as the bias term of the forget gate, with dimension h=18; S3-3. Calculate candidate memory activation value through candidate memory : , Where W C Represented as the weight matrix of candidate memory, with dimension (h+d)×h; b C It is represented as the bias term of the candidate memory, with dimension h=18; S3-4. Update unit status C t : , S3-5, calculate the output activation value O through the output gate t : , Where W O Represented as the weight matrix of the output gate, with dimension (h+d)×h; b O It is represented as the bias term of the output gate, with dimension h=18; S3-6. Update hidden state h t : ; S3-7, output the scene classification probability P(C i | F i,t ): , Where W sofmax is the weight matrix of the softmax layer, with a dimension of h×k; b sofmax is the bias term of the softmax layer, with a dimension of k=5, which is determined by the number of scene categories; S3-8. Output the demand forecast value D at the future time step t+k through the fully connected layer i,t+k : , Where W fc Represented as the weight matrix of the fully connected layer, with a dimension of h×1; b fc Represented as the bias term of the fully connected layer, with a dimension of 1.

7. The HVAC optimization control method based on scene recognition and demand prediction according to claim 6 is characterized in that: In S3, After the LSTM model is constructed, it is trained based on historical data. The training steps include: First, set the objective function of the LSTM model, where For the task of scene recognition, the cross entropy loss function L is used CE : , In the formula, yi represents the real scene category; Expressed as the predicted scene category probability; For the demand forecasting task, the mean square error loss function L is used. MSE : , In the formula, yi represents the actual demand value; Expressed as predicted demand value; Secondly, the back-propagation algorithm is used to calculate the gradient of the loss function with respect to the LSTM model parameters; and the gradient descent method is used to update the model parameters.

8. The HVAC optimization control method based on scene recognition and demand prediction according to claim 7 is characterized in that: The LSTM model is trained by adopting a learning rate η of 0.001, a time step T of 20, a batch size B of 64, an iteration E of 100 times, and continuously optimizing the weight matrix and the bias term during the initialization and training process.

9. The HVAC optimization control method based on scene recognition and demand prediction according to claim 1, characterized in that: In S5, The HVAC system control instructions include at least a temperature set point, a humidity set point, a ventilation rate set point, a temperature adjustment factor, and a ventilation volume adjustment factor, wherein: The temperature set point, humidity set point and ventilation rate set point are issued by the HVAC system control; The temperature adjustment factor △T and ventilation adjustment factor △V are dynamically adjusted and controlled by the HVAC system according to external environmental parameters: , .

Citation Information

Cited By

  • Expressway service area heating and ventilation control method based on Internet of Things and neural network

    CN119901050A

  • HVAC Control Method for Highway Service Areas Based on Internet of Things and Neural Networks

    CN119901050B