Intelligent regulation and control method based on perception and dynamic partitioning

Abstract

The invention discloses an intelligent regulation and control method based on perception and dynamic partitioning. In order to solve the problem that an existing public building air conditioning system is difficult to deal with personnel activity intensity difference, emotion fluctuation and space density dynamic change, accurate recognition of personnel behaviors and activity density is achieved by fusing YOLOv5, OpenPose and LSTM networks, and the group emotion state is obtained by combining infrared thermal imaging and expression recognition. Based on behavior-density partitioning and an emotional load correction model, space thermal load distribution is dynamically judged, an NSGA-I multi-objective optimization algorithm is introduced, the air conditioner energy efficiency ratio and the thermal comfort degree are optimized, and differential intelligent control over regular and irregular partitioning is achieved through self-adaptive regulation and control of a meta-reinforcement learning control framework. According to the method, the building energy consumption is effectively reduced, the local comfort degree is improved, good real-time performance and intelligent adjusting capacity are achieved, and the method is suitable for optimization design of an intelligent air conditioning system of a large public building.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent control of building environment, and particularly relates to an intelligent regulation method based on perception and dynamic zoning, belonging to the technical direction of the intersection of artificial intelligence, building energy conservation and air-conditioning adaptive control, and is particularly suitable for the energy conservation optimization and thermal comfort control of air-conditioning systems in large public buildings. Background Art

[0002] With the acceleration of the urbanization process and the continuous increase in the usage frequency of public buildings, the energy consumption control and environmental comfort regulation of the internal air-conditioning systems in buildings have become important research directions in building intelligence. In the prior art, centralized air-conditioning systems are usually adopted in public buildings, and their control methods are mostly fixed zoning management or global unified regulation, making it difficult to achieve precise control according to the dynamic changes in the real-time distribution, activity intensity and emotional state of people.

[0003] In traditional systems, the air-conditioning regulation strategy mainly relies on environmental temperature and humidity parameters or time period settings, lacking the ability to perceive human behavior and space density, resulting in insufficient cooling load in areas with concentrated people, while over-cooling or over-heating may occur in sparsely populated areas, thus causing energy waste and a decline in local comfort. According to relevant research statistics, the energy consumption of air-conditioning systems in public buildings accounts for more than 40% of the total building energy consumption, and the ineffective energy consumption caused by unreasonable control strategies, delayed regulation response or inaccurate space load assessment can reach 15% to 30%.

[0004] In addition, with the development of computer vision and artificial intelligence, some studies have begun to explore using cameras or sensors to monitor the number and location of people to achieve dynamic zoning regulation, but there are still the following main problems:

[0005] (1) The behavior recognition dimension is single, and it is impossible to accurately distinguish the differences in human activity intensity;

[0006] (2) The consideration of the group emotional state and psychological comfort is lacking, affecting the overall thermal comfort evaluation;

[0007] (3) The space zoning rules lack self-adaptability and are difficult to cope with irregular load fluctuations;

[0008] (4) The control strategy is difficult to dynamically balance between comfort and energy efficiency, and the regulation response lags.

[0009] Therefore, there is an urgent need for a new type of air-conditioning regulation method that integrates multi-modal perception, intelligent zoning determination and multi-objective optimization control to improve the intelligence, self-adaptability and energy-saving effect of the system to meet the needs of regulating the dynamic environment inside complex public buildings. Summary of the Invention

[0010] In view of the above-mentioned disadvantages of the existing technologies, to solve the technical problems in the existing technologies, such as the difficulty for the air conditioning systems of public buildings to dynamically perceive the differences in the intensity of personnel activities, the fluctuations in group emotions, and the changes in space density, resulting in insufficient coordinated regulation of energy efficiency optimization and thermal comfort, rigid zoning rules, and lagging responses;

[0011] In one possible implementation, an intelligent regulation method based on perception and dynamic zoning is provided, including the following steps:

[0012] S1. Collect personnel activity data, use the YOLOv5 and OpenPose algorithms to obtain limb movements, and perform activity intensity classification through a time series prediction algorithm to realize the recognition of personnel behavior patterns in the space;

[0013] S2. Divide the space according to the personnel behavior patterns and space density, and use a multi-dimensional scoring model to determine regular zones and irregular zones;

[0014] S3. Combine infrared thermal imaging and facial expression recognition to extract the group emotional state, and construct an emotion correction factor to participate in the thermal load modeling;

[0015] S4. Construct an upper-layer multi-objective optimization module, and perform coordinated optimization of the energy efficiency ratio and thermal comfort based on the NSGA-II algorithm, and dynamically adjust the weights of the optimization objectives;

[0016] S5. Based on the classification results of regular and irregular zones and the output results of the upper-layer multi-objective optimization module, apply a lower-layer reinforcement learning regulation module, and use the Meta-MPC meta-reinforcement learning algorithm to generate air conditioning regulation actions in real time to realize the adaptive adjustment of the air conditioning system.

[0017] In one possible implementation, density recognition is achieved through YOLOv5 combined with DBSCAN clustering analysis, and the density factor is calculated using the clustering area and the number of personnel, so as to be divided into high-density areas and low-density areas.

[0018] In one possible implementation, the determination of the regular zone is achieved by calculating the spatio-temporal overlap rate, load fluctuation consistency, and regulation responsiveness of personnel activities, comprehensively constructing a scoring model, and determining whether the space belongs to a regular zone according to a set threshold.

[0019] In one possible implementation, the construction of the emotion correction factor combines the skin temperature change rate detected by thermal imaging and facial expression features, and forms an emotion regulation factor for thermal load correction through cross-modal attention network fusion.

[0020] In a possible implementation manner, the upper-layer multi-objective optimization module takes the energy efficiency ratio and thermal comfort of the air-conditioning system as optimization objectives, dynamically adjusts the optimization weights based on the personnel density, indoor-outdoor temperature difference, and emotion correction factor, and iteratively generates an optimal control scheme using the NSGA-II algorithm.

[0021] In a possible implementation manner, the lower-layer reinforcement learning control module adopts the Meta-MPC meta-reinforcement learning framework, constructs a meta-policy network using the Actor-Critic structure, and dynamically generates control actions by combining historical load and real-time perception data.

[0022] In a possible implementation manner, for regular partitions, the STL algorithm is used to decompose the trend of the historical load, the LSTM network is combined for load prediction, and control actions for precooling or preheating are generated according to the prediction results.

[0023] In a possible implementation manner, for irregular partitions, the meta-policy network is fine-tuned by real-time sampling to quickly adapt to new load changes and generate a control scheme.

[0024] In a possible implementation manner, the hybrid perception network consists of a Doppler ultrasonic array and a door magnetic sensor, and is used to detect the door and window states and the opening area in real time to achieve load disturbance compensation.

[0025] Based on the above technical solutions, the multi-objective intelligent control method for air conditioners in public buildings based on multi-modal behavior-emotion perception and dynamic zoning of the present invention performs behavior recognition by integrating the YOLOv5, OpenPose, and LSTM models, realizes emotion modeling by combining infrared thermal imaging and expression recognition, dynamically partitions and determines its regularity, introduces the NSGA-II algorithm for multi-objective optimization between energy efficiency and comfort, and finally completes differential intelligent control through the meta-reinforcement learning control strategy, effectively improving the energy efficiency and comfort of the system and meeting the air-conditioning control requirements under complex crowd behavior dynamics.

[0026] The multi-objective intelligent control method for air conditioners in public buildings based on multi-modal behavior-emotion perception and dynamic zoning proposed by the present invention has the following remarkable technical effects:

[0027] 1. By combining the YOLOv5, OpenPose, and LSTM algorithms, integrating image detection, skeleton recognition, and time-series modeling, it realizes the real-time recognition of the activity intensity, behavior pattern, and movement speed of personnel in different spaces, thereby realizing a dynamic behavior partition that is more in line with the actual heat load distribution, and improving the responsiveness and partition control accuracy of the air-conditioning system.

[0028] 2. By introducing a crowd density recognition mechanism based on DBSCAN clustering, a density factor model is constructed to effectively measure the local spatial aggregation degree, and it is combined with the activity type to form a sensible heat load correction model, making the heat load calculation more targeted and real-time.

[0029] 3. In terms of comfort modeling, a multi-modal emotion recognition method that fuses infrared thermal imaging and facial expressions is introduced for the first time, and a cross-modal attention network is constructed to extract the individual physiological temperature changes and emotional arousal states, forming an emotion correction factor to participate in the load modeling, effectively making up for the deficiency of traditional thermal comfort indicators in responding to psychological factors.

[0030] 4. A joint index system based on spatio-temporal regularity discrimination and load fluctuation analysis is adopted to construct a regularity partition recognition and update mechanism, enabling the system to automatically distinguish high-frequency usage areas and sudden usage areas, and dynamically maintain the control strategy library, improving the adaptability and scalability of spatial regulation.

[0031] 5. Through the NSGA-II multi-objective optimization algorithm, the weight is adaptively adjusted between COP and PMV, and multiple factors such as personnel density, temperature difference disturbance, and emotional fluctuation are introduced to construct an optimization objective function, enabling the control strategy to achieve the collaborative optimum of energy efficiency and comfort in different operating scenarios.

[0032] 6. A Meta-MPC meta-reinforcement learning control framework is proposed. In the regular partition, policy pre-training is carried out based on historical loads, and the abnormal load recognition ability is improved through LSTM prediction and STL decomposition. In the irregular partition, task migration and rapid fine-tuning are realized to ensure that the control strategy has generalization, stability, and real-time performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is the control logic diagram of the present invention.

[0034] Figure 2 It is the control flow chart of the present invention.

[0035] Figure 3 It is the architecture diagram for determining the regular partition of the present invention.

[0036] Figure 4 It is the pre-control demonstration diagram for the regular partition of the present invention.

[0037] Figure 5 It is the schematic diagram of the upper-layer multi-objective optimization module of the present invention.

[0038] Figure 6 It is the schematic diagram of the lower-layer reinforcement learning control module of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0039] The following specific embodiments illustrate the implementation manners of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification.

[0040] An intelligent regulation method based on perception and dynamic partitioning (as Figure 1 、 2 shown) includes the following steps:

[0041] I. Regular partitioning module: (as Figure 3 、 4 shown)

[0042] In this module, based on the behavior-density partitioning principle, the internal space of the building is divided into different regulation partitions to achieve precise control of different personnel behavior states and personnel density states. The behavior-density partitioning principle includes: behavior partitioning and density partitioning.

[0043] In a possible implementation manner, the behavior partitioning realizes real-time analysis of the movement frequency and amplitude of personnel limbs by integrating the YOLOv5 real-time pose detection algorithm and the OpenPose skeleton key point tracking algorithm, and classifies continuous action sequences by combining the LSTM (Long Short-Term Memory) neural network. The classification criterion is: when the detected movement speed of personnel does not exceed 1.2 m / s, it is classified as a low-activity area (such as sitting or walking); when the movement speed exceeds 1.2 m / s, it is classified as a high-activity area (such as running or strenuous exercise).

[0044] In a possible implementation manner, the movement speed is calculated based on the displacement of human key points and the time interval. It is set that between time t1 and t2, the coordinates of a certain key point are (x1, y1), (x2, y2), then:

[0045] Δx = x2 - x1

[0046] Δy = y2 - y1

[0047]

[0048] Where: Δx and Δy are the horizontal and vertical coordinate differences between the current position and the previous position of this person;

[0049] Δt = t2 - t1 is the time interval;

[0050] v i (t) is the speed of a certain person.

[0051] In addition, the LSTM neural network is used to process the time series data output by the pose recognition algorithm, and the input gate, forget gate, and output gate control mechanisms are used to dynamically determine the activity intensity level to which the personnel belong.

[0052] Furthermore, for each moment t, key point data of each person is obtained through a pose recognition algorithm, and a time series X(t) is generated based on this data:

[0053] X(t) = [v1(t), v2(t), …, v n (t)]

[0054] The core of the LSTM network is based on the memory unit h(t), which processes the input sequence through a gating mechanism:

[0055] i(t) = σ(W i · [h(t - 1), X(t)] + b i )

[0056] f(t) = σ(W f · [h(t - 1), X(t)] + b f )

[0057] o(t) = σ(W o · [h(t - 1), X(t)] + b o )

[0058]

[0059]

[0060] h(t) = o(t) · tanh(C(t))

[0061] Where: i(t) is the input gate;

[0062] f(t) is the forget gate;

[0063] o(t) is the output gate;

[0064] C(t) is the cell state;

[0065] h(t) is the output.

[0066] The LSTM network will classify the activities of people using the output h(t), and the classification basis is:

[0067] Low - activity area: If v i (t) ≤ 1.2 m / s;

[0068] High - activity area: If v i (t) > 1.2 m / s.

[0069] Furthermore, it is determined whether each person is in the low - activity area (such as sitting still, walking) or in the high - activity area (such as running, strenuous exercise).

[0070] In a possible implementation, the density partition detects the number of people based on the YOLOv5 algorithm and completes the division of crowded areas of people through the DBSCAN clustering algorithm. A density factor f is assigned to each clustering area 密度 , which is defined in the following way by this design:

[0071]

[0072] where: N 聚类 is the number of people in the clustering area;

[0073] A 聚类 is the area of this cluster, approximately the area of a circle, i.e., A 聚类 = πr 2 ; r is the radius of the cluster.

[0074] Furthermore, the density factor changes with the density of the cluster. Let ρ 标准 be the historical mean density, then by observing and calculating the population density of multiple sample areas over a period of time, its average value is obtained.

[0075]

[0076] where: T is the total number of historical samples (i.e., the number of time periods or the number of observation areas in the historical records);

[0077] N i is the number of people in the i-th sample area;

[0078] A i is the area of the i-th sample area;

[0079] is the population density of the i-th sample area.

[0080] The density factor of this design is expressed as:

[0081]

[0082] where: ρ 聚类 is the density of the current clustering area;

[0083] ρ 标准 is the historical mean density.

[0084] Furthermore, according to the value of the density factor, the space is divided into a low-density area and a high-density area. The characteristics of the low-density area are: the clustering radius r ≥ 2m and the population density ρ 聚类 ≤ 0.3 person / m 2 ; the characteristics of the high-density area are: the clustering radius r ≤ 1m and the population density ρ 聚类 ≥ 0.8 person / m2 。

[0085] In a possible implementation, by combining the sensible heat load parameters of people with different activity types in the national standard "Design Standard for Energy Efficiency of Public Buildings" GB50189-2015, different sensible heat load values are assigned to the people in the low-activity - low-density area, low-activity - high-density area, high-activity - low-density area, and high-activity - high-density area respectively. The specific calculation method is as follows:

[0086] 1. Low-activity - low-density area

[0087]

[0088] Where: n ll (t) is the number of people in the low-activity - low-density area at time t;

[0089] h low is the sensible heat load coefficient of the low-activity area;

[0090] f 密度,ll is the density factor of the low-density area.

[0091] 2. Low-activity - high-density area

[0092]

[0093] Where: n lh (t) is the number of people in the low-activity - high-density area at time t;

[0094] f 密度,lh is the density factor of the high-density area.

[0095] 3. High-activity - low-density area

[0096]

[0097] Where: n hl (t) is the number of people in the high-activity - high-density area at time t;

[0098] h high is the sensible heat load coefficient of the high-activity area;

[0099] f 密度,hl is the density factor of the low-density area.

[0100] 4. High-activity - high-density area

[0101]

[0102] Where: n hh (t) is the number of people in the high-activity - high-density area at time t;

[0103] f密度,hh is the density factor of the high-density area.

[0104] Furthermore, in order to adapt to the changes in personnel flow and heat load at different times, a weighted summation method is adopted, combined with the weight factors of each time period, to calculate the total sensible heat load:

[0105] Q total (t) = ω t ·(L ll (t) + L lh (t) + L hl (t) + L hh (t))

[0106] where: ω t is the weight of time period t, used to adjust the influence of sensible heat load in different time periods.

[0107] In a possible implementation manner, the determination of the regular partition adopts a spatio-temporal-load multi-dimensional joint determination strategy. Specifically, it includes:

[0108] 1. Spatio-temporal regularity test: By calculating the time variance of personnel activities, the time period concentration index is obtained, and by calculating the spatial intersection-over-union ratio, the spatial overlap rate is obtained:

[0109]

[0110] where: S time is the time period concentration index;

[0111] is the occurrence time variance;

[0112] R space is the spatial overlap rate;

[0113] IoU is the intersection-over-union ratio.

[0114] Furthermore, if S time ≥0.65 and R space ≥0.7, it is determined that this area has regularity.

[0115] 2. Load characteristic matching: The dynamic time warping (DTW) algorithm is used to calculate the normalized distance of the load curve:

[0116]

[0117] where: D DTW is the dynamic time warping distance;

[0118] π is the optimal alignment path;

[0119] C cv is the load fluctuation consistency.

[0120] Further, if D DTW ≤ 0.3 and C cv ≤ 0.25, it is determined that the load characteristic meets the regularity requirement.

[0121] 3. Verification of regulation effectiveness, through the policy responsiveness index:

[0122]

[0123] where η is the policy responsiveness.

[0124] Further, if η ∈ [0.8, 1.2], it indicates that the partition has good regulation response ability.

[0125] Further, the present invention defines a comprehensive regularity index through a weighted scoring model, and the scoring dimensions include time regularity, spatial stability, load consistency, and regulation effectiveness:

[0126] Table 1 Weighted Scoring Replication Table

[0127]

[0128]

[0129]

[0130] Further, when I regular ≥ 0.6, it is determined as a regular partition; otherwise, it is determined as an irregular partition.

[0131] In a possible implementation manner, dynamic update management is performed for the regular partition. Specifically, it includes:

[0132] 1. New rule: For a physical space currently detected as an irregular partition, if it shows a high load intensity and conforms to a periodic occurrence pattern within three consecutive cycles, Fourier transform is used to analyze the periodic load coincidence degree:

[0133]

[0134] where: L 历史 (t) is the load intensity of a physical space at time t in historical data;

[0135] L 实际 (t) is the actual load intensity of a physical space at the current time t.

[0136] Further, if P 周期 (i) > 0.7, it is newly added as a regular partition.

[0137] 2. Removal Rule: For the identified regular partitions, set the deviation rate:

[0138]

[0139] where: δ(i,t) is the deviation rate.

[0140] Furthermore, if within three consecutive periods, the deviation between the predicted value and the actual value of the partition load exceeds 20%, that is then it is considered that the regulation effect of this partition is inefficient, and it will be removed from the regular partition database.

[0141] II. Upper-layer Multi-objective Optimization Module: (as Figure 5 shown)

[0142] In a possible implementation manner, an upper-layer multi-objective optimization module is provided. This module is used in the multi-objective intelligent regulation system of air conditioners in public buildings to achieve dynamic balance adjustment of energy-saving effect and thermal comfort according to environmental parameters and dynamic thermal loads.

[0143] Specifically, for the indoor and outdoor air flow caused by the opening of doors and windows in the public building environment, and thus the resulting changes in temperature and humidity, the upper-layer multi-objective optimization module obtains the internal environmental parameters of the building in real time through the arranged temperature and humidity sensors, and combines with the theoretical thermal load in the area with concentrated people to dynamically adjust the operating efficiency of the air conditioning system to ensure human thermal comfort while reducing energy consumption.

[0144] Furthermore, to achieve the optimization effect, the upper-layer multi-objective optimization module takes the energy efficiency ratio (COP) of the air conditioning system and the thermal comfort index (PMV) as dual optimization objectives, where:

[0145]

[0146] PMV(t) = f(T 室内 ,T 辐射 , humidity, air flow, P 活动 (t))

[0147] where: COP (Coefficient of Performance) is the energy efficiency index of the air conditioning system;

[0148] PMV (Predicted Mean Vote) is the thermal comfort evaluation index;

[0149] P 活动 (t) is the activity intensity, affected by emotions;

[0150]

[0151] P baseIt is the reference value of GB50189-2015.

[0152] Furthermore, weight coefficients α(t) and β(t) are set, corresponding to the priorities of energy efficiency and comfort objectives respectively, and the two are in the following functional forms:

[0153]

[0154] where: ρ(t) is the real-time personnel density (persons / m 2 ), calculated by the dynamic clustering module;

[0155] |PMV(t)| is the absolute value of the current comfort deviation, and PMVmax = 2 is the acceptable maximum deviation;

[0156] ΔT in-out (t) is the indoor-outdoor temperature difference, reflecting the intensity of heat load disturbance;

[0157] k is the slope coefficient of the Sigmoid function, controlling the sensitivity of the weight to density changes, k = 0.8;

[0158] ρ base is the reference density, ρ base = 0.5 persons / m 2 ;

[0159] T ref is the reference value for temperature difference normalization, T ref = 10°C;

[0160] σ emotion (t) is the emotion correction factor.

[0161] During the specific operation process, the weight coefficients are dynamically adjusted according to the following logic:

[0162] When the real-time personnel density ρ(t) is higher than the reference density ρbase, that is, ρ(t)>ρbase, α(t) decays exponentially with the increase of density, and β(t) increases synchronously to give priority to ensuring comfort;

[0163] When ρ(t) is lower than ρbase, that is, ρ(t)<ρbase, α(t) rises significantly and β(t) decreases to strengthen energy efficiency optimization;

[0164] If the |PMV(t)| deviation approaches the extreme value PMVmax, indicating that the comfort deteriorates, then α(t) is forced to decrease and β(t) is increased to preferentially improve thermal comfort;

[0165] If the indoor-outdoor temperature difference ΔT in-out (t) increases significantly, β(t) is additionally weighted to compensate for the heat load fluctuations caused by the opening of doors and windows.

[0166] To accurately identify the opening state of doors and windows, the upper-layer multi-objective optimization module deploys a Doppler ultrasonic array (operating frequency 40 kHz) at the air outlet of the air conditioner, and constructs a hybrid perception network in combination with a door magnetic sensor to realize the joint detection of the opening and closing state and the opening area of doors and windows:

[0167]

[0168]

[0169] where: Δt is the round-trip time of ultrasonic waves;

[0170] θ is the installation inclination angle of the sensor;

[0171] d th is the closing determination threshold, d th = 5 mm;

[0172] d max is the maximum detection distance, d max = 500 mm;

[0173] r is the radius of the air outlet, r = 0.15 m.

[0174] The opening area of the doors and windows is obtained through the above calculations, and the load disturbance compensation factor is dynamically corrected.

[0175] At the same time, the upper-layer multi-objective optimization module integrates an emotion-physiological coupling correction strategy. Through the deployed face image acquisition device, combined with infrared thermal imaging and face expression recognition technologies, the emotional state and physiological parameters of the crowd are extracted to form a dynamic sensible heat load correction model. Specifically, the thermal imaging technology is used to detect the dynamic skin temperature change rate of the face, and through the cross-modal attention network in the deep neural network, the visual features and thermal imaging data are fused to output the emotion correction factor σ emotion (t), and its steps and calculation formula are:

[0176] Thermal imaging analysis: Calculate the dynamic skin temperature within the detected face ROI (128×128 pixels):

[0177]

[0178] Establish an individualized temperature baseline:

[0179]

[0180] Calculation of the temperature change rate:

[0181]

[0182] where: is the Laplacian operator of the motion energy map.

[0183] Construct a dual-branch deep neural network to achieve feature fusion:

[0184] F emotion = MSCM-CNN(I RGB )(multi-scale features)

[0185] F thermal = TCN([ΔHR(t), T skin (t)])(temporal convolution)

[0186] α = Softmax(W α [F emotion ; F thermal )

[0187] F fusion = α ⊙ F emotion + (1 - α) ⊙ F thermal

[0188] The network output includes the emotion state with physiological calibration:

[0189] High-arousal emotions (happiness / anger): ΔHR > 0.3℃ / min and p high > 0.7

[0190] Low-arousal emotions (sadness / relaxation): ΔHR ≤ 0.1℃ / min and p low > 0.5

[0191] Furthermore, the individual emotion state is output through the emotion recognition module, and combined with the skin temperature change rate ΔHR(t), the emotion correction factor h emotion (t) is generated:

[0192]

[0193] where: σ emotion (t) is the emotion correction factor;

[0194] h emotion,i (t) is the average emotion correction of the people in each area.

[0195] Furthermore, to simultaneously optimize the energy efficiency and comfort objectives, the upper-layer multi-objective optimization module applies the multi-objective evolutionary algorithm NSGA-II to construct the following objective function:

[0196]

[0197] γ = 0.1·ReLU(ΔT in-out - 5℃)

[0198] Where: δ is the emotional stability weight, used to suppress the impact of drastic emotional fluctuations on the system, and δ = 0.05.

[0199] Furthermore, during the optimization process, the following constraint conditions are satisfied:

[0200] Weight normalization constraint: α(t) + β(t) ≤ 1.2;

[0201] Regulation action boundary: The temperature adjustment amount ΔT set ∈ [T min , T max , and the wind speed V fan ∈ [V low , V high , which is limited by the building thermal engineering code.

[0202] Furthermore, through the iterative calculation of the NSGA-II algorithm, an optimal solution set (Pareto Front) that meets the balance requirements of different energy efficiency and comfort is generated. Select the optimal solution with the highest current demand matching degree from it, and the corresponding energy efficiency - comfort weight coefficients α(t) and β(t) are used as the final output of the multi-objective optimization module.

[0203] Furthermore, adopt a dynamic elite retention strategy to preferentially retain non-dominated solutions that meet the dual-objective optimization criteria in each iteration, ensuring the stability and robustness of the optimization effect:

[0204]

[0205] Furthermore, define a dynamic fitness score function, select the solution with the highest comprehensive score as the optimization result, and then output the energy efficiency - comfort weight coefficients α(t) and β(t):

[0206]

[0207] III. Lower-level reinforcement learning regulation module: (as Figure 6 shown)

[0208] In a possible implementation manner, a lower-level reinforcement learning regulation module is provided, which is used to dynamically optimize the operation strategy of the air-conditioning system based on the dynamic changes of the loads in the regular partition and the irregular partition, and further improve the responsiveness and adaptive ability of the air-conditioning regulation system for public buildings.

[0209] Specifically, the lower-level reinforcement learning regulation module first decomposes the historical load time-series data for the regular partition through the STL (Seasonal-Trend decomposition using Loess) algorithm:

[0210] Q(t) = S(t) + T(t) + R(t)

[0211] Wherein: S(t) is the seasonal component, reflecting the periodic fluctuation;

[0212] T(t) is the trend component, reflecting the long-term load change trend;

[0213] R(t) is the residual component, reflecting the noise in the data.

[0214] Furthermore, a load prediction model based on LSTM (Long Short-Term Memory Network) is adopted to predict the change of the future load Q 预测 (t + 1):

[0215] Q 预测 (t + 1) = f LSTM (Q 历史 , S 开窗 (t), ΔT in-out (t))

[0216] Wherein: Q 历史 (t) is the historical load data;

[0217] f LSTM is the output of the model;

[0218] S 开窗 (t) is the window opening area monitored at time t.

[0219] Furthermore, considering the influence of door and window disturbances, the window opening area S 开窗 (t) detected by the sensor is used as one of the model input variables to further improve the load prediction accuracy. To identify abnormal load changes, the module defines an abnormal load threshold abnormal load threshold(t):

[0220] abnormal load threshold(t) = μ 历史负荷 (t) + n · σ 历史负荷 (t)

[0221] Wherein: μ 历史负荷 (t) is the load mean calculated according to the time window (ten minutes);

[0222] σ 历史负荷 (t) is the load standard deviation in the same period of historical data;

[0223] n is the amplification factor, n = 3.

[0224] When the predicted load Q 预测 (t + 1) exceeds the abnormal load threshold abnormal load threshold(t), the lower-layer reinforcement learning control module triggers a pre-cooling or pre-heating adjustment mechanism, and the adjustment amount calculation formula is:

[0225] Δt负荷 (t + 1)= Q 预测 (t + 1)- t 综合 + 0.2·S 开窗 ·ΔT(t) in-out

[0226] If ΔQ 负荷 (t + 1)> 0, it means pre-cooling is required; if ΔQ 负荷 (t + 1)< 0, it means pre-heating is required.

[0227] Furthermore, the time calculation formula for the adjustment action is:

[0228]

[0229] Where: Δt 预调 is the time required for adjustment;

[0230] The adjustment rate is preset according to the air conditioner parameters.

[0231] To achieve adaptive dynamic optimization, the module adopts the Meta-MPC (Meta-Model Predictive Control) algorithm, trains the meta-policy network based on historical load and disturbance data, and realizes the dynamic regulation of regular partitions.

[0232] The meta-policy network adopts the Actor-Critic architecture, including:

[0233] Actor network: Input the current state s t , and output the action distribution π θ (s t );

[0234] Critic network: Input the current state s t and the action a t , and output the long-term value estimate Q φ (s t , a t ).

[0235] Specifically, define the state space S(t) including:

[0236] S(t)= [Q 预测 (t), T 当前 (t), V 当前 (t), S 开窗 (t), ΔT in-out (t), σ emotion (t)]

[0237] Where: Q 预测 (t) is the current load;

[0238] T 当前(t) is the indoor temperature;

[0239] V 当前 (t) is the wind speed;

[0240] S 开窗 is the window opening area.

[0241] Define the action space A(t) to include:

[0242] A(t) = [ΔT 调节 , ΔV 调节

[0243] Where: ΔT 调节 is the temperature adjustment amount;

[0244] ΔV 调节 is the wind speed adjustment amount;

[0245] On this basis, construct the reward function R(t):

[0246]

[0247] Where: T 目标 is the set temperature output by the multi-objective optimization module;

[0248] T 实际 is the measured temperature by the sensor;

[0249] PMV 实际 is the real-time comfort index calculated based on temperature and humidity. The smaller the absolute value, the higher the comfort;

[0250] P ac (t) is the instantaneous power (kW) of the air conditioning system, which is inversely proportional to the COP energy efficiency ratio;

[0251] γ is the comfort penalty coefficient, fixed at 0.5;

[0252] λ(t) is the dynamic energy efficiency weight;

[0253]

[0254] K cop is the energy efficiency conversion coefficient;

[0255] t + i is to predict the future i steps by LSTM;

[0256] γ is the long-term reward discount factor, γ = 0.9.

[0257] In the model training stage, the module performs the following steps:

[0258] S1. Establish a meta-policy network and Q function ​Randomly initialize the parameters θ meta , φ meta , pre-train the LSTM model, and initialize the multi-task task distribution, that is, each task corresponds to a dynamic model of a partition type.

[0259] S2. Sample N tasks from the task distribution for training. Each task contains historical data:

[0260] Adopt the inner loop. For each task, use the current meta-policy to generate the initial action sequence.

[0261] Based on the dynamic prediction model f predict roll and optimize the actions for the next k steps:

[0262]

[0263] Collect the adaptation data

[0264] Adopt the outer loop to calculate the multi-task loss:

[0265]

[0266] where the target value is constructed as:

[0267]

[0268] Update the meta-parameters:

[0269]

[0270] S3. In the online regulation stage, input the current state S(t):

[0271] s t = [Q 预测 , T 当前 , V 当前 , S 开窗 , ΔT in-out , σ emotion

[0272] Use the dynamic model to predict the state for the next k = 5 steps:

[0273] s t+1:t+k = f predict (s t )

[0274] And call the meta-policy network to generate the candidate action sequence:

[0275] ​​

[0276] Select the optimal sequence that meets the constraints:

[0277] ΔT 调节 ∈[T min , T max , ΔV 调节 ∈[V low , V high

[0278] Furthermore, execute the first step action Observe the actual feedback r t and the next state s t+1 .

[0279] If a change in the partition type is detected (such as the emergence of a new clustering region):

[0280] Collect real-time data:

[0281]

[0282] Execute the fast fine-tuning strategy:

[0283]

[0284] Among them, the fine-tuning learning rate is set to α = 0.1.

[0285] S4. Adopt a soft update mechanism to synchronize the target network parameters every 100 steps:

[0286] θ′←τθ+(1 - τ)θ′, φ′←τφ+(1 - τ)φ′

[0287] Among them, τ = 0.01 to ensure policy stability.

[0288] When it is detected that the PMV deviation continuously exceeds the standard (|PMV(t)| > 1.5) or the coefficient of performance COP(t) drops suddenly, immediately trigger a hard update to ensure policy stability:

[0289] θ′←θ, φ′←φ

[0290] S5. Calculate and execute the adjustment amount:

[0291] Temperature adjustment amount:

[0292]

[0293] Wind speed adjustment amount:

[0294]

[0295] For non-regular partitions, the lower-layer reinforcement learning control module regards it as a new task: ​

[0296] Task = {ρ 聚类 , ΔT in-out , σ emotion}

[0297] Furthermore, through the multi-task generalization ability of meta-learning, fast policy transfer is achieved. In the online fine-tuning stage, the module utilizes the real-time data of the new partition and completes the rapid update of policy parameters within 10 minutes. The update formula is:

[0298]

[0299] where: D new is the real-time data of the new partition;

[0300] η is the fast learning rate, η = 0.1.

[0301] Furthermore, the meta-policy outputs:

[0302] ΔT 调节 = f meta (Q 综合 , ΔT in-out , σ emotion )

[0303] ΔV 调节 = f meta (Q 综合 , COP, α(t) / β(t))

[0304] Constraint conditions:

[0305]

[0306] Through the above method, the lower-layer reinforcement learning control module can achieve precise control of different types of partitions, fully respond to dynamic load changes and environmental disturbances, and achieve adaptive energy-saving optimization and thermal comfort guarantee of the air-conditioning system.

Claims

1. An intelligent regulation method based on perception and dynamic partitioning, characterized in that, It includes the following steps: S1, Collect personnel activity data, use the YOLOv5 and OpenPose algorithms to obtain limb movements, and classify the activity intensity through a time series prediction algorithm to achieve the recognition of personnel behavior patterns in space; S2, Divide the space according to the personnel behavior pattern and space density, and use a multi-dimensional scoring model to determine regular and irregular partitions; S3, Combine infrared thermal imaging and facial expression recognition to extract the group emotional state, and construct an emotion correction factor to participate in the thermal load modeling; S4, Construct an upper-layer multi-objective optimization module, and use the NSGA-II algorithm to synergistically optimize the energy efficiency ratio and thermal comfort, and dynamically adjust the optimization target weights; S5, Based on the classification results of regular and irregular partitions and the output results of the upper-layer multi-objective optimization module, apply the lower-layer reinforcement learning control module, use the Meta-MPC meta-reinforcement learning algorithm to generate air-conditioning control actions in real time, and achieve the adaptive adjustment of the air-conditioning system.

2. The method according to claim 1, characterized in that, The determination of the regular partition is to calculate the spatio-temporal overlap rate, load fluctuation consistency, and control responsiveness of personnel activities, comprehensively construct a scoring model, and determine whether the space belongs to the regular partition according to the set threshold.

3. The method according to claim 1, wherein The construction of the emotion correction factor combines the skin temperature change rate detected by thermal imaging and facial expression features, and forms an emotion regulation factor for thermal load correction through cross-modal attention network fusion.

4. The method according to claim 1, characterized in that The upper-layer multi-objective optimization module takes the energy efficiency ratio and thermal comfort of the air-conditioning system as optimization targets, dynamically adjusts the optimization weights based on personnel density, indoor-outdoor temperature difference, and emotion correction factor, and uses the NSGA-II algorithm to iteratively generate the optimal control scheme.

5. The method according to claim 1, wherein The lower-layer reinforcement learning control module uses the Meta-MPC meta-reinforcement learning framework, constructs a meta-policy network using the Actor-Critic structure, and combines historical load and real-time perception data to dynamically generate control actions.

6. The method according to claim 5, wherein For the regular partition, use the STL algorithm to decompose the trend of historical load, combine the LSTM network for load prediction, and generate pre-cooling or pre-heating control actions according to the prediction results.

7. The method according to claim 5, wherein For the irregular partition, quickly adapt to the new load change and generate a control scheme by real-time sampling and fine-tuning the meta-policy network.

8. The method according to claim 1, wherein The hybrid perception network consists of a Doppler ultrasonic array and a door magnetic sensor, which is used to detect the door and window state and opening area in real time to achieve load disturbance compensation.

Images