Building control collaborative optimization decision-making method fused with lightweight AI inference engine

Abstract

The invention provides a building control collaborative optimization decision-making method fused with a lightweight AI inference engine. The method comprises the following steps: deploying an ultra-lightweight model on an edge side; collecting a mixed data set, and performing structured processing to obtain a standard time sequence matrix; feature extraction is carried out in a data dimension division mode, and a fused high-dimensional feature set is obtained; carrying out dimension reduction and activation, dimension raising reconstruction and feature fusion on the fused high-dimensional feature set to obtain an inference feature vector; and inputting the reasoning feature vector into the ultra-light model for intelligent reasoning to obtain a control suggestion, and completing a building collaborative optimization decision. According to the invention, the defects of insufficient intelligence, adjustment lag, poor adjustment precision and high cloud intelligent response delay of a traditional edge controller can be solved, and efficient, real-time and reliable building intelligent optimization control is realized.

Description

Technical Field

[0001] This invention relates to the field of building automation control technology, and in particular to a building control collaborative optimization decision-making method that integrates a lightweight AI inference engine. Background Technology

[0002] Traditional building control edge controllers, such as DDC controllers at the end of fan coil units and combined air conditioning units, have limited decision-making capabilities and commonly use fixed logic or simple proportional-integral-derivative (PID) logic. For example, they preset the supply air temperature or adjust the inverter frequency according to a specified supply air temperature. When facing complex and dynamic environments (fluctuations in pedestrian traffic, rapid changes in external climate, load coupling), they exhibit poor adaptability and low adjustment accuracy. Their decision-making also largely relies on fixed parameters manually set in the cloud, resulting in high latency, high bandwidth consumption, and paralysis upon network outages.

[0003] Furthermore, most existing edge computing solutions rely on a few key environmental parameters to provide simple linear control logic, resulting in coarse-grained regulation schemes that are insufficient in terms of the accuracy, granularity, real-time performance, and resource utilization efficiency of the regulation strategy. Many existing control rule engines and models are designed for general-purpose devices and do not consider the special resource limitations of building control edge controllers, such as microsecond-level real-time performance, MB-level memory usage, and mW-level power consumption. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a building control collaborative optimization decision-making method that integrates a lightweight AI inference engine. This method can solve the problems of insufficient intelligence, lagging adjustment, poor adjustment accuracy, and high cloud intelligence response latency of traditional edge controllers, and achieve efficient, real-time, and reliable intelligent building optimization control.

[0005] To achieve the above objectives, the present invention provides the following solution: a building control collaborative optimization decision-making method integrating a lightweight AI inference engine, comprising: Construct an end-to-end data path for the building and select a model framework. Based on the selected model framework, perform pruning, quantization, inference engine and containerization, and effect quantization operations to obtain an ultra-lightweight model that runs on the edge. Collect real-time building data, mix historical data with real-time data to obtain a mixed dataset, and perform structured processing on the mixed dataset to obtain a standard time series matrix; Based on the standard time series matrix, feature extraction is performed by dividing the data dimensions to obtain a fused high-dimensional feature set. Based on the fused high-dimensional feature set, dimensionality reduction and activation are performed, followed by dimensionality reconstruction and feature fusion to obtain the inference feature vector; The inference feature vector is input into the ultra-lightweight model for intelligent inference to obtain control suggestions. Based on the control suggestions, closed-loop execution of building equipment, performance and resource monitoring, and model iterative optimization are performed to complete building collaborative optimization decision-making.

[0006] Optionally, an end-to-end data path for the building is constructed, and a model framework is selected. Based on the selected model framework, pruning, quantization, inference engine and containerization, and effect quantization operations are performed to obtain an ultra-lightweight model running on the edge side, including: A bidirectional MQTT link between the host computer platform and the edge box is constructed using Ethernet. The edge box is accessed using the Ethernet building automation network protocol, and the DDC is accessed using the Ethernet serial bus transmission control protocol. The edge box is then analyzed and cracked to complete the construction of the end-to-end data path. The Transformer is selected as the model architecture, and the model architecture is distilled into a multilayer perceptron; the multilayer perceptron includes an input layer, a hidden layer, an output layer, and a linear rectified function; Based on the multilayer sensor, the core channels associated with temperature, humidity, carbon dioxide and human flow are retained, redundant branches such as open / closed windows or noise are pruned, and the hardware instruction set of the edge box is adapted to complete the pruning operation. Set the temperature sensor input layer to 32-bit floating-point precision, the intermediate calculation layer to 8-bit integer, and the output layer to 4-bit integer to complete the quantization operation; Select TensorFlow Lightweight or Open Neural Network Exchange Format runtime as the inference engine, trim the general functions of the inference engine, and deploy edge boxes in a containerized manner to unify lifecycle management and complete the inference engine and containerization operations. By reducing the volume of the multilayer perceptron to 98% and increasing the inference speed by 9 times, an ultra-lightweight model that runs on the edge side is obtained.

[0007] Optionally, real-time building data is collected, and historical data is mixed with real-time data to obtain a hybrid dataset. The hybrid dataset is then subjected to structured processing to obtain a standard time-series matrix, including: Based on a preset acquisition frequency, sensors are used to collect temperature and humidity, carbon dioxide, total volatile organic compounds, human presence sensing millimeter waves, equipment operating status, and outdoor information to obtain multi-source raw data streams. The multi-source raw data streams are then mixed with historical data to obtain a mixed dataset. Based on the hybrid dataset, outlier filtering is performed using the interquartile range method, and timestamp alignment is performed by resampling to a unified window to obtain a standard time series matrix.

[0008] Optionally, based on the standard time series matrix, feature extraction is performed by partitioning the data dimensions to obtain a fused high-dimensional feature set, including: Based on the standard time-series data, indoor temperature, indoor humidity, number of people, and indoor carbon dioxide are directly extracted to obtain the original dimensional data. Based on the original dimensional data, the temperature and humidity coupling coefficient, the influence factor of the correlation between people flow and CO2, the influence factor of the correlation between people flow and temperature, and the influence factor of the correlation between people flow and humidity are calculated to obtain the combined factor dimensional data. Based on the original dimensional data, the air quality index, perceived temperature, ventilation efficiency factor, and temperature change trend are calculated to obtain the derived index dimensional data. The original dimensional data, the combined factor dimensional data, and the derived index dimensional data are integrated to obtain a fused high-dimensional feature set.

[0009] Optionally, based on the fused high-dimensional feature set, dimensionality reduction and activation are performed, followed by dimensionality increase reconstruction and feature fusion to obtain the inference feature vector, including: Based on the fused high-dimensional feature set, linear dimensionality reduction is performed to obtain a low-dimensional vector, and the ReLU function is used to perform nonlinear activation on the low-dimensional vector to obtain the environmental characterization. Based on the environmental characterization, an upgraded matrix is ​​constructed, and the fused high-dimensional feature set and its comprehensive factor of the upgraded matrix are calculated to obtain upgraded features. Then, the upgraded features are fused with the derived index dimension data to obtain an inference feature vector.

[0010] Optionally, the inference feature vector is input into the ultra-lightweight model for intelligent inference to obtain control suggestions. Based on the control suggestions, closed-loop execution of building equipment, performance and resource monitoring, and model iterative optimization are performed to complete building collaborative optimization decisions, including: The inference feature vector is input into the ultra-lightweight model for intelligent inference to obtain control suggestions. The control suggestions are then sent to the building DDC controller, and sensor and DDC controller status data are uploaded in real time to achieve closed-loop control of building terminal devices. By real-time monitoring of sensor response latency, human thermal comfort index, and model resource occupancy, a building operation profile is obtained. Based on the building operation profile, input data quantization, engine interface adaptation, sample data fine-tuning, re-distillation, re-pruning, re-quantization, and adaptive updates of rule thresholds are performed to complete model iterative optimization and obtain building collaborative optimization decisions.

[0011] This invention discloses the following technical effects by providing a building control collaborative optimization decision-making method that integrates a lightweight AI inference engine: 1. Real-time performance: End-to-end response time in milliseconds, much faster than cloud / traditional logic.

[0012] 2. Comfort: The compliance rate of the human thermal comfort index (PMV) has significantly increased to 85% to 90%.

[0013] 3. Energy efficiency and cost: While ensuring comfort, overall energy consumption is reduced through the coordination of intelligent air supply temperature / air volume / air exchange efficiency; edge deployment reduces cloud resource and bandwidth costs.

[0014] 4. Stability and security: No network outages will cause system crashes; data will not leave the domain; containerization ensures upgrades and rollbacks; there is a backup strategy.

[0015] 5. Explainability: Explainable characteristics such as human flow, carbon dioxide, perceived temperature, and ventilation efficiency drive decision-making, facilitating compliance and operations communication.

[0016] 6. Practical: Optimized for resource-constrained hardware (pruning, quantization, engine trimming), and actually runnable.

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the method flow provided in an embodiment of the present invention; Figure 2 A flowchart illustrating the method architecture provided in this embodiment of the invention; Figure 3 A data reasoning flowchart provided for embodiments of the present invention; Figure 4 This is a schematic diagram of environmental diagnosis and control decision-making provided in an embodiment of the present invention; Figure 5 A flowchart of model lightweighting processing provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the application scenario architecture provided in the embodiments of the present invention; Figure 7 This is a schematic diagram of the hardware architecture and layout design provided for an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] like Figure 1 , Figure 2 As shown, this invention provides a building control collaborative optimization decision-making method that integrates a lightweight AI inference engine, including: 1. For example Figures 5-7 As shown, an end-to-end data path for the building is constructed, and a model framework is selected. Based on the selected model framework, pruning, quantization, inference engine and containerization, and effect quantization operations are performed to obtain an ultra-lightweight model running on the edge. Specifically, this includes: 1.1 A bidirectional MQTT link between the host computer platform and the edge box is constructed using Ethernet. The edge box is accessed using the Ethernet building automation network protocol, and the DDC is accessed using the Ethernet serial bus transmission control protocol. The edge box is then analyzed and cracked to complete the construction of the end-to-end data path.

[0023] 1) Upper-level IoT building control platform: used by operation and maintenance personnel to adjust equipment, control comfort, and monitor key data of building environment operation and maintenance, such as air supply temperature and indoor temperature.

[0024] 2) Edge Box: Used to deploy the trained building comfort AI model; deploy point data protocol parsing to regulate downstream air conditioning, fresh air and other building control units, such as cracking to obtain data points such as start-up, shutdown, fault, and temperature of downstream air conditioning units; install software function packages such as start-up and shutdown control and regulation control.

[0025] The edge box is connected to the host IoT building control platform server via wired Ethernet, and the two communicate via the MQTT protocol. The edge box accesses the downstream DDC controller via Ethernet-BACnet IP protocol and Ethernet-Modbus TCP protocol. By parsing and deciphering these protocols, it obtains the unit's PID control commands from within the DDC controller, such as start / stop and adjustment of air supply temperature. It can also acquire temperature and humidity monitoring data and equipment status data from within the DDC controller.

[0026] The internal structure of the edge box includes the following components: main control chip (CPU), AI acceleration unit (NPU), 2GB DDR memory, 16GB eMMC flash memory, 32GB SD card, low-power power management module, security chip + security algorithm, and communication interface.

[0027] Main control chip CPU and its function: Quad-core Cortex-A9 @1.8GHz. Its functions include: running basic control logic (PID control), handling communication protocols (Modbus / BACnet), and managing NPU task scheduling.

[0028] AI Acceleration Unit - NPU: 5-16 TOPS computing power / latency <10ms. Its functions include: performing AI model inference (INT8 quantization), real-time feature extraction (such as perceived temperature, temperature-humidity coupling coefficient, etc.), and safety-critical decision-making (adjusting supply air temperature, adjusting unit fan frequency converters, etc.). Memory: 2GB / 64GB / s bandwidth. Its functions include: caching real-time sensor data streams, such as indoor temperature, indoor humidity, and indoor carbon dioxide levels; storing intermediate results from model inference, such as the temperature-person flow rate change rate; and creating a temporary database, such as a time-series database.

[0029] eMMC: 16GB. Its function is to store and store embedded AI model files and record device operation logs.

[0030] Low-power power management: dynamically adjusts NPU voltage and frequency to keep edge computing in a low-power state, CPU sleep mode switching, and emergency power failure data protection. Ethernet, 485: The edge box connects to the downstream DDC controller via Ethernet-BACnet IP protocol and Ethernet-Modbus TCP protocol, and the above protocols are parsed and cracked.

[0031] 3) DDC Building Automation Controller: Used to connect to terminal units, such as air conditioning units, fresh air units, and other terminal equipment, such as air conditioning valves, water valves, temperature and humidity sensors, etc. Connection methods: Digital and analog line connections such as DI, DO, AI, and AO.

[0032] 4) Building control unit terminal facilities: air conditioning unit temperature and humidity sensors, air valve actuators, water valve actuators, frequency converters, etc. Connection method: Digital and analog signals such as DI, DO, AI, and AO are connected to the upstream DDC.

[0033] 1.2 Select Transformer as the model architecture, and distill the model architecture into a multilayer perceptron; the multilayer perceptron includes an input layer, a hidden layer, an output layer, and a linear rectified function.

[0034] 1.3 Based on the aforementioned multilayer sensor, the core channels associated with temperature, humidity, carbon dioxide, and human flow are retained, redundant branches such as switch windows or noise are pruned, and the hardware instruction set of the edge box is adapted to complete the pruning operation.

[0035] 1.4 Set the temperature sensor input layer to 32-bit floating-point precision, the intermediate calculation layer to 8-bit integer, and the output layer to 4-bit integer to complete the quantization operation.

[0036] 1.5 Select TensorFlow Lightweight or Open Neural Network Exchange Format runtime as the inference engine, trim the general functions of the inference engine, and deploy edge boxes in a containerized manner to unify lifecycle management and complete the inference engine and containerization operations.

[0037] 1.6 The volume of the multilayer perceptron was reduced to 98% (2GB → 40MB), and the inference speed was increased by 9 times (greater than 1s → 100ms), resulting in an ultra-lightweight model that runs on the edge side.

[0038] 2. Collect real-time building data, mix historical data with real-time data to obtain a hybrid dataset, and perform structured processing on the hybrid dataset to obtain a standard time-series matrix. Specifically, this includes: 2.1 Based on the preset acquisition frequency, sensors are used to collect temperature and humidity, carbon dioxide, total volatile organic compounds, human presence sensing millimeter waves, equipment operating status, and outdoor information to obtain multi-source raw data streams. The multi-source raw data streams are then mixed with historical data to obtain a mixed dataset.

[0039] Specifically: Sampling frequency: 1 time / 15 seconds.

[0040] The system uses temperature and humidity sensors to acquire real-time indoor temperature, humidity, CO2, and TVOC data, which are then transmitted to the edge box via the MQTT protocol.

[0041] The system uses millimeter-wave sensors to acquire data on the presence of the human body, and then transmits this data to the edge box via the MQTT protocol.

[0042] The DDC controller collects the opening data of the end air valve actuator and water valve actuator, and then transmits it to the edge box via ModbusTcp and Bacnet IP protocols.

[0043] Outdoor sensors or third-party platforms are used to acquire time, weather forecasts, and outdoor weather data, which are then transmitted to the edge box via the MQTT protocol.

[0044] Historical data input includes: Historical data (temperature and humidity over the past N minutes), equipment status (return air temperature, valve opening); data reported by local sensors (temperature, humidity, CO2, personnel count sensor data).

[0045] External data (temperature and humidity over the past N minutes), third-party data, and outdoor sensor data.

[0046] 2.2 Based on the aforementioned hybrid dataset, outlier filtering is performed using the interquartile range method, and timestamp alignment is achieved by resampling to a unified window to obtain a standard time series matrix.

[0047] Outlier filtering: Outliers are filtered using the interquartile range (IQR) method: Q1 = 25th percentile, Q3 = 75th percentile, IQR = Q3 - Q1, lower limit = Q1 - 1.5 * IQR, upper limit = Q3 + 1.5 * IQR. Data exceeding the above upper and lower limits are filtered out as outliers.

[0048] Multi-sensor timestamp alignment: Sensor data is aligned to a unified time standard structure, standardizing indoor temperature, humidity, carbon dioxide, and personnel counts to a uniform time interval, such as reporting every 15 seconds. Temperature, humidity, and personnel counts are collected over a cumulative 15-second interval; carbon dioxide data is aligned using a 15-second interpolation method.

[0049] 3. For example Figure 3 As shown, based on the standard time series matrix, feature extraction is performed through data dimension partitioning to obtain a fused high-dimensional feature set. Specifically, this includes: 3.1 Based on the aforementioned standard time-series data, indoor temperature, indoor humidity, number of people, and indoor carbon dioxide are directly extracted to obtain the original dimensional data.

[0050] 3.2 Based on the original dimensional data, calculate the temperature and humidity coupling coefficient, the correlation factor between pedestrian flow and CO2, the correlation factor between pedestrian flow and temperature, and the correlation factor between pedestrian flow and humidity to obtain combined factor dimensional data.

[0051] The temperature and humidity coupling coefficient is the coupling relationship between temperature and humidity data. It represents a value calculated based on the coupling of temperature and humidity, which indicates the influence of humidity on ambient temperature.

[0052] The core formula is as follows: HI=c1+c2T+c3RH+c4T·RH+c5T2+c6RH2+c7T2·RH+c8T⋅RH2+c9T2·RH2 Substituting C1-C9, T=27, RH=55%, the temperature and humidity coupling coefficient H1=0.8 is obtained.

[0053] Reference coefficient values: C1= -42.379, C2 = 2.049, C3= 10.14; C4 = -0.225, C5= -0.007, C6= -0.0548; C7= 0.00123, C8= -0.00085, C9= -0.000002.

[0054] Factors influencing the relationship between pedestrian traffic and CO2 levels: Model formula: LCO2=(Coutdoor-Cindoor) / n; LCO2 = (624-600) / 20 = 1.2.

[0055] This formula represents the impact of human traffic on carbon dioxide concentration, and the degree of the final impact is expressed by the LCO2 value.

[0056] Factors influencing the relationship between pedestrian flow and temperature: The core equation is: ; This formula represents the impact of foot traffic on ambient temperature; the final degree of impact is expressed by the following formula. It is represented by numerical values.

[0057] Factors influencing the relationship between pedestrian flow and humidity: The core equation is: ; This formula represents the impact of foot traffic on environmental humidity; the final degree of impact is expressed by the following formula. It is represented by numerical values.

[0058] 3.3 Based on the original dimensional data, calculate the air quality index, perceived temperature, ventilation efficiency factor, and temperature change trend to obtain derived index dimensional data.

[0059] Air Quality Index: IAQI = ReLU(CO2 - 500) × β + ReLU(People - 15) × γ; The above formula reflects the impact of carbon dioxide and population deviations from normal values.

[0060] ReLU(x) = max(0, x) Substitute the scene values ​​into the calculation: IAQI = max(0, 600-500)*1.2 + max(0,20-15)*0.05 = 100*1.2+5*0.05=120.25.

[0061] Feeling temperature: AT = T + α·RH·ReLU(T-25); This formula represents the actual hot and cold effects that environmental humidity and temperature ultimately bring to the human body.

[0062] Ventilation efficiency factor: η = ReLU(-ΔCO2 / Δt) (assuming ΔCO2 / Δt = -20ppm / min). This formula represents the rate and trend of change of the current carbon dioxide concentration over time, and also represents the current ventilation effectiveness index.

[0063] Temperature change trend: Dt / dt =γ·ReLU(People-10)-0.5·ReLU(T-26); This formula represents the impact of the actual number of people and the actual temperature deviating from the normal value.

[0064] 3.4 The original dimensional data, the combined factor dimensional data, and the derived index dimensional data are integrated to obtain a fused high-dimensional feature set.

[0065] 4. For example Figure 4 As shown, based on the fused high-dimensional feature set, dimensionality reduction and activation are performed, followed by dimensionality upscaling reconstruction and feature fusion to obtain the inference feature vector. Specifically, this includes: 4.1 Based on the fused high-dimensional feature set, linear dimensionality reduction is performed to obtain a low-dimensional vector, and the ReLU function is used to perform nonlinear activation on the low-dimensional vector to obtain the environment characterization.

[0066] 4.2 Based on the environmental characterization, construct an upgraded matrix, calculate the fused high-dimensional feature set and its comprehensive factor of the upgraded matrix to obtain upgraded features, and then fuse the upgraded features with the derived index dimension data to obtain an inference feature vector.

[0067] 5. Input the inference feature vector into the ultra-lightweight model for intelligent inference to obtain control suggestions. Based on the control suggestions, perform closed-loop execution of building equipment, performance and resource monitoring, and model iterative optimization to complete building collaborative optimization decision-making. Specifically, this includes: 5.1 Input the inference feature vector into the ultra-lightweight model for intelligent inference to obtain control suggestions, send the control suggestions to the building DDC controller, and upload the sensor and DDC controller status data in real time to control the building terminal equipment in a closed loop.

[0068] For example: The correlation between pedestrian flow and carbon dioxide is 531.5, which is greater than the threshold of 500: it is determined that the carbon dioxide level is too high relative to the pedestrian flow → activate the strong ventilation mode. Feeling temperature 35.6℃, greater than 30℃: Temperature is too high → Set air conditioner to 24℃; Ventilation efficiency factor 1.0, less than 1.2: insufficient ventilation efficiency → issue an alarm and maintenance suggestion to "optimize the air duct layout".

[0069] Linkage command (sample landing): Fresh air system: operating at 100% power; Air conditioning: Set to 24 degrees Celsius; Warning: Ventilation efficiency needs to be optimized.

[0070] Results: Actionable control recommendations and alerts.

[0071] 5.2 Real-time monitoring of sensor response delay, human thermal comfort index, and model resource occupancy rate yields a building operation profile. Based on this profile, input data quantization, engine interface adaptation, sample data fine-tuning, re-distillation, re-pruning, re-quantization, and adaptive updates of rule thresholds are performed to complete model iterative optimization and obtain building collaborative optimization decisions.

[0072] Response latency: less than 100 milliseconds from sensor change to command issuance (about one-tenth of traditional logic).

[0073] Comfort compliance rate: The proportion of time during which the human thermal comfort index (PMV, target 0) is maintained within the ±0.5 range reaches 85% to 90% (an improvement of about 40 percentage points compared to the traditional control of 45% to 50%).

[0074] Resource consumption: The utilization rate of the neural network processing unit is about 30%; memory consumption is about 10%.

[0075] Overall data flow processing flow: Input data format conversion and entry into the inference engine interface, converting FP32 to INT8; Sensor raw data FP32 - dynamic range calibration - quantization parameter calculation - accuracy conversion - quantization format encapsulation inference engine interface.

[0076] Dynamic range calibration: Determines the optimal quantization range for FP32 data.

[0077] The maximum and minimum values ​​in the original FP32 data format are filtered out, outliers are removed, a data histogram is created, and a quantified data range set is established based on the updated data histogram and data range.

[0078] Quantization parameter calculation: Q = round ( FP 32- zeropoint ) / scale Scale = ( effectivemax - effectivemin ) / 2bits-1 zeropoint = round (- effectivemin × scale ) The quantization parameters scale and zeropoint are calculated using the above formulas. Conversion from FP32 to IN8: Load the fP32 data format, apply the quantization formula Q = round (FP32-zeropoint) / scale, calculate the INT8 value, pack and store it.

[0079] Therefore, by providing a building control collaborative optimization decision-making method that integrates a lightweight AI inference engine, this invention can solve the shortcomings of traditional edge controllers, such as insufficient intelligence, lagging adjustment, poor adjustment accuracy, and high cloud intelligence response latency, and achieve efficient, real-time, and reliable building intelligent optimization control.

[0080] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0081] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A building control collaborative optimization decision-making method integrating a lightweight AI inference engine, characterized in that, include: Construct an end-to-end data path for the building and select a model framework. Based on the selected model framework, perform pruning, quantization, inference engine and containerization, and effect quantization operations to obtain an ultra-lightweight model that runs on the edge. Collect real-time building data, mix historical data with real-time data to obtain a mixed dataset, and perform structured processing on the mixed dataset to obtain a standard time series matrix; Based on the standard time series matrix, feature extraction is performed by dividing the data dimensions to obtain a fused high-dimensional feature set. Based on the fused high-dimensional feature set, dimensionality reduction and activation are performed, followed by dimensionality reconstruction and feature fusion to obtain the inference feature vector; The inference feature vector is input into the ultra-lightweight model for intelligent inference to obtain control suggestions. Based on the control suggestions, closed-loop execution of building equipment, performance and resource monitoring, and model iterative optimization are performed to complete building collaborative optimization decision-making.

2. The building control collaborative optimization decision-making method integrating a lightweight AI inference engine according to claim 1, characterized in that, An end-to-end data path for the building is constructed, and a model framework is selected. Based on the selected model framework, pruning, quantization, inference engine and containerization, and effect quantization operations are performed to obtain an ultra-lightweight model running on the edge, including: A bidirectional MQTT link between the host computer platform and the edge box is constructed using Ethernet. The edge box is accessed using the Ethernet building automation network protocol, and the DDC is accessed using the Ethernet serial bus transmission control protocol. The edge box is then analyzed and cracked to complete the construction of the end-to-end data path. The Transformer is selected as the model architecture, and the model architecture is distilled into a multilayer perceptron; the multilayer perceptron includes an input layer, a hidden layer, an output layer, and a linear rectified function; Based on the multilayer sensor, the core channels associated with temperature, humidity, carbon dioxide and human flow are retained, redundant branches such as open / closed windows or noise are pruned, and the hardware instruction set of the edge box is adapted to complete the pruning operation. Set the temperature sensor input layer to 32-bit floating-point precision, the intermediate calculation layer to 8-bit integer, and the output layer to 4-bit integer to complete the quantization operation; Select TensorFlow Lightweight or Open Neural Network Exchange Format runtime as the inference engine, trim the general functions of the inference engine, and deploy edge boxes in a containerized manner to unify lifecycle management and complete the inference engine and containerization operations. By reducing the volume of the multilayer perceptron to 98% and increasing the inference speed by 9 times, an ultra-lightweight model that runs on the edge side is obtained.

3. The building control collaborative optimization decision-making method integrating a lightweight AI inference engine according to claim 2, characterized in that, Real-time building data is collected, and historical data is mixed with real-time data to obtain a hybrid dataset. This hybrid dataset is then subjected to structured processing to obtain a standard time-series matrix, including: Based on a preset acquisition frequency, sensors are used to collect temperature and humidity, carbon dioxide, total volatile organic compounds, human presence sensing millimeter waves, equipment operating status, and outdoor information to obtain multi-source raw data streams. The multi-source raw data streams are then mixed with historical data to obtain a mixed dataset. Based on the hybrid dataset, outlier filtering is performed using the interquartile range method, and timestamp alignment is performed by resampling to a unified window to obtain a standard time series matrix.

4. The building control collaborative optimization decision-making method integrating a lightweight AI inference engine according to claim 3, characterized in that, Based on the standard time series matrix, feature extraction is performed by partitioning the data dimensions to obtain a fused high-dimensional feature set, including: Based on the standard time-series data, indoor temperature, indoor humidity, number of people, and indoor carbon dioxide are directly extracted to obtain the original dimensional data. Based on the original dimensional data, the temperature and humidity coupling coefficient, the influence factor of the correlation between people flow and CO2, the influence factor of the correlation between people flow and temperature, and the influence factor of the correlation between people flow and humidity are calculated to obtain the combined factor dimensional data. Based on the original dimensional data, the air quality index, perceived temperature, ventilation efficiency factor, and temperature change trend are calculated to obtain the derived index dimensional data. The original dimensional data, the combined factor dimensional data, and the derived index dimensional data are integrated to obtain a fused high-dimensional feature set.

5. The building control collaborative optimization decision-making method integrating a lightweight AI inference engine according to claim 4, characterized in that, Based on the fused high-dimensional feature set, dimensionality reduction and activation are performed, followed by dimensionality increase reconstruction and feature fusion to obtain the inference feature vector, including: Based on the fused high-dimensional feature set, linear dimensionality reduction is performed to obtain a low-dimensional vector, and the ReLU function is used to perform nonlinear activation on the low-dimensional vector to obtain the environmental characterization. Based on the environmental characterization, an upgraded matrix is ​​constructed, and the fused high-dimensional feature set and its comprehensive factor of the upgraded matrix are calculated to obtain upgraded features. Then, the upgraded features are fused with the derived index dimension data to obtain an inference feature vector.

6. The building control collaborative optimization decision-making method integrating a lightweight AI inference engine according to claim 5, characterized in that, The inference feature vector is input into the ultra-lightweight model for intelligent inference to obtain control suggestions. Based on the control suggestions, closed-loop execution of building equipment, performance and resource monitoring, and model iterative optimization are performed to complete building collaborative optimization decisions, including: The inference feature vector is input into the ultra-lightweight model for intelligent inference to obtain control suggestions. The control suggestions are then sent to the building DDC controller, and sensor and DDC controller status data are uploaded in real time to achieve closed-loop control of building terminal devices. By real-time monitoring of sensor response latency, human thermal comfort index, and model resource occupancy, a building operation profile is obtained. Based on the building operation profile, input data quantization, engine interface adaptation, sample data fine-tuning, re-distillation, re-pruning, re-quantization, and adaptive updates of rule thresholds are performed to complete model iterative optimization and obtain building collaborative optimization decisions.

Images