Air conditioning intelligent control method and system based on artificial intelligence

Abstract

The present invention discloses an intelligent air-conditioning control method and system based on artificial intelligence, which relates to the field of artificial intelligence technology, including: first, constructing a first adaptation vector by fusing regional characteristics and strategy characteristics to characterize the adaptability of the controlled area and the control strategy to be enabled. The vector is input into a pre-trained intelligent air-conditioning control model, and under the constraint of the target strategy priority tensor table, the adaptability is evaluated by the working condition association method, and the optimized second adaptation vector is output, and finally the adaptive control strategy is dynamically enabled for multiple controlled areas according to the vector. The method realizes the intelligent matching of regional characteristics and strategies through the artificial intelligence model, and manages the strategy priority by using the tensor constraint mechanism, which solves the problems of insufficient multi-region adaptability and low strategy optimization efficiency in traditional control schemes, can dynamically improve the energy efficiency and user comfort of the air-conditioning system, and is suitable for air-conditioning collaborative control in multiple scenarios such as smart buildings and building automation.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to an air-conditioning intelligent control method and system based on artificial intelligence. Background Art

[0002] Existing air conditioning control systems often fail to dynamically adapt to the environmental characteristics of different regions and the differences in control strategy priorities in multi-zone collaborative control, resulting in low energy efficiency and insufficient user comfort. Traditional methods often use fixed strategies or simple threshold control, lacking intelligent assessment of regional characteristics and strategy adaptability, and lack a global constraint framework for strategy priorities, making it difficult to dynamically optimize control strategy configuration under complex operating conditions. Summary of the Invention

[0003] The purpose of the present invention is to provide an air conditioning intelligent control method and system based on artificial intelligence.

[0004] In a first aspect, an embodiment of the present invention provides an air conditioner intelligent control method based on artificial intelligence, comprising:

[0005] Determining a first adaptation vector based on regional features corresponding to the plurality of controlled regions and strategy features corresponding to the plurality of control strategies to be activated, wherein each characteristic value in the first adaptation vector represents a degree of adaptation of a controlled region to a control strategy to be activated;

[0006] The first adaptation vector is loaded into a pre-trained air conditioning intelligent control model to obtain a second adaptation vector output by the air conditioning intelligent control model; the air conditioning intelligent control model is used to evaluate the fitness value of each controlled area to the control strategy to be activated according to the working condition association method within the constraint framework of the target strategy priority tensor table, where each eigenvalue in the target strategy priority tensor table represents the priority coefficient of the corresponding control strategy to be activated;

[0007] At least one to-be-enabled control strategy is enabled for each of the plurality of controlled areas according to the second adaptation vector.

[0008] In a possible implementation, before loading the first adaptation vector into the air conditioning intelligent control model to obtain the second adaptation vector output by the air conditioning intelligent control model, the method further includes:

[0009] Determine an original strategy priority tensor table based on the activation counts corresponding to each control strategy instance included in each training data, where each activation count includes the number of times at least one controlled region instance executes the corresponding control strategy instance;

[0010] Determining a third adaptation vector based on the regional features of the controlled region instances and the strategy features of the control strategy instances respectively included in the respective training data, wherein each feature value in the third adaptation vector represents a fitness value of a controlled region instance to a control strategy instance;

[0011] The air conditioning intelligent control model is trained according to the original strategy priority tensor table and the third adaptation vector to obtain the pre-trained air conditioning intelligent control model.

[0012] In a possible implementation, the original strategy priority tensor table is determined based on the number of activations corresponding to each control strategy instance included in each training data, including:

[0013] performing an aggregation operation on the control strategy instances included in each training data according to the number of activations corresponding to each control strategy instance included in the training data, to obtain a plurality of control strategy clusters, each control strategy cluster including at least one control strategy instance;

[0014] sorting the plurality of control policy clusters by priority according to activation count values ​​corresponding to the plurality of control policy clusters, wherein each activation count value is determined according to the activation count of the control policy instances included in the corresponding control policy cluster;

[0015] According to the priority layer sorting, the priority coefficients of the plurality of control strategy clusters adopted are determined, and according to the priority coefficients of the plurality of control strategy clusters adopted, the original strategy priority tensor table is determined.

[0016] In a possible implementation, training an air conditioning intelligent control model according to the original strategy priority tensor table and the third adaptation vector includes:

[0017] Performing a policy scheduling benchmark configuration on the air-conditioning intelligent control model according to the original policy priority tensor table and the third adaptation vector;

[0018] Obtaining a fourth adaptation vector based on the air conditioning intelligent control model configured according to the policy scheduling benchmark, wherein an eigenvalue in the fourth adaptation vector represents a constraint framework of a priority coefficient currently adopted in a corresponding control strategy instance, and a fitness value of a controlled area instance to the control strategy instance;

[0019] setting an error function corresponding to the air-conditioning intelligent control model according to an expected deviation of the degree of fitness between the third fitness vector and the fourth fitness vector;

[0020] According to the error function, the Adam optimizer is used to perform model tuning on the air conditioning intelligent control model.

[0021] In one possible implementation, the parameters of the air conditioning intelligent control model include a target policy priority tensor table, a regional adaptation vector, and a policy adaptation vector; and performing a policy scheduling benchmark configuration on the air conditioning intelligent control model based on the original policy priority tensor table and the third adaptation vector includes:

[0022] Configuring the target policy priority tensor table based on the policy scheduling benchmark of the original policy priority tensor table;

[0023] Performing vector analysis on the third adaptation vector to obtain an original region feature vector and an original strategy feature vector;

[0024] configuring the regional adaptation vector based on the original regional feature vector policy scheduling benchmark, and configuring the policy adaptation vector based on the original policy feature vector policy scheduling benchmark;

[0025] According to the air conditioning intelligent control model configured according to the policy scheduling benchmark, the fourth adaptation vector is obtained, including:

[0026] According to the regional adaptation vector, the strategy adaptation vector and the target strategy priority tensor table, the adaptation value of each controlled region instance to each control strategy instance is determined to obtain the fourth adaptation vector.

[0027] In a possible implementation, determining the fitness value of each controlled area instance to each control strategy instance according to the area adaptation vector, the strategy adaptation vector, and the target strategy priority tensor table includes:

[0028] Determining a comprehensive adaptation coefficient according to all adaptation degree values ​​included in the third adaptation vector;

[0029] Determining a region adaptation coefficient corresponding to each controlled region instance according to at least one adaptation degree value corresponding to each controlled region instance in the third adaptation vector;

[0030] Determining a strategy adaptation coefficient corresponding to each control strategy instance according to at least one adaptation degree value corresponding to each control strategy instance in the third adaptation vector;

[0031] According to the regional adaptation vector, the strategy adaptation vector, the target strategy priority tensor table, the comprehensive adaptation coefficient, the regional adaptation coefficient and the strategy adaptation coefficient, the adaptation value of each controlled area instance to each control strategy instance is determined to obtain the fourth adaptation vector.

[0032] In a possible implementation, determining the first adaptation vector according to the regional features corresponding to the plurality of controlled regions and the strategy features corresponding to the plurality of control strategies to be enabled includes:

[0033] Evaluate the regional characteristics of the multiple controlled areas and the strategy characteristics of the multiple control strategies to be activated according to the response prediction model to obtain first evaluation information, where the first evaluation information is used to indicate the response tendency of the multiple controlled areas to activate the respective control strategies to be activated;

[0034] The first adaptation vector is determined according to the first evaluation information.

[0035] In a possible implementation, the response prediction model is obtained through the following process, including:

[0036] Extracting features of discrete scenes corresponding to controlled area instances included in each training data to obtain corresponding area feature vectors, and extracting features of discrete scenes corresponding to control strategy instances included in each training data to obtain corresponding strategy feature vectors;

[0037] According to the regional feature vectors and continuous scene features corresponding to each controlled area instance and the strategy feature vectors and continuous scene features corresponding to each control strategy instance, the response tendency value corresponding to each training data is evaluated;

[0038] According to the enabled response labels and corresponding response tendency values ​​included in each training data, an error function corresponding to the response prediction model is set, and the response prediction model is tuned according to the error function.

[0039] In a possible implementation, determining the first adaptation vector according to the first evaluation information includes:

[0040] Obtaining second evaluation information based on the first evaluation information and a preset environmental adjustment factor, wherein the second evaluation information is used to indicate a fitness value of the plurality of controlled areas to each of the control strategies to be activated;

[0041] Dividing the fitness value threshold into segments according to the second evaluation information to obtain a plurality of fitness value threshold segments, each fitness value threshold segment corresponding to an fitness threshold level;

[0042] Vector reconstruction is performed according to the multiple adaptation degree value threshold segments to obtain the first adaptation vector.

[0043] In a second aspect, an embodiment of the present invention provides a server system, including a server, wherein the server is configured to execute the method described in the first aspect.

[0044] Compared with the prior art, the beneficial effects provided by the present invention include: using an artificial intelligence-based air-conditioning intelligent control method and system disclosed by the present invention, a first adaptation vector is constructed by fusing regional characteristics and strategy characteristics to characterize the adaptability of the controlled area and the control strategy to be enabled. This vector is input into the pre-trained air-conditioning intelligent control model, and under the constraints of the target strategy priority tensor table, the adaptability is evaluated by the working condition association method, and the optimized second adaptation vector is output. Finally, the adaptive control strategy is dynamically enabled for multiple controlled areas based on the vector. The method realizes intelligent matching of regional characteristics and strategies through an artificial intelligence model, and manages strategy priorities using a tensor constraint mechanism, solving the problems of insufficient multi-region adaptability and low strategy optimization efficiency in traditional control schemes. It can dynamically improve the energy efficiency and user comfort of the air-conditioning system, and is suitable for collaborative air-conditioning control in multiple scenarios such as smart buildings and building automation. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly describes the drawings required for use in the embodiments. It should be understood that the following drawings illustrate only certain embodiments of the present invention and should not be construed as limiting the scope of the present invention. Those skilled in the art can, without inventive effort, derive other relevant drawings from these drawings.

[0046] Figure 1 A schematic flow chart of the steps of an air-conditioning intelligent control method based on artificial intelligence provided by an embodiment of the present invention;

[0047] Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0048] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the embodiments of the present invention. It should be understood that the described embodiments are only a portion of the embodiments of the present invention, not all of them. Generally, the components of the embodiments of the present invention described and illustrated in the drawings herein may be arranged and designed in a variety of different configurations.

[0049] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0050] In order to solve the technical problems in the above background technology, Figure 1 This is a flow chart of an air-conditioning intelligent control method based on artificial intelligence provided by an embodiment of the present disclosure. The air-conditioning intelligent control method based on artificial intelligence is introduced in detail below.

[0051] Step S201: determining a first adaptation vector based on regional features corresponding to a plurality of controlled regions and strategy features corresponding to a plurality of control strategies to be activated, wherein each characteristic value in the first adaptation vector represents a degree of adaptation of a controlled region to a control strategy to be activated;

[0052] Step S202: Loading the first adaptation vector into a pre-trained air conditioning intelligent control model to obtain a second adaptation vector output by the air conditioning intelligent control model; the air conditioning intelligent control model is used to evaluate the fitness value of each controlled area to the control strategy to be activated according to the working condition association method within the constraint framework of the target strategy priority tensor table, wherein each eigenvalue in the target strategy priority tensor table represents the priority coefficient of the corresponding control strategy to be activated;

[0053] Step S203: according to the second adaptation vector, enable at least one control strategy to be enabled to each of the plurality of controlled areas.

[0054] In this embodiment of the present invention, for example, a server, acting as the execution entity, implements an AI-based intelligent air conditioning control method for a smart office building within a smart park. The office building comprises three controlled areas: conference room A, open office area B, and employee lounge area C. Each area has distinct spatial attributes, environmental parameters, and usage scenarios. The control strategies to be activated include Strategy 1 (target temperature 25°C, medium wind speed, comfort mode), Strategy 2 (target temperature 26°C, low wind speed, energy-saving mode), and Strategy 3 (target temperature 24°C, high wind speed, extreme speed mode), covering different control requirements from various dimensions, including temperature regulation, wind speed configuration, and operating mode.

[0055] The server first collects regional characteristics for each controlled area: Conference Room A's characteristics include spatial attributes (80 square meters), real-time environmental parameters (temperature 28°C, humidity 65%), and usage scenario (a meeting at 10:00 AM with a density of 15 people per 80 square meters, a high-comfort scenario); Office Area B's characteristics include spatial attributes (300 square meters), environmental parameters (temperature 26°C, humidity 60%), and usage scenario (normal office hours with a density of 80 people per 300 square meters, a scenario requiring a balance between energy consumption and comfort); Rest Area C's characteristics include spatial attributes (50 square meters), environmental parameters (temperature 27°C, humidity 70%), and usage scenario (leisure hours with a density of 5 people per 50 square meters, a scenario requiring low energy consumption and moderate comfort). Simultaneously, the server analyzes the policy characteristics of the control policies to be activated: Policy 1 prioritizes comfort (rapid temperature stabilization, gentle airflow, and moderate energy consumption); Policy 2 prioritizes energy conservation (moderate temperature adjustment rate and low energy consumption); and Policy 3 prioritizes rapid cooling (high air speed and high energy consumption).

[0056] To quantify the "region-strategy" adaptation relationship, the server invokes a response prediction model to assess response propensity. This model is trained based on historical data: discrete scene features (such as "meeting," "work," and "rest") and continuous scene features (temperature, humidity, and number of people) from historical controlled region instances are extracted to generate region feature vectors. Discrete scene features (such as "comfort," "energy saving," and "extreme speed") and continuous scene features (temperature adjustment range and wind speed level) from historical control strategy instances are extracted to generate strategy feature vectors. An error function is constructed by combining historical "activated response tags" (such as temperature compliance rate and occupant comfort feedback) with the response propensity values ​​output by the model. Training is completed through gradient descent optimization.

[0057] In this scenario, the response prediction model evaluates the "area-strategy" combination as follows: Conference Room A has a response tendency of 0.8 to Strategy 1 (comfort mode is highly compatible with meeting scenarios), 0.5 to Strategy 2 (slow cooling in energy-saving mode), and 0.3 to Strategy 3 (noise interference in extreme mode); Office Area B has a response tendency of 0.6 to Strategy 1 (high energy consumption in comfort mode), 0.7 to Strategy 2 (energy-saving mode is compatible with office needs), and 0.4 to Strategy 3 (extreme mode is not an essential office need); Rest Area C has a response tendency of 0.4 to Strategy 1 (comfort mode consumes more energy than energy-saving mode), 0.6 to Strategy 2 (energy-saving mode is compatible with rest needs), and 0.2 to Strategy 3 (extreme mode affects rest). The server directly maps these response tendency values ​​to adaptation values, arranges them according to the dimension of "controlled area-control strategy to be activated", and generates the first adaptation vector (the dimension is 3 areas × 3 strategies, and the elements are: A-1: ​​0.8, A-2: 0.5, A-3: 0.3, B-1: 0.6, B-2: 0.7, B-3: 0.4, C-1: 0.4, C-2: 0.6, C-3: 0.2).

[0058] The air conditioning intelligent control model needs to be pre-trained. The training logic combines the historical strategy activation rules and the adaptability relationship:

[0059] 1. Generate the original policy priority tensor table: The server counts the number of times the historical control policy instances have been activated (for example, the energy-saving policy is activated 300 times during office hours, the comfort policy is activated 200 times during meeting hours, and the extreme speed policy is activated 50 times during equipment testing). The data is aggregated into "energy-saving clusters," "comfort clusters," and "extreme speed clusters." These clusters are sorted by the number of activations (energy-saving cluster > comfort cluster > extreme speed cluster). The priority coefficients are determined (0.8 for energy-saving clusters, 0.6 for comfort clusters, and 0.3 for extreme speed clusters), forming the original policy priority tensor table.

[0060] 2. Training Model Parameters: The server generates a third adaptation vector (similar in structure to the first adaptation vector) based on historical region-policy fit. Using the original policy priority tensor as a benchmark, it initializes the model's target policy priority tensor. The server parses the third adaptation vector, extracts the original region feature vectors and policy feature vectors, and initializes the region adaptation vector (weights for each region scenario) and policy adaptation vector (weights for each policy mode). The model calculates the region-policy fit under historical scenarios using the region adaptation vector, policy adaptation vector, and target policy priority tensor to generate a fourth adaptation vector. The error function is constructed based on the deviation between the third and fourth adaptation vectors, and the model parameters are iteratively tuned using the Adam optimizer.

[0061] During the real-time control phase, after loading the first adaptation vector into the trained air conditioning intelligent control model, the model dynamically adjusts the adaptation degree within the constraints of the target strategy priority tensor, taking into account the combined working condition of "10:00 AM meeting, office, and rest." In the target strategy priority tensor, the priority coefficient of Strategy 2 (Energy Saving Cluster) in Office Area B is increased to 0.8 (energy saving is the primary demand during office hours), the priority coefficient of Strategy 1 (Comfort Cluster) in Conference Room A is increased to 0.9 (comfort is a high requirement in meeting scenarios), and the priority coefficient of Strategy 3 (Extreme Speed ​​Cluster) in Rest Area C is reduced to 0.3 (extra speed is not a requirement in rest scenarios). Using a working condition association method (fusion time and scenario weighting), the model performs a weighted fusion of the first adaptation vector with the target strategy priority tensor, the regional adaptation vector, and the strategy adaptation vector. The "region-strategy" adaptation degree is recalculated to generate the second adaptation vector. For example: In the conference room, the adaptability of strategy 1 is adjusted from 0.8 to 0.85 (the weight of meeting scenarios and comfort mode is increased), strategy 2 is adjusted to 0.4 (the priority of meeting scenarios in energy-saving mode is reduced), and strategy 3 is adjusted to 0.2 (the extreme mode scenario does not match); in the office area, strategy 2 is adjusted to 0.8 (the priority of energy-saving in office scenarios is increased), strategy 1 is adjusted to 0.5 (the weight of energy consumption in office scenarios in comfort mode is reduced), and strategy 3 is adjusted to 0.3 (the extreme mode is not an essential requirement for office work); in the rest area, strategy 2 is adjusted to 0.7 (the priority of energy-saving in rest scenarios is increased), strategy 1 is adjusted to 0.5 (the energy consumption in comfort mode is higher than that in energy-saving mode), and strategy 3 is adjusted to 0.1 (the extreme mode scenario does not match).

[0062] The server performs a threshold check on the second fitness vector (setting fitness ≥ 0.6 to indicate "good fitness") and iterates through the vector elements: In Conference Room A, if Strategy 1 has a fitness of 0.85 ≥ 0.6, it sends a command to the air conditioning system to activate Strategy 1 (temperature 25°C, medium fan speed, comfort mode); in Office Area B, if Strategy 2 has a fitness of 0.8 ≥ 0.6, it sends a command to activate Strategy 2 (temperature 26°C, low fan speed, energy-saving mode); and in Rest Area C, if Strategy 2 has a fitness of 0.7 ≥ 0.6, it sends a command to activate Strategy 2 (temperature 26°C, low fan speed, energy-saving mode). If a given area has multiple policies with fitness ≥ the threshold (for example, in Rest Area C, if Strategy 1 has a fitness of 0.55 and the threshold is 0.5, dual policies are activated), the server will issue a combined policy command (e.g., first extreme cooling followed by comfort maintenance). In this scenario, each area has a unique policy with the highest fitness, so a single policy is used.

[0063] Through the above process, the server conducts intelligent and precise air conditioning control in multiple scenarios, from region-strategy fit analysis and model-constrained fit optimization to precise policy activation. This method, by continuously updating training data and expanding policy dimensions, can adapt to more complex scenarios such as large commercial complexes and industrial parks, providing efficient support for energy management and environmental control in smart buildings, minimizing energy consumption while meeting user comfort needs.

[0064] In an embodiment of the present invention, before loading the first adaptation vector into the air-conditioning intelligent control model to obtain the second adaptation vector output by the air-conditioning intelligent control model, the following implementation manner is further provided.

[0065] Determine an original strategy priority tensor table based on the activation counts corresponding to each control strategy instance included in each training data, where each activation count includes the number of times at least one controlled region instance executes the corresponding control strategy instance;

[0066] Determining a third adaptation vector based on the regional features of the controlled region instances and the strategy features of the control strategy instances respectively included in the respective training data, wherein each feature value in the third adaptation vector represents a fitness value of a controlled region instance to a control strategy instance;

[0067] The air conditioning intelligent control model is trained according to the original strategy priority tensor table and the third adaptation vector to obtain the pre-trained air conditioning intelligent control model.

[0068] In an embodiment of the present invention, during the training phase of the intelligent air conditioning control model, the server executes a process based on nearly a year's worth of air conditioning control operation data from a smart office building. First, to determine the original policy priority tensor table, the server retrieves historical logs, which record the number of executions of control policy instances such as "Energy Saving Mode (corresponding to Policy 2)," "Comfort Mode (corresponding to Policy 1)," and "Extreme Speed ​​Mode (corresponding to Policy 3)" for controlled areas such as Conference Room A, Office Area B, and Rest Area C. Energy Saving Mode was activated 280 times due to daily office energy consumption control requirements in Office Area B; Comfort Mode was activated 220 times due to meeting comfort requirements in Conference Room A; and Extreme Speed ​​Mode was activated only during equipment testing or extreme high temperatures, for a total of 50 times. The server aggregates policies with similar functions and scenarios into "Energy Saving Policy Cluster," "Comfort Policy Cluster," and "Extreme Speed ​​Policy Cluster," sorting them by total activation count (Energy Saving Policy Cluster > Comfort Policy Cluster > Extreme Speed ​​Policy Cluster). Priority coefficients of 0.8 are assigned to the Energy Saving Cluster, 0.6 to the Comfort Cluster, and 0.3 to the Extreme Speed ​​Cluster, generating the original policy priority tensor table.

[0069] Next, to determine the third adaptation vector, the server extracts regional characteristics of historical controlled area instances (such as the spatial dimensions and occupancy density of Conference Room A; the workstation distribution and energy consumption threshold of Office Area B) and policy characteristics of control policy instances (temperature adjustment gradients in energy-saving mode and airflow uniformity parameters in comfort mode). Using a weighted algorithm based on scenario-specific matching and energy efficiency, the server quantifies the degree of adaptation: Conference Room A's adaptability to comfort mode is 0.85 (rapid temperature uniformity meets physical requirements), Office Area B's adaptability to energy-saving mode is 0.8 (balancing energy consumption and comfort), and Rest Area C's adaptability to energy-saving mode is 0.75 (reducing standby power consumption). All the "region instance-policy instance" adaptations are arranged by dimension to form the third adaptation vector, providing a data foundation for the model to learn historical patterns.

[0070] Finally, when training the air conditioning intelligent control model, the server uses the original policy priority tensor as a benchmark and inputs the third adaptation vector into the model, driving the model to adjust parameters such as the target policy priority tensor, regional adaptation vector, and policy adaptation vector. The model first simulates the "region-policy" fitness calculation logic to generate the fourth fitness vector (the predicted fitness during the training phase). The model then compares the deviation between the third fitness vector (the actual historical fitness) and the fourth fitness vector, constructs a mean squared error function, and iteratively tunes the parameters using the Adam optimizer until the deviation converges. After multiple rounds of training, the model accurately learns the historical policy priorities and regional-policy adaptation patterns, becoming a pre-trained model supporting real-time control.

[0071] In an embodiment of the present invention, the original strategy priority tensor table is determined according to the number of times each control strategy instance included in each training data is enabled, which can be implemented through the following example.

[0072] performing an aggregation operation on the control strategy instances included in the training data according to the number of activations, to obtain a plurality of control strategy clusters, each control strategy cluster including at least one control strategy instance;

[0073] sorting the plurality of control policy clusters by priority according to activation count values ​​corresponding to the plurality of control policy clusters, wherein each activation count value is determined according to the activation count of the control policy instances included in the corresponding control policy cluster;

[0074] According to the priority layer sorting, the priority coefficients of the plurality of control strategy clusters adopted are determined, and according to the priority coefficients of the plurality of control strategy clusters adopted, the original strategy priority tensor table is determined.

[0075] In an embodiment of the present invention, exemplarily, in an air conditioning control training data processing scenario of a smart office building, when the server executes the process of determining the original policy priority tensor table, it first performs an aggregation operation on the control policy instance. The server traversed the control strategy instances recorded in the historical training data and found that strategy instances such as "Summer Energy Saving Mode (set temperature 26°C, low wind speed)" and "Daily Energy Saving Mode (set temperature 25°C, low wind speed)" were all designed around the core goal of "reducing energy consumption and maintaining basic comfort", so these strategy instances were aggregated into an "energy saving strategy cluster"; strategy instances such as "Conference Comfort Mode (set temperature 24°C, medium wind speed)" and "Reception Comfort Mode (set temperature 25°C, medium wind speed)" focused on "rapid temperature equalization and improved airflow softness" to meet the needs of crowded scenes, and were aggregated into a "comfort strategy cluster"; strategy instances such as "Equipment Testing Extreme Mode (set temperature 22°C, high wind speed)" and "Extreme High Temperature Extreme Mode (set temperature 20°C, high wind speed)" were targeted at special scenarios of "extreme cooling in a short period of time" and were aggregated into an "Extreme Speed ​​Strategy Cluster". Each control strategy cluster thus contains at least one control strategy instance with similar functional logic.

[0076] Next, the server counts the activation counts for each control policy cluster. For the "Energy Savings" cluster, it extracts 150 records of "Summer Energy Savings Mode" being activated in Office Area B between 9:00 and 18:00 on weekdays, and 130 records of "Daily Energy Savings Mode" being activated in Rest Area C during off-peak hours, totaling 280 activations for this cluster. For the "Comfort" cluster, it extracts 120 records of "Conference Comfort Mode" being activated in Conference Room A during meeting hours, and 100 records of "Reception Comfort Mode" being activated in the temporary exhibition hall (a controlled area instance) during reception hours, totaling 220 activations. For the "Extreme Speed" cluster, it extracts 30 records of "Equipment Test Extreme Speed" being activated in the test area near the equipment room, and 20 records of "Extreme Heat Extreme Speed" being activated throughout the office building on an extremely hot day, totaling 50 activations. Based on these three activation counts, the server prioritizes the control policy clusters: "Energy Savings" cluster (280 times) > "Comfort" cluster (220 times) > "Extreme Speed" cluster (50 times)."

[0077] Finally, the server determines the priority coefficients of each control strategy cluster based on the priority hierarchy. Because the "Energy Saving Strategy Cluster" is activated the most frequently and provides the strongest support for daily operational energy consumption control, it is assigned a priority coefficient of 0.8. The "Comfort Strategy Cluster" meets the comfort needs of high-frequency meetings and receptions, is activated the second most frequently, and is assigned a priority coefficient of 0.6. The "Extreme Speed ​​Strategy Cluster" serves only special scenarios, is activated the least frequently, and is assigned a priority coefficient of 0.3. The server associates these priority coefficients with the corresponding control strategy clusters, organizing them into a two-dimensional table structure with "Energy Saving, Comfort, Extreme Speed" as the strategy dimensions and "0.8, 0.6, 0.3" as the priority coefficients. This completes the determination of the original strategy priority tensor table, providing a strategy priority benchmark for the subsequent training of the air conditioning intelligent control model.

[0078] In an embodiment of the present invention, the air-conditioning intelligent control model is trained according to the original strategy priority tensor table and the third adaptation vector, which can be implemented through the following examples.

[0079] Performing a policy scheduling benchmark configuration on the air-conditioning intelligent control model according to the original policy priority tensor table and the third adaptation vector;

[0080] Obtaining a fourth adaptation vector based on the air conditioning intelligent control model configured according to the policy scheduling benchmark, wherein an eigenvalue in the fourth adaptation vector represents a constraint framework of a priority coefficient currently adopted in a corresponding control strategy instance, and a fitness value of a controlled area instance to the control strategy instance;

[0081] setting an error function corresponding to the air-conditioning intelligent control model according to an expected deviation of the degree of fitness between the third fitness vector and the fourth fitness vector;

[0082] According to the error function, the Adam optimizer is used to perform model tuning on the air conditioning intelligent control model.

[0083] In an embodiment of the present invention, illustratively, in the training process of the smart office building air conditioning intelligent control model, the server first executes the policy scheduling benchmark configuration. Based on the original policy priority tensor table (priority coefficient 0.8 for the energy-saving policy cluster, 0.6 for the comfort policy cluster, and 0.3 for the extreme speed policy cluster), the server directly uses this table as the initial benchmark for the "target policy priority tensor table" within the model. At the same time, the server parses the third adaptation vector (recording the "region instance-policy instance" adaptation in historical scenarios, such as 0.85 for the comfort policy instance in conference room A and 0.8 for the energy-saving policy instance in office area B). The server extracts the spatial attributes and scene labels of the controlled area instances to generate the "original region feature vector" (such as a comfort weight of 0.7 for the meeting scene, an energy consumption weight of 0.6 for the office scene, and a leisure weight of 0.5 for the rest scene). The server also extracts the temperature adjustment range and wind speed level of the control policy instances to generate the "original policy feature vector" (such as a energy-saving mode energy consumption weight of 0.8, a comfort mode airflow weight of 0.9, and an extreme speed mode cooling weight of 0.7). These two vectors are used as the initial configurations of the model's "region adaptation vector" and "policy adaptation vector," respectively, completing the baseline initialization of the model parameters.

[0084] The server then obtains the fourth adaptation vector. The model, after the policy scheduling baseline configuration, recalculates the fitness of each controlled area instance-control policy instance based on the constraints of the "target policy priority tensor table" (for example, the energy-saving policy cluster's priority of 0.8 strengthens the energy-saving policy weight in the office area B scenario), combined with the "regional adaptation vector" (the comfort demand weight of the meeting scenario) and the "strategy adaptation vector" (the comfort policy's weight for temperature uniformity). For example, for Conference Room A and the comfort policy instance, the model, within the framework of the comfort policy cluster's priority of 0.6, superimposes the regional weight of 0.7 for the meeting scenario and the policy weight of 0.9 for the comfort policy, resulting in a predicted fitness value of 0.82 for this combination (the actual value of this combination in the third adaptation vector is 0.85). Similarly, the fitness calculations for all region-policy combinations are performed, and the results are arranged by dimension to form the fourth adaptation vector. Each eigenvalue in this vector reflects both the policy priority constraint and the dynamic adaptation relationship between the region and the policy.

[0085] Next, the server sets the error function. The third fitness vector (the set of historically accurate fitness values) is compared element-by-element with the fourth fitness vector (the set of model-predicted fitness values). The fitness deviation for each "area-strategy" combination is calculated (for example, the deviation for Conference Room A - Comfort Strategy is |0.85-0.82| = 0.03). All deviations are integrated using the mean square error formula (MSE = Σ(actual value - predicted value)² / N, where N is the total number of combinations). This error function is then constructed for the air conditioning intelligent control model, quantifying the degree of deviation between the model predictions and historically accurate data.

[0086] Finally, the server uses the Adam optimizer to fine-tune the model. Based on the error function, the gradients of the model parameters (target policy priority tensor, regional adaptation vector, and policy adaptation vector) are calculated. The Adam optimizer updates these parameters at an adaptive learning rate. For example, to address the predicted deviation in the fitness of "Conference Room A - Comfort Policy," the Adam optimizer adjusts the comfort weight of the meeting scene in the regional adaptation vector and the temperature uniformity weight of the comfort policy in the policy adaptation vector. Simultaneously, the priority coefficients of the comfort policy cluster in the target policy priority tensor are fine-tuned, gradually bringing the model's predicted value closer to the true value of the third fitness vector. After multiple iterations (e.g., 100 epochs of training), when the error function stabilizes at a minimum (e.g., MSE ≤ 0.001), the model is considered fine-tuned and capable of accurately assessing the fitness of "region-policy" within the constraints of policy priorities, providing reliable pre-trained model support for subsequent real-time air conditioning control.

[0087] In an embodiment of the present invention, the parameters of the air-conditioning intelligent control model include a target policy priority tensor table, a regional adaptation vector, and a policy adaptation vector; then, according to the original policy priority tensor table and the third adaptation vector, the air-conditioning intelligent control model is subjected to a policy scheduling benchmark configuration, and the package can be implemented through the following example.

[0088] Configuring the target policy priority tensor table based on the policy scheduling benchmark of the original policy priority tensor table;

[0089] Performing vector analysis on the third adaptation vector to obtain an original region feature vector and an original strategy feature vector;

[0090] configuring the regional adaptation vector based on the original regional feature vector policy scheduling benchmark, and configuring the policy adaptation vector based on the original policy feature vector policy scheduling benchmark;

[0091] According to the air conditioning intelligent control model configured according to the policy scheduling benchmark, the fourth adaptation vector is obtained, including:

[0092] According to the regional adaptation vector, the strategy adaptation vector and the target strategy priority tensor table, the adaptation value of each controlled region instance to each control strategy instance is determined to obtain the fourth adaptation vector.

[0093] In an embodiment of the present invention, exemplarily, during the training phase of the smart office building air conditioning intelligent control model, when the server executes the policy scheduling benchmark configuration and the fourth adaptation vector generation process, the target policy priority tensor is first configured with the original policy priority tensor: the server directly uses the original policy priority tensor obtained by historical aggregation (priority coefficient of "energy-saving policy cluster" 0.8, "comfort policy cluster" 0.6, "extreme speed policy cluster" 0.3) as the initial benchmark value of the "target policy priority tensor" in the model, so that the priority distribution law of the historical policy is inherited in the initial stage of model training - for example, in the office scenario, the energy-saving policy is activated at a high frequency, and its priority weight is retained in the target policy priority tensor, providing a basis for the priority constraint of subsequent policy scheduling.

[0094] Next, the third adaptation vector is vector-parsed: the third adaptation vector records the adaptability of the "controlled area instance-control strategy instance" in the historical scenario (for example, the adaptability of conference room A to the comfort strategy instance is 0.85, and the adaptability of office area B to the energy-saving strategy instance is 0.8, etc.). The server extracts information such as discrete scene labels ("meeting," "office," and "rest") and continuous environmental parameters (area, temperature, and occupancy density) from the controlled area instances to generate original regional feature vectors. For example, since meeting scenes require high temperature uniformity due to the large number of people, the corresponding weight of the "meeting scene-comfort" dimension is set to 0.7; since office scenes need to balance energy consumption and comfort, the weight of the "office scene-energy balance" dimension is set to 0.6; since rest scenes focus on low power consumption, the weight of the "rest scene-low power consumption" dimension is set to 0.5. The server also extracts information such as discrete mode labels ("energy saving," "comfort," and "extreme speed") and continuous control parameters (temperature adjustment range, wind speed level) from the control strategy instances to generate original strategy feature vectors. For example, since energy-saving mode is frequently used for energy consumption control during office hours, the weight of the "energy-saving mode-energy consumption optimization" dimension is set to 0.8; since comfort mode requires high airflow softness in meeting scenes, the weight of the "comfort mode-airflow uniformity" dimension is set to 0.9; since extreme speed mode only serves special scenarios, the weight of the "extreme speed mode-cooling rate" dimension is set to 0.7.

[0095] Subsequently, the regional adaptation vector and the policy adaptation vector are configured: the server uses the original regional feature vector as the initialization benchmark for the "regional adaptation vector", allowing the model to prioritize learning the historical demand weights for regional regulation in various scenarios - for example, the "comfort" weight of the conference scene is high in the original regional feature vector, and the comfort-related weight of "Conference Room A-Conference Scene" in the corresponding regional adaptation vector is simultaneously strengthened; similarly, the "policy adaptation vector" is initialized based on the original policy feature vector, so that the model inherits the contribution weights of each policy to the regulation target - for example, the "energy consumption optimization" weight of the energy-saving mode is high in the original policy feature vector, and the related weight of "energy-saving policy-energy consumption control" in the policy adaptation vector is strengthened accordingly, ensuring that the initial state of the model parameters fits the historical regulation logic.

[0096] Finally, the fourth adaptation vector is generated: the model after the policy scheduling benchmark configuration recalculates the fitness for each "controlled area instance-control strategy instance" combination based on the "regional adaptation vector" (weight of each regional scenario demand), "strategy adaptation vector" (weight of each strategy control target) and "target strategy priority tensor" (strategy priority coefficient). Taking conference room A (weight of "conference scene" in the regional adaptation vector is 0.7) and the comfort strategy instance (weight of "air flow uniformity" in the strategy adaptation vector is 0.9, and coefficient of "comfort strategy cluster" in the target strategy priority tensor is 0.6) as an example, the model calculates the predicted fitness value of this group as 0.82 through multi-dimensional weighted fusion (such as fitness = regional scenario weight × strategy control weight × strategy priority coefficient + environmental parameter correction term) (the actual fitness of this combination in the third adaptation vector is 0.85); similarly, office area B ("office" in the regional adaptation vector) The predicted fitness value for the "public scene" (weight 0.6) and the energy-saving strategy instance (weight 0.8 for "energy consumption optimization" in the strategy adaptation vector and coefficient 0.8 for "energy-saving strategy cluster" in the target strategy priority tensor table) is 0.6×0.8×0.8+temperature difference correction term=0.384+0.016=0.4 (this is only an example logic; the actual calculation needs to be combined with complex scenario parameters). After completing the fitness calculation for all "region-strategy" combinations, the results are arranged by dimension to form the fourth fitness vector, which provides predictive data support for subsequent model error calculation and parameter tuning.

[0097] Through the above steps, the server completes the baseline configuration of core parameters at the initial stage of model training, and generates a prediction fitness vector based on multi-dimensional weight fusion to ensure that the model training process conforms to historical control rules and scenario demand logic.

[0098] In an embodiment of the present invention, the adaptation value of each controlled area instance to each control strategy instance is determined according to the area adaptation vector, the strategy adaptation vector and the target strategy priority tensor table, which can be implemented through the following examples.

[0099] Determining a comprehensive adaptation coefficient according to all adaptation degree values ​​included in the third adaptation vector;

[0100] Determining a region adaptation coefficient corresponding to each controlled region instance according to at least one adaptation degree value corresponding to each controlled region instance in the third adaptation vector;

[0101] Determining a strategy adaptation coefficient corresponding to each control strategy instance according to at least one adaptation degree value corresponding to each control strategy instance in the third adaptation vector;

[0102] According to the regional adaptation vector, the strategy adaptation vector, the target strategy priority tensor table, the comprehensive adaptation coefficient, the regional adaptation coefficient and the strategy adaptation coefficient, the adaptation value of each controlled area instance to each control strategy instance is determined to obtain the fourth adaptation vector.

[0103] In an embodiment of the present invention, for example, in the training process of the smart office building air conditioning intelligent control model, when the server executes the step of "determining the fitness value of each controlled area instance to the control strategy instance", it is necessary to integrate the multi-dimensional coefficients and model parameters in stages:

[0104] First, the comprehensive adaptation coefficient is determined: the server traverses the adaptation values ​​of all "region instance-policy instance" included in the third adaptation vector (for example, Conference Room A has an adaptation of 0.85, 0.5, and 0.3 for the comfort, energy saving, and extreme speed policies; Office Area B has an adaptation of 0.6, 0.7, and 0.4 for the three policies; and Rest Area C has an adaptation of 0.4, 0.6, and 0.2 for the three policies). The server calculates the arithmetic mean of these nine adaptation values, that is, ((0.85 + 0.5 + 0.3 + 0.6 + 0.7 + 0.4 + 0.4 + 0.6 + 0.2) ÷ 9 ≈ 0.52). This value serves as the comprehensive adaptation coefficient, reflecting the overall adaptation level of the "region-policy" combination in the historical data and providing a benchmark for subsequent adaptation calculations.

[0105] Next, the regional adaptation coefficient is determined: for each controlled area instance, the mean adaptation degree for all control policy instances is calculated. Conference Room A's adaptation degrees for the comfort, energy-saving, and speed policies are 0.85, 0.5, and 0.3, respectively. The mean ((0.85 + 0.5 + 0.3) ÷ 3 ≈ 0.55) is calculated as the regional adaptation coefficient for Conference Room A. The mean adaptation degree for the three policies in Office Area B is ((0.6 + 0.7 + 0.4) ÷ 3 ≈ 0.57), which is the regional adaptation coefficient for Office Area B. The mean adaptation degree for the three policies in Rest Area C is ((0.4 + 0.6 + 0.2) ÷ 3 ≈ 0.4), which is the regional adaptation coefficient for Rest Area C. These coefficients reflect the average historical adaptation tendency of each area to different policies. For example, Conference Room A's regional adaptation coefficient of 0.55 is higher than the overall adaptation coefficient of 0.52, indicating that its overall adaptation degree is slightly higher than the average for all areas.

[0106] Next, the strategy adaptation coefficient was determined: for each control strategy instance, the mean adaptation value for all controlled area instances was calculated. The energy-saving strategy (corresponding to strategy 2) had adaptation values ​​of 0.5, 0.7, and 0.6 in conference room A, office area B, and rest area C, respectively. The mean ((0.5 + 0.7 + 0.6) ÷ 3 ≈ 0.6) was calculated as the strategy adaptation coefficient for the energy-saving strategy. The comfort strategy (corresponding to strategy 1) had a mean adaptation value of ((0.85 + 0.6 + 0.4) ÷ 3 ≈ 0.62) for the three areas, which was used as the strategy adaptation coefficient for the comfort strategy. The extreme speed strategy (corresponding to strategy 3) had a mean adaptation value of ((0.3 + 0.4 + 0.2) ÷ 3 ≈ 0.3) for the three areas, which was used as the strategy adaptation coefficient for the extreme speed strategy. These coefficients reflect the average adaptation characteristics of each strategy for different areas over time. For example, the comfort strategy's strategy adaptation coefficient of 0.62 was higher than the overall adaptation coefficient of 0.52, indicating that its overall adaptation was superior to the average level of all strategies.

[0107] Finally, the fourth adaptation vector is generated by fusion: the model parameters (region adaptation vector, policy adaptation vector, target policy priority tensor table) after the server calls the policy scheduling benchmark configuration are combined with the above coefficients for multi-dimensional weighted calculation. Taking the calculation of the compatibility between Conference Room A and the comfort strategy instance as an example: the "conference scene comfort weight" of Conference Room A in the regional adaptation vector is 0.7, the "airflow uniformity weight" of the comfort strategy in the strategy adaptation vector is 0.9, and the priority coefficient of the comfort strategy cluster in the target strategy priority tensor is 0.6. Furthermore, introducing the comprehensive adaptation coefficient of 0.52, the regional adaptation coefficient of Conference Room A of 0.55, and the strategy adaptation coefficient of the comfort strategy of 0.62, the formula (fitness = (regional adaptation vector value × strategy adaptation vector value × target priority coefficient) × (regional adaptation coefficient ÷ comprehensive adaptation coefficient) × (strategy adaptation coefficient ÷ comprehensive adaptation coefficient)) yields ((0.7 × 0.9 × 0.6) × (0.55 ÷ 0.52) × (0.62 ÷ 0.52) ≈ 0.378 × 1.057 × 1.192 ≈ 0.477) is used. Similarly, the fitness calculations for all "area-strategy" combinations, including office area B and the energy-saving strategy (calculated fitness ≈0.58) and rest area C and the energy-saving strategy (calculated fitness ≈0.27), are completed. The results are arranged by dimension to form the fourth fitness vector, providing predictive data support for subsequent model error assessment and parameter tuning.

[0108] Through multi-coefficient fusion and model parameter linkage, the server accurately simulates the historical laws of "region-strategy" adaptability during training, ensuring that the fourth adaptation vector inherits the original strategy priority and region-strategy characteristics, and conforms to the real scenario logic through coefficient correction.

[0109] In the embodiment of the present invention, determining the first adaptation vector according to the area features corresponding to the multiple controlled areas and the strategy features corresponding to the multiple control strategies to be enabled can be implemented through the following examples.

[0110] Evaluate the regional characteristics of the multiple controlled areas and the strategy characteristics of the multiple control strategies to be activated according to the response prediction model to obtain first evaluation information, where the first evaluation information is used to indicate the response tendency of the multiple controlled areas to activate the respective control strategies to be activated;

[0111] The first adaptation vector is determined according to the first evaluation information.

[0112] In an exemplary embodiment of the present invention, in a real-time scenario of intelligent air conditioning control in a smart office building, when the server executes the "determine the first adaptation vector" process, it first relies on a response prediction model to evaluate the responsiveness of regions and strategies. This model has been trained using historical data. It extracts features from discrete scene labels (e.g., "meeting," "work," "rest") and continuous environmental parameters (area, temperature, and occupancy density) for controlled area instances such as Conference Room A, Office Area B, and Rest Area C to generate regional feature vectors. It also extracts features from discrete mode labels (e.g., "comfort," "energy saving," and "extreme speed") and continuous control parameters (e.g., temperature adjustment range and wind speed level) for control strategy instances such as Strategy 1 (Comfort Mode, including parameters such as temperature 25°C and medium wind speed), Strategy 2 (Energy Saving Mode, temperature 26°C and low wind speed), and Strategy 3 (Extreme Speed ​​Mode, temperature 24°C and high wind speed) to generate strategy feature vectors. The model then constructs an error function based on historical "activated response labels" (e.g., time to reach temperature target and employee satisfaction feedback) and optimizes the model to accurately predict the responsiveness of each "region-strategy" combination.

[0113] During the real-time control phase, the server collects regional characteristics of the currently controlled area: Conference Room A, in a 10:00 AM meeting scenario, has an 80-square-meter space, a real-time temperature of 28°C, a humidity of 65%, and a occupancy density of 15 people per 80 square meters (high comfort requirements); Office Area B, during normal office hours, has a 300-square-meter space, a temperature of 26°C, a humidity of 60%, and a occupancy density of 80 people per 300 square meters (balancing energy consumption and comfort requirements); Rest Area C, during leisure time, has a 50-square-meter space, a temperature of 27°C, a humidity of 70%, and a occupancy density of 5 people per 50 square meters (low energy consumption and moderate comfort requirements). The server also analyzes the characteristics of the control strategies to be activated: Strategy 1 (Comfort mode, focusing on rapid temperature averaging, gentle airflow, and moderate energy consumption); Strategy 2 (Energy-saving mode, focusing on gentle temperature regulation and low energy consumption); and Strategy 3 (Extreme mode, focusing on high-speed cooling and high energy consumption).

[0114] The server inputs the aforementioned regional and policy characteristics into the response prediction model. Based on the scenario-policy matching logic learned through training, the model outputs the first evaluation information (i.e., the response propensity of each "region-policy" combination): Conference Room A's response propensity to Strategy 1 is 0.8 (Conference scenarios require rapid creation of a comfortable environment, and Strategy 1's temperature adjustment rate and wind speed level are highly matched); the response propensity to Strategy 2 is 0.5 (Energy-saving mode has low cooling efficiency and cannot meet the timely temperature control requirements of meetings); and the response propensity to Strategy 3 is 0.3 (Extreme mode's high wind speeds easily generate noise and airflow discomfort, which does not meet the quiet requirements of conference scenarios). Office Area B's response propensity to Strategy 1 is 0.6 (Comfort mode consumes a relatively high amount of energy, and office scenarios require a balance between energy consumption and basic comfort); the response propensity to Strategy 2 is 0.7 (Energy-saving mode meets office energy consumption control targets and maintains basic comfort); and the response propensity to Strategy 3 is 0.4 (Extreme mode is not a necessity for office work, and its high energy consumption conflicts with daily control logic). Rest area C has a response tendency of 0.4 to strategy 1 (the energy consumption in comfort mode is higher than that in energy-saving mode, and the low load in rest scenarios does not require high comfort investment); the response tendency to strategy 2 is 0.6 (the energy-saving mode meets the low power consumption and moderate cooling requirements of rest scenarios); and the response tendency to strategy 3 is 0.2 (the high wind speed and noise in extreme speed mode will interfere with the rest experience).

[0115] The server then determines the first adaptation vector based on the first evaluation information. This involves directly mapping the response propensity values ​​to fitness values ​​(higher response propensity values ​​indicate higher fitness values) and arranging them according to the "controlled area - control policy to be activated" dimension. The specific values ​​are: Conference Room A - Policy 1: 0.8, A - Policy 2: 0.5, A - Policy 3: 0.3; Office Area B - Policy 1: 0.6, B - Policy 2: 0.7, B - Policy 3: 0.4; Rest Area C - Policy 1: 0.4, C - Policy 2: 0.6, C - Policy 3: 0.2. The server organizes these values ​​into a vector structure according to the dimensions, completing the construction of the first adaptation vector and providing initial "area-policy" fitness quantification data for subsequent input into the air conditioning intelligent control model.

[0116] By responding to the scenario-based evaluation and numerical mapping of the prediction model, the server accurately captures the "region-strategy" adaptation relationship in the current scenario, allowing the first adaptation vector to reflect real-time scenario requirements while inheriting the historical laws accumulated by model training, laying a solid data foundation for the precise scheduling of intelligent air-conditioning control.

[0117] In an embodiment of the present invention, the response prediction model is obtained through the following process, including:

[0118] Extracting features of discrete scenes corresponding to controlled area instances included in each training data to obtain corresponding area feature vectors, and extracting features of discrete scenes corresponding to control strategy instances included in each training data to obtain corresponding strategy feature vectors;

[0119] According to the regional feature vectors and continuous scene features corresponding to each controlled area instance and the strategy feature vectors and continuous scene features corresponding to each control strategy instance, the response tendency value corresponding to each training data is evaluated;

[0120] According to the enabled response labels and corresponding response tendency values ​​included in each training data, an error function corresponding to the response prediction model is set, and the response prediction model is tuned according to the error function.

[0121] In an embodiment of the present invention, illustratively, in the training process of the smart office building air conditioning response prediction model, the server sequentially performs the following steps:

[0122] First, discrete scene features are extracted to generate vectors. The training data comes from air conditioning control logs from the past year. These logs contain examples of controlled areas, such as Conference Room A (discrete scene label "Conference"), Office Area B (labeled "Office"), and Rest Area C (labeled "Rest"), as well as control strategy examples such as "Comfort Mode (labeled 'Comfort')", "Energy Saving Mode (labeled 'Energy Saving')," and "Extreme Speed ​​Mode (labeled 'Extreme Speed')". The server uses one-hot encoding to map "Conference," "Office," and "Rest" into discrete feature vectors for the regions (e.g., the vector [1,0,0] for Conference Room A), and "Comfort," "Energy Saving," and "Extreme Speed" into discrete feature vectors for the strategies (e.g., the vector [1,0,0] for Comfort Mode). This completes the initial construction of the regional and strategy feature vectors, transforming the discrete scene labels into numerical expressions that the model can compute.

[0123] Next, continuous features are integrated to assess response propensity. The server extracts continuous scene features from the controlled area instance: Conference Room A: 80 square meters, real-time temperature: 28°C, occupancy density: 0.1875 people / square meter (15 people divided by 80 square meters); Office Area B: 300 square meters, temperature: 26°C, occupancy density: 0.2667 people / square meter (80 people divided by 300 square meters); Rest Area C: 50 square meters, temperature: 27°C, occupancy density: 0.1 people / square meter (5 people divided by 50 square meters). Continuous scene features are also extracted for the control strategy instance: Comfort mode: target temperature: 25°C, medium wind speed (digitized wind speed coefficient: 0.5); Energy Saving mode: target temperature: 26°C, low wind speed (wind speed coefficient: 0.3); and Extreme Speed ​​mode: target temperature: 24°C, high wind speed (wind speed coefficient: 0.8). The server inputs the "regional discrete feature vector + continuous feature" and the "strategy discrete feature vector + continuous feature" into a multi-layer perceptron (MLP) model, and evaluates the matching degree through feature intersection and weight calculation. For example, Conference Room A (meeting scene + high temperature + high crowd density) and the comfort mode (rapid temperature averaging + medium wind speed) are highly consistent with the scenario requirements. The model outputs a response tendency value of 0.85 for this combination (historical data shows that this combination quickly reaches the temperature standard and has high employee satisfaction). Office Area B (office scene + normal temperature + medium crowd density) and the energy-saving mode (gentle temperature adjustment + low wind speed) balance energy consumption and comfort needs, outputting a response tendency value of 0.8. Rest Area C (rest scene + slightly high temperature + low crowd density) matches the leisure needs of the energy-saving mode (low power consumption temperature adjustment + low wind speed), outputting a response tendency value of 0.7, thus completing the response tendency quantification of all training data.

[0124] Finally, an error function was constructed and the model was tuned. The "enabled response labels" in the training data represent historical real-world feedback: the label for Conference Room A (Comfort Mode) is "Temperature reached standard in 15 minutes, satisfaction score 9 (out of 10)," corresponding to a quantitative value of 0.9; the label for Office Area B (Energy Saving Mode) is "Temperature reached standard in 20 minutes, satisfaction score 7," corresponding to a quantitative value of 0.7; and the label for Rest Area C (Energy Saving Mode) is "Temperature reached standard in 25 minutes, satisfaction score 8," corresponding to a quantitative value of 0.8. The server calculates the mean squared error (MSE) between the model's predicted response propensity (e.g., 0.85 for Conference Room A (Comfort Mode) vs. the true label of 0.9), using the mean squared error (MSE) as the error function. The Adam optimizer was used to iteratively adjust the MLP model's weight parameters, gradually reducing the error function value. The initial MSE was 0.05, and after 50 rounds of training, it dropped below 0.01. The model accurately learned the mapping between "regional scenario characteristics, policy control characteristics, and response effects," becoming a pre-trained model supporting real-time "region-policy" response propensity assessment.

[0125] Through full-dimensional extraction of discrete and continuous features, quantitative evaluation of multi-scenario matching, and model tuning based on real feedback, the response prediction model trained by the server can accurately capture the "region-strategy" response logic in the air-conditioning control scenario, providing reliable evaluation capabilities for real-time generation of the first adaptation vector.

[0126] In the embodiment of the present invention, determining the first adaptation vector according to the first evaluation information may be implemented through the following examples.

[0127] Obtaining second evaluation information based on the first evaluation information and a preset environmental adjustment factor, wherein the second evaluation information is used to indicate a fitness value of the plurality of controlled areas to each of the control strategies to be activated;

[0128] Dividing the fitness value threshold into segments according to the second evaluation information to obtain a plurality of fitness value threshold segments, each fitness value threshold segment corresponding to an fitness threshold level;

[0129] Vector reconstruction is performed according to the multiple adaptation degree value threshold segments to obtain the first adaptation vector.

[0130] In an exemplary embodiment of the present invention, in a smart office building air conditioning intelligent control scenario, when the server executes the "determining the first adaptation vector based on the first evaluation information" process, it first generates second evaluation information in combination with environmental adjustment factors. The first evaluation information is the "area-strategy" response tendency output by the response prediction model (e.g., Conference Room A's response tendency to Strategy 1 is 0.8, Strategy 2 is 0.5, and Strategy 3 is 0.3; Office Area B's response tendency to Strategy 1 is 0.6, Strategy 2 is 0.7, and Strategy 3 is 0.4; Rest Area C's response tendency to Strategy 1 is 0.4, Strategy 2 is 0.6, and Strategy 3 is 0.2). The server retrieves real-time environmental parameters as environmental adjustment factors: On a summer day, the outdoor temperature is 35°C (high temperature) and the humidity is 80% (high humidity). These are quantified as a "high temperature enhancement coefficient of 1.2" (increasing the adaptability of rapid cooling strategies), a "high humidity attenuation coefficient of 0.8" (reducing the adaptability of strong airflow strategies), and an "energy-saving attenuation coefficient of 0.9" (increasing the cooling efficiency of energy-saving mode at high temperatures, requiring a moderate adjustment of the adaptability). Weighted correction of the response tendency of each "area-strategy" combination: the adaptability value of conference room A-strategy 1 (comfort mode, rapid temperature equalization meets the needs of high temperature scenes) is (0.8×1.2=0.96); A-strategy 2 (energy-saving mode, the cooling rate at high temperature cannot meet the timeliness requirements of the meeting) is (0.5×0.9=0.45); A-strategy 3 (extreme speed mode, strong airflow in high humidity can easily cause physical discomfort) is (0. The values ​​for B-Strategy 1 in the office area are (0.6 × 1.2 = 0.72), B-Strategy 2 is (0.7 × 0.9 = 0.63), and B-Strategy 3 is (0.4 × 0.8 = 0.32). The values ​​for C-Strategy 1 in the rest area are (0.4 × 1.2 = 0.48), C-Strategy 2 is (0.6 × 0.9 = 0.54), and C-Strategy 3 is (0.2 × 0.8 = 0.16). These revised values ​​constitute the second evaluation information, which directly indicates the level of suitability of each "area-strategy" combination.

[0131] Next, the server divides the fitness threshold into segments. The preset rule is: a fitness value (≥0.8) is defined as "high fitness" (corresponding to level 3), (0.6leq fitness <0.8) is defined as "medium fitness" (corresponding to level 2), and (fitness <0.6) is defined as "low fitness" (corresponding to level 1). Traversing the second evaluation information values: Conference Room A - Strategy 1 (0.96) is high fitness, A-Strategy 2 (0.45) and A-Strategy 3 (0.24) are low fitness; Office Area B - Strategy 1 (0.72) and B-Strategy 2 (0.63) are medium fitness, and B-Strategy 3 (0.32) is low fitness; Rest Area C - Strategy 1 (0.48), C-Strategy 2 (0.54), and C-Strategy 3 (0.16) are all low fitness.

[0132] Finally, the server vector is reconstructed to generate the first adaptation vector. Each segment's adaptation level is mapped to a numerical value (high adaptation = 3, medium adaptation = 2, low adaptation = 1) and arranged in order according to the "controlled area - control strategy to be activated" dimension: Conference Room A corresponds to strategies 1 to 3, strategy 2 to 1, and strategy 3 to 1; Office Area B corresponds to strategies 1 to 2, strategy 2 to 2, and strategy 3 to 1; and Rest Area C corresponds to strategies 1 to 1, strategy 2 to 1, and strategy 3 to 1. This forms the first adaptation vector ([3, 1, 1, 2, 2, 1, 1, 1, 1]), completing the conversion from response propensity to a quantitative adaptation vector and providing structured data support for subsequent input into the air conditioning intelligent control model.

[0133] Through dynamic correction of environmental factors, scenario-based division of threshold segments, and vector reconstruction, the server ensures that the first adaptation vector not only reflects real-time environmental requirements, but also simplifies the complexity of model input through level quantification, thereby achieving accurate digital expression of the "region-strategy" adaptation relationship.

[0134] The embodiment of the present invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned air conditioning intelligent control method based on artificial intelligence. Figure 2 As shown, Figure 2 This is a block diagram of the structure of a computer device 100 provided in an embodiment of the present invention. Computer device 100 includes a memory 111, a processor 112, and a communication unit 113. To enable data transmission or exchange, memory 111, processor 112, and communication unit 113 are electrically connected to each other, directly or indirectly. For example, these components can be electrically connected via one or more communication buses or signal lines.

[0135] For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Numerous modifications and variations are possible in light of the above teachings. These embodiments have been selected and described in order to best illustrate the principles of the present disclosure and its practical application, thereby enabling those skilled in the art to best utilize the present disclosure and to utilize various embodiments with various modifications as appropriate for the specific application contemplated.

Claims

1. An air conditioning intelligent control method based on artificial intelligence, characterized in that: include: Evaluate the regional characteristics of the plurality of controlled areas and the strategy characteristics of the plurality of control strategies to be activated according to the response prediction model to obtain first evaluation information, where the first evaluation information is used to indicate a response tendency of the plurality of controlled areas to activate a response to each of the control strategies to be activated; Determining a first adaptation vector according to the first evaluation information, wherein each eigenvalue in the first adaptation vector represents a degree of adaptation of a controlled area to a control strategy to be enabled; The first adaptation vector is loaded into a pre-trained air conditioning intelligent control model to obtain a second adaptation vector output by the air conditioning intelligent control model; the air conditioning intelligent control model is used to evaluate the fitness value of each controlled area to the control strategy to be activated according to the working condition association method within the constraint framework of the target strategy priority tensor table, where each eigenvalue in the target strategy priority tensor table represents the priority coefficient of the corresponding control strategy to be activated; activating at least one to-be-activated control strategy to each of the plurality of controlled areas according to the second adaptation vector; Before loading the first adaptation vector into the air-conditioning intelligent control model to obtain a second adaptation vector output by the air-conditioning intelligent control model, the method further includes: Determine an original strategy priority tensor table based on the activation counts corresponding to each control strategy instance included in each training data, where each activation count includes the number of times at least one controlled region instance executes the corresponding control strategy instance; Determining a third adaptation vector based on the regional features of the controlled region instances and the strategy features of the control strategy instances respectively included in the respective training data, wherein each feature value in the third adaptation vector represents a fitness value of a controlled region instance to a control strategy instance; Performing a policy scheduling benchmark configuration on the air-conditioning intelligent control model according to the original policy priority tensor table and the third adaptation vector; Obtaining a fourth adaptation vector based on the air conditioning intelligent control model configured according to the policy scheduling benchmark, wherein an eigenvalue in the fourth adaptation vector represents a constraint framework of a priority coefficient currently adopted in a corresponding control strategy instance, and a fitness value of a controlled area instance to the control strategy instance; setting an error function corresponding to the air-conditioning intelligent control model according to an expected deviation of the degree of fitness between the third fitness vector and the fourth fitness vector; According to the error function, the air conditioning intelligent control model is tuned using the Adam optimizer to obtain the pre-trained air conditioning intelligent control model.

2. The method according to claim 1, characterized in that According to the number of times each control strategy instance included in each training data is enabled, the original strategy priority tensor table is determined, including: performing an aggregation operation on the control strategy instances included in each training data according to the number of activations corresponding to each control strategy instance included in the training data, to obtain a plurality of control strategy clusters, each control strategy cluster including at least one control strategy instance; sorting the plurality of control policy clusters by priority according to activation count values ​​corresponding to the plurality of control policy clusters, wherein each activation count value is determined according to the activation count of the control policy instances included in the corresponding control policy cluster; According to the priority layer sorting, the priority coefficients of the plurality of control strategy clusters adopted are determined, and according to the priority coefficients of the plurality of control strategy clusters adopted, the original strategy priority tensor table is determined.

3. The method according to claim 1, characterized in that The parameters of the air conditioning intelligent control model include a target policy priority tensor table, a regional adaptation vector, and a policy adaptation vector; then, according to the original policy priority tensor table and the third adaptation vector, a policy scheduling benchmark configuration is performed on the air conditioning intelligent control model, including: Configuring the target policy priority tensor table based on the policy scheduling benchmark of the original policy priority tensor table; Performing vector analysis on the third adaptation vector to obtain an original region feature vector and an original strategy feature vector; configuring the regional adaptation vector based on the original regional feature vector policy scheduling benchmark, and configuring the policy adaptation vector based on the original policy feature vector policy scheduling benchmark; According to the air conditioning intelligent control model configured according to the policy scheduling benchmark, the fourth adaptation vector is obtained, including: According to the regional adaptation vector, the strategy adaptation vector and the target strategy priority tensor table, the adaptation value of each controlled region instance to each control strategy instance is determined to obtain the fourth adaptation vector.

4. The method according to claim 3, characterized in that Determining the fitness value of each controlled area instance to each control strategy instance according to the area fitness vector, the strategy fitness vector, and the target strategy priority tensor table, including: Determining a comprehensive adaptation coefficient according to all adaptation degree values ​​included in the third adaptation vector; Determining a region adaptation coefficient corresponding to each controlled region instance according to at least one adaptation degree value corresponding to each controlled region instance in the third adaptation vector; Determining a strategy adaptation coefficient corresponding to each control strategy instance according to at least one adaptation degree value corresponding to each control strategy instance in the third adaptation vector; According to the regional adaptation vector, the strategy adaptation vector, the target strategy priority tensor table, the comprehensive adaptation coefficient, the regional adaptation coefficient and the strategy adaptation coefficient, the adaptation value of each controlled area instance to each control strategy instance is determined to obtain the fourth adaptation vector.

5. The method according to claim 1, wherein The response prediction model is obtained through the following process, including: Extracting features of discrete scenes corresponding to controlled area instances included in each training data to obtain corresponding area feature vectors, and extracting features of discrete scenes corresponding to control strategy instances included in each training data to obtain corresponding strategy feature vectors; According to the regional feature vectors and continuous scene features corresponding to each controlled area instance and the strategy feature vectors and continuous scene features corresponding to each control strategy instance, the response tendency value corresponding to each training data is evaluated; According to the enabled response labels and corresponding response tendency values ​​included in each training data, an error function corresponding to the response prediction model is set, and the response prediction model is tuned according to the error function.

6. The method according to claim 1, characterized in that Determining the first adaptation vector according to the first evaluation information includes: Obtaining second evaluation information based on the first evaluation information and a preset environmental adjustment factor, wherein the second evaluation information is used to indicate a fitness value of the plurality of controlled areas to each of the control strategies to be activated; Dividing the fitness value threshold into segments according to the second evaluation information to obtain a plurality of fitness value threshold segments, each fitness value threshold segment corresponding to an fitness threshold level; Vector reconstruction is performed according to the multiple adaptation degree value threshold segments to obtain the first adaptation vector.

7. A server system, characterized in that: The method comprises a server, wherein the server is configured to execute the method according to any one of claims 1 to 6.

Images