A single-body air conditioner load adjustment potential evaluation method based on transfer learning and multi-task learning

Abstract

The application discloses a kind of monomer air conditioner load regulation potential evaluation methods based on transfer learning and multi-task learning, it is related to air conditioner management technical field, the method includes: based on the air conditioner operation data under general scene constructs regional association model, and the regional association model is analyzed, the collaborative regulation relationship between air conditioner equipment is obtained;Utilize deep residual neural network to construct air conditioner load regulation potential evaluation model;According to the collaborative regulation relationship between air conditioner equipment, formulate multi-task learning framework, and utilize multi-task learning framework to optimize air conditioner load regulation potential evaluation model under general scene;Combined with transfer learning technology, the weight of optimized air conditioner load regulation potential evaluation model is adjusted, and the air conditioner load regulation potential evaluation model under specific scene is obtained.The application is by constructing regional association model and introducing self-supervised learning and counter training, significantly improve the adaptability and generalization ability of air conditioner load regulation potential evaluation model.

Description

Technical Field

[0001] This invention relates to the field of air conditioning management technology, and more specifically, to a method for evaluating the load adjustment potential of a single air conditioner based on transfer learning and multi-task learning. Background Technology

[0002] Demand response, as a load management strategy, typically guides electricity users to adjust their electricity consumption behavior during peak or off-peak periods of grid demand through incentives, price signals, or direct control. This helps grid operators balance supply and demand more efficiently, improving grid stability and economics. Demand-side response not only helps reduce reliance on traditional power generation resources but also promotes the integration of renewable energy and provides economic compensation or incentives to participating users.

[0003] Air conditioning load, as a high-quality flexible load resource, plays a crucial role in demand-side response. During peak electricity consumption periods, adjusting air conditioning load can significantly reduce the load pressure on the power grid. Air conditioning systems possess strong control flexibility; by adjusting parameters such as temperature setpoint and compressor frequency, they can achieve rapid response to electricity demand. Simultaneously, the heat accumulation effect of temperature allows air conditioning systems to fine-tune load within a short period, minimizing the impact on user comfort and keeping indoor temperature changes within an acceptable range.

[0004] Currently, data-driven methods utilize big data analytics and machine learning techniques to extract patterns in air conditioning load regulation from historical operational data and identify load regulation potential. However, existing technologies have the following limitations:

[0005] 1) High cost of obtaining sample data: In specific scenarios (such as newly built buildings or special environments), it is difficult to obtain a large number of labeled samples, which makes model training difficult.

[0006] 2) Insufficient model generalization ability: Due to the diversity of building environment and user habits, the existing model is difficult to adapt to complex real-world scenarios, and the evaluation results have poor stability in different environments.

[0007] 3) User comfort is not fully considered: Existing methods often only focus on load adjustment potential and fail to take user experience into account, which may lead to reduced user acceptance and affect the demand response effect.

[0008] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0009] To address the problems in related technologies, this invention proposes a method for evaluating the load regulation potential of a single air conditioner based on transfer learning and multi-task learning, in order to overcome the aforementioned technical problems existing in the current related technologies.

[0010] Therefore, the specific technical solution adopted by the present invention is as follows:

[0011] A method for assessing the load regulation potential of a single air conditioner unit based on transfer learning and multi-task learning, comprising:

[0012] S1. Construct a regional correlation model based on air conditioning operation data under general scenarios, and analyze the regional correlation model to obtain the coordinated adjustment relationship between air conditioning equipment;

[0013] S2. Construct an air conditioning load regulation potential assessment model using a deep residual neural network, and train the air conditioning load regulation potential assessment model based on a predefined training dataset to obtain an air conditioning load regulation potential assessment model for general scenarios.

[0014] S3. Develop a multi-task learning framework based on the coordinated adjustment relationship between air conditioning equipment, and use the multi-task learning framework to optimize the air conditioning load adjustment potential assessment model in general scenarios.

[0015] S4. Based on sample data of air conditioning adjustment potential under specific scenarios, and combined with transfer learning technology to adjust and optimize the weights of the air conditioning load adjustment potential assessment model, an air conditioning load adjustment potential assessment model under specific scenarios is obtained to assess the air conditioning load adjustment potential under different scenarios.

[0016] Preferably, the air conditioning operation data in general scenarios includes air conditioning equipment data, building environment data, external environment data, user behavior data, and air conditioning control data;

[0017] The data for air conditioning equipment includes production date, installation date, refrigerant type, filter status, and equipment energy efficiency.

[0018] Building environment data includes thermal insulation performance, structural layout, usage functions, and construction year;

[0019] External environmental data include outdoor temperature, indoor humidity, and solar radiation intensity;

[0020] User behavior data includes users' indoor temperature settings, adjustment habits, and usage frequency;

[0021] Air conditioning control data includes power adjustment deviation, load adjustment rate, standard adjustment rate, and adjustment accuracy response time.

[0022] Preferably, a deep residual neural network is used to construct an air conditioning load regulation potential assessment model, and the model is trained based on a predefined training dataset to obtain an air conditioning load regulation potential assessment model for general scenarios, including:

[0023] S21. Preprocess the air conditioner operation data, and sequentially perform feature selection, feature construction and feature transformation on the preprocessed air conditioner operation data to obtain initial training features;

[0024] S22. Use principal component analysis dimensionality reduction to identify key features from the initial training features that represent the trend of the training dataset, and construct the training dataset based on the key features;

[0025] S23. Construct an air conditioning load regulation potential assessment model based on a deep residual neural network, and train the air conditioning load regulation potential assessment model using a training dataset to obtain an air conditioning load regulation potential assessment model under a general scenario.

[0026] Preferably, the process of using principal component analysis (PCA) dimensionality reduction to identify key features representing trends in the training dataset from the initial training features, and constructing the training dataset based on these key features, includes:

[0027] S221. Transform the feature values ​​of the initial training features to a preset range, and calculate the Pearson correlation coefficient between the initial training features to obtain the linear correlation between the initial training features;

[0028] S222. Determine the degree of nonlinear relationship between the initial training features based on linear correlation. If the degree of nonlinear relationship between the initial training features is high, calculate the Spearman rank correlation coefficient between the initial training features to replace the Pearson correlation coefficient. Otherwise, proceed to step S223.

[0029] S223. Calculate the correlation coefficient matrix between the initial training features, and perform eigenvalue decomposition on the correlation coefficient matrix to obtain the eigenvalue decomposition results;

[0030] S224. Calculate the principal component cumulative contribution rate of the initial training features based on the eigenvalue decomposition results, select key features based on the principal component cumulative contribution rate, and integrate the key features to obtain the training dataset.

[0031] Preferably, the cumulative contribution rate of the principal components of the initial training features is calculated based on the eigenvalue decomposition results. Key features are then selected based on the cumulative contribution rate of the principal components, and the key features are integrated to obtain the training dataset, which includes:

[0032] S2241. Obtain the eigenvectors corresponding to the eigenvalue decomposition results, sort the eigenvectors, and filter out the eigenvectors within a preset range.

[0033] S2242. Calculate the proportion of the selected feature vectors to the total sum of feature vectors, use the calculation result as the principal component contribution rate, and sum the principal component contribution rates to obtain the principal component cumulative contribution rate.

[0034] S2243. Compare the cumulative contribution rate of the principal components with a preset threshold. If the cumulative contribution rate of the principal components is greater than the preset threshold, the feature vector corresponding to the cumulative contribution rate of the principal components is transformed to a new feature space to obtain a key feature set, and the key features are integrated to obtain a training dataset. Otherwise, after adjusting the preset range, step S2241 is executed again.

[0035] Preferably, an air conditioning load regulation potential assessment model is constructed based on a deep residual neural network, and the model is trained using a training dataset to obtain an air conditioning load regulation potential assessment model for general scenarios, including:

[0036] S231. Construct multi-level residual modules in a deep residual neural network, introduce a dense connection mechanism in each residual module, and pass the output of the front layer to the back layer to improve the feature transfer efficiency of the air conditioning load regulation potential assessment model.

[0037] S232. Channel attention mechanism and spatial attention mechanism are introduced into the deep residual neural network to improve the ability of the air conditioning load regulation potential assessment model to pay attention to key features.

[0038] S233. A max pooling layer is introduced into the output of the residual module. The key features are processed sequentially through the residual module, the max pooling layer and the fully connected layer. The processing results are then passed to the output layer for output.

[0039] S234. Input the training dataset into a deep residual neural network for training to obtain a general scenario air conditioning load regulation potential assessment model.

[0040] Preferably, channel attention mechanism and spatial attention mechanism are introduced into the deep residual neural network to improve the ability of the air conditioning load regulation potential assessment model to focus on key features, including:

[0041] S2321. Calculate the global features of each channel in the deep residual neural network using global average pooling, and assign weights to each channel based on the global features;

[0042] S2322. Use convolution operations to generate spatial weights in a deep residual neural network, enabling the deep residual neural network to focus on key features at different locations.

[0043] Preferably, the multi-task learning framework developed based on the coordinated adjustment relationship between air conditioning devices includes:

[0044] S31. Based on the coordinated adjustment relationship between air conditioning equipment, generate several sub-tasks and establish hard shared constraints to capture the common characteristics of each sub-task;

[0045] S32. Introduce soft shared constraints for each subtask-specific layer, calculate the parameter distance between subtask-specific layers, compare the similarity of parameter distances between subtask-specific layers, and generate task loss based on the similarity results.

[0046] S33. Introduce a regularization term into the soft sharing constraint to obtain the soft sharing regularization loss, and combine the task loss and the soft sharing regularization loss to construct a multi-task joint learning framework.

[0047] S34. Based on the predefined learning difficulty of the sub-tasks, dynamically adjust the loss weights of hard shared constraints and soft shared constraints in the multi-task joint learning framework.

[0048] Preferably, based on sample data of air conditioning regulation potential under specific scenarios, and combined with transfer learning techniques to adjust and optimize the weights of the air conditioning load regulation potential assessment model, an air conditioning load regulation potential assessment model for specific scenarios is obtained to assess the air conditioning load regulation potential under different scenarios, including:

[0049] S41. Use the field testing method to test the adjustment potential of air conditioning equipment in a specific scenario, and obtain sample data of the air conditioning adjustment potential in the specific scenario.

[0050] S42. Using transfer learning technology, the model weights of the optimized air conditioning load regulation potential assessment model are transferred to the target scenario to obtain an air conditioning load regulation potential assessment model for a specific scenario, so as to evaluate the air conditioning load regulation potential under different scenarios.

[0051] Preferably, the air conditioning equipment under specific scenarios is tested using on-site testing methods to obtain sample data on the air conditioning regulation potential under specific scenarios, including:

[0052] S411. Determine whether the air conditioning equipment is in the operating state under a specific scenario; if the air conditioning equipment is not turned on, it means there is no adjustment potential and the process ends; if the air conditioning equipment is turned on, enter the adjustment test stage and execute step S412.

[0053] S412. Record the air conditioner's operating status before adjustment, adjust the air conditioner's power adjustment range and temperature setting range, and when the air conditioner's power and temperature tend to stabilize, record the air conditioner's operating status after adjustment, and base the difference in air conditioner power before and after adjustment on the air conditioner's operating status before and after adjustment.

[0054] S413. Calculate the air conditioning load adjustment potential based on the air conditioning power difference, and obtain sample data of air conditioning adjustment potential under specific scenarios based on the air conditioning load adjustment potential.

[0055] Preferably, the model weights of the optimized air conditioning load regulation potential assessment model are transferred to the target scenario using transfer learning technology, resulting in an air conditioning load regulation potential assessment model for a specific scenario, including:

[0056] S421. The source domain feature selection technique and the weight adaptive adjustment technique are used to optimize the air conditioning regulation potential sample data respectively, and pseudo-labels are generated by self-supervised learning to expand the feature expression ability of the optimized air conditioning regulation potential sample data.

[0057] S422. The structure of the air conditioning load regulation potential assessment model is updated and optimized using a progressive training strategy and a meta-learning strategy, and trained based on the air conditioning regulation potential sample data to obtain an air conditioning load regulation potential assessment model under a specific scenario.

[0058] S423. Introduce adversarial training technology into the air conditioning load adjustment potential assessment model under specific scenarios, learn domain-invariant features through a domain discriminator, and optimize the performance of the air conditioning load adjustment potential assessment model under specific scenarios based on the domain-invariant features.

[0059] Preferably, the formula for calculating the Pearson correlation coefficient is as follows:

[0060]

[0061] In the formula, r XY X represents the Pearson correlation coefficient between the initial training features; i Y represents the observed values ​​of the initial training feature X; i represents the observed values ​​of the initial training feature Y; i represents the number of initial training features; This represents the mean of the initial training feature X; This represents the mean of the initial training feature Y.

[0062] Preferably, the formula for calculating the Spearman rank correlation coefficient is as follows:

[0063]

[0064] In the formula, r s d represents the Spearman rank correlation coefficient; i This represents the difference in rank between the two initial training features in the i-th sample; n represents the number of samples.

[0065] The beneficial effects of this invention are as follows:

[0066] 1. The single air conditioning load regulation potential assessment method based on transfer learning and multi-task learning provided by this invention significantly improves the adaptability and generalization ability of the air conditioning load regulation potential assessment model by constructing a regional correlation model and introducing self-supervised learning and adversarial training. This effectively solves the problems of high sample acquisition cost and insufficient model generalization ability in the prior art, and thus realizes accurate assessment and prediction of air conditioning load regulation potential in a variety of complex scenarios.

[0067] 2. This invention constructs an air conditioning load regulation potential assessment model for general scenarios based on a large amount of air conditioning operation data, and introduces technologies such as regional correlation modeling, multi-task learning, self-supervised learning and adversarial training, which significantly enhances the adaptability and prediction accuracy of the air conditioning load regulation potential assessment model for general scenarios. At the same time, the model is optimized by using transfer learning technology, taking into account the spatial correlation characteristics between equipment and user comfort, so as to achieve accurate load regulation potential assessment of air conditioning equipment in different scenarios. Attached Figure Description

[0068] 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.

[0069] Figure 1 This is a flowchart of a method for evaluating the load regulation potential of a single air conditioner based on transfer learning and multi-task learning according to an embodiment of the present invention;

[0070] Figure 2 This is a schematic diagram of the thermal comfort model in a method for evaluating the load adjustment potential of a single air conditioner based on transfer learning and multi-task learning according to an embodiment of the present invention.

[0071] Figure 3 This is a schematic diagram of the air conditioning load adjustment potential assessment model based on deep residual neural network in a single air conditioning load adjustment potential assessment method based on transfer learning and multi-task learning according to an embodiment of the present invention.

[0072] Figure 4 This is a flowchart illustrating the process of obtaining sample data of air conditioning regulation potential based on field testing in a single air conditioning load regulation potential assessment method based on transfer learning and multi-task learning according to an embodiment of the present invention.

[0073] Figure 5 This is a flowchart of the training process for deep residual network transfer learning in a method for evaluating the load regulation potential of a single air conditioner based on transfer learning and multi-task learning according to an embodiment of the present invention. Detailed Implementation

[0074] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0075] According to an embodiment of the present invention, a method for evaluating the load regulation potential of a single air conditioner based on transfer learning and multi-task learning is provided.

[0076] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figures 1-5 As shown, according to an embodiment of the present invention, a method for assessing the load adjustment potential of a single air conditioner based on transfer learning and multi-task learning includes:

[0077] S1. Construct a regional correlation model based on air conditioning operation data under general scenarios, and analyze the regional correlation model to obtain the coordinated adjustment relationship between air conditioning equipment.

[0078] It should be noted that air conditioning operation data in general scenarios includes air conditioning equipment data, building environment data, external environment data, user behavior data, and air conditioning control data;

[0079] The data includes: air conditioning equipment data such as production date, installation date, refrigerant type, filter status, and equipment energy efficiency; building environment data such as insulation performance, structural layout, usage function, and construction year; external environment data such as outdoor temperature, indoor humidity, and solar radiation intensity; user behavior data such as user settings, adjustment habits, and usage frequency for indoor temperature; and air conditioning control data such as power adjustment deviation, load adjustment rate, standard adjustment rate, and adjustment accuracy response time.

[0080] It should be noted that air conditioning load is affected by the performance of air conditioning equipment, the working environment, the building's insulation effect, and user habits; fluctuations in environmental factors such as outdoor temperature, humidity, and solar radiation will have a significant impact on air conditioning load; the insulation performance, structural layout, usage function, and age of different buildings will lead to significant differences in their thermal characteristics; users' setting and adjustment behavior of indoor temperature will also affect air conditioning load. Based on the above factors, a physical model of building air conditioning load is constructed, which includes:

[0081] 1) Thermodynamic model:

[0082] Considering the impact of heat exchange between the outside and inside temperatures on the air conditioner's power, a thermodynamic model is established to calculate the air conditioner's power requirement. The thermodynamic model expression is as follows:

[0083] Q cond =k×A×(T) out -T in );

[0084] In the formula, Q cond T represents the amount of heat conducted through heat transfer, k represents the thermal conductivity of the building material, A represents the external surface area of ​​the building, and T represents the amount of heat transferred through heat transfer. out and T in These represent the outdoor temperature and the indoor temperature, respectively.

[0085] 2) Building energy efficiency model:

[0086] Considering the impact of building heat capacity and thermal resistance on air conditioning power demand, a building energy efficiency model is established. The expression for the building energy efficiency model is as follows:

[0087]

[0088] In the formula, P building T represents the building's load power demand. in Indicates indoor temperature, T set S represents the set temperature, S represents the set wind speed, R represents the equivalent thermal resistance, and C represents the building's heat capacity.

[0089] 3) Air Conditioner Power Consumption Model:

[0090] A power consumption model for air conditioners is established to consider the impact of cooling and heating demands on air conditioner power consumption. The expression for the air conditioner power consumption model is as follows:

[0091]

[0092] In the formula, P t Q represents the power consumption at time t. t η represents the cooling (heating) capacity at time t. t Let t represent the energy efficiency ratio at time t, and E represent the total power consumption.

[0093] 4) Thermal comfort model:

[0094] Considering the thermal comfort of the human body under different temperature conditions, a thermal comfort model is established. Thermal comfort indices are commonly used to evaluate thermal comfort using predicted average thermal perception (PMV) and predicted percentage dissatisfaction (PPD). The thermal comfort model expression is as follows:

[0095] PMV = (0.303e -0.036M +0.028)L;

[0096] In the formula, PMV represents the predicted average thermal sensation value, and M represents the human metabolic rate (unit: W / m³). 2L represents the amount of heat exchanged between the human body and the environment.

[0097] The predicted dissatisfaction rate (PPD) is calculated based on the PMV value, using the following formula:

[0098]

[0099] In the formula, PPD represents the predicted dissatisfaction rate (the percentage of people who are dissatisfied with the environment), and PMV represents the predicted average thermal sensation value.

[0100] 5) User behavior model:

[0101] Analyze users' habits of setting and adjusting indoor temperature, as these user behaviors affect the trend of air conditioning load changes; establish a user behavior model to predict users' temperature setting preferences, and incorporate user comfort constraints to improve the accuracy of user behavior model predictions and user experience.

[0102] It should be noted that the analysis of users' indoor temperature setting and adjustment habits includes:

[0103] Step 1: Collect relevant data on temperature settings and adjustments, and extract key behavioral features.

[0104] Step 2: Analyze different user behavior patterns based on user temperature settings and operation frequency, and group these patterns using a clustering algorithm. The grouped user behavior patterns include:

[0105] Stable type: The set value fluctuates little and requires minimal adjustment.

[0106] Frequent adjustment type: Frequent operation and large changes in temperature settings.

[0107] Temperature-sensitive type: The set value closely follows changes in outdoor temperature.

[0108] Time-sensitive: There are significant differences in settings for different time periods.

[0109] Step 3: Automatically generate behavioral tags for users based on clustering results:

[0110] Fixed temperature type: Prefers a constant temperature (e.g., always set to 24℃).

[0111] High-temperature preference type: The average setting value is relatively high (such as 26℃ and above).

[0112] Low temperature preference type: The average set value is relatively low (such as 22°C and below).

[0113] Frequent adjustment type: The number of adjustment operations exceeds the set threshold (e.g., more than 5 times per day).

[0114] Further refine the labels using logical rules combined with clustering results:

[0115] Time-dependent: The temperature difference between night and day is set to exceed a certain value (e.g., more than 2°C).

[0116] Seasonally sensitive: Significant temperature differences between winter and summer (e.g., a difference of more than 3°C between set values).

[0117] Behavioral pattern extraction is based on decision tree parameters, and the decision tree is trained based on these parameters. Behavioral pattern rules are automatically extracted from the paths generated by the decision tree. The decision tree parameters include:

[0118] Maximum depth (max_depth): Limits the depth of the tree to avoid overfitting; Minimum number of sample splits (min_samples_split): Controls the number of samples in a leaf node. Use the target label (behavior type) as the classification objective, for example:

[0119] Low temperature stable type: average set temperature below 23.5℃; adjustment frequency less than 3.5 times / day.

[0120] Low-temperature frequent adjustment type: average set temperature below 23.5℃; adjustment frequency above 3.5 times / day.

[0121] High-temperature fixed type: The average set temperature is higher than 23.5℃; the temperature does not change over time.

[0122] Time-sensitive: The average set temperature is higher than 23.5℃; the set value is significantly adjusted according to the time period.

[0123] 6) Regional Association Model:

[0124] A graph neural network (GNN) is introduced to establish a regional correlation model, which captures the spatial coordinated adjustment relationship between different air conditioning equipment. By identifying the mutual influence between air conditioning equipment through the regional correlation model, the model's load adjustment potential prediction capability is improved.

[0125] It should be noted that the introduction of Graph Neural Networks (GNNs) to establish a regional association model and capture the spatial coordinated adjustment relationship between different air conditioning devices is specifically represented as a graph structure, capturing the spatial coordinated adjustment relationship, which includes:

[0126] Step 1: Extract data on the operating status of different air conditioning equipment, building spatial layout, user behavior data, and thermodynamic model input data, including equipment set temperature, actual temperature, operating power, indoor and outdoor environmental conditions (temperature, humidity, airflow), building heat exchange characteristics (such as heat transfer coefficient, spatial distance), and user behavior data (adjustment frequency, comfort feedback).

[0127] The physical relationships between building spaces and equipment are represented by an adjacency matrix, where nodes represent air conditioning equipment and edges represent spatial connections (with weights indicating heat exchange efficiency).

[0128] Step 2: Construct a graph using an adjacency matrix, where each node represents an air conditioning device or area, with features including the current set temperature, actual temperature, operating power, and user preference characteristics; the weights of the edges are determined by the heat exchange intensity calculated by the thermodynamic model, physical distance, and building energy efficiency characteristics, and a time-series dynamic graph is constructed to represent the device status and its relationships at different time steps.

[0129] Step 3: Using a graph convolutional network, node features and edge features (spatial association characteristics) are taken as input. Neighbor information is aggregated through graph convolutional layers to capture the spatial collaborative adjustment relationships between different air conditioning devices. The expression for aggregating neighbor information in the graph convolutional layer is as follows:

[0130]

[0131] In the formula, Let N(i) represent the features of the i-th node in the (k+1)-th convolutional layer of the graph; N(i) represent the neighbors of the i-th node. Represents the normalized weights; W represents the weight matrix of the graph convolutional layer; represents the feature of the j-th node in the k-th convolutional layer; b represents the bias vector.

[0132] Step 4: Based on the spatial coordinated adjustment relationship between different air conditioning devices, analyze the temporal dynamic relationship between the air conditioning devices, including:

[0133] The time series dynamic graph describes the current state of air conditioning equipment and spatial relationships. Recurrent neural networks (such as LSTM and GRU) are used to process the time dimension. Based on the processing results of the time dimension, the spatial coordination relationship is dynamically updated. At the same time, optimization objectives are defined, including spatial coordination (minimizing the temperature difference between adjacent equipment), thermal comfort (PMV index close to the target range), user behavior matching degree (setting temperature close to user preference), and energy consumption optimization, etc.

[0134] S2. Construct an air conditioning load regulation potential assessment model using a deep residual neural network, and train the air conditioning load regulation potential assessment model based on a predefined training dataset to obtain an air conditioning load regulation potential assessment model for general scenarios.

[0135] It should be noted that a training dataset for a general air conditioning load regulation potential assessment model is created using air conditioning operation data. This involves collecting operating condition data such as indoor temperature, outdoor temperature, indoor humidity, room size, temperature setpoint, air conditioning operating mode, and air conditioning fan speed during daily operation. Power data before and after adjusting air conditioning operating parameters is recorded, and the power difference is used to assess the air conditioning's power regulation potential. Furthermore, operational data during the control process is collected, including dispatched control commands, actual air conditioning load, control start time, control end time, and control deviation value. Based on this data, air conditioning regulation indicators are calculated, such as power regulation deviation, load regulation rate, standard regulation rate, regulation accuracy, and response time.

[0136] The formula for calculating the power regulation deviation is as follows:

[0137] ΔP = Pg - Pactual;

[0138] In the formula, ΔP represents the adjustment deviation value; Pg represents the target power value; and Pactual represents the actual measured air conditioning power value.

[0139] The formula for calculating the load regulation rate is:

[0140]

[0141] In the formula, S r P1 represents the power at the start of the adjustment; P2 represents the power at the end of the adjustment; t1 and t2 represent the start and end times of the adjustment, respectively.

[0142] The formula for calculating the standard adjustment rate is:

[0143]

[0144] In the formula, S g P represents the standard adjustment rate; g P1 represents the target power; P1 represents the power at the start of adjustment; T g Indicates the standard response time.

[0145] The formula for calculating adjustment accuracy is:

[0146]

[0147] In the formula, η represents the adjustment accuracy, ranging from 0 to 1; ΔP represents the adjustment deviation value; P g This indicates the target power value.

[0148] The formula for calculating response time is:

[0149] Tresponse = t2 - t1;

[0150] In the formula, Tresponse represents the response time of the air conditioning system; t1 represents the adjustment start time; and t2 represents the adjustment end time.

[0151] Specifically, a deep residual neural network is used to construct an air conditioning load regulation potential assessment model, and the model is trained based on a predefined training dataset to obtain a general-scenario air conditioning load regulation potential assessment model, including:

[0152] S21. Preprocess the air conditioning operation data, and then perform feature selection, feature construction and feature transformation on the preprocessed air conditioning operation data in sequence to obtain the initial training features.

[0153] It should be noted that the air conditioning operation data undergoes preprocessing to ensure the accuracy and consistency of the dataset, including:

[0154] Missing data augmentation: Missing data is augmented using linear interpolation.

[0155] Outlier handling: Identify, correct, or remove outliers.

[0156] Deduplication: Removes duplicate records to prevent data discrepancies.

[0157] Data standardization and normalization: Data is transformed using normalization methods, and the calculation formula is as follows:

[0158]

[0159] In the formula, x represents the normalized air conditioning operating data value; x represents the original air conditioning operating data value; x min and x max These represent the minimum and maximum values ​​of the data, respectively.

[0160] It should be noted that key features that significantly impact the model's predictive ability are extracted through feature selection, feature construction, and feature transformation. Principal component analysis (PCA) dimensionality reduction techniques are then used to identify and extract the key features that best represent the trends and structure of the dataset to construct the training dataset. Feature selection, feature construction, and feature transformation include:

[0161] Feature selection: Taking into account the physical model of the air conditioning load and the availability of data in the actual system, the following features are determined and selected: indoor temperature, indoor humidity, outdoor temperature, outdoor humidity, rated cooling power of the air conditioner, rated heating power of the air conditioner, building exterior area, building volume, temperature setpoint, building orientation, thermal insulation performance, window area ratio, start-up frequency, and usage period, etc.

[0162] Feature construction: Combining selected features to generate new derived features, including:

[0163] Daily average temperature and average temperature of the past two days: Calculate the average of the temperature data for the current day and the previous two days to smooth out short-term fluctuations and enhance the capture of temperature trends.

[0164] Temperature and humidity index: A comprehensive index calculated based on temperature and humidity data. The formula for calculating the temperature and humidity index expresses the impact of environmental comfort on air conditioning load. The formula is as follows:

[0165] Temperature and humidity index = T + 0.1RH;

[0166] In the formula, T represents temperature, reflecting the degree of hotness or coldness of the environment; RH represents relative humidity; and the weighting coefficient of 0.1 indicates that humidity has a relatively small impact on comfort.

[0167] Date Feature Tag: Generate binary feature tags for "statutory holidays", "working days" and "non-working days" to reflect load demand changes with the date.

[0168] Time feature construction: Time information is mapped to hours (0 to 23) or segments (morning, noon, evening) to capture load variation patterns in different time periods.

[0169] Feature transformation: Some features are transformed to improve the predictive ability and adaptability of the model. For example, building orientation is transformed into azimuth angle to quantify the impact of different orientations on the load; thermal insulation performance is quantified on a percentage scale, and the thermal insulation effect of different buildings is represented by the usage period and start time, which improves the interpretability of time features.

[0170] S22. Use principal component analysis dimensionality reduction to identify key features from the initial training features that represent the trend of the training dataset, and construct the training dataset based on the key features.

[0171] The process of using principal component analysis (PCA) dimensionality reduction to identify key features representing trends in the training dataset from the initial training features, and then constructing the training dataset based on these key features, includes:

[0172] S221. Transform the feature values ​​of the initial training features to a preset range, and calculate the Pearson correlation coefficient between the initial training features to obtain the linear correlation between the initial training features;

[0173] S222. Determine the degree of nonlinear relationship between the initial training features based on linear correlation. If the degree of nonlinear relationship between the initial training features is high, calculate the Spearman rank correlation coefficient between the initial training features to replace the Pearson correlation coefficient. Otherwise, proceed to step S223.

[0174] S223. Calculate the correlation coefficient matrix between the initial training features, and perform eigenvalue decomposition on the correlation coefficient matrix to obtain the eigenvalue decomposition results;

[0175] S224. Calculate the principal component cumulative contribution rate of the initial training features based on the eigenvalue decomposition results, select key features based on the principal component cumulative contribution rate, and integrate the key features to obtain the training dataset.

[0176] The process involves calculating the cumulative contribution rate of the principal components of the initial training features based on the eigenvalue decomposition results, selecting key features based on the cumulative contribution rate of the principal components, and integrating these key features to obtain the training dataset, which includes:

[0177] S2241. Obtain the eigenvectors corresponding to the eigenvalue decomposition results, sort the eigenvectors, and filter out the eigenvectors within a preset range.

[0178] S2242. Calculate the proportion of the selected feature vectors to the total sum of feature vectors, use the calculation result as the principal component contribution rate, and sum the principal component contribution rates to obtain the principal component cumulative contribution rate.

[0179] S2243. Compare the cumulative contribution rate of the principal components with a preset threshold. If the cumulative contribution rate of the principal components is greater than the preset threshold, the feature vector corresponding to the cumulative contribution rate of the principal components is transformed to a new feature space to obtain a key feature set, and the key features are integrated to obtain a training dataset. Otherwise, after adjusting the preset range, step S2241 is executed again.

[0180] It should be noted that the process of using principal component analysis (PCA) dimensionality reduction to identify key features representing trends in the training dataset from the initial training features, and constructing the training dataset based on these key features, includes:

[0181] Standardization: Convert all initial training feature values ​​to the range of 0 to 1 to eliminate the influence of different units on data comparison.

[0182] Calculate the correlation coefficient: Calculate the Pearson correlation coefficient between the initial training features to measure linear correlation. The formula for calculating the Pearson correlation coefficient is as follows:

[0183]

[0184] In the formula, r XY X represents the Pearson correlation coefficient between the initial training features; i Y represents the observed values ​​of the initial training feature X; i represents the observed values ​​of the initial training feature Y; i represents the number of initial training features; This represents the mean of the initial training feature X; This represents the mean of the initial training feature Y.

[0185] When data exhibits nonlinear relationships or has numerous outliers, the Spearman rank correlation coefficient can be used as an alternative method. The formula for calculating the Spearman rank correlation coefficient is as follows:

[0186]

[0187] In the formula, r s d represents the Spearman rank correlation coefficient; i This represents the difference in rank between the two initial training features in the i-th sample; n represents the number of samples.

[0188] The covariance between the initial training features is calculated to obtain the covariance matrix between the features. Then, eigenvalue decomposition is performed on this matrix. Based on the decomposed eigenvalues ​​and the cumulative contribution rate of the principal components, key features are selected, including:

[0189] Each element of the matrix represents the degree of linear correlation between two features. The matrix is ​​symmetric, and the main diagonal represents the variance of each feature.

[0190] The covariance matrix is ​​decomposed into eigenvalues ​​to obtain each eigenvalue and its corresponding eigenvector. The eigenvalues ​​reflect the degree of variation of the data in the direction of the corresponding eigenvector, and the eigenvectors indicate the direction of the principal components of the data.

[0191] The eigenvalues ​​are sorted from largest to smallest; the corresponding eigenvectors are also arranged in this order, indicating that the importance of the principal components is from high to low.

[0192] The proportion of each eigenvalue to the total number of eigenvalues ​​is called the principal component contribution rate. The principal component contribution rates are summed sequentially to obtain the cumulative principal component contribution rate, which is used to determine the proportion of data variation that the selected principal components can explain.

[0193] Based on the cumulative contribution rate, set a threshold (such as 85% or 95%), select the first few principal components required for the cumulative contribution rate to reach the threshold; view the feature vectors corresponding to these principal components, and extract the original features with larger weights in each principal component as key features.

[0194] S23. Construct an air conditioning load regulation potential assessment model based on a deep residual neural network, and train the air conditioning load regulation potential assessment model using a training dataset to obtain an air conditioning load regulation potential assessment model under a general scenario.

[0195] Among them, an air conditioning load regulation potential assessment model is constructed based on a deep residual neural network, and the model is trained using a training dataset to obtain an air conditioning load regulation potential assessment model for general scenarios, including:

[0196] S231. Construct multi-level residual modules in a deep residual neural network, introduce a dense connection mechanism in each residual module, and pass the output of the front layer to the back layer to improve the feature transfer efficiency of the air conditioning load regulation potential assessment model.

[0197] S232. Channel attention mechanism and spatial attention mechanism are introduced into the deep residual neural network to improve the ability of the air conditioning load regulation potential assessment model to pay attention to key features.

[0198] Among these measures, channel attention and spatial attention mechanisms are introduced into the deep residual neural network to enhance the ability of the air conditioning load regulation potential assessment model to focus on key features, including:

[0199] S2321. Calculate the global features of each channel in the deep residual neural network using global average pooling, and assign weights to each channel based on the global features;

[0200] S2322. Use convolution operations to generate spatial weights in a deep residual neural network, enabling the deep residual neural network to focus on key features at different locations.

[0201] S233. A max pooling layer is introduced into the output of the residual module. The key features are processed sequentially through the residual module, the max pooling layer and the fully connected layer. The processing results are then passed to the output layer for output.

[0202] S234. Input the training dataset into a deep residual neural network for training to obtain a general scenario air conditioning load regulation potential assessment model.

[0203] It should be noted that an improved deep residual neural network (DRN) model is used to construct the air conditioning load regulation potential assessment model. This model achieves efficient capture and analysis of complex load characteristics through multi-level residual modules, dense connection structure, and adaptive attention mechanism, specifically including:

[0204] ① Multi-level residual modules and dense connection structure:

[0205] Each residual module contains multi-scale convolutional layers (such as 3×3 and 5×5 convolutional kernels) to ensure detailed analysis of air conditioning load characteristics. The calculation formula for the multi-level residual module is as follows:

[0206] y = F(q, W) + q;

[0207] In the formula, y represents the output of the residual block; F(q, W) represents the mapping of the convolutional layer; W represents the weight of the convolutional kernel; and q is the input feature.

[0208] To fully utilize the feature information extracted from each layer, the model introduces dense connections in the residual module, directly passing the outputs of all preceding layers to subsequent layers. The expression for feature passing is as follows:

[0209]

[0210] In the formula, h l+1 represents the output of the (l+1)th layer, and f represents the activation function (ReLU).

[0211] ② Channel and spatial attention mechanisms:

[0212] Channel attention mechanism: Global average pooling (GAP) is used to calculate the global features of each channel, and different weights are assigned to each channel. The calculation formula is as follows:

[0213] w c =σ(GAP(F) c ));

[0214] In the formula, w c σ represents the channel weights; F represents the sigmoid activation function; c This is a channel feature map.

[0215] Spatial attention mechanism: Spatial weights are generated through convolution operations, enabling the model to focus on load features at different locations. The calculation formula is as follows:

[0216] w s =σ(Conv(F));

[0217] Among them, w s represents spatial weights; Conv represents the convolution operation; F represents the input feature map.

[0218] ③ Pooling layers and nonlinear activation functions:

[0219] A 2×2 max pooling layer with a stride of 2 is added after the improved residual block to reduce the spatial dimension and extract key information. All intermediate layers use the modified linear unit (ReLU) as the activation function, calculated as follows:

[0220] f(q) = max(0, q);

[0221] Where f(q) represents the output of the ReLU activation function; q represents the input feature.

[0222] ④ Fully connected layer and Softmax output:

[0223] After processing by residual blocks and pooling layers, the prediction results are passed through three fully connected layers (configured as FC64, FC32, and FC16) and then to the Softmax layer. The expression for the Softmax function is as follows:

[0224]

[0225] Where Softmax represents the Softmax function; z i Z represents the score of the i-th class; K represents the number of classes; z j This represents the score for the j-th class.

[0226] It should be noted that training and validation sets are constructed based on key features. The training data is then input into the improved deep residual neural network, and the output of the improved deep residual neural network is calculated, including:

[0227] 1) Loss calculation and backpropagation:

[0228] The error between the network output and the actual adjustment potential is calculated using the mean squared error (MSE) as the loss function, and the calculation formula is as follows:

[0229]

[0230] Where L represents the loss value; y represents the model's predicted value; i This represents the actual value; N represents the sample size.

[0231] 2) Gradient calculation and parameter update:

[0232] The gradient is calculated using the chain rule, and the formula is as follows:

[0233]

[0234] Based on the calculated gradient, select an appropriate learning rate α to update the weights:

[0235]

[0236] In the formula, W new The weights are updated after the learning rate; α represents the learning rate; and L represents the loss value.

[0237] 3) Model evaluation and validation:

[0238] Evaluate the model performance on the validation set, calculate the loss and accuracy, and adjust the learning rate and hyperparameters based on the validation results. Repeat the training process until the performance stabilizes or the predetermined number of iterations is reached.

[0239] S3. Develop a multi-task learning framework based on the coordinated adjustment relationship between air conditioning equipment, and use the multi-task learning framework to optimize the air conditioning load adjustment potential assessment model in general scenarios.

[0240] The multi-task learning framework developed based on the coordinated adjustment relationship between air conditioning equipment includes:

[0241] S31. Based on the coordinated adjustment relationship between air conditioning equipment, generate several sub-tasks and establish hard shared constraints to capture the common characteristics of each sub-task;

[0242] S32. Introduce soft shared constraints for each subtask-specific layer, calculate the parameter distance between subtask-specific layers, compare the similarity of parameter distances between subtask-specific layers, and generate task loss based on the similarity results.

[0243] S33. Introduce a regularization term into the soft sharing constraint to obtain the soft sharing regularization loss, and combine the task loss and the soft sharing regularization loss to construct a multi-task joint learning framework.

[0244] S34. Based on the predefined learning difficulty of the sub-tasks, dynamically adjust the loss weights of hard shared constraints and soft shared constraints in the multi-task joint learning framework.

[0245] It should be noted that the multi-task learning framework is developed based on the coordinated adjustment relationship between air conditioning equipment, and the optimization of the air conditioning load adjustment potential assessment model under general scenarios using the multi-task learning framework includes:

[0246] Step 1: Define subtasks and subtask divisions (using subtasks 1-4 as examples only):

[0247] Subtask 1: Load forecasting, predicting the operating load (power) of air conditioning equipment;

[0248] Subtask 2: Coordinate the modeling of joint relationships, capturing the spatial collaboration characteristics and thermal dynamics between devices;

[0249] Subtask 3: Thermal comfort analysis and evaluation, the impact of adjustment schemes on indoor thermal comfort;

[0250] Subtask 4: Energy consumption optimization, quantifying the overall system energy consumption trend.

[0251] Step 2: Design a multi-task learning framework:

[0252] Hard shared architecture: Construct a shared underlying layer to capture common characteristics of all subtasks, such as:

[0253] Basic status of air conditioning equipment (temperature, load).

[0254] Environmental characteristics (indoor and outdoor temperature and humidity).

[0255] Building heat exchange characteristics (spatial distance, heat transfer coefficient).

[0256] Hard sharing reduces redundant computations and improves model efficiency.

[0257] Soft shared architecture: Task-specific layers. Based on the shared layer, independent, task-specific layers are established for each subtask. For example:

[0258] Load forecasting module: forecasts equipment power.

[0259] Collaborative Adjustment Module: Models the collaborative relationships between devices.

[0260] Thermal comfort module: outputs PMV index.

[0261] Energy consumption optimization module: Calculates overall energy consumption.

[0262] Introduce soft shared constraints for specific layers of each subtask and add regularization terms to bring the specific layer parameters of the subtasks closer together.

[0263] Specifically, soft sharing constraints are introduced for specific layers of each subtask, and regularization terms are added to make the parameters of specific layers between subtasks closer to each other, including:

[0264] Step 1: Define soft sharing for specific layers of a subtask. For example, for specific layers of subtask A and subtask B, they may be two independent fully connected layers or convolutional layers, and each specific layer has an independent set of parameters.

[0265] Step 2: Design a method for sharing parameters: Introduce a global parameter center (such as an implicitly shared parameter set); implicitly align the specific layer parameters of each subtask with the parameters of other subtasks, so that the parameters are close but not completely the same.

[0266] Step 3: Add soft sharing constraints: Calculate the parameter distance between specific layers of each subtask and compare the similarity of parameters in specific layers of each subtask. For example: compare the parameters of specific layers of subtask A and subtask B, and use the distance between the parameters as a constraint to reduce this distance during training.

[0267] Step 4: Introduce a regularization term: Add a soft constraint to the loss function specifically designed to measure and reduce parameter differences between subtasks. For example, penalize instances where parameters in specific layers of each subtask deviate too far from their intended purpose. The weight of this regularization term should be controlled by hyperparameters to avoid excessive interference with the model's ability to personalize its subtasks.

[0268] Step 5: Optimize the training process: Introducing soft sharing constraints during the training process of multi-subtask learning requires the following steps:

[0269] Initialize the parameters of the shared layer and the subtask-specific layer to ensure that the parameter distribution of the specific layer is independent in the early stage of training.

[0270] The specific losses for each subtask are calculated separately, such as load forecasting losses and thermal comfort losses, and constraints are applied to specific layer parameters for each subtask, for example:

[0271] Compare the parameters of subtask A and subtask B to measure their similarity; if multiple subtasks are strongly correlated, specific layer parameters of multiple subtasks can be constrained simultaneously.

[0272] By combining subtask loss and soft-shared regularization loss, a joint optimization objective for multiple subtasks is constructed. During gradient updates, both subtask loss and soft-shared regularization loss are reduced simultaneously, gradually bringing the parameters of specific layers of subtasks closer to each other.

[0273] By adjusting the weight hyperparameter of the regularization term, the degree of sharing of specific layers among subtasks can be balanced. During the model tuning process, the regularization weights are gradually adjusted to find the best balance between sharing and subtask personalization.

[0274] Step 3: Utilize uncertainty weighting to dynamically adjust the loss weights between hard sharing and soft sharing based on the learning difficulty and importance of the sub-tasks.

[0275] It's important to note that hard sharing, by sharing the underlying feature extractor, distributes general information across all subtasks, making it suitable for scenarios where subtasks are highly correlated. Soft sharing, on the other hand, performs personalized modeling for each subtask at specific layers, while introducing constraints or regularization to promote collaborative learning among subtasks. Hard sharing provides a general foundation, reducing redundant computation and feature extraction, while soft sharing achieves targeted optimization at specific layers to meet the needs of multiple subtasks. Combining the two, by establishing a general feature representation at the shared layer and optimizing the independent objectives of each subtask at specific layers, achieves a balance between global efficiency and local accuracy.

[0276] S4. Based on sample data of air conditioning adjustment potential under specific scenarios, and combined with transfer learning technology to adjust and optimize the weights of the air conditioning load adjustment potential assessment model, an air conditioning load adjustment potential assessment model under specific scenarios is obtained to assess the air conditioning load adjustment potential under different scenarios.

[0277] Specifically, based on sample data of air conditioning load regulation potential under specific scenarios, and combined with the weights of the optimized air conditioning load regulation potential assessment model using transfer learning techniques, an air conditioning load regulation potential assessment model for specific scenarios is obtained to evaluate the air conditioning load regulation potential under different scenarios, including:

[0278] S41. Use the field testing method to test the adjustment potential of air conditioning equipment in a specific scenario, and obtain sample data of the air conditioning adjustment potential in the specific scenario.

[0279] Among them, the on-site testing method was used to test the regulation potential of air conditioning equipment in specific scenarios, and the sample data of air conditioning regulation potential in specific scenarios included:

[0280] S411. Determine whether the air conditioning equipment is in the operating state under a specific scenario; if the air conditioning equipment is not turned on, it means there is no adjustment potential and the process ends; if the air conditioning equipment is turned on, enter the adjustment test stage and execute step S412.

[0281] S412. Record the air conditioner's operating status before adjustment, adjust the air conditioner's power adjustment range and temperature setting range, and when the air conditioner's power and temperature tend to stabilize, record the air conditioner's operating status after adjustment, and base the difference in air conditioner power before and after adjustment on the air conditioner's operating status before and after adjustment.

[0282] S413. Calculate the air conditioning load adjustment potential based on the air conditioning power difference, and obtain sample data of air conditioning adjustment potential under specific scenarios based on the air conditioning load adjustment potential.

[0283] It should be noted that, as Figure 4 As shown, for air conditioning equipment in specific scenarios, a field testing method is used to test the regulation potential in order to obtain a small amount of regulation potential sample data and to obtain sample data for evaluating the air conditioning regulation potential. The specific process includes:

[0284] 1) Power-on status detection: Determine if the air conditioner is running. If the air conditioner is not turned on, there is no adjustment potential, and the process ends. If the air conditioner is turned on, proceed to the adjustment test stage, preparing to perform power adjustment operations.

[0285] 2) Determine the adjustment range. Based on factors such as the type of air conditioning equipment, current operating power, and temperature setting, determine a reasonable adjustment range to avoid over-adjustment or unattainable adjustment targets.

[0286] 3) Derive the set temperature: Based on historical data and empirical calculations, determine the temperature set range required to achieve the target power adjustment.

[0287] In cooling mode, the set temperature is usually lowered to reduce power consumption;

[0288] In heating mode, the set temperature is usually increased to increase power consumption.

[0289] 4) Collect and record data before adjustment. Before adjustment, collect and record the real-time operating data of the air conditioner, including key parameters such as current power, set temperature, ambient temperature and humidity.

[0290] 5) Perform setting adjustments: Based on the target power value, adjust the air conditioner's temperature setting and initiate the power adjustment process. Record temperature and power changes in real time throughout the adjustment process to ensure the continuity and integrity of data acquisition.

[0291] 6) After adjusting the set temperature, wait for the power and temperature to stabilize, then collect and record the real-time operating data of the air conditioner, including the new power value, set temperature and ambient temperature.

[0292] 7) Calculate the load adjustment potential. Based on the power difference during the adjustment process, calculate the load adjustment potential of the air conditioner at different set temperatures and record the actual power adjustment effect.

[0293] 8) Generate an evaluation report and save key data from each adjustment process, including set temperature, actual power, and adjustment time; generate an adjustment potential evaluation report based on data analysis and record the power adjustment effect of the air conditioner under different setting conditions.

[0294] The formula for calculating load regulation potential is as follows:

[0295]

[0296] In the formula, P potential This indicates the load regulation potential, which is the change in air conditioning load power per unit time. This indicator is used to quantify the regulation effect of air conditioning under different temperature setting conditions; P initial This indicates the initial power at the start of adjustment, i.e., the actual power value before adjusting the set temperature; P final This indicates the power value after adjustment, i.e., the stable power value after the air conditioner's set temperature is adjusted; T adjust This indicates the adjustment time, which is the time required from the start of adjustment until the power reaches a stable state at the new temperature setting, expressed in seconds.

[0297] S42. Using transfer learning technology, the model weights of the optimized air conditioning load regulation potential assessment model are transferred to the target scenario to obtain an air conditioning load regulation potential assessment model for a specific scenario, so as to evaluate the air conditioning load regulation potential under different scenarios.

[0298] Among them, the model weights of the optimized air conditioning load regulation potential assessment model are transferred to the target scenario using transfer learning technology, resulting in an air conditioning load regulation potential assessment model for a specific scenario, including:

[0299] S421. The source domain feature selection technique and the weight adaptive adjustment technique are used to optimize the air conditioning regulation potential sample data respectively, and pseudo-labels are generated by self-supervised learning to expand the feature expression ability of the optimized air conditioning regulation potential sample data.

[0300] It should be noted that source domain feature selection and weight adaptive adjustment techniques are used to optimize the air conditioning regulation potential sample data, and self-supervised learning is used to generate pseudo-labels, thereby expanding the feature representation capabilities of the optimized air conditioning regulation potential sample data, including:

[0301] For the source domain (historical air conditioning equipment data) and the target domain (current or new environment air conditioning regulation data), feature alignment maps the features of the source domain to the target domain, while retaining features that are significant in both domains, thus eliminating domain bias.

[0302] Weights are assigned to different features or samples in the sample data, with higher weights given to high-frequency and important features, and lower weights given to outliers or low-relevance samples; for sub-tasks in multi-task learning such as load prediction and thermal comfort analysis, the weight of each task in the total loss is dynamically adjusted, including:

[0303] By using instance-weighted methods, weights are assigned based on the importance of samples (such as error rate weighting in the Boosting algorithm), and the model is retrained after weight adjustment, so that the model focuses more on high-weight samples or features.

[0304] Unlabeled data (e.g., unlabeled load potential or thermal comfort data) is selected from the air conditioning regulation potential sample data, and an initial label structure is constructed using unsupervised methods (such as clustering or dimensionality reduction techniques). Based on the initial label structure, pseudo-labels are generated using self-supervised learning techniques, including:

[0305] Design a pretext task, such as predicting future values ​​of features, performing feature comparisons, or time series analysis.

[0306] By completing the pre-task training encoder, the embedding representation of unlabeled data is extracted. The encoder output is used to classify or regress the unlabeled data to generate pseudo-labels. The pseudo-labels are added to the training data, and the original labeled and pseudo-labeled samples are merged. The data scale is expanded through pseudo-labels, and the feature distribution is optimized.

[0307] S422. The structure of the air conditioning load regulation potential assessment model is updated and optimized using a progressive training strategy and a meta-learning strategy, and trained based on the air conditioning regulation potential sample data to obtain an air conditioning load regulation potential assessment model for a specific scenario.

[0308] It should be noted that meta-learning is a technique for learning general learning rules across multiple tasks, suitable for models that need to quickly adapt to new scenarios or tasks. The structure of the air conditioning load regulation potential assessment model updated and optimized using meta-learning strategies includes:

[0309] Step 1: Treat the sub-tasks of air conditioning load adjustment as meta-tasks (using meta-tasks 1-3 as an example only):

[0310] Meta-task 1: Single device load forecasting.

[0311] Meta-task 2: Coordinated adjustment of equipment within the region.

[0312] Meta-task 3: Thermal comfort optimization.

[0313] The model learns shared parameters across different meta-tasks. The shared layers serve as common parameters for meta-learning, while the specific layers for each meta-task serve as adaptation parameters for that meta-task.

[0314] Step 2: Learn the initial parameters of the shared layer so that it can quickly adapt to new tasks with a small number of samples. This includes:

[0315] Each meta-task is initialized based on the current shared layer parameters and performs gradient updates for the task-specific layer; the shared layer parameters are updated using the cumulative gradients of each meta-task: the task-specific layer parameters are fine-tuned with a small number of samples to quickly adapt to new requirements.

[0316] Step 3: Perform multiple gradient updates in each meta-task to obtain the optimized task parameters, and then roll back the optimized task parameters to the shared initial parameters:

[0317] Specifically, in the initial stage: a progressive strategy is used to train on simple tasks and simple samples to learn basic features and simple relationships.

[0318] Mid-term stage: Incorporate complex samples and complex tasks, optimize task-specific layers through multi-task learning, and gradually incorporate meta-learning strategies in mid-to-late-stage tasks to enable the shared layer to adapt across tasks.

[0319] Later stage: Utilize meta-learning strategies to optimize shared parameters, enabling the model to maintain high predictive performance in diverse scenarios.

[0320] S423. Introduce adversarial training technology into the air conditioning load adjustment potential assessment model under specific scenarios, learn domain-invariant features through a domain discriminator, and optimize the performance of the air conditioning load adjustment potential assessment model under specific scenarios based on the domain-invariant features.

[0321] It should be noted that the adversarial training technique, which incorporates the ideas of Generative Adversarial Networks (GANs), consists of the following parts:

[0322] Feature extractor: learns the feature representation of the raw data.

[0323] Domain discriminator: distinguishes the domain from which samples originate.

[0324] Task Predictor: A task for assessing the potential for air conditioning load regulation based on feature optimization.

[0325] During adversarial training, the feature extractor learns domain-invariant features by deceiving the domain discriminator, thereby improving the model's adaptability to the target scene.

[0326] By jointly optimizing the feature extractor, the domain discriminator, and the task predictor, the following objectives are achieved:

[0327] Feature extractors learn domain-invariant features.

[0328] The task predictor improves task performance in target scenarios based on domain-invariant features.

[0329] The domain discriminator correctly distinguishes the domain, forcing the features output by the feature extractor to be unable to distinguish the domain.

[0330] Specifically, Phase 1: Domain Discriminator Training: Fix the feature extractor parameters and train only the domain discriminator to accurately distinguish the domain.

[0331] Phase 2: Adversarial Optimization of Feature Extractor and Task Predictor: With fixed parameters for the domain discriminator, the feature extractor and task predictor are trained; the feature extractor learns domain-invariant features by adversarially challenging the domain discriminator; the task predictor optimizes the target task based on the domain-invariant features.

[0332] Phase 3: Alternating Training: Alternately optimize the neighborhood discriminator and feature extractor until the model converges.

[0333] It should be noted that, as Figure 5 As shown, the transfer learning method based on Deep Residual Networks (DRN) further optimizes the pre-trained deep convolutional network using a small amount of field test data, transferring some network structure and weights to the target scenario to construct an air conditioning load regulation potential assessment model for a specific scenario, including:

[0334] 1) Input a small amount of sample data and perform source domain optimization:

[0335] A small amount of sample data was obtained from the field testing method. Data utilization was optimized by source domain feature selection and adaptive weight adjustment. The input feature matrix is ​​as follows:

[0336] X selected = [x1ω1, x2ω2, ..., x n ω n};

[0337] In the formula, X selected x represents the selected input feature matrix. i Let w represent the i-th feature. i This represents the adaptive weight of the i-th feature.

[0338] 2) Load the general model and enhance the network structure:

[0339] Based on the general scenario-based air conditioning load regulation potential assessment model, multi-scale convolution and attention mechanisms are added to improve feature extraction capabilities. The residual module output is as follows:

[0340]

[0341] In the formula, F(x) represents the output characteristics of the residual module, σ represents the activation function, and W k This represents the kernel weights, i.e., kernels of different scales, b k For bias.

[0342] The formula for calculating channel attention weights is as follows:

[0343] w c =σ(GAP(X) c ));

[0344] In the formula, w c Xc represents the channel attention weight, and Xc represents the channel feature.

[0345] 3) Convolutional layer fine-tuning and self-supervised learning:

[0346] The fully connected layers are fixed, and the weights of the convolutional layers are fine-tuned. Pseudo-labels are generated through self-supervised learning to expand the dataset and enhance feature representation. Self-supervised learning tasks include time series prediction and missing data recovery, used to leverage unlabeled data to improve model performance.

[0347] 4) Fine-tuning of fully connected layers and multi-task learning:

[0348] Fixed convolutional layers, fine-tuned fully connected layers, added a multi-task learning module, and defined the total loss function:

[0349]

[0350] In the formula, L represents the total loss function, L i Let α represent the loss of the i-th task. i This represents the loss weight of the i-th task.

[0351] 5) Progressive training and meta-learning optimization:

[0352] A progressive training strategy is adopted, combined with a meta-learning strategy to quickly adapt to a small amount of sample data. The parameter update formula is as follows:

[0353]

[0354] In the formula, θ represents the current model parameters, θ′ represents the updated model parameters, and β is the learning rate of meta-learning. L represents the meta-learning loss meta The gradient with respect to the parameter θ.

[0355] 6) Generate the final adjustment potential assessment model:

[0356] Based on the optimized model structure, an assessment model for the air conditioning load regulation potential under specific scenarios is generated and validated.

[0357] In summary, by utilizing the above-mentioned technical solutions of this invention, the single air conditioning load regulation potential assessment method based on transfer learning and multi-task learning provided by this invention significantly improves the adaptability and generalization ability of the air conditioning load regulation potential assessment model by constructing a regional correlation model and introducing self-supervised learning and adversarial training. This effectively solves the problems of high sample acquisition cost and insufficient model generalization ability in the prior art, thereby achieving accurate assessment and prediction of air conditioning load regulation potential in various complex scenarios. This invention constructs an air conditioning load regulation potential assessment model for general scenarios based on a large amount of air conditioning operation data, and introduces regional correlation modeling, multi-task learning, self-supervised learning, and adversarial training techniques, which significantly enhances the adaptability and prediction accuracy of the air conditioning load regulation potential assessment model for general scenarios. At the same time, it uses transfer learning technology to optimize the model, comprehensively considering the spatial correlation characteristics between devices and user comfort, so as to achieve accurate load regulation potential assessment of air conditioning equipment in different scenarios.

[0358] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for evaluating the load regulation potential of a single air conditioner based on transfer learning and multi-task learning, characterized in that, include: S1. Construct a regional correlation model based on air conditioning operation data under general scenarios, and analyze the regional correlation model to obtain the coordinated adjustment relationship between air conditioning equipment; S2. Preprocess the air conditioner operation data, and then perform feature selection, feature construction and feature transformation on the preprocessed air conditioner operation data in sequence to obtain the initial training features; The feature values ​​of the initial training features are transformed to a preset range, and the Pearson correlation coefficient between the initial training features is calculated to obtain the linear correlation between the initial training features. The degree of nonlinear relationship between the initial training features is determined based on linear correlation. If the degree of nonlinear relationship between the initial training features is high, the Spearman rank correlation coefficient is calculated to replace the Pearson correlation coefficient. Otherwise, the correlation coefficient matrix between the initial training features is calculated, and the correlation coefficient matrix is ​​decomposed into eigenvalues ​​to obtain the eigenvalue decomposition results. Obtain the eigenvectors corresponding to the eigenvalue decomposition results, sort the eigenvectors, and filter out the eigenvectors within a preset range; Calculate the proportion of the selected feature vectors to the total number of feature vectors, use the calculation result as the principal component contribution rate, and sum the principal component contribution rates to obtain the principal component cumulative contribution rate. The cumulative contribution rate of the principal components is compared with a preset threshold. If the cumulative contribution rate of the principal components is greater than the preset threshold, the feature vector corresponding to the cumulative contribution rate of the principal components is transformed to a new feature space to obtain a key feature set, and the key features are integrated to obtain a training dataset. Otherwise, the feature vectors are re-selected after adjusting the preset range. An air conditioning load regulation potential assessment model is constructed based on a deep residual neural network, and the model is trained using a training dataset to obtain an air conditioning load regulation potential assessment model for general scenarios. S3. Develop a multi-task learning framework based on the coordinated adjustment relationship between air conditioning equipment, and use the multi-task learning framework to optimize the air conditioning load adjustment potential assessment model in general scenarios. S4. Based on sample data of air conditioning adjustment potential under specific scenarios, and combined with transfer learning technology to adjust and optimize the weights of the air conditioning load adjustment potential assessment model, an air conditioning load adjustment potential assessment model under specific scenarios is obtained to assess the air conditioning load adjustment potential under different scenarios. The multi-task learning framework developed based on the coordinated adjustment relationship between air conditioning devices includes: S31. Based on the coordinated adjustment relationship between air conditioning equipment, generate several sub-tasks and establish hard shared constraints to capture the common characteristics of each sub-task; S32. Introduce soft shared constraints for each subtask-specific layer, calculate the parameter distance between subtask-specific layers, compare the similarity of parameter distances between subtask-specific layers, and generate task loss based on the similarity results. S33. Introduce a regularization term into the soft sharing constraint to obtain the soft sharing regularization loss, and combine the task loss and the soft sharing regularization loss to construct a multi-task joint learning framework. S34. Based on the predefined learning difficulty of sub-tasks, dynamically adjust the loss weights of hard shared constraints and soft shared constraints in the multi-task joint learning framework. The air conditioning load regulation potential assessment model is obtained by combining sample data of air conditioning regulation potential under specific scenarios with the weights of the optimized air conditioning load regulation potential assessment model using transfer learning techniques. This model is used to assess the air conditioning load regulation potential under different scenarios, including: S41. Use the field testing method to test the adjustment potential of air conditioning equipment in a specific scenario, and obtain sample data of the air conditioning adjustment potential in the specific scenario. S42. Using transfer learning technology, the model weights of the optimized air conditioning load regulation potential assessment model are transferred to the target scenario to obtain an air conditioning load regulation potential assessment model for a specific scenario, so as to evaluate the air conditioning load regulation potential under different scenarios.

2. The method for evaluating the load regulation potential of a single air conditioner based on transfer learning and multi-task learning according to claim 1, characterized in that, The air conditioning operation data in the general scenario includes air conditioning equipment data, building environment data, external environment data, user behavior data, and air conditioning control data; The air conditioning equipment data includes production date, installation date, refrigerant type, filter status, and equipment energy efficiency. The building environment data includes thermal insulation performance, structural layout, usage functions, and construction year; The external environmental data includes outdoor temperature, indoor humidity, and solar radiation intensity; The user behavior data includes users' indoor temperature settings, adjustment habits, and usage frequency; The air conditioning control data includes power adjustment deviation, load adjustment rate, standard adjustment rate, and adjustment accuracy response time.

3. The method for evaluating the load regulation potential of a single air conditioner based on transfer learning and multi-task learning according to claim 1, characterized in that, The method involves constructing an air conditioning load regulation potential assessment model based on a deep residual neural network, and training the model using a training dataset to obtain a general-scenario air conditioning load regulation potential assessment model, including: Multi-level residual modules are constructed in a deep residual neural network, and a dense connection mechanism is introduced in each residual module to pass the output of the front layer to the back layer, so as to improve the feature transfer efficiency of the air conditioning load regulation potential assessment model. Channel attention mechanism and spatial attention mechanism are introduced into deep residual neural network to improve the ability of air conditioning load regulation potential assessment model to pay attention to key features; A max pooling layer is introduced into the output of the residual module. Key features are processed sequentially through the residual module, the max pooling layer, and the fully connected layer. The processing results are then passed to the output layer for output. The training dataset is input into a deep residual neural network for training, resulting in a model for evaluating the air conditioning load regulation potential under general scenarios.

4. The method for evaluating the load regulation potential of a single air conditioner based on transfer learning and multi-task learning according to claim 3, characterized in that, The method of introducing channel attention and spatial attention mechanisms into the deep residual neural network to enhance the ability of the air conditioning load regulation potential assessment model to focus on key features includes: Global average pooling is used to calculate the global features of each channel in the deep residual neural network, and weights are assigned to each channel based on the global features; Spatial weights in a deep residual neural network are generated using convolution operations, enabling the deep residual neural network to focus on key features at different locations.

5. The method for evaluating the load regulation potential of a single air conditioner based on transfer learning and multi-task learning according to claim 1, characterized in that, The method of using on-site testing to test the adjustment potential of air conditioning equipment in a specific scenario, and obtaining sample data of the air conditioning adjustment potential in that specific scenario, includes: S411. Determine whether the air conditioning equipment is in the operating state under a specific scenario; if the air conditioning equipment is not turned on, it means there is no adjustment potential and the process ends; if the air conditioning equipment is turned on, enter the adjustment test stage and execute step S412. S412. Record the air conditioner's operating status before adjustment, adjust the air conditioner's power adjustment range and temperature setting range, and when the air conditioner's power and temperature tend to stabilize, record the air conditioner's operating status after adjustment, and base the difference in air conditioner power before and after adjustment on the air conditioner's operating status before and after adjustment. S413. Calculate the air conditioning load adjustment potential based on the air conditioning power difference, and obtain sample data of air conditioning adjustment potential under specific scenarios based on the air conditioning load adjustment potential.

6. The method for evaluating the load regulation potential of a single air conditioner based on transfer learning and multi-task learning according to claim 5, characterized in that, The step of using transfer learning technology to transfer the model weights of the optimized air conditioning load regulation potential assessment model to the target scenario, thereby obtaining an air conditioning load regulation potential assessment model for a specific scenario, includes: S421. The source domain feature selection technique and the weight adaptive adjustment technique are used to optimize the air conditioning regulation potential sample data respectively, and pseudo-labels are generated by self-supervised learning to expand the feature expression ability of the optimized air conditioning regulation potential sample data. S422. The structure of the air conditioning load regulation potential assessment model is updated and optimized using a progressive training strategy and a meta-learning strategy, and trained based on the air conditioning regulation potential sample data to obtain an air conditioning load regulation potential assessment model under a specific scenario. S423. Introduce adversarial training technology into the air conditioning load adjustment potential assessment model under specific scenarios, learn domain-invariant features through a domain discriminator, and optimize the performance of the air conditioning load adjustment potential assessment model under specific scenarios based on the domain-invariant features.

7. The method for evaluating the load regulation potential of a single air conditioner based on transfer learning and multi-task learning according to claim 1, characterized in that, The formula for calculating the Pearson correlation coefficient is as follows: ； In the formula, r XY X represents the Pearson correlation coefficient between the initial training features; i Y represents the observed values ​​of the initial training feature X; i The observed values ​​represent the initial training features Y; i represents the initial number of training features; This represents the mean of the initial training feature X; This represents the mean of the initial training feature Y.

8. The method for evaluating the load regulation potential of a single air conditioner based on transfer learning and multi-task learning according to claim 1, characterized in that, The formula for calculating the Spearman rank correlation coefficient is as follows: ； In the formula, r s d represents the Spearman rank correlation coefficient; i This represents the difference in rank between the two initial training features in the i-th sample; n represents the number of samples.

Images