Monomer air conditioner load regulation potential evaluation method based on transfer learning and multi-task learning

Abstract

The invention discloses a single air conditioner load regulation potential evaluation method based on transfer learning and multi-task learning, and relates to the technical field of air conditioner management.The method comprises the steps that a regional association model is built based on air conditioner operation data in a general scene, and the regional association model is analyzed to obtain a cooperative regulation relation between air conditioner equipment; constructing an air conditioner load regulation potential evaluation model by using a deep residual neural network; a multi-task learning framework is formulated according to the cooperative adjustment relation between the air conditioning equipment, and the air conditioning load adjustment potential evaluation model in the general scene is optimized by using the multi-task learning framework; and adjusting the weight of the optimized air conditioner load regulation potential evaluation model in combination with a transfer learning technology to obtain an air conditioner load regulation potential evaluation model in a specific scene. According to the method, the adaptability and generalization ability of the air conditioner load regulation potential evaluation model are remarkably improved by constructing the regional correlation model and introducing self-supervised learning and adversarial training.

Description

Technical Field

[0001] The present invention relates to the technical field of air conditioning management, and in particular to a method for evaluating the load regulation potential of a single air conditioner based on transfer learning and multi-task learning. Background Art

[0002] As a load management strategy, demand response usually guides power users to adjust their electricity consumption behavior during peak or trough periods of grid demand through incentives, price signals or direct control, so as to help grid operators balance supply and demand more efficiently and improve the stability and economy of the grid. Demand-side response not only helps reduce dependence on traditional power generation resources, but also promotes the consumption of renewable energy and provides economic compensation or incentives for participating users.

[0003] As a high-quality flexible load resource, air conditioning load plays an important role in demand-side response. During peak hours of electricity consumption, adjusting air conditioning load can significantly reduce the load pressure of the power grid. The air conditioning system has strong control flexibility and can achieve rapid response to electricity demand by adjusting parameters such as temperature set point and compressor frequency. At the same time, the thermal accumulation effect of temperature allows the air conditioner to fine-tune the load in a short time, minimizing the impact on user comfort and keeping the temperature change of the indoor environment within the acceptable range for users.

[0004] At present, the data-driven method uses big data analysis and machine learning technology to mine the regulation rules of air conditioning load from historical operation data and identify the load regulation potential. However, the existing technology has the following limitations:

[0005] 1) High cost of acquiring sample data: In certain scenarios (such as new buildings and 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 environments and user habits, existing models are difficult to adapt to complex actual scenarios, and the evaluation results are less stable in different environments.

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

[0008] Currently, no effective solution has been proposed for the problems in the related technologies. Summary of the invention

[0009] In view of the problems in the related technology, the present invention proposes a single air-conditioning load regulation potential evaluation method based on transfer learning and multi-task learning to overcome the above-mentioned technical problems existing in the existing related technology.

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

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

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

[0013] S2. Use a deep residual neural network to build an air conditioning load regulation potential evaluation model, and train the air conditioning load regulation potential evaluation model based on a predefined training data set to obtain an air conditioning load regulation potential evaluation model under a general scenario;

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

[0015] S4. Based on the sample data of air-conditioning regulation potential in specific scenarios, and combined with the transfer learning technology to adjust the weights of the optimized air-conditioning load regulation potential evaluation model, an air-conditioning load regulation potential evaluation model in specific scenarios is obtained to evaluate the air-conditioning load regulation potential in 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] Among them, the air conditioning equipment data includes production date, installation date, refrigerant type, filter status and equipment energy efficiency;

[0018] Building environmental data include thermal insulation performance, structural layout, usage function and construction year;

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

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

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

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

[0023] S21, preprocessing the air-conditioning operation data, and performing feature selection, feature construction and feature conversion on the preprocessed air-conditioning operation data in sequence to obtain initial training features;

[0024] S22, using principal component analysis dimensionality reduction technology to identify key features for representing the trend of the training data set from the initial training features, and constructing the training data set based on the key features;

[0025] S23. Construct an air conditioning load regulation potential evaluation model based on a deep residual neural network, and use the training data set to train the air conditioning load regulation potential evaluation model to obtain the air conditioning load regulation potential evaluation model under general scenarios.

[0026] Preferably, using principal component analysis dimensionality reduction technology to identify key features for representing the trend of the training data set from the initial training features, and constructing the training data set based on the key features includes:

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

[0028] S222, judging 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, calculating the Spearman rank correlation coefficient between the initial training features instead of the Pearson correlation coefficient, otherwise, executing step S223;

[0029] S223, calculating the correlation coefficient matrix between the initial training features, and performing eigenvalue decomposition on the correlation coefficient matrix to obtain an eigenvalue decomposition result;

[0030] S224. Calculate the principal component cumulative contribution rate of the initial training features based on the eigenvalue decomposition results, screen out key features according to the principal component cumulative contribution rate, and integrate the key features to obtain a training data set.

[0031] Preferably, calculating the principal component cumulative contribution rate of the initial training features based on the eigenvalue decomposition results, screening out key features according to the principal component cumulative contribution rate, and integrating the key features to obtain a training data set includes:

[0032] S2241, obtaining eigenvectors corresponding to the eigenvalue decomposition result, and sorting the eigenvectors to filter out eigenvectors within a preset range;

[0033] S2242, calculating the ratio of the selected eigenvectors to the total eigenvectors, taking the calculation result as the principal component contribution rate, and accumulating the principal component contribution rates to obtain the principal component cumulative contribution rate;

[0034] S2243. Compare the cumulative contribution rate of the principal component with the preset threshold. If the cumulative contribution rate of the principal component is greater than the preset threshold, convert the feature vector corresponding to the cumulative contribution rate of the principal component to a new feature space to obtain a key feature set, and integrate the key features to obtain a training data set; otherwise, adjust the preset range and re-execute step S2241.

[0035] Preferably, an air conditioning load regulation potential evaluation model is constructed based on a deep residual neural network, and the air conditioning load regulation potential evaluation model is trained using a training data set, and the air conditioning load regulation potential evaluation model under a general scenario includes:

[0036] S231. Construct a multi-layer residual module in the deep residual neural network, introduce a dense connection mechanism in each layer of the residual module, and transfer 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. Introducing channel attention mechanism and spatial attention mechanism into the deep residual neural network respectively to enhance the ability of the air conditioning load regulation potential assessment model to focus on key features;

[0038] S233, introducing a maximum pooling layer at the output of the residual module, processing the key features in sequence through the residual module, the maximum pooling layer and the fully connected layer, and passing the processing results to the output layer for output;

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

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

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

[0042] S2322. Use convolution operation to generate spatial weights in the deep residual neural network, so that the deep residual neural network focuses on key features at different positions.

[0043] Preferably, formulating a multi-task learning framework according to the coordinated adjustment relationship between air-conditioning equipment includes:

[0044] S31, generating several subtasks based on the coordinated adjustment relationship between air-conditioning equipment, and establishing hard sharing constraints to capture the common features of each subtask;

[0045] S32, introducing soft sharing constraints for each subtask specific layer, and calculating the parameter distances between the subtask specific layers, comparing the similarities of the subtask specific layer parameter distances, and generating task losses based on similar results;

[0046] S33. A regularization term is introduced into the soft sharing constraint to obtain a soft sharing regularization loss, and a multi-task joint learning framework is constructed by combining the task loss and the soft sharing regularization loss.

[0047] S34. According to the predefined subtask learning difficulty, dynamically adjust the loss weights of the hard sharing constraint and the soft sharing constraint in the multi-task joint learning framework.

[0048] Preferably, based on the air conditioning regulation potential sample data in a specific scenario, the weight of the air conditioning load regulation potential evaluation model after optimization is adjusted in combination with the transfer learning technology, and the air conditioning load regulation potential evaluation model in a specific scenario is obtained to evaluate the air conditioning load regulation potential in different scenarios, including:

[0049] S41, using a field test method to test the regulation potential of the air-conditioning equipment in a specific scenario, and obtaining sample data of the air-conditioning regulation potential in the specific scenario;

[0050] S42. Use transfer learning technology to transfer the model weights of the optimized air-conditioning load regulation potential evaluation model to the target scenario, and obtain the air-conditioning load regulation potential evaluation model under a specific scenario to evaluate the air-conditioning load regulation potential under different scenarios.

[0051] Preferably, the air conditioning equipment in a specific scenario is tested for its regulation potential by using a field test method, and the air conditioning regulation potential sample data in the specific scenario is obtained, including:

[0052] S411, determining whether the air conditioning device in a specific scenario is in a powered-on state; if the air conditioning device is not powered on, it indicates that there is no adjustment potential, and the process ends; if the air conditioning device is powered on, the adjustment test phase is entered, and step S412 is executed;

[0053] S412, recording the air conditioner operating state before adjustment, adjusting the air conditioner power adjustment range and temperature setting interval, and when the air conditioner power and temperature tend to be stable, recording the air conditioner operating state after adjustment, and based on the air conditioner power difference before and after adjustment;

[0054] S413. Calculate the air conditioning load regulation potential based on the air conditioning power difference, and obtain air conditioning regulation potential sample data in a specific scenario according to the air conditioning load regulation potential.

[0055] Preferably, the model weights of the optimized air conditioning load regulation potential evaluation model are transferred to the target scene by using transfer learning technology, and the air conditioning load regulation potential evaluation model under the specific scene is obtained, including:

[0056] S421, optimizing the air conditioning regulation potential sample data by using source domain feature selection technology and weight adaptive adjustment technology respectively, and generating pseudo labels by using self-supervised learning to expand the feature expression capability of the optimized air conditioning regulation potential sample data;

[0057] S422, using a progressive training strategy and a meta-learning strategy to update the optimized air conditioning load regulation potential evaluation model structure, and training based on the air conditioning regulation potential sample data to obtain an air conditioning load regulation potential evaluation model under a specific scenario;

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

[0059] Preferably, the calculation formula of the Pearson correlation coefficient is:

[0060]

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

[0062] Preferably, the calculation formula of the Spearman rank correlation coefficient is:

[0063]

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

[0065] The beneficial effects of the present invention are:

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

[0067] 2. The present invention constructs an air-conditioning load regulation potential evaluation model in 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 evaluation model in general scenarios. At the same time, the model is optimized using transfer learning technology, and the spatial correlation characteristics between equipment and user comfort are comprehensively considered to achieve accurate load regulation potential evaluation of air-conditioning equipment in different scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0069] Figure 1 is a flow chart 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 is a schematic diagram of a thermal comfort model 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;

[0071] Figure 3 It is a schematic diagram of an air conditioning load regulation potential assessment model based on a deep residual neural network 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;

[0072] Figure 4 It is a flow chart of obtaining air conditioning regulation potential sample data based on a field test method 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 It is a training flow chart of deep residual network transfer learning 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. DETAILED DESCRIPTION

[0074] To further illustrate each embodiment, the present invention provides drawings, which are part of the disclosure of the present invention and 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 contents, ordinary technicians in the field should be able to understand other possible implementations and advantages of the present invention. The components in the figures 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 with reference to the accompanying drawings and specific embodiments. Figure 1-Figure 5 As shown, according to the single air conditioner load regulation potential evaluation method based on transfer learning and multi-task learning in an embodiment of the present invention, the single air conditioner load regulation potential evaluation method includes:

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

[0078] It should be noted that 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;

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

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

[0081] 1) Thermodynamic model:

[0082] Considering the influence of heat exchange between the outside and indoor temperature on the air conditioning power, a thermodynamic model is established to calculate the power demand of the air conditioning. The thermodynamic model expression is as follows:

[0083] Q cond = k × A × (T out -T in );

[0084] In the formula, Q cond represents the amount of heat transferred by heat conduction, k represents the thermal conductivity of the building material, A represents the building surface area, T out and T in Represents outdoor temperature and indoor temperature respectively.

[0085] 2) Building Energy Efficiency Model:

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

[0087]

[0088] Where P building represents the load power demand of the building, T in Indicates the indoor temperature, T set represents the set temperature, S represents the wind speed setting value, R represents the equivalent thermal resistance, and C represents the heat capacity of the building.

[0089] 3) Air conditioning power model:

[0090] Considering the impact of cooling and heating demand on air conditioning power consumption, an air conditioning power consumption model is established. The expression of the air conditioning power consumption model is as follows:

[0091]

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

[0093] 4) Thermal comfort model:

[0094] Considering the thermal comfort of human body under different temperature conditions, a thermal comfort model is established. Thermal comfort index is usually evaluated by predicted mean thermal sensation (PMV) and predicted dissatisfaction rate (PPD). The thermal comfort model expression is as follows:

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

[0096] Where PMV represents the predicted mean thermal sensation value, M represents the human metabolic rate (unit: W / m 2), L represents the heat exchange between the human body and the environment.

[0097] The predicted dissatisfaction rate (PPD) is calculated based on the PMV value. The calculation formula is as follows:

[0098]

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

[0100] 5) User behavior model:

[0101] Analyze users' settings and adjustment habits for indoor temperature, as these user behaviors will affect the changing trend of air-conditioning load; establish a user behavior model to predict users' temperature setting preferences, and add user comfort constraints to improve the accuracy of user behavior model predictions and user experience.

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

[0103] Step 1: Collect temperature setting and adjustment related data and extract key behavior features.

[0104] Step 2: Analyze different user behavior patterns based on the user's temperature setting and operation frequency, and use a clustering algorithm to group different user behavior patterns, where the grouped user behavior patterns include:

[0105] Stable type: small fluctuations in set values ​​and few adjustments.

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

[0107] Temperature difference sensitive type: The set value changes closely with the outdoor temperature.

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

[0109] Step 3: Automatically generate behavior labels for users based on clustering results:

[0110] Fixed temperature type: prefers a constant temperature (such as always set to 24°C).

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

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

[0113] Frequent adjustment type: The number of adjustment operations is higher than the set threshold (such as more than 5 times per day).

[0114] Use logical rules combined with clustering results to further subdivide labels:

[0115] Time-dependent: The temperature difference between night and day exceeds a certain value (such as more than 2°C).

[0116] Seasonal sensitivity: There is a significant temperature difference between winter and summer (e.g. the set value differs by more than 3°C).

[0117] The decision tree parameters are set based on the behavior pattern extracted by the decision tree, and the decision tree is trained according to the decision tree parameters. The behavior pattern rules are automatically extracted through the path generated by the decision tree. The decision tree parameters include:

[0118] Maximum depth (max_depth): limit the depth of the tree to avoid overfitting; minimum sample split number (min_samples_split): control the number of samples of leaf nodes. Use the target label (behavior type) as the classification target, for example:

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

[0120] Low temperature and frequent adjustment type: the average set temperature is lower than 23.5℃; the adjustment frequency is higher than 3.5 times / day.

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

[0122] Time-sensitive: The average set temperature is higher than 23.5°C; 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 association model to capture the spatial coordinated regulation relationship between different air-conditioning equipment. The mutual influence between air-conditioning equipment is identified through the regional association model, thereby improving the model's load regulation potential prediction ability.

[0125] It should be noted that the introduction of graph neural network (GNN) to establish a regional association model to capture the spatial coordinated regulation relationship between different air-conditioning equipment is specifically expressed as representing the air-conditioning equipment in the region and their mutual relationships as a graph structure to capture the spatial coordinated regulation relationship, which includes:

[0126] Step 1: Extract the operating status of different air-conditioning equipment, building spatial layout, user behavior data and thermodynamic model input data, etc., 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 relationship between building space and equipment is represented by an adjacency matrix, with nodes representing air-conditioning equipment and edges representing spatial connections (with weights representing heat exchange efficiency).

[0128] Step 2: Construct a graph through the adjacency matrix, in which each node represents an air-conditioning device or area, and the characteristics include the current set temperature, actual temperature, operating power and user preference characteristics; the weight of the edge is determined by the heat exchange intensity calculated by the thermodynamic model, the physical distance and the building energy efficiency characteristics, and a time series dynamic graph is constructed to represent the equipment status and its relationship at different time steps.

[0129] Step 3: Use the graph convolutional network to take node features and edge features (spatial correlation characteristics) as input, aggregate neighbor information through the graph convolution layer and capture the spatial collaborative regulation relationship between different air-conditioning equipment. The expression of the graph convolution layer aggregating neighbor information is:

[0130]

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

[0132] Step 4: Based on the spatial coordinated regulation relationship between different air-conditioning devices, analyze the temporal dynamic relationship between air-conditioning devices, which includes:

[0133] A time series dynamic graph is used to describe the current status and spatial relationship of the air-conditioning equipment, and a recurrent neural network (such as LSTM, GRU) is used to process the time dimension. The spatial coordination relationship is dynamically updated based on the time dimension processing results, and the optimization goals are defined, including spatial coordination (minimizing the temperature difference between adjacent equipment), thermal comfort (PMV index is close to the target range), user behavior matching (the set temperature is close to the user preference) and energy consumption optimization, etc.

[0134] S2. Use a deep residual neural network to build an air-conditioning load regulation potential evaluation model, and train the air-conditioning load regulation potential evaluation model based on a predefined training data set to obtain an air-conditioning load regulation potential evaluation model under general scenarios.

[0135] It should be noted that a general training data set for the air-conditioning load regulation potential assessment model is created using air-conditioning operation data; operating condition data such as indoor temperature, outdoor temperature, indoor humidity, space size, temperature setting value, air-conditioning operation mode, and air-conditioning wind speed are collected during daily operation of the air-conditioner; power data before and after the air-conditioning operation parameters are adjusted are recorded, and the power regulation potential of the air-conditioner is evaluated using the power difference; in addition, operating data during the regulation process is also collected, including the regulation instructions issued by the dispatcher, the actual load of the air-conditioner, the regulation start time, the regulation end time, and the regulation deviation value; based on the above 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 calculation formula of power regulation deviation is:

[0137] ΔP=Pg-Pactual;

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

[0139] The calculation formula of load regulation rate is:

[0140]

[0141] In the formula, S r Indicates the air conditioning load adjustment rate; P1 indicates the power at the start of adjustment; P2 indicates the power at the end of adjustment; t1 and t2 indicate the start and end time of adjustment respectively.

[0142] The standard regulation rate is calculated as:

[0143]

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

[0145] The calculation formula for adjustment accuracy is:

[0146]

[0147] Where η represents the adjustment accuracy, ranging from 0 to 1; ΔP represents the adjustment deviation value; P g Indicates the target power value.

[0148] The response time is calculated as:

[0149] Tresponse = t2-t1;

[0150] Where Tresponse represents the response time of the air conditioning system; t1 represents the start time of regulation; and t2 represents the end time of regulation.

[0151] Among them, a deep residual neural network is used to build an air conditioning load regulation potential evaluation model, and the air conditioning load regulation potential evaluation model is trained based on a predefined training data set. The air conditioning load regulation potential evaluation model under general scenarios includes:

[0152] S21, preprocessing the air-conditioning operation data, and performing feature selection, feature construction and feature conversion on the preprocessed air-conditioning operation data in sequence to obtain initial training features.

[0153] It should be noted that the air conditioning operation data is preprocessed to ensure the accuracy and consistency of the data set, including:

[0154] Missing data supplementation: Linear interpolation method is used to supplement missing data.

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

[0156] Deduplication: Remove duplicate records to avoid data deviation.

[0157] Data standardization and normalization: The data is converted using the normalization method, and the calculation formula is as follows:

[0158]

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

[0160] It should be noted that the key features that have an important impact on the model's prediction ability are extracted through feature selection, feature construction and feature transformation, and the principal component analysis dimensionality reduction technology is used to identify and extract the key features that best represent the trend and structure of the data set to construct the training data set. Feature selection, feature construction and feature transformation include:

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

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

[0163] Daily average temperature, 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 short-term fluctuations and enhance the capture of temperature trends.

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

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

[0166] In the formula, T is temperature, which reflects the degree of hotness or coldness of the environment; RH is relative humidity; and the weighting coefficient of 0.1 indicates that humidity has little effect on comfort.

[0167] Date feature tags: Generate binary feature tags of "legal holidays", "working days" and "non-working days" to reflect the changes in load demand with dates.

[0168] Time feature construction: Map the time information into hours (0 to 23) or segments (morning, noon, and evening) to capture the load variation pattern in different time periods.

[0169] Feature conversion: convert some features to improve the predictive ability and adaptability of the model. For example, the building orientation is converted to azimuth to quantify the impact of different orientations on the load; the thermal insulation performance is quantified in percentage to represent the thermal insulation effect of different buildings; the use period is converted to usage duration and start time to improve the interpretability of time features.

[0170] S22. Use principal component analysis dimensionality reduction technology to identify key features used to represent the trend of the training data set from the initial training features, and construct a training data set based on the key features.

[0171] Among them, the principal component analysis dimensionality reduction technology is used to identify key features used to represent the trend of the training data set from the initial training features, and the training data set is constructed based on the key features, including:

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

[0173] S222, judging 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, calculating the Spearman rank correlation coefficient between the initial training features instead of the Pearson correlation coefficient, otherwise, executing step S223;

[0174] S223, calculating the correlation coefficient matrix between the initial training features, and performing eigenvalue decomposition on the correlation coefficient matrix to obtain an eigenvalue decomposition result;

[0175] S224. Calculate the principal component cumulative contribution rate of the initial training features based on the eigenvalue decomposition results, screen out key features according to the principal component cumulative contribution rate, and integrate the key features to obtain a training data set.

[0176] Among them, the principal component cumulative contribution rate of the initial training features is calculated based on the eigenvalue decomposition results, the key features are screened out according to the principal component cumulative contribution rate, and the key features are integrated to obtain the training data set including:

[0177] S2241, obtaining eigenvectors corresponding to the eigenvalue decomposition result, and sorting the eigenvectors to filter out eigenvectors within a preset range;

[0178] S2242, calculating the ratio of the selected eigenvectors to the total eigenvectors, taking the calculation result as the principal component contribution rate, and accumulating the principal component contribution rates to obtain the principal component cumulative contribution rate;

[0179] S2243. Compare the cumulative contribution rate of the principal component with the preset threshold. If the cumulative contribution rate of the principal component is greater than the preset threshold, convert the feature vector corresponding to the cumulative contribution rate of the principal component to a new feature space to obtain a key feature set, and integrate the key features to obtain a training data set; otherwise, adjust the preset range and re-execute step S2241.

[0180] It should be noted that the principal component analysis dimensionality reduction technique is used to identify key features for representing the trend of the training data set from the initial training features, and the training data set is constructed based on the key features, including:

[0181] Standardization: All initial training feature values ​​are converted to the range of 0 to 1 to eliminate the impact of different dimensions on data comparison.

[0182] Calculate the correlation coefficient: Calculate the Pearson correlation coefficient between the initial training features to measure the linear correlation. The calculation formula of the Pearson correlation coefficient is:

[0183]

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

[0185] When there is a nonlinear relationship in the data or there are many outliers, the Spearman rank correlation coefficient is used as an alternative method. The calculation formula of the Spearman rank correlation coefficient is:

[0186]

[0187] In the formula, r s represents the Spearman rank correlation coefficient; d i represents the difference between the ranks of 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, and the eigenvalue decomposition is performed on it. The key features are screened out based on the decomposed eigenvalues ​​and the cumulative contribution rate of the principal components, including:

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

[0190] Perform eigenvalue decomposition on the covariance matrix to obtain each eigenvalue and the corresponding eigenvector; the eigenvalue reflects the degree of variation of the data in the direction of the corresponding eigenvector, and the eigenvector indicates the direction of the principal component of the data.

[0191] Sort by eigenvalue from large to small; the corresponding eigenvectors are also arranged in this order, indicating that the importance of the principal components is from high to low.

[0192] Calculate the proportion of each eigenvalue to the total eigenvalues, which is called the principal component contribution rate; add up the principal component contribution rates in order to get the principal component cumulative contribution rate, which is used to determine the proportion of data variation that the selected principal component can explain.

[0193] According to the cumulative contribution rate, a threshold (such as 85% or 95%) is set, and the first few principal components required for the cumulative contribution rate to reach the threshold are selected; the eigenvectors corresponding to these principal components are checked, and the original features with larger weights in each principal component are extracted as key features.

[0194] S23. Construct an air conditioning load regulation potential evaluation model based on a deep residual neural network, and use the training data set to train the air conditioning load regulation potential evaluation model to obtain the air conditioning load regulation potential evaluation model under general scenarios.

[0195] Among them, an air conditioning load regulation potential evaluation model is constructed based on a deep residual neural network, and the air conditioning load regulation potential evaluation model is trained using a training data set. The air conditioning load regulation potential evaluation model under a general scenario includes:

[0196] S231. Construct a multi-layer residual module in the deep residual neural network, introduce a dense connection mechanism in each layer of the residual module, and transfer 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 respectively to enhance the ability of the air conditioning load regulation potential assessment model to focus on key features.

[0198] Among them, the channel attention mechanism and spatial attention mechanism are introduced into the deep residual neural network to improve the air conditioning load regulation potential evaluation model's ability to focus on key features, including:

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

[0200] S2322. Use convolution operation to generate spatial weights in the deep residual neural network, so that the deep residual neural network focuses on key features at different positions.

[0201] S233, introducing a maximum pooling layer at the output of the residual module, processing the key features in sequence through the residual module, the maximum pooling layer and the fully connected layer, and passing the processing results to the output layer for output;

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

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

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

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

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

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

[0208] In order to make full use of the feature information extracted in each layer, the model introduces dense connection in the residual module to directly transfer the output of all previous layers to the subsequent layers. The expression of feature transfer is as follows:

[0209]

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

[0211] ②Channel and spatial attention mechanism:

[0212] Channel attention mechanism: The global features of each channel are calculated through global average pooling (GAP), 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 represents the channel weight; σ represents the Sigmoid activation function; F c is the channel feature map.

[0215] Spatial attention mechanism: Generate spatial weights through convolution operations to make the model pay attention to the load characteristics at different positions. The calculation formula is as follows:

[0216] w s =σ(Conv(F));

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

[0218] ③Pooling layer and nonlinear activation function:

[0219] The improved residual block is followed by a 2×2 maximum pooling layer with a step size of 2 to reduce the spatial dimension and extract key information. All intermediate layers use the rectified linear unit (ReLU) as the activation function, and the calculation formula is as follows:

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

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

[0222] ④Fully connected layer and Softmax output:

[0223] After the residual block and pooling layer processing, it passes through 3 layers of fully connected layers (configured as FC64, FC32 and FC16) and is passed to the Softmax layer to output the prediction result. The expression of the Softmax function is as follows:

[0224]

[0225] Among them, Softmax represents the Softmax function; z i represents the score of the i-th category; K represents the number of categories; z j represents the score of the jth class.

[0226] It should be noted that the training set and validation set are constructed according to the key features, the training data is 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 back propagation:

[0228] The error between the network output and the actual regulation potential is calculated. The loss function is the mean square error (MSE), which is calculated as follows:

[0229]

[0230] Where L represents the loss value; Represents the model prediction value; y i represents the actual value; N represents the number of samples.

[0231] 2) Gradient calculation and parameter update:

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

[0233]

[0234] According to the calculated gradient, choose an appropriate learning rate α to update the weights:

[0235]

[0236] Where W new represents the weight after learning rate update; α represents the learning rate; L represents the loss value.

[0237] 3) Model evaluation and validation:

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

[0239] S3. A multi-task learning framework is formulated based on the collaborative adjustment relationship between air-conditioning equipment, and the multi-task learning framework is used to optimize the air-conditioning load adjustment potential evaluation model in general scenarios.

[0240] Among them, the multi-task learning framework is formulated according to the coordinated adjustment relationship between air-conditioning equipment, including:

[0241] S31, generating several subtasks based on the coordinated adjustment relationship between air-conditioning equipment, and establishing hard sharing constraints to capture the common features of each subtask;

[0242] S32, introducing soft sharing constraints for each subtask specific layer, and calculating the parameter distances between the subtask specific layers, comparing the similarities of the subtask specific layer parameter distances, and generating task losses based on similar results;

[0243] S33. A regularization term is introduced into the soft sharing constraint to obtain a soft sharing regularization loss, and a multi-task joint learning framework is constructed by combining the task loss and the soft sharing regularization loss.

[0244] S34. According to the predefined subtask learning difficulty, dynamically adjust the loss weights of the hard sharing constraint and the soft sharing constraint in the multi-task joint learning framework.

[0245] It should be noted that a multi-task learning framework is formulated according to the coordinated adjustment relationship between air-conditioning equipment, and the air-conditioning load adjustment potential evaluation model in general scenarios is optimized using the multi-task learning framework, including:

[0246] Step 1: Define subtasks and subtask divisions (only subtask 1 to subtask 4 are used as examples):

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

[0248] Subtask 2: Coordinate node relationship modeling to capture the spatial collaboration characteristics and thermal dynamic relationships 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 changing trend of the overall system energy consumption.

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

[0252] Hard shared structure: build a shared bottom layer to capture the common features of all subtasks, for example:

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

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

[0255] Heat exchange characteristics of buildings (spatial distance, heat transfer coefficient).

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

[0257] Soft sharing structure: Task-specific layer. On the basis of the shared layer, an independent specific layer for each subtask is established. For example:

[0258] Load forecasting module: predicts equipment power.

[0259] Collaborative regulation module: modeling the collaborative relationship between devices.

[0260] Thermal comfort module: output PMV index.

[0261] Energy consumption optimization module: calculates the overall energy consumption.

[0262] Soft sharing constraints are introduced for the specific layers of each subtask, and regularization terms are added to make the specific layer parameters between subtasks close to each other.

[0263] Specifically, a soft sharing constraint is introduced for the specific layer of each subtask, and a regularization term is added to make the specific layer parameters between subtasks close to each other, including:

[0264] Step 1: Define soft sharing of subtask specific layers. For example, the specific layers of subtask A and subtask B 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 way to share 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 exactly the same.

[0266] Step 3: Add soft sharing constraints: Calculate the parameter distance between subtask-specific layers and compare the similarity of the parameters of each subtask-specific layer. For example: compare the parameters of the 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 to measure and reduce the parameter differences between subtasks. For example, penalize the situation where the parameters of a specific layer of each subtask deviate too far; the regularization term should control its weight through hyperparameters to avoid excessive interference with the model's subtask personalization ability.

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

[0269] Initialize the parameters of shared layers and subtask-specific layers to ensure that the parameter distributions of specific layers are independent at the beginning of training.

[0270] The specific losses of each subtask are calculated separately, such as load forecasting losses, thermal comfort losses, etc., and constraints are imposed on the specific layer parameters of each subtask, for example:

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

[0272] Combining the subtask loss and the soft-shared regularization loss, a multi-subtask joint optimization objective is constructed. When updating the gradient, the subtask loss and the soft-shared regularization loss are reduced at the same time, gradually bringing the parameters of the subtask specific layers closer to each other.

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

[0274] Step 3: Use uncertainty weighting to dynamically adjust the loss weight between hard sharing and soft sharing according to the learning difficulty and importance of the subtask.

[0275] It should be noted that hard sharing shares common information among all subtasks by sharing the underlying feature extractor, which is suitable for scenarios with strong correlations between subtasks; soft sharing models each subtask individually through the subtask-specific layer, while introducing constraints or regularization to promote collaborative learning between subtasks; hard sharing provides a common foundation to reduce repeated calculations and redundant feature extraction, and soft sharing achieves targeted optimization through the subtask-specific layer to meet the needs of multiple subtasks. The combination of the two, by establishing a common feature representation in the shared layer and optimizing the independent goals of each subtask in the specific layer, achieves a balance between global efficiency and local precision.

[0276] S4. Based on the sample data of air-conditioning regulation potential in specific scenarios, and combined with the transfer learning technology to adjust the weights of the optimized air-conditioning load regulation potential evaluation model, an air-conditioning load regulation potential evaluation model in specific scenarios is obtained to evaluate the air-conditioning load regulation potential in different scenarios.

[0277] Among them, based on the air conditioning regulation potential sample data in specific scenarios, and combined with the transfer learning technology to adjust the weights of the optimized air conditioning load regulation potential evaluation model, the air conditioning load regulation potential evaluation model in specific scenarios is obtained to evaluate the air conditioning load regulation potential in different scenarios, including:

[0278] S41. Use the on-site testing method to test the regulation potential of the air-conditioning equipment in a specific scenario to obtain sample data of the air-conditioning regulation potential in the specific scenario.

[0279] Among them, the field test method is 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 are obtained, including:

[0280] S411, determining whether the air conditioning device in a specific scenario is in a powered-on state; if the air conditioning device is not powered on, it indicates that there is no adjustment potential, and the process ends; if the air conditioning device is powered on, the adjustment test phase is entered, and step S412 is executed;

[0281] S412, recording the air conditioner operating state before adjustment, adjusting the air conditioner power adjustment range and temperature setting interval, and when the air conditioner power and temperature tend to be stable, recording the air conditioner operating state after adjustment, and based on the air conditioner power difference before and after adjustment;

[0282] S413. Calculate the air conditioning load regulation potential based on the air conditioning power difference, and obtain air conditioning regulation potential sample data in a specific scenario according to the air conditioning load regulation potential.

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

[0284] 1) Power-on status detection to determine whether the air conditioner is in operation. If the air conditioner is not turned on, there is no adjustment potential and the process ends. If the air conditioner is turned on, it enters the adjustment test phase and prepares 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) Derivation of set temperature: Determine the temperature setting range required to achieve target power adjustment through historical data and experience:

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

[0290] 5) Perform setting adjustment, adjust the temperature setting of the air conditioner according to the target power value, and start the power adjustment process. Record the temperature and power changes in real time during the entire adjustment process to ensure the continuity and integrity of data collection.

[0291] 6) After adjusting the set temperature, when the power and temperature tend to stabilize, 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 regulation potential. According to the power difference during the adjustment process, calculate the load regulation potential of the air conditioner at different set temperatures, and record the actual power regulation effect.

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

[0294] The calculation formula of load regulation potential is as follows:

[0295]

[0296] Where P potential It indicates the load regulation potential, that is, the load power change of the air conditioner in unit time. This indicator is used to quantify the regulation effect of the air conditioner under different temperature setting conditions; P initial Indicates the initial power at the beginning of regulation, that is, the actual power value before adjusting the set temperature; P final Indicates the power value after adjustment, that is, the stable power value after the air conditioner set temperature is adjusted; T adjust Represents the adjustment time, i.e. the time from the start of adjustment until the power reaches a stable state at the new temperature setting, in seconds.

[0297] S42. Use transfer learning technology to transfer the model weights of the optimized air-conditioning load regulation potential evaluation model to the target scenario, and obtain the air-conditioning load regulation potential evaluation model under a specific scenario 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 evaluation model are transferred to the target scene by using the transfer learning technology, and the air-conditioning load regulation potential evaluation model under the specific scene is obtained, including:

[0299] S421. Optimize the air conditioning regulation potential sample data using source domain feature selection technology and weight adaptive adjustment technology respectively, and generate pseudo labels using self-supervised learning to expand the feature expression capabilities of the optimized air conditioning regulation potential sample data.

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

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

[0302] Assign weights to different features or samples in the sample data, give higher weights to high-frequency important features, and give lower weights to outliers or low-correlation samples; for subtasks such as load forecasting and thermal comfort analysis in multi-task learning, dynamically adjust the weight of each task in the total loss, including:

[0303] Using the instance weighting method, weights are assigned according to the importance of samples (such as error rate weighting in the Boosting algorithm), and the model is retrained after the weight adjustment to make the model more focused on high-weight samples or features.

[0304] Select unlabeled data (such as unlabeled load potential or thermal comfort data) from the air conditioning potential sample data, and use an unsupervised method (such as clustering or dimensionality reduction technology) to construct a preliminary label structure, and use a self-supervised learning technology to generate pseudo labels based on the preliminary label structure, which includes:

[0305] Design pretext tasks, for example, predicting feature ranking, feature contrast (contrastive learning), or future values ​​of time series.

[0306] By completing the pre-task training encoder, extracting the embedded representation of the unlabeled data, using the output of the encoder to classify or regress the unlabeled data, generating pseudo labels, adding the pseudo labels to the training data, merging the original label and pseudo label samples, expanding the data scale through pseudo labels, and optimizing the feature distribution.

[0307] S422. Use progressive training strategy and meta-learning strategy to update the optimized air conditioning load regulation potential evaluation model structure, and train it based on air conditioning regulation potential sample data to obtain the air conditioning load regulation potential evaluation model under specific scenarios.

[0308] It should be noted that meta-learning is a technology that learns general learning rules on multiple tasks. It is suitable for the needs of rapid adaptation of models in new scenarios or tasks. The air conditioning load regulation potential evaluation model structure after updating and optimizing using the meta-learning strategy includes:

[0309] Step 1: Treat the subtasks of air conditioning load regulation as metatasks (only metatasks 1-3 are used as examples):

[0310] Meta-task 1: Single equipment 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 in different meta-tasks, with the shared layers serving as common parameters for meta-learning and the specific layers of each meta-task serving as meta-task adaptation parameters.

[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, which includes:

[0315] Each meta-task is initialized based on the current shared layer parameters and performs gradient updates on the task-specific layers. The shared layer parameters are updated using the cumulative gradients of each meta-task. The task-specific layer parameters are fine-tuned using 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 return the optimized task parameters to the shared initial parameters:

[0317] Specifically, in the initial stage: use progressive strategies to train simple tasks and simple samples to learn basic features and simple relationships.

[0318] Mid-term stage: Add complex samples and complex tasks, optimize task-specific layers through multi-task learning, and gradually add meta-learning strategies in mid- and late-stage tasks to enable shared layers to have cross-task adaptability.

[0319] Later stage: Meta-learning strategies are used to optimize shared parameters so that the model maintains efficient prediction performance in diverse scenarios.

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

[0321] It should be noted that adversarial training technology combines the idea of ​​generative adversarial network (GAN) and is divided into the following parts:

[0322] Feature Extractor: Learn feature representations of raw data.

[0323] Domain discriminator: distinguish the domain from which the samples come.

[0324] Task predictor: An assessment task based on feature-optimized air conditioning load regulation potential.

[0325] During the adversarial training process, the feature extractor learns domain-invariant features by deceiving the domain discriminator, thereby improving the adaptability of the model in the target scenario.

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

[0327] The feature extractor learns domain-invariant features.

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

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

[0330] Specifically, stage 1: domain discriminator training: fix the feature extractor parameters and only train the domain discriminator to accurately distinguish domains.

[0331] Stage 2: Adversarial optimization of feature extractor and task predictor: fix the domain discriminator parameters, train the feature extractor and task predictor; the feature extractor learns domain-invariant features by adversarial domain discriminator; the task predictor optimizes the target task based on the domain-invariant features.

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

[0333] It should be noted that if Figure 5 As shown in the figure, the transfer learning method based on the deep residual network (DRN) uses a small amount of field test data to continue optimizing the pre-trained deep convolutional network, migrate part of the network structure and weights to the target scene, and build an air conditioning load regulation potential evaluation model in a specific scene, including:

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

[0335] A small amount of sample data is obtained from the field test method, and the source domain feature selection and weight adaptive adjustment are used to optimize data utilization. The input feature matrix is ​​as follows:

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

[0337] Where, X selected represents the selected input feature matrix, x i represents the i-th feature, w i represents the adaptive weight of the i-th feature.

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

[0339] Based on the air conditioning load regulation potential evaluation model in general scenarios, multi-scale convolution and attention mechanism are added to improve the feature extraction capability. The output of the residual module is as follows:

[0340]

[0341] In the formula, F(x) represents the output feature of the residual module, σ represents the activation function, and W k represents the convolution kernel weight, that is, the convolution kernel of different scales, b k For bias.

[0342] The calculation formula of channel attention weight is as follows:

[0343] w c =σ(GAP(X c ));

[0344] In the formula, w c represents the 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, and the dataset is expanded to enhance feature expression capabilities. Self-supervised learning tasks include time series prediction and missing data recovery, which are used to enhance model performance using unlabeled data.

[0347] 4) Fully connected layer fine-tuning and multi-task learning:

[0348] Fix the convolutional layer, fine-tune the fully connected layer, add the multi-task learning module, and define the total loss function:

[0349]

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

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

[0352] Adopt progressive training strategy, combined with meta-learning strategy to quickly adapt to a small amount of sample data, parameter update formula:

[0353]

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

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

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

[0357] To sum up, with the help of the above-mentioned technical scheme of the present invention, the single air-conditioning load regulation potential evaluation method based on transfer learning and multi-task learning provided by the present invention, by constructing a regional association model and introducing self-supervised learning and adversarial training, significantly improves the adaptability and generalization ability of the air-conditioning load regulation potential evaluation model, thereby effectively solving the problems of high sample acquisition cost and insufficient model generalization ability in the prior art, and thus realizing accurate evaluation and prediction of the air-conditioning load regulation potential in various complex scenarios; the present invention constructs an air-conditioning load regulation potential evaluation model in a general scenario based on a large amount of air-conditioning operation data, and introduces regional association modeling, multi-task learning, self-supervised learning and adversarial training and other technologies, which significantly enhances the adaptability and prediction accuracy of the air-conditioning load regulation potential evaluation model in a general scenario, and at the same time uses transfer learning technology to optimize the model, comprehensively considering the spatial correlation characteristics and user comfort between equipment, so as to realize accurate load regulation potential evaluation 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 principle of the present invention should be included in 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: The single air conditioning load regulation potential assessment method includes: S1. Build a regional correlation model based on air-conditioning operation data in general scenarios, and analyze the regional correlation model to obtain the coordinated adjustment relationship between air-conditioning equipment; S2. Use a deep residual neural network to build an air conditioning load regulation potential evaluation model, and train the air conditioning load regulation potential evaluation model based on a predefined training data set to obtain an air conditioning load regulation potential evaluation model under a general scenario; S3. Develop a multi-task learning framework based on the collaborative regulation relationship between air-conditioning equipment, and use the multi-task learning framework to optimize the air-conditioning load regulation potential evaluation model in general scenarios; S4. Based on the sample data of air-conditioning regulation potential in specific scenarios, and combined with the transfer learning technology to adjust the weights of the optimized air-conditioning load regulation potential evaluation model, an air-conditioning load regulation potential evaluation model in specific scenarios is obtained to evaluate the air-conditioning load regulation potential in different scenarios.

2. According to the method for evaluating the load regulation potential of a single air conditioner based on transfer learning and multi-task learning in claim 1, it is 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 include thermal insulation performance, structural layout, usage function and construction year; The external environment data includes outdoor temperature, indoor humidity and solar radiation intensity; The user behavior data includes the user's settings, adjustment habits and usage frequency of indoor temperature; The air conditioning control data includes power adjustment deviation, load adjustment rate, standard adjustment rate and adjustment accuracy response time.

3. According to the method for evaluating the load regulation potential of a single air conditioner based on transfer learning and multi-task learning in claim 1, it is characterized in that: The air conditioning load regulation potential evaluation model is constructed by using a deep residual neural network, and the air conditioning load regulation potential evaluation model is trained based on a predefined training data set to obtain an air conditioning load regulation potential evaluation model in a general scenario, including: S21, preprocessing the air-conditioning operation data, and performing feature selection, feature construction and feature conversion on the preprocessed air-conditioning operation data in sequence to obtain initial training features; S22, using principal component analysis dimensionality reduction technology to identify key features for representing the trend of the training data set from the initial training features, and constructing the training data set based on the key features; S23. Construct an air conditioning load regulation potential evaluation model based on a deep residual neural network, and use the training data set to train the air conditioning load regulation potential evaluation model to obtain the air conditioning load regulation potential evaluation model under general scenarios.

4. According to claim 3, a method for evaluating the load regulation potential of a single air conditioner based on transfer learning and multi-task learning is characterized in that: The method of using principal component analysis dimensionality reduction technology to identify key features for representing the trend of the training data set from the initial training features, and constructing the training data set based on the key features includes: S221, transforming the feature values ​​of the initial training features into a preset range, and calculating the Pearson correlation coefficient between the initial training features to obtain the linear correlation between the initial training features; S222, judging 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, calculating the Spearman rank correlation coefficient between the initial training features instead of the Pearson correlation coefficient, otherwise, executing step S223; S223, calculating the correlation coefficient matrix between the initial training features, and performing eigenvalue decomposition on the correlation coefficient matrix to obtain an eigenvalue decomposition result; S224. Calculate the principal component cumulative contribution rate of the initial training features based on the eigenvalue decomposition results, screen out key features according to the principal component cumulative contribution rate, and integrate the key features to obtain a training data set.

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 4 is characterized in that: The method of calculating the principal component cumulative contribution rate of the initial training features based on the eigenvalue decomposition results, screening out key features according to the principal component cumulative contribution rate, and integrating the key features to obtain a training data set includes: S2241, obtaining eigenvectors corresponding to the eigenvalue decomposition result, and sorting the eigenvectors to filter out eigenvectors within a preset range; S2242, calculating the ratio of the selected eigenvectors to the total eigenvectors, taking the calculation result as the principal component contribution rate, and accumulating the principal component contribution rates to obtain the principal component cumulative contribution rate; S2243. Compare the cumulative contribution rate of the principal component with the preset threshold. If the cumulative contribution rate of the principal component is greater than the preset threshold, convert the feature vector corresponding to the cumulative contribution rate of the principal component to a new feature space to obtain a key feature set, and integrate the key features to obtain a training data set; otherwise, adjust the preset range and re-execute step S2241.

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 air conditioning load regulation potential evaluation model is constructed based on the deep residual neural network, and the air conditioning load regulation potential evaluation model is trained using the training data set to obtain the air conditioning load regulation potential evaluation model in a general scenario, including: S231. Construct a multi-layer residual module in the deep residual neural network, introduce a dense connection mechanism in each layer of the residual module, and transfer 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; S232. Introducing channel attention mechanism and spatial attention mechanism into the deep residual neural network respectively to enhance the ability of the air conditioning load regulation potential assessment model to focus on key features; S233, introducing a maximum pooling layer at the output of the residual module, processing the key features in sequence through the residual module, the maximum pooling layer and the fully connected layer, and passing the processing results to the output layer for output; S234. Input the training data set into a deep residual neural network for training to obtain an air conditioning load regulation potential assessment model under a general scenario.

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 6, characterized in that: The channel attention mechanism and the spatial attention mechanism are introduced into the deep residual neural network to improve the ability of the air conditioning load regulation potential evaluation model to pay attention to key features, including: S2321, calculating the global features of each channel in the deep residual neural network using global average pooling, and assigning weights to each channel based on the global features; S2322. Use convolution operation to generate spatial weights in the deep residual neural network, so that the deep residual neural network focuses on key features at different positions.

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 multi-task learning framework formulated according to the coordinated adjustment relationship between air-conditioning equipment includes: S31, generating several subtasks based on the coordinated adjustment relationship between air-conditioning equipment, and establishing hard sharing constraints to capture the common features of each subtask; S32, introducing soft sharing constraints for each subtask specific layer, and calculating the parameter distances between the subtask specific layers, comparing the similarities of the subtask specific layer parameter distances, and generating task losses based on similar results; S33. A regularization term is introduced into the soft sharing constraint to obtain a soft sharing regularization loss, and a multi-task joint learning framework is constructed by combining the task loss and the soft sharing regularization loss. S34. According to the predefined subtask learning difficulty, dynamically adjust the loss weights of the hard sharing constraint and the soft sharing constraint in the multi-task joint learning framework.

9. 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 load regulation potential evaluation model weights are adjusted based on the air conditioning regulation potential sample data in a specific scenario and optimized by combining the transfer learning technology to obtain the air conditioning load regulation potential evaluation model in a specific scenario to evaluate the air conditioning load regulation potential in different scenarios, including: S41, using a field test method to test the regulation potential of the air-conditioning equipment in a specific scenario, and obtaining sample data of the air-conditioning regulation potential in the specific scenario; S42. Use transfer learning technology to transfer the model weights of the optimized air-conditioning load regulation potential evaluation model to the target scenario, and obtain the air-conditioning load regulation potential evaluation model under a specific scenario to evaluate the air-conditioning load regulation potential under different scenarios.

10. The method for evaluating the load regulation potential of a single air conditioner based on transfer learning and multi-task learning according to claim 9, characterized in that: The air conditioning equipment in a specific scenario is tested for its regulation potential by using the field test method to obtain the air conditioning regulation potential sample data in the specific scenario, including: S411, determining whether the air conditioning device in a specific scenario is in a powered-on state; if the air conditioning device is not powered on, it indicates that there is no adjustment potential, and the process ends; if the air conditioning device is powered on, the adjustment test phase is entered, and step S412 is executed; S412, recording the air conditioner operating state before adjustment, adjusting the air conditioner power adjustment range and temperature setting interval, and when the air conditioner power and temperature tend to be stable, recording the air conditioner operating state after adjustment, and based on the air conditioner power difference before and after adjustment; S413: Calculate the air conditioning load regulation potential based on the air conditioning power difference, and obtain air conditioning regulation potential sample data in a specific scenario according to the air conditioning load regulation potential.

11. The method for evaluating the load regulation potential of a single air conditioner based on transfer learning and multi-task learning according to claim 10, characterized in that: The method of using transfer learning technology to transfer the model weights of the optimized air conditioning load regulation potential evaluation model to the target scene to obtain the air conditioning load regulation potential evaluation model in a specific scene includes: S421, optimizing the air conditioning regulation potential sample data by using source domain feature selection technology and weight adaptive adjustment technology respectively, and generating pseudo labels by using self-supervised learning to expand the feature expression capability of the optimized air conditioning regulation potential sample data; S422, using a progressive training strategy and a meta-learning strategy to update the optimized air conditioning load regulation potential evaluation model structure, and training based on the air conditioning regulation potential sample data to obtain an air conditioning load regulation potential evaluation model under a specific scenario; S423. Introduce adversarial training technology into the air-conditioning load regulation potential evaluation model in a specific scenario, learn domain-invariant features through a domain discriminator, and optimize the performance of the air-conditioning load regulation potential evaluation model in a specific scenario based on the domain-invariant features.

12. The method for evaluating the load regulation potential of a single air conditioner based on transfer learning and multi-task learning according to claim 4, characterized in that: The calculation formula of the Pearson correlation coefficient is: In the formula, r XY represents the Pearson correlation coefficient between the initial training features; X i represents the observed value of the initial training feature X; Y i Represents the observed value of the initial training feature Y; i represents the number of initial training features; Represents the mean of the initial training feature X; Represents the mean of the initial training feature Y.

13. The method for evaluating the load regulation potential of a single air conditioner based on transfer learning and multi-task learning according to claim 4, characterized in that: The calculation formula of the Spearman rank correlation coefficient is: In the formula, r s represents the Spearman rank correlation coefficient; d i represents the difference between the ranks of the two initial training features in the i-th sample; n represents the number of samples.

Images