Method and system for evaluating day-ahead / intra-day load regulation capacity of building air conditioner cluster

Abstract

The invention discloses a day-ahead / intra-day load regulation capability evaluation method and system for a building air conditioner cluster, and the method comprises the specific steps: (1) collecting open source simulation data and local actual measurement data, and processing the data to obtain a standardized data set; secondly, calculating mobility scores of the candidate source domain and the target domain, and screening an optimal source domain; (2) constructing an evaluation model which takes an improved time convolution network as a core and integrates maximum mean value difference and depth correlation alignment joint field adaptive loss, and generating simulation label reinforcement training through a physical information neural network when labels are insufficient; and (3) respectively applying the trained model to day-ahead and intra-day load regulation potential evaluation, and outputting corresponding results. According to the method, cross-industry migration deviation is reduced through multi-dimensional data normalization and scientific source domain screening, and reliable labels are generated in combination with a physical mechanism, so that simulation evaluation is closer to an actual operation scene, and evaluation precision and practicability are improved.

Description

Technical Field

[0001] This invention relates to the field of equipment load regulation capacity assessment, and in particular to a method and system for assessing the day-ahead / intra-day load regulation capacity of a building air conditioning cluster. Background Technology

[0002] In the practice of electricity demand-side management and energy optimization, building air conditioning systems, as the core electricity load of public buildings and commercial complexes, require accurate assessment of their load regulation potential as a fundamental basis for achieving demand response and balancing the power grid's supply and demand. However, current technologies for assessing the load regulation potential of building air conditioning systems face significant bottlenecks in cross-industry applications, data adaptation, and model generalization, making it difficult to meet the online assessment needs in real-world scenarios.

[0003] From a data perspective, air conditioning load data from different building types exhibit significant heterogeneity. The scale-dominant effect of raw power data can mask the true differences in energy consumption behavior and load patterns, leading to misjudgments. Meteorological and seasonal disturbances cause exogenous shifts in the load curves between the source domain and the target domain of the collected data, increasing cross-domain assessment errors. Intraday pattern misalignment, peak-hour misalignment, and short-term spikes further exaggerate inter-domain differences. Existing technologies lack systematic data normalization solutions, processing only a single dimension and failing to comprehensively isolate these non-essential differences, making it difficult to directly use cross-industry data for model training and transfer.

[0004] Regarding label data, load regulation potential assessment relies on load change labels under demand response scenarios. However, in reality, most buildings lack historical response data, and the data from the few participating buildings is difficult to form a large-scale label set due to the uniqueness of their response strategies and external environments. Existing simulation label generation methods do not fully integrate the physical characteristics of air conditioning systems, resulting in significant deviations from real-world scenarios. Consequently, traditional supervised learning models suffer from poor generalization ability in the target domain due to insufficient training data.

[0005] At the model application level, traditional time-series models are prone to gradient vanishing and low parallel efficiency when processing long-term dependent data on air conditioning loads. Existing TCN applications do not optimize the structure by incorporating the physical characteristics of air conditioning loads, and fail to adequately capture nonlinear relationships. Furthermore, cross-domain models often employ single-domain adaptive loss, failing to simultaneously achieve high-order and second-order statistical alignment between the source and target domains. They also lack scientific indicators for source domain transferability, easily incorporating data from significantly different source domains, exacerbating bias and increasing the risk of negative transfer. In addition, existing assessments do not differentiate between day-ahead and intraday demand scales, resulting in poor practicality and difficulty in supporting actual demand response decisions.

[0006] Therefore, there is an urgent need for a highly accurate load regulation capability assessment technology that is cross-industry adaptable and compatible with fewer labels. Summary of the Invention

[0007] Purpose of the invention: To address the problems existing in the prior art, this invention proposes a method and system for assessing the day-ahead / intra-day load regulation capacity of building air conditioning clusters, enabling accurate and real-time assessment of cross-industry air conditioning load regulation potential under conditions of few labels / no samples, and providing reliable technical support for demand response decision-making on the power demand side.

[0008] Technical solution:

[0009] This invention proposes a method for evaluating the day-ahead / intra-day load regulation capacity of a building air conditioning cluster, comprising:

[0010] Step 1: Collect the operational data of the building air conditioning cluster, and perform standardization and normalization processing on the operational data to obtain a standardized dataset; the operational data includes open-source simulation data of the source domain and the target domain, as well as local measured data;

[0011] Step 2: Construct daily vector features based on the standardized dataset, calculate individual indicators, and obtain a transferability score after component normalization and weighted fusion. Select the optimal source domain based on the transferability score and a preset threshold. The individual indicators include morphological similarity, temperature sensitivity difference, and autocorrelation difference.

[0012] Step 3: Construct a load regulation capability assessment model with an improved temporal convolutional network as the core and an integrated joint domain adaptive loss. Train the load regulation capability assessment model using the standardized datasets of the optimal source domain and target domain. If the labels of the standardized dataset of the target domain are insufficient, generate simulated load labels and reinforce the load regulation capability assessment model through training.

[0013] Step 4: Apply the trained load regulation capacity assessment model to the day-ahead / intra-day load regulation potential assessment of the target domain building air conditioning cluster, and output the assessment results.

[0014] Furthermore, the standardization process includes: uniform time resolution, uniform field caliber, uniform units, and handling of missing values; the missing value handling uses time series linear interpolation for gaps of no more than 3 consecutive sampling points, and fills gaps of more than 3 consecutive sampling points with the average of observations three days before and after the same time.

[0015] The normalization process includes:

[0016] Scale normalization is used to eliminate the difference between building volume and air conditioning installation scale;

[0017] Meteorological normalization is used to isolate exogenous disturbances caused by climate and seasons;

[0018] Morphology normalization; used to reduce the interference of intraday load morphology misalignment.

[0019] Furthermore, the daily vector features are in matrix form, including the source domain feature matrix. Target domain feature matrix ,in , These represent the number of sample days in the source domain and the target domain, respectively. The number of hours in a day.

[0020] Furthermore, the individual indicators include morphological similarity, temperature sensitivity difference, and autocorrelation difference, specifically:

[0021] (1) The morphological similarity is obtained by dynamic time-normalized distance. Peak time difference The weighted average is obtained as follows:

[0022] ,

[0023]

[0024] ,

[0025] in These are weighting coefficients. The daily average curve of the source domain. For the target domain daily average curve, For the peak hour of the source domain, For the peak hour of the target domain;

[0026] (2) The temperature sensitivity difference Represented as:

[0027]

[0028]

[0029] in The coefficients for regression of normalized power are given by the normalization factor, with the subscript s indicating the source domain and t indicating the target domain.

[0030] (3) The autocorrelation difference Represented as:

[0031]

[0032]

[0033]

[0034] Where k represents the time lag, which is the length of the time interval between two points in time. The autocorrelation coefficient is the value with a lag of k time units. For normalized power, When the standardization degree is .

[0035] Furthermore, the transferability score obtained after component normalization and weighted fusion includes:

[0036] For each individual indicator, component normalization is performed using a robust normalization formula:

[0037] ,

[0038] in For a single indicator The original value, , These are the 10th and 90th percentiles of the indicator across all candidate source domains, respectively. The function truncates the result to the interval [0,1].

[0039] The normalized individual indicators are weighted and fused to obtain the total difference. :

[0040] ,

[0041] in Weights for each indicator;

[0042] Transferability score Represented as:

[0043]

[0044] when When the time is right, the source domain is determined to be the optimal source domain.

[0045] Furthermore, the improved temporal convolutional network adopts a one-dimensional fully convolutional network structure, introducing dilated causal convolution and a dilation factor. Adopting an exponential growth strategy, i.e. , The network layer depth;

[0046] In the residual learning framework integrated into the temporal convolutional network, the residual module includes two dilated causal convolutional layers equipped with ReLU activation functions, as well as weight normalization and dropout mechanisms. A 1×1 convolutional layer is added to maintain consistency between the model output and input lengths. The residual mapping calculation formula is as follows: ,in For the first Layer input, For the first Layer parameter set, where F is the residual function.

[0047] Furthermore, the joint neighborhood adaptive loss is expressed as:

[0048]

[0049] Using cosine ascending to proceed from 0 to Weighted scheduling, The alignment is dynamically adjusted based on the validation set alignment effect; among which... The maximum mean difference loss is expressed as:

[0050]

[0051] In the above formula, k is the Gaussian kernel function. For feature space mapping function, The total number of samples contained in the source domain dataset. The total number of samples contained in the target domain dataset;

[0052] The depth-related alignment loss is expressed as:

[0053]

[0054] In the above formula , These are the feature covariance matrices of the source domain and the target domain, respectively. For the Frobenius norm, The dimension of the feature.

[0055] Furthermore, the generation of simulation load labels includes:

[0056] Based on the physical information neural network and the original load data of the target domain building air conditioning, the parameters of the known air conditioning load model are identified. The original load data includes air conditioning power, indoor and outdoor temperature data and indoor set temperature data. The model parameters include equivalent thermal resistance and equivalent heat capacity.

[0057] The identified model parameters are input into the air conditioning load model. The indoor temperature setpoint is changed while keeping other input features unchanged. The simulated load label after the demand response is calculated.

[0058] Furthermore, the day-ahead / intra-day load regulation potential assessment applied to the target domain building air conditioning cluster includes:

[0059] Collect meteorological forecast data, building basic parameters and historical load data for the next 24 hours of the target area, and input them into the load regulation capacity assessment model. The model outputs the day-ahead load regulation capacity assessment results of the building air conditioning cluster in the target area based on historical load patterns and future meteorological forecasts.

[0060] The system collects load data, real-time meteorological data, and air conditioning operating parameters of the target area building air conditioning cluster in real time. The real-time data is input into the load regulation capacity assessment model at preset time intervals. The model updates the prediction results on a rolling basis and dynamically outputs the load regulation capacity assessment results for each time period of the day.

[0061] This invention also proposes a day-ahead / intra-day load regulation capacity assessment system for building air conditioning clusters, comprising:

[0062] The data acquisition and preprocessing module is used to collect the operating data of the building air conditioning cluster, and to standardize and normalize the operating data to obtain a standardized dataset; the operating data includes open-source simulation data from the source domain and the target domain, as well as local measured data;

[0063] The filtering and adaptation module is used to construct daily vector features based on the standardized dataset, calculate individual indicators, obtain a transferability score after component normalization and weighted fusion, and filter out the optimal source domain based on the transferability score and a preset threshold; the individual indicators include morphological similarity, temperature sensitivity difference, and autocorrelation difference.

[0064] The load adjustment potential assessment model is based on an improved temporal convolutional network and integrates joint domain adaptive loss. The load adjustment capacity assessment model is trained using the standardized datasets of the optimal source domain and target domain. If the labels of the standardized dataset of the target domain are insufficient, simulated load labels are generated to reinforce the load adjustment capacity assessment model during training.

[0065] The evaluation result output module is used to apply the trained load regulation capacity evaluation model to the day-ahead / intra-day load regulation potential evaluation of the target domain building air conditioning clusters and output the evaluation results.

[0066] Beneficial effects:

[0067] (1) This invention effectively removes non-essential differences in cross-industry data by normalizing multi-dimensional data domains and filtering source domain transferability, significantly reducing the risk of negative migration, and can be adapted to the evaluation scenarios of air conditioning clusters in multi-business buildings such as commercial, office, and hotel buildings.

[0068] (2) This invention combines physical information simulation label generation with semi-supervised cross-domain training. By using PINN to identify physical parameters such as equivalent thermal resistance and heat capacity of air conditioners, it generates simulation load labels that fit the real scene. Even in scenarios where the missing label rate in the target domain exceeds 80%, it can still maintain high evaluation accuracy and break through the data dependence bottleneck of traditional supervised models.

[0069] (3) This invention improves the TCN module to take into account both long-term dependency capture and parallel computing efficiency, and achieves accurate cross-domain alignment by combining domain adaptive loss. The root mean square error (RMSE) of the evaluation results is reduced by 18%-25% compared with traditional CNN and LSTM models. The evaluation of a single building cluster can be completed within 15 minutes, meeting the needs of online applications. Attached Figure Description

[0070] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0071] Figure 2 This is a schematic diagram of the air conditioning load regulation capability evaluation architecture based on deep domain adaptive design of the present invention.

[0072] Figure 3 Weighting factors of this invention Schematic diagram of optimization analysis. Detailed Implementation

[0073] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments. Figure 1 The diagram shows a flowchart of a method for evaluating the day-ahead / intra-day load regulation capacity of a building air conditioning cluster according to the present invention, which specifically includes the following steps:

[0074] Step 1: Collect open-source simulation data and local measured data of the building air conditioning cluster through the data acquisition and preprocessing module. Perform time resolution unification, field standardization, unit unification, and missing value processing on the multi-source data to obtain a standardized dataset. Then, perform scale normalization, meteorological normalization, and morphological normalization on the standardized dataset. Specifically, this includes the following sub-steps:

[0075] (1.1) Collect two types of core data to construct a multi-source dataset:

[0076] First, there is open-source simulation data. This data is based on real HVAC systems and building prototypes and is constructed using reinforcement learning methods. It covers multiple business types such as commercial, office, and hotel. Each business type contains multiple building samples. Each building is continuously recorded for a preset duration (such as 30 days). The original time resolution is 1 minute. The data includes the total load at the control point (as a label) and input features such as indoor temperature, outdoor temperature, building area, and air conditioning set temperature.

[0077] Second, the data is based on local measurements and is taken from the operation monitoring systems of various public buildings and commercial complexes, covering a variety of building types such as complex buildings, museums, and campus buildings. The duration of continuous recording for each building is consistent with the open-source simulation data, with a sampling interval of 15 minutes. The data includes input features such as the total load of the building's total incoming line (label, unit kW), outdoor temperature, relative humidity, wind speed, day type, and total building area. It also provides metadata such as air conditioning equipment parameters, sensor and control system configurations, and building envelope information.

[0078] (1.2) Standardize the collected multi-source data:

[0079] The open-source simulation data was aggregated and resampled at 15-minute intervals to form a uniform time series of "96 sampling points per day, with a fixed data point for a continuous preset duration" for each building sample. This was consistent with the sampling interval of the local measured data. Both types of data were unified to the standard timestamps under their respective local time zones to ensure that all samples were aligned hourly on a fixed time grid. Daily type information such as "weekday / weekend" was uniformly encoded (e.g., weekdays were marked as 1 and weekends as 0), and the description criteria and classification rules of data fields were unified. Core indicators such as power, temperature, area, wind speed, and humidity were standardized in terms of units (power in kW, temperature in ℃, area in m², wind speed in m / s, and humidity in %). Shortages of no more than 3 consecutive sampling points were filled using linear interpolation of time series data. Longer gaps of more than 3 consecutive sampling points were filled with the average observation values ​​of the three days before and after the same time point, resulting in a standardized dataset.

[0080] (1.3) To address the heterogeneity of cross-industry data, three types of normalization are carried out in sequence: scale normalization, meteorological normalization, and morphological normalization.

[0081] The scale normalization prioritizes area normalization: if the standardized dataset contains reliable building usable area data, the formula is used:

[0082]

[0083] Calculation, where A represents the total power consumption (kW) and A represents the building's usable area (m²). This refers to power per unit area (kW / m²). If the building's usable area is missing or the data is unreliable, a capacity normalization method is used, and the calculation formula is as follows:

[0084]

[0085] in Rated capacity (kW) of the air conditioning system. As a relative power (dimensionless), scale normalization is used to eliminate the interference of differences between building volume and air conditioning installation scale on subsequent analysis.

[0086] (1.4) When calculating the cooling load using meteorological normalization, the following formula is used:

[0087]

[0088] in Outdoor temperature (°C). The reference temperature is ℃. The sampling interval is (h). The time above the reference temperature; then, the cooling load is standardized using the following formula:

[0089]

[0090] in , The mean and standard deviation of cooling load degrees (in °C·h) estimated on the unlabeled union set of source industries are separated from exogenous disturbances caused by climate and season by meteorological normalization.

[0091] (1.5) The robust normalization is based on quantile scaling to reduce the impact of spikes, calculates the optimal phase difference of the daily average curve of the source / target domain, adjusts the peak and valley positions, and reduces the interference of intraday load shape misalignment.

[0092] Step 2: The source domain screening and adaptation module constructs "daily vector" features from the normalized source domain and target domain data, calculates individual indicators such as morphological similarity, temperature sensitivity difference, and autocorrelation difference, and obtains a transferability score after component normalization and weighted fusion. The optimal source domain is selected based on the transferability score and a set threshold.

[0093] (2.1) After the source and target domain data have been normalized, features are constructed in the form of "daily vectors". That is, load data of 24 hours a day are spliced ​​together to form daily feature vectors, and the source domain feature matrix is ​​obtained. Target domain feature matrix ,in , These represent the number of sample days in the source domain and the target domain, respectively. The number of hours in a day. .

[0094] (2.2) Calculate the three types of unsupervised individual indicators:

[0095] First, morphological similarity is used, starting with the daily average curves of the source and target domains. , Calculate the dynamic time warping (DTW) distance Peak time difference:

[0096]

[0097] In the formula Peak hour. (Using the formula) ( The morphological similarity index is obtained by weighting. These are weighting coefficients. The daily average curve of the source domain. For the target domain daily average curve, For the peak hour of the source domain, The peak hour for the target domain.

[0098] Second, the temperature sensitivity difference, expressed in terms of standardized degrees. For normalized power Regression is performed, and we have:

[0099]

[0100]

[0101]

[0102] in The coefficients of the regression on normalized power are given by the standardization degree. The constant term in the linear regression equation represents the baseline level value of the normalized power when the normalization degree is 0; the subscript s indicates the source domain and t indicates the target domain.

[0103] Thirdly, the autocorrelation difference, calculated using the formula for the 1-24 hour autocorrelation function:

[0104]

[0105] Then, the autocorrelation error index is obtained using the formula:

[0106]

[0107] Where k represents the time lag, which is the length of the time interval between two points in time. The autocorrelation coefficient is the value with a lag of k time units. For normalized power, When the standardization degree is .

[0108] (2.3) Normalize the components of each individual indicator using a robust normalization formula:

[0109] ,

[0110] in These are the original values ​​for a single indicator. , These are the 10th and 90th percentiles of the indicator across all candidate source domains, respectively. The function truncates the results to the [0,1] interval to ensure that the indicators are comparable.

[0111] (2.4) Weighted fusion of the normalized individual indicators, and calculation of the total difference using the formula:

[0112] ( )

[0113] in Assign weights to each indicator; then, according to the formula... Calculate the transferability score, and The closer the score is to 1, the stronger the transferability. A threshold is set. Screening for transferability scores The source domain is identified as the optimal source domain, thereby reducing the negative migration risk of cross-industry migration from a data perspective.

[0114] Step 3: The load adjustment potential assessment model module is constructed with an improved temporal convolutional network as the core and an integrated joint domain adaptive loss. The model is trained using labeled data from the optimal source domain and unlabeled data from the target domain.

[0115] like Figure 2 The diagram illustrates the architecture for evaluating air conditioning load regulation capability based on deep domain adaptation according to this invention. Tc1-Tc5 are temporal feature extraction layers, responsible for extracting low- and mid-level temporal features from the input historical response data. Fc6-Fc8 are fully connected layers, mapping the features extracted by the Tc layers to a more abstract high-dimensional feature space. Tc1 corresponds to layer 1, Tc2 to layer 2, and so on up to Tc5, which corresponds to layer 5. Fc6 to Fc8 correspond to layers 6 to 8, respectively.

[0116] Figure 2The source domain data input box connects steps 1 and 2, containing both the standardized and normalized source domain data from step 1 and the optimal source domain samples selected through transferability scoring in step 2, providing a high-quality labeled data foundation for subsequent training. The target domain data input box corresponds to the same standard preprocessing in step 1, serving as the source of unlabeled data for cross-domain distribution alignment in step 3. The shared feature extraction layers (Tc1~Tc5, Fc6) are the implementation modules for improving the TCN model in step 3. The temporal feature extraction of the Tc layer matches the dilated causal convolution design in step 3, while parameter sharing echoes the cross-domain shared extractor requirement in step 3. The Fc7~Fc8 layers correspond to the feature abstraction and deepening in step 3, providing adaptive features for loss calculation. The loss calculation module directly corresponds to the total loss construction in step 3. The regression loss follows the MSE loss optimization prediction accuracy in step 3, while the joint domain adaptive loss matches the MMD and DeepCORAL loss design in step 3, driving the model parameter update. The evaluation result output box corresponds to step 4 below, which receives day-ahead and intraday data and outputs the adjustment potential results to support power demand-side decision-making.

[0117] Specifically, step 3 includes the following sub-steps:

[0118] (3.1) An improved TCN is built using a one-dimensional fully convolutional network structure, and dilated causal convolution is introduced, with the dilation factor increasing exponentially. (where i is the network layer depth). By increasing the expansion factor, the receptive field of the model is expanded, thereby effectively capturing the long-term dependence of air conditioning load.

[0119] An integrated residual learning framework is used, with each residual module containing two dilated causal convolutional layers equipped with ReLU activation functions. Weight normalization is added to promote rapid network convergence, dropout is introduced to prevent overfitting, and a 1×1 convolutional layer is added to maintain consistency between the model output and input lengths. The residual mapping calculation formula is as follows: , For the input of layer l, Let F be the parameter set of the l-th layer, and F be the residual function. An efficient time-series feature extraction module is constructed by improving TCN to extract nonlinear time-series features of air conditioning load from source and target domain data.

[0120] (3.2) Design a joint domain adaptive loss, including maximum mean difference (MMD) loss and deep correlation alignment (DeepCORAL) loss. The MMD loss calculation formula is as follows:

[0121]

[0122] in For the maximum mean difference loss, k is the Gaussian kernel function. For feature space mapping function, The total number of samples contained in the source domain dataset. The total number of samples contained in the target domain dataset;

[0123] The depth-dependent alignment loss is calculated using the following formula:

[0124]

[0125] , These are the feature covariance matrices of the source domain and the target domain, respectively. For the Frobenius norm, The dimension of the feature is the number of dimensions of the feature vector corresponding to each sample after nonlinear transformation.

[0126] (3.3) Construct the joint domain adaptive loss, the formula is:

[0127]

[0128] in Using a cosine rising strategy to schedule from 0 to , The alignment effect of the validation set is dynamically adjusted to balance the contributions of the two types of losses.

[0129] like Figure 3 The following are the weighting factors. The optimization analysis diagram illustrates the optimal performance balance. When the weight factor λ is in the range of 0.4 to 0.6, the model achieves its best performance balance—at this point, the MAE value drops to its lowest level (close to 0.05), indicating the highest prediction accuracy; the total loss is also low (approximately 1.0 to 1.5), indicating the best overall optimization effect; simultaneously, the domain adaptive loss remains within a reasonable range (approximately 0.05 to 0.10), proving that the feature distributions of the source and target domains have been effectively aligned. When λ < 0.4, the domain adaptive loss is too high, and the cross-domain alignment effect is insufficient, resulting in poor generalization ability of the model in the target domain and an increase in the MAE value. When λ > 0.6, although the domain adaptive loss further decreases, the excessive emphasis on cross-domain alignment causes the model to be biased towards adapting to the target domain distribution, ignoring the learning of prediction patterns in labeled data from the source domain, which in turn causes the MAE value to rise again and the prediction accuracy to decrease.

[0130] Figure 3 The results verify that the cosine rising scheduling λ from 0 to... The rationality of the model is demonstrated, and the optimal range of λ values ​​in practical applications is clarified, providing key parameter basis for model training. This ensures that the model can achieve cross-domain adaptation through joint domain adaptive loss and guarantee prediction accuracy through regression loss, ultimately achieving the core goal of "high-precision evaluation in cross-domain scenarios".

[0131] Using improved TCN as the core of feature extraction, this method combines joint domain adaptive loss with mean squared error (MSE) regression loss based on labeled data from the source domain:

[0132]

[0133] in For real labels, To construct a complete load adjustment potential assessment model based on the model's predicted values.

[0134] (3.4) Train the model using the optimal labeled data from the source domain and the unlabeled data from the target domain, so as to... The total loss function is used, and the Adam optimizer is employed to update the network parameters. An early-stop strategy is introduced to prevent overfitting. During training, source domain data is randomly sampled, while limited target domain data is reused to enhance the model's generalization ability, ensuring that the amount of source and target domain data is consistent in each batch.

[0135] If the target domain labels are insufficient, the Physical Information Label Generation Module is activated to supplement them. Based on the Physical Information Neural Network (PINN), and combined with the original air conditioning load data of the target domain buildings (air conditioning power, indoor and outdoor temperatures, indoor set temperature), the equivalent thermal resistance R and equivalent heat capacity C of the air conditioning load model are identified. The identified parameters are input into the air conditioning load model to simulate the demand response temperature setting strategy (changing the indoor temperature set value while keeping other input features unchanged), and the simulated load label after demand response is calculated. Combining the simulated load label and the corresponding input features, a data sample set after response is constructed, and mixed with a limited number of real labels in proportion for model reinforcement training to improve the data density of supervised training of the model.

[0136] Step 4: The trained load regulation potential assessment model is applied to the day-ahead and intraday load regulation potential assessment of the target domain building air conditioning cluster, and the assessment results are output through the assessment result output module.

[0137] (4.1) Apply the trained load regulation potential assessment model to the day-ahead assessment of the air conditioning cluster of the target domain building: collect the meteorological forecast data (outdoor temperature, relative humidity, wind speed, etc.), building basic parameters (building area, air conditioning rated capacity, etc.) and historical load data of the target domain for the next 24 hours, input them into the model, and the model outputs the day-ahead load regulation potential results of the air conditioning cluster of the target domain building based on the historical load pattern and future meteorological forecast, including key indicators such as interruptible power, adjustable duration and adjustment cost for each time period, so as to provide a reference for the load regulation potential 24 hours in advance for the formulation of power grid dispatch plan.

[0138] (4.2) Conduct intraday load regulation potential assessment: Real-time collection of load data (total power consumption, air conditioning sub-item power, etc.), real-time meteorological data (outdoor temperature, relative humidity, etc.) and air conditioning operating parameters (indoor set temperature, equipment operating status, etc.) of the target area building air conditioning cluster. Real-time data is input into the model at preset time intervals (e.g., 15 minutes). The model updates the prediction results on a rolling basis and dynamically outputs the load regulation potential for each time period of the day. It promptly reflects the impact of intraday load fluctuations and changes in the external environment on the regulation potential, and meets the real-time assessment requirements for intraday emergency demand response.

[0139] (4.3) The evaluation results are output in two forms through the evaluation result output module: a visual interface and a standardized data interface. The visual interface uses line charts, bar charts, heat maps and other forms to intuitively show the changing trend of load regulation potential at different times during the day and day, and marks key constraints (such as indoor temperature comfort range and air conditioning equipment operation safety threshold). The standardized data interface supports data interaction with the power grid dispatching system and the load aggregator management platform. The output evaluation results are packaged in a unified data format and include evaluation timestamp, building identification, regulation potential index and confidence interval, etc., which are convenient for downstream systems to call directly.

[0140] (4.4) At the same time, the assessment results are adapted to different user needs: regional-level building air conditioning cluster aggregation regulation potential is provided to the power grid dispatch center to support power grid supply and demand balance decision-making; refined regulation potential data of single buildings or building groups are provided to load aggregators to assist in the formulation of demand response implementation plans; and air conditioning system regulation potential and energy consumption optimization suggestions are provided to building property management parties to achieve efficient building energy management and ensure that the assessment results can directly serve the actual decision-making of all aspects of power demand-side management.

Claims

1. A method for evaluating the day-ahead / intra-day load regulation capacity of a building air conditioning cluster, characterized in that, include: Step 1: Collect the operational data of the building air conditioning cluster, and perform standardization and normalization processing on the operational data to obtain a standardized dataset; the operational data includes open-source simulation data of the source domain and the target domain, as well as local measured data; Step 2: Construct daily vector features based on the standardized dataset, calculate individual indicators, and obtain a transferability score after component normalization and weighted fusion. Select the optimal source domain based on the transferability score and a preset threshold. The individual indicators include morphological similarity, temperature sensitivity difference, and autocorrelation difference. Step 3: Construct a load regulation capability assessment model with an improved temporal convolutional network as the core and an integrated joint domain adaptive loss. Train the load regulation capability assessment model using the standardized datasets of the optimal source domain and target domain. If the labels of the standardized dataset of the target domain are insufficient, generate simulated load labels and reinforce the load regulation capability assessment model through training. Step 4: Apply the trained load regulation capacity assessment model to the day-ahead / intra-day load regulation potential assessment of the target domain building air conditioning cluster, and output the assessment results.

2. The load regulation capacity assessment method according to claim 1, characterized in that, The standardization process includes: uniform time resolution, uniform field caliber, uniform units, and handling of missing values; the missing value handling uses time series linear interpolation for gaps of no more than 3 consecutive sampling points, and fills gaps of more than 3 consecutive sampling points with the average of observations three days before and after the same time. The normalization process includes: Scale normalization is used to eliminate the difference between building volume and air conditioning installation scale; Meteorological normalization is used to isolate exogenous disturbances caused by climate and seasons; Morphology normalization; used to reduce intraday load morphology misalignment interference.

3. The load regulation capacity assessment method according to claim 2, characterized in that, The daily vector features are in matrix form, including the source domain feature matrix. Target domain feature matrix ,in , These represent the number of sample days in the source domain and the target domain, respectively. The number of hours in a day.

4. The load regulation capacity assessment method according to claim 3, characterized in that, Individual indicators include morphological similarity, temperature sensitivity difference, and autocorrelation difference, specifically: The morphological similarity is determined by dynamic time-normalized distance. Peak time difference The weighted average is obtained as follows: ， ， in These are weighting coefficients. The daily average curve of the source domain. For the target domain daily average curve, For the peak hour of the source domain, For the peak hour of the target domain; The temperature sensitivity difference When using standardization For normalized power Performing regression, represented as: in The coefficients of the regression on normalized power are given by the standardization factor. For the constant term of the linear regression equation, the subscript s represents the source domain and t represents the target domain; The autocorrelation difference Represented as: Where k represents time lag, which is the length of the time interval between two points in time; The autocorrelation coefficient is the result of a lag of k time units, with the subscript s indicating the source domain and t indicating the target domain.

5. The load regulation capacity assessment method according to claim 4, characterized in that, The transferability score obtained after component normalization and weighted fusion includes: For each individual indicator, component normalization is performed using a robust normalization formula: ， in For a single indicator The original value, , These are the 10th and 90th percentiles of the indicator across all candidate source domains, respectively. The function truncates the result to the interval [0,1]. The normalized individual indicators are weighted and fused to obtain the total difference. : ， in Weights for each indicator; Transferability score Represented as: when When the time is right, the source domain is determined to be the optimal source domain.

6. The load regulation capacity assessment method according to claim 5, characterized in that, The improved temporal convolutional network adopts a one-dimensional fully convolutional network structure, introducing dilated causal convolution and a dilation factor. Adopting an exponential growth strategy, , The network layer depth; In the residual learning framework integrated by the temporal convolutional network, the residual module includes two dilated causal convolutional layers equipped with ReLU activation functions, as well as weight normalization and dropout mechanisms, and adds a 1×1 convolutional layer. The residual mapping calculation formula is as follows: ,in For the first Layer input, For the first Layer parameter set, where F is the residual function.

7. The load regulation capacity assessment method according to claim 6, characterized in that, The joint domain adaptive loss is expressed as: Using cosine ascending to proceed from 0 to Weighted scheduling, The alignment is dynamically adjusted based on the validation set alignment effect; among which... The maximum mean difference loss is expressed as: In the above formula, k is the Gaussian kernel function. For feature space mapping function, The total number of samples contained in the source domain dataset. The total number of samples contained in the target domain dataset; The depth-related alignment loss is expressed as: In the above formula , These are the feature covariance matrices of the source domain and the target domain, respectively. For the Frobenius norm, The dimension of the feature.

8. The load regulation capacity assessment method according to claim 7, characterized in that, The generation of simulation load labels includes: Based on the physical information neural network and the original load data of the target domain building air conditioning, the parameters of the known air conditioning load model are identified. The original load data includes air conditioning power, indoor and outdoor temperature data and indoor set temperature data. The model parameters include equivalent thermal resistance and equivalent heat capacity. The identified model parameters are input into the air conditioning load model. The indoor temperature setpoint is changed while keeping other input features unchanged. The simulated load label after the demand response is calculated.

9. The load regulation capacity assessment method according to claim 8, characterized in that, The day-ahead / intraday load regulation potential assessment applied to the target domain building air conditioning clusters includes: Collect meteorological forecast data, building basic parameters and historical load data for the next 24 hours of the target area, and input them into the load regulation capacity assessment model. The model outputs the day-ahead load regulation capacity assessment results of the building air conditioning cluster in the target area based on historical load patterns and future meteorological forecasts. The system collects load data, real-time meteorological data, and air conditioning operating parameters of the target area building air conditioning cluster in real time. The real-time data is input into the load regulation capacity assessment model at preset time intervals. The model updates the prediction results on a rolling basis and dynamically outputs the load regulation capacity assessment results for each time period of the day.

10. A system for assessing the day-ahead / intra-day load regulation capacity of a building air conditioning cluster, characterized in that, include: The data acquisition and preprocessing module is used to collect the operating data of the building air conditioning cluster, and to standardize and normalize the operating data to obtain a standardized dataset. The operational data includes open-source simulation data from the source domain and the target domain, as well as local measured data. The filtering and adaptation module is used to construct daily vector features based on the standardized dataset, calculate individual indicators, obtain a transferability score after component normalization and weighted fusion, and filter out the optimal source domain based on the transferability score and a preset threshold; the individual indicators include morphological similarity, temperature sensitivity difference, and autocorrelation difference. The load adjustment potential assessment model is based on an improved temporal convolutional network and integrates joint domain adaptive loss. The load adjustment capacity assessment model is trained using the standardized datasets of the optimal source domain and target domain. If the labels of the standardized dataset of the target domain are insufficient, simulated load labels are generated to reinforce the load adjustment capacity assessment model during training. The evaluation result output module is used to apply the trained load regulation capacity evaluation model to the day-ahead / intra-day load regulation potential evaluation of the target domain building air conditioning clusters and output the evaluation results.

Images