Cold region residential building group energy consumption prediction method integrating deep learning and transfer learning

Abstract

The invention provides a cold region residential building group energy consumption prediction method integrating deep learning and transfer learning, and belongs to the field of residential building group energy consumption prediction. According to the method, a deep learning model, namely TabNet, based on an attention mechanism is adopted, a building energy consumption prediction task is carried out in combination with a transfer learning (TL) technology, and a TabNet-TL residential building group energy consumption prediction model is constructed. According to the method, the energy consumption of the residential building group can be accurately predicted based on short-term nonlinear data, small sample data and different feature set data, the adaptability of the residential building group energy consumption prediction model to different scenes is improved, prediction method support can be provided for the energy conservation and emission reduction design decision process of the residential building group in the cold region, and the method is suitable for popularization and application. And model support is provided for residential building grouping form design practice.

Description

Technical Field

[0001] The present invention belongs to the field of energy consumption prediction of residential building clusters, and in particular relates to an energy consumption prediction method for residential building clusters in cold regions that integrates deep learning and transfer learning. Background Art

[0002] Against the backdrop of global climate change and energy crisis, the issue of building energy consumption has attracted great attention from all over the world. The energy consumption of the building industry accounts for a large proportion of the total global energy consumption, and the energy consumption of buildings in my country accounts for about 50% of the total energy consumption. Therefore, accurately predicting building energy consumption has become a key issue in energy-saving design.

[0003] Due to the impact of the environment and the growing demand for energy resources, building energy consumption has become an important issue in my country's energy conservation and emission reduction. my country has implemented a series of building energy-saving standards and measures, and has gradually made significant progress in improving building energy efficiency through mandatory building energy-saving specifications, design, construction and operation and maintenance technology innovations, green building material innovations, energy efficiency labeling programs and financial incentives.

[0004] Although my country has made great progress in the field of building energy conservation, building energy consumption prediction still faces the following challenges: First, cold-region buildings consume a lot of energy throughout their life cycle, and the existing energy consumption prediction methods for residential building groups do not fully consider the uncertainty and complexity of the heat and mass transfer processes of residential building groups; second, when the machine learning-based building single-unit energy consumption prediction method is applied to the energy consumption prediction of residential building groups, it is still easy to cause large deviations due to error superposition, and the energy consumption prediction modeling is complex and time-consuming; finally, the existing methods rely more on learning from case data, which is not conducive to adapting to different building types or environmental changes. Summary of the invention

[0005] In view of the above situation, in order to overcome the problems in the prior art, the present invention proposes a method for predicting energy consumption of residential building clusters in cold regions that integrates deep learning and transfer learning. In view of the problems of strong data dependence, limited model generalization ability, separation of feature extraction and prediction, and limited application of transfer learning in the existing energy consumption prediction technology for residential building clusters, this scheme is based on the regional characteristics of residential building clusters in cold regions. Through long-term detection data of relevant parameters of energy consumption levels of residential building cluster spaces and residential space types, an energy consumption prediction method for residential building clusters is proposed, an energy consumption prediction model for residential building clusters is constructed, and a prediction model program that can be linked to commonly used building performance simulation platforms is developed, which is helpful to improve the adaptability of the energy consumption prediction model for residential building clusters to different scenarios, and can provide prediction method support for the energy-saving and emission reduction design decision-making process of residential building clusters, and provide model support for the morphological design practice of residential building clusters.

[0006] The technical solution adopted by the present invention is as follows: The present invention proposes a method for predicting energy consumption of residential buildings in cold regions by integrating deep learning and transfer learning, and the method comprises the following steps:

[0010] S1: Collecting building unit characteristic information and residential area characteristic information; in S1, the building unit characteristic information includes geometric and non-geometric characteristics of the building, and the residential area characteristics include residential building group characteristics and monthly energy consumption;

[0011] S2: Process the data set obtained in S1, extract key energy consumption features, coordinate the feature data set to the TabNet model, and build an energy consumption prediction model;

[0012] S3: Use part of the data for TL training and build a TabNet-TL model based on the TabNet model;

[0013] S4: Based on the prediction model obtained in step S3, an energy-saving potential assessment model is established to simulate and verify the performance of residential building groups in the same city and different cities under the same climate zone, with the same city serving as an internal verification set and different cities serving as an external verification set;

[0014] S5: Focus on adjusting the parameters that have a significant positive impact on energy consumption, and use the adjusted residential building cluster parameters as the input of the TabNet-TL model to predict the energy consumption of residential building clusters.

[0015] S6: The optimized residential building cluster energy consumption prediction data is used as the model output to obtain the residential building cluster energy consumption prediction result, and the result is compared with the data before optimization to obtain the optimization improvement rate of the TabNet-TL model.

[0016] Furthermore, the step S2 comprises:

[0017] S21: Select a wide range of residential building group energy consumption data sets and perform raw data cleaning and preprocessing;

[0018] S22: Identify key factors affecting energy consumption through careful feature selection and feature engineering;

[0019] S23: Building TabNet model based on feature selection and feature engineering;

[0020] S24: input the preprocessed data into the TabNet model after conversion, and use the TabNet model to predict energy consumption;

[0021] S25: Determine the loss function and hyperparameter optimizer, and train and verify the performance of the TabNet model.

[0022] Furthermore, the step S3 comprises:

[0023] S31: Migrate the pre-trained TabNet model to a specific target residential building group;

[0024] S32: Fine-tune the model to meet the needs of the target residential building cluster, ensure the accuracy and adaptability of the prediction, and further optimize the model performance;

[0025] S33: Evaluate the prediction results of the final model on the validation dataset and build the TabNet-TL model.

[0026] Furthermore, the step S4 comprises:

[0027] S41: Establish energy saving potential assessment model;

[0028] S42: Evaluate the accuracy of the prediction model for residential building clusters in the same city;

[0029] S43: Generalization ability assessment of residential building clusters in different cities.

[0030] The data processing includes removing missing values, outliers and duplicate records to ensure the integrity and reliability of the data. The missing values ​​are caused by sensor failure or data transmission problems, and the outliers are caused by data entry errors or equipment failures.

[0031] The dataset is based on meteorological data, building geometry data, building non-geometry data and energy consumption data. The dataset is divided into a pre-training set pretrain, a fine-tuning dataset finetune and a test set test, and a label encoding LabelEncoder is performed on a specific feature set.

[0032] The TabNet model is a deep learning model that combines the Feature transformer and the Attentive transformer. The TabNet model identifies the interactions between the features of residential building groups through the attention mechanism, and reveals the impact of building density, floor area ratio, and building type on energy consumption. Building types include point-type and plate-type, high-rise, multi-story, and low-rise.

[0033] The TL refers to transfer learning, the core idea of ​​which is to use the knowledge learned on one task to help solve another related but not completely identical task. In addition, the core structure of the transfer learning model includes the source task model, the target task model and the transfer layer. The source task model is the source domain, and the target task model is the target domain, where:

[0034] The source task model is a model pre-trained in a data-rich environment to capture general knowledge;

[0035] The target task model is suitable for new environments with less or different data;

[0036] The migration layer is responsible for transferring the knowledge in the source task and adapting it to the requirements of the target task.

[0037] The characteristic types of the residential building cluster are divided into the following four categories: residential geometric characteristics, residential non-geometric characteristics, residential area characteristics and other characteristics.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] 1. The present invention can accurately predict the energy consumption of residential building groups based on short-term nonlinear data, small sample data, and different feature set data, thereby improving the adaptability of the residential building group energy consumption prediction model to different scenarios. Compared with traditional prediction methods, this method has better prediction accuracy.

[0040] 2. The present invention can solve the problem of insufficient sample data for learning of newly built buildings or old residential buildings through sharing of energy consumption prediction data of residential building groups and migration of models. This further shows that when faced with new and unseen data, this method has stronger versatility and stability while maintaining the absolute value of prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to clearly explain the embodiments of the present invention or the technical solutions of the prior art, the following briefly introduces the drawings required for use. The drawings in the following description are only used to further understand the present invention and constitute a part of the specification, and together with the embodiments of the present invention, explain the present invention. For ordinary technicians in this field, based on the provided drawings, other drawings can be obtained without creative work.

[0042] Figure 1 This is a flow chart of the method for predicting energy consumption of residential buildings in cold regions that integrates deep learning and transfer learning.

[0043] Figure 2 This is a basic flow chart of transfer learning of the present invention. DETAILED DESCRIPTION

[0044] The present invention is described in detail below in conjunction with the accompanying drawings and specific implementation cases. This embodiment is implemented based on the technical solution of the present invention, and a detailed implementation method and a specific operation process are given. Obviously, the embodiments are only partial examples of the present invention, not all. Based on these embodiments, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present invention, but the scope of protection of the present invention is not limited to the following embodiments.

[0045] The present invention provides a method for predicting energy consumption of residential buildings in cold regions by integrating deep learning and transfer learning. Figure 1 As shown, the following steps are included:

[0046] S1: Collecting building unit characteristic information and residential area characteristic information; in step S1, the building unit characteristic information includes geometric and non-geometric characteristics of the building, and the residential area characteristics include residential building group characteristics and monthly energy consumption;

[0047] In this embodiment, 209 residential building groups in Ci County, Hebei Province were selected as cases, and monthly energy consumption data since 2018 were collected. At the same time, data of 16 typical parameters were also included, covering single buildings, group buildings, and geometric and non-geometric factors, providing rich data support for in-depth analysis and model building. The 209 residential building groups in Ci County are divided into training sets (70%), test sets (20%), and internal validation sets (10%). The external validation set selected 5 residential building groups in the neighboring Handan area. In addition to 16 typical parameters, it also contains energy consumption data for the whole year of 2022. The training set is used for pre-training of the model, learning the general features of energy consumption prediction through large-scale data; the test set is used to fine-tune the model to adapt to specific prediction tasks; the internal validation set is used to evaluate the performance of the model on the training data to monitor whether the model is overfitting and ensure its accuracy. The external validation set is used to test the performance of the model on completely different external data sets to verify its generalization ability.

[0048] S2: Process the data obtained in S1, extract key energy consumption features, coordinate the feature data set to the TabNet model, and build an energy consumption prediction model;

[0049] Specifically, S2 includes the following steps:

[0050] S21: Select a wide range of residential building group energy consumption data sets and perform raw data cleaning and preprocessing;

[0051] This method adopts a transfer learning strategy that combines pretraining and fine-tuning. The raw data cleaning and preprocessing include different stages such as data merging, feature screening, data cleaning, data conversion and data normalization to ensure the integrity and rationality of the processing process. The raw data cleaning and preprocessing methods can effectively deal with the main problems of data outliers, missing data values, inconsistent time scales and invalid detection points.

[0052] The missing data are supplemented by the following formula: A(d,t)=average(A(d-1,t),A(d+1,t),A(d,t-1),A(d,t+1))

[0053] Among them, A(d,t) is the data of month t in the dth year; A(d-1,t) is the data of month t in the d-1th year; A(d+1,t) is the data of month t in the d+1th year; A(d,t-1) is the data of month t-1 in the dth year; A(d,t+1) is the data of month t+1 in the dth year. This formula is calculated as the average value of the four moments. If any of the four data is also missing, the average value of the remaining data is taken as the missing value filling. The moment t in the formula is the recording interval of the data, and t-1 and t+1 refer to the previous recording moment and the next recording moment of the current recording moment respectively.

[0054] Data merging includes merging the input data sets and output data of different residential building groups, detecting and deleting duplicate records in the data to avoid the impact of data redundancy on model analysis, and matching the information of each building with the corresponding energy consumption data and date.

[0055] Feature screening refers to using the random forest method to rank the importance of features, evaluate the contribution of each feature to the model's predictive ability, and reduce the dimension of the data set.

[0056] Data cleaning refers to eliminating outliers in the data set through the 3σ criterion, identifying and processing outliers or noise in the data through statistical analysis methods, and preventing the model from being interfered by inaccurate data.

[0057] Data conversion refers to the assignment of values ​​to text variables in order to generate data suitable for machine learning methods, using numbers to replace the parameters represented by different text type data sets.

[0058] Data normalization processing means to make the evaluation results more intuitive, narrow the range of the data set, change the processed values ​​to the range of 0 to 1, and eliminate the dimensional and nonlinear effects between data.

[0059] The data normalization in the present invention specifically adopts the following formula:

[0060] Where Xscaled is the normalized data, and Xm and Xmax are the minimum and maximum power consumption in the training set.

[0061] S22: Identify key factors affecting energy consumption through careful feature selection and feature engineering;

[0062] In this embodiment, the following 16 general energy consumption features are selected: residential geometric features (number of residential floors, average household usage area, indoor partition ratio, total roof area, southeast northwest window-to-wall ratio), residential non-geometric features (wall U value, building orientation, door and window U value), residential features (residential type ratio, residential age, building density, volume ratio, greening rate, total number of households, geographical location, surrounding business formats). These features comprehensively consider the basic attributes, time series information, physical characteristics and some specific design parameters of the building. By selecting these features for deep transfer learning, it is ensured that the model fully learns important information about energy consumption while maintaining the generalization ability and prediction accuracy of the model.

[0063] S23: Building TabNet model based on feature selection and feature engineering;

[0064] The TabNet model can not only effectively process high-dimensional data, but also automatically select and combine features. The model is mainly implemented through pytorch_tabnet. The key components of the model include: input layer, hidden layer and output layer. The input layer defines the dimension of the input features, that is, the data processed by feature selection and feature engineering mentioned above. The hidden layer constructs multiple hidden layers, and each layer uses the ReLU activation function to introduce nonlinearity to capture the complex interactions between features. The output layer generates the predicted results, that is, the energy consumption value of the residential building group. TabNet's unique sequential attention mechanism is implemented in these layers, enabling the model to automatically select the most important features for prediction.

[0065] Specifically, the core of deep transfer learning technology is to use pre-trained deep learning models, CNN and RNN, as basic models to extract feature representations from source tasks. These feature representations can better solve the target task.

[0066] The feature representation calculation process in the source task in the present invention specifically adopts the following formula: F t =F s (θ s )+F adapt (θ adapt )

[0067] Among them, Ft is the feature representation of the target task, Fs is the feature representation of the source task, θs is the parameter of the source task model, Fadapt is the new feature representation adapted to the target task, and θadapt is the parameter of the new feature representation.

[0068] S24: input the preprocessed data into the TabNet model after conversion, and use the TabNet model to predict energy consumption;

[0069] The preprocessed data is input into the Feature Transformer Network, which can learn the internal representation of the input data and transform the original features into a form that is more beneficial to the prediction task.

[0070] The output of the feature transformation network is fed into the AttentiveTransformer. The AttentiveTransformer determines which features are most important for the current prediction task. Based on the output of the AttentiveTransformer, certain features are selected and "masked" for prediction in the current decision step.

[0071] In each decision step, TabNet makes partial decisions based on masked features and finally aggregates them into prediction outputs to generate more accurate and comprehensive energy consumption predictions.

[0072] S25: Determine the loss function and hyperparameter optimizer, and train and verify the performance of the TabNet model.

[0073] The present invention uses the Adam optimizer and a specific learning rate to optimize the model parameters, and sets a suitable batch size and virtual batch size to effectively manage memory usage and improve training efficiency. The Adam optimizer refers to a method that considers the first-order moment (mean) and second-order moment (uncentered variance) of the gradient at the same time, and assigns a different learning rate to each model parameter.

[0074] The present invention performs self-supervisory pre-training on the model. Pre-training refers to training the TabNet model through a large-scale data set to enable it to have basic energy consumption prediction capabilities. Selecting the appropriate mask type and setting the number of decoders is to adjust the model's attention mechanism and feature learning ability. Use TabNetPretrainer for pre-training, select entmax as the mask type, and enhance the model's ability to select features.

[0075] The present invention uses a verification method to evaluate the performance of the model, ensure the accuracy of the prediction results and the generalization ability of the model, and finally complete the establishment of the TabNet model.

[0076] The Adam optimizer used in the present invention specifically adopts the following formula: a t =a0×sqrt(1-B2) / (1-B1)

[0077] Among them, α0 is the initial learning rate, t represents the current number of iterations, β1 and β2 are the exponential decay rates of the first-order moment estimate and the second-order moment estimate of the parameter update.

[0078] S3: Use part of the data for TL training and build a TabNet-TL model based on the TabNet model;

[0079] Specifically, S3 includes the following steps:

[0080] S31: Migrate the pre-trained TabNet model to a specific target residential building cluster.

[0081] When constructing the prediction model for the target urban area, TabNet adjusts the importance of features based on data relevance. By introducing the data of the target urban area, the model realizes the transition between the source domain and the target domain, dynamically adjusts the importance of the input layer features, and makes the model of the source domain closer to the target domain.

[0082] In this embodiment, the basic process of transfer learning includes: 1) obtaining a feature data set; 2) splitting the pre-training data; 3) training a deep learning model and evaluating the prediction results; 4) judging whether the transfer learning conditions are satisfied, if not, returning to the previous step, if satisfied, proceeding to the next step; 5) processing the downstream fine-tuning data set to improve the prediction model and obtain the energy consumption prediction results; 6) evaluating the final prediction results and proposing corresponding energy-saving transformation strategies.

[0083] S32: Fine-tune the model to meet the needs of the target residential building cluster, ensure the accuracy and adaptability of the prediction, and further optimize the model performance;

[0084] After selecting the TabNet model, design and implement a fine-tuning strategy to optimize the performance of the pre-trained model and ensure that it can adapt to the specific target task. Fine-tuning refers to adjusting some layers in the pre-trained model to better adapt to the target task. Finally, the knowledge of the pre-trained model is transferred to the fine-tuned model to enhance the generalization ability of the model.

[0085] S33: Evaluate the prediction results of the final model on the validation dataset and build the TabNet-TL model.

[0086] The accuracy change assessment of the present invention specifically adopts the following formula:

[0087] Where Rd is the accuracy change rate, Ri is the evaluation index R of the internal validation set 2 The result of Re is the evaluation index R of the external validation set. 2 The above formula is used to quantify the accuracy variation of the model on different data sets.

[0088] This paper selects 5 residential building groups as the external validation set and calculates the R 2The value, combined with the above formula, can quantitatively evaluate the effect of transfer learning on the generalization ability of the model, and finally obtain the TabNet-TL energy consumption prediction model. Generalization ability means that the model can not only show high-precision prediction on the training set or internal validation set, but also maintain stable prediction performance on unseen data (external validation set). The accuracy of the model is evaluated on the validation data set.

[0089] S4: Based on the prediction model obtained in step S3, an energy-saving potential assessment model is established to simulate and verify the performance of residential building groups in the same city and different cities under the same climate zone, with the same city being the internal verification set and the different cities being the external verification set;

[0090] Specifically, S4 includes the following steps:

[0091] S41: Establish a model for evaluating the energy saving potential of residential building clusters;

[0092] Based on the identified key energy consumption influencing factors, a quantitative model is established to evaluate the potential energy-saving effects of different energy-saving measures. The model takes into account the specific contribution of each influencing factor to energy consumption, as well as the proportion of energy consumption reduction achieved by implementing specific energy-saving measures. Using the established energy-saving potential evaluation model, different energy-saving measures are quantitatively analyzed, and the energy-saving potential of each measure is evaluated by simulating the changes in energy consumption after the implementation of these measures. The energy-saving potential evaluation includes the prediction model accuracy evaluation and generalization ability evaluation.

[0093] S42, the prediction model accuracy is evaluated for residential building clusters in the same city (internal validation set);

[0094] In this example, 10% of the Cixian dataset was used as an internal validation set to evaluate the model accuracy. A variety of prediction models (including random forest, XGBoost, linear regression, Lasso linear regression, multi-layer perceptron (MLP), TabNet and its transfer learning optimized version TabNet-TL) were used to evaluate the energy consumption of residential building clusters. 2 ) three key indicators, and conducted a comprehensive model accuracy evaluation.

[0095] The calculation process of the mean absolute error (MAE) in the present invention specifically adopts the following formula:

[0096] The calculation process of the root mean square error (RMSE) in the present invention specifically adopts the following formula:

[0097] The coefficient of determination (R 2) is calculated using the following formula:

[0098] S43, generalization ability evaluation of residential building clusters in different cities (external validation set).

[0099] In this embodiment, five typical residential building groups in Handan City are selected as external validation sets to evaluate the generalization ability of the model, mainly for TabNet-TL and XGBoost for small sample validation, and the control group is the result of direct prediction without training, and the selected models are TabNet-TL and XGBoost. When evaluating generalization ability, prediction accuracy is a key indicator. The present invention further evaluates the generalization ability of the model by comparing the difference in prediction accuracy between the model in the training set or internal validation set and the external validation set.

[0100] The model accuracy change evaluation in the present invention specifically adopts the following formula:

[0101] Where Rd is the accuracy change rate, Ri is the evaluation index R of the internal validation set 2 The result of Re is the evaluation index R of the external validation set. 2 result.

[0102] S5: Focus on adjusting the parameters that have a significant positive impact on energy consumption, and use the adjusted building parameters as the input of the TabNet-TL model to predict the energy consumption of residential building groups.

[0103] This embodiment records the corresponding residential building groups and related energy consumption parameters before and after optimization. The optimization is through the coefficients of the linear regression model, reflecting the degree of influence of each feature on energy consumption. A positive coefficient indicates that energy consumption increases when the characteristic value increases, and a negative coefficient indicates that energy consumption decreases when the characteristic value increases. The present invention focuses on adjusting the parameters that have a significant positive impact on energy consumption, such as reducing the window-to-wall ratio, reducing the U value of walls and doors and windows to improve thermal insulation performance, optimizing the building shape coefficient to an appropriate range (such as below 0.4), and reducing the volume ratio. At the same time, the present invention supports the above optimization measures by increasing the greening rate and building density of the site. After optimization and adjustment, this embodiment predicts the optimized energy consumption through deep transfer learning.

[0104] S6: The optimized residential building cluster energy consumption prediction data is used as the model output to obtain the residential building cluster energy consumption prediction result, and the result is compared with the data before optimization to obtain the optimization improvement rate of the TabNet-TL model.

[0105] According to the optimization adjustment, the present invention predicts the energy consumption before and after the optimization by deep transfer learning, and compares the energy consumption prediction results, which is helpful for formulating energy-saving optimization strategies for each residential building group. The present invention helps to reduce the annual total energy consumption of residential building groups and improve overall energy efficiency.

[0106] The above description of the present invention and its embodiments is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art can design similar structures and embodiments without creativity based on the inspiration without departing from the purpose of the present invention, they should all fall within the protection scope of the present invention.

Claims

1. A method for predicting energy consumption of residential buildings in cold regions that integrates deep learning and transfer learning, characterized in that: The method comprises the following steps: S1: Collecting building unit characteristic information and residential area characteristic information; in S1, the building unit characteristic information includes geometric and non-geometric characteristics of the building, and the residential area characteristics include residential building group characteristics and monthly energy consumption; S2: Process the data set obtained in S1, extract key energy consumption features, coordinate the feature data set to the TabNet model, and build an energy consumption prediction model; S3: Use part of the data for TL training and build a TabNet-TL model based on the TabNet model; S4: Based on the prediction model obtained in S3, an energy-saving potential assessment model is established to simulate and verify the performance of residential building groups in the same city and different cities under the same climate zone. The same city is the internal verification set, and the different cities are the external verification set. S5: Focus on adjusting the parameters that have a significant positive impact on energy consumption, and use the adjusted residential building cluster parameters as the input of the TabNet-TL model to predict the energy consumption of residential building clusters; S6: The optimized residential building cluster energy consumption prediction data is used as the model output to obtain the residential building cluster energy consumption prediction result, and the result is compared with the data before optimization to obtain the optimization improvement rate of the TabNet-TL model.

2. The method according to claim 1, characterized in that The step S2 comprises: S21: Select a wide range of residential building group energy consumption data sets and perform raw data cleaning and preprocessing; S22: Identify key factors affecting energy consumption through careful feature selection and feature engineering; S23: Building TabNet model based on feature selection and feature engineering; S24: input the preprocessed data into the TabNet model after conversion, and use the TabNet model to predict energy consumption; S25: Determine the loss function and hyperparameter optimizer, and train and verify the performance of the TabNet model.

3. The method according to claim 1, characterized in that The step S3 comprises: S31: Migrate the pre-trained TabNet model to a specific target residential building group; S32: Fine-tune the model to meet the needs of the target residential building cluster, ensure the accuracy and adaptability of the prediction, and further optimize the model performance; S33: Evaluate the prediction results of the final model on the validation dataset and build the TabNet-TL model.

4. The method according to claim 1, characterized in that: The step S4 comprises: S41: Establish a model for evaluating the energy saving potential of residential building clusters; S42: Evaluate the accuracy of the prediction model for residential building clusters in the same city; S43: Generalization ability assessment of residential building clusters in different cities.

5. The method according to claim 1, characterized in that: The data processing includes removing missing values, outliers and duplicate records to ensure the integrity and reliability of the data. The missing values ​​are caused by sensor failure or data transmission problems, and the outliers are caused by data entry errors or equipment failures.

6. The method according to claim 1, characterized in that The dataset is based on meteorological data, building geometry data, building non-geometry data and energy consumption data. The dataset is divided into a pre-training set pretrain, a fine-tuning dataset finetune and a test set test, and a label encoding LabelEncoder is performed on a specific feature set.

7. The method according to claim 1, characterized in that The TabNet model is a deep learning model that combines Feature transformer and Attentive transformer. The TabNet model identifies the interaction between building features through the attention mechanism and reveals the impact of building density, volume ratio, and building type on energy consumption. Building types include point-type and plate-type, high-rise, multi-story, and low-rise.

8. The method according to claim 1, characterized in that: The TL refers to transfer learning, the core idea of ​​which is to use the knowledge learned on one task to help solve another related but not identical task. In addition, the core structure of the transfer learning model includes the source task model, the target task model and the transfer layer. The source task model is the source domain, and the target task model is the target domain, where: The source task model is a model pre-trained in a data-rich environment to capture general knowledge; the target task model is suitable for new environments with less or different data; the migration layer is responsible for transferring the knowledge in the source task and adapting to the needs of the target task.

9. The method according to claim 1, characterized in that: The characteristic types of the residential building cluster are divided into the following four categories: residential geometric characteristics, residential non-geometric characteristics, residential area characteristics and other characteristics.

Images