Building air conditioner cooling load prediction method based on RFE-VaDE-XGBoost feature selection and HN-Adam-STGAT

Abstract

The invention discloses a building air conditioner cooling load prediction method based on RFE-VaDE-XGBoost feature selection and HN-Adam-STGAT, and the method comprises the following steps: 1, carrying out the preprocessing and cleaning of data; 2, determining the optimal number of input parameters of the prediction model by adopting an RFE algorithm; 3, performing clustering analysis on the cleaned data by using a VaDE algorithm; 4, based on results of the step 2 and the step 3, factors which have small influence on the cold load are eliminated in combination with an XGBoost algorithm; 5, dynamically adjusting the learning rate in the STGAT model training process by using an HN-Adam algorithm; and 6, predicting the cooling load of the air conditioner by using the adjusted model. According to the method, the problems of low prediction speed, insufficient precision and the like in a traditional building air conditioner cooling load prediction method are effectively solved, and an efficient and accurate new method is provided for air conditioner cooling load prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of building cooling load prediction, and particularly relates to a building air-conditioning cooling load prediction method based on RFE-VaDE-XGBoost feature selection and HN-Adam-STGAT. Background Technique

[0002] In recent years, the global building energy consumption has shown a rapid growth trend, and the energy consumed by buildings has accounted for more than one-third of the global total energy. Among them, the energy consumption of commercial buildings has been steadily increasing, and the heating, ventilation, and air conditioning (HVAC) system accounts for 40% - 50% of the total energy consumption of commercial buildings. Therefore, reducing the energy consumption of the HVAC system is of great significance for saving global energy, and the key to improving the efficiency of the HVAC system lies in accurately predicting the cooling load of building air conditioners.

[0003] At present, building cooling load prediction mainly relies on two methods: physics-based modeling methods and data-driven methods. Physics-based modeling methods describe the heat transfer process of buildings through thermodynamic principles and calculate the cooling load accordingly, which can objectively reflect the heat exchange situation inside and outside the building. However, this method usually requires a large number of building physical parameters and professional knowledge, and the modeling process is complex and time-consuming, limiting its application in real-time prediction and optimal control. In contrast, data-driven methods can build efficient and accurate prediction models only relying on building operation data, and with the continuous improvement of building energy management systems, the collection of cooling load data has become more convenient, making it more widely used in practical applications. Nevertheless, existing building air-conditioning cooling load prediction models still have certain limitations. For example, Mao Yun et al. proposed a method based on the nonlinear chaotic harris hawk optimization (NCHHO) algorithm to optimize the fully Elman neural network (FENN) for commercial building cooling load prediction. This method improves the accuracy of the prediction model to a certain extent, but due to the relatively complex calculation process of NCHHO and the long training time, it limits its application in more complex scenarios (Yun Mao, Junqi Yu, Na Zhang, et al. A hybrid model of commercial building cooling load prediction based on the improved NCHHO-FENN algorithm. Journal of Building Engineering, 2023, 78: 107660). On the other hand, Huang Xiaofei et al. proposed a hybrid model based on empirical mode decomposition (EMD) and an improved long short-term memory (LSTM) network with Markov chain for building cooling load prediction. However, this model fails to fully consider the importance differences of the characteristics corresponding to different building air-conditioning cooling load intervals, resulting in being affected by noise during the feature reduction process, thus reducing the accuracy of the prediction model (Xiaofei Huang, Yangming Han, Junwei Yan, et al. Hybrid forecasting model of building cooling load based on EMD-LSTM-Markov algorithm. Energy and Buildings, 2024, 321: 114670,).In summary, while existing methods improve prediction accuracy, they still suffer from problems such as high computational complexity, long training time, or insufficient feature extraction. Therefore, there is an urgent need for an efficient, stable, and accurate building cooling load prediction method to overcome challenges such as the insufficient local search ability of traditional artificial neural networks, the tendency to fall into local optima, and the overly long training time, so as to better apply to complex practical application scenarios. Summary of the Invention

[0004] The present invention aims to provide a building air-conditioning cooling load prediction method based on RFE-VaDE-XGBoost feature selection and HN-Adam-STGAT to solve problems in the prior art such as insufficient local search ability of the prediction model, the tendency to fall into local optima, and the overly long training time, thereby significantly improving prediction accuracy and response speed.

[0005] To solve the above problems, the method adopted by the present invention is to first use the recursive feature elimination (RFE) algorithm to determine the optimal number of input parameters of the prediction model, and use the variation deep embedding (VaDE) algorithm to perform clustering analysis on the data set after feature reduction; subsequently, according to the results of the RFE and VaDE algorithms, the factors with less influence on the building cooling load are eliminated through the extreme gradient boosting (XGBoost); finally, during the training process of the spatial-temporal graph attention networks (STGAT), the Hybrid-Norm Adaptive Moment Estimation (HN-Adam) algorithm is introduced to dynamically adjust the learning rate, thereby significantly improving the speed and accuracy of building air-conditioning cooling load prediction.

[0006] To achieve the above technical objectives, the present invention specifically adopts the following technical solutions:

[0007] A building air-conditioning cooling load prediction method based on RFE-VaDE-XGBoost feature selection and HN-Adam-STGAT, comprising the following steps:

[0008] Step 1: Preprocess and clean the data used for feature selection and prediction;

[0009] Step 2: Use the RFE algorithm to determine the optimal number of input parameters of the prediction model;

[0010] Step 3: Use the VaDE algorithm to perform clustering analysis on the cleaned data;

[0011] Step 4: Based on the results of Step 2 and Step 3, combine the XGBoost algorithm to eliminate the factors with less influence on the building cooling load;

[0012] Step 5: Use the HN-Adam algorithm to dynamically adjust the learning rate during the training process of the STGAT model;

[0013] Step 6: Use the model optimized in Step 5 to predict the building air-conditioning cooling load.

[0014] Furthermore, in Step 1, use the density-based spatial clustering of applications with noise (DBSCAN) algorithm to detect and process outliers in the dataset. The process includes the following steps:

[0015] (1) Set the parameters of the DBSCAN algorithm;

[0016] (2) Perform density clustering on the building cooling load dataset to detect and process abnormal data.

[0017] Furthermore, in Step 2, use the RFE algorithm to determine the optimal number of input parameters of the prediction model. The process includes the following steps:

[0018] (1) Sort all the cooling load influence factors according to the importance scores of each feature;

[0019] (2) Sequentially form feature subsets with higher-scoring influence factors;

[0020] (3) Use GBRT to perform cross-validation on different feature subsets, and select the feature subset with the highest cross-validation score. The number of features it contains is the optimal number of input parameters of the prediction model.

[0021] Furthermore, in Step 3, use the VaDE algorithm to perform clustering analysis on the cleaned data. The process includes the following steps:

[0022] (1) Preset the air-conditioning cooling load dataset to be divided into K clusters;

[0023] (2) Calculate the cluster category to which each cooling load data point belongs;

[0024] (3) Evaluate the clustering effect through the reconstruction error to determine the final clustering result.

[0025] Furthermore, in Step 4, based on the results of Step 2 and Step 3, combine the XGBoost algorithm to eliminate the factors with less influence on the building cooling load. The process includes the following steps:

[0026] (1) Set the parameters of the XGBoost algorithm;

[0027] (2) Use the XGBoost algorithm to evaluate the importance of the influencing factors in the cooling load data of different types of building air conditioners;

[0028] (3) Weight the scores of the influencing factors of each category to obtain the final importance score;

[0029] (4) Select the top-ranked key influencing factors according to the number of optimal input parameters determined in step 2.

[0030] Furthermore, the XGBoost algorithm formula:

[0031]

[0032] F = {f(v) = w q(v)} (2)

[0033] Where, is the predicted value of the cooling load of the i-th building, v i is the feature vector of the influencing factors of the building cooling load, F is the space of regression trees, f k represents the k-th regression tree, K represents the number of regression trees included in the integrated model, q represents the tree structure that maps the building cooling load samples to the corresponding leaf nodes, and T is the total number of leaf nodes in the tree.

[0034] Furthermore, the process of using the HN-Adam algorithm to adjust the learning rate during the training of the STGAT model in step 5 includes the following steps:

[0035] (1) Initialize the momentum for recording the gradient history information according to the basic operation of the Adam optimizer;

[0036] (2) Calculate the parameters to be optimized in the STGAT neural network;

[0037] (3) Dynamically adjust the learning rate at each parameter update according to the gradient change amplitude;

[0038] (4) Calculate the dynamically adjusted normalization factor, determine the final learning rate, and thus obtain the optimized STGAT prediction model.

[0039] Furthermore, STGAT is a neural network that introduces the attention mechanism into the Transformer model.

[0040] Furthermore, the STGAT neural network projects the features affecting the cooling load prediction value into three vectors: query (Q), key (K), value (V), to calculate the similarity between the cooling load influencing factors and calculate the weighted output based on this.

[0041] Further, in Step 6, the optimized model obtained in Step 5 is used for cold load prediction, and the process includes the following steps:

[0042] (1) Divide the prediction data set into a training set and a test set;

[0043] (2) Set the parameters of the prediction model;

[0044] (3) Train the prediction model and test its performance.

[0045] Compared with the prior art, the present invention has the following technical advantages:

[0046] (1) Efficient processing of data outliers

[0047] Compared with the prior art, the present invention uses the DBSCAN algorithm to detect and process outliers in the building air-conditioning cold load data set, and automatically selects the optimal number of clusters in combination with the CH index. The DBSCAN algorithm can effectively identify and classify abnormal building air-conditioning cold load data points, improving the accuracy of the data; the CH index can automatically determine the number of clusters, ensuring data accuracy and clustering quality. This method can effectively remove outliers in the building air-conditioning cold load data set.

[0048] (2) Automatically determine the input parameters of the prediction model

[0049] Compared with the prior art, the present invention uses the RFE algorithm to determine the optimal number of input parameters of the prediction model. RFE is a feature selection algorithm based on a greedy strategy, which relies on a machine learning model to construct a classifier and evaluate the importance of each predictor. Its core idea is to iteratively remove features that contribute less to the model, automatically screen out the optimal feature subset, and thus determine the optimal number of input parameters of the prediction model. Compared with the traditional method that relies on expert experience to select input parameters, RFE can effectively reduce human intervention, avoid problems of improper feature selection caused by empirical biases, and improve the stability and reliability of the prediction model. This automated feature screening mechanism not only simplifies the modeling process, but also enhances the generalization ability of the model, making it more adaptable in different application scenarios.

[0050] (3) Innovative clustering method reduces data interference

[0051] Compared with the prior art, the present invention uses the VaDE algorithm to perform clustering analysis on the data set to optimize the feature selection process of the cooling load prediction model. Compared with traditional shallow clustering methods such as K-means and spectral clustering, VaDE can automatically learn deep feature representations, more accurately capture the internal structure of the data, thereby improving the clustering effect and data analysis ability. In the current research on cooling load prediction, the mutual interference of air-conditioning cooling load data in different time intervals or spatial regions is often ignored, which has a greater impact on the subsequent feature selection process. Therefore, the present invention uses the VaDE algorithm to perform clustering analysis on the data set, effectively distinguishing the cooling load data in different intervals and reducing the feature selection error caused by data mixing. At the same time, this method can reduce the influence of noise in the feature dimensionality reduction process, improve the accuracy of data processing, and thus enhance the stability and reliability of the cooling load prediction model.

[0052] (4) Accurate importance assessment of influencing factors

[0053] Compared with the prior art, the present invention combines the analysis results of the RFE and VaDE algorithms to construct a more efficient building cooling load prediction method. First, the XGBoost algorithm proposed by the University of Washington is used to train the influencing factors in the cooling load using its decision trees and regression trees, thereby mining the complex relationships between features. Then, a three-dimensional evaluation system is introduced to comprehensively analyze the interaction between features and deeply evaluate the performance of different feature subsets to achieve a more accurate ranking of the importance of influencing factors. Finally, based on the optimal input parameters determined by the RFE algorithm, the features with less influence on the prediction results are removed, thereby achieving effective dimensionality reduction and data compression, significantly improving the computational efficiency and accuracy of the cooling load prediction. This optimization strategy not only reduces redundant data, improves the generalization ability of the model, but also greatly improves the response speed of subsequent predictions, making it more suitable for actual application scenarios.

[0054] (5) High-precision cooling load prediction model

[0055] Compared with the prior art, the present invention uses STGAT as the basic model to capture the spatio-temporal dynamic dependence of building air-conditioning cooling loads. This algorithm combines a fully connected layer and a multi-head attention mechanism for feature modeling. Among them, the multi-head attention mechanism can quickly capture the long-distance time dependence between influencing factors, while the convolution operation can aggregate adjacent information to further improve the accuracy of cooling load prediction. In addition, in order to dynamically adjust the learning rate of the STGAT neural network, the present invention introduces the HN-Adam optimizer into the prediction model. This optimizer can adaptively adjust the learning rate according to the current and historical gradients, effectively avoiding the training bottleneck caused by a fixed learning rate, thereby greatly improving the prediction performance of the model. This optimization strategy not only improves the computational efficiency and convergence speed of the cooling load prediction, but also enhances the adaptability of the model to complex application scenarios. Brief Description of the Drawings

[0056] Figure 1 This is the flowchart of the working process of the building air-conditioning cooling load prediction method based on RFE-VaDE-XGBoost feature selection and HN-Adam-STGAT of the present invention;

[0057] Figure 2 This is the result diagram of the present invention using the RFE algorithm to determine the optimal number of input parameters of the cooling load prediction model;

[0058] Figure 3 This is the result diagram of the present invention using the VaDE algorithm to perform cluster analysis on the building air-conditioning cooling load data for feature reduction;

[0059] Figure 4 This is the effect diagram of data processing after feature reduction of the present invention;

[0060] Figure 5 This is the flowchart of the specific application of the STAGT algorithm of the present invention;

[0061] Figure 6 This is the analysis and comparison diagram of the prediction results for highlighting feature reduction of the present invention;

[0062] Figure 7 This is the analysis and comparison diagram of the prediction results for highlighting the HN-Adam optimizer of the present invention;

[0063] Figure 8 This is the diagram of the comparison of the final prediction results of the prediction model of the present invention. Detailed Embodiment

[0064] In order to make the objectives, technical solutions and advantages of the present invention clearer and more distinct, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with specific embodiments and with reference to the accompanying drawings. It should be noted that the described embodiments of the present invention are illustrative, but not restrictive of the present invention. Therefore, the present invention is not limited to the above embodiments. Based on the principles of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts are considered to be within the protection scope of the present invention.

[0065] The technical solution of the building air-conditioning cooling load prediction method based on RFE-VaDE-XGBoost feature selection and HN-Adam-STGAT of the present invention is mainly divided into three parts, as Figure 1As shown below. First, to avoid the influence of abnormal data points in the dataset, the dataset used for feature reduction and prediction is preprocessed and cleaned. Second, the RFE algorithm is used to determine the optimal number of input parameters for the prediction model, and the VaDE clustering method is used to perform clustering analysis on the data for feature reduction. Based on the XGBoost algorithm, combined with the Shapley additive explanations (SHAP) interaction value, mean squared error (MSE) value, and coefficient of determination (R 2 ) for three-dimensional evaluation, thus effectively reducing the influencing factors that contribute little or are redundant to the cooling load prediction, and obtaining an optimal data sample set with non-redundant features. Finally, the prediction performance of the HN-Adam-STGAT model is verified using this optimal data sample set, fully demonstrating the effectiveness of the proposed solution of the present invention. The method of the present invention includes the following steps:

[0066] Step 1: Preprocess and clean the data used for feature selection and prediction;

[0067] Step 2: Use the RFE algorithm to determine the optimal number of input parameters for the prediction model;

[0068] Step 3: Use the VaDE algorithm to perform clustering analysis on the cleaned data;

[0069] Step 4: Based on the results of Step 2 and Step 3, combine the XGBoost algorithm to eliminate the factors with little influence on the building cooling load;

[0070] Step 5: Use the HN-Adam algorithm to dynamically adjust the learning rate during the training process of the STGAT model;

[0071] Step 6: Use the model optimized in Step 5 to predict the building air-conditioning cooling load.

[0072] In Step 1: Clean the data used for feature selection and prediction.

[0073] Principle of the DBSCAN algorithm:

[0074] DBSCAN is a density-based clustering algorithm. Its core idea is to judge whether a building cooling load data sample belongs to a certain clustering cluster by setting the neighborhood parameter eps and the minimum number of samples. When a cooling load data sample contains at least the preset minimum number of samples within the eps neighborhood, this sample is regarded as a core point. DBSCAN forms complete clustering clusters by continuously expanding the neighborhood of the core point, and can automatically identify and exclude the outliers in the building air-conditioning cooling load dataset, thereby improving the accuracy and reliability of data processing.

[0075] DBSCAN algorithm formula:

[0076] (1) Core point determination: If a sample point of building cooling load data contains MinPts sample points, then this sample point of cooling load data is called a core point.

[0077] |N(p)|≥MinPts (3)

[0078] Among them, N(p) = {q|d(p,q)≤ε} represents the neighborhood of the sample point p of building cooling load data, and d(p,q) represents the distance between the sample points p and q of cooling load data.

[0079] (2) Cluster construction: Starting from the core point, add the points in its neighborhood to the same cluster, and repeat expanding the neighborhood until all neighborhoods of the core points are included.

[0080] The specific steps are as follows:

[0081] In the present invention, in order to detect and process outliers in the building air-conditioning cooling load data set, first set the parameters of the DBSCAN algorithm: the neighborhood parameter eps is set to 0.5, and the minimum number of samples is set to 5. The building cooling load sample data includes 14 cooling load influencing factors and the actual value of the building cooling load. These 14 influencing factors are respectively: the previous moment's air-conditioning cooling load (t-1), the current moment's outdoor humidity (t), the previous moment's outdoor humidity (t-1), the current moment's outdoor temperature (t), the previous moment's outdoor temperature (t-1), the current moment's solar radiation intensity (t), the previous moment's solar radiation intensity (t-1), the current moment's rainfall (t), the current moment's wind speed (t), the current moment's air pressure (t), the current moment's indoor humidity (t), the current moment's indoor temperature (t), equipment heat generation, and lighting heat generation. Using the DBSCAN algorithm to perform density clustering on the above sample data can detect and eliminate outliers, thereby improving the accuracy of subsequent feature reduction and prediction.

[0082] In step two, use the RFE algorithm to determine the optimal number of input parameters of the prediction model.

[0083] Principle of the RFE algorithm:

[0084] The RFE algorithm is a greedy algorithm for finding the optimal variable subset, and its core lies in relying on a machine learning algorithm to construct a classifier and evaluate the importance of each predictor. In the present invention, the gradient boosting regression tree (GBRT) is selected as the model classifier in the RFE process.

[0085] The RFE mainly includes three steps:

[0086] (1) Sort all the cooling load impact factors according to their importance scores, and sequentially add the impact factors with higher scores to the feature subset;

[0087] (2) Use this feature subset to train the GBRT model and obtain the model score through cross-validation;

[0088] (3) Repeat the above process for the remaining impact factors to obtain the cross-validation scores of all feature subsets. Finally, select the feature subset with the optimal cross-validation score, and the corresponding number of features is the optimal number of input parameters for the prediction model.

[0089] The specific steps are as follows:

[0090] Determine the optimal number of input parameters through cross-validation, and the results are as Figure 2 shown. When the number of features is 9, the model performance is optimal, and the MSE value of cross-validation reaches the minimum, indicating that 9 features can not only ensure the prediction accuracy of the model but also effectively reduce the complexity of the building air-conditioning cooling load prediction model, laying a foundation for the subsequent feature reduction process.

[0091] In step three, the VaDE algorithm is used to perform clustering analysis on the data for feature selection.

[0092] Principle of the VaDE algorithm:

[0093] The VaDE algorithm cleverly combines the deep representation learning ability of the variational autoencoder (VAE) with the probability clustering advantage of the Gaussian mixture model (GMM), effectively making up for the deficiencies of traditional clustering methods in learning and generation capabilities. This algorithm can automatically learn deep feature representations beneficial for clustering and quickly capture the internal structure of the data, so it is very suitable for clustering analysis of high-dimensional cooling load data sets.

[0094] The specific steps are as follows:

[0095] Select a data sample x D from the air-conditioning cooling load data set R i for feature reduction, select a cluster c ∼ Cat(π) from K' clusters, and represent the internal features of the cooling load data sample x through the latent vector z, where D Calculate the mean vector of the data distribution of the sample x i and the covariance matrix through formula (4);

[0096]

[0097] Where K' is a predefined parameter, f(z; θ) is a neural network for calculating the mean and variance of cooling load data, with input z and parameterized by θ, where θ represents the weights and biases of the neural network. I is an identity matrix. Cat(π) is a categorical distribution parameterized by π, where π represents the probability of selecting cluster c, and μ c , represents the mean vector and covariance matrix of the Gaussian distribution corresponding to cluster c. is the Gaussian distribution parameterized by μ c , σ c .

[0098] According to the above VaDE process, the joint probability p(x, z, c) can be decomposed as:

[0099] p(x, z, c) = p(x|z)p(z|c)p(c) (5)

[0100] where x is a sample of building air-conditioning cooling load data, z is the intrinsic feature of sample x, and c is a clustering result of air-conditioning cooling load we selected.

[0101] Since x and c are independent given z, the probability is defined as:

[0102] p(c) = Cat(c|π) (6)

[0103]

[0104] where Ber(x|μ x ) is the multivariate Bernoulli distribution parameterized by μ x .

[0105] Determine the cluster category to which each cooling load data point belongs by calculating the posterior probability p(z|c); subsequently, evaluate the clustering effect of the air-conditioning cooling load sample set through the reconstruction error, and its calculation method is to statistically calculate the error of the distance between the building cooling load data points within each category and the corresponding cluster center, and the specific calculation process is shown in formula (9).

[0106]

[0107] where represents the total number of clusters of cooling load data, represents all data points of the -th cluster, is the mean of the -th cluster.

[0108] The specific steps are as follows:

[0109] The VaDE algorithm uses the reconstruction error as the objective function to evaluate the clustering effect, which can intuitively show the impact of different K values on the data reconstruction quality. Figure 3 The classification results of the air-conditioning cooling load are shown: when the K value increases from 1 to 6, the reconstruction error generally shows a downward trend, while when K>6, the reconstruction error tends to be stable and only fluctuates within a small range. It should be noted that when K = 2, there is an obvious "elbow" drop in the reconstruction error curve. According to the "elbow method" and the actual application requirements of the clustering results, K = 2 is selected as the optimal number of classifications. Therefore, when using the VaDE algorithm to classify the air-conditioning cooling load data, the building air-conditioning cooling load data set is divided into two categories: high load and low load.

[0110] In step four: Based on the results of step two and step three, the XGBoost algorithm is used to eliminate the unimportant influencing factors in the building cooling load.

[0111] Principle of the XGBoost algorithm:

[0112] The XGBoost algorithm is an efficient, accurate and widely used ensemble learning algorithm optimized based on the gradient boosting decision tree (GBDT).

[0113] Formula of the XGBoost algorithm:

[0114]

[0115] F = {f(v) = w q ( v )} (11)

[0116] Among them, is the predicted value of the i-th building cooling load, v i is the feature vector of the influencing factors of the building cooling load, F is the space of the regression tree. f k represents the k-th regression tree. K represents the number of regression trees included in the ensemble model. q represents the tree structure that maps the building cooling load samples to the corresponding leaf nodes. T is the total number of leaf nodes in the tree.

[0117] The specific steps are as follows:

[0118] When the present invention uses the XGBoost algorithm for feature screening and reduction, it first uses the VaDE algorithm to perform clustering analysis on the building air-conditioning cooling load data, and determines that the number of discrete intervals of the air-conditioning cooling load is 2. Based on this, combining the basic features of the data with the concentration degree of the load distribution, fully considering the different attributes of the air-conditioning cooling load and the influence of discrete points, the equal-distance division method is used to divide the air-conditioning cooling load data set of 50 large commercial buildings into two intervals, namely the high-load data set and the low-load data set.

[0119] To comprehensively evaluate the importance of features, the present invention constructs a three-dimensional evaluation framework, including the following three key dimensions: (1) Prediction accuracy evaluation - the mean squared error (MSE) is used to measure the prediction accuracy of the model; (2) Model fitting degree evaluation - the coefficient of determination R 2 is used to evaluate the fitting effect of the evaluation model; (3) Feature interaction importance evaluation - the interaction strength between features is calculated by combining the SHAP (Shapley additive explanations) interaction values, and the synergy effect of different feature combinations is evaluated.

[0120] Based on the above three-dimensional evaluation framework, the comprehensive score results of each feature under different data sets are calculated. Subsequently, according to the scoring conditions of the high- and low-load data sets, the core features affecting the air-conditioning cooling load are screened respectively. Among them, for the high-load data set: a feature subset with a comprehensive score of more than 70 points (ranked top 20) is selected; for the low-load data set: a feature subset with a comprehensive score of more than 80 points (ranked top 20) is selected. On this basis, the present invention further adopts a weighted aggregation method to perform weighted calculation on the top 20 feature subsets in the high- and low-load data sets to obtain the final feature scores. Subsequently, the final scores of the corresponding features in the high- and low-load data sets are weighted and averaged to obtain the comprehensive feature score affecting the air-conditioning cooling load prediction. Combining the analysis results of the RFE algorithm, the optimal number of features affecting the air-conditioning cooling load prediction is determined to be 9. Finally, according to the ranking of the comprehensive feature scores, the top 9 features are selected as input variables for building air-conditioning cooling load prediction. The specific selected features are as follows (ranked according to the comprehensive score): the outdoor temperature at the previous moment (t-1), the wind speed at the current moment (t), the rainfall at the current moment (t), the air-conditioning cooling load at the previous moment (t-1), the outdoor temperature at the current moment (t), the atmospheric pressure at the current moment (t), the solar radiation at the previous moment (t-1), equipment heat generation, and the indoor humidity at the current moment (t). The specific ranking is as Figure 4 shown.

[0121] In step five: The HN-Adam algorithm is used to adjust the learning rate during the training of the STGAT algorithm.

[0122] Principle of the STGAT algorithm:

[0123] The STGAT neural network projects the features affecting the cooling load prediction value into three vectors: query (Q), key (K), and value (V) to calculate the similarity between the cooling load influence factors, and based on this, calculates the weighted output. We can calculate the attention output through formula (12):

[0124]

[0125] Among them, Q represents the cold load impact factor at the current or future moment, K represents the impact factor corresponding to the historical cold load data, V represents the actual cold load value corresponding to the historical impact factor, and d k is the dimension of the cold load impact factor, softmax is the activation function used to generate attention weights, and T represents the transpose of the matrix.

[0126] The present invention is directed to the building air-conditioning cold load prediction task. When designing the STGAT model, its core spatio-temporal multi-head self-attention mechanism is retained to model the spatio-temporal dependence relationship between input features. However, since the data of the present invention does not contain graph structure information, the original STGAT model is simplified, the graph neural network part is deleted, and a combination of a fully connected layer and a multi-head attention mechanism is used for feature modeling. This improvement not only simplifies the model structure, but also significantly shortens the prediction time, and better fits the data characteristics and actual requirements of the building air-conditioning cold load prediction task. As Figure 5 shown, it shows the specific application process of the STGAT model in the field of building air-conditioning cold load invention.

[0127] Principle of the HN-Adam algorithm:

[0128] HN-Adam is an optimizer based on a hybrid of adaptive norm technology, the original Adam algorithm, and the AMSGrad algorithm. Among them, the letter "H" represents the hybrid mechanism, and "N" represents the adaptive norm. This optimizer introduces an adaptive (or dynamic) norm function into the Adam algorithm, enabling it to automatically adjust the learning rate and step size of the prediction model according to the adaptability of the current and historical gradients, effectively avoiding the model falling into local minima.

[0129] When using the HN-Adam optimizer to optimize the STGAT neural network, first initialize the momentum parameters for recording gradient historical information according to the basic operations of the Adam optimizer, that is: the cumulative direction m0 of the cold load prediction error adjustment = 0, the fluctuation amplitude v of the error adjustment = 0, and the iteration number t of the model = 0. Since a smaller K value usually leads to poor optimization results, while a higher K value brings expensive computational costs and little performance improvement, the initial adjustment coefficient B0 is set within the range of [1, 4].

[0130] Immediately afterwards, use formula (13) to calculate the current parameter H in the building air-conditioning cold load prediction model t and the gradient g t .

[0131]

[0132] Among them, g tIndicates the sensitivity of the difference between the predicted cooling load value by STGAT and the actual cooling load value to the parameters to be optimized, h t are the parameters to be optimized in the STGAT neural network, and J is the loss function.

[0133] After obtaining the current gradient, in order to retain the "memory" of the past gradients in the cooling load prediction training model and to avoid excessive fluctuations in a single update during the model training process, formula (14) is used to calculate the first-order momentum m t .

[0134] m t = β1·m t-1 + (1 - β1)·g t (14)

[0135] where m t represents the "memory" of the change direction of the past cooling load prediction error, and β1 is the decay factor of the first-order momentum, usually set to 0.9.

[0136] Subsequently, in order to further help the STGAT neural network adjust the learning rate, formula (15) is used to calculate the second-order momentum v t , which is used to capture the change amplitude of the gradient, and then helps to dynamically adjust the learning rate of each update.

[0137]

[0138] where v t represents the amplitude of the cooling load prediction error fluctuation, and β2 is the decay factor of the first-order momentum, usually set to 0.999.

[0139] On the above basis, in order to dynamically adjust the learning step during the training process of the building air-conditioning cooling load prediction model, combining the current gradient and historical momentum information, the dynamic adjustment normalization factor K(t) is calculated through formula (16). This normalization factor can adaptively adjust the learning rate according to the gradient change amplitude, further improving the convergence speed and global search ability of the model training, and thus effectively preventing getting stuck in local minima.

[0140]

[0141] Finally, the target optimization parameters in the optimized model are calculated through formula (17).

[0142]

[0143] where η is the learning rate that controls the amplitude of each parameter update, and in order to prevent the denominator in the calculation formula from being zero, a smoothing factor ε is introduced.

[0144] Through the adaptive adjustment of gradients and momentum by the HN-Adam optimizer, the finally optimized STGAT neural network will be obtained.

[0145] In step six, the model obtained in step five is used for cold load prediction.

[0146] The present invention collects 2,176 groups of data from large commercial buildings, among which 2,076 groups of data are selected for training the RFE-VaDE-XGBoost-HN-Adam-STGAT prediction model, and another 100 groups of data are used to test the prediction performance of the model.

[0147] In existing algorithms, the deep belief network (DBN), as one of the most widely used deep neural networks, is suitable for processing large-scale and high-dimensional data; the Elman neural network, as a typical dynamic recurrent neural network, can capture the dynamic characteristics of time series and is applicable to dealing with dynamic changes in the environment; while the STGAT neural network combines spatial and temporal characteristics and can more accurately capture the law of cold load changing with time by introducing the attention mechanism in the time dimension. Therefore, to verify the superiority of the RFE-VaDE-XGBoost-HN-Adam-STGAT prediction model, the present invention conducts comparative experiments on this model with multiple prediction models such as STGAT, HN-Adam-STGAT, RFE-VaDE-XGBoost-STGAT, DBN, HN-Adam-DBN, RFE-VaDE-XGBoost-DBN, RFE-VaDE-XGBoost-HN-Adam-DBN, Elman, HN-Adam-Elman, RFE-VaDE-XGBoost-Elman, and RFE-VaDE-XGBoost-HN-Adam-Elman.

[0148] The number of neuron nodes in the input layer and output layer of each prediction model is calculated according to the empirical formula to obtain the estimated interval of the number of neuron nodes in the hidden layer. The specific process is as follows: First, substitute the number of neuron nodes in each evaluation interval into the neural network; then, by comparing the mean squared error (MSE) results during the network training process, select the number of neuron nodes corresponding to the minimum error, so as to determine the optimal structure of each neural network.

[0149] During the construction of the building air-conditioning cold load prediction model, to verify the necessity of feature reduction, we take whether there is redundancy in the feature variables input to the prediction model as the only variable for comparative analysis. Figure 6It shows the comparison of the prediction performance of the STGAT neural network architecture optimized by the HN-Adam optimizer under the conditions of with and without feature reduction (based on RFE-VaDE-XGBoost). The comparison results show that the model after feature reduction exhibits better prediction performance when dealing with complex time-series data, significantly reducing the interference of redundant data noise. At the same time, it verifies the correctness of removing the five input features of the current solar radiation intensity, the previous outdoor humidity, the current outdoor humidity, the current indoor temperature, and lighting heat.

[0150] To verify the necessity of the HN-Adam optimizer, the present invention takes the presence or absence of the HN-Adam optimizer in the prediction model as the only variable for comparative analysis, and the results are as Figure 7 shown. It can be seen from the figure that both the RFE-VaDE-XGBoost-STGAT and RFE-VaDE-XGBoost-HN-Adam-STGAT prediction models can better capture the trend of building air-conditioning cooling load, but the prediction curve of the latter is more consistent with the actual value. This shows that the optimization of hyperparameters such as the learning rate in the STGAT neural network by the HN-Adam optimization algorithm has greatly improved the learning rate and step size of the model, thereby enhancing the search accuracy, convergence speed, and stability, and significantly improving the prediction accuracy.

[0151] The prediction performances of different building air-conditioning cooling load prediction models are as Figure 8 shown. The figure shows that compared with the RFE-VaDE-XGBoost-HN-Adam-DBN and RFE-VaDE-XGBoost-HN-Adam-Elman, the predicted values of the RFE-VaDE-XGBoost-HN-Adam-STGAT model are highly consistent with the true values. Especially in the peak and trough regions of the cooling load, its prediction deviation is smaller. While the models using the DBN or Elman algorithms have certain deviations when capturing the rapid fluctuations of the cooling load, reflecting the deficiencies of these two algorithms in capturing dynamic change laws. This proves that the prediction model containing the STGAT algorithm has higher prediction accuracy and stability in capturing the complex change laws of building air-conditioning cooling load over time.

[0152] To further and accurately evaluate the performance improvement of each prediction model after introducing feature reduction and the HN-Adam optimizer, the present invention uses two metrics, mean absolute error (MAE) and root mean squared error (RMSE), to comprehensively evaluate the building air-conditioning cooling load prediction model. As shown in Table 1, the RFE-VaDE-XGBoost-HN-Adam-STGAT model outperforms other models in both MAE and RMSE metrics, intuitively indicating that the building air-conditioning cooling load prediction model after feature reduction and HN-Adam optimization has significantly improved prediction accuracy.

[0153] Table 1 Performance Evaluation of Building Air-Conditioning Cooling Load Prediction Model

[0154]

[0155]

[0156] The above experimental results of the prediction analysis of the cooling load data of a large commercial building in Jiaxing City show that the building air-conditioning cooling load prediction method based on RFE-VaDE-XGBoost feature selection and HN-Adam-STGAT proposed by the present invention can effectively and accurately predict the building air-conditioning cooling load, with high accuracy and practical value. This method helps to optimize the control strategy of the HVAC system, significantly improve the energy efficiency of the air-conditioning system, and thus effectively reduce the energy waste of the building air-conditioning cooling load.

[0157] Although the exemplary embodiments that are regarded as the present invention have been described and recited, those skilled in the art will understand that various changes and substitutions can be made thereto without departing from the spirit of the present invention. Additionally, many modifications can be made to adapt a particular situation to the teachings of the present invention without departing from the central concept described herein. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but the present invention may also include all embodiments and equivalents within the scope of the present invention.

Claims

1. A building air conditioning cooling load prediction method based on RFE-VaDE-XGBoost feature selection and HN-Adam-STGAT, characterized in that: The following steps are involved: Step 1: Preprocess and clean the data for feature selection and prediction; Step 2: Use the RFE algorithm to determine the optimal number of input parameters for the prediction model; Step 3: Use VaDE algorithm to perform cluster analysis on the cleaned data; Step 4: Based on the results of steps 2 and 3, the XGBoost algorithm is used to eliminate factors that have little impact on the building cooling load; Step 5: Use the HN-Adam algorithm to dynamically adjust the learning rate during STGAT model training; Step 6: Use the model optimized in step 5 to predict the building air conditioning cooling load.

2. The building air conditioning cooling load prediction method based on RFE-VaDE-XGBoost feature selection and HN-Adam-STGAT according to claim 1 is characterized in that: In step 1, the DBSCAN algorithm is used to detect and process outliers in the data set. The process includes the following steps: (1) Set the parameters of the DBSCAN algorithm; (2) Density clustering is performed on the building cooling load dataset to detect and process abnormal data.

3. The building air conditioning cooling load prediction method based on RFE-VaDE-XGBoost feature selection and HN-Adam-STGAT according to claim 1 is characterized in that: The process of determining the optimal number of input parameters of the prediction model using the RFE algorithm in step 2 includes the following steps: (1) Sort all cooling load influencing factors according to the importance score of each feature; (2) The influencing factors with higher scores are sequentially used to form feature subsets; (3) Use GBRT to cross-validate different feature subsets and select the feature subset with the highest cross-validation score. The number of features it contains is the optimal number of input parameters for the prediction model.

4. The building air conditioning cooling load prediction method based on RFE-VaDE-XGBoost feature selection and HN-Adam-STGAT according to claim 1 is characterized in that: The process of clustering the data for feature selection using the VaDE algorithm in step 3 includes the following steps: (1) The preset air conditioning cooling load data set is divided into K clusters; (2) Calculate the cluster category to which each cooling load data point belongs; (3) Evaluate the clustering effect through reconstruction error and determine the final clustering result.

5. The building air conditioning cooling load prediction method based on RFE-VaDE-XGBoost feature selection and HN-Adam-STGAT according to claim 1 is characterized in that: In step 4, based on the results of step 2 and step 3, the process of using the XGBoost algorithm to eliminate unimportant influencing factors in the building cooling load includes the following steps: (1) Set the parameters of the XGBoost algorithm; (2) The XGBoost algorithm was used to evaluate the importance of influencing factors in the air conditioning cooling load data of different types of buildings; (3) Weight the scores of each category of impact factors to obtain the final importance score; (4) Based on the optimal number of input parameters determined in step 2, select the top-ranked key influencing factors.

6. The building air conditioning cooling load prediction method based on RFE-VaDE-XGBoost feature selection and HN-Adam-STGAT according to claim 5 is characterized in that: XGBoost algorithm formula: F={f(v)=w q(v) } in, is the predicted cooling load of the ith building, v i is the characteristic vector of the factors affecting the building cooling load, F is the space of the regression tree, and f k represents the kth regression tree, K represents the number of regression trees included in the integrated model, q represents the tree structure in which the building cooling load samples are mapped to the corresponding leaf nodes, and T is the total number of leaf nodes in the tree.

7. The building air conditioning cooling load prediction method based on RFE-VaDE-XGBoost feature selection and HN-Adam-STGAT according to claim 1 is characterized in that: The process of adjusting the learning rate during STGAT model training using the HN-Adam algorithm in step 5 includes the following steps: (1) Initialize the momentum used to record the gradient history information according to the basic operation of the Adam optimizer; (2) Calculate the parameters to be optimized in the STGAT neural network; (3) Dynamically adjust the learning rate at each parameter update based on the magnitude of the gradient change; (4) Calculate the dynamically adjusted normalization factor and determine the final learning rate to obtain the optimized STGAT prediction model.

8. The building air conditioning cooling load prediction method based on RFE-VaDE-XGBoost feature selection and HN-Adam-STGAT according to claim 7 is characterized in that: The STGAT is a neural network that introduces the attention mechanism into the Transformer model.

9. The building air conditioning cooling load prediction method based on RFE-VaDE-XGBoost feature selection and HN-Adam-STGAT according to claim 8 is characterized in that: The STGAT neural network calculates the similarity between the factors affecting the cooling load by projecting the features that affect the cooling load prediction value into three vectors: query (Q), key (K), and value (V), and calculates the weighted output based on this.

10. The building air conditioning cooling load prediction method based on RFE-VaDE-XGBoost feature selection and HN-Adam-STGAT according to claim 1, characterized in that: The process of using the optimized model to predict the building air conditioning cooling load in step 6 includes the following steps: (1) Divide the prediction data set into a training set and a test set; (2) Setting the parameters of the prediction model; (3) Train the prediction model and test its performance.

Images