Method for predicting cooling load of air conditioner of large commercial building based on ERT-RFE, cross validation and Radam-Autoformer

Abstract

The invention discloses a method for predicting the cooling load of an air conditioner of a large commercial building based on ERT-RFE, cross validation and Radam-Autoformer, and the method comprises the following steps: (1) collecting data for predicting the cooling load of the air conditioner of the large commercial building, carrying out the detection and processing of an abnormal value, and carrying out the normalization preprocessing; (2) performing feature screening on the data set by adopting ERT-RFE, and determining an optimal feature subset in combination with cross validation; (3) dividing the data set after feature selection into a training set and a test set; (4) training the Radam-Autoformer by using the training set, and continuously adjusting model parameters; (5) finishing final prediction by using the test set; and (6) comparing the predicted value with the true value by using the evaluation index. The method solves the problems that a traditional large commercial building air conditioner cooling load prediction model is unstable in training and insufficient in generalization ability, and a new method is provided for accurate prediction of the cooling load.

Description

Technical Field

[0001] The invention belongs to the field of large commercial building air conditioning cooling load prediction, and in particular relates to a large commercial building air conditioning cooling load prediction method based on ERT-RFE, cross validation and RAdam-Autoformer. Background Art

[0002] With the rapid development of the global economy, air conditioning systems in large commercial buildings are becoming more and more widely used. According to statistics, the global energy consumption of building air conditioning systems accounts for nearly 40%. Therefore, how to effectively reduce the energy consumption of air conditioning systems in large commercial buildings and improve energy efficiency has become an important issue in global energy conservation and environmental protection. Accurate and efficient air conditioning cooling load prediction can not only significantly improve the energy efficiency of air conditioning systems, but also provide reliable technical support for emergency management of emergencies. Therefore, the development of scientific and reasonable air conditioning cooling load prediction and control methods is of great significance to achieving energy conservation and emission reduction goals and promoting global sustainable development.

[0003] At present, there are two main methods for predicting the cooling load of air conditioners in large commercial buildings: one is the physical model, which is based on building structure information, outdoor climate conditions, and building internal environmental data. However, this method has a strong dependence on building characteristics, air conditioning system configuration, and environmental factors, and has high computational complexity and high computational cost. The second is the data-driven model, which can be specifically divided into statistical models, machine learning models, and deep learning models. Data-driven models capture data change trends more flexibly and efficiently by mining the time correlation and inherent regularity in historical data without relying on detailed building physical information. Especially with the expansion of data scale, deep learning models have gradually become the mainstream method for predicting cooling load of air conditioners in large commercial buildings with their powerful processing capabilities for high-dimensional complex data.

[0004] However, when dealing with long time series data and multi-level dependency structure prediction tasks, traditional deep learning models often struggle to fully realize their potential due to inappropriate optimizer selection. For example, in 2024, Yan et al. proposed a building air conditioning cooling load prediction method based on an improved bidirectional long short-term memory network. This method uses a hybrid strategy improved whale optimization algorithm to optimize network parameters. Although it improves the prediction accuracy to a certain extent, it is prone to falling into local optima, limiting the further improvement of model performance (XYan, X Ji, Q Meng, et al. A hybrid prediction model of improved bidirectional long short-term memory network for cooling load based on PCANet and attention mechanism. Energy, 2024, 292:130388). In another study, in 2024, He et al. proposed a method to optimize the extreme gradient boosting model using the satin bowerbird optimization algorithm, which significantly improved the prediction accuracy of building air conditioning cooling load. However, the parameter tuning process is complex, limiting the convenience of the model's practical application (D He, Y Zhang, M Bin Ashab. Proposing hybrid prediction approaches with the integration of machine learning models and metaheuristic algorithms to forecast the cooling and heating load of buildings. Energy, 2024, 291:130297)

[0005] In addition, the performance of the air-conditioning cooling load prediction model does not continuously improve with the increase in the number of input features. In fact, too many features may introduce redundant information or noise, thus affecting the prediction accuracy of the model. For example, the cooling load prediction method proposed by Gao et al. in 2024, which integrates machine learning and optimization algorithms, directly uses the integration algorithm for prediction without feature reduction, resulting in a significant increase in data complexity, a substantial increase in computational cost, and also being affected in prediction accuracy (T Gao, X Han, J Wang, et al. Enhancing building energy efficiency: An integrated approach to predicting heating and cooling loads using machine learning and optimization algorithms. Journal of Building Engineering, 2024, 98: 110759). Therefore, how to establish a large commercial building air-conditioning cooling load prediction model with higher prediction accuracy and stronger generalization ability on the basis of reducing redundant features has become an urgent need to improve the operating efficiency of building air-conditioning systems and the level of cooling load management. Summary of the Invention

[0006] The present invention aims to propose a prediction method for the air-conditioning cooling load of large commercial buildings based on ERT-RFE, cross-validation, and RAdam-Autoformer, to solve the problems of feature redundancy, low prediction accuracy, difficulty in optimizing model parameters, and insufficient prediction stability existing in the existing large commercial building air-conditioning cooling load prediction methods. The method can significantly improve the prediction accuracy of the air-conditioning cooling load of large commercial buildings, thus effectively improving the energy utilization efficiency of building air-conditioning systems and providing reliable technical support for coping with emergencies.

[0007] To address the above problems, the present invention proposes a prediction method for optimizing Autoformer based on extremely randomized trees (ERT) - recursive feature elimination (RFE), cross - validation, and RAdam (RectifiedAdam). Specifically, this method first uses the ERT - RFE algorithm to rank the importance of all initial features affecting the air - conditioning cooling load of large commercial buildings, clarifying the contribution degree of each feature to the prediction result. Secondly, based on the above - mentioned feature importance ranking, the cross - validation method is used to evaluate the prediction performance of different feature subsets, and the best feature combination is selected according to the evaluation index to achieve feature reduction of the dataset, eliminate redundant information, and reduce the model complexity. Then, the reduced dataset is reasonably divided into a training set and a test set, and the training set data is input into the Autoformer deep - learning model. The RAdam optimizer is used to dynamically adjust the learning rate during model training, quickly and stably optimize the hyperparameter configuration of the model to ensure the optimal model performance. Finally, the test set data is input into the RAdam - Autoformer model after training and parameter optimization to achieve high - precision prediction of the air - conditioning cooling load of large commercial buildings.

[0008] To achieve the above technical objectives, the technical solution of the present invention specifically includes the following steps:

[0009] Step 1: Data collection and pre - processing. Collect various feature data affecting the air - conditioning cooling load of large commercial buildings and the corresponding air - conditioning cooling load data, including meteorological data, the air - conditioning cooling load of large commercial buildings, and indoor environmental parameters. Perform pre - processing operations such as data cleaning, outlier detection, and normalization on the collected data to ensure the accuracy of subsequent feature reduction and cooling load prediction.

[0010] Step 2: Feature reduction. Use the ERT - RFE method combined with cross - validation to evaluate and screen the importance of each feature of the air - conditioning cooling load of large commercial buildings, eliminate redundant features, and reduce data complexity, thereby improving the efficiency and accuracy of the model.

[0011] Step 3: Dataset division. Divide the data obtained after feature reduction into a training set and a test set for model training and performance evaluation. The training set is used for model parameter learning and optimization, and the test set is used to verify the actual prediction effect of the model.

[0012] Step 4: Model training. Input the training set into the RAdam - Autoformer model, and use the RAdam optimizer to dynamically adjust and optimize the hyperparameters of Autoformer, enhancing the convergence stability and prediction accuracy of the model.

[0013] Step 5: Prediction of air-conditioning cooling load. Using the trained and optimized RAdam-Autoformer model, predict the feature data in the test set to obtain the prediction results of the air-conditioning cooling load of large commercial buildings.

[0014] Step 6: Evaluation of prediction performance. Quantitatively evaluate the error between the prediction results and the actual values. The specific evaluation indicators used include mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of variation of the root mean square error (CV-RMSE), so as to scientifically and objectively measure the prediction performance of the method.

[0015] Furthermore, the data collection and preprocessing in Step 1 include:

[0016] (1) Collect key feature data affecting the air-conditioning cooling load of large commercial buildings and the corresponding air-conditioning cooling load data, where the data covers meteorological factors, indoor environmental parameters, and the actual load of building air conditioners, etc.;

[0017] (2) Clean the collected data and detect outliers, and normalize the data that meets the conditions to ensure the accuracy of feature reduction and subsequent cooling load prediction.

[0018] Furthermore, in Step 2, ERT-RFE and cross-validation are used to reduce the features of the air-conditioning cooling load of large commercial buildings, specifically including:

[0019] (1) Use ERT-RFE to rank the importance of all initial features;

[0020] (2) According to the importance ranking, select different numbers of features in descending order for cross-validation, calculate and compare the prediction scores of each feature set;

[0021] (3) According to the cross-validation results, determine the feature set with the highest score, eliminate redundant features, and reduce data complexity.

[0022] Furthermore, in Step 3, the reduced dataset is divided into a training set and a test set, and the specific functions are as follows:

[0023] (1) The training set is used to adjust and optimize the hyperparameters of the model;

[0024] (2) The test set is used to evaluate the prediction performance of the model and examine its generalization ability on unknown data.

[0025] Furthermore, in step four, the training set is input into the RAdam-Autoformer model, and the hyperparameters of the Autoformer model are dynamically adjusted through the RAdam algorithm, specifically including:

[0026] (1) Input the feature data and air-conditioning cooling load data in the training set into RAdam-Autoformer;

[0027] (2) During the preliminary training process, RAdam adjusts the learning rate in real time according to the prediction results and corrects the model;

[0028] (3) When the training accuracy of the model reaches the preset requirement, output the optimal parameters and apply them to the Autoformer model.

[0029] Furthermore, in step five, the feature data in the test set is input into the trained RAdam-Autoformer model to obtain the predicted value of the air-conditioning cooling load of large commercial buildings. Since RAdam-Autoformer organically combines the optimizer with the deep learning framework, it has strong prediction ability.

[0030] Furthermore, in step six, by comparing the differences between the predicted values and the true values, the evaluation indicators include MAE, RMSE, MAPE, and CV-RMSE, which measure the prediction accuracy and stability of the model from multiple dimensions.

[0031] The innovation points of the present invention are mainly reflected in the following aspects: (1) Adopting a feature reduction strategy that combines ERT-RFE and cross-validation, effectively removing redundant features and reducing data complexity; (2) Introducing the RAdam optimizer, which can dynamically adjust the learning rate according to the training process, greatly improving the speed and prediction stability of the hyperparameter optimization of deep learning models; (3) Utilizing the Autoformer deep learning framework to accurately capture the changing rules of the air-conditioning cooling load of large commercial buildings in terms of trend terms and seasonal terms, and being able to predict the air-conditioning cooling load more accurately. Compared with traditional prediction methods, the method of the present invention has significant improvements in prediction accuracy, model stability, and feature utilization efficiency, providing strong technical support for the efficient operation and emergency response of the air-conditioning system in large commercial buildings.

[0032] Compared with the prior art, the present invention has significant technical advantages in the following three aspects:

[0033] (1) Multi-step feature reduction

[0034] The present invention first uses ERT-RFE to rank the importance of all features, and then determines the optimal feature combination through cross-validation to perform multi-step reduction on the air-conditioning cooling load features of large commercial buildings, thereby eliminating redundant features. ERT is an ensemble learning algorithm that introduces more randomness based on traditional random trees. By randomly selecting features and their division thresholds, it can better characterize the interaction relationships between complex features in air-conditioning cooling load prediction. RFE can effectively reduce the model complexity and calculation cost by iteratively deleting irrelevant or weakly relevant features. Cross-validation evaluates the performance of different feature combinations in terms of prediction accuracy and model complexity, helping to balance the model performance and generalization ability. Through the above process, the present invention can more accurately screen out the feature variables that have a significant impact on the air-conditioning system of large commercial buildings, while ensuring the prediction accuracy, reducing data redundancy and improving the training efficiency.

[0035] (2) Powerful optimizer

[0036] The present invention uses the RAdam optimizer in the model training process, significantly improving the training convergence speed and stability. RAdam was proposed by Liu et al. in 2019. By correcting the variance in the adaptive learning rate, it solves the problem that the traditional Adam algorithm is prone to unstable update directions in the early stage of training. Compared with commonly used algorithms such as Adam or stochastic gradient descent, RAdam realizes a warm-up mechanism in the early stage of model training by introducing an automated variance correction factor, reducing the dependence on manual adjustment of the learning rate. Its characteristic of dynamically adjusting the learning rate effectively balances global exploration and fast convergence, reducing the risk of falling into local optima due to improper hyperparameter settings. In diverse applications such as air-conditioning cooling load prediction, RAdam reduces the severe loss fluctuations in the early stage of training, achieving a more stable and efficient optimization process. Therefore, by introducing the RAdam optimizer, the present invention maintains high prediction accuracy and generalization ability in complex scenarios, and combines high efficiency and reliability.

[0037] (3) Novel deep learning framework

[0038] The present invention uses the Autoformer model for predicting the air-conditioning cooling load of large commercial buildings, significantly improving the prediction accuracy and calculation efficiency. Autoformer was proposed by research teams such as Peking University in 2021. Through an adaptive decomposition framework and a self-correlation mechanism, it decomposes the time series into trend terms and seasonal terms, and at the same time uses an improved self-attention mechanism to better capture long-term periodicity and global dependencies. Compared with traditional Transformer or long short-term memory (LSTM) models, Autoformer abandons the standard attention mechanism and uses a delayed stacked self-correlation module to handle the inherent periodicity of the sequence, which can not only reduce the computational complexity but also enhance the modeling ability for long-term dependence information. In the face of complex scenarios such as noise and missing data, Autoformer has stronger robustness, avoiding feature omission caused by local attention bias in traditional methods. Therefore, through the Autoformer model, the present invention can still accurately identify trends and seasonal changes in the cooling load prediction over a long time span and provide more reliable prediction results. In summary, through the integration of three core technologies: multi-step feature reduction, RAdam optimizer, and Autoformer deep learning framework, the present invention significantly improves the accuracy, stability, and efficiency of the air-conditioning cooling load prediction of large commercial buildings, and has high practical value and promotion potential. Description of the Drawings

[0039] Figure 1 Fig. is the overall workflow diagram of the large commercial building air-conditioning cooling load prediction method based on ERT-RFE, cross-validation, and RAdam-Autoformer proposed by the present invention;

[0040] Figure 2 Fig. is a schematic diagram of the importance scores of the air-conditioning cooling load features obtained by applying the ERT-RFE algorithm;

[0041] Figure 3 Fig. is a comparison chart of the scores obtained by evaluating different numbers of features through cross-validation;

[0042] Figure 4 Fig. is a schematic diagram of the workflow of the Autoformer model;

[0043] Figure 5 Fig. is a schematic diagram of the workflow of the RAdam optimizer in the present invention;

[0044] Figure 6 Fig. is a comparison chart of the air-conditioning cooling load prediction errors when comparing the use and non-use of the feature reduction method of the present invention;

[0045] Figure 7The distribution diagram of the prediction error of the air-conditioning cooling load when different optimization algorithms are adopted;

[0046] Figure 8 The radar chart of the prediction evaluation indexes of the air-conditioning cooling load when different prediction algorithms are adopted. Specific implementation manners

[0047] To better understand the present invention, the following is described in conjunction with specific embodiments. These embodiments are only used to illustrate the present invention and fall within the protection scope of the present invention, but do not limit its applicable scope.

[0048] As Figure 1 shown, the present invention provides a method for predicting the air-conditioning cooling load of large commercial buildings based on ERT-RFE, cross-validation, and RAdam-Autoformer, which specifically includes the following steps:

[0049] Step 1: Data collection and preprocessing

[0050] First, building meteorological data, air-conditioning cooling load data of large commercial buildings, and indoor environment parameters are obtained through multi-source data collection. To ensure the accuracy of subsequent feature reduction and cooling load prediction, the collected data is systematically preprocessed, including: (1) cleaning outliers using statistical methods and reasonably filling in missing data; (2) normalizing all features, and finally forming a normalized air-conditioning cooling load data set.

[0051] Step 2: Feature reduction

[0052] To accurately identify the key features affecting the air-conditioning cooling load and improve the generalization ability and accuracy of the model, the present invention proposes a feature screening method combining ERT-RFE and cross-validation. The specific operations are as follows: (1) Use the ERT model to score the importance of each feature and adopt the RFE strategy. The ERT model can evaluate the contribution degree of different features to the air-conditioning cooling load prediction by randomly selecting features and their division thresholds. RFE removes the features with the lowest scores during the iteration process, thereby completing the importance ranking of all features; (2) Under the framework of 5-fold cross-validation, dynamically evaluate the prediction performance of different feature subsets. Cross-validation can reduce the accidental influence caused by data division and ensure the stability and generalization ability of the selected features on different data sets; (3) Finally, comprehensively compare the performance of different feature subsets in terms of prediction accuracy and model complexity, screen out the optimal feature combination, achieve effective feature reduction, improve the model calculation efficiency, and enhance the accuracy of air-conditioning cooling load prediction.

[0053] Step 3: Data set division

[0054] After completing feature reduction, the dataset is divided into a training set and a test set. The training set is used to train the RAdam-Autoformer model and tune hyperparameters, helping the model better learn the data feature patterns and obtain the optimal hyperparameter configuration. The test set is used as an independent dataset for final prediction performance verification to evaluate the generalization ability and actual prediction effect of the model.

[0055] Step Four: Model Training

[0056] When training Autoformer, the RAdam optimization algorithm is introduced to fully explore the potential of deep learning models. Identify the main optimization parameters, including the weight matrix of the self-attention layer, the projection weights of the factorized moving average module, the cross-attention scaling factor in the decoder, and the dynamic adjustment parameters of each normalization layer; initialize the RAdam optimizer and set the base learning rate, weight decay coefficient, and warm-up steps; balance the direction and step size of parameter updates by dynamically adjusting the learning rate, and perform real-time iterative optimization by combining the exponential moving averages of the first and second moments of the gradient, continuously monitoring the loss function during the training process; when the prediction accuracy reaches the preset threshold, stop training and retain the optimal model parameters. Since RAdam introduces an adaptive learning rate and warm-up mechanism, it can effectively alleviate the sensitivity problem of Autoformer in the initial stage of the attention weight matrix, thereby improving the convergence speed and generalization ability of the model.

[0057] Step Five: Air Conditioning Cooling Load Prediction

[0058] Input the feature data in the test set (including historical values of air conditioning cooling load, meteorological data, etc.) into the trained and optimized RAdam-Autoformer model to output the final air conditioning cooling load prediction value. Through the coupling of Autoformer and RAdam, the present invention can not only capture long-term time series dependencies and complex relationships between features, but also automatically smooth the learning rate fluctuations in the early stage of model training, making the overall prediction more accurate and stable.

[0059] Step Six: Prediction Performance Evaluation

[0060] After completing the air conditioning cooling load prediction, quantitatively analyze the model performance through multiple evaluation metrics, including: MAE - measuring the average deviation between the prediction result and the true value; RMSE - being sensitive to large errors and reflecting the overall level of prediction errors; MAPE - measuring the relative error size and intuitively reflecting the proportional impact of the prediction deviation on the true value; CV-RMSE - evaluating the consistency between the prediction fluctuations and the true load fluctuations. Through the comprehensive consideration of the above multi-dimensional metrics, the accuracy and stability of the model can be comprehensively tested, providing a reference basis for subsequent model improvement and air conditioning system optimization.

[0061] To verify the effectiveness of the present invention, the present invention conducts a simulation experiment on the cooling load prediction of large commercial buildings. The specific implementation process is as follows:

[0062] In the data collection and preprocessing of Step 1, approximately 2,100 hours of observation data were collected, involving 14 feature variables, which were divided into three major categories: meteorological data, building equipment and energy system parameters, and indoor environment parameters. (1) Meteorological data (9 variables) include the outdoor temperature OT(t) at the current moment, the outdoor relative humidity ORH(t) at the current moment, the solar radiation SR(t) at the current moment, the outdoor temperature OT(t - 1) at the previous moment, the outdoor relative humidity ORH(t - 1) at the previous moment, the solar radiation SR(t - 1) at the previous moment, the wind speed WS(t) at the current moment, the precipitation P(t) at the current moment, and the atmospheric pressure AP(t) at the current moment. These variables are closely related to the building's heat load and cooling load demands and are important external factors affecting cooling load prediction. (2) Building equipment and energy system parameters (3 variables) include the equipment heat generation EH(t) at the current moment, the lighting heat LH(t) at the current moment, and the building cooling load CL(t - 1) at the previous moment. These indicators can reflect the main heat sources and energy consumption characteristics inside the building and are of great significance for evaluating the cooling load demand. (3) Indoor environment parameters (2 variables) include the indoor temperature IT(t) at the current moment and the indoor relative humidity IRH(t) at the current moment. These two parameters, as direct indicators of the cooling load demand, can accurately reflect the comfort requirements and operating status inside the building.

[0063] The units, value ranges, and specific meanings of the above variables are listed in Table 1. By covering data in multiple dimensions such as meteorology, building equipment and energy systems, and the indoor environment, this study was able to comprehensively capture the key driving factors affecting the cooling load of large commercial buildings, providing a solid data foundation for subsequent model training and prediction.

[0064] Table 1 Variables and Descriptions of the Original Dataset

[0065]

[0066] In Step 2 (feature reduction), the present invention adopts a feature screening method based on ERT - RFE combined with 5 - fold cross - validation. This process first scores the importance of each feature through the ERT model and evaluates the contribution of the feature to the building air - conditioning cooling load prediction using the Gini index; subsequently, relying on RFE, the feature with the lowest score is removed in each iteration to complete the importance ranking of the features and identify the key features (see Figure 2)。On this basis, the prediction performance of different feature subsets is dynamically evaluated through 5-fold cross-validation. Cross-validation can effectively avoid the contingency brought by data division, ensuring the stability and generalization ability of the finally selected feature subset in various data scenarios. By comprehensively comparing the prediction scores of different feature subsets (see Figure 3 ), the optimal feature combination can be screened out; this not only minimizes redundant features, improves the model calculation efficiency, but also enhances the prediction accuracy of the air-conditioning cooling load.

[0067] In step three (dataset division), the large commercial building air-conditioning cooling load dataset obtained after feature reduction is further divided into a training set and a test set. Among them, the training set is used for the parameter learning and tuning of the subsequent RAdam-Autoformer model, and the test set is specifically used for performance verification in the final model evaluation stage to ensure that the model still maintains high prediction accuracy and generalization ability on unknown data.

[0068] In step four (model training), by inputting the training set into RAdam-Autoformer, the internal parameters of Autoformer are dynamically tuned in combination with the RAdam optimizer.

[0069] The series decomp module in the Autoformer model decomposes the time series into a trend term and a seasonal term, effectively distinguishing long-term trends and periodic fluctuations, which helps to capture the load change rules of large commercial buildings within daily / seasonal cycles. The Auto-correlation mechanism replaces the traditional self-attention with auto-correlation, which can automatically identify and aggregate historical period information with strong correlation in the time dimension, so as to accurately capture long-term dependencies and avoid large-scale calculations for long sequence data. For air-conditioning cooling load prediction, Autoformer can naturally decompose daily cycles, seasonal cycles, and short-term random fluctuations, and at the same time use the auto-correlation mechanism to model periodic phenomena such as peak loads that appear multiple times, significantly improving the model's ability to identify and predict long-term load fluctuations.

[0070] The encoder-decoder structure is as Figure 4As shown in the figure, each layer of the Autoformer mainly includes functional modules such as Auto-correlation, Series Decomp, and Feed Forward. Auto-correlation calculates the similarity between Query (Q) and Key (K) in the input sequence in the time dimension, uses the delay coefficient to weight and aggregate the historical information between Q and K, and normalizes the similarity using the Softmax function. Finally, it extracts the historical moments that are most valuable for predicting the current air-conditioning cooling load. Q usually represents the characteristics of the target time period or the moment to be predicted, and K refers to the set of historical characteristics that may be associated with it. This process can not only significantly reduce the computational complexity on long sequences but also effectively capture the periodic characteristics of the air-conditioning cooling load, showing excellent performance in identifying the occurrence of peak loads and coping with long-term load fluctuations. Through the above mechanism, the Autoformer can significantly reduce the computational complexity while fully mining the important periodic and trend characteristics in the time-series data of the air-conditioning cooling load. With the help of the RAdam optimizer to dynamically adjust the learning rate, it can obtain faster and more stable convergence. Therefore, the present invention significantly improves the prediction efficiency and accuracy in the task of predicting the air-conditioning cooling load of large commercial buildings, providing reliable technical support for practical engineering applications.

[0071] In the Auto-correlation mechanism, first record the length of the time series as L and define a set of delay coefficients {τ1, τ2, …, τ k} to represent different historical lag steps. For the query vector Q of the current time period and the corresponding historical time period data K, the original correlation quantity of each delay τ i can be calculated first and its general form can be expressed as:

[0072]

[0073] where, Q t represents the key feature reflecting the cooling load state at the t-th moment, represents the historical feature value after shifting backward by τ i steps, and τ i is the selected delay step. This original correlation quantity quantifies the degree of association between the current time period and its historical time period in the case of delay τ i , providing a basis for subsequent weighted normalization and autocorrelation aggregation. By comparing the correlation quantities of different delays τ i , the model can more flexibly capture the long-period and repetitive characteristics in the time series of the cooling load, thereby improving the prediction accuracy and stability of the air-conditioning load of large commercial buildings while reducing the overall computational complexity.

[0074] After completing the calculation of the original correlation quantities for each delay τ i the original correlation quantities After calculation, by introducing exponential weighting and Softmax normalization, the attention weight α(τ i ) corresponding to each lag τ can be obtained i . Its expression form can usually be written as:

[0075]

[0076] where represents the importance of a certain historical lag moment for the current cooling load in the final prediction. β is usually an adjustable hyperparameter used to control the sensitivity of different correlation quantities during weighting

[0077] Through the above Softmax function, the original correlation quantities at each lag time step can be transformed into an interpretable probability distribution, thus emphasizing the historical moments that contribute most to the current prediction. In the prediction of air-conditioning cooling load, the lag steps with larger correlation quantities will be assigned higher weights, thus highlighting the historical information that is more sensitive to the load trend and periodic characteristics, and improving the model's ability to capture complex time series patterns. Finally, by performing a rolling operation on the Key sequence according to the delay amount τ i (denoted as Roll(K,τ i )) and multiplying and accumulating with the corresponding attention weights , the final weighted synthesis result can be obtained, and its mathematical expression is as follows:

[0078]

[0079] In the case of long sequence calculation, the above Auto-correlation mechanism can highlight the historical information that is most critical to the current demand, and effectively identify seasonal peaks or periodic patterns, thus improving the prediction accuracy and robustness for long-term load fluctuations and peak loads

[0080] Subsequently, the Series decomp module will decompose the original cooling load sequence into a trend term and a seasonal term:

[0081] y = T + S (4)

[0082] where T represents the long-term trend obtained through operations such as moving average, and S represents the seasonal fluctuations of a shorter period. Compared with directly modeling the original sequence, decomposing the cooling load sequence into a trend term and a seasonal term not only has better physical interpretability, but also helps the model focus on its respective key patterns at different frequency levels. Such a hierarchical processing method has a significant effect on accurately capturing multiple periodic fluctuations and enhancing the model's ability to identify different seasonal peaks

[0083] The last step is the Feed forward process. Usually, a two-layer neural network in the following form is used to perform a non-linear mapping on the intermediate features:

[0084] H = σ(W1X + b1) (5)

[0085]

[0086] Among them, X represents the feature vector output by the previous module, which can be regarded as "the deep representation of the current and historical air-conditioning cooling load status". W1 and W2 are respectively learnable weight matrices. b1 and b2 are the corresponding bias vectors. σ(·) represents the non-linear activation function. H is the high-dimensional abstract feature extracted after the first-layer non-linear transformation, which further captures the complex mapping relationship between different physical quantities. is the final air-conditioning cooling load prediction result, which can be understood as "the optimal estimate of the cooling load value at the future moment given all the important features at the previous moment".

[0087] Thanks to the non-linear mapping of the above two-layer structure, the Feed forward module can establish a more flexible functional relationship between the original input and the predicted output, thereby improving the prediction accuracy of the air-conditioning cooling load under different scenarios (such as seasonal changes, peak loads, and special working conditions, etc.). Through such deep modeling, the present invention can perform more refined analysis and control on the operating status of complex and variable building air-conditioning systems, and further meet the actual needs of energy conservation and emission reduction and accurate load control.

[0088] The core idea of the RAdam optimizer is to introduce a correction factor to correct the second-order moment estimation on the basis of the traditional Adam optimizer, thereby smoothing the learning rate fluctuation in the initial stage of training. The first-order moment and the second-order moment of the traditional Adam algorithm are respectively defined as follows:

[0089] m t = β1m t-1 +(1 - β1)g t (7)

[0090] v t = β2v t-1 +(1 - β2)g t 2 (8)

[0091] Among them, m t and v t respectively represent the estimation of the gradient momentum and variance. β1 and β2 are hyperparameters (usually taken as 0.9 and 0.999), which are used to control the memory decay rate of the historical gradient when the model updates the parameters.

[0092] In the task of predicting the cooling load of air conditioners in large commercial buildings, the model often needs to process multi-dimensional feature inputs (such as temperature, humidity, solar radiation, and historical load values, etc.). Let g t represent the gradient information contributed by each feature in the current batch of data to the update of the model parameters. When there is a sudden increase in the load of the building in a short period of time or the gradient shows significant fluctuations due to drastic changes in meteorological conditions, a stable and reliable optimizer is needed to capture these mutation features and prevent the model from overshooting. RAdam automates the warm-up of the learning rate by correcting the second-order moment estimation, effectively improving the smoothness of the training process and the generalization ability of the model.

[0093] The main improvement of RAdam lies in introducing a learnable correction term ρ t , and by detecting the effective difference of the second-order moment v t , the size and direction of the learning rate are adjusted in an adaptive manner. Its core correction form can be summarized as:

[0094]

[0095]

[0096] where ρ ∞ is the limit value of the theoretical second-order moment degree of freedom (related to β2), and are the estimates after bias correction of the gradient momentum and variance respectively.

[0097] RAdam decides whether to perform a complete adaptive correction by judging whether ρ t exceeds a pre-set threshold, so as to avoid the learning rate being too large or too small due to insufficient variance estimation in the initial stage of training. Finally, the update method of the model parameter θ t can be described as:

[0098]

[0099] where η is the base learning rate. ∈ is a constant to prevent division by zero. f(ρ t ) represents a function that dynamically adjusts the learning rate based on the correction factor ρ t . θ t can be regarded as the initial parameter values of the model after the first iteration, and they will directly affect the dynamic learning of the model for the multi-dimensional inputs of the building air conditioning system (such as temperature, humidity, solar radiation amount, or historical load values, etc.). When there are significant changes in the external environment (such as extreme weather or sudden load surges) and the gradient update of g t fluctuates greatly, the stable and appropriate update strategy provided by RAdam is needed to ensure the rapid convergence of the model and maintain a high generalization performance.

[0100] Figure 5 It shows the overall process schematic of the RAdam optimizer. It can be seen that RAdam can not only effectively avoid the parameter oscillation caused by the unstable learning rate in the early training stage, but also make full use of variance correction in the later stage to ensure the model convergence speed and prediction accuracy. By applying this optimizer to the air-conditioning cooling load prediction of large commercial buildings, the present invention further enhances the adaptability and robustness to complex time-varying features while ensuring stability.

[0101] In step five (air-conditioning cooling load prediction), the feature data of the test set (including historical air-conditioning cooling load, meteorological data, etc.) is input into the trained RAdam-Autoformer model to output the final air-conditioning cooling load prediction value.

[0102] In step six (prediction performance evaluation), based on the composite evaluation system, by quantitatively analyzing the distribution and fluctuation characteristics of the prediction error, the comprehensive advantages of the constructed model in terms of air-conditioning cooling load prediction accuracy and anti-interference ability are verified from multiple index perspectives. To evaluate the effectiveness of the present invention, first, the influence of "using the feature reduction method" and "not using the feature reduction method" on the air-conditioning cooling load prediction results is compared and analyzed, and the relevant evaluation indexes are shown in Table 2. To more intuitively display this result, Figure 6 The comparison of the prediction errors between ERT-RFE-CV-RAdam-Autoformer and RAdam-Autoformer is presented, and the results show that using ERT-RFE-CV for feature reduction can obtain better prediction performance.

[0103] To further examine the influence of the optimizer on the prediction performance, the Autoformer model is trained using different optimization algorithms respectively, and the MAE, RMSE, MAPE, and CV-RMSE between its prediction results and the true values are compared. The specific values are shown in Table 3. Figure 7 The comparison of the prediction error distributions between the ERT-RFE-CV-RAdam-Autoformer model and other models (such as ERT-RFE-CV-Adam-Autoformer, ERT-RFE-CV-Autoformer, ERT-RFE-CV-PSO-Autoformer, ERT-RFE-CV-CNN-Autoformer) is given. As Figure 7 (a) shows, the prediction error of ERT-RFE-CV-RAdam-Autoformer is significantly smaller than that of other models, indicating that the dynamic variance correction brought by the RAdam optimizer effectively avoids the large-scale manual tuning that may occur during training and achieves better performance in terms of accuracy and stability.

[0104] After analyzing the performance of feature reduction methods and optimizers, it is equally important to compare the deep learning frameworks themselves. To this end, on the premise of the same feature selection method and optimizer, Autoformer is compared with models commonly used in air-conditioning cooling load prediction, such as DELM, Elman, and TCN. The evaluation metrics of the four deep learning frameworks are shown in Table 4. Figure 8 The radar chart comparison of the four metrics (RMSE, MAPE, MAE, CV-RMSE) between ERT-RFE-CV-RAdam-Autoformer and other models (ERT-RFE-CV-RAdam-DLEM, ERT-RFE-CV-RAdam-Elman, ERT-RFE-CV-RAdam-TCN) is shown. It can be seen that the area enclosed by ERT-RFE-CV-RAdam-Autoformer is much smaller than that of other models, indicating that the combined model performs better in the prediction accuracy of air-conditioning cooling load.

[0105] In summary, this step comprehensively evaluates the feature reduction method, optimizer, and deep learning framework selected in the present invention, and conducts a comparative analysis with international advanced algorithms. The results show that the present invention has significant advantages in the accuracy, stability, and generalization ability of air-conditioning cooling load prediction, laying a solid foundation for its application in practical engineering.

[0106] Table 2 Comparison of model evaluation metrics with and without the ERT-RFE-CV feature reduction method

[0107]

[0108] Table 3 Comparison of model prediction performance under different optimizers

[0109]

[0110] Table 4 Comparison of model evaluation metrics under different deep learning frameworks

[0111]

[0112] The present invention discloses a prediction method for the air conditioning cooling load of large commercial buildings based on ERT-RFE, cross-validation, and RAdam-Autoformer. Specifically, first, data required for predicting the air conditioning cooling load of large commercial buildings is collected, including meteorological data, the air conditioning cooling load of large commercial buildings, and indoor environmental parameters. After detecting outliers and missing values, they are uniformly normalized to ensure the effectiveness and consistency of the data. Subsequently, ERT-RFE is used to sort all features according to their importance, and cross-validation is combined to evaluate the performance of different numbers of feature subsets, so as to obtain the optimal feature set and reduce the complexity of the data set. After feature reduction is completed, the obtained data set is divided into a training set and a test set. The RAdam-Autoformer model is trained and dynamically parameter-adjusted using the training set, and finally, the cooling load is predicted on the test set. By comparing the prediction results with the true values and combining various evaluation indicators, the performance of the model in terms of accuracy and stability can be comprehensively measured. Compared with traditional prediction methods for the air conditioning cooling load of large commercial buildings, the present invention can effectively improve problems such as unstable model training, significantly improve the prediction accuracy of the air conditioning cooling load, and provide a new and efficient solution for the load prediction of the air conditioning system in large commercial buildings.

[0113] The above content cannot determine that the specific implementation of the present invention is limited only to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, and all should be regarded as belonging to the scope of patent protection determined by the claims submitted for the present invention.

Claims

1. A prediction method for the air conditioning cooling load of large commercial buildings based on ERT-RFE, cross-validation, and RAdam-Autoformer, characterized in that It includes the following steps: Step 1: Data collection and preprocessing; Step 2: Feature reduction; Step 3: Divide the reduced dataset into a training set and a test set; Step 4: Use the training set to train the RAdam-Autoformer model; Step 5: Air conditioning cooling load prediction; Step 6: Evaluation of prediction results.

2. The prediction method for the air conditioning cooling load of large commercial buildings based on ERT-RFE, cross-validation, and RAdam-Autoformer according to claim 1, wherein It includes the following steps: Step 1: Collect historical data samples for feature reduction and perform data preprocessing; Step 2: Use ERT-RFE combined with cross-validation to perform feature reduction on the air conditioning cooling load data; Step 3: Divide the data after feature reduction obtained in Step 2 into a training set and a test set; Step 4: Use the training set to train the RAdam-Autoformer model; Step 5: Input the feature data of the test set into the trained RAdam-Autoformer model for air conditioning cooling load prediction; Step 6: Evaluate the prediction results and compare and analyze the predicted values with the true values.

3. The prediction method of the air conditioning cooling load of large commercial buildings based on ERT-RFE, cross-validation and RAdam-Autoformer according to claim 2, characterized in that, The data collected in Step 1 specifically includes meteorological data, air conditioning cooling load data of large commercial buildings, and indoor environment parameter data.

4. The prediction method for the air conditioning cooling load of large commercial buildings based on ERT-RFE, cross-validation, and RAdam-Autoformer according to claim 2, wherein The data preprocessing in Step 1 specifically includes data cleaning, outlier detection and processing, and data normalization to ensure the accuracy of subsequent feature reduction and model prediction.

5. The prediction method for the air conditioning cooling load of large commercial buildings based on ERT-RFE, cross-validation, and RAdam-Autoformer according to claim 2, wherein The specific steps of feature reduction in Step 2 are as follows: (1) Use the ERT-RFE method to rank the importance of all features related to air conditioning cooling load; (2) Based on the feature importance ranking results, gradually select different numbers of features from high to low for cross-validation to determine the prediction accuracy of different feature subsets; (3) According to the cross-validation results, select the optimal feature subset with the highest prediction accuracy.

6. The prediction method of the air conditioning cooling load of large commercial buildings based on ERT-RFE, cross-validation and RAdam-Autoformer according to claim 2, wherein The specific division of the dataset in Step 3 is to divide all the data into a modeling training set for model training and a test dataset for model performance verification.

7. The prediction method for the air conditioning cooling load of large commercial buildings based on ERT-RFE, cross-validation and RAdam-Autoformer according to claim 2, characterized in that, The specific model training in Step 4 includes: (1) Input the feature data in the training set into the RAdam-Autoformer model; (2) Automatically adjust the learning rate and other hyperparameters of the Autoformer model through the RAdam optimization algorithm to ensure the stability of the model during training; after meeting the termination conditions, output the optimal parameter settings of the model to effectively improve the stability and robustness of the model.

8. The prediction method of the air conditioning cooling load of large commercial buildings based on ERT-RFE, cross-validation, and RAdam-Autoformer according to claim 2, wherein The specific air conditioning cooling load prediction in Step 5 is to input the feature data in the test set into the trained RAdam-Autoformer model to output the corresponding air conditioning cooling load prediction values of large commercial buildings.

9. The prediction method for the air conditioning cooling load of large commercial buildings based on ERT-RFE, cross-validation, and RAdam-Autoformer according to claim 2, wherein, The evaluation of the prediction results in Step 6 is carried out by calculating the mean absolute error, root mean square error, mean absolute percentage error, and coefficient of variation of the root mean square error between the predicted values and the true values.

10. The prediction method for the air conditioning cooling load of large commercial buildings based on ERT-RFE, cross-validation, and RAdam-Autoformer according to claim 9, characterized in that, The calculation formulas of the evaluation indexes are as follows respectively: Where: n is the total number of cold load data samples, and y i is the true value of the i-th cold load, and is the predicted value of the i-th cold load.

Images