HVAC load pattern prediction method based on time series decomposition and ensemble learning

Abstract

The present invention relates to the technical field of HVAC load pattern prediction, and discloses an HVAC load pattern prediction method based on time series decomposition and ensemble learning. Based on time series decomposition, trend components and seasonal components are extracted, which enhances the in-depth exploration of long-term load variation patterns and periodic characteristics, making up for the defect of traditional methods that fail to fully capture time series characteristics, and providing a more discriminative enhanced data set for cluster analysis. Furthermore, when generating classification labels through cluster analysis, a dynamic determination strategy for the optimal number of clusters is introduced, which significantly improves the objectivity and accuracy of load pattern classification. Multi-dimensional feature extraction is combined to screen feature sets that are strongly correlated with classification labels, avoiding the interference of redundant features on the model. Finally, a multi-classifier ensemble learning model is used in combination with a voting mechanism for prediction. By complementing the advantages of different models and integrating the results, the stability and generalization ability of the prediction results are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of HVAC load pattern prediction, and in particular to an HVAC load pattern prediction method based on time series decomposition and ensemble learning. Background Art

[0002] Heating, Ventilation, and Air Conditioning (HVAC) systems account for a significant portion of energy consumption in the operation and management of buildings such as subway stations and rail transit hubs. The rational classification of HVAC load patterns to facilitate the development of more precise energy-saving control strategies is a key topic in current research on energy conservation in intelligent buildings and rail transit. Subway station loads exhibit significant time-dependency and cyclical characteristics, influenced by multiple factors such as passenger flow, ambient temperature and humidity. Traditional load pattern classification methods struggle to accurately identify different operating modes, resulting in significant limitations in energy optimization.

[0003] Existing technologies have already conducted some research on load pattern classification, with the main methods including those based on cluster analysis and machine learning. For example, unsupervised clustering methods such as K-means and DBSCAN are used to directly classify load data. Traditional methods directly cluster raw load data, resulting in insufficient utilization of features such as temporal periodicity and trend characteristics, and poor stability of classification results. Machine learning-based classification methods have been applied to load pattern recognition, but these methods often fail to combine time series decomposition and feature engineering, using only raw load data for classification without fully considering the time series characteristics of load data. Furthermore, the generalization ability of a single model is limited and it is easily affected by changes in data distribution.

[0004] Although the above methods have achieved the classification of HVAC load patterns to a certain extent, they still have the following shortcomings: (1) Lack of feature engineering. Most methods directly use raw load data for clustering or classification, and fail to fully extract the long-term trend and periodic characteristics of the load, resulting in unstable classification results; (2) Without combining time series decomposition, traditional clustering methods find it difficult to accurately capture the trend and periodic characteristics of the load; (3) Lack of ensemble learning. Most existing classification methods are based on a single model, resulting in poor adaptability under different data conditions and insufficient generalization ability. Therefore, it can be seen that most existing methods focus on a certain aspect of load pattern classification and lack a comprehensive and systematic framework.

[0005] Therefore, there is an urgent need for an HVAC load pattern prediction method based on time series decomposition and ensemble learning, which can more accurately identify different operating modes, improve the stability and generalization ability of the prediction model, and provide more reliable basic data support for subsequent energy-saving optimization. Summary of the Invention

[0006] In order to solve the above technical problems, the present invention provides, on one hand, an HVAC load pattern prediction method based on time series decomposition and ensemble learning, comprising the following steps: Step S1: obtaining original HVAC load data, preprocessing the original HVAC load data, and generating preprocessed load time series data; Step S2: performing time series decomposition based on the preprocessed load time series data, extracting trend components and seasonal components, and generating an enhanced data set; Step S3: performing cluster analysis on the enhanced data set by obtaining an optimal number of clusters to obtain HVAC load pattern classification labels; Step S4: obtaining a feature set by performing correlation analysis on the HVAC load pattern classification labels and the original HVAC load data;

[0007] Step S5: Based on the feature set, a multi-classifier ensemble learning model is used for prediction, combined with a voting mechanism, to obtain a load pattern prediction result.

[0008] Furthermore, in step S2, time series decomposition is performed based on the preprocessed load time series data, trend components and seasonal components are extracted, and an enhanced data set is generated, specifically including: S21: decomposing the preprocessed load time series data into trend components, seasonal components and residual components using the STL method; S22: merging the trend components and the seasonal components with the original HVAC load data to form the enhanced data set.

[0009] Furthermore, the calculation formula of the STL method is:

[0010] ;

[0011] Where Y t Represents the given load data, T t represents the trend component, S t represents the seasonal component, R t represents the residual component.

[0012] Furthermore, in step S3, based on the enhanced data set, cluster analysis is performed by obtaining the optimal number of clusters to obtain HVAC load pattern classification labels, including:

[0013] S31: Determine a range of cluster numbers based on the data volume of the enhanced data set;

[0014] S32: performing cluster analysis based on the enhanced data set using each cluster number within the cluster number range;

[0015] S33: Calculate the evaluation index of the clustering result corresponding to each cluster number respectively;

[0016] S34: According to the evaluation index of the clustering result corresponding to each cluster number, the normalized clustering evaluation index fusion method is used to determine the optimal cluster number;

[0017] S35: Performing cluster analysis on the enhanced data set with the optimal number of clusters to obtain the HVAC load pattern classification label.

[0018] Furthermore, in S34, according to the evaluation index of the clustering result corresponding to each cluster number, a normalized clustering evaluation index fusion method is adopted to determine the optimal cluster number, including:

[0019] S341: Obtaining evaluation indicators of clustering results corresponding to each cluster number, wherein the evaluation indicators include silhouette coefficient, gap statistic value, DBI value and CH value;

[0020] S342: normalizing the evaluation index of the clustering result corresponding to each cluster number to obtain a normalized score of the evaluation index for each cluster number;

[0021] S343: Selecting the cluster number with the highest normalized evaluation index score as the optimal cluster number.

[0022] Furthermore, in step S4, based on the HVAC load pattern classification label and the original HVAC load data, a feature set is obtained by performing correlation analysis and screening, including:

[0023] performing a continuous variable correlation analysis based on the HVAC load pattern classification label and the original HVAC load data to obtain a preliminary feature set;

[0024] A discrete variable correlation analysis is performed based on the initially selected feature set to obtain the feature set.

[0025] Furthermore, before performing the continuous variable correlation analysis, the method further includes:

[0026] Using sine and cosine functions to periodically encode the time information in the original HVAC load data, and constructing a characteristic variable reflecting the time periodicity;

[0027] Performing discrete encoding on the week information in the original HVAC load data to generate a 7-dimensional one-hot encoding feature;

[0028] Performing a nonlinear transformation on the passenger flow data in the original HVAC load data to eliminate scale differences;

[0029] The characteristic variables reflecting the time periodicity, the 7-dimensional one-hot encoding features, the nonlinearly transformed passenger flow data, and other data of the original HVAC load data are subjected to correlation analysis and screening.

[0030] Furthermore, in step S5, based on the feature set, a multi-classifier ensemble learning model is used for prediction, combined with a voting mechanism, to obtain a load pattern prediction result, including:

[0031] Based on the feature set, the HVAC load pattern classification labels predicted by each classifier are obtained through decision tree, random forest and XGBoost respectively;

[0032] Through the hard voting mechanism, the prediction results of each classifier are aggregated and the HVAC load pattern classification label with the highest frequency is selected as the final prediction result.

[0033] Furthermore, when training each classifier, the grid search cross-validation method is used to tune the hyperparameters of each classifier.

[0034] The embodiments of the present invention have the following technical effects:

[0035] The HVAC load pattern prediction method based on time series decomposition and ensemble learning provided by the present invention first preprocesses the original HVAC load data, unifies the data time intervals, and eliminates the interference of outliers and missing values, thereby providing a high-quality time series data foundation for subsequent analysis. Second, based on time series decomposition, trend components and seasonal components are extracted, which enhances the in-depth exploration of long-term load variation patterns and periodic characteristics, makes up for the defect that traditional methods do not fully capture time series characteristics, and provides a more discriminative enhanced data set for cluster analysis. Furthermore, when generating classification labels through cluster analysis, a dynamic determination strategy for the optimal number of clusters is specifically introduced to solve the pattern division bias problem caused by the fixed or empirical setting of the cluster number in traditional methods, significantly improves the objectivity and accuracy of load pattern classification, and provides a more reliable classification label foundation for subsequent feature screening and model training. Combined with multi-dimensional feature extraction, feature sets with strong correlation with classification labels are screened to avoid the interference of redundant features on the model. Finally, a multi-classifier ensemble learning model combined with a voting mechanism is used for prediction. By complementing the advantages of different models and fusing the results, the stability and generalization ability of the prediction results are enhanced. This solution provides a highly reliable load pattern recognition basis for building energy-saving control through the collaboration of time series feature enhancement, adaptive selection of the optimal clustering number and integrated learning framework, solving the technical problems of unstable classification effect, strong clustering subjectivity and poor model adaptability of traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0037] Figure 1 1 is a diagram showing the relationship between the steps of the HVAC load pattern prediction method based on time series decomposition and ensemble learning provided by an embodiment of the present invention;

[0038] Figure 2 This is a flow chart of an HVAC load pattern prediction method based on time series decomposition and ensemble learning provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0039] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention are described clearly and completely below. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are also within the scope of protection of the present invention.

[0040] By integrating techniques such as time series decomposition, ensemble learning, multi-level feature extraction, periodic transformation, and feature enhancement, this invention achieves accurate and stable classification of subway HVAC load patterns. Each technique is introduced to address the shortcomings of traditional methods. By decomposing data, integrating multiple learning models, considering time dependencies, and enhancing the comprehensiveness of feature data, this method achieves precise identification and efficient classification of load patterns. Ultimately, this provides a solid foundation for energy-saving optimization of subway HVAC systems, with strong applicability and generalization capabilities.

[0041] Depend on Figure 1 and Figure 2 It can be seen that, on one hand, the present invention provides an HVAC load pattern prediction method based on time series decomposition and ensemble learning, comprising the following steps: In order to solve the above technical problem, on one hand, the present invention provides an HVAC load pattern prediction method based on time series decomposition and ensemble learning, comprising the following steps: Step S1: obtaining original HVAC load data, preprocessing the original HVAC load data, and generating preprocessed load time series data;

[0042] Collect relevant time series data as raw HVAC load data, including but not limited to load, passenger flow, outdoor air temperature, outdoor air humidity, date, and other data. Each set of data corresponds to a timestamp to identify the change pattern of the data over time.

[0043] Raw HVAC load data may be obtained from different sources. The time intervals for data collection from different sources may not be uniform, resulting in different time intervals for the load data obtained from each source. In the case of irregularly sampled data, it is necessary to convert the data with different time intervals into data with the same time interval. Resampling the time series data can convert the irregular time interval data into time series data with the same time interval. The following methods can be used:

[0044] (1) Forward filling: fill missing values ​​with the values ​​of the previous moment.

[0045] (2) Linear interpolation: interpolate missing values ​​according to the changing trend between time points to obtain a smooth time series.

[0046] (3) Mean interpolation: fill missing values ​​according to the mean value of adjacent data points.

[0047] In addition, data cleaning is required to complete data preprocessing, including filling missing values ​​with a simple moving average method and removing outlier data.

[0048] The moving average method fills missing values ​​by calculating the average value of the data within a rolling time window around the missing value. Its effectiveness depends on the selected window size. Smaller windows are more sensitive to data changes but are susceptible to noise; larger windows help smooth data but may delay the capture of information. The window size is determined based on the rate of data change. Cross-validation using simulated missing values ​​is used to evaluate the effectiveness of the imputation method with different window sizes. Specifically, a portion of data is randomly deleted from the original complete data as "simulated missing values." Moving averages with different window sizes, 𝑛, are then used to fill missing values. The root mean square error (RMSE) and coefficient of variation (CV) between the imputed results and the original data are then calculated to comprehensively measure the accuracy and stability of the imputation.

[0049] After missing values ​​are supplemented, outliers can also be removed. For example, other methods such as the interquartile range rule and local outliers can be used. Removing outliers provides a high-quality time series data foundation for subsequent analysis.

[0050] Step S2: performing time series decomposition based on the preprocessed load time series data, extracting trend components and seasonal components, and generating an enhanced data set; In some implementations, in step S2, performing time series decomposition based on the preprocessed load time series data, extracting trend components and seasonal components, and generating an enhanced data set specifically includes: S21: decomposing the preprocessed load time series data into trend components, seasonal components, and residual components using the STL method;

[0051] The STL method is a technique for decomposing time series using the locally weighted regression (LOESS) method. t Decompose and output as the sum of three parts: trend component T t , seasonal component S t and the residual component R t , respectively extracting the long-term evolution trend, periodic operation law and random disturbance components, effectively reducing the interference between different components. In some embodiments, the calculation formula of the STL method is:

[0052] ;

[0053] Where Y t Represents the given load data, T t represents the trend component, S t represents the seasonal component, R t represents the residual component.

[0054] Trend component T t It represents the long-term change pattern in time series data and reflects the trend of data over time. It is obtained by applying Loess smoothing to the data after removing the seasonal term, as shown in the following formula:

[0055] ;

[0056] Seasonal ingredient S t It represents the periodic fluctuations in the time series and reflects the repetitive pattern of the data within the period. It is obtained by applying Loess smoothing to the data after removing the trend term. The length of the seasonal cycle is represented by P, as shown in the following formula:

[0057] ;

[0058] Residual component R t It represents the random fluctuations remaining in the data after removing the trend and seasonal terms, which usually reflects noise, abnormal data or unpredictable factors, as shown in the following formula:

[0059] ;

[0060] Seasonal and Trend decomposition using Loess (STL, seasonal and trend decomposition method based on local weighted regression) is a key step in mining the inherent laws of subway HVAC load data. Characterizes the slow changes in load over time, such as the baseline energy consumption shift caused by changes in the thermal performance of the building envelope; the calculation uses a local weighted regression algorithm to achieve smooth extraction of trend characteristics by fitting a low-order polynomial function within a sliding window. Seasonal component Reflects repeated fluctuations within a fixed period, such as the surge in cooling demand during peak hours in the morning and evening. When calculating, the detrended data is first segmented by period length, and the average of the same phase points in each segment is taken to form a baseline seasonal pattern. High-frequency noise is then eliminated through low-pass filtering. Random interference, such as measurement error and unexpected events, is included as a byproduct of the decomposition process and can be used to assess data quality. Large residuals often indicate abnormal events or unique patterns not captured by the model. This formula establishes an additive decomposition model for load data, providing a theoretical basis for subsequent component analysis and feature engineering.

[0061] This decomposition method overcomes the limitation of the traditional Fourier transform, which is only applicable to stationary signals, and can effectively handle the nonlinear and non-stationary characteristics of building loads. The explicit separation of trend and seasonal components provides structured features for subsequent analysis. For example, the slope of the trend component can be used to assess the rate of change of equipment energy efficiency, and the amplitude of the seasonal component can be used to quantify the intensity of cyclical fluctuations, making load pattern classification more physically interpretable.

[0062] By decomposing complex fluctuations into time series, long-term trends, seasonality, and residual components are extracted to enhance data characterization capabilities. The trend component reflects the impact of long-term factors such as building equipment aging and operating strategy adjustments on the load. The seasonal component captures the repetitive fluctuation patterns within fixed periods such as days and weeks, and the residual component retains random interference and unmodeled factors. It effectively analyzes the inherent periodic characteristics of load data (such as daily, weekly, and seasonal changes) and long-term change trends. The decomposed components allow for more accurate analysis of load patterns, avoiding mixing periodic and trend changes into noise. Clustering of long-term trends can identify the long-term operating patterns of subway HVAC systems, while clustering of seasonal components can identify seasonal changes in daily operations. In this way, the potential connections between different load patterns can be deeply explored, improving classification accuracy.

[0063] S22: Merge the trend component and the seasonal component with the original HVAC load data to form the enhanced data set.

[0064] The decomposed trend component T t, seasonal component S t The enhanced data set is constructed by fusing with the preprocessed load time series data, providing multi-dimensional feature input for cluster analysis, breaking through the limitation of traditional methods that only rely on raw data, and improving prediction accuracy.

[0065] Step S3: Based on the enhanced data set, performing cluster analysis by obtaining an optimal number of clusters to obtain an HVAC load pattern classification label. In some embodiments, in step S3, performing cluster analysis by obtaining an optimal number of clusters to obtain an HVAC load pattern classification label based on the enhanced data set includes:

[0066] S31: Determine a range of cluster numbers based on the data volume of the enhanced data set;

[0067] For example, if the data volume of the enhanced data set is M groups, the range of the number of clusters is set to [2, M / 5], and an integer within the range is taken as the number of clusters. The range of the number of clusters can be appropriately expanded or reduced according to actual conditions.

[0068] S32: performing cluster analysis based on the enhanced data set using each cluster number within the cluster number range;

[0069] S33: Calculate the evaluation index of the clustering result corresponding to each cluster number respectively;

[0070] S34: According to the evaluation index of the clustering result corresponding to each cluster number, the normalized clustering evaluation index fusion method is used to determine the optimal cluster number;

[0071] In some implementations, in S34, the optimal number of clusters is determined using a normalized clustering evaluation index fusion method based on the evaluation index of the clustering result corresponding to each number of clusters, including:

[0072] S341: Obtaining evaluation indicators of clustering results corresponding to each cluster number, wherein the evaluation indicators include silhouette coefficient, gap statistic value, DBI value and CH value;

[0073] The silhouette coefficient is an indicator for evaluating the quality of each sample clustering, taking into account the tightness within the cluster and the separation between clusters. The range of the silhouette coefficient is -1 to 1, and the larger the value, the better the clustering effect. By calculating the silhouette coefficient under different K values, the K value with the maximum silhouette coefficient value is selected as the first cluster number, and the similarity difference between each sample and the same cluster and the nearest different cluster is calculated. The overall mean maximum corresponds to the optimal balance between classification tightness and separation. The present invention directly applies the existing silhouette coefficient calculation method to calculate and obtain the silhouette coefficient of the clustering result corresponding to each cluster number.

[0074] Gap Statistic is used to evaluate the rationality of the number of clusters. It determines whether the number of clusters is optimal by comparing the clustering results with the clustering effect of random data. If the clustering effect is significantly better than the clustering effect of random data, the current number of clusters is appropriate. The larger the Gap Statistic value, the better the number of clusters. First, multiple random data sets are generated and the sum of squared errors of these data sets under different numbers of clusters are calculated. The present invention directly applies the existing Gap Statistic value calculation method to obtain the Gap Statistic value of the clustering result corresponding to each number of clusters.

[0075] The Davies-Bouldin Index (DBI) is used to assess the compactness and separation of clusters. The smaller the DBI value, the more compact the clusters and the greater the separation between clusters, indicating a better clustering effect. The K value with the smallest DBI value is selected as the third cluster number, and the classification quality is evaluated by calculating the ratio of the inter-cluster distance to the intra-cluster diameter. The smaller the value, the higher the separation between clusters and the better the compactness within the cluster. The DBI measures the clustering effect by calculating the similarity of each pair of clusters. The present invention directly applies the existing DBI value calculation method to obtain the DBI value corresponding to the clustering result for each cluster number.

[0076] The Calinski-Harabasz index (CH index, also known as the variance ratio criterion) assesses clustering quality by comparing the compactness within clusters with the separation between clusters. A higher CH index indicates better clustering results. The ratio of the inter-cluster dispersion to the intra-cluster dispersion measures classification effectiveness, with a higher value indicating a clearer cluster structure. This paper directly applies existing CH value calculation methods to obtain a CH value for each cluster number.

[0077] S342: normalizing the evaluation index of the clustering result corresponding to each cluster number to obtain a normalized score of the evaluation index for each cluster number;

[0078] The evaluation index of the clustering result corresponding to each cluster number is normalized separately, and each index is normalized to the interval [0, 1] as follows.

[0079] For the maximization index, that is, the larger the value, the better the clustering result: Silhouette coefficient, GapStatistic value and CH value, when normalizing the scores of these three evaluation indicators corresponding to each cluster number, the following formula is used:

[0080] ;

[0081] Where, It represents the normalized score of the indicator when the number of clusters is K, ranging from [0,1], M KIt represents the original calculated value of an evaluation indicator when the number of clusters is K. min(M) represents the minimum value of an indicator under all K values, and max(M) represents the maximum value of an indicator under all K values.

[0082] Taking the silhouette coefficient as an example, the above formula is used to calculate the silhouette coefficient of the clustering result corresponding to each cluster number. It represents the normalized score of the silhouette coefficient when the number of clusters is K, corresponding to the following formula , M K It represents the original calculated value of the silhouette coefficient when the number of clusters is K, min(M) represents the minimum value of the silhouette coefficient under all K values, and max(M) represents the maximum value of the silhouette coefficient under all K values.

[0083] Similarly, calculate the normalized score of the Gap Statistic value and the normalized score of CH value .

[0084] For the minimization index, the smaller the value, the better the clustering result, that is, the DBI value, is calculated using the following formula.

[0085] ;

[0086] It represents the normalized score of DBI value when the number of clusters is K, ranging from [0,1]. K It represents the original calculated value of DBI value when the number of clusters is K, min(DBI) represents the minimum value of DBI value under all K values, and max(DBI) represents the maximum value of DBI value under all K values.

[0087] Then, according to the normalized score of each evaluation indicator, the following formula is used to calculate the normalized score of the evaluation indicator for each cluster number:

[0088] ;

[0089] S K Indicates the normalized score of the evaluation index when the number of clusters is K.

[0090] S343: Select the cluster number with the highest normalized evaluation index score as the optimal cluster number. This normalized cluster evaluation index fusion method ensures that the load mode division is consistent with the inherent structure of the data and meets the requirements of engineering applications for the stability of classification results.

[0091] S35: Performing cluster analysis on the enhanced data set with the optimal number of clusters to obtain the HVAC load pattern classification label.

[0092] Exemplarily, the trend component and seasonal component in the enhanced dataset are first standardized and normalized to eliminate the influence of different feature scales.

[0093] Normalization is to make the data follow a standard normal distribution by subtracting the mean and dividing by the standard deviation. Normalization is to scale the data to the interval [0,1].

[0094] For example, an appropriate clustering method (such as K-Means, K-Medoids, GMM, etc.) is selected to cluster the standardized and normalized trend and seasonal component data. The clustering results assign a cluster label to each data point, indicating its category. During the clustering process, load data with similar patterns are grouped together based on data similarity, and corresponding cluster labels are generated to serve as HVAC load pattern classification labels. For example, matching patterns are classified into Mode 1, Mode 2, etc., or peak mode, flat mode, etc.

[0095] Step S4: Based on the HVAC load mode classification label and the original HVAC load data, a feature set is obtained by screening through correlation analysis;

[0096] The obtained HVAC load pattern classification labels are used as new features and added to the original HVAC load data to form a new data set for correlation analysis.

[0097] In some embodiments, step S4, based on the HVAC load pattern classification label and the original HVAC load data, obtaining a feature set through correlation analysis and screening, includes:

[0098] In some embodiments, before performing the continuous variable correlation analysis, the method further comprises:

[0099] S4A: The time information in the original HVAC load data is periodically encoded using sine and cosine functions to construct characteristic variables reflecting time periodicity. The characteristic variables reflecting time periodicity retain the periodic characteristics of time, making it easier for the model to identify the time dependence of the load pattern.

[0100] S4B: Discretely encode the week information in the original HVAC load data to generate a 7-dimensional one-hot encoded feature. Discretely encode the week information in the original HVAC load data to generate a 7-dimensional one-hot encoded feature. The original week data (such as the numbers 1 to 7) may mislead the model into thinking that the relationship between week 1 and week 2 is linear or sequential, when in fact they are just different categories. This turns each week into an independent feature, eliminating this erroneous "sequential relationship."

[0101] S4C: Perform a nonlinear transformation on the passenger flow data in the original HVAC load data to eliminate scale differences. A logarithmic transformation of the number of people entering and exiting the station is performed to prevent the large magnitude difference between the passenger flow data and other data from affecting the analysis results, thereby improving classification accuracy and enhancing the model's ability to learn the relationships between different features.

[0102] S4D: performing correlation analysis and screening on the characteristic variables reflecting the time periodicity, the 7-dimensional one-hot encoding features, the nonlinearly transformed passenger flow data, and other data of the original HVAC load data.

[0103] Multi-dimensional feature extraction aims to construct a comprehensive system of influencing factors that characterize load patterns. During time feature conversion, hourly information is mapped to a continuous periodic variable using a sine-cosine transform to avoid feature abrupt changes caused by treating 00:00 and 23:59 as discrete points. Weekday information is one-hot encoded to generate a vector of 0 / 1, preserving the independent influence of each weekday. Passenger flow data is log-encoded to prevent large magnitude differences from affecting analysis results.

[0104] S41: performing a continuous variable correlation analysis based on the HVAC load pattern classification label and the original HVAC load data to obtain a preliminary feature set;

[0105] Continuous variable correlation analysis uses variance analysis to test whether there are significant mean differences between different categories of continuous variables. If the mean differences between different categories are large, it indicates that the relationship between the continuous variable and the categorical output variable is strong. When the input variables are continuous and the output variables are categorical, one-way analysis of variance (One-Way ANOVA) can be used. ANOVA tests the difference between each continuous input variable and the output category to obtain the correlation between each category of data in the original HVAC load data and the HVAC load pattern classification label. In other words, factors that have an impact on the HVAC load pattern classification label are screened for subsequent prediction calculations to avoid inputting too many irrelevant factors and reducing prediction efficiency. The present invention takes a significance level of p ≤ (e.g., 0.05) to screen and obtain a preliminary feature set from the original HVAC load data.

[0106] S42: Performing discrete variable correlation analysis based on the preliminarily selected feature set to obtain the feature set.

[0107] Similarly, a discrete variable correlation analysis is performed based on the preliminary feature set. For example, a Chi-Square Test is selected to evaluate the correlation between each category of data and the HVAC load mode classification label in the preliminary feature set, and the significance level (e.g., 0.05) is also taken.

[0108] The preliminary feature set screened by discrete variable correlation analysis is merged with the HVAC load mode classification label to generate a feature set.

[0109] Step S5: Based on the feature set, a multi-classifier ensemble learning model is used for prediction, combined with a voting mechanism, to obtain a load pattern prediction result.

[0110] In some implementations, in step S5, based on the feature set, a multi-classifier ensemble learning model is used for prediction, combined with a voting mechanism, to obtain a load pattern prediction result, including:

[0111] S51: Based on the feature set, obtain the HVAC load pattern classification label predicted by each classifier through decision tree, random forest and XGBoost respectively;

[0112] Ensemble learning frameworks enhance predictive robustness through the complementary nature of heterogeneous models. Decision trees use information gain ratios to select split features and construct intuitive load pattern discrimination rules, but are sensitive to noisy data and prone to overfitting. Random forests construct multiple decision trees through bootstrap sampling, introducing feature randomness to reduce variance but potentially overlooking subtle local patterns. XGBoost uses a gradient boosting strategy to iteratively optimize the loss function, controlling model complexity through regularization. While it excels at capturing nonlinear relationships, it takes a long time to train.

[0113] In some implementations, when training each classifier, a grid search cross-validation method is used to tune the hyperparameters of each classifier.

[0114] S52: Through the hard voting mechanism, the prediction results of each classifier are aggregated and the HVAC load pattern classification label with the highest frequency is selected as the final prediction result.

[0115] The hard voting mechanism requires each base classifier to make independent predictions, ultimately selecting the category with the highest number of votes. This democratic decision-making mechanism effectively mitigates the influence of misjudgments by individual models. For example, if two classifiers identify a "peak mode" and one as a "flat mode," the system will ultimately output a "peak mode." Stratified sampling during the training phase ensures balanced samples across categories, preventing voting from favoring high-frequency categories.

[0116] Ensemble learning improves the robustness and generalization of the model by combining multiple base classifiers for voting. Each classifier can analyze the load pattern from a different perspective, and the voting mechanism can reduce the error of individual classifiers by aggregating the judgments of different models. The model integration strategy fully leverages the advantages of each algorithm. The decision tree provides highly interpretable rules, the random forest ensures generalization performance, and XGBoost captures complex nonlinear relationships. This combination enables the system to handle both linearly separable basic patterns and identify complex working conditions with multiple coupled factors, significantly improving its adaptability in different building scenarios. This improves the robustness and generalization of the model, allowing it to be applied in different subway stations and different operating conditions, thereby improving the applicability of the method.

[0117] For example, the construction and training of the ensemble learning model are as follows:

[0118] Features with high feature importance contribute significantly to model predictions and are generally preferred during feature selection. Classifiers are trained to predict the class labels for different features. Ensembling different classifiers improves classification accuracy and generalization.

[0119] (1) Data division

[0120] Divide the dataset into training and test sets. For example, use 70% of the data for training and 30% for testing to ensure that the training and test datasets are representative and effectively avoid overfitting.

[0121] (2) Model construction and training

[0122] Step A: Hyperparameter Tuning

[0123] Adjust the hyperparameters for each classifier. The specific hyperparameters to be adjusted include:

[0124] Decision Trees: Adjust the maximum depth of the tree and the minimum number of samples per node split.

[0125] Random Forest: Adjust the number of decision trees and the maximum number of features used per tree.

[0126] XGBoost: Adjust hyperparameters such as learning rate, maximum tree depth, number of trees, and explicitly set evaluation metrics.

[0127] A grid search with cross-validation method was used. This method exhaustively enumerates preset hyperparameter combinations and performs cross-validation evaluation under each combination, selecting the parameter configuration with the best overall performance as the final model setting. The performance evaluation of each parameter combination is achieved through K-fold cross-validation, which divides the dataset into Z subsets, uses (Z-1) subsets for training in each iteration, and uses the remaining subset for validation. After Z cycles, the average evaluation index is taken as the final score of the parameter combination. This method can effectively reduce the accidental effects caused by different data partitioning and improve the stability and reliability of parameter selection. Finally, the optimal parameter configuration selected by this method is used in subsequent model training and integration processes.

[0128] Step B: Model training

[0129] In some embodiments, when training each classifier, a grid search cross-validation method is used to tune the hyperparameters of each classifier separately.

[0130] The training set is trained using three classification models: decision tree, random forest, and XGBoost. Each classifier model uses a different learning strategy to capture the patterns in the data:

[0131] Decision tree: By recursively splitting the feature space, the best partitioning features are selected using information gain or Gini index.

[0132] Random Forest: Improves model stability by integrating multiple decision trees and performing voting.

[0133] XGBoost: Uses the gradient boosting tree method to gradually optimize the model to improve prediction accuracy.

[0134] (3) Ensemble learning and classification prediction

[0135] Ensemble learning methods were used to improve model performance. A voting classifier was constructed by integrating various classifiers (decision tree, random forest, and XGBoost), completing the overall construction of the ensemble learning model. Ensembling during training improved the prediction stability and accuracy of the final model. The voting classifier used a hard voting method, making decisions based on the predictions of each classifier and selecting the most frequent HVAC load pattern as the final prediction.

[0136] Train the ensemble learning model and make predictions on the test set. Use multiple metrics to evaluate the performance of the ensemble learning model.

[0137] Accuracy: The proportion of samples that the model predicts correctly to the total samples.

[0138] Precision: The proportion of samples predicted to be positive that are actually positive.

[0139] Recall: The proportion of samples that are actually positive that are correctly predicted to be positive.

[0140] F1-Score: The harmonic mean of precision and recall, which comprehensively considers the performance of the model on positive samples.

[0141] Cross-validation: Use cross-validation to evaluate the ensemble learning model to ensure that the model performs stably on different data subsets and further avoid overfitting.

[0142] This paper constructs a systematic framework of "time series decomposition-cluster analysis-feature engineering-integrated prediction." First, by preprocessing the raw HVAC load data, unifying the data time intervals and eliminating outliers and missing values, this method provides a high-quality time series data foundation for subsequent analysis. Second, by extracting trend and seasonal components based on time series decomposition, this method enhances the in-depth exploration of long-term load variation patterns and cyclical characteristics, addressing the shortcomings of traditional methods in fully capturing time series characteristics and providing a more discriminative enhanced dataset for cluster analysis. Furthermore, when generating classification labels through cluster analysis, a dynamic strategy for determining the optimal number of clusters is introduced, addressing the pattern division bias caused by fixed or empirically set cluster numbers in traditional methods. This significantly improves the objectivity and accuracy of load pattern classification and provides a more reliable classification label foundation for subsequent feature screening and model training. Multi-dimensional feature extraction is combined with the selection of feature sets that are strongly correlated with the classification labels, avoiding interference from redundant features on the model. Finally, a multi-classifier ensemble learning model combined with a voting mechanism is used for prediction. By complementing the strengths of different models and integrating their results, the stability and generalization of the prediction results are enhanced. This solution provides a highly reliable load pattern recognition basis for building energy-saving control through the collaboration of time series feature enhancement, adaptive selection of the optimal clustering number and integrated learning framework, solving the technical problems of unstable classification effect, strong clustering subjectivity and poor model adaptability of traditional methods.

[0143] Another aspect of the present invention provides an HVAC load pattern prediction system based on time series decomposition and ensemble learning, which implements any of the above-mentioned HVAC load pattern prediction methods based on time series decomposition and ensemble learning, including: a raw data acquisition module, a data enhancement module, an HVAC load pattern classification label determination module, a feature set generation module, and an output module;

[0144] The original data acquisition module is used to acquire original HVAC load data, preprocess the original HVAC load data, and generate preprocessed load time series data; the data enhancement module is connected to the original data acquisition module, and is used to perform time series decomposition based on the preprocessed load time series data, extract trend components and seasonal components, and generate an enhanced data set; the HVAC load pattern classification label determination module is connected to the data enhancement module, and is used to obtain the HVAC load pattern classification label by performing cluster analysis based on the enhanced data set by obtaining the optimal number of clusters; the feature set generation module is connected to the HVAC load pattern classification label determination module and the original data acquisition module, and is used to obtain a feature set based on the HVAC load pattern classification label and the original HVAC load data through correlation analysis screening; the output module is connected to the feature set generation module, and is used to obtain a load pattern prediction result based on the feature set through a multi-classifier ensemble learning model prediction combined with a voting mechanism.

[0145] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the technical solutions of the embodiments of the present invention.

Claims

1. HVAC load pattern prediction method based on time series decomposition and ensemble learning, characterized by: The following steps are involved: Step S1: obtaining original HVAC load data, preprocessing the original HVAC load data, and generating preprocessed load time series data; Step S2: performing time series decomposition based on the preprocessed load time series data, extracting trend components and seasonal components, and generating an enhanced data set; The enhanced data set includes the trend component, the seasonal component, and the original HVAC load data; Step S3: Based on the enhanced data set, performing cluster analysis by obtaining the optimal number of clusters to obtain HVAC load pattern classification labels; Specifically include: S31: Determine a range of cluster numbers based on the data volume of the enhanced data set; S32: performing cluster analysis based on the enhanced data set using each cluster number within the cluster number range; S33: Calculate the evaluation index of the clustering result corresponding to each cluster number respectively; S34: According to the evaluation index of the clustering result corresponding to each cluster number, the normalized clustering evaluation index fusion method is used to determine the optimal cluster number; S35: performing cluster analysis on the enhanced data set with the optimal number of clusters to obtain the HVAC load mode classification label; Step S4: obtaining a feature set by performing correlation analysis and screening based on the HVAC load mode classification label and the original HVAC load data; Specifically include: Based on the HVAC load pattern classification label and the original HVAC load data, a continuous variable correlation analysis is performed to obtain a preliminary feature set; Performing discrete variable correlation analysis based on the preliminary feature set to obtain the feature set; Step S5: Based on the feature set, a multi-classifier ensemble learning model is used for prediction, combined with a voting mechanism, to obtain a load pattern prediction result.

2. The HVAC load pattern prediction method based on time series decomposition and ensemble learning according to claim 1 is characterized in that: In the step S2, time series decomposition is performed based on the preprocessed load time series data, trend components and seasonal components are extracted, and an enhanced data set is generated, which specifically includes: S21: decomposing the preprocessed load time series data into trend components, seasonal components and residual components using the STL method; S22: merging the trend components and the seasonal components with the original HVAC load data to form the enhanced data set.

3. The HVAC load pattern prediction method based on time series decomposition and ensemble learning according to claim 2 is characterized in that: The calculation formula of the STL method is: ； Where Y t Represents the given load data, T t represents the trend component, S t represents the seasonal component, R t represents the residual component.

4. The HVAC load pattern prediction method based on time series decomposition and ensemble learning according to claim 1 is characterized in that: In S34, based on the evaluation index of the clustering result corresponding to each cluster number, a normalized clustering evaluation index fusion method is used to determine the optimal cluster number, including: S341: Obtain evaluation indicators of the clustering results corresponding to each cluster number, wherein the evaluation indicators include silhouette coefficient, GapStatistic value, DBI value and CH value; S342: normalizing the evaluation index of the clustering result corresponding to each cluster number to obtain a normalized score of the evaluation index for each cluster number; S343: Selecting the cluster number with the highest normalized evaluation index score as the optimal cluster number.

5. The HVAC load pattern prediction method based on time series decomposition and ensemble learning according to claim 1 is characterized in that: Before performing the continuous variable correlation analysis, the method further includes: Using sine and cosine functions to periodically encode the time information in the original HVAC load data, and constructing a characteristic variable reflecting the time periodicity; Performing discrete encoding on the week information in the original HVAC load data to generate a 7-dimensional one-hot encoding feature; Performing a nonlinear transformation on the passenger flow data in the original HVAC load data to eliminate scale differences; The characteristic variables reflecting the time periodicity, the 7-dimensional one-hot encoding features, the nonlinearly transformed passenger flow data, and other data of the original HVAC load data are subjected to correlation analysis and screening.

6. The HVAC load pattern prediction method based on time series decomposition and ensemble learning according to claim 1 is characterized in that: In step S5, based on the feature set, a multi-classifier ensemble learning model is used for prediction, combined with a voting mechanism, to obtain a load pattern prediction result, including: Based on the feature set, the HVAC load pattern classification labels predicted by each classifier are obtained through decision tree, random forest and XGBoost respectively; Through the hard voting mechanism, the prediction results of each classifier are aggregated and the HVAC load pattern classification label with the highest frequency is selected as the final prediction result.

7. The HVAC load pattern prediction method based on time series decomposition and ensemble learning according to claim 6 is characterized in that: When training each classifier, the grid search cross-validation method is used to tune the hyperparameters of each classifier.

Images