Building energy consumption prediction method and system based on multi-frequency feature decomposition and double-branch modeling

Abstract

The application provides a building energy consumption prediction method and system based on multi-frequency feature decomposition and double-branch modeling. The method introduces a complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) algorithm to perform multi-scale decomposition on the building energy consumption signal, significantly improving the decomposability and signal stability of the original energy consumption sequence. The CEEMDAN decomposition mechanism ensures that different frequency components are effectively distinguished before modeling, avoiding the mode aliasing and feature distortion problems caused by directly modeling non-stationary signals. Furthermore, through an adaptive grouping strategy based on variance entropy, the decomposed signals are automatically classified and reconstructed, enabling the high-frequency disturbance and low-frequency trend to be modeled specifically, further improving the feature expression ability and generalization performance of the model.

Description

Technical Field

[0001] This invention relates to the field of energy consumption control technology, and in particular to a method and system for predicting building energy consumption using multi-frequency feature decomposition and bi-branch modeling. Background Technology

[0002] Accurate building energy consumption forecasting can not only improve building energy efficiency, but also provide decision support for building energy consumption optimization scheduling, intelligent control and low-carbon operation.

[0003] Building energy consumption series typically exhibit significant randomness, nonlinearity, and multi-scale fluctuations, influenced by various factors such as meteorological conditions, human activities, equipment operation, and differences in building type. This complexity makes it difficult for traditional linear models to fit effectively, leading to decreased prediction accuracy and insufficient generalization ability.

[0004] In Building Energy Management Systems (BEMS), energy consumption forecasts directly impact the scheduling of Heating, Ventilation, and Air Conditioning (HVAC) systems, demand response strategies, and the optimized operation of distributed energy systems. Therefore, constructing high-accuracy, robust, and interpretable building energy consumption forecasting models is of significant practical importance for achieving intelligent building energy management and carbon emission reduction.

[0005] In recent years, with the development of artificial intelligence and signal processing technologies, researchers have begun to explore building energy consumption prediction methods from the perspective of integrating multi-scale feature modeling with deep learning. By combining signal decomposition, multi-frequency modeling, and nonlinear feature learning, the non-stationarity of energy consumption signals can be alleviated to some extent, improving the predictive performance of the model across multiple time scales. However, how to effectively distinguish different frequency features at the decomposition level and achieve the synergistic fusion of high- and low-frequency features at the modeling level remains the main challenge and direction of current research.

[0006] Currently, research on building energy consumption prediction has formed a relatively systematic technical framework, mainly including four categories: statistical modeling methods, machine learning methods, deep learning methods, and hybrid modeling methods that combine signal decomposition and fusion. Different technical approaches have their own characteristics in terms of model structure, prediction mechanisms, and applicable scenarios, but they still have significant shortcomings in addressing the nonlinear, multi-scale, and non-stationary characteristics of building energy consumption data.

[0007] In early research, statistical models were widely used due to their simple structure, high computational efficiency, and strong interpretability. Typical examples include autoregressive (AR), autoregressive moving average (ARMA), and seasonal autoregressive integrated moving average (SARIMA) models. These models model energy consumption time series using linear equations, capable of depicting short-term fluctuations and seasonal periodicity. However, they rely on the assumption of time series stationarity, are sensitive to noise, and struggle to capture the nonlinear effects of external environmental factors (such as weather variables, human behavior, and equipment start-up and shutdown) on energy consumption. Therefore, their predictive performance significantly deteriorates when dealing with non-stationary and complex dynamic load data.

[0008] With the development of data-driven technologies, machine learning methods have gradually become the mainstream approach for building energy consumption prediction. Through nonlinear mapping mechanisms, models such as Support Vector Regression (SVR), Random Forest (RF), Decision Tree (DT), and Gradient Boosting Decision Tree (GBDT) have achieved superior performance compared to traditional linear models across various building types. They can automatically learn the complex relationships between input variables and energy output, exhibiting strong generalization ability, especially when considering weather characteristics, time characteristics, and holiday effects. However, machine learning models generally assume that samples are independent of each other, lacking explicit modeling of time-dependent structures. Furthermore, these methods typically rely on manually constructed features (such as time delay terms and interaction variables), which can easily lead to overfitting and computational complexity issues in high-dimensional, multi-feature scenarios.

[0009] In recent years, deep learning methods have been widely adopted in the field of building energy consumption prediction due to their outstanding capabilities in temporal modeling and feature self-learning. Recurrent Neural Networks (RNNs) and their improved structures—Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRUs)—can automatically capture long- and short-term time-dependent features in energy consumption data, improving the ability to describe nonlinear dynamic changes. Meanwhile, combined models of Convolutional Neural Networks (CNNs) and Bidirectional Long Short-Term Memory (BiLSTM) and Bidirectional Gated Recurrent Units (BiGRUs) further enhance the joint extraction capability of local and global temporal features. Although deep models have advantages in accuracy, their training process is highly sensitive to data scale, hyperparameter configuration, and computational resources. When high-frequency noise or sudden fluctuations exist in the input data, the model is prone to overfitting or gradient degradation, leading to insufficient prediction stability. Furthermore, the "black box" nature of deep models limits their interpretability in practical engineering applications.

[0010] To overcome non-stationarity issues, signal decomposition and prediction fusion methods have gained increasing attention. These methods improve overall prediction accuracy by performing multi-scale decomposition on the energy consumption sequence before prediction, breaking down complex signals into several frequency components and modeling them separately. Typical examples include Empirical Mode Decomposition (EMD), Ensemble Empirical Mode Decomposition (EEMD), and Complete Ensemble Empirical Mode Decomposition Adaptive Noise (CEEMDAN). Researchers often combine these decomposition algorithms with machine learning or deep learning models to form a hybrid prediction framework of "decomposition-modeling-reconstruction." For example, models such as CEEMDAN–SVR, CEEMDAN–RF, CEEMDAN–LSTM, and CEEMDAN–GRU have shown superior performance compared to single models on various types of building load data. Through decomposition, complex energy consumption sequences are transformed into intrinsic mode functions (IMFs) at different time scales. Each component can be independently modeled to capture dynamic features in a specific frequency band, thereby effectively alleviating the non-stationarity and mode coupling problems of the original sequence.

[0011] However, existing decomposition and fusion models generally use a uniform predictor to model all frequency components, ignoring the differences in statistical characteristics, dynamic behavior, and physical meaning between different frequency bands. For example, high-frequency components typically reflect short-term disturbances and random fluctuations, while low-frequency components contain long-term trends and seasonal variations. Treating both with the same type of model limits modeling accuracy and may even introduce information interference. Furthermore, most existing methods only use simple weighting or linear superposition in the fusion stage, failing to achieve deep coupling of cross-frequency information at the feature level, making it difficult to fully leverage the complementary advantages of different models. Therefore, while existing technologies have improved the accuracy of building energy consumption prediction to some extent, there is still significant room for improvement in multi-frequency modeling and fusion mechanisms. Summary of the Invention

[0012] The embodiments of the present invention provide a building energy consumption prediction method and system based on multi-frequency feature decomposition and bi-branch modeling, which is used to solve the problems existing in the prior art.

[0013] To achieve the above objectives, the present invention adopts the following technical solution.

[0014] A building energy consumption prediction method based on multi-frequency eigenvalue decomposition and bi-branch modeling includes: S1 performs multi-scale decomposition on the target building's energy consumption time series to obtain the intrinsic mode function and residual trend term; S2 divides the intrinsic mode functions into frequency bands based on the variance entropy index of each intrinsic mode function. The intrinsic mode functions above the threshold are divided into high-frequency component groups, and the intrinsic mode functions below or equal to the threshold and the residual trend term are divided into low-frequency component groups. The high-frequency signal and the low-frequency signal are then reconstructed respectively. S3 uses high-frequency signals as input to train the first branch prediction model and low-frequency signals as input to train the second branch prediction model. S4 performs an additive fusion operation on the prediction results of the first branch prediction model and the second branch prediction model to obtain the energy consumption prediction results of the target building. Building energy consumption prediction results are used to provide load prediction basis for the air conditioning control system of the target building, so as to achieve optimized control of the air conditioning system of the target building.

[0015] Preferably, the multi-scale decomposition operation in step S1 is performed using a fully ensemble empirical mode decomposition adaptive noise algorithm, specifically including: Through

[0016] The energy consumption time series of the target building is decomposed into multiple non-interfering components, each corresponding to the energy consumption oscillation characteristics at different time scales; where, Indicates the first One modal component, This represents the residual trend term.

[0017] Preferably, step S2 includes: Through

[0018] The variance entropy index of the intrinsic mode functions is calculated; With the median of all entropy values As a dynamic threshold, through the formula

[0019] The target building's energy consumption time series is reconstructed into high-frequency component groups and low-frequency component groups; Through

[0020] The high-frequency component group and the low-frequency component group are superimposed to form a high-frequency signal and a low-frequency signal, respectively.

[0021] Preferably, step S3 includes: Through

[0022] Based on the preset sliding window length L The first input signal for constructing the high-frequency signal; The first input signal is fed into the first branch model for training, and the first predicted value at the corresponding time step is obtained. ; Through

[0023] The time-series features of the low-frequency signal are updated using gated recurrent units to obtain the hidden state sequence representation for training the second branch model; where, To update the door, In candidate activation state; The hidden state sequence representation corresponding to the low-frequency signal is input into the second branch model for training, and intermediate features are obtained. ; Based on intermediate features The mapping operation is performed to obtain the second predicted value. ; For the first predicted value Second predicted value Perform uniform labeling processing on each, so that the first predicted value Second predicted value At the predicted time The above corresponds one-to-one and is used in step S4.

[0024] Preferably, step S4 includes: Through

[0025] The first and second predicted values ​​with unified labels output by the first and second branch models are fused together.

[0026] Secondly, the present invention provides a building energy consumption prediction system based on multi-frequency eigenvalue decomposition and bi-branch modeling, comprising: The signal decomposition module is used to: perform multi-scale decomposition operations on the energy consumption time series of the target building in real time to obtain the intrinsic mode function and residual trend term; The feature grouping module is used to: divide the intrinsic mode functions into frequency bands based on the variance entropy index of each intrinsic mode function, divide the intrinsic mode functions above the threshold into high frequency component groups, divide the intrinsic mode functions below or equal to the threshold and residual trend terms into low frequency component groups, and reconstruct the high frequency signal and low frequency signal respectively. The model training module is used to: train the first branch prediction model with high-frequency signals as input, and train the second branch prediction model with low-frequency signals as input. The predictive fusion module is used for: An additive fusion operation is performed on the prediction results of the first branch prediction model and the second branch prediction model to obtain the energy consumption prediction results of the target building. The building energy consumption forecast results are input into the air conditioning control system of the target building to provide load forecast basis for the optimized control of the air conditioning of the target building.

[0027] As can be seen from the technical solutions provided by the embodiments of the present invention above, the present invention provides a building energy consumption prediction method and system based on multi-frequency feature decomposition and bi-branch modeling. By introducing the Complete Ensemble Empirical Mode Decomposition Adaptive Noise Algorithm (CEEMDAN) to decompose building energy consumption signals at multiple scales, the decomposability and signal stability of the original energy consumption sequence are significantly improved. The CEEMDAN decomposition mechanism ensures that different frequency components are effectively distinguished before modeling, avoiding the mode aliasing and feature distortion problems caused by directly modeling non-stationary signals. Furthermore, through an adaptive grouping strategy based on variance entropy, the decomposed signals are automatically classified and reconstructed, enabling high-frequency disturbances and low-frequency trends to be modeled in a targeted manner, further improving the feature representation ability and generalization performance of the model.

[0028] A two-branch collaborative modeling approach, employing a Lightweight Gradient Boosting Machine (LightGBM) and a Bidirectional Gated Recurrent Unit (BiGRU), significantly reduces the fitting error of a single prediction model across different frequency features. Furthermore, combining this with Bayesian optimization enables adaptive hyperparameter search, improving training efficiency and global optimality on complex building energy consumption data. The two-branch structure captures sudden load changes in high-frequency channels and extracts long-term trend features in low-frequency channels, enhancing the system's robustness and predictive stability across multiple time scales.

[0029] A more accurate multi-scale modeling strategy enables building energy consumption prediction models to achieve finer fitting based on actual energy consumption fluctuations, reducing prediction bias caused by the coupling of short-term fluctuations and long-term trends. This invention has achieved significant beneficial effects in improving the accuracy and stability of building energy consumption prediction models, as demonstrated by experimental results. Figure 3-5 As shown. The method of this invention, through the synergistic effect of high and low frequency branches and the additive fusion mechanism, enables the model to better reproduce the trend of energy consumption curve changes, while effectively reducing the overall prediction error and improving the load prediction accuracy and energy utilization efficiency of the building energy management system.

[0030] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of the invention. Attached Figure Description

[0031] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 The flowchart of the building energy consumption prediction method based on multi-frequency feature decomposition and bi-branch modeling provided by the present invention is shown. Figure 2 A process diagram of a preferred embodiment of the building energy consumption prediction method based on multi-frequency eigendecomposition and bi-branch modeling provided by the present invention; Figure 3 A comparison chart of prediction results for various types of building energy consumption datasets in a preferred embodiment of the building energy consumption prediction method based on multi-frequency feature decomposition and bi-branch modeling provided by the present invention. Figure 4Ablation experiment results of a key design module of a preferred embodiment of the building energy consumption prediction method of multi-frequency eigendecomposition and bi-branch modeling provided by the present invention; Figure 5 A comparison of MAPE error distributions for different model categories in a preferred embodiment of the building energy consumption prediction method based on multi-frequency eigendecomposition and bi-branch modeling provided by the present invention. Figure 6 The logical block diagram of the building energy consumption prediction system based on multi-frequency feature decomposition and bi-branch modeling provided by this invention. Detailed Implementation

[0033] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0034] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or couplings. The term “and / or” as used herein includes any and all combinations of one or more of the associated listed items.

[0035] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0036] To facilitate understanding of the embodiments of the present invention, the following will provide further explanation and description with reference to the accompanying drawings and several specific embodiments. These embodiments do not constitute a limitation on the embodiments of the present invention.

[0037] This invention provides a building energy consumption prediction method and system based on multi-frequency eigenvalue decomposition and bi-branch modeling, to solve the following technical problems existing in the prior art: While existing decomposition-fusion-based building energy consumption prediction methods have improved the non-stationarity and complexity of energy consumption sequences to some extent, they still have significant limitations overall. These shortcomings are mainly reflected in the model's frequency feature processing, branch matching, multi-scale fusion mechanism, adaptive optimization capability, and the interpretability of prediction results, which seriously restrict the promotion and application of the model in complex building scenarios.

[0038] First, regarding frequency feature processing mechanisms, existing methods generally use isomorphic predictors to model each frequency component obtained from signal decomposition, failing to differentiate the statistical characteristics and dynamic differences at different frequency levels. Building energy consumption signals typically contain high-frequency disturbances, low-frequency trends, and periodic components. High-frequency components reflect random fluctuations such as equipment start-up and shutdown, and human behavior, while low-frequency components characterize long-term trends such as climate change and seasonal cycles. Using the same modeling framework for all components leads to insufficient fitting ability for high-frequency components, decreased accuracy in capturing low-frequency trends, and significantly limited generalization performance of the model at different scales.

[0039] Secondly, regarding model structure matching and task allocation mechanisms, most existing studies adopt an "algorithm stacking" fusion strategy, simply stacking machine learning or deep learning models after decomposition modules, lacking a systematic design for matching model affiliation based on frequency features. This non-targeted combination approach ignores the differences in inductive bias among different models. For example, tree models are better at handling discontinuous high-frequency fluctuations, while recurrent networks are better at capturing smooth time-dependent structures. Mixing two types of features into the same model not only leads to information interference and gradient competition but may also compromise model convergence, causing instability and accumulated bias in prediction results.

[0040] Furthermore, regarding multi-scale feature fusion and cross-frequency interaction mechanisms, most existing methods employ simple linear weighting or averaging fusion strategies, directly superimposing the prediction results from different frequency branches to obtain the final output. This type of fusion fails to achieve deep alignment and semantic coupling at the feature level, lacking the ability to explore the potential complementarity between high- and low-frequency features, resulting in insufficient model capabilities in overall information representation and global trend reconstruction. Moreover, because features in different frequency bands have non-linear dependencies in the time and frequency domains, simple additive fusion cannot effectively recover the true dynamic structure of the energy consumption signal, causing biases in model predictions at extreme points or abrupt changes.

[0041] Furthermore, regarding model parameter optimization and adaptability, some existing works rely on empirical manual hyperparameter tuning or grid search to determine hyperparameters, lacking automated and global optimization mechanisms. When faced with datasets of different building types, climate zones, or sampling frequencies, these models often require retraining or manual parameter tuning to maintain stable performance, limiting the universality and engineering scalability of the methods. For high-frequency components of the model, without suitable regularization and hyperparameter search strategies, it is prone to overfitting in local noise, further weakening the model's generalization ability.

[0042] Finally, regarding model interpretability and generalizability, although deep learning models have a clear advantage in accuracy, the lack of transparency in their internal feature extraction process makes it difficult to trace the prediction results back to causality and physical explanation at the frequency level. This "black box" characteristic prevents the model's decision-making basis from being directly mapped to the building's operating mechanism, weakening its credibility and usability in building energy consumption optimization, anomaly detection, and energy management systems (BEMS). Especially in actual deployments involving multiple building types and across regions, insufficient model interpretability often becomes a core obstacle to application and promotion.

[0043] In summary, existing building energy consumption prediction methods still suffer from systemic deficiencies in multi-scale signal decomposition and feature fusion: insufficient frequency hierarchical modeling, mismatch between model structure and signal features, inadequate cross-frequency information integration, limited adaptive optimization capabilities, and weak interpretability. Therefore, it is necessary to propose a novel multi-scale fusion prediction framework. By distinguishing between high- and low-frequency features during the signal decomposition stage and employing a differentiated bi-branch learning structure during the modeling stage, collaborative modeling and additive fusion of multi-frequency information can be achieved, thereby significantly improving the accuracy, stability, and interpretability of building energy consumption prediction.

[0044] In view of this, the object of the present invention is: 1. A building energy consumption prediction framework based on multi-frequency feature decomposition and bi-branch modeling is designed. This framework introduces the Complete Ensemble Empirical Mode Decomposition Adaptive Noise Algorithm (CEEMDAN) to decompose complex non-stationary building energy consumption sequences into multiple subsequences with independent frequency features, effectively separating high-frequency disturbances from low-frequency trends. This fundamentally alleviates the mismatch problem in multi-scale feature modeling of traditional models and significantly improves the interpretability and stability of energy consumption data modeling.

[0045] 2. To address the dynamic characteristics of different frequency components, a dual-branch prediction network combining high and low frequencies was constructed. The high-frequency branch employs the LightGBM model to capture nonlinear short-term fluctuations in building energy consumption, while the low-frequency branch uses a bidirectional gated recurrent unit network (BiGRU) to characterize long-term trends and temporal dependencies. This dual-branch structure effectively realizes the correspondence between model structure and signal characteristics, fully leveraging the complementary advantages of the two models at different time scales, and improving overall prediction accuracy.

[0046] 3. To ensure physical consistency between high- and low-frequency branches during the fusion phase, this invention proposes an adaptive grouping criterion based on variance entropy and an additive fusion mechanism. By dynamically dividing frequency bands using entropy features and reconstructing prediction results additively, cross-frequency feature alignment and energy conservation are achieved, enabling the prediction output to maintain physical interpretability while possessing good numerical stability. This method can be widely applied to energy consumption prediction, energy efficiency management, and intelligent scheduling scenarios for various types of buildings.

[0047] See Figure 1 This invention provides a building energy consumption prediction method based on multi-frequency eigenvalue decomposition and bi-branch modeling, comprising the following steps: A building energy consumption prediction method based on multi-frequency eigenvalue decomposition and bi-branch modeling, characterized by comprising: S1 performs multi-scale decomposition on the target building's energy consumption time series to obtain the intrinsic mode function and residual trend term; S2 divides the intrinsic mode functions into frequency bands based on the variance entropy index of each intrinsic mode function. The intrinsic mode functions above the threshold are divided into high-frequency component groups, and the intrinsic mode functions below or equal to the threshold and the residual trend term are divided into low-frequency component groups. The high-frequency signal and the low-frequency signal are then reconstructed respectively. S3 uses high-frequency signals as input to train the first branch prediction model and low-frequency signals as input to train the second branch prediction model. S4 performs an additive fusion operation on the prediction results of the first branch prediction model and the second branch prediction model to obtain the energy consumption prediction results of the target building. Building energy consumption prediction results are used to provide load prediction basis for the air conditioning control system of the target building, so as to achieve optimized control of the air conditioning system of the target building.

[0048] In a preferred embodiment provided by the present invention, the specific execution process of the above steps is as follows.

[0049] 1. Signal decomposition

[0050] This module is used to transform the original building energy consumption time series into a set of subsequences with independent spectral characteristics, solving the structural aliasing problem when directly modeling non-stationary signals.

[0051] This invention employs the Complete Ensemble Empirical Mode Decomposition Adaptive Noise Algorithm (CEEMDAN) to perform multi-scale decomposition on the input energy consumption sequence. In some feasible embodiments, CEEMDAN is first used to adaptively decompose the energy consumption time series, and then the variance entropy index is introduced, using the median as a dynamic threshold to automatically group the IMFs and reconstruct high-frequency and low-frequency signals.

[0052] The basic idea of ​​CEEMDAN is to introduce Gaussian white noise with different random seeds in each round of decomposition, and generate robust intrinsic mode functions (IMFs) through ensemble averaging, ensuring that the components are approximately orthogonal in the frequency domain. Let the original sequence be... This can be represented using CEEMDAN as:

[0053] in, Indicates the first One modal component, This represents the residual trend term.

[0054] In the algorithm implementation, the noise amplitude parameter ε is set to 0.2, the number of iterations M = 100, and the iteration termination condition is the monotonicity of the residual. Through this process, the complex energy consumption sequence is decomposed into... Each of the following components is independent and corresponds to the energy consumption oscillation characteristics at different time scales. Among them: (1) High-frequency IMFs reflect factors such as short-term fluctuations in building energy consumption, equipment start-up and shutdown, and changes in occupancy behavior; (2) Mid-frequency IMFs exhibit medium-cycle patterns such as weekdays and weekends; (3) Low-frequency IMFs and residual trend terms reflect long-term trends such as weather, seasons and building operating conditions.

[0055] Therefore, this decomposition process is essentially an adaptive reconstruction process based on the statistical fluctuation intensity. Based on the linear invertibility of CEEMDAN, high-frequency prediction and low-frequency prediction are additively reconstructed to ensure the consistency of the prediction results in energy distribution and time structure, and to provide a clear time-frequency structure for subsequent frequency band identification and differentiated modeling.

[0056] II. Feature Grouping

[0057] This module aims to solve the frequency classification problem of multimodal functions after decomposition, and to achieve automatic identification and differentiation of high-frequency and low-frequency features, thereby ensuring that different features enter the most suitable prediction model.

[0058] (1) Calculation of variance entropy: For each IMF component Calculate the standard deviation of its first-order difference sequence as the variance entropy index:

[0059] Variance entropy can effectively measure the fluctuation intensity of each component. High-frequency components have higher entropy values, while low-frequency components have smaller fluctuation intensity.

[0060] (2) Adaptive grouping strategy: With the median of all entropy values As a dynamic threshold, it automatically divides high-frequency groups into high-frequency and low-frequency groups:

[0061] To prevent all components from falling into the same group in extreme cases, this invention introduces a redundancy protection mechanism, which forces the highest entropy component to be assigned to the high-frequency group and the lowest entropy component to be assigned to the low-frequency group, in order to ensure the integrity of the decomposed signal.

[0062] (3) Signal reconstruction: Based on the grouping results, the high-frequency component quantities are re-superimposed into a high-frequency signal. The low-frequency components are superimposed to form a low-frequency signal. :

[0063] This grouping mechanism is entirely driven by signal characteristics, requiring no manual parameter tuning. It automatically adjusts the grouping threshold based on the fluctuation amplitude and spectral characteristics of different building energy consumption signals. Verification has shown that this method can effectively separate high- and low-frequency characteristics while preserving the signal energy structure.

[0064] III. Two-branch prediction network

[0065] To achieve targeted modeling of signals in different frequency bands, this invention designs a dual-branch prediction network structure. The two branches correspond to different time-scale features and are characterized by structural independence and complementary functions.

[0066] (1) High-frequency branch (LightGBM model)

[0067] The high-frequency component mainly includes short-term disturbances and highly volatile signals. To capture nonlinear relationships and local spike characteristics, this invention selects the Light Gradient Boosting Machine (LightGBM) as the modeling core.

[0068] LightGBM is an efficient ensemble learning algorithm based on gradient boosting trees (GBDT). It employs histogram-based binning, gradient-based one-side sampling (GOSS), and exclusive feature bundling (EFB) mechanisms, which significantly improves computational efficiency and generalization performance.

[0069] The model training input is a sliding window feature matrix of high-frequency signals:

[0070] It works by using a preset sliding window length. L Construct, and output the predicted value for the next time step. In the embodiments provided by the present invention, it is referred to as the first predicted value.

[0071] Model parameters (learning rate η, number of leaf nodes, maximum tree depth, regularization terms λ1 and λ2, feature sampling rate, etc.) are automatically searched globally using the Optuna (Bayesian Optimization Framework) hyperparameter optimization framework. The evaluation function is the root mean square error (RMSE) of the validation set, and an early stopping strategy is employed to avoid overfitting.

[0072] The high-frequency branch model focuses on learning local nonlinear relationships and abrupt change modes, and has high sensitivity to short-term load fluctuations, abnormal peaks and equipment operation disturbances.

[0073] (2) Low-frequency branch (BiGRU model)

[0074] Low-frequency signals reflect the long-term patterns and trends of building energy consumption, making them suitable for time-dependent modeling using recurrent neural network (RNN) models. This invention employs a bidirectional gated recurrent unit (BiGRU) network, whose structure consists of two layers of bidirectional GRU units, containing 128 and 64 hidden units respectively, enabling it to capture time-dependent features simultaneously from both forward and backward directions.

[0075] At each time step t, the update equation for BiGRU is:

[0076] It updates the temporal features of the low-frequency signal using gated recurrent units to obtain the hidden state sequence representation for training the second branch model. To update the door, These are candidate activation states. This equation describes the recursive process of the hidden state of a BiGRU under low-frequency input.

[0077] The forward and backward hidden states are concatenated to obtain the global temporal feature representation: This represents the joint feature representation of the forward and backward hidden states of the low-frequency signal formation, which serve as intermediate features for subsequent predictions. The network employs batch normalization and dropout (p = 0.2) to prevent overfitting. The optimizer is an Adaptive Moment Estimation (Adam), and the loss function is Mean Squared Error (MSE). The learning rate uses a dynamic adjustment mechanism, ReduceLROnPlateau, which automatically decays the learning rate if the validation error does not decrease for several consecutive rounds.

[0078] Then, based on the aforementioned bidirectional hidden state, the low-frequency prediction result is obtained through output mapping. This is used to characterize the predicted value of the low-frequency energy consumption component at the next time step, and is fused with the high-frequency prediction result in subsequent steps. In the embodiments provided by this invention, it is referred to as the second predicted value.

[0079] The low-frequency branch focuses on extracting the smooth trend of energy consumption changes and seasonal cycle information, and has good long-term forecasting ability and stability.

[0080] Then, the first and second predicted values ​​are uniformly labeled to obtain... and It is used to characterize the prediction results obtained by the first branch model and the second branch model within the current prediction period, corresponding to the two prediction values, and is used for subsequent additive fusion.

[0081] (3) Two-branch collaborative mechanism

[0082] This dual-branch collaborative mechanism is built upon the aforementioned consistency constraints of signal decomposition and reconstruction. The high-frequency branch and the low-frequency branch model the high-frequency signal and the low-frequency signal, respectively, obtained by CEEMDAN decomposition and reconstruction based on the variance entropy criterion. The two types of signals satisfy a linear additive relationship in the time domain. At the model structure level, the high-frequency branch adopts a LightGBM model based on gradient boosting trees, focusing on characterizing the local nonlinear relationships and sudden changes in high-frequency disturbances; the low-frequency branch adopts a bidirectional gated recurrent unit network (BiGRU), focusing on learning the long-term trends and time-dependent structures contained in the low-frequency signal. The two types of models have significant differences in inductive bias and modeling focus, and can respectively adapt to the statistical properties of different frequency characteristics.

[0083] After the two branches are trained independently, inference can be performed in parallel, and high-frequency prediction results can be output separately. Compared with low-frequency prediction results .

[0084] In this invention, the high-frequency and low-frequency branches remain structurally independent during both the training and inference phases, without cross-branch feature transfer, pseudo-label supervision, or inference path selection. The two branches collaborate only at the prediction result level, achieving overall reconstruction of the original building energy consumption signal through additive fusion of the high-frequency and low-frequency prediction results. This collaborative mechanism ensures the physical consistency of the prediction results in energy distribution and temporal structure, avoiding interference between information from different frequency bands. Furthermore, it leverages the complementarity of LightGBM and BiGRU in terms of bias and variance characteristics, statistically reducing the overall prediction error and improving the model's prediction accuracy and stability in complex building energy consumption scenarios. Moreover, since LightGBM prioritizes accuracy in local modeling while BiGRU prioritizes smoothness in global trends, their complementary bias and variance levels statistically reduce the overall prediction error.

[0085] IV. Predictive Fusion

[0086] To obtain the final prediction result and maintain the physical consistency of the signal, this invention adopts an additive reconstruction prediction fusion strategy.

[0087]

[0088] This method follows the linear reversibility principle of CEEMDAN, ensuring that the predicted signal remains consistent with the original energy consumption signal in terms of energy conservation, time structure, and spectral characteristics.

[0089] Unlike weighted averaging or multilayer sensing fusion, additive reconstruction does not introduce additional learnable parameters, thus avoiding the distortion risks associated with nonlinear superposition. Furthermore, the fused results offer greater physical interpretability: the high-frequency component contributes to sudden loads and short-term responses, while the low-frequency component reflects long-term energy consumption trends and meteorological driving factors.

[0090] Furthermore, this invention also achieves model diagnosis through error decomposition: if high-frequency errors dominate, the LightGBM parameters need to be adjusted; if low-frequency deviations are significant, the BiGRU structure needs to be optimized. This diagnostic capability enables the model to have self-feedback optimization ability.

[0091] The present invention also provides an embodiment, such as Figure 2 The diagram illustrates a preferred execution process of the present invention.

[0092] Step 1: Preparing the input data for training the model. This invention selects a building energy consumption dataset as the experimental object, which includes energy consumption records of buildings over continuous time periods. Through data cleaning and standardization, outliers are removed and missing values ​​are filled in to maintain the continuity of the time series. The data is divided into training, validation, and test sets in chronological order to ensure the model's generalization ability and predictive stability, providing high-quality input data for subsequent modeling.

[0093] Step 2: Prepare the algorithm modules required for signal decomposition and model training. This invention employs the Complete Ensemble Empirical Mode Decomposition Adaptive Noise Algorithm (CEEMDAN) to perform multi-scale decomposition of the energy consumption sequence to separate feature components at different frequencies. Simultaneously, a dual-branch prediction structure is established, where the high-frequency branch uses the LightGBM model, and the low-frequency branch uses a BiGRU network. LightGBM models short-term nonlinear perturbations through a gradient boosting tree mechanism, while BiGRU extracts long-term trend features through a bidirectional recurrent structure. Key parameters for each model are set before training, including the learning rate, window length, number of hidden units, and regularization coefficient, to ensure the stability and reproducibility of the training process.

[0094] Step 3: Perform CEEMDAN decomposition on the input building energy consumption time series. For the input standardized energy consumption series... The energy consumption signal is decomposed layer by layer using the CEEMDAN algorithm, yielding several intrinsic mode functions (IMFs) and a residual trend term. Each IMF represents the oscillation mode of the energy consumption signal at a specific time scale; the high-frequency component reflects short-term fluctuations, while the low-frequency component describes the long-term trend. The decomposition result can be expressed as follows: Each component is independent and their energies do not overlap. Through this step, the complex non-stationary signal is converted into a set of sub-signals that can be modeled separately, laying the foundation for subsequent frequency band division.

[0095] Step 4: Calculate variance entropy and group high- and low-frequency features. Calculate the variance entropy index for each IMF component obtained from the decomposition, which is used to characterize the signal fluctuation intensity. Based on a threshold set according to the median of the variance entropy, the IMFs are automatically divided into high-frequency and low-frequency groups. The high-frequency group mainly contains short-term random disturbance signals, while the low-frequency group mainly contains periodic variations and trend information. The high-frequency signals are then reconstructed through signal superposition. and low frequency signals This process requires no manual intervention and can adaptively group data based on the actual fluctuation characteristics of energy consumption data, providing input data for the model's two-branch modeling.

[0096] Step 5: Train the high-frequency branch model. Input the high-frequency signal and its external features into the LightGBM model for training. The model uses a gradient boosting mechanism to fit short-term changes in energy consumption, and automatically adjusts parameters such as the learning rate and tree depth through Bayesian optimization to improve fitting accuracy and generalization ability. The high-frequency branch can capture rapid fluctuations, abnormal peaks, and nonlinear disturbances in the building energy consumption sequence, providing local correction capabilities for overall prediction.

[0097] Step 6: Train the low-frequency branch model. The low-frequency signal is input into a Bidirectional Gated Recurrent Unit (BiGRU) network for training. BiGRU learns the long-term trend and periodic variation of building energy consumption by modeling the time dependencies of both forward and backward passes. The model uses the Adam optimizer and mean squared error loss function for iterative updates, combined with Batch Normalization and Dropout mechanisms to prevent overfitting, thus achieving smooth and stable trend prediction capabilities.

[0098] Step 7: Fusion of high-frequency and low-frequency prediction results. In the prediction phase, the trained LightGBM and BiGRU models are used to predict the test set data respectively, obtaining the high-frequency prediction results. and low-frequency prediction results Based on the principle of linear invertibility of CEEMDAN decomposition, the final predicted output is obtained through additive fusion. This fusion mechanism ensures the energy conservation and physical consistency of the prediction results, and achieves the collaborative expression of multi-scale features.

[0099] Step 8: Evaluate Model Performance. The predicted values ​​are compared with the actual energy consumption values, and the model is evaluated using metrics such as Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE). The results show that the model of this invention exhibits high prediction accuracy and stability in various building scenarios, especially outperforming traditional models when dealing with non-stationary energy consumption sequences and sudden fluctuations.

[0100] Step 9: Deploy the model and perform continuous optimization. Deploy the trained model into the building energy management system and run it automatically according to the process of "data acquisition → preprocessing → CEEMDAN decomposition → feature grouping → bi-branch prediction → additive fusion → output prediction results". The system can periodically retrain the model or update the parameters to adapt to seasonal changes and changes in building operating conditions, achieving real-time, high-precision prediction of building energy consumption.

[0101] Step 10: End the entire process. Through the above steps, this invention completes the entire process design from data input, signal decomposition, feature grouping, branch modeling to result fusion, and constructs a building energy consumption prediction method with high accuracy and strong robustness. This method can effectively solve the problems of non-stationarity and multi-scale characteristics of energy consumption sequences, providing reliable technical support for energy optimization and scheduling in intelligent buildings.

[0102] In constructing this model, the Complete Ensemble Empirical Mode Decomposition Adaptive Noise Algorithm (CEEMDAN) is introduced to decompose the building energy consumption signal at multiple scales, significantly improving the decomposability and signal stability of the original energy consumption sequence. The CEEMDAN decomposition mechanism ensures that different frequency components are effectively distinguished before modeling, avoiding mode aliasing and feature distortion problems caused by directly modeling non-stationary signals. Furthermore, through an adaptive grouping strategy based on variance entropy, the decomposed signal is automatically classified and reconstructed, enabling targeted modeling of high-frequency disturbances and low-frequency trends, further enhancing the model's feature representation ability and generalization performance.

[0103] The dual-branch collaborative modeling method, composed of LightGBM and BiGRU, significantly reduces the fitting error of a single prediction model across different frequency features. Furthermore, it incorporates Bayesian optimization to achieve adaptive hyperparameter search, improving training efficiency and global optimality of the model under complex building energy consumption data. The dual-branch structure captures sudden load changes in the high-frequency channel and extracts long-term trend features in the low-frequency channel, enhancing the system's robustness and predictive stability across multiple time scales.

[0104] A more accurate multi-scale modeling strategy enables building energy consumption prediction models to achieve finer fitting based on actual energy consumption fluctuations, reducing prediction bias caused by the coupling of short-term fluctuations and long-term trends. This invention has achieved significant beneficial effects in improving the accuracy and stability of building energy consumption prediction models, as demonstrated by experimental results. Figures 3 to 5 As shown. The method of this invention, through the synergistic effect of high and low frequency branches and the additive fusion mechanism, enables the model to better reproduce the trend of energy consumption curve changes, while effectively reducing the overall prediction error and improving the load prediction accuracy and energy utilization efficiency of the building energy management system.

[0105] Based on recent research progress in the field of building energy consumption prediction, the construction of building energy consumption prediction models typically includes three key steps: (1) Reduce noise interference and data instability of energy consumption signals through preprocessing and feature extraction; (2) Capture the dynamic evolution of energy consumption data through time series models or deep networks; (3) By integrating prediction results from different sources through model integration or post-processing mechanisms, the overall prediction accuracy and stability can be improved.

[0106] However, traditional methods still have two core bottlenecks: First, energy consumption signals have significant multi-level characteristics on a time scale, meaning that high-frequency disturbances and low-frequency trends coexist, and a single model can hardly take both modes into account in the same structure; Second, although existing hybrid models introduce signal decomposition techniques, they still use a unified predictor in the post-decomposition stage, failing to perform structure matching according to different frequency characteristics, which limits the prediction accuracy.

[0107] To address the aforementioned problems, this invention proposes a building energy consumption prediction system based on multi-frequency feature decomposition and bi-branch modeling. This method comprehensively utilizes signal decomposition, information entropy measurement, ensemble learning, and recurrent neural network techniques to achieve structured representation of energy consumption data across multiple time scales, frequency adaptive modeling, and energy consistency fusion. Figure 6 As shown, the overall scheme includes four main parts: signal decomposition module 601, feature grouping module 602, model training module 603, and prediction fusion module 604.

[0108] The signal decomposition module 601 is used to: perform multi-scale decomposition operations on the energy consumption time series of the target building in real time to obtain the intrinsic mode function and residual trend term; The feature grouping module 602 is used to: obtain the variance entropy index of the intrinsic mode function in real time through calculation, and reconstruct the target building energy consumption time series into high-frequency signals and low-frequency signals based on the variance entropy index; The model training module 603 has a dual-branch prediction network, specifically used to: train the first branch model in real time using a high-frequency signal group and train the second branch model using a low-frequency signal group; Predictive fusion module 604 is used for: The first branch model and the second branch model are fused in real time to obtain a primary prediction model; The primary prediction model is iteratively optimized using the outputs of the first and second branch models in real time to obtain the building energy consumption prediction model. Deploy building energy consumption prediction models into the air conditioning control system of the target building.

[0109] In summary, this invention provides a building energy consumption prediction method and system based on multi-frequency eigenvalue decomposition and bi-branch modeling. By introducing the Complete Ensemble Empirical Mode Decomposition Adaptive Noise Algorithm (CEEMDAN) and variance entropy grouping strategy, it achieves multi-scale decomposition of building energy consumption signals and adaptive partitioning of high- and low-frequency features. Based on the CEEMDAN decomposition mechanism, by introducing adaptive noise into the signal and performing multiple ensemble averaging, the high-frequency disturbances and low-frequency trends in the energy consumption sequence are effectively separated. For complex non-stationary building energy consumption signals, the aliased time-frequency features are reorganized into structurally clear and spectrally independent intrinsic mode functions, enhancing the interpretability and modeling stability of the signal features.

[0110] This invention not only reduces the uncertainty caused by manual division thresholds in traditional decomposition models through an adaptive grouping method based on variance entropy, but also uses signal energy distribution characteristics as the basis for grouping, eliminating frequency errors caused by manually setting grouping boundaries, ensuring the physical integrity and reconstruction consistency of the decomposed signal, and enabling the model to adopt matching prediction structures for different frequency features, thereby achieving differentiated modeling of features at multiple time scales.

[0111] Based on the differences in high- and low-frequency characteristics, this invention constructs a dual-branch prediction structure composed of LightGBM and BiGRU. LightGBM models the rapid fluctuations and nonlinear changes of high-frequency components, while BiGRU captures the long-term trends and periodic patterns of low-frequency components, achieving collaborative prediction across multiple time scales. This dual-branch structure not only reduces interference between different frequency components, improving local accuracy and overall smoothness of prediction, but also integrates the complementary features of the two models, ensuring high accuracy and robustness in building energy consumption prediction.

[0112] This invention designs a parameter adaptive adjustment mechanism based on Bayesian optimization, which allows the model to automatically search for the optimal hyperparameter combination under different building types and data features, improving training efficiency and avoiding local optima problems. Through an additive fusion mechanism, energy-conserving reconstruction is performed on the high- and low-frequency prediction results, ensuring that the final predicted value remains consistent with the original energy consumption signal in both the time and frequency domains, thus achieving physical interpretability and controllability of the model output.

[0113] The overall design enables the prediction model to accurately model and adaptively learn building energy consumption signals, reducing the fitting bias and lag effect of traditional models under non-stationary data. Through the synergistic effect of core technologies such as decomposition, grouping, bi-branch modeling, and energy consistency fusion, the system can maintain the stability and accuracy of predictions at different time scales, significantly improving the accuracy of building energy consumption prediction and energy utilization efficiency.

[0114] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention.

[0115] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of the present invention.

[0116] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for apparatus or system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The apparatus and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0117] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A building energy consumption prediction method based on multi-frequency eigenvalue decomposition and bi-branch modeling, characterized in that, include: S1 performs multi-scale decomposition on the target building's energy consumption time series to obtain the intrinsic mode function and residual trend term; The multi-scale decomposition operation is performed using a fully ensemble empirical mode decomposition adaptive noise algorithm, specifically including: Through The energy consumption time series of the target building is decomposed into multiple non-interfering components, each corresponding to the energy consumption oscillation characteristics at different time scales; where, Indicates the first One modal component, This is the residual trend term; S2 divides the intrinsic mode functions (IMFs) into frequency bands based on the variance entropy index of each IMF. IMFs above a threshold are divided into high-frequency component groups, and IMFs below or equal to the threshold, along with the residual trend term, are divided into low-frequency component groups. High-frequency and low-frequency signals are then reconstructed respectively. This includes: Through The variance entropy index of the intrinsic mode functions is calculated; With the median of all entropy values As a dynamic threshold, through the formula The energy consumption time series of the target building is reconstructed into a high-frequency component group and a low-frequency component group; among them, the maximum entropy component is forcibly assigned to the high-frequency component group, and the minimum entropy component is forcibly assigned to the low-frequency component group. Through The high-frequency component group and the low-frequency component group are superimposed to form a high-frequency signal and a low-frequency signal, respectively. S3 uses the high-frequency signal as input to train the first branch prediction model, and uses the low-frequency signal as input to train the second branch prediction model; including: Through Based on the preset sliding window length L The first input signal for constructing the high-frequency signal; The first input signal is fed into the first branch model for training. By capturing nonlinear relationships and local spike features, the first predicted value at the corresponding time step is obtained. ; Through The time-series features of the low-frequency signal are updated using gated recurrent units to obtain the hidden state sequence representation for training the second branch model; where, To update the door, In candidate activation state; The hidden state sequence representation corresponding to the low-frequency signal is input into the second branch model for training. By simultaneously capturing time-dependent features from both the forward and backward directions, intermediate features are obtained. ; Based on intermediate features The mapping operation is performed to obtain the second predicted value. ; For the first predicted value and the second predicted value Perform uniform labeling processing on each, so that the first predicted value and the second predicted value At the predicted time The above corresponds one-to-one and is used in step S4; S4 performs an additive fusion operation on the prediction results of the first branch prediction model and the second branch prediction model to obtain the energy consumption prediction results of the target building. Building energy consumption prediction results are used to provide load prediction basis for the air conditioning control system of the target building, so as to achieve optimized control of the air conditioning system of the target building.

2. The method according to claim 1, characterized in that, Step S4 includes: Through The first and second predicted values ​​with unified labels output by the first and second branch models are fused together.

3. A building energy consumption prediction system based on multi-frequency eigenvalue decomposition and bi-branch modeling, used to perform the method as described in claim 1 or 2, characterized in that, include: The signal decomposition module is used to: perform multi-scale decomposition operations on the energy consumption time series of the target building in real time to obtain the intrinsic mode function and residual trend term; The feature grouping module is used to: divide the intrinsic mode functions into frequency bands based on the variance entropy index of each intrinsic mode function, divide the intrinsic mode functions above the threshold into high frequency component groups, divide the intrinsic mode functions below or equal to the threshold and the residual trend term into low frequency component groups, and reconstruct the high frequency signal and low frequency signal respectively. The model training module is used to: train a first branch prediction model with the high-frequency signal as input, and train a second branch prediction model with the low-frequency signal as input; The predictive fusion module is used for: An additive fusion operation is performed on the prediction results of the first branch prediction model and the second branch prediction model to obtain the energy consumption prediction results of the target building. The building energy consumption forecast results are input into the air conditioning control system of the target building to provide load forecast basis for the optimized control of the air conditioning of the target building.

Images