A building energy consumption prediction method based on adaptive decomposition and intelligent reconstruction

Abstract

The application provides a building energy consumption prediction method based on adaptive decomposition and intelligent reconstruction, and relates to the technical fields of building energy management and artificial intelligence, which comprises the following steps: firstly, through adaptive decomposition parameter optimization based on energy entropy, the optimal modal number and penalty factor of VMD or CEEMDAN are automatically determined, and the original energy consumption sequence is decomposed into several intrinsic modal components; then, according to the sample entropy and average period of each component, the components are divided into high-frequency noise, detail components and trend components, and a differentiated PSO-XGBoost strategy is used for modeling; then, the component prediction results are weighted and fused through a dynamic weighted intelligent reconstruction mechanism based on error feedback; finally, through an incremental learning and federal migration collaborative framework, the online updating of the model and the multi-building knowledge migration are realized. The application realizes the automatic determination of the decomposition parameters and the differentiated modeling of the components, and significantly improves the prediction accuracy and robustness.

Description

Technical Field

[0001] This invention relates to the fields of building energy management and artificial intelligence technology, specifically to a building energy consumption prediction method based on adaptive decomposition and intelligent reconstruction. Background Technology

[0002] Building energy consumption accounts for a significant proportion of global energy consumption, and accurate energy consumption forecasting is crucial for energy conservation, cost control, and grid demand response. However, building energy consumption data is influenced by a combination of factors, including seasonal changes, weather conditions, human activity, and equipment operation, exhibiting strong nonlinearity, non-stationarity, and multi-timescale characteristics. Traditional single-model forecasting approaches struggle to effectively capture these complex dynamic patterns. To address this challenge, the "decomposition followed by integration" forecasting approach has emerged and is gradually becoming the mainstream technology in this field.

[0003] Currently, commonly used signal decomposition methods mainly include Empirical Mode Decomposition (EMD), wavelet decomposition, Variational Mode Decomposition (VMD), and Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN). Existing technical solutions, represented by "short-term load forecasting methods based on EMD and support vector machines," typically involve: first, using EMD to decompose the original load sequence into several intrinsic mode functions (IMFs); then, establishing a prediction model (such as a support vector machine) for each component; and finally, linearly superimposing the prediction results of all components to obtain the final predicted value.

[0004] However, existing technologies still suffer from the following core shortcomings in practical applications. First, decomposition algorithms such as VMD or CEEMDAN are extremely sensitive to parameters like the number of modes K and the penalty factor α. Existing methods often rely on manual pre-setting, leading to significant fluctuations in decomposition quality and difficulty in matching the inherent characteristics of the data itself. Second, the modeling strategy is too simplistic, applying the same treatment to all IMF components, ignoring the inherent physical differences between high-frequency and low-frequency components, easily causing overfitting of high-frequency components and underfitting of low-frequency components. Third, in the reconstruction stage, existing methods typically use simple linear superposition, assuming that the prediction results of each component have the same confidence level, failing to dynamically adjust the fusion weights, severely weakening the model's robustness. Fourth, commonly used particle swarm optimization algorithms are prone to premature convergence in high-dimensional parameter spaces, resulting in insufficient parameter optimization and affecting the overall model performance. Fifth, in terms of feature construction, existing methods are mostly limited to time-domain features, failing to fully utilize the frequency-domain information of each decomposed component, resulting in low information utilization. Sixth, once deployed, existing models struggle to adapt to new buildings or working conditions, lacking cross-scenario knowledge transfer capabilities, requiring retraining before use. Seventh, due to the combined effect of the above-mentioned defects, existing methods are difficult to simultaneously meet the dual requirements of prediction accuracy and generalization ability in practical applications.

[0005] In view of the above, this application is hereby submitted. Summary of the Invention

[0006] This invention provides a building energy consumption prediction method based on adaptive decomposition and intelligent reconstruction, which can at least partially improve the above-mentioned problems.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A building energy consumption prediction method based on adaptive decomposition and intelligent reconstruction, comprising: Acquire historical energy consumption data of the target building, along with corresponding meteorological data and timestamp information, and perform data preprocessing to obtain the original energy consumption sequence; Based on the original energy consumption sequence, adaptive decomposition parameter optimization based on energy entropy is performed to determine the optimal number of modes and the optimal penalty factor. Based on the optimal number of modes and the optimal penalty factor, variational mode decomposition is performed on the original energy-consuming sequence to obtain multiple intrinsic mode function components. Each intrinsic mode function component is then classified, and a differentiated PSO (Particle Swarm Optimization) parameter search space corresponding to each category is designed. A particle population is established based on the differential PSO parameter search space and initialized. Adaptive mode switching is performed based on the initialized particle population to update the particle positions. This process is repeated multiple times to obtain the final global optimal solution and the trained XGBoost (Extreme Gradient Boosting) model of the intrinsic mode function components of each category. The XGBoost model, trained using the intrinsic mode function components of each category, is used to predict the validation set data, resulting in a sequence of predicted values. Dynamic weights are constructed using an exponential function, and the predicted value sequences are weighted and summed according to the dynamic weights to obtain the final reconstructed prediction result.

[0009] In summary, this invention provides a building energy consumption prediction method based on adaptive decomposition and intelligent reconstruction. First, it automatically determines the optimal decomposition parameters for VMD / CEEMDAN using an "adaptive decomposition parameter optimization based on energy entropy" technique, decomposing the original energy consumption data into a series of IMFs. Second, based on the sample entropy and average period of each IMF, it classifies them into high-frequency noise, detail components, and trend components, and employs differentiated PSO-XGBoost modeling strategies for different categories. The PSO optimization process is performed by an improved version of "HPSOWOA (Hybrid Particle Swarm Optimization and Whale Optimization Algorithm)" to improve parameter search efficiency. Then, the prediction error of each component model is calculated on the validation set, and weights are dynamically calculated using a "dynamic weighted intelligent reconstruction mechanism based on error feedback" to weightedly fuse the component prediction results, obtaining the final prediction value. Furthermore, to adapt to practical applications, this invention also introduces an "incremental learning and federated transfer collaborative framework" to support online model updates and knowledge transfer among multiple buildings.

[0010] Compared with existing technologies, this method has the following advantages: First, it significantly improves prediction accuracy. By using adaptive decomposition parameter optimization based on energy entropy and the coupling theorem of modal aliasing and prediction accuracy, the mathematical correlation between decomposition quality and prediction task is ensured, resulting in reduced prediction error compared to fixed-parameter decomposition methods. Second, it enhances model robustness. The dynamic weighted intelligent reconstruction mechanism based on error feedback can dynamically adjust the contribution of each component model according to its real-time performance, avoiding the impact of a single model failure on the overall model and overcoming the shortcomings of traditional linear superposition methods. Third, it improves optimization efficiency and effectiveness. The HPSOWOA optimization algorithm, through an adaptive switching mechanism, combines the global exploration capability of particle swarm optimization and the local development capability of whale optimization, improving optimization speed by 30% and making it easier to find the global optimum. Fourth, it achieves high adaptability. From determining decomposition parameters to model selection and reconstruction, the entire process is automated, reducing the workload and subjectivity of manual parameter tuning and facilitating deployment to different buildings. Fifth, it expands application scenarios. The incremental learning and federated transfer collaborative framework enables the model to adapt to changes in data distribution and achieve knowledge sharing among multiple buildings while protecting privacy. It solves the problems of cold start and insufficient generalization ability and has high engineering practical value. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating the building energy consumption prediction method based on adaptive decomposition and intelligent reconstruction provided in an embodiment of the present invention. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0013] refer to Figure 1 As shown, the first embodiment of the present invention discloses a building energy consumption prediction method based on adaptive decomposition and intelligent reconstruction, which can be executed by a building energy consumption prediction device based on adaptive decomposition and intelligent reconstruction (hereinafter referred to as the prediction device), specifically, by one or more processors within the prediction device, to implement the following method: It should be noted that the data involved in this invention comes from the ASHRAE Great Energy Predictor III competition dataset, which is one of the most authoritative benchmark datasets in the field of building energy consumption prediction. The hardware environment used is a server configured with an AMD Ryzen 5 7500F processor, an RTX 4070 Super graphics card, and 32GB of RAM. The software environment is Python 3.13, with integrated open-source libraries such as Scikit-learn, XGBoost, and Pandas.

[0014] S1: Obtain historical energy consumption data and corresponding meteorological data (such as temperature and humidity) and timestamp information of the target building, and perform data preprocessing to obtain the original energy consumption sequence; Specifically, step S1 further includes: the data preprocessing includes cleaning, handling of missing and outlier values, and normalization.

[0015] S2, based on the original energy consumption sequence, perform adaptive decomposition parameter optimization based on energy entropy to determine the optimal number of modes and the optimal penalty factor respectively; Specifically, step S2 further includes: targeting the original energy consumption sequence The initial value of the loop attempt number m is set to 2, and the mode number K = m. The loop attempt number m is gradually increased with a step size of 1. For each value of the loop attempt number m, the mode number K = m is set, and variational mode decomposition is performed once to obtain a eigenmode function components. , For the first intrinsic mode function component, Let a be the a-th intrinsic mode function component, where a is the number of intrinsic mode function components, and a = K; The formula for calculating the energy density and relative energy percentage of each eigenmode function component is as follows: , , ,in, Let i be the energy density of the i-th eigenmode function component. Let j be the energy density of the j-th eigenmode function component. For the integral sign, This indicates differentiating with respect to time t. Let i be the value of the i-th intrinsic mode function component at time t. The relative energy percentage of the i-th eigenmode function component; Calculate the total energy entropy of the variational mode decomposition corresponding to the loop attempt number m. And calculate the rate of change of energy entropy under adjacent mode numbers. , The total energy entropy for the variational mode decomposition corresponding to index m-1 in the cyclic attempts. It is a logarithmic function; where energy entropy measures the degree of dispersion of information across the frequency band.

[0016] When the first judgment is made When the information gain from further increasing the number of modes is negligible, determine the optimal number of modes. for: , For the preset threshold, .

[0017] Based on the modal aliasing-prediction accuracy coupling theorem, the modal aliasing index is minimized. The modal aliasing index is defined. The sum of the absolute values ​​of the cross-correlation coefficients among the variational modal components is given by the formula: , The Pearson correlation coefficient is used. For the i-th intrinsic mode function component (i.e., given...) The i-th eigenmode function component obtained by decomposition. For the j-th intrinsic mode function component (i.e., given...) The j-th eigenmode function component obtained from the decomposition); where... The smaller the value, the lower the linear correlation between modes, the less aliasing, and the higher the decomposition quality.

[0018] Within the predefined search range Within, the modal aliasing index will be minimized. As an optimization objective, use grid search or the HPSOWOA algorithm to find the optimal penalty factor. , As a penalty factor, This is the lower limit of the search range. This represents the upper limit of the search range. Experiments show that minimizing... It can effectively reduce the final prediction error RMSE.

[0019] In this embodiment, this step aims to automatically determine the optimal number of modes and penalty factor for VMD (or CEEMDAN). Specifically, the original energy consumption sequence undergoes adaptive decomposition parameter optimization based on energy entropy. The initial value of the loop attempt number m is set to 2, the number of modes K is set to m, and m is gradually increased by a step size of 1. For each value of m, variational mode decomposition is performed once, obtaining an eigenmode function component equal to the number of modes. The energy density and relative energy proportion of each component are calculated, and then the total energy entropy of the current decomposition and the rate of change of energy entropy under adjacent modes are calculated. When the rate of change of energy entropy is first less than a preset threshold, the optimal number of modes is determined to be the current number of modes minus 1. This adaptive optimization method eliminates the subjectivity of manually preset mode numbers, making the decomposition quality match the inherent characteristics of the data. Based on this, the mode aliasing index is defined as the sum of the absolute values ​​of the cross-correlation coefficients between each component. Within a predefined search range, the optimal penalty factor is found by minimizing this index. By minimizing the mode aliasing, the information overlap between components is effectively reduced, laying the foundation for improving prediction accuracy.

[0020] It should be noted that, to overcome the subjectivity of a single threshold, this invention introduces an automatic threshold determination method based on inflection point detection. The energy entropy change rate sequence is calculated. The Kneedle algorithm is used to detect the point of maximum curvature change in the sequence, and the rate of change corresponding to this point is used as a threshold. Theoretical benchmark value Experiments show that, On the ASHRAE dataset, the value is stable between 0.047 and 0.053, therefore we take... As a unified threshold.

[0021] Specifically, the Kneedle algorithm finds the inflection point (the point of maximum curvature change) of sequence S. The x-coordinate corresponding to this inflection point is... Then the theoretical benchmark threshold Here is the ordinate value corresponding to this point: , Threshold sensitivity analysis: To verify the threshold Its robustness makes Within the range [0.01, 0.10], the optimal K value and the final predicted RMSE are recorded by varying the K value in steps of 0.01. Analysis shows that when... At that time, due to an overly strict threshold, the value of K is overestimated (K=7), introducing redundant modes. When the optimal K value is stable at 5, the RMSE fluctuation is less than ±1.2%. At that time, the threshold was too wide, which led to the underestimation of the K value (K=3), resulting in information loss. Located at the center of the performance stability platform region, it is proven that the threshold selection has theoretical robustness.

[0022] Furthermore, in this embodiment, the penalty factor directly controls the bandwidth and orthogonality of each intrinsic mode function (IMF). Theorem 1 (Modal Aliasing-Prediction Accuracy Coupling): For the VMD decomposition-ensemble prediction framework, the final prediction error RMSE is coupled with the modal aliasing degree. satisfy: Where β>0 is the coupling coefficient, C is a constant related to the inherent complexity of the data, and ξ is the random error term. As the VMD penalty factor, compared with XGBoost Irrelevant.

[0023] Experiments verify that the theorem has a goodness-of-fit R² of 0.94 on the ASHRAE dataset, and the coupling coefficient is [missing value]. The constant term C = 0.21.

[0024] S3. Based on the optimal number of modes and the optimal penalty factor, variational mode decomposition is performed on the original energy-consuming sequence to obtain multiple intrinsic mode function components. Each intrinsic mode function component is classified, and a differentiated PSO parameter search space corresponding to each category is designed. Specifically, step S3 further includes: based on the optimal number of modes and optimal penalty factor For the original energy consumption sequence Variational mode decomposition is performed to obtain Each intrinsic mode function component ; The sample entropy and average period (i.e., the principal period of the component calculated based on spectral analysis) of each intrinsic mode function component are calculated using the following formula: , , , , Let be the sample entropy of the i-th intrinsic mode function component. It is the natural logarithm. Given a tolerance r, the length is The number of self-similar templates matched by the vector. Given a tolerance r, the length is The number of self-similar templates matched by the vector. Let be the average period of the i-th intrinsic mode function component. Let be the dominant frequency of the i-th eigenmode function component. For Fast Fourier Transform; Based on sample entropy and average period, The intrinsic mode function components are divided into three categories, and the classification formula is as follows: ,in, For high threshold of sample entropy, For low threshold of sample entropy, For long-period thresholds, This is the short-period threshold; the threshold can be set according to the actual data characteristics, and the example value is: , and Determined based on the sample entropy distribution.

[0025] For different categories of intrinsic mode function components, a corresponding differentiated PSO parameter search space is designed for their XGBoost prediction models to match the physical nature and complexity of each component. For each intrinsic mode function component, the PSO algorithm is used to minimize the prediction error within its corresponding differential PSO parameter search space. The objective function is: , Let i be the objective function corresponding to the i-th intrinsic mode function component. For the root mean square error on the validation set, For XGBoost prediction models, The training set portion for the i-th intrinsic mode function component, after decomposition, yields a one-dimensional time series. Divided chronologically, the earlier portion is used to train the model. For the hyperparameter combination to be optimized, Let be the search space for the differentiated PSO parameters corresponding to the i-th intrinsic mode function component. Wherein, The learning rate of the XGBoost model Maximum depth of the tree Subsampling ratio wait.

[0026] The search space for differentiated PSO parameters for high-frequency noise components (prone to overfitting, requiring strong regularization) is: , , , , The search space for differentiated PSO parameters for detailed components (moderate complexity, balancing fit and generalization) is: , , , , The search space for differentiated PSO parameters for the trend component (which has higher complexity and requires thorough learning of long-term dependencies) is as follows: , , , , ,in, The maximum depth of the tree. For learning rate, The subsampling ratio, The feature sampling ratio, The L1 regularization parameter of XGBoost is related to the VMD penalty factor. Irrelevant; These are the L2 regularization parameters for XGBoost. For the number of trees, here This is the L1 regularization parameter for XGBoost.

[0027] In this embodiment, based on the determined optimal number of modes and optimal penalty factor, variational mode decomposition is performed on the original energy consumption sequence to obtain multiple intrinsic mode function (IMF) components. These components separate fluctuation patterns at different time scales from the original complex energy consumption signal, providing a foundation for subsequent refined modeling. Subsequently, the sample entropy and average period of each IMF component are calculated. Sample entropy is used to measure the complexity of each component; a higher entropy value indicates a more random sequence and weaker regularity. The average period is obtained based on spectral analysis to obtain the dominant frequency of the component, reflecting its fluctuation time scale. Based on these two indicators, all components are automatically divided into three categories: high-frequency noise components, detail components, and trend components. Among them, high-frequency noise components have high sample entropy and short period characteristics, indicating weak regularity and strong randomness; trend components have low sample entropy and long period characteristics, reflecting the long-term variation law of energy consumption; detail components are in between. This classification mechanism enables subsequent modeling to differentiate based on the physical nature of different components, effectively avoiding the high-frequency overfitting and low-frequency underfitting problems caused by the traditional method of uniformly modeling all components.

[0028] Differentiated PSO parameter search spaces are designed for different categories of components. For high-frequency noise components, which are prone to overfitting, the search space is limited to a smaller tree depth, a lower subsampling ratio, and a lower feature sampling ratio. L1 regularization is introduced to enhance model sparsity, thereby suppressing the risk of overfitting. For detail components, a moderate tree depth and subsampling ratio are used, and L2 regularization is introduced to balance fitting accuracy and generalization ability. For trend components, which require sufficient learning of long-term dependencies, the search space is set to a larger tree depth, a larger number of decision trees, and a higher subsampling ratio and feature sampling ratio to enhance the model's ability to capture long-term trends. This differentiated design ensures that the model complexity of each component matches its own characteristics, avoiding overfitting of complex models on simple components while ensuring that simple models have sufficient expressive power on complex components.

[0029] Finally, for each intrinsic mode function component, the PSO algorithm is used to find the XGBoost hyperparameter combination that minimizes the prediction error on the validation set within its corresponding differential search space. By decoupling and optimizing the modeling tasks of each component separately, the overall prediction accuracy is significantly improved, and the model's generalization ability on different building data is further enhanced.

[0030] S4. A particle population is established based on the differential PSO parameter search space and initialized. Adaptive mode switching is performed based on the initialized particle population to update the particle positions. This process is repeated multiple times to obtain the final global optimal solution and the trained XGBoost model of the intrinsic mode function components of each category. Specifically, step S4 further includes: using Latin hypercube sampling to generate a particle set accounting for 80% of the particle population size, ensuring that the particle population is uniformly covered in the differentiated PSO parameter search space, and injecting the remaining 20% ​​of particles with an experience-based elite seed parameter set (that is, directly introducing excellent parameter combinations determined by previous experience or heuristic rules into the population to accelerate convergence), thus obtaining the particle population. ,in, The set of particles generated by Latin hypercube sampling. A collection of experienced elite seeds; Each particle in the particle swarm is initialized, where the initial velocity of the nth particle is... The initial individual best position of the nth particle is The initial global optimal position of the particle swarm is , Let n be the initial position of the nth particle. Let n be the fitness function (i.e., the prediction error RMSE on the validation set), and n be the number of particles. Based on the initialized particle population, adaptive mode switching is performed, where, for the... In the next iteration, the adaptive switching probability (which decreases linearly with each iteration) is calculated. , , To maximize the number of iterations, a random number following a uniform distribution is generated for each particle. When it is determined Execute PSO mode to update particle positions when it is determined that The WOA (Whale Optimization Algorithm) mode is executed to update particle positions. After the particle positions are updated, calculate the fitness value of the new particle position for the nth particle. When it is determined At that time, the updated best historical position of the individual is obtained as follows: The updated global optimal position is When it is determined At that time, maintain the individual's best historical position as ,in, Let be the position of the nth particle at iter+1 iteration. Let be the individual historical best position of the nth particle at the iter iteration. Let be the individual historical best position of the nth particle at iter+1 iterations. The global optimal position of the particle swarm at iter+1 iterations; Repeat the above steps (except for the initialization step) until the maximum number of iterations is reached, and obtain the final global optimal solution. And use it as the optimal hyperparameter combination of the XGBoost model for the i-th intrinsic mode function component. We obtain the trained XGBoost models with intrinsic mode function components for each category.

[0031] In PSO mode, particle positions are updated, specifically using the standard particle swarm optimization algorithm. The velocity update formula is as follows: The position update formula is: , ,in, Let n be the velocity of the nth particle in the (iter+1)th iteration. For inertial weights, Let n be the velocity of the nth particle at the iter iteration. , All are learning factors. , All are uniformly distributed random numbers. This represents the position of the nth particle at the iter iteration. Execute WOA mode to update particle positions, specifically by using the spiral update mechanism of the whale optimization algorithm to simulate the bubble-net feeding behavior of humpback whales. The position update formula is as follows: , where e is the base of the natural logarithm, b is a constant controlling the shape of the logarithmic spiral (usually taken as 1), and l is a uniformly distributed random number used to control the step size and direction of the spiral update. This represents the distance between the current particle and the current global optimum. This mechanism allows the particle to perform a spiral-like, refined search around the current global optimum, enhancing its local exploitation capabilities.

[0032] In this embodiment, to improve the search efficiency of PSO in high-dimensional parameter spaces, an improved WOA strategy is introduced into the PSO framework, proposing HPSOWOA for parameter optimization of the HPSOWOA model. Specifically, for each classified IMF component, a hybrid particle swarm optimization-whale optimization algorithm is used to adaptively optimize the hyperparameters of its XGBoost model. Specifically, a particle population is established based on the differentiated PSO parameter search space and initialized. Latin hypercube sampling is used to generate particles accounting for 80% of the population size, ensuring uniform particle coverage in the search space. Simultaneously, the remaining 20% ​​of particles are injected with an empirically-based elite seed parameter set, that is, optimal parameter combinations determined by heuristic rules are directly introduced into the population. This hybrid initialization strategy ensures population diversity, preventing the search from getting trapped in local optima, and accelerates the convergence process through the guidance of elite seeds, thereby improving optimization efficiency.

[0033] After initialization, each particle is set to an initial velocity of zero, and its initial position is set to its individual historical best position. The position with the best fitness among all particles is then set as the global best position for the population. The adaptive mode switching and position update phase then begins. For each iteration, an adaptive switching probability that decreases linearly with the iteration number is calculated, allowing the algorithm to focus on global exploration in the early stages and local development in the later stages. A uniform random number is generated for each particle. If this random number is less than the switching probability, the PSO mode is executed to update the particle's position; otherwise, the WOA spiral mode is executed. The PSO mode adjusts the direction of motion using a velocity-position update formula, combining the particle's own historical best and the global best, demonstrating strong global search capabilities. The WOA mode simulates the bubble-net hunting behavior of a humpback whale, allowing particles to perform a spiral-like fine-grained search around the current global best solution, enhancing local development capabilities. The adaptive switching between the two modes enables the algorithm to both broadly explore unknown regions in the high-dimensional parameter space and finely search near the optimal solution, effectively avoiding premature convergence. After each position update, the fitness value of the new position is calculated. If it is better than the particle's historical best position, the individual best position is updated; if it is also better than the global best position of the population, the global best position is updated. This iterative process is repeated until the maximum number of iterations is reached. The final output global optimal solution is the optimal hyperparameter combination of the XGBoost model for that eigenmode function component. Through this optimization process, each component obtains optimal model parameters that match its characteristics, laying the model foundation for subsequent high-precision predictions. Compared with the traditional PSO algorithm, this optimization method improves the optimization speed by about 30% and is more likely to obtain the global optimal solution, thus significantly improving the prediction performance of each component model.

[0034] S5. The XGBoost model trained with the intrinsic mode function components of each category (whose hyperparameters are optimized by the HPSOWOA algorithm in step S4) is used to predict the validation set data to obtain the corresponding predicted value sequence. The dynamic weights are constructed using an exponential function, and the predicted value sequence is weighted and summed according to the dynamic weights to obtain the final reconstructed prediction result.

[0035] Specifically, step S5 further includes: using the XGBoost model trained with the intrinsic mode function components of each category to predict the validation set data, and obtaining the predicted value sequence of the i-th intrinsic mode function component at time t. , , , This represents the total number of time indices in the validation set. Calculate the root mean square error of the prediction of the i-th intrinsic mode function component on the validation set. , This is represented as the true value of the i-th component (obtained by VMD decomposition); Based on the root mean square error (RMSE) of each intrinsic mode function component, an exponential function is used to construct dynamic weights, so that components with higher prediction accuracy (i.e., smaller RMSE) receive a larger weight contribution; the formula is as follows: , The dynamic weights of the i-th intrinsic mode function component are... It is an exponential function. Let be the root mean square error of the prediction of the j-th intrinsic mode function component on the validation set, where the dynamic weights satisfy the normalization condition: ; The predicted value sequences of each intrinsic mode function component are weighted and summed according to dynamic weights to obtain the final reconstruction prediction result. .

[0036] In this embodiment, the dynamic weighted reconstruction mechanism uses the validation set prediction error. This serves as the basis for weight calculation, enabling error feedback and self-correction; through By mapping errors to non-negative weights, high-precision components acquire exponentially higher weights; simultaneously, the weights... It can dynamically and adaptively adjust with the validation set error, breaking through the limitations of traditional fixed or equal weights. By suppressing the contribution of high error components, it weakens the influence of noise and improper modeling, ultimately improving prediction robustness and making the combined result statistically approach the optimal solution with the minimum overall prediction error variance.

[0037] Specifically, the trained XGBoost models for each component are used to predict the validation set data, obtaining a sequence of predicted values ​​for each intrinsic mode function component. Then, the root mean square error (RMSE) of each component model on the validation set is calculated; this error directly reflects the prediction accuracy of each component model. Based on the prediction errors of each component, an exponential function is used to construct dynamic weights, giving greater weight to components with higher prediction accuracy (i.e., smaller RMSE). All weights are normalized to ensure the sum of all weights is 1. Finally, the predicted value sequences of each component are weighted and summed according to the calculated dynamic weights to obtain the final reconstructed prediction result. This dynamic weighting mechanism achieves self-correction based on error feedback: when a component model performs poorly on the validation set, its weight is automatically reduced, thus suppressing the negative impact of that component's error on the final result; conversely, components with high prediction accuracy receive higher weights, fully leveraging their advantages. Compared to the traditional method of linearly superimposing all components with equal or fixed weights, the dynamic weighted reconstruction mechanism of this invention significantly enhances the robustness and accuracy of the overall prediction model.

[0038] Preferably, in this embodiment, the method further includes: designing an online incremental learning mechanism and a federated transfer learning mechanism for online data stream scenarios and multi-building joint prediction scenarios, respectively, to solve the problems of "post-deployment stagnation" and "cold start" in traditional methods; wherein, for online data stream scenarios, the online incremental learning mechanism sets a sliding window (such as a week's length) to achieve continuous model updates and adaptation, and for multi-building joint prediction scenarios, the federated transfer learning mechanism introduces the federated learning framework FedProx-Dyn (Federated Proximal with Dynamic Regularization) to achieve cross-building knowledge sharing and model collaboration under privacy protection; The online incremental learning mechanism is as follows: When new data enters the window, adaptive decomposition parameter optimization based on energy entropy is used to determine the optimal number of modes and the optimal penalty factor. Variational mode decomposition is then performed on the new data to obtain the values ​​of each intrinsic mode function component in the new interval. For the prediction model of each intrinsic mode function component, the gradient of the loss function is calculated using the new data, and one or more steps of gradient descent update are performed. The dynamic weights of each component are recalculated based on the new error within the sliding window. The predicted values ​​of each component within the new data interval are weighted and reconstructed to obtain the final prediction result. The federated transfer learning mechanism works as follows: a building calculates the update amount of model parameters based on local data, or directly calculates the parameter change amount after gradient update; collects the parameter update amounts of each client, and aggregates them according to the proportion of each building's data amount to the total data amount to obtain new global model parameters; adds a proximate term to the local loss function to form a federated proximate mechanism; for newly added buildings, the global model obtained by federated learning is directly used as the initial parameters, and a small amount of local data is used for rapid fine-tuning.

[0039] Specifically, in this embodiment, to expand application scenarios, the present invention proposes an enhanced framework supporting online learning and multi-building collaboration. The triggering mechanism for the online incremental learning mechanism is designed as follows: a dual-threshold sliding window strategy is adopted, triggering incremental model updates when any of the following conditions are met: the time window condition is: periodic updates are triggered every N new data points received (N=168 hours, i.e., one week); the performance drift condition is: continuous monitoring of prediction errors, if the RMSE increases by more than a threshold compared to the baseline value within the sliding window... This triggered an emergency update.

[0040] Specifically, the first step is new data decomposition, when newly arrived window data... Enter the window, and using the optimal decomposition parameters determined in step S2, perform VMD decomposition on the new data to obtain the values ​​of each component in the new interval. The second step is model gradient fine-tuning. For the prediction model of the i-th component (using an elastic weight consolidation method to prevent catastrophic forgetting), the gradient of the loss function is calculated using the new data, and one or more steps of gradient descent update are performed: ,in: The parameters of the i-th component model before updating; These are the parameters of the updated i-th component model; Let L be the gradient of the loss function L with respect to the new data under the old parameters.

[0041] The third step is dynamic weight update, which recalculates the dynamic weights of each component based on the new error within the sliding window. The formula is as follows: , Finally, online reconstruction of predictions: the predicted values ​​of each component within the new data interval are weighted and reconstructed to obtain the final prediction result. .

[0042] In this embodiment, a multi-building federated migration mechanism is implemented: For multi-building joint prediction scenarios, the federated learning framework FedProx-Dyn is introduced to achieve cross-building knowledge sharing and model collaboration while protecting privacy. A non-independent, identically distributed data processing strategy is adopted: To address the issue of large differences in the distribution of data across different buildings, dynamic regularization coefficient adjustment is used. ,in Let be the regularization coefficient for the j-th building; Local data distribution for the j-th building With global distribution The KL divergence is used to measure the degree of data heterogeneity; This is the scaling factor. The greater the difference in data distribution, the larger the regularization coefficient, to prevent the local model from deviating excessively from the global consensus.

[0043] Specifically, 1. Local model update: The j-th building is updated based on local data. Calculate the update amount of the model parameters Alternatively, the parameter changes after gradient updates can be calculated directly. 2. Server-side weighted aggregation: The central server collects parameter updates from each client and aggregates them according to the amount of data for each building. Percentage of total data volume We perform weighted aggregation based on the proportions to obtain new global model parameters: 3. Local loss function regularization: To prevent client models from deviating excessively from global consensus under non-independent and identically distributed data, a proximate term is added to the local loss function, forming a federated proximate mechanism. That is, the j-th building is based on local data Calculate the update amount of the model parameters and the local loss function. Where: Let be the original loss function for the local data of the j-th building; This is the regularization coefficient, which controls the penalty strength for local parameters deviating from the global model; This is the squared Euclidean distance between the local and global parameters.

[0044] 4. Communication efficiency optimization: To reduce communication overhead, gradient compression and sparsity strategies are adopted: only the k% gradient parameters with the largest absolute value are uploaded (k=10), and Gaussian noise is added using a differential privacy mechanism to protect local data privacy. 5. Cold start and knowledge transfer for new buildings: For newly added buildings (with no or very little historical data), a meta-learning initialization strategy is adopted: the global model obtained from federated learning can be used directly. As initial parameters, combined with a small amount of local data Perform quick fine-tuning: This enables the transfer of learning from existing architectural knowledge to new buildings, fundamentally alleviating the cold start problem.

[0045] In this embodiment, to further verify the effectiveness and superiority of the method of the present invention, a systematic comparative experiment was conducted on the ASHRAE GreatEnergy Predictor III competition dataset. This dataset contains hourly energy consumption data for various building types (office buildings, shopping malls, hotels, etc.) and corresponding meteorological and temporal characteristics, spanning three years. The training set, validation set, and test set are divided in chronological order (ratio 7:1:2). The prediction task is a rolling prediction of hourly energy consumption for the next 24 hours. Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination (R²) are used as evaluation metrics. The prediction performance comparison is shown in Table 1.

[0046] Table 1: Comparison of Predictive Performance

[0047] First, a comparison method was set up, selecting three representative methods as baselines. All methods used the same data partitioning: Baseline model: PSO-XGBoost, which directly predicts the original energy consumption sequence, with hyperparameters optimized by standard PSO. Traditional decomposition and ensemble method 1: EMD-PSO-XGBoost, which uses empirical mode decomposition to decompose the energy consumption sequence into several intrinsic mode function components, and establishes an XGBoost model for each component, with hyperparameters optimized by standard PSO. Traditional decomposition and ensemble method 2: Wavelet-PSO-XGBoost, which uses wavelet decomposition instead of EMD, with the remaining settings the same as above. Method 1 of this invention: VMD-complete framework, which uses variational mode decomposition, automatically determines the optimal number of modes K and penalty factor α through adaptive decomposition parameter optimization based on energy entropy; it performs differentiated modeling for each component based on sample entropy and average period, with hyperparameters optimized by the HPSOWOA algorithm; it introduces frequency domain-time domain fusion features in feature construction; and finally, it obtains the prediction result through dynamic weighted reconstruction based on error feedback. Method 2 of the present invention: CEEMDAN-complete framework, which replaces VMD in the above framework with fully adaptive noise set empirical mode decomposition, while keeping the other modules unchanged, to verify the universality of the framework of the present invention for different decomposition methods.

[0048] Secondly, to quantify the independent contribution of each innovation module, based on the VMD-complete framework, key modules were removed one by one for ablation experiments, and the results are shown in Table 2.

[0049] Table 2 Ablation Experiment Results:

[0050] Ablation experiments show that each module contributes positively to the final accuracy, with the adaptive decomposition and frequency-time domain fusion features having the most significant impact, thus verifying the necessity of the various innovations in the technical solution of this invention.

[0051] In summary, this invention aims to solve the following problems: (1) Decomposition parameter adaptation problem: How to automatically determine the optimal decomposition parameters of VMD / CEEMDAN based on the inherent characteristics of the original energy consumption data, and eliminate the subjectivity and uncertainty of manual parameter tuning. (2) Component differentiation modeling problem: How to design differentiated prediction models and parameter search strategies based on the frequency domain characteristics and complexity of each IMF, and achieve refined modeling. (3) Intelligent reconstruction and fusion problem: How to design a dynamic weighted reconstruction mechanism with "self-correction" capability, and dynamically adjust the fusion weights based on the real-time performance of each component model. (4) Parameter optimization efficiency problem: How to integrate the advantages of PSO and WOA, improve the efficiency and global optimality of high-dimensional parameter search, and avoid getting trapped in local optima. (5) Feature dimension expansion problem: How to automatically generate high-order meta-features of fused frequency domain-time domain information, and fully explore the information value of decomposition components. (6) Application scenario expansion problem: How to support online incremental learning and multi-building collaborative prediction, and solve the cold start and knowledge isolation problems after model deployment.

[0052] Compared with existing technologies, this invention has the following beneficial effects: 1. Improved prediction accuracy: By using adaptive decomposition parameter optimization based on energy entropy and the coupling theorem of modal aliasing degree and prediction accuracy, the mathematical correlation between decomposition quality and subsequent prediction tasks is ensured, improving information utilization efficiency from the source. Compared with fixed parameter decomposition methods, the prediction error can be reduced by more than 15%. 2. Enhanced model robustness: The dynamic weighted intelligent reconstruction mechanism based on error feedback can dynamically adjust the contribution based on the real-time performance of each component model, avoiding the impact of a single model failure on the whole, and effectively overcoming the defects of traditional linear superposition methods. 3. Improved optimization efficiency and effect: The HPSOWOA optimization algorithm combines the global exploration capability of PSO and the local development capability of WOA through an adaptive switching mechanism, and uses a hybrid initialization strategy to improve the optimization speed in high-dimensional parameter space by 30%, and is more likely to find the global optimum. 4. Reduced manual intervention and adaptive implementation: The entire process, from determining decomposition parameters to model selection and reconstruction, is highly automated, reducing the workload and subjectivity of manual parameter tuning, making the method easy to deploy to different buildings. 5. Expanding application scenarios and possessing engineering practicality: The incremental learning and federated transfer collaborative framework enables the model to adapt to changes in data distribution and achieve knowledge sharing among multiple buildings while protecting privacy. It solves the problems of insufficient cold start and generalization ability of the model and has high practical application value.

[0053] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A building energy consumption prediction method based on adaptive decomposition and intelligent reconstruction, characterized in that, include: Acquire historical energy consumption data of the target building, along with corresponding meteorological data and timestamp information, and perform data preprocessing to obtain the original energy consumption sequence; Based on the original energy consumption sequence, adaptive decomposition parameter optimization based on energy entropy is performed to determine the optimal number of modes and the optimal penalty factor. Based on the optimal number of modes and the optimal penalty factor, variational mode decomposition is performed on the original energy-consuming sequence to obtain multiple intrinsic mode function components. Each intrinsic mode function component is then classified, and a differentiated PSO parameter search space corresponding to each category is designed. A particle population is established based on the differential PSO parameter search space, initialized, and then adaptive mode switching is performed based on the initialized particle population to update the particle positions. This process is repeated multiple times to obtain the final global optimal solution and the trained XGBoost model with the intrinsic mode function components of each category. The XGBoost model, trained using the intrinsic mode function components of each category, is used to predict the validation set data, resulting in a sequence of predicted values. Dynamic weights are constructed using an exponential function, and the predicted value sequences are weighted and summed according to the dynamic weights to obtain the final reconstructed prediction result.

2. The building energy consumption prediction method based on adaptive decomposition and intelligent reconstruction according to claim 1, characterized in that, The data preprocessing includes cleaning, handling of missing and outlier values, and normalization.

3. The building energy consumption prediction method based on adaptive decomposition and intelligent reconstruction according to claim 1, characterized in that, Based on the original energy consumption sequence, adaptive decomposition parameter optimization based on energy entropy is performed to determine the optimal number of modes and the optimal penalty factor, specifically: For the original energy consumption sequence The initial value of the loop attempt number m is set to 2, and the mode number K = m. The loop attempt number m is gradually increased with a step size of 1. For each value of the loop attempt number m, the mode number K = m is set, and variational mode decomposition is performed once to obtain a eigenmode function components. , For the first intrinsic mode function component, Let a be the a-th intrinsic mode function component, where a is the number of intrinsic mode function components, and a = K; The formula for calculating the energy density and relative energy percentage of each eigenmode function component is as follows: , , ,in, Let i be the energy density of the i-th eigenmode function component. Let j be the energy density of the j-th eigenmode function component. For the integral sign, This indicates differentiating with respect to time t. Let i be the value of the i-th intrinsic mode function component at time t. The relative energy percentage of the i-th eigenmode function component; Calculate the total energy entropy of the variational mode decomposition corresponding to the loop attempt number m. And calculate the rate of change of energy entropy under adjacent mode numbers. , The total energy entropy for the variational mode decomposition corresponding to index m-1 in the cyclic attempts. It is a logarithmic function; When the first judgment is made When determining the optimal number of modes. for: , For the preset threshold, .

4. The building energy consumption prediction method based on adaptive decomposition and intelligent reconstruction according to claim 3, characterized in that, Also includes: Define modal aliasing index The sum of the absolute values ​​of the cross-correlation coefficients among the variational modal components is given by the formula: , The Pearson correlation coefficient is used. For the i-th intrinsic mode function component, Let j be the j-th eigenmode function component; Within the predefined search range Within, the modal aliasing index will be minimized. As an optimization objective, use grid search or the HPSOWOA algorithm to find the optimal penalty factor. , As a penalty factor, This is the lower limit of the search range. This is the upper limit of the search range.

5. The building energy consumption prediction method based on adaptive decomposition and intelligent reconstruction according to claim 4, characterized in that, Based on the optimal number of modes and the optimal penalty factor, variational mode decomposition is performed on the original energy-consuming sequence to obtain multiple intrinsic mode function components. These intrinsic mode function components are then classified, and a differentiated PSO parameter search space is designed for each category. Specifically: Based on optimal mode number and optimal penalty factor For the original energy consumption sequence Variational mode decomposition is performed to obtain Each intrinsic mode function component ; The sample entropy and average period of each intrinsic mode function component are calculated using the following formula: , , , , Let be the sample entropy of the i-th intrinsic mode function component. It is the natural logarithm. Given a tolerance r, the length is The number of self-similar templates matched by the vector. Given a tolerance r, the length is The number of self-similar templates matched by the vector. Let be the average period of the i-th intrinsic mode function component. Let be the dominant frequency of the i-th eigenmode function component. For Fast Fourier Transform; Based on sample entropy and average period, The intrinsic mode function components are divided into three categories, and the classification formula is as follows: ,in, For high threshold of sample entropy, For low threshold of sample entropy, For long-period thresholds, For short-cycle thresholds; For different categories of intrinsic mode function components, a corresponding differentiated PSO parameter search space is designed for their XGBoost prediction models to match the physical nature and complexity of each component. For each intrinsic mode function component, the PSO algorithm is used to minimize the prediction error within its corresponding differential PSO parameter search space. The objective function is: , Let i be the objective function corresponding to the i-th intrinsic mode function component. For the root mean square error on the validation set, For XGBoost prediction models, This is a portion of the training set for the i-th intrinsic mode function component. For the hyperparameter combination to be optimized, Let be the search space for the differentiated PSO parameters corresponding to the i-th intrinsic mode function component.

6. The building energy consumption prediction method based on adaptive decomposition and intelligent reconstruction according to claim 5, characterized in that, The search space for differentiated PSO parameters for high-frequency noise components is: , , , , The search space for differentiated PSO parameters for detailed components is: , , , , ; The search space for differentiated PSO parameters for trend components is: , , , , ,in, The maximum depth of the tree. For learning rate, The subsampling ratio, The feature sampling ratio, These are the L1 regularization parameters for XGBoost. These are the L2 regularization parameters for XGBoost. The number of trees.

7. The building energy consumption prediction method based on adaptive decomposition and intelligent reconstruction according to claim 5, characterized in that, A particle population is established based on the differentiated PSO parameter search space and initialized. Adaptive mode switching is then performed based on the initialized particle population to update particle positions. This process is repeated multiple times to obtain the final global optimum, resulting in a trained XGBoost model with eigenmode function components for each category. Specifically: Latin hypercube sampling is used to generate a particle set that accounts for 80% of the particle population size, ensuring that the particle population is uniformly covered in the differentiated PSO parameter search space. The remaining 20% ​​of particles are then injected with an empirically based elite seed parameter set to obtain the particle population. ,in, The set of particles generated by Latin hypercube sampling. A collection of experienced elite seeds; Each particle in the particle swarm is initialized, where the initial velocity of the nth particle is... The initial individual best position of the nth particle is The initial global optimal position of the particle swarm is , Let n be the initial position of the nth particle. The fitness function is denoted by n, where n is the number of particles. Based on the initialized particle population, adaptive mode switching is performed, where, for the... In the next iteration, the adaptive switching probability is calculated. , , To maximize the number of iterations, a random number following a uniform distribution is generated for each particle. When it is determined Execute PSO mode to update particle positions when it is determined that Execute WOA mode to update particle positions; After the particle positions are updated, calculate the fitness value of the new particle position for the nth particle. When it is determined At that time, the updated best historical position of the individual is obtained as follows: The updated global optimal position is When it is determined At that time, maintain the individual's best historical position as ,in, Let be the position of the nth particle at iter+1 iteration. Let be the individual historical best position of the nth particle at the iter iteration. Let be the individual historical best position of the nth particle at iter+1 iterations. This represents the globally optimal position of the particle swarm at iter+1 iterations. Repeat the above steps until the maximum number of iterations is reached, and the final global optimal solution is obtained. And use it as the optimal hyperparameter combination of the XGBoost model for the i-th intrinsic mode function component. We obtain the trained XGBoost models with intrinsic mode function components for each category.

8. The building energy consumption prediction method based on adaptive decomposition and intelligent reconstruction according to claim 7, characterized in that, In PSO mode, particle positions are updated, specifically using the standard particle swarm optimization algorithm. The velocity update formula is as follows: The position update formula is: , ,in, Let n be the velocity of the nth particle in the (iter+1)th iteration. For inertial weights, Let n be the velocity of the nth particle at the iter iteration. , All are learning factors. , All are uniformly distributed random numbers. This represents the position of the nth particle at the iter iteration. Execute WOA mode to update particle positions, specifically by using the spiral update mechanism of the whale optimization algorithm to simulate the bubble-net feeding behavior of humpback whales. The position update formula is as follows: , where e is the base of the natural logarithm, b is a constant controlling the shape of the logarithmic spiral, and l is a uniformly distributed random number used to control the step size and direction of the spiral update. This represents the distance between the current particle and the current global best position.

9. The building energy consumption prediction method based on adaptive decomposition and intelligent reconstruction according to claim 8, characterized in that, The XGBoost model, trained using the intrinsic mode function components of each category, is used to predict the validation set data, resulting in a sequence of predicted values. Dynamic weights are constructed using an exponential function, and the predicted value sequences are then weighted and summed according to these dynamic weights to obtain the final reconstructed prediction result. Specifically: The XGBoost model, trained with intrinsic mode function components of each category, is used to predict the validation set data, resulting in the predicted value sequence of the i-th intrinsic mode function component at time t. , , , This represents the total number of time indices in the validation set. Calculate the root mean square error of the prediction of the i-th intrinsic mode function component on the validation set. ; Based on the root mean square error of the prediction of each intrinsic mode function component, an exponential function is used to construct dynamic weights, the formula of which is: , The dynamic weights of the i-th intrinsic mode function component are... It is an exponential function. Let be the root mean square error of the prediction of the j-th intrinsic mode function component on the validation set, where the dynamic weights satisfy the normalization condition: ; The predicted value sequences of each intrinsic mode function component are weighted and summed according to dynamic weights to obtain the final reconstruction prediction result. .

10. The building energy consumption prediction method based on adaptive decomposition and intelligent reconstruction according to claim 1, characterized in that, Also includes: For online data stream scenarios and multi-building joint prediction scenarios, an online incremental learning mechanism and a federated transfer learning mechanism are designed respectively. For online data stream scenarios, the online incremental learning mechanism sets a sliding window to realize continuous model updates and adaptation. For multi-building joint prediction scenarios, the federated transfer learning mechanism introduces the federated learning framework FedProx-Dyn to achieve cross-building knowledge sharing and model collaboration under privacy protection. The online incremental learning mechanism is as follows: When new data enters the window, adaptive decomposition parameter optimization based on energy entropy is used to determine the optimal number of modes and the optimal penalty factor. Variational mode decomposition is then performed on the new data to obtain the values ​​of each intrinsic mode function component in the new interval. For the prediction model of each intrinsic mode function component, the gradient of the loss function is calculated using the new data, and one or more steps of gradient descent update are performed. The dynamic weights of each component are recalculated based on the new error within the sliding window. The predicted values ​​of each component within the new data interval are weighted and reconstructed to obtain the final prediction result. The federated transfer learning mechanism works as follows: a building calculates the update amount of model parameters based on local data, or directly calculates the parameter change amount after gradient update; collects the parameter update amounts of each client, and aggregates them according to the proportion of each building's data amount to the total data amount to obtain new global model parameters; adds a proximate term to the local loss function to form a federated proximate mechanism; for newly added buildings, the global model obtained by federated learning is directly used as the initial parameters, and a small amount of local data is used for rapid fine-tuning.

Images