Energy consumption decomposition method based on Bayesian optimization and hidden Markov model

Abstract

The invention relates to the technical field of energy load monitoring, in particular to an energy consumption decomposition method based on Bayesian optimization and a hidden Markov model, and the method comprises the following steps: S1, carrying out the data loading of electric equipment and the construction of a feature library, carrying out the steady-state voltage and current signal collection of a plurality of pieces of electric equipment, recording the electrical parameters of the electric equipment in different operation states, calculating statistical characteristics of active power and reactive power for each electric device and the state of the electric device, describing power characteristics of each electric device in different states through the statistical characteristics, and estimating a state transition probability matrix of the electric device according to an operation sequence of the electric device; the state transition probability matrix describes a rule of transition of the electric equipment from one operation state to another state, and provides a basis for subsequent time sequence modeling. According to the method, discriminant learning and time sequence modeling are combined, high precision and interpretability are achieved, and the method is suitable for energy consumption identification and energy efficiency optimization in the building and industrial fields.

Description

Technical Field

[0001] This invention relates to the field of energy load monitoring technology, specifically to an energy consumption decomposition method based on Bayesian optimization and hidden Markov models. Background Technology

[0002] With the accelerating electrification of the building and industrial sectors and increasingly stringent energy accountability requirements, energy management systems face new challenges in refined monitoring and energy efficiency optimization. Traditional invasive load monitoring methods require the installation of independent sensors on each electrical device, resulting in high equipment costs and complex maintenance, making them unsuitable for widespread application in large-scale building environments. Non-invasive load monitoring technology, by analyzing total current and voltage signals, infers the operating status and energy consumption level of each electrical device. Due to its lower cost and ease of installation, it has become a key tool in smart grid and building energy conservation management. In recent years, with the development of machine learning technology, non-invasive load monitoring methods based on deep learning and ensemble learning have achieved significant improvements in decomposition accuracy.

[0003] Existing discriminative models can provide intuitive state determination criteria through feature importance analysis, but they are difficult to accurately reflect the temporal characteristics of electrical equipment operation. Generative models can characterize the time transition law of electrical equipment operation state, but they have high computational complexity in the scenario of multiple electrical devices operating in parallel and have limited ability to express high-dimensional features. The above methods cannot simultaneously balance decomposition accuracy, computational efficiency and interpretability. Therefore, they do not meet the current requirements. To address this, we propose an energy consumption decomposition method based on Bayesian optimization and hidden Markov models. Summary of the Invention

[0004] The purpose of this invention is to provide an energy consumption decomposition method based on Bayesian optimization and hidden Markov models, in order to solve the problems mentioned in the background art. Existing discriminative models can provide intuitive state judgment criteria through feature importance analysis, but they are difficult to accurately reflect the temporal characteristics of electrical equipment operation. Generative models can characterize the time transition law of electrical equipment operation state, but they have high computational complexity in the scenario of multiple electrical devices operating in parallel and have limited ability to express high-dimensional features. The above methods are difficult to simultaneously take into account the problems of decomposition accuracy, computational efficiency and interpretability.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an energy consumption decomposition method based on Bayesian optimization and hidden Markov models, comprising the following steps: S1: Loading and building a feature database for electrical equipment. This involves collecting steady-state voltage and current signals from multiple electrical devices and recording their electrical parameters under different operating conditions. Then, for each electrical device and its state, the statistical characteristics of active and reactive power are calculated. These statistical characteristics can characterize the power characteristics of each electrical device under different conditions. At the same time, the state transition probability matrix is ​​estimated based on the operating sequence of the electrical devices. The state transition probability matrix describes the pattern of the electrical devices transitioning from one operating state to another, providing a foundation for subsequent time series modeling. The final equipment feature database contains power statistical characteristics and state transition probabilities, providing prior knowledge support for the entire energy consumption decomposition process. S2: Bayesian-optimized boosting model pruning. Based on the feature library, a Bayesian-optimized boosting model is used to determine the status of electrical equipment. The boosting model takes feature vectors as input and outputs predicted values ​​through the superposition of multiple regression trees. Bayesian optimization can be used to improve the hyperparameters of the model. Perform adaptive adjustments and minimize the acquisition function. After training, the model is improved to predict the probability of whether each electrical device is in the on state. If the probability is lower than the set threshold, the state branch of the electrical device is pruned at that time point, thereby significantly reducing the state space that needs to be considered in subsequent inferences. S3: Factor Hidden Markov Model Construction. The pruned set of electrical devices is input into the factor hidden Markov model for time series modeling. The total power is then analyzed using the factor hidden Markov model. We assume that the observation probability can be defined as a multivariate Gaussian distribution. Then, through an approximate inference algorithm, the most likely state sequence of each electrical device at each time point can be solved; S4: Posterior state inference and joint estimation. Posterior estimation is performed on the joint state of the electrical equipment at each time step. The maximum posterior probability of each electrical equipment state is calculated. Combined with the state transition probability constraints of each electrical equipment, the optimal state path of the entire time series can be obtained. In this process, approximate dynamic programming or Viterbi algorithm can be used to ensure the continuity of time series and physical rationality. S5: Energy scaling and balance correction. In order to meet the energy conservation constraint, the power estimation results of each electrical device are globally balanced and adjusted, and a scaling factor is defined. Then, the corrected power of each electrical device is estimated to ensure that the decomposition results are strictly consistent at the total power level, while maintaining the relative power ratio between multiple electrical devices. S6: Noise injection and smoothing correction introduces random noise terms to simulate power fluctuations in actual operation and enhances the realism of the power sequence. This can effectively avoid the power curve from being too smooth and make the output results closer to the real working conditions. S7: Results Output and Performance Evaluation. Outputs the operating status and decomposed power sequence of each electrical device at each time step. The accuracy of the factor hidden Markov model can be evaluated by the mean absolute error and mean relative error. The evaluation results are visualized using curves and error fluctuation graphs to show the energy consumption changes and decomposition accuracy of each electrical device.

[0006] Preferably, the electrical parameters of the electrical equipment in S1 include active power and reactive power, wherein the active power is determined by the actual active power. and estimated active power The statistical characteristics comprise the mean and standard deviation; ; ; ; ; in, The operating sequence of electrical equipment. and These represent the active power and reactive power of the sample points, respectively. The average of the actual active power and reactive power is given. The standard deviation of the actual active power and reactive power is given by the following. To estimate the mean of active power and reactive power, the... To estimate the standard deviation of active power and reactive power; The state transition law of the electrical equipment is a state transition probability matrix. ; .

[0007] Preferably, the objective function of the boosting model in S2 includes a loss function. and regularization term ; ; ; ; ; in, The number of leaf nodes. This represents the node weight.

[0008] Preferably, the total power in S3 Gaussian noise is superimposed on the linear combination of the power output of each electrical device; ; ; ; in, Let be the expected power vector of the electrical equipment in a certain state.

[0009] Preferably, the maximum posterior probability of the electrical equipment state in S4 is ; .

[0010] Preferably, the scaling factor in S5 is ; ; ; in, The total measured power of the system, Corrected power for each electrical device.

[0011] Preferably, the random noise term in S6 is ; ; in, This is the noise intensity scaling factor.

[0012] Preferably, the mean absolute error of the factorial hidden Markov model in S7 is The average relative error of the factorial hidden Markov model is ; ; .

[0013] Compared with the prior art, the beneficial effects of the present invention are: This invention improves the discriminative pruning mechanism of the model through Bayesian optimization, enabling real-time screening of electrical equipment states, significantly compressing the state space, reducing combinatorial search complexity, and improving computational efficiency. It facilitates rapid deployment in scenarios with multiple electrical equipment. Through temporal modeling and energy scaling of the factorial hidden Markov model, the physical rationality of the electrical equipment state sequence and the conservation of total power are ensured. At the same time, appropriate random perturbations are introduced to enhance the authenticity and robustness of the results. The overall process balances discrimination accuracy, temporal continuity, and interpretability, making it suitable for non-intrusive load decomposition tasks in real-world scenarios such as buildings, industries, and smart grids. Attached Figure Description

[0014] Figure 1 This is a flowchart of the entire invention. Detailed Implementation

[0015] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0016] Please see Figure 1 The present invention provides an embodiment of an energy consumption decomposition method based on Bayesian optimization and hidden Markov models, comprising the following steps: S1: Loading and building a feature database for electrical equipment. This involves collecting steady-state voltage and current signals from multiple electrical devices and recording their electrical parameters under different operating conditions. Then, for each electrical device and its state, the statistical characteristics of active and reactive power are calculated. These statistical characteristics can characterize the power characteristics of each electrical device under different conditions. At the same time, the state transition probability matrix is ​​estimated based on the operating sequence of the electrical devices. The state transition probability matrix describes the pattern of the electrical devices transitioning from one operating state to another, providing a foundation for subsequent time series modeling. The final equipment feature database contains power statistical characteristics and state transition probabilities, providing prior knowledge support for the entire energy consumption decomposition process. S2: Bayesian-optimized boosting model pruning. Based on the feature library, a Bayesian-optimized boosting model is used to determine the status of electrical equipment. The boosting model takes feature vectors as input and outputs predicted values ​​through the superposition of multiple regression trees. Bayesian optimization can be used to improve the hyperparameters of the model. Perform adaptive adjustments and minimize the acquisition function. After training, the model is improved to predict the probability of whether each electrical device is in the on state. If the probability is lower than the set threshold, the state branch of the electrical device is pruned at that time point, thereby significantly reducing the state space that needs to be considered in subsequent inferences. S3: Factor Hidden Markov Model Construction. The pruned set of electrical devices is input into the factor hidden Markov model for time series modeling. The total power is then analyzed using the factor hidden Markov model. We assume that the observation probability can be defined as a multivariate Gaussian distribution. Then, through an approximate inference algorithm, the most likely state sequence of each electrical device at each time point can be solved; S4: Posterior state inference and joint estimation. Posterior estimation is performed on the joint state of the electrical equipment at each time step. The maximum posterior probability of each electrical equipment state is calculated. Combined with the state transition probability constraints of each electrical equipment, the optimal state path of the entire time series can be obtained. In this process, approximate dynamic programming or Viterbi algorithm can be used to ensure the continuity of time series and physical rationality. S5: Energy scaling and balance correction. In order to meet the energy conservation constraint, the power estimation results of each electrical device are globally balanced and adjusted, and a scaling factor is defined. Then, the corrected power of each electrical device is estimated to ensure that the decomposition results are strictly consistent at the total power level, while maintaining the relative power ratio between multiple electrical devices. S6: Noise injection and smoothing correction introduces random noise terms to simulate power fluctuations in actual operation and enhances the realism of the power sequence. This can effectively avoid the power curve from being too smooth and make the output results closer to the real working conditions. S7: Results Output and Performance Evaluation. Outputs the operating status and decomposed power sequence of each electrical device at each time step. The accuracy of the factor hidden Markov model can be evaluated by the mean absolute error and mean relative error. The evaluation results are visualized using curves and error fluctuation graphs to show the energy consumption changes and decomposition accuracy of each electrical device.

[0017] The electrical parameters of electrical equipment include active power and reactive power. Active power is determined by the actual active power. and estimated active power Composition, statistical characteristics include mean and standard deviation; ; ; ; ; in, The operating sequence of electrical equipment. and These represent the active power and reactive power of the sample points, respectively. This is the average of the actual active power and reactive power. The standard deviation of actual active power and reactive power. To estimate the mean of active power and reactive power, To estimate the standard deviation of active power and reactive power; The state transition rules of electrical equipment are represented by the state transition probability matrix. ; .

[0018] The objective function of the improved model includes the loss function. and regularization term ; ; ; ; ; in, The number of leaf nodes. This represents the node weight.

[0019] Total power Gaussian noise is superimposed on the linear combination of the power output of each electrical device; ; ; ; in, Let be the expected power vector of the electrical equipment in a certain state.

[0020] The maximum posterior probability of the state of electrical equipment is ; .

[0021] Scaling factor is ; ; ; in, The total measured power of the system, Corrected power for each electrical device.

[0022] The random noise term is ; ; in, This is the noise intensity scaling factor.

[0023] The mean absolute error of the factorial hidden Markov model is The mean relative error of the factorial hidden Markov model is ; ; .

[0024] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. An energy consumption decomposition method based on Bayesian optimization and hidden Markov models, characterized in that, Includes the following steps: S1: Loading and building a feature database for electrical equipment. This involves collecting steady-state voltage and current signals from multiple electrical devices and recording their electrical parameters under different operating conditions. Then, for each electrical device and its state, the statistical characteristics of active and reactive power are calculated. These statistical characteristics can characterize the power characteristics of each electrical device under different conditions. At the same time, the state transition probability matrix is ​​estimated based on the operating sequence of the electrical devices. The state transition probability matrix describes the pattern of the electrical devices transitioning from one operating state to another, providing a foundation for subsequent time series modeling. The final equipment feature database contains power statistical characteristics and state transition probabilities, providing prior knowledge support for the entire energy consumption decomposition process. S2: Bayesian-optimized boosting model pruning. Based on the feature library, a Bayesian-optimized boosting model is used to determine the status of electrical equipment. The boosting model takes feature vectors as input and outputs predicted values ​​through the superposition of multiple regression trees. Bayesian optimization can be used to improve the hyperparameters of the model. Perform adaptive adjustments and minimize the acquisition function. After training, the model is improved to predict the probability of whether each electrical device is in the on state. If the probability is lower than the set threshold, the state branch of the electrical device is pruned at that time point, thereby significantly reducing the state space that needs to be considered in subsequent inferences. S3: Factor Hidden Markov Model Construction. The pruned set of electrical devices is input into the factor hidden Markov model for time series modeling. The total power is then analyzed through the factor hidden Markov model. We assume that the observation probability can be defined as a multivariate Gaussian distribution. Then, through an approximate inference algorithm, the most likely state sequence of each electrical device at each time point can be solved; S4: Posterior state inference and joint estimation. Posterior estimation is performed on the joint state of the electrical equipment at each time step. The maximum posterior probability of each electrical equipment state is calculated. Combined with the state transition probability constraints of each electrical equipment, the optimal state path of the entire time series can be obtained. In this process, approximate dynamic programming or Viterbi algorithm can be used to ensure the continuity of time series and physical rationality. S5: Energy scaling and balance correction. In order to meet the energy conservation constraint, the power estimation results of each electrical device are globally balanced and adjusted, and a scaling factor is defined. Then, the corrected power of each electrical device is estimated to ensure that the decomposition results are strictly consistent at the total power level, while maintaining the relative power ratio between multiple electrical devices. S6: Noise injection and smoothing correction introduces random noise terms to simulate power fluctuations in actual operation and enhances the realism of the power sequence. This can effectively avoid the power curve from being too smooth and make the output results closer to the real working conditions. S7: Results Output and Performance Evaluation. Outputs the operating status and decomposed power sequence of each electrical device at each time step. The accuracy of the factor hidden Markov model can be evaluated by the mean absolute error and mean relative error. The evaluation results are visualized using curves and error fluctuation graphs to show the energy consumption changes and decomposition accuracy of each electrical device.

2. The energy consumption decomposition method based on Bayesian optimization and hidden Markov model according to claim 1, characterized in that: The electrical parameters of the electrical equipment in S1 include active power and reactive power, wherein the active power is determined by the actual active power. and estimated active power The statistical characteristics comprise the mean and standard deviation; ； ； ； ； in, The operating sequence of electrical equipment. and These represent the active power and reactive power of the sample points, respectively. The average of the actual active power and reactive power is given. The standard deviation of the actual active power and reactive power is given by the following. To estimate the mean of active power and reactive power, the... To estimate the standard deviation of active power and reactive power; The state transition law of the electrical equipment is a state transition probability matrix. ; 。 3. The energy consumption decomposition method based on Bayesian optimization and hidden Markov model according to claim 2, characterized in that: The objective function of the boosting model in S2 includes a loss function. and regularization term ; ； ； ； ； in, The number of leaf nodes. This represents the node weight.

4. The energy consumption decomposition method based on Bayesian optimization and hidden Markov model according to claim 3, characterized in that: The total power in S3 Gaussian noise is superimposed on the linear combination of the power output of each electrical device; ； ； ； in, Let be the expected power vector of the electrical equipment in a certain state.

5. The energy consumption decomposition method based on Bayesian optimization and hidden Markov model according to claim 4, characterized in that: The maximum posterior probability of the electrical equipment status in S4 is: ; 。 6. The energy consumption decomposition method based on Bayesian optimization and hidden Markov model according to claim 5, characterized in that: The scaling factor in S5 is: ; ； ； in, The total measured power of the system, Corrected power for each electrical device.

7. The energy consumption decomposition method based on Bayesian optimization and hidden Markov model according to claim 6, characterized in that: The random noise term in S6 is: ; ； in, This is the noise intensity scaling factor.

8. The energy consumption decomposition method based on Bayesian optimization and hidden Markov model according to claim 7, characterized in that: The mean absolute error of the factorial hidden Markov model in S7 is: The average relative error of the factorial hidden Markov model is ; ； 。

Images