Power load decomposition capable of explaining decision tree and hidden Markov model
By combining interpretable decision trees and hidden Markov models, the problem of insufficient interpretability of NILM models is solved, and accurate decomposition and state identification of power load are achieved, improving the decomposition accuracy and reliability.
Patent Information
- Application Number
- CN202511727888.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-24
- Publication Date
- 2026-05-01
AI Technical Summary
The NILM study suffers from insufficient model interpretability. Deep learning methods are highly accurate but lack explicit mapping, making it difficult to verify the credibility of the results.
By employing interpretable decision trees and hidden Markov models, and through the construction of a device feature library, decision tree pruning, hidden Markov model construction, optimal state combination search, and result visualization, accurate decomposition of power load is achieved.
It achieves accurate decomposition and state identification of power load, provides explicit rule sets and temporal logic constraints, improves decomposition accuracy and interpretability, and meets the needs of power consumption management.
Smart Images
Figure CN121958798A_ABST
Abstract
Description
An Interpretable Decomposition of Power Load Using Decision Trees and Hidden Markov Models Technical Field
[0001] This invention relates to the field of power energy consumption monitoring technology, specifically to a power load decomposition method using interpretable decision trees and hidden Markov models. Background Technology
[0002] As one of the high-carbon emission sectors, the power system faces enormous pressure to reduce emissions and undergo a profound energy transition. According to global energy development plans, the development of smart grid technology has become an inevitable trend. Under this premise, accurate monitoring of user-side equipment energy consumption has become the key to optimizing the energy efficiency of traditional power grids and managing the energy consumption of new power systems. Non-intrusive load monitoring (NILM) technology predicts equipment operating status and real-time energy consumption by analyzing total electricity consumption data. Due to its ease of use, low cost, and easy deployment, it has become an important tool in the fields of smart grids and building energy conservation.
[0003] While NILM research focuses on improving algorithm performance and decomposition accuracy, the problem of insufficient model interpretability is becoming increasingly apparent. Although deep learning methods have high accuracy, their "black box" nature leads to a lack of explicit mapping between input and output, making it difficult to verify the credibility of the results. Therefore, they do not meet the existing needs. To address this, we propose an interpretable decision tree and hidden Markov model for power load decomposition. Summary of the Invention
[0004] The purpose of this invention is to provide an interpretable decision tree and hidden Markov model for power load decomposition, in order to address the problem mentioned in the background art that while NILM research focuses on improving algorithm performance and decomposition accuracy, the problem of insufficient model interpretability is becoming increasingly apparent. Although deep learning methods have high accuracy, their "black box" characteristics lead to a lack of explicit mapping between input and output, making it difficult to verify the credibility of the results.
[0005] To achieve the above objectives, the present invention provides the following technical solution: an interpretable decision tree and hidden Markov model for power load decomposition, comprising the following steps: S1: construction of equipment feature library; S2: training and application of decision tree pruning model; S3: construction of hidden Markov model; S4: search for optimal state combination and power decomposition; S5: result output and visualization.
[0006] Preferably, the construction of the S1 device feature library specifically includes: 1) data acquisition and grouping: collecting time-series data of steady-state electrical parameters of each electrical device under different operating states; 2) feature calculation: for each group of data, calculating the mean of its eight key steady-state features ( ) and standard deviation ( ); 3) Complete the feature library formation: Compile the calculated feature mean and standard deviation of all devices into a table to form the device feature library.
[0007] Preferably, the S2 decision tree pruning model training and application specifically includes: 1) Model training: training a binary classification decision tree model independently for each device; 2) Threshold setting: determining an optimal pruning threshold for the decision tree model of each device during training; 3) Real-time pruning: inputting the total power data into the decision tree model of each device during real-time decomposition.
[0008] Preferably, the construction of the S3 Hidden Markov Model specifically includes: 1) Calculating state transition probabilities: Based on the historical state sequence data of the device, the probability of switching from one state to another is statistically analyzed to form the state transition probability matrix of the device; 2) Constructing time constraints: Using the state transition probability matrix, time constraints are established.
[0009] Preferably, the optimal state combination search and power decomposition in S4 specifically includes: 1) Constructing a search space: combining the possible device state combinations reduced by decision tree pruning in S2 with the state transition probability constraints provided by HMM in S3 to form a joint search space with controllable scale and conforming to the time sequence; 2) Defining and solving the objective function: in the joint search space, finding an optimal device state combination such that the error between the sum of the power of each device under this combination and the measured total power is minimized, while the state transition probability of this combination is as high as possible; 3) Power allocation: based on the optimal state combination determined by defining and solving the objective function, searching for the typical power values of the corresponding devices in the corresponding states from the device feature library, and allocating them as the decomposed power of each device.
[0010] Preferably, S5 specifically includes: 1) Outputting results: outputting the working status of each device at each time point and the power value obtained by decomposition; 2) Completing the visualization display module: visually displaying the decomposition results in the form of time series curves (power curve graphs), state sequence graphs, etc., and providing indicators such as mean absolute error (MAE) and mean relative error (MRE) to evaluate the decomposition accuracy.
[0011] The preferred formula for calculating typical power values is as follows: in ( )and ( ) represent the active power and reactive power of the device in state , respectively, is the joint state transition probability of the Hidden Markov Model (HMM), and and are the weight coefficients of the objective function.
[0012] Preferably, the data decomposition results are evaluated using two commonly used error metrics: mean squared absolute error (MAE) and mean squared relative error (MRE). The calculation formulas are as follows: in This is the actual power value. To estimate the power value, This represents the number of samples.
[0013] Compared with existing technologies, the beneficial effects of this invention are as follows: 1. This invention achieves accurate decomposition and state identification of mixed loads by constructing an intelligent two-level analysis framework, providing technical support for energy consumption management in the power sector, establishing a complete equipment feature database, recording the power characteristic parameters of various electrical equipment under different operating states, and using a pre-pruned decision tree algorithm to construct an explicit rule set for the feature database to achieve rapid screening of equipment states; 2. This invention enhances the rationality of temporal logic by introducing a hidden Markov model. This model analyzes historical data of equipment states to establish a state transition probability matrix that conforms to actual operating rules. This temporal constraint effectively avoids inverse logic state sequences, making the decomposition results more consistent with actual operating scenarios. Attached Figure Description
[0014] Figure 1. Logic diagram of two-level load decomposition operation using decision tree and hidden Markov model. 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 refer to Figure 1. An embodiment of the present invention is shown in Figure 1, which describes the decomposition of electrical load using a hidden Markov model. The method of this decomposition is as follows.
[0017] S1: Equipment Feature Library Construction: First, data acquisition and grouping are performed: Steady-state electrical parameter time-series data (including current, voltage, active power, reactive power, etc.) of each electrical device under different operating states (such as different speeds, high and low power modes) are collected, and the data are grouped according to the device and its specific operating state; then, feature calculation is performed: For each group of data (i.e., each operating state of each device), the mean of its eight key steady-state features (active power, reactive power, power factor, etc.) is calculated. ) and standard deviation ( Finally, the feature library is formed: the calculated feature mean and standard deviation of all devices are compiled into a table to form the device feature library. This library records the "load fingerprint" of each device under different states, which serves as the benchmark for subsequent decomposition.
[0018] S2: Decision Tree Pruning Model Training and Application: First, model training is performed: a binary classification decision tree model is independently trained for each device. This model learns to predict the probability that the device is in an "on" state at a specific time based on the total power data. Second, threshold setting is performed: during training, an optimal pruning threshold is determined for the decision tree model of each device. Finally, real-time pruning is performed: during real-time decomposition, the total power data is input into the decision tree model of each device. If the model predicts that the probability of a device being on is lower than its threshold, the pruning will be performed. If the device is forced to be "off" at that time, a large number of inactive devices can be quickly eliminated, significantly reducing the space of possible state combinations to be searched.
[0019] S3: Hidden Markov Model Construction: On the one hand, it is necessary to calculate the state transition probability: based on the historical state sequence data of the device, the probability of switching from one state to another is statistically analyzed to form the state transition probability matrix of the device. This matrix quantifies the law of evolution of the device state over time. On the other hand, it is necessary to construct time constraints: using the state transition probability matrix, time constraints are established. For the prediction of the device state at the current moment, its probability is significantly affected by the state at the previous moment. The model tends to select the state with the highest transition probability compared to the previous moment, thereby avoiding unreasonable and frequent state jumps and achieving temporal smoothness.
[0020] S4: Optimal State Combination Search and Power Decomposition: First, construct the search space: Combine the possible device state combinations reduced by decision tree pruning in S2 with the state transition probability constraints provided by HMM in S3 to form a joint search space with controllable scale and conforming to temporal patterns; Second, define and solve the objective function: In the joint search space, find an optimal device state combination that minimizes the error between the sum of the power of each device under this combination and the measured total power, while maximizing the state transition probability of this combination. The objective function is usually set as the weighted sum of power matching error and state transition probability. Solve this function through an optimization algorithm to obtain the optimal device state combination at the current moment; Third, power allocation: Based on the optimal state combination determined in step S4.2, find the typical power values (such as feature mean) of the corresponding devices in the corresponding states from the device feature library and allocate them as the decomposed power of each device.
[0021] (1) where ( )and ( ) represent the equipment In state Active power and reactive power at time It is the joint state transition probability of a Hidden Markov Model (HMM), while and These are the weighting coefficients of the objective function.
[0022] S5: Result Output and Visualization: The final output is as follows: output the working status of each device at each time point and the power value obtained by decomposition; finally, complete the visualization display module: visualize the decomposition results in the form of time series curves (power curve graphs), state sequence graphs, etc., and provide indicators such as mean absolute error (MAE) and mean relative error (MRE) to evaluate the decomposition accuracy. This algorithm is used to decompose the total power consumption data, obtain the energy consumption data and operating status of each device, and then use two commonly used error indicators, mean absolute error (MAE) and mean relative error (MRE), to evaluate the data decomposition results.
[0023] The calculation formula is as follows: (2) (3) Among them This is the actual power value. To estimate the power value, This represents the number of samples.
[0024] As a preferred embodiment: the starting module "Device Data Acquisition" is the starting point of the process, responsible for inputting the original power consumption data. Next, the "Device Feature Library Construction" module plays a key technical foundation role. Its task is to establish a "feature fingerprint" database for each device based on steady-state characteristics (such as active power and reactive power) containing different operating states (such as on, off, and different speeds). The completeness and accuracy of this database directly determine the reliability of subsequent decomposition steps.
[0025] The diamond-shaped decision node in the decision tree pruning module is the core of the entire process for achieving "spatial pruning" and reducing computational complexity. In this stage, the system calculates an "on" probability for each device in real time, and the process splits here: if the on probability of a device meets the preset threshold condition, it is determined that the device may be in an active state, and its state is "preserved" and enters the subsequent search space; if the condition is not met, it is determined that the device is in a "off" state, and its state will be "reduced" from the current possible combinations. This filtering mechanism greatly reduces the number of device state combinations that need to be traversed, solves the problem of the "curse of dimensionality" of state combinations in multi-device scenarios, and improves the real-time performance of the algorithm.
[0026] The possible state combinations after pruning will be input into the "Optimal State Combination Search Module" along with the timing constraint information provided by the "HMM State Transition Probability Calculation Module". The role of the HMM module is to use the continuity of device states in time and the state transition probability matrix to avoid unrealistic and frequent state jumps (such as repeated switching of the device in an instant), thereby adding smoothness and rationality in the time dimension to the decomposition results. Finally, under the dual constraints of pruning and HMM, the search module will find an optimal combination of device states that minimizes the error between the theoretical total power and the actual measured total power of the combination.
[0027] The process culminates in the "Decomposition Result Output and Visualization Module," which outputs the optimal solution obtained from the search, namely the status and power value of each device at each moment, and displays it in an intuitive form such as charts, providing users with clear energy consumption data for each device.
[0028] 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 interpretable decision tree and hidden Markov model for power load decomposition, characterized in that: The process includes the following steps: S1: Construction of the device feature library; S2: Training and application of the decision tree pruning model; S3: Construction of the hidden Markov model; S4: Search for the optimal state combination and power decomposition; S5: Output and visualization of results.
2. The power load decomposition using interpretable decision trees and hidden Markov models according to claim 1, characterized in that: The construction of the S1 device feature library specifically includes: 1) data acquisition and grouping: collecting time-series data of steady-state electrical parameters of each electrical device under different operating conditions; 2) feature calculation: for each group of data, calculating the mean of its eight key steady-state features ( ) and standard deviation ( ); 3) Complete the feature library formation: Compile the calculated feature mean and standard deviation of all equipment into a table to form the equipment feature library.
3. The power load decomposition using interpretable decision trees and hidden Markov models according to claim 2, characterized in that: The S2 decision tree pruning model training and application specifically includes: 1) Model training: Training a binary classification decision tree model independently for each device; 2) Threshold setting: Determining an optimal pruning threshold for the decision tree model of each device during training; 3) Real-time pruning: Inputting total power data into the decision tree model of each device during real-time decomposition.
4. The power load decomposition using interpretable decision trees and hidden Markov models according to claim 3, characterized in that: The construction of the S3 Hidden Markov Model specifically includes: 1) Calculating state transition probabilities: Based on the historical state sequence data of the device, the probability of switching from one state to another is statistically analyzed to form the state transition probability matrix of the device; 2) Constructing time constraints: Using the state transition probability matrix, time constraints are established.
5. The power load decomposition using interpretable decision trees and hidden Markov models according to claim 4, characterized in that: The S4 optimal state combination search and power decomposition specifically includes: 1) Constructing the search space: Combining the possible device state combinations reduced by the decision tree pruning in S2 with the state transition probability constraints provided by the HMM in S3 to form a joint search space with controllable scale and conforming to the time sequence; 2) Defining and solving the objective function: In the joint search space, finding an optimal device state combination that minimizes the error between the sum of the power of each device under this combination and the measured total power, while maximizing the state transition probability of this combination; 3) Power allocation: Based on the optimal state combination determined by defining and solving the objective function, searching for the typical power values of the corresponding devices in the corresponding states from the device feature library, and allocating them as the decomposed power of each device.
6. The power load decomposition using interpretable decision trees and hidden Markov models according to claim 5, characterized in that: S5 specifically includes: 1) Outputting results: Outputting the working status of each device at each time point and the power value obtained by decomposition; 2) Completing the visualization display module: Visualizing the decomposition results in the form of time series curves (power curves), state sequence diagrams, etc., and providing indicators such as mean absolute error (MAE) and mean relative error (MRE) to evaluate the decomposition accuracy.
7. The power load decomposition using interpretable decision trees and hidden Markov models according to claim 5, characterized in that: The typical power value calculation formula is as follows: in （ )and （ ) represent the equipment In state Active power and reactive power at time It is the joint state transition probability of a Hidden Markov Model (HMM), while and These are the weighting coefficients of the objective function.
8. The power load decomposition using interpretable decision trees and hidden Markov models according to claim 6, characterized in that: The mean squared absolute error (MAE) and mean squared relative error (MRE), two commonly used error metrics, are used to evaluate the data decomposition results. Their calculation formulas are as follows: in This is the actual power value. To estimate the power value, This represents the number of samples.