Non-intrusive load decomposition method based on multi-dimensional low-frequency characteristics and accelerated Viterbi algorithm

Abstract

The invention discloses a non-intrusive load decomposition method based on multi-dimensional low-frequency characteristics and an acceleration Viterbi algorithm. The method comprises the following steps: acquiring total power information under a multi-load situation; inputting the total power information into the load decomposition model to obtain a load combination state sequence, and obtaining the working state of each load through the load combination state sequence; wherein the load decomposition model is constructed by a factor hidden Markov model (FHMM), the factor hidden Markov model (FHMM) is constructed based on a multi-dimensional hidden Markov model (HMM), and the FHMM is optimized by using a Viterbi optimization method. According to the method, the load state space and the state transition path are effectively constrained, so that the calculation efficiency can be improved under the condition of ensuring that the load decomposition precision is not changed.

Description

Technical Field

[0001] The present invention relates to the technical field of non-intrusive load monitoring, and in particular to a non-intrusive load decomposition method based on multi-dimensional low-frequency features and an accelerated Viterbi algorithm. Background Art

[0002] As an important part of the advanced measurement system of smart grid, non-intrusive load monitoring (NILM) technology is of great significance to improving energy efficiency. Since Hidden Markov Model (HMM) has the ability to model power data and characterize load status, it is widely used in the field of NILM.

[0003] At present, power load monitoring is mainly divided into two methods: intrusive load monitoring (ILM) and non-intrusive load monitoring (NILM). ILM technology realizes real-time monitoring of electricity consumption by deploying sensors on each load device, which requires a large amount of money to purchase and maintain hardware. This technology has poor versatility. The sensor device needs to be installed at home and it is difficult to adapt to the physical layout changes of the load equipment, so it cannot be promoted and applied on a large scale. NILM technology obtains real-time energy consumption information of each power load within the power user by analyzing and processing the total load meter data. Due to the privacy of the home environment, NILM technology can protect user privacy while saving sensor hardware costs. With the continuous development of NILM research, machine learning methods have been introduced to improve the accuracy, real-time and versatility of monitoring.

[0004] Depending on whether it is necessary to identify load switching events, NILM can be divided into two different types of methods: event-based and non-event-based. The event-based NILM method focuses on the differentiated features in the signal components of a single load device and the event features extracted from the differential signal in the observed total metering data, and realizes the monitoring process by identifying and matching the differentiated features and event features. The non-event-based NILM method mostly considers the mapping relationship between the signal features of a single load device and the observed total metering data, so as to directly use the observed total metering data to obtain the signal components of a single load device. Since the event-based NILM method converts the NILM problem into a classification problem, it is easy to introduce errors in the intermediate links. It needs to rely on high-frequency differentiated features and load event labels to ensure the accuracy of event detection and the accuracy of load event matching, which is limited in practical application scenarios with low sampling rates. Although the non-event-based NILM method omits the event detection link, it needs to consider the relationship between the solution efficiency and the number of loads in the monitored range during operation.

[0005] The current electric meters can easily obtain the steady-state characteristics of the load at low frequency. Using the easily obtained steady-state characteristics to improve the recognition rate of the load will be beneficial to the popularization of NILM technology. Among the non-event methods, the hidden Markov model (HMM) has a good ability to handle time series problems. In addition, since the model characterizes the load state through observations, it requires less training time and data, and has received widespread attention in the NILM field. However, the HMM-based methods mostly use active power characteristics to construct load models, which makes it difficult to effectively distinguish power-approximate loads. In addition, the main problem with using HMM for load decomposition is that the number of states and the amount of calculations processed by the model increase exponentially with the increase in the number of devices and the number of states of each device, resulting in low computational efficiency. Summary of the invention

[0006] In order to improve the decomposition accuracy of power-similar devices and realize efficient calculation of load decomposition, the present invention proposes a non-intrusive load decomposition method based on multi-dimensional low-frequency features and accelerated Viterbi algorithm.

[0007] To achieve the above object, the present invention provides a non-intrusive load decomposition method based on multi-dimensional low-frequency features and accelerated Viterbi algorithm, comprising:

[0008] Obtain total power information under multiple load scenarios;

[0009] Inputting the total power information into a load decomposition model to obtain a load combination state sequence, and obtaining the working state of each load through the load combination state sequence;

[0010] The load decomposition model is constructed by a factor hidden Markov FHMM model, which is constructed based on a multidimensional hidden Markov HMM model, and the FHMM model is optimized using a Viterbi optimization method.

[0011] Preferably, the method for constructing the factor hidden Markov model FHMM includes:

[0012] Determine the dynamic change of energy consumption of load equipment, obtain auxiliary variables for load model parameter iteration through forward probability and backward probability, and iterate and solve through Baum-Welch method to obtain the HMM model parameter update expression;

[0013] The FHMM model is constructed based on the HMM model parameter update expression.

[0014] Preferably, determining the dynamic change of energy consumption of the load device includes:

[0015] The dynamic change of energy consumption of the load equipment is determined according to the unobservable state sequence and the observable observation sequence, characterized by the Markov homogeneity and observation independence assumptions, and the emission probability distribution is represented by a multidimensional Gaussian model; wherein the unobservable state sequence includes the standby state, cooling state and heating state of the air conditioner, and the observable observation sequence is the power sequence generated by the load in different working states.

[0016] Preferably, the multidimensional Gaussian model is used to represent the emission probability distribution as:

[0017]

[0018] In the formula, is the expected vector under state j; is the covariance matrix under state j; b j (O) is the emission probability matrix.

[0019] Preferably, the method for obtaining the auxiliary variables for load model parameter iteration is:

[0020]

[0021] In the formula, γ t (i) is the probability that the load is in state i at time t; ξ t (i, j) is the probability that the load is in state i at time t and in state j at time t+1; α t (i) The observation sequence at time t is O 1:t And the forward probability of state i; β t (i) is in state i at time t, and the observation sequence from time t to time T is O t+1:T The backward probability of t (j) is the observation sequence at time t is O 1:t And the forward probability of state j; β t (j) is in state j at time t, and the observation sequence from time t to time T is O t+1:T The backward probability of ij is the transition probability of the load switching from state i to state j, b j (O t+1 ) is the load in state j in power consumption information O t+1 The observation probability of t+1 (j) is in state j at time t+1, and the observation sequence from time t+2 to time T is O t+2:T The backward probability of; N is the number of load states.

[0022] Preferably, the HMM model parameter update expression includes:

[0023] π i =γ1(i),

[0024]

[0025] In the formula, π i is the probability that the load is in state i at the initial moment, γ1(i) is the probability that the load is in state i at the initial moment, T is the total time step, t is the current moment, μ i is the expected vector in state i, O t is the observation vector at time t, Σ i is the covariance matrix under state i, and μ is the expected vector.

[0026] Preferably, the parameter set of the FHMM model is defined as λ=(π, A, μ, Σ), wherein parameters π and A are obtained by performing Kronecker product operation on corresponding parameters in each load model, and parameters μ and Σ are obtained by accumulating active power and reactive power components of corresponding parameters of each load model.

[0027] Preferably, the Viterbi algorithm is optimized, including:

[0028] The load combination state sequence Q is inferred from the total active power and reactive power sequence 1:T , find an optimal path with the highest probability among all possible load combination state sequences, backtrack the optimal path at the final moment to obtain the load combination state sequence, and introduce the state space constraint coefficient and the state transfer path constraint coefficient to optimize the Viterbi method;

[0029] Among them, the state space and state transition path of the Viterbi algorithm are constrained based on the operating state of the load, so that the power level in the new state space meets the user's historical maximum power or maximum load power requirements. When the power change is a positive value, for the calculation of the load combination state i at time t+1, only the case where the load combination state index at time t is less than or equal to i+βM′ is considered; when the change is a negative value, for the calculation of the load combination state i at time t+1, only the case where the load combination state index value at time t is greater than or equal to i-βM′ is considered, where β is the state transition path constraint coefficient, and M′ is the new state space size.

[0030] Preferably, after obtaining the load combination state sequence, the working state of each load is obtained by a combination state decomposition formula; wherein the state decomposition and historical average active power mapping are:

[0031]

[0032] In the formula, is the state index of the i-th load at time t, is the floor operator, mod is the remainder operator, Q t is the combined state index value at time t, K j is the number of states of the jth load, K i is the number of states of the i-th load, and N is the number of loads.

[0033] Preferably, the method further comprises constructing a power decomposition model based on convex quadratic programming according to the constraint condition that the sum of the power of each load is approximately equal to the total power. The specific energy consumption allocation process is:

[0034]

[0035] In the formula, is the state index of the i-th load at time t, is the floor operator, P i (t) is the energy consumption of the ith load, and P(t) is the total energy consumption.

[0036] Compared with the prior art, the present invention has the following advantages and technical effects:

[0037] The present invention extends the observation perspective of the traditional HMM model to the active-reactive plane, thereby improving the distinguishability of power approximation devices on the basis of increasing the load feature dimension; through effective constraints on the load state space and state transition path, it can achieve improved computational efficiency while ensuring that the load decomposition accuracy remains unchanged. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0039] Figure 1 This is a flow chart of a non-intrusive load decomposition method based on multi-dimensional low-frequency features and accelerated Viterbi algorithm according to an embodiment of the present invention;

[0040] Figure 2 A state transition topology diagram of a finite state machine (FSM) according to an embodiment of the present invention;

[0041] Figure 3 It is a basic structure diagram of HMM in an embodiment of the present invention;

[0042] Figure 4 This is a diagram of the FHMM structure integrating multiple electrical features according to an embodiment of the present invention;

[0043] Figure 5 It is a schematic diagram of partial operation of three types of load equipment according to an embodiment of the present invention;

[0044] Figure 6 Schematic diagram of state space constraints and state transition path constraints according to an embodiment of the present invention. DETAILED DESCRIPTION

[0045] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0046] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0047] In order to improve the decomposition accuracy of power-similar devices and realize efficient calculation of load decomposition, the present invention extends the observation perspective of the traditional HMM model to the active-reactive plane, and improves the distinguishability of power-similar devices on the basis of increasing the load feature dimension. At the same time, in the quantification of load state conversion information, the effective value of the load current is additionally introduced to achieve effective complementarity of load knowledge information. In the algorithm flow, the HMM related parameters of a single load are first learned through multi-dimensional electrical steady-state characteristics, and the relevant information of load state changes is simultaneously characterized from multiple feature perspectives. Then, the relevant parameters of FHMM are modeled by Kronecker product to characterize the combined operating state of multiple loads. When inferring the combined load state sequence, the active power and reactive power that meet the additivity are used as the observation dimension, and on this basis, a two-stage optimization accelerated Viterbi algorithm is proposed. Finally, by decomposing the combined load state sequence, the working state of each electrical appliance can be obtained, and the active power under the corresponding state can be obtained.

[0048] A non-intrusive load decomposition method based on multi-dimensional low-frequency features and accelerated Viterbi algorithm, such as Figure 1 , including:

[0049] A factor hidden Markov FHMM model is constructed based on a multidimensional hidden Markov HMM model, and the FHMM model is optimized based on a Viterbi optimization method to obtain a load combination state sequence, and the working state of each load is obtained through the load combination state sequence.

[0050] The load model is a quantitative description and reflection of the power consumption behavior and characteristics of the equipment. Figure 2As shown in Figure 1, the hidden Markov model can effectively represent the state transition information of the load and the corresponding load characteristic relationship, so it is widely used in the representation of the load model topology. Among the many variants of finite state machines, the hidden Markov model has the ability to handle load time series data problems and can accurately reflect the statistical laws of the mutual conversion between different working states of the load during operation. Therefore, this model has received widespread attention in load modeling and load decomposition problems.

[0051] When using HMM to construct a load model, the dynamic changes in energy consumption of load equipment are represented by two interrelated random sequences (such as Figure 3 ), that is, the unobservable state sequence q 1:T and the observable observation sequence O 1:T The state sequence q 1:T Refers to the working state of the load within the time range T, such as the standby state, cooling state and heating state of the air conditioner; the observation sequence O 1:T Refers to the power sequence generated by the load in different working states. According to the Markov homogeneity and observation independence assumptions, the HMM with a total number of K states can be characterized by the parameter λ = (π, A, B). is the initial state probability vector, which represents the state probability distribution of the load at the initial moment; The state transition matrix is ​​composed of the transition probability a representing the load state i switching to the load state j ij composition; For element b j (O) constitutes the emission probability matrix, b j (O) represents the probability that the load is at power level O in state j, and M is the number of power levels. Since the power corresponding to the load state is a continuously changing random variable rather than a fixed value, the Gaussian model is often used to represent the transmission probability distribution.

[0052] When using HMM models for load modeling, we often face the situation of partial information missing, such as only the load power sequence but lack of load state transition information. Therefore, HMM and its derivative models usually need to deal with the following three basic issues when applied:

[0053] (1) Learning problem: Given a known observation sequence O 1:T In the case of , the model parameter λ is estimated. That is, the parameter information used to characterize the load state and the mapping relationship between the state and the load energy consumption is estimated through the acquired load sequence data.

[0054] (2) Evaluation problem: Given the observation sequence O 1:T and model parameters λ, calculate the probability of the observation sequence occurring when the model parameters are given, and use this to evaluate the model parameters that best match the observation sequence.

[0055] (3) Prediction problem: Given the observation sequence O 1:T and model parameter λ, determine the corresponding state sequence q 1:T When decomposing the load, the load state sequence is inferred through the total power sequence and the load model parameters, and then the mapping relationship between the load state and the load energy consumption is used to further obtain the specific energy consumption of each load device.

[0056] For an HMM model with multiple observation dimensions, its state sequence corresponds to multiple observation sequences at the same time. In this embodiment, taking the three characteristics of active power, reactive power and current as examples, the working state of a load device can simultaneously reflect three types of load information, that is, the observation sequence is composed of observation vector O1: T =[O P ,O Q ,O I ]1: T Indicates that the emission probability distribution of HMM can be represented by a multidimensional Gaussian model:

[0057]

[0058] in, is the expected vector under state j; is the covariance matrix under state j. At this time, the parameters of the load model can be expressed as λ = (π, A, μ, Σ). Multidimensional electrical characteristics can simultaneously characterize the changes in load states from different dimensions, so that parameters π and A can better reflect the statistical laws of mutual conversion between different working states.

[0059] In order to solve the learning problem of the HMM model, that is, to train the load model parameters from a known observation sequence, we need to use the forward probability α t (i) and the backward probability β t (i) Calculation. Given the initial parameters λ of the load model, the forward probability α t (i) indicates that the observation sequence at time t is O1,…,O t And the probability of load state i; backward probability β t (i) indicates that under the condition that the load state at time t is i, the partial observation sequence from time t+1 to T is O t+1 ,…,O T The probability of. Under the initial condition α1(i)=π i p(O1|q1=i), β T When (i) = 1, α t (i) and β t The recursive calculation formula for (i) is as follows:

[0060]

[0061]

[0062] Furthermore, through the forward probability α t (i) and the backward probability β t (i) Two auxiliary variables γ can be obtained for load model parameter iteration t (i) and t (i,j). γ t (i) represents the probability that the load is in state i at time t, ξ t (i, j) represents the probability that the load is in state i at time t and in state j at time t+1, and the expression is as follows:

[0063]

[0064] Finally, the Baum-Welch method is used to iteratively solve the problem, and the following HMM model parameter update expression is obtained without state sequence data. The initial values ​​of μ and Σ are determined by the K-means clustering algorithm, and the number of clusters is consistent with the number of load states; the initial values ​​of π and A are randomly selected.

[0065] π i =γ1(i),

[0066]

[0067] Since load decomposition needs to consider multiple electrical appliances at the same time, the Factorial Hidden Markov Model (FHMM) is used to represent the topological structure containing multiple independent load state sequences in the same time period, such as Figure 4 In the FHMM that integrates multiple electrical features, the load observation value at each moment is no longer the power of a single electrical appliance, but the power superposition of multiple electrical appliances under different load conditions. Since the effective value of the load current does not meet the additivity, in this embodiment, active power and reactive power that meet the additivity are selected as the observation values.

[0068] Similarly, the parameter set of FHMM can also be defined as λ = (π, A, μ, Σ), where the parameters π and A can be obtained by performing Kronecker product operations on the corresponding parameters in each load model, and the parameters μ and Σ are obtained by accumulating the active power and reactive power components of the corresponding parameters of each load model. From the perspective of observable dimensions, although the method proposed in the present invention only includes active power and reactive power, current characteristics are added when constructing a single load model, and the current when the load is running can be embedded in the parameters π and A to achieve effective characterization of load state changes.

[0069] After constructing the load model and obtaining the FHMM model parameters, in the load decomposition stage, the load combination state sequence Q must first be inferred from the total active power and reactive power sequences. 1:T .Q t is the combined state index value at time t, which is related to the working state index of each load The relationship is expressed by the following formula:

[0070]

[0071] Where: is the state index of the nth load, K i is the number of states of the i-th load, and N is the number of loads.

[0072] When the total number of load states is M, the complexity of using the traversal solution method to solve the above prediction problem is relatively high. Therefore, the Viterbi algorithm can usually be used to solve the combined state sequence Q 1:T , that is, using the idea of ​​dynamic programming to retain the optimal path for each node To obtain the optimal path Two auxiliary variables δ are needed t (i) and Ψ t (i) Calculate. t (i) represents the maximum probability among all single paths with state i at time t, Ψ t (i) represents the t-1th node of the path with the highest probability among all single paths with state i at time t. Under the initial condition δ1(i) = π i b i When (O1) and Ψ1(i) = 0, use the following formula recursively to find:

[0073]

[0074] At the final moment T, the optimal path By backtracking, we can get the load combination state sequence It can be seen that when the power sequence length is T and the total number of load states is M, the Viterbi algorithm needs to compare M at each time step. 2 times to obtain the optimal state of each time step at the previous moment, so the algorithm complexity is O(TM 2 ). Since the time complexity of the Viterbi algorithm is proportional to the square of the total number of load states, it is slow to process long power sequences containing multiple electrical appliances, making it difficult to meet the needs of real-time load decomposition. In the load decomposition task, the operating state of the load implies the numerical value of the power level, which can be used to constrain the state space and state transition path of the Viterbi algorithm to reduce the complexity of the algorithm.

[0075] The upper limit of household load power is limited by the cross-sectional size of the current-carrying wire, which makes it difficult for all devices to operate at maximum power at the same time. In most cases, users will use different power loads at different times. Figure 5 Taking the operating time of three high-power electrical appliances as an example, it is rare that the three electrical appliances are turned on at the same time during this time period.

[0076] The HMM-based load decomposition method independently models the power level of each load. When constructing the state space of load combinations, some combination states that do not exist in actual power consumption scenarios may be introduced. In order to make full use of the power value to constrain the state space, the load combination states can be arranged in ascending order according to the power value of the load combination state. Then, the state space constraint coefficient α (0<α<1) is introduced so that the power level in the new state space (M′=αM) meets the user's historical maximum power or maximum load power requirements. After the first stage of optimization, the complexity of the Viterbi algorithm is reduced from O(TM 2 ) becomes O(α 2 TM 2 ).

[0077] Since the load combination state transition is accompanied by a change in total power, the state transition path can be further constrained by the total power change. Since the load combination states have been arranged in ascending order according to the power value in advance, when the power change is a positive value, for the load combination state i at time t+1, only the load combination state index at time t is less than or equal to i needs to be considered. At the same time, in order to avoid the influence of metering errors on load state prediction, the state transition path constraint coefficient β (0<β<1) is introduced to further relax the constraints of the second stage optimization, that is:

[0078]

[0079] Similarly, when the power change is a negative value, for the calculation of the load combination state i at time t+1, it is only necessary to consider the case where the load combination state index value at time t is greater than or equal to i-βM, that is:

[0080]

[0081] Since the total load power is in a fluctuating state most of the time, after the second stage of optimization, the algorithm complexity is reduced from O(α 2 TM 2 ) becomes O((-β 2 +2β+1)α 2 TM 2 / 2+(1-β)αTM / 2), the state space constraints and state transfer path constraints are shown as follows Figure 6 .

[0082] Through the above Viterbi optimization algorithm, the load combination state sequence Q can be inferred 1:T On this basis, it is necessary to further obtain the working state of each load through the combined state decomposition formula. The state decomposition and historical average active power mapping are expressed by the following formula:

[0083]

[0084] Where: represents the state index of the i-th load at time t, It represents the floor operator and mod is the remainder operator.

[0085] Since the power of the load fluctuates in a certain operating state, approximating the use of historical average active power to characterize the energy consumption information of the load working state may affect the accuracy of sub-item metering. Therefore, this embodiment constructs a power decomposition model based on convex quadratic programming according to the constraint that the sum of the power of each load is approximately equal to the total power. The specific energy consumption allocation process is as follows:

[0086]

[0087] This embodiment further illustrates the technical solution by setting experimental data:

[0088] The experimental data is selected from the public dataset AMPds in the NILM field, which collects electricity consumption information of a Canadian family for one year. The sampling time of this dataset is from April 2012 to April 2013, with a sampling frequency of 1 minute. The electrical parameters include 11 electrical parameters such as voltage, current and power. The long-term data span is of great value for studying household electricity consumption patterns and NILM algorithms. During the experiment, we mainly focus on six types of load equipment that account for a relatively high proportion of the overall energy consumption, namely heat pumps (HPE), refrigerators (FGE), dryers (CDE), fireplaces (WOE), dishwashers (DWE) and washing machines (CWE). These six types of load equipment cover different power levels and have quite different working methods.

[0089] In the experiment, six types of loads, including clothes dryers (CDE), washing machines (CWE), dishwashers (DWE), refrigerators (FGE), fireplaces (WOE) and heat pumps (HPE) in the AMPds dataset were used as research objects, and the load states were set to 3, 5, 4, 2, 4, and 3 according to the power level of the load curve. A total of 20,160 data samples were used for model training during the experimental period. During the experimental period, the normalized decomposition error (NDE) and F1 score were used to quantify and compare the performance of each load decomposition method.

[0090] In order to verify the performance of the method, three load decomposition algorithms, namely exact factor hidden Markov model (FHMMExact), factor hidden semi-Markov model (FHSMM) and piecewise integer quadratic programming (SIQCP), were selected as FSM-based comparison algorithms. In addition, WindowGRU, RNN and Seq2Seq were selected as neural network load decomposition methods with strong nonlinear expression capabilities.

[0091] The computational efficiency and decomposition effect of the proposed load decomposition algorithm are investigated through four case studies. Case 1 analyzes the impact of load state space and state transition path constraints on computational time and decomposition accuracy. Case 2 analyzes the performance comparison between the proposed method and the load decomposition algorithm based on the finite state machine class. Case 3 analyzes the performance comparison between the proposed method and the load decomposition algorithm based on the neural network, with special attention paid to the accuracy of the load decomposition results for power approximate devices. Case 4 analyzes the impact of the training sample size on the performance of the proposed method.

[0092] To quantify the performance of load decomposition, the normalized decomposition error (NDE) and F1 score are used to quantify and compare the performance of each load decomposition method. NDE provides a direct measure of the algorithm's ability to decompose the power curve, defined as:

[0093]

[0094] Where: and p(t) are the predicted power value and the true value at time t, respectively. From the perspective of energy consumption division, the recall rate Recall and the precision rate Precision reflect the energy consumption part that is correctly classified and the energy consumption part that is correctly divided to the corresponding equipment, respectively. The F1 score is the harmonic mean of the precision rate and the recall rate. The calculation method is defined as follows:

[0095]

[0096] The present invention can improve the computing efficiency while ensuring that the load decomposition accuracy remains unchanged by effectively constraining the load state space and the state transfer path.

[0097] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A non-intrusive load decomposition method based on multi-dimensional low-frequency features and accelerated Viterbi algorithm, characterized in that: include: Obtain total power information under multiple load scenarios; Inputting the total power information into a load decomposition model to obtain a load combination state sequence, and obtaining the working state of each load through the load combination state sequence; The load decomposition model is constructed by a factor hidden Markov FHMM model, which is constructed based on a multidimensional hidden Markov HMM model, and the FHMM model is optimized using a Viterbi optimization method.

2. The non-intrusive load decomposition method based on multi-dimensional low-frequency features and accelerated Viterbi algorithm according to claim 1 is characterized in that: The method for constructing the factor hidden Markov model includes: Determine the dynamic change of energy consumption of load equipment, obtain auxiliary variables for load model parameter iteration through forward probability and backward probability, and iterate and solve through Baum-Welch method to obtain the HMM model parameter update expression; The FHMM model is constructed based on the HMM model parameter update expression.

3. The non-intrusive load decomposition method based on multi-dimensional low-frequency features and accelerated Viterbi algorithm according to claim 2 is characterized in that: Determining the dynamic change of energy consumption of the load equipment includes: The dynamic change of energy consumption of the load equipment is determined according to the unobservable state sequence and the observable observation sequence, characterized by the Markov homogeneity and observation independence assumptions, and the emission probability distribution is represented by a multidimensional Gaussian model; wherein the unobservable state sequence includes the standby state, cooling state and heating state of the air conditioner, and the observable observation sequence is the power sequence generated by the load in different working states.

4. The non-intrusive load decomposition method based on multi-dimensional low-frequency features and accelerated Viterbi algorithm according to claim 3 is characterized in that: The multidimensional Gaussian model is used to represent the emission probability distribution as: In the formula, is the expected vector under state j; is the covariance matrix under state j; b j (O) is the emission probability matrix.

5. The non-intrusive load decomposition method based on multi-dimensional low-frequency features and accelerated Viterbi algorithm according to claim 2 is characterized in that: The method for obtaining the auxiliary variables used for load model parameter iteration is: In the formula, γ t (i) is the probability that the load is in state i at time t; ξ t (i, j) is the probability that the load is in state i at time t and in state j at time t+1; α t (i) The observation sequence at time t is O 1:t And the forward probability of state i; β t (i) is in state i at time t, and the observation sequence from time t to time T is O t+1:T The backward probability of t (j) is the observation sequence at time t is O 1:t and the forward probability of state j; βt(j) is the state j at time t, and the observation sequence from time t to time T is O t+1:T aij is the backward probability of the load switching from state i to state j, b j (O t+1 ) is the load in state j in power consumption information O t+1 The observation probability of t+1 (j) is in state j at time t+1, and the observation sequence from time t+2 to time T is O t+2:T The backward probability of; N is the number of load states.

6. The non-intrusive load decomposition method based on multi-dimensional low-frequency features and accelerated Viterbi algorithm according to claim 5, characterized in that: The HMM model parameter update expression includes: p i =γ1(i), In the formula, π i is the probability that the load is in state i at the initial moment, γ1(i) is the probability that the load is in state i at the initial moment, T is the total time step, t is the current moment, μ i is the expected vector in state i, O t is the observation vector at time t, Σ i is the covariance matrix under state i, and μ is the expected vector.

7. The non-intrusive load decomposition method based on multi-dimensional low-frequency features and accelerated Viterbi algorithm according to claim 6, characterized in that: The parameter set of the FHMM model is defined as λ=(π, A, μ, Σ), wherein the parameters π and A are obtained by performing Kronecker product operations on the corresponding parameters in each load model, and the parameters μ and Σ are obtained by accumulating the active power and reactive power components of the corresponding parameters of each load model.

8. The non-intrusive load decomposition method based on multi-dimensional low-frequency features and accelerated Viterbi algorithm according to claim 1, characterized in that: The Viterbi optimization method comprises: The load combination state sequence Q is inferred from the total active power and reactive power sequence 1:T , find an optimal path with the highest probability among all possible load combination state sequences, backtrack the optimal path at the final moment to obtain the load combination state sequence, and introduce the state space constraint coefficient and the state transfer path constraint coefficient to optimize the Viterbi method; Among them, the state space and state transition path of the Viterbi algorithm are constrained based on the operating state of the load, so that the power level in the new state space meets the user's historical maximum power or maximum load power requirements. When the power change is a positive value, for the calculation of the load combination state i at time t+1, only the case where the load combination state index at time t is less than or equal to i+βM′ is considered; when the change is a negative value, for the calculation of the load combination state i at time t+1, only the case where the load combination state index value at time t is greater than or equal to i-βM′ is considered, β is the state transition path constraint coefficient, and M′ is the new state space size.

9. The non-intrusive load decomposition method based on multi-dimensional low-frequency features and accelerated Viterbi algorithm according to claim 8, characterized in that: After obtaining the load combination state sequence, the working state of each load is obtained through the combination state decomposition formula; wherein the state decomposition and historical average active power mapping are: In the formula, is the state index of the i-th load at time t, is the floor operator, mod is the remainder operator, Q t is the combined state index value at time t, K j is the number of states of the jth load, K i is the number of states of the i-th load, and N is the number of loads.

10. The non-intrusive load decomposition method based on multi-dimensional low-frequency features and accelerated Viterbi algorithm according to claim 1, characterized in that: The method further includes constructing a power decomposition model based on convex quadratic programming according to the constraint condition that the sum of the power of each load is approximately equal to the total power. The specific energy consumption allocation process is: In the formula, is the state index of the i-th load at time t, is the floor operator, P i (t) is the energy consumption of the ith load, and P(t) is the total energy consumption.

Images