Non-intrusive load decomposition method based on hidden Markov model

Abstract

The invention relates to a non-intrusive load decomposition method based on a hidden Markov model, and the non-intrusive load decomposition method based on the hidden Markov model is characterized in that the hidden Markov model is expanded into a factor hidden Markov model which is applied to non-intrusive load monitoring; wherein the task of non-intrusive load monitoring corresponds to a parameter learning problem and a state inference problem in the hidden Markov model; a public data set is used for experimenting a load decomposition method based on a hidden Markov model to confirm a decomposition effect, the hidden Markov model (HMM) is used for solving a non-intrusive load decomposition problem, a Viterbi algorithm is adopted for step-by-step solving, the Viterbi algorithm searches for an optimal state sequence by traversing all state transition conditions, and a non-intrusive load decomposition problem is solved. And applying dynamic programming to efficiently solve the state sequence with the maximum probability, and speculating the load state condition at the t + 1 moment according to the load condition at the t moment and the total power at the t + 1 moment in the step-by-step decomposition process.

Description

Technical Field

[0001] The invention relates to the technical field of non-intrusive load monitoring in an electric power system, and discloses a non-intrusive load decomposition method based on a hidden Markov model. Background Art

[0002] Non-intrusive load monitoring is of great significance in the management and optimization of power systems. Traditional intrusive load monitoring methods require the installation of sensors on each load device, which is costly and complex to implement. Non-intrusive load monitoring (NILM) is a more economical and convenient way to decompose the working status and energy consumption of each load device by collecting total load data at the power entrance. However, the existing non-intrusive load monitoring methods still have certain challenges in terms of accuracy and efficiency. Non-intrusive load monitoring methods based on statistical states, especially hidden Markov models and their variations, have great application potential, but when they are actually applied to power non-intrusive load decomposition scenarios, further improvements and optimizations are still needed to better meet the needs. Summary of the invention

[0003] In order to solve the technical problems in the above-mentioned background technology, the present invention provides a non-invasive load decomposition method based on the hidden Markov model. The present invention expands the hidden Markov model into a factor hidden Markov model and applies it to non-invasive load monitoring. It clarifies that the task of non-invasive load monitoring corresponds to the parameter learning problem and state inference problem in the hidden Markov model. By establishing this corresponding relationship, a theoretical foundation is laid for accurate load monitoring and decomposition using the hidden Markov model and its deformation.

[0004] A non-intrusive load monitoring and decomposition method based on a hidden Markov model and its deformations, including: a load decomposition method based on a hidden Markov model is experimented with using a public data set to experiment with a load decomposition method based on a hidden Markov model, and a large amount of experimental data is used to verify the decomposition effect of the method in practical applications, thereby providing a basis for subsequent applications in practical scenarios.

[0005] The use of HMM (Hidden Markov Model) to solve the non-intrusive load decomposition problem and the application of Viterbi algorithm In the actual scenario of applying the hidden Markov model to the non-intrusive load decomposition of electricity, HMM is used to solve the non-intrusive load decomposition problem. The Viterbi algorithm is used to solve step by step. The algorithm searches for the optimal state sequence by traversing all state transitions. The Viterbi algorithm uses dynamic programming to efficiently solve the optimal path, that is, the state sequence with the highest probability. In the step-by-step decomposition process, the load state at time t+1 is inferred based on the load state at time t and the total power at time t+1. This process is similar to the mapping from input to output of a deep network.

[0006] Construction and function realization of non-intrusive load decomposition network A non-intrusive load decomposition network is constructed, whose input is the status of each load at time t and the non-intrusive power at time t+1, and the output is the predicted parameter value of each load at time t+1. According to the universal approximation characteristics of the deep network, if there is enough training data, the network can fully learn the operating characteristics of the load, thereby realizing the function of the Viterbi algorithm, providing an effective implementation method for non-intrusive load decomposition.

[0007] In the Hidden Markov Model, the Viterbi algorithm is used to find the most likely sequence of hidden states. The following is an explanation of the general steps of the Viterbi algorithm and the related formulas:

[0008] 1. Symbol Definition

[0009] 1.0 = {o1, o2, ..., or}: observation sequence, where ot represents the observation value at time t;

[0010] 2. S = {s1, s2, ..., sn}: a set of hidden states, where Sn represents the nth hidden state;

[0011] 3. A = {aij}: state transition probability matrix, where aij = P(sj|si) represents the probability of transitioning from state si to state sj;

[0012] 4. B = {bj(ot)}: emission probability matrix, where bj(ot) = P(ot|sj) represents the probability of generating the observation value ot in state sj.

[0013] 5.π={πi}: Initial state probability distribution, where πi=P(si|t=1) represents the probability of being in state s: at the initial moment.

[0014] 6.δt(i): The probability of the maximum probability path to state si at time t.

[0015] 7.δt(i): The index of the previous state on the path with the maximum probability of reaching state s at time t.

[0016] 2. Algorithm steps

[0017] 1. Initialization:

[0018] At time t = 1, for each state si, calculate the initial probability

[0019] δ1(i)=π i s i b i (o1).

[0020] Setting φ1(i) to 0 indicates that the initial state has no previous state.

[0021] 2. Recursive calculation;

[0022] For time t=2 to T.

[0023] For each state sj.

[0024]

[0025] 3. Termination:

[0026] Calculate the probability of the final maximum probability path: 0

[0027]

[0028] Find the index of the final state:

[0029]

[0030] 4. Backtracking:

[0031]

[0032] Finally, q*={q*1,q2,...,q*T} is the most likely hidden state sequence found by the Viterbi algorithm.

[0033] In the non-intrusive load decomposition problem, the observation sequence can be the total load data, and the hidden state sequence can be the state of each load device. By applying the Viterbi algorithm, the most likely state sequence of each load device can be inferred from the total load data, thereby achieving non-intrusive load decomposition.

[0034] The beneficial effects of the present invention are:

[0035] 1. Improving the accuracy of non-intrusive load monitoring

[0036] Advantages of Task Mapping Based on Factor Hidden Markov Model By mapping the non-intrusive load monitoring task to the parameter learning and state inference problem in the factor hidden Markov model, the working state and energy consumption of the load equipment can be modeled more accurately. This mapping method fully considers various complex factors in load monitoring, such as the relationship between different load equipment, the operating characteristics of load equipment under different working conditions, etc., thereby improving the accuracy of the state and parameter estimation of each load equipment.

[0037] Public data set validation guarantees Experiments with the load decomposition method based on the hidden Markov model using public data sets can verify the effectiveness of the method in a variety of practical scenarios. Through testing and optimization of a large number of different types of load data, the method can adapt to different load characteristics and power usage patterns, further improving the decomposition accuracy in actual non-intrusive load monitoring and reducing decomposition errors.

[0038] 2. Improve load decomposition efficiency

[0039] The efficient solution of the Viterbi algorithm has significant efficiency advantages when using the hidden Markov model to solve the non-intrusive load decomposition problem. The Viterbi algorithm efficiently traverses all state transitions through dynamic programming to find the optimal state sequence, avoiding the huge amount of calculation caused by brute force search. This makes it possible to quickly and accurately determine the most likely state of each load device at each moment when processing a large amount of load data and complex state transition relationships, greatly improving the speed of load decomposition and meeting the needs of real-time monitoring and management of power systems.

[0040] The constructed non-intrusive load decomposition network utilizes the universal approximation characteristics of the deep network, and can quickly learn the operating characteristics of the load when there is enough training data, realizing the function of the Viterbi algorithm. This deep network-based implementation method can utilize the parallel computing power of modern computers to further improve computing efficiency, quickly and accurately predict the parameter values ​​of each load, and realize efficient non-intrusive load decomposition.

[0041] 3. Reduce monitoring costs and complexity

[0042] Cost-effectiveness of non-invasive monitoring methods The present invention adopts a non-invasive load monitoring method. Compared with the traditional invasive load monitoring method, there is no need to install sensors on each load device. This not only reduces the number of hardware devices installed and reduces hardware costs, but also avoids the complex operations caused by the installation and maintenance of a large number of sensors, reduces labor costs and system complexity, and improves the economy and maintainability of the entire load monitoring system.

[0043] 4. Promote the Optimal Management of Power Systems

[0044] Accurate load monitoring and decomposition facilitates decision-making. Accurate non-invasive load monitoring and decomposition results provide reliable data support for the operation and management of the power system. Power system operators can optimize power distribution, formulate reasonable electricity price strategies, and conduct effective grid dispatching based on the detailed status and energy consumption of each load device. For example, according to the usage of each load device in different time periods, the power generation plan can be reasonably adjusted to improve the utilization rate of power generation equipment and reduce energy waste, thereby improving the operating efficiency and economic benefits of the entire power system.

[0045] Adapting to the dynamic changes of the power system The method of the present invention can adapt to the dynamic changes of load equipment in the power system. As users continue to add new load equipment or the operating characteristics of existing equipment change, the method can adapt to these changes by continuously updating model parameters or retraining the network, always maintaining accurate monitoring and decomposition capabilities of the load, and helping to maintain the stable operation and sustainable development of the power system. DETAILED DESCRIPTION

[0046] The method based on hidden Markov model and its deformation belongs to the statistical state-based non-intrusive load monitoring, which is also the current mainstream state-based non-intrusive load monitoring method. First, we study how to expand it into a factor hidden Markov model and apply it to non-intrusive load monitoring, including the task of non-intrusive load monitoring corresponding to the parameter learning problem and state inference problem in the hidden Markov model; then we use public data sets to experiment with the load decomposition method based on hidden Markov model to confirm the decomposition effect.

[0047] The hidden Markov model is applied to the actual scenario of non-intrusive load decomposition of electricity. The HMM is used to solve the non-intrusive load decomposition problem. The Viterbi algorithm is used to solve it step by step, and the optimal state sequence is found by traversing all state transitions. The Viterbi algorithm uses dynamic programming to efficiently solve the optimal path, that is, the state sequence with the highest probability. In the step-by-step decomposition process, the Viterbi algorithm infers the load state at time t+1 based on the load state at time t and the total power at time t+1. This process can be regarded as a mapping from input to output of the deep network. According to the universal approximation characteristics, when there is enough training data, the deep network can fully learn the operating characteristics of the load and realize the function of the Viterbi algorithm. The input of the non-intrusive load decomposition network is the state of each load at time t and the non-intrusive power at time t+1, and the output is the predicted parameter value of each load at time t+1.

[0048] Example 1, construction and task mapping of factor hidden Markov model First, a factor hidden Markov model is constructed according to the characteristics of non-intrusive load monitoring in the power system. The hidden state in the model is determined, for example, the different working states of each load device (such as on, off, different power operation states, etc.) are defined as hidden states. Then, the tasks in non-intrusive load monitoring, such as determining the working state and energy consumption parameters of each load device at different times, are mapped to parameter learning and state inference problems in the hidden Markov model. For parameter learning, the parameters such as the transition probability matrix and the emission probability matrix in the hidden Markov model are determined by analyzing a large amount of historical load data. For state inference, the known total load data and the learned model parameters are used to infer the hidden state of each load device at each moment.

[0049] The experimental process based on public data sets collects multiple public power load data sets, which contain load data of different types of load equipment in different scenarios. The load decomposition method based on the hidden Markov model is applied to these data sets. During the experiment, different experimental parameters are set, such as different initial model parameters, different data partitioning methods (ratios of training set, validation set, and test set), etc. By comparing and analyzing the experimental results, the decomposition effect of the method on different data sets is evaluated, such as calculating the error indicators between the decomposed load and the actual load (such as mean square error, mean absolute error, etc.) to determine the accuracy and stability of the method.

[0050] In the actual non-intrusive load decomposition scenario of electricity, the total load data at the power entrance is collected as input. The total load data is divided according to the time series, and the total load power value at time t and time t+1 is obtained at fixed time intervals (such as every minute, every hour, etc.). Based on the constructed hidden Markov model, the Viterbi algorithm is used for solution. At each time step, the Viterbi algorithm traverses all possible state transitions based on the load conditions at time t (which can be the various load states obtained by the previous decomposition or the default state initially set) and the total power at time t+1, and calculates the state sequence with the highest probability through dynamic programming. This state sequence represents the most likely working state combination of each load device at time t+1.

[0051] Training and function realization of non-intrusive load decomposition network To construct a non-intrusive load decomposition network, the network structure can adopt a multi-layer perceptron (MLP) or other suitable neural network structure. The conditions of each load at time t (such as the state information of each load obtained from the previous step) and the non-intrusive power at time t+1 are used as input data to train the network. During the training process, a large amount of historical load data is used to optimize the network, and the network parameters such as weights and bias are adjusted to minimize the error between the network output (the predicted parameter values ​​of each load at time t+1) and the actual value. According to the universal approximation characteristics of the deep network, when there are enough training data, the network can accurately learn the operating characteristics of the load, thereby realizing functions similar to the Viterbi algorithm, that is, accurately predicting the parameter values ​​of each load at time t+1, and then realizing non-intrusive load decomposition.

[0052] The application of factor hidden Markov model in non-intrusive load monitoring is extended to factor hidden Markov model for non-intrusive load monitoring. It is clear that the task of non-intrusive load monitoring corresponds to the parameter learning problem and state inference problem in hidden Markov model.

[0053] For parameter learning, a large amount of historical load data is analyzed to determine the parameters such as the transition probability matrix and emission probability matrix in the hidden Markov model. For state inference, the known total load data and the learned model parameters are used to infer the hidden state of each load device at each moment.

[0054] The experiment on the load decomposition method based on the hidden Markov model uses a public data set to experiment with the load decomposition method based on the hidden Markov model. A large amount of experimental data is used to verify the decomposition effect of this method in practical applications, thereby providing a basis for subsequent applications in practical scenarios.

[0055] Using HMM to solve the non-intrusive load decomposition problem and Viterbi algorithm application The hidden Markov model is applied to the actual scenario of non-intrusive load decomposition of power, and HMM is used to solve the non-intrusive load decomposition problem.

[0056] The Viterbi algorithm is used to solve the problem step by step. The algorithm searches for the best state sequence by traversing all state transitions. The Viterbi algorithm uses dynamic programming to efficiently solve the best path, that is, the state sequence with the highest probability. In the step-by-step decomposition process, the load state at time t+1 is inferred based on the load state at time t and the total power at time t+1. This process is similar to the mapping of deep networks from input to output.

[0057] Construction and function realization of non-intrusive load decomposition network A non-intrusive load decomposition network is constructed, whose input is the status of each load at time t and the non-intrusive power at time t+1, and the output is the predicted parameter value of each load at time t+1.

[0058] According to the universal approximation characteristics of the deep network, when there is sufficient training data, the network can fully learn the operating characteristics of the load, thereby realizing the function of the Viterbi algorithm and providing an effective implementation method for non-intrusive load decomposition.

[0059] HMM parameter learning The observation sequence and hidden state sequence are related in HMM, but usually there is only an observation sequence without a corresponding hidden state sequence. The parameter learning problem of HMM is to estimate the model parameters (initial state probability, state transition probability and generated observation probability) given the observation sequence, so that the probability of the observation sequence occurring under these parameters is maximized.

[0060] Viterbi learning algorithm (hard alignment): First, the optimal path is solved by the Viterbi algorithm, that is, "hard alignment", to obtain a new observation sequence and its corresponding state sequence. Then, these known state sequences are used to calculate the model parameters. This process is repeated until the parameters converge.

[0061] Baum-Welch learning algorithm (soft alignment): When the hidden state corresponding to the observation sequence is unknown, the Baum-Welch algorithm is used to estimate the model parameters. The parameters are estimated iteratively using the EM algorithm (expectation maximization algorithm). In the E step, the expected value of the unknown variable is calculated based on the assumed value of the parameter; in the M step, the maximum likelihood estimate of the parameter is calculated based on the estimated value of the unknown variable.

[0062] Improving the accuracy of non-intrusive load monitoring Based on the task mapping advantage of the factor hidden Markov model, it is possible to more accurately model the working status and energy consumption of load equipment and improve the accuracy of the estimation of the status and parameters of each load device.

[0063] The decomposition accuracy is improved and the decomposition error is reduced in actual non-intrusive load monitoring by using public data sets for verification.

[0064] Improve load decomposition efficiency The Viterbi algorithm uses dynamic programming to efficiently traverse all state transitions to find the optimal state sequence, thereby improving the speed of load decomposition.

[0065] The non-intrusive load decomposition network utilizes the universal approximation characteristics of the deep network to improve the computational efficiency and quickly and accurately predict the parameter values ​​of each load.

[0066] Reduce monitoring costs and complexity. Non-intrusive monitoring methods do not require the installation of sensors on each load device, reducing hardware and labor costs and reducing system complexity.

[0067] Accurate load monitoring and decomposition results provide reliable data support for the operation and management of power systems, optimize power distribution, formulate reasonable electricity price strategies, and conduct effective grid dispatching.

[0068] Adapt to the dynamic changes of load equipment in the power system and maintain accurate load monitoring and decomposition capabilities by continuously updating model parameters or retraining the network.

[0069] The method based on hidden Markov model and its deformation belongs to the statistical state-based non-intrusive load monitoring, which is also the current mainstream state-based non-intrusive load monitoring method. First, we study how to expand it into a factor hidden Markov model and apply it to non-intrusive load monitoring, including the task of non-intrusive load monitoring corresponding to the parameter learning problem and state inference problem in the hidden Markov model; then we use public data sets to experiment with the load decomposition method based on hidden Markov model to confirm the decomposition effect.

[0070] The hidden Markov model is applied to the actual scenario of non-intrusive load decomposition of electricity. The HMM is used to solve the non-intrusive load decomposition problem. The Viterbi algorithm is used to solve it step by step, and the optimal state sequence is found by traversing all state transitions. The Viterbi algorithm uses dynamic programming to efficiently solve the optimal path, that is, the state sequence with the highest probability. In the step-by-step decomposition process, the Viterbi algorithm infers the load state at time t+1 based on the load state at time t and the total power at time t+1. This process can be regarded as a mapping from input to output of the deep network. According to the universal approximation characteristics, when there is enough training data, the deep network can fully learn the operating characteristics of the load and realize the function of the Viterbi algorithm. The input of the non-intrusive load decomposition network is the state of each load at time t and the non-intrusive power at time t+1, and the output is the predicted parameter value of each load at time t+1.

Claims

1. A non-intrusive load monitoring and decomposition method based on hidden Markov model and its variants, characterized in that: include: S1. Extending the hidden Markov model into a factor hidden Markov model and applying it to non-intrusive load monitoring, where the task of non-intrusive load monitoring corresponds to the parameter learning problem and state inference problem in the hidden Markov model; S2, use public data sets to experiment with the load decomposition method based on the hidden Markov model to confirm the decomposition effect; S3. Using Hidden Markov Model (HMM) to solve the non-intrusive load decomposition problem, the Viterbi algorithm is used to solve it step by step. The Viterbi algorithm searches for the best state sequence by traversing all state transitions, and uses dynamic programming to efficiently solve the state sequence with the highest probability. In the step-by-step decomposition process, the load state at time t+1 is inferred based on the load state at time t and the total power at time t+1. S4. Construct a non-intrusive load decomposition network, whose input is the status of each load at time t and the non-intrusive power at time t+1, and the output is the predicted parameter value of each load at time t+1. According to the universal approximation characteristics of the deep network, when there is sufficient training data, the function of the Viterbi algorithm is implemented to perform non-intrusive load decomposition.

2. The non-intrusive load monitoring and decomposition method based on hidden Markov model and its variants according to claim 1, characterized in that: When the hidden Markov model is expanded into a factor hidden Markov model and applied to non-intrusive load monitoring, the parameter learning determines parameters such as a transition probability matrix and a transmission probability matrix in the hidden Markov model by analyzing historical load data.

3. The non-intrusive load monitoring and decomposition method based on hidden Markov model and its variants according to claim 1, characterized in that: In an experiment using a public data set to test a load decomposition method based on a hidden Markov model, the decomposition effect is evaluated by calculating an error index between the decomposed load and the actual load, wherein the error index includes at least one of a mean square error and a mean absolute error.

4. The non-intrusive load monitoring and decomposition method based on hidden Markov model and its variants according to claim 1, characterized in that: In solving the non-intrusive load decomposition problem using a hidden Markov model (HMM), the Viterbi algorithm traverses all possible state transitions at each time step to calculate the state sequence with the highest probability based on the load condition at time t and the total power at time t+1. The load condition at time t includes the load conditions obtained by the previous decomposition or the default state initially set.

5. The non-intrusive load monitoring and decomposition method based on hidden Markov model and its variants according to claim 1, characterized in that: In constructing a non-intrusive load decomposition network, the network structure adopts a multi-layer perceptron (MLP) or other suitable neural network structures, and the error between the network output and the actual value is minimized by adjusting parameters such as network weights and biases.

Images