A non-intrusive load decomposition method with feature enhancement
Through feature enhancement and deep neural network training, combined with lag, sliding window and extended window features, the problem of low accuracy in traditional non-invasive load decomposition methods is solved, significantly improving the accuracy and technical maturity of load decomposition.
Patent Information
- Application Number
- CN202210466293.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-04-27
- Publication Date
- 2024-12-20
- Estimated Expiration
- 2042-04-27
AI Technical Summary
The traditional non-invasive load decomposition method has large calculation volume, single training characteristics, difficult accurate reasoning, and low nonlinear learning ability, resulting in low load decomposition accuracy and incomplete technology.
Non-invasive load decomposition methods with feature enhancement, including load data acquisition, data preprocessing, feature enhancement, feature splicing and construction of electrical load decomposition models. Through deep neural network training, hysteresis features, sliding window features or extended window features are used. The combination of features and model learning ability is improved.
It effectively improves the accuracy of load decomposition, achieves more accurate power load decomposition, and improves the technical maturity and practicality of non-invasive load decomposition.
Smart Images
Figure QLYQS_1 
Figure QLYQS_2 
Figure QLYQS_3
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of power load analysis, and in particular to a feature-enhanced non-invasive load decomposition method. Background Art
[0002] Non-intrusive load decomposition is to analyze the total meter data of a specific area to obtain the power consumption and energy consumption values of each power load within the range. Compared with the traditional invasive load decomposition method, non-intrusive load decomposition does not require the installation of measurement devices for each user's power equipment to obtain the operating conditions of each power equipment. On the one hand, it saves the purchase cost of sensor equipment, and on the other hand, it avoids the trouble of modifying and maintaining the existing power equipment lines one by one. It is a convenient, low-cost and highly versatile method.
[0003] Non-intrusive load decomposition has great practical value for all electricity users. For ordinary users, it helps guide users to use electricity rationally, thereby achieving energy conservation and consumption reduction, and saving electricity bills; for power companies, non-intrusive load decomposition can improve the safety and economy of the power grid, and also help to more accurately model user behavior and provide differentiated and precise services to users; for power equipment manufacturers, the equipment status and corresponding energy consumption information provided by non-intrusive load decomposition can provide a basis for equipment fault diagnosis, thereby maintaining the equipment.
[0004] However, the traditional non-intrusive load decomposition method has the problems of large amount of computation, single training features, difficulty in accurate reasoning, and low nonlinear learning ability, which leads to low load decomposition accuracy. At present, the research in this field is not mature and the technology is not perfect, so there is still a lot of room for improvement. Summary of the invention
[0005] In order to solve the problems involved in the above background, the technical solution used in the present invention is: a feature-enhanced non-intrusive load decomposition method, comprising the following steps:
[0006] Step 1: Load data collection: Collect the active power data of the total load and each power load over a period of time to construct the original data set;
[0007] Step 2: Data preprocessing: fill in and normalize the missing values of the total load and each power load data in the original data set;
[0008] Step 3, feature enhancement: constructing enhanced features for the total load active power sequence preprocessed in step 2, the enhanced features including hysteresis features, sliding window features, extended window features, or a combination of several features;
[0009] Step 4: Feature splicing: Splice the original features and the enhanced features constructed in step 3 to form a multi-dimensional feature sequence. The specific process includes:
[0010] The data set processed by the method described in step 2 has two columns, the first column is time, and the second column is the normalized corresponding active power. Now add a third column of data, and set the value of the third column to the value obtained by subtracting the total load active power of each row from the total load active power of the previous row;
[0011] Set a sliding window of length n, given an input sequence of length l X = (X1, X2, ..., X i ,…,X l ), X i is the total load active power at the i-th time point, and the length of each is and Fill with zero values;
[0012] After the above processing, the input is a processed main load window H with a length of n. i:i+n-1 , H i =(X i ,X i -X i-1 ), the output is the midpoint element of each corresponding load window is the active power of the power load j at the kth time point, where And the input sequence dimension is (l,n,2), and the output sequence dimension is (l,1);
[0013] Step 5: Construct a power load decomposition model: Take the multi-dimensional feature sequence obtained in step 4 as input, and the active power sequence of each power load as output, train the deep neural network model, and obtain the decomposition model corresponding to each power load. The specific process includes:
[0014] Establish a deep neural network model for each power load;
[0015] Given an input sequence X of length l = (X1, X2, ..., X i ,…,X l ), X i Represents the total load active power at time point i. The eigenvector With the original feature vector X i Splice to get the final input sample
[0016] Construct sample vector W = (W1, W2, ..., W i ,…,W l) is used as the input of the deep neural network model of each power load, and its label vector is Y j =(Y j,1 ,Y j,2 ,L,Y j,i ,L,Y j,l ), Y j represents the jth load vector in the power load to be decomposed, Y j,i is the active power of load j at time point i, l represents the length of input sequence;
[0017] Using W and Y j Train the power load decomposition model and iteratively update the load decomposition model parameters according to the mean square error function gradient information;
[0018] Step 6, power load decomposition: perform feature enhancement and feature concatenation on the total load active power sequence according to steps 3 and 4 to form a multi-dimensional feature sequence, which is input into the power load decomposition model constructed in step 5 to obtain the decomposed active power sequence of each power load.
[0019] In the above-mentioned feature-enhanced non-intrusive load decomposition method, the data normalization method in step 2 is:
[0020]
[0021] Among them, x i represents the active power value of the power load at time i, x max Indicates the maximum active power value in the power load time series, It represents the normalized active power value of the power load at time i.
[0022] In the above-mentioned feature-enhanced non-intrusive load decomposition method, the specific steps of feature enhancement in step 3 are as follows:
[0023] By processing the power load window data, a new feature is constructed for the total load active power sequence preprocessed in step 2. Feature enhancement includes hysteresis feature, sliding window feature, extended window feature, or a combination of several features.
[0024] Constructing lag features is a classic way to transform time series prediction problems into supervised learning problems, where the lag value of the lag feature is determined by autocorrelation analysis. In the present invention, through autocorrelation analysis, it is found that the feature correlation is highest when the feature is one time step apart, so the lag value is set to 1.
[0025] The sliding window feature uses the statistical information of the previous time observation value as a feature, and the value of the sliding window is jointly determined by the type and autocorrelation of the power load. In the present invention, different sliding windows are set for different power loads.
[0026] The expanding window feature is different from the sliding window feature in that the window size is constant while the expanding window feature considers all historical values of the load. At each step, the size of the expanding window increases, taking into account each new value added to the sequence.
[0027] In the above-mentioned feature-enhanced non-intrusive load decomposition method, the specific steps of feature splicing in step 4 are as follows:
[0028] After step 2, the data set has two columns. The first column is time, and the second column is the corresponding normalized active power. Now add a third column of data and set the value of the third column to the value obtained by subtracting the total load active power of the previous row from the total load active power of each row. Set a sliding window of length n. Given an input sequence of length l, X = (X1, X2, ..., X i ,…,X l ), X i is the total load active power at the i-th time point, and the length of each is and The input is a processed main load window H of length n. i:i+n-1 , H i =(X i ,X i -X i-1 ), the output is the midpoint element of each corresponding load window is the active power of the power load j at the kth time point, where The input sequence dimension is (l,n,2) and the output sequence dimension is (l,1).
[0029] In the above-mentioned feature-enhanced non-intrusive load decomposition method, the specific steps of step 5 for training the corresponding load decomposition model for each power load are as follows:
[0030] A sliding window of length n is set, and the dimension of the first input layer of the neural network model is (n,2); the second and third layers are one-dimensional convolution layers, consisting of 16 convolution kernels of length 4, the convolution step size is 1, the activation function is a linear function, and the padding value is set to same; the fourth layer is a maximum pooling layer, and the size of the pooling window is set to 2; the fifth layer is an LSTM layer, the output dimension is 128, the return sequence is set to True, and the return status is set to False; the sixth layer is a flattening layer, which "flattens" the multi-dimensional input into one dimension; the seventh and eighth layers are fully connected layers, the output dimension of the seventh layer is 128, the activation function is the tanh function, and the eighth layer is the output layer, the output dimension is 1, and the activation function is a linear function; finally, the loss function mse and the optimizer adam are selected to train the network model. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] The present invention is further described below in conjunction with the accompanying drawings and embodiments.
[0032] Figure 1 It is a flow chart of the present invention.
[0033] Figure 2 Schematic diagram of the deep neural network model structure used in the specific implementation.
[0034] Figure 3 Schematic diagram of the feature enhancement proposed in the present invention. DETAILED DESCRIPTION
[0035] The technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the present invention.
[0036] according to Figure 1 The present invention discloses a feature-enhanced non-intrusive load decomposition method, which specifically comprises the following steps:
[0037] Step 1: Load data collection
[0038] This embodiment selects five typical appliances, kettle, refrigerator, washing machine, microwave oven, and dishwasher, for load decomposition. These include simple switch-type appliances such as kettles, as well as complex multi-state appliances such as refrigerators, washing machines, microwave ovens, and dishwashers. The total load active power is the sum of the active power of these five appliances and the rest, i.e., P 总 =P 水壶 +P 冰箱 +P 洗衣机 +P 微波炉 +P 洗碗机 +P 其它电器 The active power and total active power data of the five electrical appliances were collected from October 1, 2013 to January 10, 2014 at a time point of 6 seconds. The data from October 1, 2013 to December 31, 2014 were used as the training set, with a total of 1,324,800 time points; the data from January 1, 2014 to January 10, 2014 were used as the test set, with a total of 144,000 time points.
[0039] Step 2: Data preprocessing
[0040] For the missing values of total load and active power of each power load, use the fillna method of pandas, and select the parameter ffill, that is, fill the missing value with the previous non-missing value. Then normalize the filled data, and the normalization formula is:
[0041]
[0042] Among them, x i represents the active power value of the power load at time i, x max Indicates the maximum active power value in the power load time series, It represents the normalized active power value of the power load at time i.
[0043] Step 3: Feature Enhancement
[0044] Feature enhancement includes hysteresis features, sliding window features, extended window features, or a combination of several features.
[0045] Constructing hysteresis features: In the present invention, after autocorrelation analysis, the feature correlation is the highest when the feature is separated by one time step, so the hysteresis value is set to 1. The data set processed by step 2 has two columns, the first column is time, and the second column is the corresponding normalized active power. Now add a third column of data, and set the value of the third column to the value obtained by subtracting the active power of the previous row from the active power of each row;
[0046] Constructing sliding window features: In the present invention, different sliding window values are set for different power loads. For a specific power load, a sliding window of length n is set, and the sliding step is set to 1. Given an input sequence of length l, X = (X1, X2, ..., X i ,…,X l ), X i is the total load active power at the i-th time point. The model of the present invention is a sequence-to-point model. For the time points at the beginning and end of the sequence, the length of and Fill with zero values so that the power points can form a complete sliding window, and these power points are located exactly in the middle of the sliding window;
[0047] Constructing Expanding Window Features: Unlike sliding window features, where the window size is constant, expanding window features take into account all past values. At each step, the size of the expanding window increases, taking into account each new value in the sequence.
[0048] The feature enhancement adopted by the present invention is a combination of a hysteresis feature and a sliding window feature, but the protection scope of the present invention is not limited thereto, and includes a hysteresis feature, a sliding window feature, an extended window feature, or a combination of several features.
[0049] Step 4: Feature stitching
[0050] The original features and the new features constructed in step three are concatenated to form a multi-dimensional feature sequence.
[0051] Specific as Figure 3As shown in the figure, a sliding window of length n is set, with 6 seconds as a time point, given an input sequence of length l X = (X1, X2, ..., X i ,…,X l ), X i is the total load active power at the i-th time point, is the active power of the power load j at the kth time point. Each input is a total load active power window H with a length of n and two-dimensional features. i:i+n-1 , that is, H i =(X i ,X i -X i-1 ), H i It consists of two features: the total load active power at the i-th time point and the difference between the total load active power at the i-th time point and the total load active power at the i-1-th time point. The output is the midpoint element of the main load window corresponding to the power load j. in The value of the total load active power window H with a length of n i:i+n-1 The input sequence dimension is (l,n,2) and the output sequence dimension is (l,1).
[0052] Step 5: Build a power load decomposition model
[0053] The total load active power constructed in step 3 and step 4 and the difference sequence of total load active power at different time points are used as input, the data from October 1, 2013 to December 31, 2013 are used as training sets, and the active power of kettle, refrigerator, washing machine, microwave oven and dishwasher are used as output. Figure 2 The deep neural network model shown is trained, where the first layer is the input layer; the second and third layers are CNN layers; the fourth layer is the maximum pooling layer; the fifth layer is the LSTM layer; the sixth layer is the flattening layer; the seventh and eighth layers are fully connected layers, and the eighth layer is the output layer. The specific parameters are set as follows:
[0054] Set a sliding window of length n, and the first input layer dimension of the neural network model is (n,2); the second and third layers are one-dimensional convolution layers, consisting of 16 convolution kernels of length 4, the convolution step is 1, the activation function is a linear function, and the padding value is set to same; the fourth layer is a maximum pooling layer, and the size of the pooling window is set to 2; the fifth layer is an LSTM layer, the output dimension is 128, the return sequence is set to True, and the return status is set to False; the sixth layer is a flattening layer, which "flattens" the multi-dimensional input into one dimension; the seventh and eighth layers are fully connected layers, the output dimension of the seventh layer is 128, the activation function is the tanh function, and the eighth layer is the output layer, the output dimension is 1, and the activation function is a linear function; finally, the loss function mse and the optimizer adam are selected to train the network model. The load decomposition models of kettles, refrigerators, washing machines, microwave ovens, and dishwashers are trained respectively.
[0055] Step 6: Power load decomposition
[0056] The data from January 1, 2014 to January 10, 2014 are used as the test set to decompose the five types of power loads, namely kettle, refrigerator, washing machine, microwave oven and dishwasher. The non-intrusive load decomposition results of the method of the present invention are compared with the results of other classic non-intrusive load decomposition methods, as shown in the following table. The evaluation indicators include MAE mean absolute error and F1 score.
[0057] Table 1
[0058]
[0059]
[0060] The above results show that the method of the present invention effectively improves the accuracy of non-invasive load decomposition.
[0061] Method 1 Source: Batra N, Kelly J, Parson O, et al. NILMTK: An open sourcetoolkit for non-intrusive load monitoring[C] / / Proceedings of the5thinternational conference on Future energy systems.2014:265-276.
[0062] Method 2 Source: Bonfigli R, Felicetti A, Principi E, et al. Denoising autoencoders for non-intrusive load monitoring: improvements and comparative evaluation[J]. Energy and Buildings, 2018, 158: 1461-1474.
[0063] Method 3 Source: Zhang C, Zhong M, Wang Z, et al. Sequence-to-point learning with neural networks for non-intrusive load monitoring[C] / / Proceedings of the AAAI Conference on Artificial Intelligence. 2018, 32(01): 2604-2611.
[0064] Method 4 Source: Shin C, Joo S, Yim J, et al. Subtask gated networks for non-intrusive load monitoring[C] / / Proceedings of the AAAI Conference on Artificial Intelligence. 2019, 33(01): 1150-1157.
[0065] Method 5 Source: Chen K, Zhang Y, Wang Q, et al. Scale-and context-aware convolutional non-intrusive load monitoring[J]. IEEE Transactions on Power Systems, 2019, 35(3): 2362-2373.
Claims
1. A feature-enhanced non-intrusive load decomposition method, It is characterized in that The following steps are involved: Step 1: Load data collection: Collect the active power data of the total load and each power load over a period of time to construct the original data set; Step 2: Data preprocessing: fill in and normalize the missing values of the total load and each power load data in the original data set; Step 3, feature enhancement: constructing enhanced features for the total load active power sequence preprocessed in step 2, the enhanced features including hysteresis features, sliding window features, extended window features, or a combination of several features; Step 4: Feature splicing: Splice the original features and the enhanced features constructed in step 3 to form a multi-dimensional feature sequence. The specific process includes: The data set processed by the method described in step 2 has two columns, the first column is time, and the second column is the normalized corresponding active power. Now add a third column of data, and set the value of the third column to the value obtained by subtracting the total load active power of each row from the total load active power of the previous row; Set a sliding window of length n, given an input sequence of length l X = (X 1 ,X 2 ,…,X i ,…,X l ), X i is the total load active power at the i-th time point, and the length of each is and Fill with zero values; After the above processing, the input is a processed main load window H with a length of n. i:i+n-1 , H i =(X i ,X i -X i-1 ), the output is the midpoint element of each corresponding load window is the active power of the power load j at the kth time point, where And the input sequence dimension is (l,n,2), and the output sequence dimension is (l,1); Step 5: Construct a power load decomposition model: Take the multi-dimensional feature sequence obtained in step 4 as input, and the active power sequence of each power load as output, train the deep neural network model, and obtain the decomposition model corresponding to each power load. The specific process includes: Establish a deep neural network model for each power load; Given an input sequence X of length l, 1 ,X 2 ,…,X i ,…,X l ), X i Represents the total load active power at time point i; the enhanced eigenvector With the original feature vector X i Splice to get the final input sample Construct sample vector W = (W 1 ,W 2 ,…,W i ,…,W l ) is used as the input of the deep neural network model of each power load, and its label vector is Y j =(Y j,1 ,Y j,2 ,L,Y j,i ,L,Y j,l ), Yj represents the jth load vector in the power load to be decomposed, Y j,i is the active power of load j at time point i, l represents the length of input sequence; Using W and Y j Train the power load decomposition model and iteratively update the load decomposition model parameters according to the mean square error function gradient information; Step 6, power load decomposition: perform feature enhancement and feature concatenation on the total load active power sequence according to steps 3 and 4 to form a multi-dimensional feature sequence, which is input into the power load decomposition model constructed in step 5 to obtain the decomposed active power sequence of each power load.
2. A feature-enhanced non-intrusive load decomposition method according to claim 1, It is characterized in that The network structure of the model is: The first layer is the input layer; the second and third layers are CNN layers; the fourth layer is the pooling layer; The fifth layer is an LSTM layer; the sixth layer is a flat layer; the seventh and eighth layers are fully connected layers, and the eighth layer is the output layer.