Sequence Mapping Load Identification and Decomposition Method Based on Multi-Scale Residual Neural Network

Abstract

The present invention relates to the technical field of load identification, and specifically relates to a sequence mapping load identification and decomposition method based on a multi-scale residual neural network, including constructing a load identification model, the network structure of the load identification model being a multi-scale residual network based on dilated convolution; obtaining an aggregated power dataset and an appliance-level power dataset in a public dataset, preprocessing the aggregated power dataset and the appliance-level power dataset to obtain a preprocessed training set and a test set; using the preprocessed training set and test set to train and test the load identification model to obtain a trained load identification model. By constructing a load identification model based on a multi-scale residual network of dilated convolution, the present invention can more specifically learn the characteristics of mixed data and improve the load identification effect of the load identification model.

Description

Technical Field

[0001] The present invention relates to the technical field of load identification, and particularly to a sequence mapping load identification and decomposition method based on a multi-scale residual neural network. Background Art

[0002] With the increasing complexity of household power consumption demands, smart grids and load monitoring technologies have become indispensable. Among them, load monitoring includes intrusive load monitoring (ILM) and non-intrusive load monitoring (NILM). Intrusive load monitoring refers to installing sensors on each household appliance. Although the measurement data obtained through intrusive load monitoring is accurate and reliable, it has defects such as high investment costs, poor operability, and low user acceptance. Non-intrusive load monitoring only sets a measurement point at the entrance of the user's home. And with the increasing demand for the economy and efficiency of measurement technologies, non-intrusive load monitoring, as a cost-effective power measurement method, has received more and more attention.

[0003] Existing non-intrusive load monitoring methods mainly include methods based on key signal features and data-driven methods. The former does not require statistical learning of the model and only needs to analyze the signal spectrum features to achieve the best results, while the latter relies on a certain amount of labeled data to learn a representative model, which can be further roughly divided into traditional machine learning, deep learning, and online transfer learning.

[0004] For the method based on key signal features, it relies on accurately detecting the switching events of electrical appliances. Signal processing techniques can be used to detect these key features. Among them, electrical features such as current, voltage, and active power are time-domain features based on waveforms. On the other hand, advanced strongly time-varying signal processing algorithms introduce more transform-domain features, such as spectrum-based features and wavelet-based features. Through sufficient artificial feature description labels, an optimization framework can be applied to non-training methods. However, the generalization ability of the model always depends on prior expert knowledge, and this method has obvious defects in the deployment of some large-scale energy management systems.

[0005] For data-driven methods, they adopt statistical learning frameworks such as single-label classification and multi-label classification. In deep neural networks, as the network depth increases continuously, better performance can be achieved, but the network will face degradation problems, including gradient explosion and gradient disappearance. Although some existing models can achieve good load identification performance in the face of simple electrical appliance hybrid application scenarios, the identification effect in complex scenarios is not satisfactory. Summary of the Invention

[0006] To solve the technical problem of low identification effect existing in the above-mentioned existing non-intrusive load monitoring method, the purpose of the present invention is to provide a sequence mapping load identification and decomposition method based on a multi-scale residual neural network, and the specific technical solution adopted is as follows:

[0007] An embodiment of the present invention provides a sequence mapping load identification and decomposition method based on a multi-scale residual neural network, and the method includes the following steps:

[0008] Construct a load identification model, and the network structure of the load identification model is a multi-scale residual network based on dilated convolution;

[0009] Obtain the aggregated power dataset and the appliance-level power dataset in the public dataset, and preprocess the aggregated power dataset and the appliance-level power dataset to obtain the preprocessed training set and test set;

[0010] Use the preprocessed training set and test set to train and test the load identification model to obtain a trained load identification model.

[0011] Further, the multi-scale residual network based on dilated convolution includes: dilated convolution residual blocks;

[0012] The input of the dilated convolution residual block sequentially passes through two convolutional layers and their corresponding normalization layers, dropout layers, and non-linear activation functions; then, the original input is added to the temporal output through a shortcut connection to form the final output of the dilated convolution residual block; wherein, the two convolutional layers in the dilated convolution residual block share the same dilation rate.

[0013] Further, the calculation formula for the receptive field of the dilated convolution residual block is:

[0014] ;

[0015] In the formula, S represents the receptive field of the dilated convolution residual block, N represents the number of dilated convolutions, n represents the network layer number of the dilated convolution, and k represents the convolution kernel size.

[0016] Further, the obtaining of the aggregated power dataset and the appliance-level power dataset in the public dataset includes:

[0017] Obtain the UK-DALE public dataset, and select the aggregated power data and the appliance-level power data of each target appliance in the first preset number of target families in the UK-DALE public dataset at the same sampling frequency; wherein, the first preset number of target families are families with the second preset number of the same type of appliances, and the target appliances are appliances of the same type;

[0018] Determine the electrical appliance activation time period for each target household; based on the obtained aggregated power dataset and appliance-level power dataset, use the Electric.get_activation function of non-intrusive load monitoring and the parameters provided by UK-DALE during the activation time period to obtain activation information, and then obtain a pair of aggregated power sequences and labeled power sequences corresponding to each time index.

[0019] Further, the preprocessing of the aggregated power dataset and the appliance-level power dataset to obtain the preprocessed training set and test set includes:

[0020] First, perform normalization processing on the aggregated power dataset and the appliance-level power dataset to obtain the normalized aggregated power dataset and appliance-level power dataset;

[0021] Then, based on the normalized aggregated power dataset and appliance-level power dataset, starting from the activation start point, use a window of a preset scale to divide the power data that does not contain the activation information of the target appliance, obtain the updated aggregated power dataset and appliance-level power dataset, and use the updated appliance-level power dataset as the preprocessed test set;

[0022] Perform filtering processing on the updated aggregated power dataset to obtain the filtered aggregated power dataset as the preprocessed training set.

[0023] Further, the filtering processing of the updated aggregated power dataset to obtain the filtered aggregated power dataset includes:

[0024] Remove the updated aggregated power data of a preset scale with an activation length less than the first preset multiple, and the updated aggregated power data of a preset scale where the ratio of the aggregated power point to the maximum value of the appliance power is greater than the third preset multiple and the activation length is greater than the second preset multiple, to obtain the filtered aggregated power dataset; where the first preset multiple is less than the second preset multiple, and the second preset multiple is less than the third preset multiple.

[0025] The present invention has the following beneficial effects:

[0026] The present invention provides a sequence mapping load identification and decomposition method based on a multi-scale residual neural network. The method constructs a load identification model with a network structure of a multi-scale residual network based on dilated convolution. This load identification model can avoid the common degradation problems that occur in traditional networks when increasing the number of layers to learn more complex features. At the same time, the proposed dilated convolution can reduce a large number of model parameters and obtain a larger receptive field and multi-scale data structure, so as to more specifically learn the mixed data features, enhancing the identification ability of the load identification model. Through the multi-scale structure, the model can achieve the best performance when facing electrical appliances with complex power characteristics. Preprocess the aggregated power dataset and the appliance-level power dataset to obtain the preprocessed training set and test set, which helps to improve the efficiency of the load identification model during the training process and further enhance the identification performance of the load identification model. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following described drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0028] Figure 1 It is a flowchart of the steps of a sequence mapping load identification and decomposition method based on a multi-scale residual neural network according to an embodiment of the present invention;

[0029] Figure 2 It is a schematic diagram of the load identification process based on NILM in an embodiment of the present invention;

[0030] Figure 3 It is a structural diagram of a dilated convolution residual block in an embodiment of the present invention;

[0031] Figure 4 It is a schematic diagram of a non-causal dilated convolution in an embodiment of the present invention;

[0032] Figure 5 It is the active power curve of the washing machine working cycle in an embodiment of the present invention;

[0033] Figure 6 It is a schematic diagram of the overall structure of a multi-scale residual network based on dilated convolution in an embodiment of the present invention;

[0034] Figure 7 It is a structural diagram of BiLSTM;

[0035] Figure 8 It is a structural diagram of a denoising autoencoder;

[0036] Figure 9 It is the structural diagram of a traditional convolutional network;

[0037] Figure 10 It is the result of testing different types of electrical appliances with different models in the embodiments of the present invention Figure 1 ;

[0038] Figure 11 It is the result of testing different types of electrical appliances with different models in the embodiments of the present invention Figure 2 。 Detailed implementation manners

[0039] In order to further elaborate on the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following, in combination with the accompanying drawings and preferred embodiments, details the specific implementation manners, structures, features and their effects of the technical solutions proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, the specific features, structures or characteristics in one or more embodiments can be combined in any suitable form.

[0040] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0041] Traditional load detection methods are event detection-based. The existing proposed latent factor Markov model is superior to basic statistical models, but when applied to complex implementation scenarios, there may be some limitations; while neural networks have powerful non-linear modeling capabilities when the amount of knowledge data is large enough, but due to the simple model structure, they lack learning ability; therefore, the identification effect of traditional load detection methods is low.

[0042] In order to improve the identification effect of load detection, this embodiment provides a sequence mapping load identification and decomposition method based on a multi-scale residual neural network, as Figure 1 shown, including the following steps:

[0043] S1, construct a load identification model with a network structure of a multi-scale residual network based on dilated convolution.

[0044] In order to separately extract the active power data of each electrical appliance from the total active power data, an independent neural network is trained for all types of electrical appliances. The input of the neural network is the total power sequence, and the network outputs a sequence of the corresponding length, and the output content only contains the active power consumption of a single electrical appliance. Among them, the schematic diagram of the load identification process based on NILM is as Figure 2 shown.

[0045] First of all, regarding the dilated convolution residual block in the load identification model, it includes:

[0046] This embodiment proposes to use the network architecture based on dilated convolutional residual blocks as the basic structural unit. Residual blocks can avoid the problem of vanishing or exploding gradients when the number of network layers increases. At the same time, residual blocks can ensure that the network makes full use of its depth to enhance and improve the overall learning performance. Among them, the structural diagram of the dilated convolutional residual block is as shown in Figure 3 the following.

[0047] In Figure 3 it, the input of a single dilated convolutional residual block passes through two convolutional layers and their corresponding normalization layers, dropout layers, and non-linear activation functions in sequence; then, the original input is added to the temporal output through a shortcut connection to form the final output of the dilated convolutional residual block. It should be noted that if the number of input channels is different from the number of temporal output channels, a convolutional layer (kernel size of 1) will be used to change the number of channels to make it equal to the number of temporal output channels, and then addition is performed.

[0048] It should be noted that the shortcut connection can enable each dilated convolutional residual block to directly learn features from the original input, which greatly improves the learning ability of the network; connecting all the dilated convolutional residual blocks in series to form a residual block aggregate can expand the depth of the residual network; compared with traditional convolutions, the dilated convolution in the residual block has two additional characteristics: First, due to the sequence-to-sequence nature, it is necessary to ensure that the sequence lengths before and after each convolution are the same; Second, there will be holes (i.e., zero values) in the convolutional kernel, and the number of these holes depends on the dilation rate of this layer. This characteristic enables the network to obtain a larger receptive field by selectively skipping some parts of the input without changing the size of the convolutional kernel or increasing the parameters. For example, if d represents the dilation rate, then d - 1 values will be skipped between every two actual convolutional values, that is, when d = 1, the dilated convolution degenerates significantly into an ordinary 1D convolution.

[0049] In the dilated convolutional residual block of this embodiment, the two convolutional layers share the same dilation rate, and as the network depth increases, the dilation rate of the residual block will gradually increase exponentially, and the receptive field increases simultaneously. The schematic diagram of the non-causal dilated convolution is as shown in Figure 4 the following, and the case with a kernel size of 3 is shown in Figure 4 it.

[0050] In Figure 4 it, let be the original input sequence, and be the final output sequence. It can be shown that for time series, convolution has non-causal characteristics, that is, its output value at time t is determined by past and future information.

[0051] In this embodiment, when load identification is regarded as a denoising problem, more information from the sample context can help the network better learn the characteristics of the pure signal. Different from the RNN that calculates strictly in order, that is, the RNN will result in a large consumption of training and prediction time, while the dilated convolutional network has the advantage of processing multiple timelines simultaneously, and the receptive field has higher flexibility. In particular, if the dilation rate d of the last residual block is , its receptive field can be calculated according to the following formula, and its calculation formula can be:

[0052] ;

[0053] In the formula, S represents the receptive field of the dilated convolutional residual block, N represents the number of dilated convolutions, n represents the number of network layers of the dilated convolution, and k represents the convolution kernel size. When the convolution kernel size represented by k is odd, then 2 in the calculation formula n represents the dilation rate in the residual block body.

[0054] Secondly, regarding the multi-scale residual network structure in the load identification model, it includes:

[0055] First of all, considering that different electrical appliances have different working modes, even for the same type of electrical appliances, there will be different sub-modes within a working cycle. In different working modes and sub-modes, the power curve characteristics including duration, power, and fluctuation are significantly different from each other. For example, the active power curve of the washing machine during the working cycle is as Figure 5 shown, Figure 5 The working modes in the two dashed boxes in are washing and auxiliary heating respectively, and the power characteristics of these two modes are significantly different.

[0056] If a receptive field can only cover a certain specific working mode, then the network will have a better representation of this specific feature mode. However, if the receptive field is blindly increased, the learning may become aimless, and the characteristics of different working modes may be mixed and overlapped with each other, thus deteriorating the final identification performance. Therefore, this embodiment proposes a new structure that can combine different receptive fields together, enabling the network to learn representations, and at the same time combining the activation durations of different electrical appliances from multiple receptive fields to reduce information loss. The new structure can be a multi-scale residual network structure. The schematic diagram of the overall structure of the multi-scale residual network based on dilated convolution is as Figure 6 shown.

[0057] In Figure 6Among them, c represents the output channel, d represents the dilation rate, and k represents the convolutional kernel size; the input sequence is fed into four residual block bodies, each with a convolutional kernel size of 5, respectively containing two blocks, three blocks, four blocks, and five blocks for feature extraction; in each residual block body, the dilation rate and output channel of the first block and the last block are specified; all the intermediate blocks are multiplied in sequence; based on the calculation formula of the receptive field, the receptive fields contain 25, 57, 121, and 249 sampling points respectively.

[0058] The dataset used in the verification process of this embodiment is the application-level power dataset (UK-DALE), whose sampling frequency is 1 point per 6 seconds. Therefore, in this embodiment, the size of the receptive field can be corresponding to the actual time duration as 150 seconds, 342 seconds, 726 seconds, and 1494 seconds; then, the outputs of each residual block body are concatenated in depth, fed into two fully connected layers, and output through the sigmoid activation function; the activation functions of all other layers are ReLU (Rectified Linear Unit), and the stride of all convolutional layers is 1.

[0059] It should be noted that the proposed novel convolutional model of the residual module can avoid the common degradation problem that occurs when traditional networks increase the number of layers to learn more complex features; at the same time, the proposed dilated convolution can reduce a large number of model parameters and obtain a larger receptive field and multi-scale data structure, so as to more specifically learn the mixed data features.

[0060] So far, this embodiment has obtained the load identification model.

[0061] S2. Obtain the aggregated power dataset and the appliance-level power dataset in the public dataset, and preprocess the aggregated power dataset and the appliance-level power dataset to obtain the preprocessed training set and test set.

[0062] The above step S2 can be implemented through the following steps S21 to S22:

[0063] S21. Obtain the aggregated power dataset and the appliance-level power dataset in the public dataset.

[0064] First step, obtain the UK-DALE public dataset, and select the aggregated power data and the appliance-level power data of each target appliance in the first preset number of target families in the UK-DALE public dataset at the same sampling frequency.

[0065] Here, the first preset number of target families are families with the second preset number of the same type of appliances, and the target appliances are the same type of appliances. Among them, the first preset number can take an empirical value as a positive integer greater than or equal to 3, and the second preset number is 5.

[0066] In this embodiment, the dataset used is the publicly available dataset UK-DALE. UK-DALE records the appliance-level power data of five households, which includes various appliances with different power ratings for selection. All the appliance-level power data is the active power sampled every 6 seconds, while the aggregated power data is the apparent power sampled every 6 seconds, as well as the active and reactive powers sampled every 1 second. To facilitate subsequent model training, all the active powers are downsampled to a resolution of 6 seconds, and in the case where no active power is available, only the apparent power is selected.

[0067] Considering the sample size, the proportion of power consumption, and the representativeness of power characteristics, this embodiment selects the following five appliances for data analysis: kettle, refrigerator, washing machine, microwave oven, and dishwasher. These five appliances are distributed in at least three households in UK-DALE. The specific appliance distribution can be viewed in Table 1.

[0068] Table 1

[0069]

[0070] In this embodiment, the data of House No. 5 is used as the test set, and the data of other houses is used as the training set. In addition, since the washing machine and the microwave oven in House No. 4 share the same electricity meter, the power data of the washing machine and the microwave oven in House No. 4 will not participate in the training.

[0071] In the second step, determine the appliance turning-on time period for each target household; based on the obtained aggregated power dataset and appliance-level power dataset, use the Electric.get_activation function of non-intrusive load monitoring and the parameters provided by UK-DALE during the turning-on time period to obtain activation information, and then obtain a pair of aggregated power sequences and labeled power sequences corresponding to each time index.

[0072] In this embodiment, in order to provide training samples and labels, it is necessary to first determine the appliance turning-on time period. Use the Electric.get_activation function of NILMTK and the parameters provided by UK-DALE to obtain activation information, which results in a pair of aggregated power sequences from the aggregated data and labeled power sequences from the appliance-level data for each time index. Since the lengths of each pair of sequences are different, a constant window length is uniformly set for each appliance. Then, by randomly placing the window on the premise that the entire activation interval is included, training and test samples are generated. The parameters of the Electric.get_activation function are shown in Table 2:

[0073] Table 2

[0074]

[0075] The usage durations of different electrical appliances are shown in Table 3 as follows:

[0076] Table 3

[0077]

[0078] S22 preprocesses the aggregated power dataset and the appliance-level power dataset to obtain the preprocessed training set and test set.

[0079] In this embodiment, before starting to train the model, the following preprocessing is performed on the dataset.

[0080] 1) First, normalize the aggregated power dataset and the appliance-level power dataset to obtain the normalized aggregated power dataset and appliance-level power dataset.

[0081] In this embodiment, each pair of training samples and labels is normalized to facilitate network training. Specifically, all data in each pair is divided by the maximum power value of the clustering power sequence of the pair, so that the value range of all values is between 0 and 1. When generating the test set, the same normalized data processing method is also adopted.

[0082] 2) Then, based on the normalized aggregated power dataset and appliance-level power dataset, starting from the starting point of activation, use a window of a preset scale to divide the power data that does not contain the activation information of the target appliance, to obtain the updated aggregated power dataset and appliance-level power dataset, and use the updated appliance-level power dataset as the preprocessed test set.

[0083] In this embodiment, the window is set in advance by a window length, and starting from the starting point of activation, some training samples that do not contain the activation of the target appliance are divided to expand the learning ability of the network. When generating the test set, the same data processing process is also adopted.

[0084] 3) Filter the updated aggregated power dataset to obtain the filtered aggregated power dataset as the preprocessed training set.

[0085] Specifically, remove the updated aggregated power data of the preset scale with an activation length less than the first preset multiple, and the updated aggregated power data of the preset scale where the ratio of the aggregated power point to the maximum appliance power is greater than the third preset multiple and the activation length is greater than the second preset multiple, to obtain the filtered aggregated power dataset. Among them, the first preset multiple is less than the second preset multiple, and the second preset multiple is less than the third preset multiple.

[0086] In this embodiment, the training samples are filtered as follows: Samples with an activation length less than 1 / 3 of the window length, and samples with a ratio of the aggregated power point to the maximum electrical power exceeding three times and an activation length exceeding half of the window length are discarded and not included in the training set. It should be noted that this rule is not applied when generating the test set in order to obtain the most representative test results.

[0087] So far, this embodiment has obtained the preprocessed training set and test set for training the load identification model.

[0088] S3. Using the preprocessed training set and test set, train and test the load identification model to obtain a trained load identification model.

[0089] In this embodiment, when training the load identification model, the optimizer used is the Adam optimizer, which can adaptively accelerate the convergence of the network; the loss function used during the training process is the cross-entropy loss function, which shows better performance compared to the mean squared error and helps to improve the identification effect of the load identification model.

[0090] After obtaining the trained load identification model, the aggregated power data of the household to be identified can be directly input into the trained load identification model to obtain the load identification result of the household to be identified.

[0091] To verify the effectiveness of a sequence mapping load identification and decomposition method based on a multi-scale residual neural network of the present invention, comparative analysis is carried out with three existing models, namely BiLSTM and the denoising autoencoder, as well as the traditional convolutional network. Among them, the structural diagram of BiLSTM is as Figure 7 shown, the structural diagram of the denoising autoencoder is as Figure 8 shown, and the structural diagram of the traditional convolutional network is as Figure 9 shown. To improve the reliability of the verification results, for the traditional convolutional network, the Adam optimizer is used; for dBiLSTM and the denoising autoencoder, stochastic gradient descent with Nesterov momentum is used, and the momentum value is 0.9; the mean squared error is used as the loss function for training all three existing models.

[0092] During the verification process, the definitions of the indicators for evaluating the identification performance include: true positive (TP, True Positive), false positive (FP, False Positive), true negative (TN, True Negative), and false negative (FN, False Negative). Denote the actual power of the electrical appliance at time t as , and the estimated power as . The calculation formulas for the four static evaluation indicators can be:

[0093] ;

[0094] ;

[0095] ;

[0096] ;

[0097] where Recall represents the recall rate, Precision represents the precision rate, F1 Score represents the F1 score, Mean Absolute Error represents the mean absolute error, T represents the time period, represents the absolute value function.

[0098] Among them, the recall rate and precision rate are usually regarded as intermediate indicators, because even if high scores are obtained on these two indicators, it may lead to a poor F1 score, while the F1 score can reflect the overall performance of the network in classifying the electrical switch states; on the other hand, the mean absolute error (MAE, Mean Absolute Error) reflects the performance of the network in more specific power deviation aspects.

[0099] The verification process is carried out in the experimental environment shown in Table 4, and the results of testing different types of electrical appliances using different models are shown in Figure 10 and Figure 11 in, in Figure 10 a represents the test result of the electric kettle, b represents the test result of the microwave oven, c represents the test result of the refrigerator, d represents the test result of the dishwasher, e represents the test result of the washing machine; the number of parameters of each model is counted using Tensorflow Profiler, and the time.time() function is used to measure the prediction time of each sample; the parameter quantity and time results of the model are shown in Tables 5 and 6 respectively.

[0100] Table 4

[0101]

[0102] Table 5

[0103]

[0104] Table 6

[0105]

[0106] All MAE results for each appliance are divided by the maximum MAE value of the corresponding appliance, so that they are normalized to the range [0, 1] and clearly shown in the figure. The denominator values are shown in Table 7:

[0107] Table 7

[0108]

[0109] Appliances with simple power characteristics, such as kettles and dishwashers, usually exhibit better identification performance than other appliances; while appliances with more difficult-to-capture characteristics, such as microwave ovens, usually exhibit poorer identification performance. Specifically for each indicator, the kettle has the best F1 score among the five models, but the MAE is relatively poor. Because the activation of the kettle is concentrated in a smaller window and the power is relatively high (exceeding 2000W), after the output of the neural network passes through the sigmoid activation, small errors will be amplified to a very high level. Similarly, the refrigerator (with lower power) has better MAE performance. The power thresholds of washing machines and dishwashers are relatively low, so there are fewer false negative or false positive points, resulting in a recall rate that is usually higher than the precision rate. For more comprehensive indicators such as F1 and MAE, the situation is different.

[0110] From the verification results, it can be seen that the traditional convolutional network performs better than the autoencoder on kettles and washing machines, but worse on dishwashers. The performance of BiLSTM is mediocre, with a large MAE, and occasionally has a better F1 score on the three appliances. At the same time, the proposed multi-scale model performs better on all five appliances. Comparing the residual network based on dilated convolution with other networks based on CNN (Convolutional Neural Network), its F1 and MAE results on the five appliances exceed those of other models. It can be seen that the residual network and dilated convolution can solve the problems of insufficient receptive field and performance degradation in ordinary convolution. Therefore, a sequence mapping load identification and decomposition method based on a multi-scale residual neural network proposed in the present invention is faster and better than RNN in terms of identification performance, and compared with other CNN-based models, while reducing the number of parameters, it also reduces the time cost.

[0111] In summary, the network structure proposed in this embodiment is a load identification model of a multi-scale residual network based on dilated convolution. Compared with traditional models, the dilated convolution and multi-scale structure of this load identification model can be successfully applied to time series load identification, effectively solving the NILM problem. Moreover, the large receptive field generated by dilated convolution can obtain more information from the background to assist identification, while the multi-scale structure enables the model to achieve optimal performance when facing electrical appliances with complex power characteristics. In addition, this embodiment shows good potential in terms of small size, helping to shorten the time required for prediction. In subsequent developments, this load identification model can be extended to a sequence-to-point regression task to directly estimate the identified energy consumption.

[0112] The above-described embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments or perform equivalent replacements for some of the technical features. However, such modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the various embodiments of the present invention and should all be included within the protection scope of the present invention.

Claims

1. A sequence mapping load identification and decomposition method based on a multi-scale residual neural network, characterized in that Including the following steps: Construct a load identification model, and the network structure of the load identification model is a multi-scale residual network based on dilated convolution; Obtain the aggregated power dataset and the appliance-level power dataset in the public dataset, and preprocess the aggregated power dataset and the appliance-level power dataset to obtain the preprocessed training set and test set; Use the preprocessed training set and test set to train and test the load identification model to obtain a trained load identification model; The multi-scale residual network based on dilated convolution includes: dilated convolution residual blocks; The input of the dilated convolution residual block sequentially passes through two convolutional layers and their corresponding normalization layers, dropout layers, and non-linear activation functions; then, the original input is added to the output over time through a shortcut connection to form the final output of the dilated convolution residual block; among them, the two convolutional layers in the dilated convolution residual block share the same dilation rate; The calculation formula for the receptive field of the dilated convolution residual block is: ； In the formula, S represents the receptive field of the dilated convolution residual block, N represents the number of dilated convolutions, n represents the number of network layers of the dilated convolution, and k represents the convolution kernel size.

2. The sequence mapping load identification and decomposition method based on a multi-scale residual neural network according to claim 1, characterized in that The obtaining of the aggregated power dataset and the appliance-level power dataset in the public dataset includes: Obtain the UK-DALE public dataset, and select the aggregated power data and the appliance-level power data of each target appliance in the first preset number of target households in the UK-DALE public dataset at the same sampling frequency; among them, the first preset number of target households are households with a second preset number of the same type of appliances, and the target appliances are the same type of appliances; Determine the appliance-on time period of each target household; based on the obtained aggregated power dataset and the appliance-level power dataset, use the Electric.get_activation function of non-intrusive load monitoring and the parameters provided by UK-DALE to obtain activation information during the on time period, so as to obtain a pair of aggregated power sequences and labeled power sequences corresponding to each time index.

3. The method for sequence mapping load identification and decomposition based on a multi-scale residual neural network according to claim 1, characterized in that, The preprocessing of the aggregated power dataset and the appliance-level power dataset to obtain the preprocessed training set and test set includes: First, perform normalization processing on the aggregated power dataset and the appliance-level power dataset to obtain the normalized aggregated power dataset and the appliance-level power dataset; Then, based on the normalized aggregated power dataset and the appliance-level power dataset, use a window of a preset scale to start from the starting point of activation, and divide the power data that does not contain the activation information of the target appliance to obtain the updated aggregated power dataset and the appliance-level power dataset, and use the updated appliance-level power dataset as the preprocessed test set; Perform filtering processing on the updated aggregated power dataset to obtain the filtered aggregated power dataset as the preprocessed training set.

4. The method for identifying and decomposing sequence mapping load based on a multi-scale residual neural network according to claim 3, wherein The performing of the filtering processing on the updated aggregated power dataset to obtain the filtered aggregated power dataset includes: Remove the updated aggregated power data of the preset scale with an activation length less than the first preset multiple, and the updated aggregated power data of the preset scale with a ratio of the aggregated power point to the maximum electrical power greater than the third preset multiple and an activation length greater than the second preset multiple, to obtain a filtered aggregated power data set; wherein, the first preset multiple is less than the second preset multiple, and the second preset multiple is less than the third preset multiple.

Images