CN114510992A - Equipment switch state detection method based on deep learning - Google Patents
Equipment switch state detection method based on deep learning Download PDFInfo
- Publication number
- CN114510992A CN114510992A CN202111616875.0A CN202111616875A CN114510992A CN 114510992 A CN114510992 A CN 114510992A CN 202111616875 A CN202111616875 A CN 202111616875A CN 114510992 A CN114510992 A CN 114510992A
- Authority
- CN
- China
- Prior art keywords
- data
- network
- equipment
- sequence
- load
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000013135 deep learning Methods 0.000 title claims abstract description 18
- 238000001514 detection method Methods 0.000 title claims abstract description 17
- 238000000034 method Methods 0.000 claims abstract description 33
- 238000000354 decomposition reaction Methods 0.000 claims abstract description 23
- 238000012545 processing Methods 0.000 claims abstract description 23
- 230000005236 sound signal Effects 0.000 claims abstract description 4
- 238000012549 training Methods 0.000 claims description 25
- 230000006870 function Effects 0.000 claims description 15
- 238000011156 evaluation Methods 0.000 claims description 12
- 230000002776 aggregation Effects 0.000 claims description 9
- 238000004220 aggregation Methods 0.000 claims description 9
- 238000012544 monitoring process Methods 0.000 claims description 8
- 238000013528 artificial neural network Methods 0.000 claims description 7
- 238000005070 sampling Methods 0.000 claims description 7
- 238000007781 pre-processing Methods 0.000 claims description 6
- 238000005265 energy consumption Methods 0.000 claims description 5
- 238000005259 measurement Methods 0.000 claims description 4
- 238000005406 washing Methods 0.000 claims description 4
- 238000004364 calculation method Methods 0.000 claims description 3
- 230000000694 effects Effects 0.000 claims description 3
- 230000007787 long-term memory Effects 0.000 claims description 3
- 239000011159 matrix material Substances 0.000 claims description 3
- 238000012216 screening Methods 0.000 claims description 3
- 230000006403 short-term memory Effects 0.000 claims description 3
- 238000013527 convolutional neural network Methods 0.000 description 3
- 238000011160 research Methods 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 238000012952 Resampling Methods 0.000 description 2
- 125000004122 cyclic group Chemical group 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 108091071337 20 family Proteins 0.000 description 1
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 description 1
- 238000004851 dishwashing Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 230000001052 transient effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/06—Energy or water supply
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Economics (AREA)
- Strategic Management (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Human Resources & Organizations (AREA)
- Health & Medical Sciences (AREA)
- Marketing (AREA)
- Bioinformatics & Cheminformatics (AREA)
- General Business, Economics & Management (AREA)
- General Health & Medical Sciences (AREA)
- Tourism & Hospitality (AREA)
- Evolutionary Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Game Theory and Decision Science (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Entrepreneurship & Innovation (AREA)
- Development Economics (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Public Health (AREA)
- Water Supply & Treatment (AREA)
- Primary Health Care (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention belongs to the technical field of non-invasive load identification, and particularly relates to an equipment switch state detection method based on deep learning. The method improves a sequence-point processing mode, namely, the prediction of a single point is changed into the prediction of a certain target sequence length, the target sequence is closer to a middle point and a value near the middle point, and the prediction efficiency is improved on the premise of ensuring better prediction performance; meanwhile, a network module for modeling an original audio signal is introduced, and a receiving domain of the network is expanded by adopting stack hole convolution, so that efficient modeling and prediction output of a long data sequence are realized; the network module includes: the regression network is used for carrying out load decomposition on the aggregated load data to obtain load data of different equipment; and classifying the network to obtain the load state of the corresponding equipment, and integrally outputting and integrating the information of the two parts to obtain a more accurate equipment switch state prediction result. The invention has remarkable advantages in the aspect of non-invasive equipment switch state identification.
Description
Technical Field
The invention belongs to the technical field of non-invasive load identification, and particularly relates to an equipment switch state detection method based on deep learning.
Background
Non-intrusive load identification is to identify the current operating state of the equipment by resolving the power of the corresponding equipment from the total power of the electric meter, so that how to deduce the power load of the specific equipment, perform power load resolution and identification is a main research target at present. At present, deep learning is deep into various industries, various problems are solved by utilizing a neural network, and how to improve the accuracy of load identification by utilizing a deep learning related method is the hot of current research.
The total power data is a section of sequence data, and the current processing modes of the sequence data mainly comprise sequence-sequence and sequence-point, wherein the sequence-sequence is data with the same length as an input sequence in prediction, and the sequence-point is a midpoint value of the prediction sequence. Researches find that the middle point and the value near the middle point in the prediction result are more accurate for the sequence value, and the sequence-point has a better prediction result than the sequence-point, but the efficiency of predicting the middle point value only for a section of sequence data is lower, and a more efficient processing mode is needed. The traditional neural network is divided into a convolutional neural network and a cyclic neural network, when an input sequence is long, the receiving domain of the network is increased, the square expression of the computational complexity of the convolutional neural network is increased, the computational complexity of the cyclic neural network is linearly increased, but only sequential computation can be performed, and effective parallel computation processing cannot be performed, so that long-sequence modeling is always a challenge faced by the traditional neural network.
Disclosure of Invention
The invention aims to provide a device switch state detection method based on deep learning, which is suitable for a long-sequence modeling network and improves efficiency on the premise of ensuring a better prediction result on a sequence value.
The invention improves the sequence-point processing mode, adopts an improved mode to change the prediction of a single point into the prediction of a certain target sequence length, the target sequence is closer to the middle point and the value near the middle point, and the prediction efficiency is improved on the premise of ensuring better prediction performance. Meanwhile, a network structure for modeling an original audio signal is introduced, a receiving domain of the network is expanded by adopting stack hole convolution, and the whole network can perform parallel computing processing due to an improved sequence-point processing mode, so that efficient modeling and prediction output of a long data sequence can be realized. According to the invention, by using an improved deep learning method, the constructed network module is divided into two main parts, one part utilizes the characteristics of a regression network to carry out load decomposition on aggregated load data to obtain load data of different equipment, the other part utilizes the result of a classification network to obtain the load state of corresponding equipment, the whole output of the network integrates the information of the two parts, a more accurate equipment switch state prediction result is obtained, and the method has superiority in the aspect of non-invasive equipment switch state identification.
The invention provides a device switch state detection method based on deep learning, which comprises the following specific steps:
step 1: determining input data;
the load data of the equipment aimed at by the load decomposition task is mainly divided into two types, one is high-frequency sampling data which is mainly expressed as signal waveforms sampled for thousands of times or tens of thousands of times per second and comprises the load data of equipment in a state switching between a steady state and a transient state and a continuous operation state; and secondly, low-frequency sampling data mainly represents current, voltage, power and other related data when the equipment runs, and the sampling frequency is very low or is the root mean square value of data with higher sampling frequency.
The network adopted by the invention aims at the situation that the data is the total power recorded when the equipment runs and is the data sampled at low frequency, and the data with lower sampling rate is adopted to ensure that the method can carry out long-time load monitoring. In a practical use scene, the use time of most household appliances is not distributed uniformly relative to the total energy consumption time, and a long-time monitoring is needed to more accurately identify the on and off states of the specific equipment from the aggregated load data, so that the method provided by the invention has more use scenes in households.
The input data of the network is obtained from long sequence data in a sliding window mode, and a modified mode is adopted in a sequence-point processing mode, so that the prediction of a single point is changed into the prediction of a certain target sequence length, and the target sequence is closer to the middle point and the value near the middle point. The reception field is marked as L, the target field is marked as r, the length of data input each time is L + r-1 node data, the size of the reception field L is 255, the size of the target field r is 100, and the data input each time is 354 node data. And simultaneously r is the size of a sliding window, a plurality of data segments are divided by taking r as a step length according to a data sequence with the time length of T, the data segments are output as sequence data with the same length as the T, and the output result of each time node in the sequence is empty or in an opening/closing state of a certain device.
Step 2: preprocessing data;
step 2.1: because the invention processes long-time sequence data, the data in the circuit data monitoring can be missed in part of time due to some environmental factors, and therefore, a preprocessing operation needs to be carried out on the data. Firstly, adopting forward filling for data loss caused by a frequency-emission problem, namely copying data recorded after the missing data segment to fill the missing segment, wherein the length of the filled data segment does not exceed three minutes; and meanwhile, zero padding is adopted under the condition of data missing caused by equipment switching, namely the missing data of the section is padded into 0, and the length of the padded data exceeds three minutes.
Step 2.2: meanwhile, training data are required to be prepared to train the neural network, the training data not only comprise aggregated load data of long-time sequences, but also comprise specific equipment load data corresponding to the time sequences, so that the performance of the regression network is measured by adopting corresponding evaluation indexes in the training process, network parameters are adjusted, and the load decomposition capability of the network is improved. Processing one, deleting the data segment of which the load value of the specific equipment is greater than the corresponding aggregation load value, so as to reduce the influence of the data segment with the monitoring error on the network; and the second processing is to perform binarization processing on the equipment load data sequence through the determined equipment power threshold, record the time period when the load value is greater than the threshold as an on state, and record the time period when the load value is greater than the threshold as an off state, so that equipment on-off state records corresponding to the time sequence are obtained, and the performance of the classification network is measured by adopting corresponding evaluation indexes in the training process, so that network parameters are adjusted, and the state detection capability of the network is improved.
And step 3: determining a network structure;
the network structure of the invention mainly comprises two parts, namely a regression sub-network and a classification sub-network, wherein the main parts of the two parts are cavity convolution residual modules, the output results of the two parts are finally subjected to a screening operation to obtain final sequence data with the same input length, and the output result of each time node in the sequence is empty or the on/off state of a certain device.
Step 3.1: the structure of the hole convolution residual module is shown in fig. 2, wherein a hole convolution layer is formed by a layer of hole convolution, and a super parameter is added to a hole rate, namely the number of intervals between convolution kernels, compared with the traditional convolution. The output obtained after the input passes through the cavity convolution layer is divided into two parts, one part passes through a tanh function, the other part passes through a sigmoid function, then matrix corresponding bit multiplication operation is carried out on the calculation results of the two parts, the convolution is carried out by one layer of 1x1, similarly, the output is divided into two parts, one part is directly output as the step output, and the other part is added with the initial input of the residual error module to obtain the residual error output.
Step 3.2: the structure of the regression subnetwork is shown in fig. 3, and 6 layers of cavity convolution residual modules are stacked from the input end to the output end, and as can be known from the structural description of the cavity convolution residual modules, except the last layer of residual module, the residual output of each layer of module is input to the next layer of module, and simultaneously except the first layer of residual module, the step output of each layer of module is added with the step output of the previous layer and transmitted to the next layer. And finally, the step output result of the addition of the last layer is subjected to ReLU function, convolution of 1x1 and full connection layer to obtain the final output of the regression subnetwork.
Step 3.3: the structure of the classification sub-network is shown in fig. 3, the first half of the network is the same as the regression sub-network, 6 layers of cavity convolution residual modules are stacked from the input end to the output end, the structural description of the cavity convolution residual modules shows that the residual output of each layer of module is input to the next layer of module except the last layer of residual module, and the step output of each layer of module is added with the step output of the previous layer and transmitted to the next layer except the first layer of residual module. And finally, the step output result of the addition of the last layer is subjected to a ReLU function and then a layer of 1x1 convolution. Two layers of gate control circulation units are connected in sequence, and the gate control circulation units are long and short term memory networks with simplified parameter quantity and are used for obtaining dependence on information before and after the sequence. And then, connecting a full connection layer, and finally obtaining the final output of the classification sub-network through a Sigmoid function.
And 4, step 4: determining an evaluation index, and training a network model;
step 4.1: the regression network module realizes load decomposition of aggregated load data, the standard of evaluation in the network training process is the error between the generated data and the real data, the average absolute error MAE and the normalized signal aggregation error are specifically adopted, and the smaller the numerical value is, the better the model decomposition effect is. Let T be the number of time nodes,
and ytRespectively predicted power and real power at time node t,
r＝∑tytrespectively representing the sum of predicted power and real power of a certain device in a certain time period, comprising:
step 4.2: the classification network module realizes the detection of the light-on state of the equipment, the data adopted in the training process of the network is more inclined to the load data record of the family, the service time of most household appliances is distributed quite unevenly relative to the total energy consumption time, therefore, the invention adopts the measurement F1 score based on the state to measure the capability of the module for predicting the equipment state, and in each time step, if the power supply of the equipment exceeds the preset power threshold value, the equipment is considered to be in the on state. The power threshold is determined according to the type of appliance and general power representatives, for example 50, 2000, 200, 20 and 10 (watts) for refrigerators, kettles, microwave ovens, washing machines and dishwashers, respectively. The F1 score can be regarded as a harmonic mean value of the precision rate P and the recall rate R, and if TP is positive sample number predicted correctly, FP is negative sample number predicted positively, and FN is positive sample number predicted negatively, then:
and 5: and obtaining a load decomposition result and judging the on-off state of the equipment.
The method of the invention finally detects the on-off state of the equipment, and according to the network model obtained by the steps and the training means, the regression network module can carry out load decomposition on the aggregated load signal of the equipment, the network receives input power data and outputs load signals with equal time duration, the output of the part is the reconstructed target equipment load signal, the current operation state of the equipment can be judged according to the load signal, and the part exceeding the set threshold value can be determined to start the equipment to operate in the time period. The load decomposition result obtained by the regression network module is an aid to the classification result, the error condition that the predicted load value of some specific equipment is larger than the corresponding aggregation load value is filtered, and the state prediction result in the period of time is recorded as an error, so that the accuracy of classification result prediction is improved.
The invention improves the sequence-point processing mode, adopts an improved mode to change the prediction of a single point into the prediction of a certain target sequence length, and the target sequence is closer to the middle point and the value near the middle point, thereby improving the prediction efficiency on the premise of ensuring better prediction performance. Meanwhile, a network structure for modeling an original audio signal is introduced, a receiving domain of the network is expanded by adopting stacking expansion convolution, and the whole network can perform parallel computing processing due to an improved sequence-point processing mode, so that efficient modeling and prediction output of a long data sequence can be realized. According to the invention, by using an improved deep learning method, the constructed network module is divided into two main parts, one part utilizes the characteristics of a regression network to carry out load decomposition on aggregated load data to obtain load data of different equipment, the other part utilizes the result of a classification network to obtain the load state of corresponding equipment, the whole output of the network integrates the information of the two parts, a more accurate equipment switch state prediction result is obtained, and the method has remarkable superiority in the aspect of non-invasive equipment switch state recognition.
Drawings
Fig. 1 is a flowchart of an apparatus switch state detection method based on deep learning according to the present invention.
FIG. 2 is a diagram of a hole convolution residual block according to the present invention.
Fig. 3 is a diagram of a network architecture of the present invention.
Detailed Description
The technical scheme of the invention is explained in detail by combining the drawings and the embodiment.
Example (b): a method for detecting the on-off state of equipment based on deep learning is disclosed, the flow chart of which is shown in figure 1, and the method comprises the following specific steps:
step 1: determining input data;
step 2: preprocessing data;
and step 3: determining a network structure;
and 4, step 4: determining an evaluation index, and training a network model;
and 5: and obtaining a load decomposition result, and judging the on-off state of the equipment.
The steps are further specifically described below.
1. Determining input data
The network adopted by the invention aims at the data which is the total power recorded when the equipment runs and is about low-frequency sampled data, and the data with lower sampling rate is adopted to enable the method to carry out long-time load monitoring. Since the usage time of most household appliances is not distributed uniformly relative to the total energy consumption time in a real usage scenario, and a long monitoring time is required to identify the on/off state of a specific device more accurately from the aggregated load Data, the present embodiment uses the existing common Data set REFIT (David multiray, line Stankovic, and vladimi stankovic.2017.an electronic local measurement Data set of a unified Kingdom house from a two-year long distance test. scientific Data (2017)) for network training and testing, the Data set collects load Data from 20 households in the United Kingdom, the collected Data is recorded about once every 8 seconds, and the total recording time exceeds 2 years.
The input data of the network is acquired from long sequence data in a sliding window mode, a reception field is recorded as L, a target domain is recorded as r, the length of the data input each time is L + r-1 node data, the size of the reception field L is 255, the size of the target domain r is 100, and the data input each time is 354 node data. Meanwhile, r is the size of a sliding window, and for a data sequence with the time length of T, a plurality of data segments are divided by taking r as a step length.
2. Data pre-processing
For preprocessing of the REFIT data set, first, resampling is performed on all continuously recorded data to reduce the fluctuation influence of the time interval between original data, and the resampling result takes 10 seconds as an interval to obtain about 93976578 data nodes. For the data missing condition, firstly, the data missing caused by the frequency-radio problem adopts forward filling, namely, the data recorded after the missing data segment is copied and filled into the missing segment, and the length of the filled data segment is not more than three minutes; and meanwhile, zero padding is adopted under the condition of data missing caused by equipment switching, namely the missing data of the section is padded into 0, and the length of the padded data exceeds three minutes.
The on-off state detection of one device is carried out on four electric appliances, namely a kettle, a microwave oven, a washing machine and a dish washing machine, in 20 family data, wherein 10 family data are randomly selected as training data of a network, and 4 test data are randomly selected from the rest family data as the network. The first step of processing the data is deleting a data segment of which the load value of the specific equipment is greater than the corresponding aggregate load value, so that the influence of the data segment with the monitoring error on the network is reduced; and the second processing is to perform binarization processing on the equipment load data sequence through the determined equipment power threshold, record the time period when the load value is greater than the threshold as an on state, and record the time period when the load value is greater than the threshold as an off state, so that equipment on-off state records corresponding to the time sequence are obtained, and the performance of the classification network is measured by adopting corresponding evaluation indexes in the training process, so that network parameters are adjusted, and the state detection capability of the network is improved.
3. Determining network structure
The network structure of the invention mainly comprises two parts, namely a regression sub-network and a classification sub-network, wherein the main parts of the two parts are cavity convolution residual modules, the output results of the two parts are finally subjected to a screening operation to obtain final sequence data with the same input length, and the output result of each time node in the sequence is empty or the on/off state of a certain device.
Step 3.1: the structure of the hole convolution residual module is shown in fig. 2, wherein a hole convolution layer is formed by a layer of hole convolution, and a super parameter is added to a hole rate, namely the number of intervals between convolution kernels, compared with the traditional convolution. The output obtained after the input passes through the cavity convolution layer is divided into two parts, one part passes through a tanh function, the other part passes through a sigmoid function, then matrix corresponding bit multiplication operation is carried out on the calculation results of the two parts, the convolution is carried out by one layer of 1x1, similarly, the output is divided into two parts, one part is directly output as the step output, and the other part is added with the initial input of the residual error module to obtain the residual error output.
Step 3.2: the structure of the regression subnetwork is shown in fig. 3, and 6 layers of cavity convolution residual modules are stacked from the input end to the output end, and as can be known from the structural description of the cavity convolution residual modules, except the last layer of residual module, the residual output of each layer of module is input to the next layer of module, and simultaneously except the first layer of residual module, the step output of each layer of module is added with the step output of the previous layer and transmitted to the next layer. And finally, the step output result of the last layer of addition is subjected to ReLU function, convolution by 1x1 and full connection layer to obtain the final output of the regression subnetwork.
Step 3.3: the structure of the classification sub-network is shown in fig. 3, the first half of the network is the same as the regression sub-network, 6 layers of cavity convolution residual modules are stacked from the input end to the output end, the structural description of the cavity convolution residual modules shows that the residual output of each layer of module is input to the next layer of module except the last layer of residual module, and the step output of each layer of module is added with the step output of the previous layer and transmitted to the next layer except the first layer of residual module. And finally, the step output result of the addition of the last layer is subjected to a ReLU function and then a layer of 1x1 convolution. Two layers of gate control circulation units are connected in sequence, and the gate control circulation units are long and short term memory networks with simplified parameter quantity and are used for obtaining dependence on information before and after the sequence. And then, connecting a full connection layer, and finally obtaining the final output of the classification sub-network through a Sigmoid function.
4. Determining evaluation index and training network model
Step 4.1: the regression network module realizes load decomposition of aggregated load data, the standard of evaluation in the network training process is the error between the generated data and the real data, the average absolute error MAE and the normalized signal aggregation error are specifically adopted, and the smaller the numerical value is, the better the model decomposition effect is. Let T be the number of time nodes,
and ytRespectively predicted power and real power at time node t,
r＝∑tytrespectively representing the sum of predicted power and real power of a certain device in a certain time period, comprising:
step 4.2: the classification network module realizes the detection of the light-on state of the equipment, the data adopted in the training process of the network is more inclined to the load data record of the family, the service time of most household appliances is distributed quite unevenly relative to the total energy consumption time, therefore, the invention adopts the measurement F1 score based on the state to measure the capability of the module for predicting the equipment state, and in each time step, if the power supply of the equipment exceeds the preset power threshold value, the equipment is considered to be in the on state. The power threshold is determined according to the type of appliance and general power representatives, and the power thresholds of the four appliances in this embodiment, the kettle, the microwave oven, the washing machine and the dishwasher, are 2000, 200, 20 and 10 (watts), respectively. The F1 score can be regarded as a harmonic mean value of the precision rate P and the recall rate R, and if TP is positive sample number predicted correctly, FP is negative sample number predicted positively, and FN is positive sample number predicted negatively, then:
5. obtaining the load decomposition result and judging the on-off state of the equipment
The method of the invention finally detects the on-off state of the equipment, and according to the network model obtained by the steps and the training means, the regression network module can carry out load decomposition on the aggregated load signal of the equipment, the network receives input power data and outputs load signals with equal time duration, the output of the part is the reconstructed target equipment load signal, the current operation state of the equipment can be judged according to the load signal, and the part exceeding the set threshold value can be determined to start the equipment to operate in the time period. The load decomposition result obtained by the regression network module is an aid to the classification result, the error condition that the predicted load value of some specific equipment is larger than the corresponding aggregation load value is filtered, and the state prediction result in the period of time is recorded as an error, so that the accuracy of classification result prediction is improved.
In this case, the method is compared with the conventional CNN and RNN methods in experiments, and the accuracy of judging the on-off state of the electrical equipment according to the load decomposition result is compared, and three indexes are adopted for evaluation as shown in Table 1. The method has excellent performance on most indexes, and proves the superiority of the network in processing long sequence data.
TABLE 1 comparison of switch State detection results
Claims (6)
1. A device switch state detection method based on deep learning is characterized in that improvement is carried out on the basis of a sequence-point processing mode, namely, prediction of a single point is changed into prediction of a certain target sequence length, the target sequence is closer to a middle point and values near the middle point, and the prediction efficiency is improved on the premise of ensuring better prediction performance; meanwhile, a network module structure for modeling an original audio signal is introduced, a receiving domain of the network is expanded by adopting stack hole convolution, and the whole network carries out parallel computing processing due to an improved sequence-point processing mode, so that efficient modeling and prediction output of a long data sequence can be realized; the built network module is divided into two parts: the regression network is used for carrying out load decomposition on the aggregated load data to obtain load data of different equipment; the method comprises the steps that firstly, a network is classified, the load state of corresponding equipment is obtained according to the result of the network classification, and the whole network outputs and synthesizes information of two parts, so that a more accurate equipment switch state prediction result is obtained.
2. The method for detecting the on-off state of the equipment based on the deep learning as claimed in claim 1, comprises the following specific steps:
step 1: determining input data;
the method comprises the steps that on the basis of the total power recorded when the adopted network model aims at data, low-frequency sampling data are adopted for input data, wherein the data comprise current, voltage and power related data when the equipment operates;
the input data of the network model is acquired from long sequence data in a sliding window mode, and particularly, an improved mode is adopted on a sequence-point processing mode, namely, the prediction of a single point is changed into the prediction of a certain target sequence length, and the target sequence is closer to the midpoint and values near the midpoint;
step 2: preprocessing data;
step 2.1: firstly, for data loss caused by a frequency-radio problem, forward filling is adopted, namely, data recorded after the data segment is lost is copied and filled into the lost segment, and the length of the filled data segment is not more than three minutes; for data loss caused by equipment switching, zero padding is adopted, namely the missing data of the section is padded into 0, and the length of the padded data exceeds three minutes;
step 2.2: preparing training data for training a neural network, wherein the training data comprises aggregation load data of a long-time sequence and specific equipment load data corresponding to the time sequence, so that the performance of the regression network is measured by adopting corresponding evaluation indexes in the training process, thereby adjusting network parameters and improving the load decomposition capability of the network;
and step 3: determining a network structure;
the network structure mainly comprises two parts, namely a regression sub-network and a classification sub-network, wherein the main parts of the two parts are a cavity convolution residual error module, the output results of the two parts are finally subjected to a screening operation to finally obtain sequence data with the same input length, and the output result of each time node in the sequence is empty or is in an on/off state of a certain device;
and 4, step 4: determining an evaluation index, and training a network model;
step 4.1: the regression network module is used for carrying out load decomposition on the aggregated load data, the standard evaluated in the network training process is the error between the generated data and the real data, specifically, the average absolute error MAE and the normalized signal aggregation error are adopted, and the smaller the numerical value is, the better the model decomposition effect is; let T be the number of time nodes,
and ytRespectively predicted power and real power at time node t,
r＝∑tytrespectively representing the sum of predicted power and real power of a certain device in a certain time period, comprising:
step 4.2: the classification network module is used for detecting the light-on state of the equipment, data adopted in the training process of the network is load data records of a family, the service time of most household electrical appliances is quite uneven relative to the total energy consumption time, so that the capacity of the module for predicting the equipment state is measured by adopting a state-based measurement F1 score, and in each time step, if the power supply of the equipment exceeds a preset power threshold value, the equipment is considered to be in an on state; the F1 score can be regarded as a harmonic mean value of the precision rate P and the recall rate R, and if TP is positive sample number predicted correctly, FP is negative sample number predicted positively, and FN is positive sample number predicted negatively, then:
and 5: obtaining a load decomposition result, and judging the on-off state of the equipment;
inputting power data of the trained network model into a regression network, outputting load signals with equal time length, namely reconstructed target equipment load signals, judging the current running state of the equipment according to the load signals, and determining that the equipment is started to run in the time period if the power data of the trained network model exceeds a set threshold; the load decomposition result obtained by the regression network module is an aid to the classification result of the classification network module, the error condition that the predicted load value of some specific equipment is larger than the corresponding aggregation load value is filtered, and the state prediction result of the period of time is marked as an error, so that the accuracy of classification result prediction is improved.
3. The method for detecting the on-off state of the device based on deep learning of claim 2, wherein the step 1 is to obtain the input data from the long sequence data by means of a sliding window, specifically: the reception field is marked as L, the target domain is marked as r, and the length of data input each time is L + r-1 node data; the size of the receptive field L is 255, the size of the target field r is 100, and data input each time is 354 node data; and simultaneously r is the size of a sliding window, a plurality of data segments are divided by taking r as a step length according to a data sequence with the time length of T, the data segments are output as sequence data with the same length as the T, and the output result of each time node in the sequence is empty or in an opening/closing state of a certain device.
4. The method for detecting the on-off state of the device based on the deep learning of claim 2, wherein the step 2.2 of preparing the training data comprises: deleting the data segment of which the load value of the specific equipment is greater than the corresponding aggregation load value, and reducing the influence of the data segment with the monitoring error on the network; carrying out binarization processing on the equipment load data sequence through the determined equipment power threshold, recording the time period when the load value is greater than the threshold as an open state, and otherwise, recording the time period as a closed state, so as to obtain an equipment switch state record corresponding to the time sequence; and in the training process, the performance of the classification network is measured by adopting the corresponding evaluation indexes, so that the network parameters are adjusted, and the state detection capability of the network is improved.
5. The deep learning-based device switch state detection method according to claim 2, wherein the hole convolution residual module in step 3 is characterized in that a hole convolution layer is formed by a layer of hole convolution, and a super parameter is added to a traditional convolution, namely a hole rate, namely the number of intervals between convolution kernels; the output obtained after the input passes through the cavity convolution layer is divided into two parts, one part passes through a tanh function, the other part passes through a sigmoid function, then matrix corresponding bit multiplication operation is carried out on the calculation results of the two parts, the convolution is carried out by one layer of 1x1, the output is also divided into two parts, one part is directly output as step output, and the other part is added with the initial input of a residual error module to obtain residual error output;
the regression subnetwork firstly stacks 6 layers of cavity convolution residual modules from an input end to an output end, the structure of the cavity convolution residual modules is known, the last layer of residual module is removed, the residual output of each layer of module is input to the next layer of module, the first layer of residual module is removed, and the step output of each layer of module is added with the step output of the previous layer and transmitted to the next layer; the final layer of added stride output results are firstly subjected to a ReLU function, then subjected to a layer of 1x1 convolution, and then subjected to a layer of full connection layer to obtain the final output of the regression subnetwork;
the classification sub-network comprises a first half part of a network and a regression sub-network, wherein 6 layers of cavity convolution residual modules are stacked from an input end to an output end, the structure of the cavity convolution residual modules is known, the residual output of each layer of module is input into the next layer of module except the last layer of residual module, and the step output of each layer of module is added with the step output of the previous layer and transmitted to the next layer except the first layer of residual module; the step output result of the last layer of addition is firstly processed by a ReLU function and then is convoluted by a layer of 1x 1; two layers of gate control circulation units are connected in sequence, and the gate control circulation units are long and short term memory networks with simplified parameter quantity and are used for obtaining dependence on information before and after the sequence; and then, connecting a full connection layer, and finally obtaining the final output of the classification sub-network through a Sigmoid function.
6. The deep learning based device switch state detection method of claim 2, wherein the power threshold in step 4.2 is determined according to the type of appliance and general power representation, and the power threshold for refrigerator, kettle, microwave oven, washing machine and dishwasher is 50, 2000, 200, 20 and 10, watt respectively.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111616875.0A CN114510992A (en) | 2021-12-27 | 2021-12-27 | Equipment switch state detection method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111616875.0A CN114510992A (en) | 2021-12-27 | 2021-12-27 | Equipment switch state detection method based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114510992A true CN114510992A (en) | 2022-05-17 |
Family
ID=81547404
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111616875.0A Pending CN114510992A (en) | 2021-12-27 | 2021-12-27 | Equipment switch state detection method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114510992A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116504005A (en) * | 2023-05-09 | 2023-07-28 | 齐鲁工业大学(山东省科学院) | Perimeter security intrusion signal identification method based on improved CDIL-Bi-LSTM |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110376457A (en) * | 2019-06-28 | 2019-10-25 | 同济大学 | Non-intrusion type load monitoring method and device based on semi-supervised learning algorithm |
| CN112033463A (en) * | 2020-09-02 | 2020-12-04 | 哈尔滨工程大学 | Nuclear power equipment state evaluation and prediction integrated method and system |
| CN112434799A (en) * | 2020-12-18 | 2021-03-02 | 宁波迦南智能电气股份有限公司 | Non-invasive load identification method based on full convolution neural network |
| CN112904220A (en) * | 2020-12-30 | 2021-06-04 | 厦门大学 | UPS (uninterrupted Power supply) health prediction method and system based on digital twinning and machine learning, electronic equipment and storable medium |
| CN113822467A (en) * | 2021-08-24 | 2021-12-21 | 华南理工大学 | Graph neural network prediction method for electric power area load |
| CN113837894A (en) * | 2021-08-06 | 2021-12-24 | 国网江苏省电力有限公司南京供电分公司 | Non-invasive resident user load decomposition method based on residual convolution module |
-
2021
- 2021-12-27 CN CN202111616875.0A patent/CN114510992A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110376457A (en) * | 2019-06-28 | 2019-10-25 | 同济大学 | Non-intrusion type load monitoring method and device based on semi-supervised learning algorithm |
| CN112033463A (en) * | 2020-09-02 | 2020-12-04 | 哈尔滨工程大学 | Nuclear power equipment state evaluation and prediction integrated method and system |
| CN112434799A (en) * | 2020-12-18 | 2021-03-02 | 宁波迦南智能电气股份有限公司 | Non-invasive load identification method based on full convolution neural network |
| CN112904220A (en) * | 2020-12-30 | 2021-06-04 | 厦门大学 | UPS (uninterrupted Power supply) health prediction method and system based on digital twinning and machine learning, electronic equipment and storable medium |
| CN113837894A (en) * | 2021-08-06 | 2021-12-24 | 国网江苏省电力有限公司南京供电分公司 | Non-invasive resident user load decomposition method based on residual convolution module |
| CN113822467A (en) * | 2021-08-24 | 2021-12-21 | 华南理工大学 | Graph neural network prediction method for electric power area load |
Non-Patent Citations (3)
| Title |
|---|
| MIN XIA: "Non intrusive load disaggregation based on deep dilated residual network", 《ELECTRIC POWER SYSTEMS RESEARCH》 * |
| ZIYUE JIA: "Sequence to Point Learning Based on Bidirectional Dilated Residual Network for Non Intrusive Load Monitoring", 《INTERNATIONAL JOURNAL OF ELECTRICAL POWER AND ENERGY SYSTEMS》 * |
| 谢晓兰等: "基于三次指数平滑法和时间卷积网络的云资源预测模型", 《通信学报》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116504005A (en) * | 2023-05-09 | 2023-07-28 | 齐鲁工业大学(山东省科学院) | Perimeter security intrusion signal identification method based on improved CDIL-Bi-LSTM |
| CN116504005B (en) * | 2023-05-09 | 2024-02-20 | 齐鲁工业大学(山东省科学院) | Perimeter security intrusion signal identification method based on improved CDIL-Bi-LSTM |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111639586B (en) | Non-invasive load identification model construction method, load identification method and system | |
| CN113109715B (en) | Battery health condition prediction method based on feature selection and support vector regression | |
| CN112215406B (en) | Non-invasive resident electricity load decomposition method based on time convolution neural network | |
| CN104297577B (en) | Ageing state estimation detecting system and method based on ultracapacitor | |
| CN114091536A (en) | Load decomposition method based on variational self-encoder | |
| CN108804696A (en) | The creation method and its system of power consumer electric appliance fingerprint base | |
| CN114510992A (en) | Equipment switch state detection method based on deep learning | |
| CN114037178A (en) | Non-invasive load decomposition method based on unsupervised pre-training neural network | |
| CN112379273A (en) | Lithium ion battery charging curve reconstruction and state estimation method based on artificial intelligence | |
| CN115510612A (en) | Method for predicting remaining service life of lithium ion battery | |
| Han et al. | Non‐intrusive load monitoring by voltage–current trajectory enabled asymmetric deep supervised hashing | |
| Athanasiadis et al. | Real-time non-intrusive load monitoring: A machine-learning approach for home appliance identification | |
| Tabanelli et al. | A feature reduction strategy for enabling lightweight non-intrusive load monitoring on edge devices | |
| CN114740360A (en) | Lithium ion battery residual service life prediction method based on SAE-CEEMDAN-LSTM | |
| CN115859190A (en) | Non-invasive household electrical classification method based on causal relationship | |
| Li et al. | Adaptive fusion feature transfer learning method for NILM | |
| CN113901726B (en) | Non-invasive load decomposition method based on seq2point model | |
| Zheng et al. | Non-intrusive load monitoring based on the graph least squares reconstruction method | |
| Ridi et al. | User interaction event detection in the context of appliance monitoring | |
| CN116628620A (en) | Non-invasive load identification calculation method | |
| Massidda et al. | Non-intrusive load disaggregation via a fully convolutional neural network: improving the accuracy on unseen household | |
| CN115222018A (en) | Non-invasive load decomposition method and device | |
| CN113687182A (en) | Household load identification method, program and system based on noise reduction automatic encoder | |
| Dash et al. | Multi-Head Attention Based Model for Non-Intrusive Appliance State Detection in Smart Buildings | |
| CN114091576A (en) | Load decomposition method based on bidirectional expansion convolution |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20220517 |
|
| RJ01 | Rejection of invention patent application after publication |