A load decomposition method based on inception neural network

Abstract

The present invention relates to a load decomposition method based on an Inception neural network. The method comprises: obtaining reference load data for a number of electrical devices, including reference total energy consumption data for all devices and reference meteorological data for a corresponding time period; inputting the reference load data into a trained Inception neural network to obtain a load decomposition result, including the operating status type of each device and the corresponding active power estimate; and analyzing the energy consumption and electricity usage patterns of all devices based on the load decomposition result. Compared with existing technologies, the present invention improves the accuracy and precision of total energy consumption decomposition.

Description

Technical Field

[0001] The present invention relates to a non-invasive load decomposition technology, and in particular to a load decomposition method based on an Inception neural network. Background Art

[0002] Building a new power system dominated by new energy and a clean, low-carbon, safe, and efficient energy system will be the development direction of my country's energy and power sector for a considerable period of time in the future. As a key component of this "new power system," the Advanced Metering Infrastructure (AMI) is a complete network and system that measures, collects, stores, analyzes, and utilizes user electricity usage information. Load monitoring, one of AMI's most important applications, primarily decomposes and categorizes loads, enabling targeted management of user electricity usage, enabling refined energy management, improving energy efficiency, and reducing energy costs.

[0003] In the 1980s, Professor Hart first proposed the concept of non-intrusive load monitoring (NILM). Using event-based monitoring, he decomposed the total power information recorded at the user's power point, thereby obtaining information on the energy consumption and patterns of the user's electrical equipment. However, due to technical limitations at the time, accuracy was difficult to achieve. Currently, research on load decomposition can be categorized into high-frequency data and low-frequency data. High-frequency data captures transient information such as current, voltage, and harmonics at the moment electrical appliances are turned on and off, and ranges from 10 to 100 MHz. However, high-frequency data is often expensive to acquire. Currently, low-frequency data, which is commonly used to decompose loads using voltage, current, or active power, is less expensive to acquire and is therefore more readily applicable in building load decomposition applications.

[0004] Traditional load decomposition algorithms mainly include the Hidden Markov model (HMM) and its variants, machine learning algorithms, etc. However, existing load decomposition algorithms only extract features of the object to be decomposed, without considering the correlation between the object's energy consumption and other data. They only use a single-structure neural network to mine features for load decomposition, resulting in low accuracy. Summary of the Invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a load decomposition method based on the Inception neural network, thereby improving the accuracy and precision of the total energy consumption decomposition.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] A load decomposition method based on an Inception neural network, comprising:

[0008] Obtain reference load data for several electrical devices, including the reference total energy consumption data of all electrical devices and reference meteorological data for the corresponding time period. Input the reference load data into the trained Inception neural network to obtain load decomposition results, including the operating status type of each electrical device and the corresponding active power estimate. Analyze the energy consumption and electricity usage patterns of all electrical devices based on the load decomposition results.

[0009] The training process of the inception neural network includes:

[0010] 1) Obtain the historical total energy consumption data and historical metering time series data of all electrical equipment, as well as the historical meteorological data of the corresponding time period,

[0011] 2) Perform data preprocessing;

[0012] 3) Based on the historical metering time series data of all electrical equipment, the sliding window algorithm and Gaussian mixture model clustering algorithm are combined to obtain the classification results of the working status of electrical equipment with respect to active power;

[0013] 4) The historical total energy consumption data, historical metering time series data, working status classification results and real-time meteorological data of the corresponding time period constitute a training set, and the training set inception neural network is used.

[0014] Furthermore, the inception neural network includes a convolutional neural network, a GRU neural network and a BP neural network.

[0015] The reference load data is input into each sub-neural network respectively, and the load decomposition results of each sub-neural network are fused to obtain the final load decomposition result.

[0016] Furthermore, step 3) includes:

[0017] 31) Based on the historical metering time series data of the electrical equipment, the start and stop times of the electrical equipment are obtained through the sliding window algorithm;

[0018] 32) Based on historical metering time series data, extract the active power waveform of the electrical equipment between the start time and the stop time. The active power waveforms of all electrical equipment constitute a waveform data set;

[0019] 33) Based on the waveform data set, the working status of the electrical equipment is clustered using the Gaussian mixture model clustering algorithm to obtain the classification results of the working status of the electrical equipment regarding active power.

[0020] Furthermore, step 31) includes:

[0021] Create a data buffer and store it in sequence N Each time a new measurement data is collected, the earliest stored measurement data will be deleted;

[0022] calculate t Working status characteristics at all times , the calculation formula is:

[0023]

[0024] in, and In the data buffer N The mean and variance of the measurement data;

[0025] When satisfied and When the power consumption equipment is started, it is determined that the power consumption equipment is started and the active power waveform of the power consumption equipment is recorded. 、 and is the set value, and ;

[0026] When satisfied again When the power consumption equipment is stopped, it is determined that the power consumption equipment has stopped, and the recording of the active power waveform of the power consumption equipment is stopped.

[0027] Furthermore, the waveform data set is represented as ,in Indicates the j Active power waveform of each electrical device;

[0028] The expression of the Gaussian mixture model is:

[0029]

[0030] in, is the probability density function, and Respectively i The mean vector and covariance matrix of the Gaussian mixture components, is the mixing coefficient, satisfying and , is the number of Gaussian mixture components, that is, the number of cluster centers, is the Gaussian mixture model i A probability distribution model.

[0031] Furthermore, step 33) includes:

[0032] Based on the waveform data set The parameters in the Gaussian mixture model 、 and

[0033] Perform iterative calculations until convergence or the maximum number of iterations is reached. The iterative calculation formula is:

[0034]

[0035]

[0036]

[0037] in, is the posterior distribution probability, and the calculation formula is:

[0038]

[0039] After the iterative calculation is completed, the Gaussian mixture model will transform the waveform data set Divided into K Cluster , each cluster is labeled Determine as follows:

[0040]

[0041] A working state type label is set for each cluster to obtain the working state classification result of the electrical equipment with respect to active power.

[0042] Furthermore, the data preprocessing includes filling missing values, removing outliers and denoising.

[0043] Furthermore, the process of filling missing values includes:

[0044] Fill in the deleted and missing values. For discrete missing values, use the data of the previous moment to fill in the missing values.

[0045] Furthermore, the process of removing outliers includes:

[0046] Set the maximum user load value , will exceed 1.2 The user load data of the selected items are deleted as outliers.

[0047] Furthermore, the meteorological data includes temperature, wind speed and air pressure.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] (1) The present invention obtains reference load data of several electrical equipment, which includes reference total energy consumption data of all electrical equipment and reference meteorological data of the corresponding time period. The reference load data is input into the trained Inception neural network to obtain load decomposition results, which include the working state type of each electrical equipment and the corresponding active power estimation value. The energy consumption and power consumption patterns of all electrical equipment are analyzed based on the load decomposition results. Since the power consumption patterns of electrical equipment are highly correlated with meteorological data, the total energy consumption of users and meteorological data are used as inputs of the Inception neural network together. Multiple sub-neural networks are used to deeply explore the characteristics of the input data, which significantly improves the accuracy of the total energy consumption decomposition.

[0050] (2) The present invention obtains the start and stop times of the electrical equipment through a sliding window algorithm based on the historical metering time series data of the electrical equipment, thereby eliminating a large amount of redundant data in the active power data of the electrical equipment and improving the accuracy and efficiency of extracting the active power waveform of the electrical equipment;

[0051] (3) Based on the waveform data set, the present invention clusters the working status of the electrical equipment through the Gaussian mixture model clustering algorithm to obtain the classification results of the working status of the electrical equipment with respect to active power. By utilizing the soft classification and flexible clustering advantages of the Gaussian mixture model clustering algorithm, the load state characteristics are fully explored, and the classification accuracy of the working status of the electrical equipment is improved, thereby improving the accuracy of active power estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a diagram of the training process of the inception neural network;

[0053] Figure 2 This is the classification result of dishwasher equipment using Gaussian mixture model clustering algorithm;

[0054] Figure 3 This is the classification result of the dishwasher equipment using the K-Means algorithm. DETAILED DESCRIPTION

[0055] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0056] A load decomposition method based on an Inception neural network, comprising:

[0057] Obtain reference load data for several electrical devices. This data includes the reference total energy consumption of all devices and reference meteorological data for the corresponding time period. This data is then fed into a trained Inception neural network to obtain load decomposition results. These results include the operating state type of each device and the corresponding active power estimate. Operating state types include off, high-power, low-power, and standby. The energy consumption and electricity usage patterns of all devices are analyzed based on the load decomposition results.

[0058] like Figure 1 , the training process of the inception neural network includes:

[0059] 1) Obtain the historical total energy consumption data and historical metering time series data of all electrical equipment, as well as the historical meteorological data of the corresponding time period,

[0060] 2) Perform data preprocessing;

[0061] 3) Based on the historical metering time series data of all electrical equipment, the sliding window algorithm and Gaussian mixture model clustering algorithm are combined to obtain the classification results of the working status of electrical equipment with respect to active power;

[0062] 4) The historical total energy consumption data, historical metering time series data, working status classification results and real-time meteorological data of the corresponding time period constitute the training set, and the training set inception neural network is used.

[0063] Weather data includes temperature, wind speed, and air pressure.

[0064] The inception neural network includes convolutional neural network, GRU neural network and BP neural network;

[0065] The reference load data is input into each sub-neural network respectively. The last layer of each sub-neural network has the same output unit so that its output tensor size is the same. The average of the three output tensors is taken and output to the next layer. The three tensors are merged into one tensor, and the synthesized tensor is then input into the logistic regression layer to obtain the final load decomposition result.

[0066] The convolutional neural network uses a one-dimensional convolution formula. An important feature of the convolution operation is that it can enhance the original signal characteristics and reduce noise. The formula is:

[0067]

[0068] in, is the input sequence of the convolutional layer, g is the convolution kernel, N For the input sequence length,M is the length of the convolution kernel, S is the convolution result sequence, and the length of the convolution result sequence is , the spatial feature information of the input sequence can be obtained through the convolution result sequence;

[0069] The GRU sub-neural network is an improvement to the long short-term memory network, with fewer parameters and faster training speed. and update gate , the calculation formula is:

[0070]

[0071]

[0072] in, is the hidden state at the previous moment, is the input at the current moment, The activation function is sigmoid, and the output of the GRU unit is the hidden state at the current moment , which is calculated as follows (“﹡” represents the product of matrices):

[0073]

[0074]

[0075] The BP sub-neural network is a combination of multiple fully connected layers, and its neuron propagation formula is as follows:

[0076]

[0077] in, H is the hidden layer output set, is the weight matrix, is the output set of the previous fully connected layer, is the deviation parameter.

[0078] Step 3) includes:

[0079] 31) Based on the historical metering time series data of the electrical equipment, the start and stop times of the electrical equipment are obtained through the sliding window algorithm;

[0080] 32) Based on historical metering time series data, extract the active power waveform of the electrical equipment between the start time and the stop time. The active power waveforms of all electrical equipment constitute a waveform data set;

[0081] 33) Based on the waveform data set, the working status of the electrical equipment is clustered using the Gaussian mixture model clustering algorithm to obtain the classification results of the working status of the electrical equipment regarding active power.

[0082] Step 31) includes:

[0083] Create a data buffer and store it in sequence N Each time a new measurement data is collected, the earliest stored measurement data will be deleted;

[0084] calculate t Working status characteristics at all times , the calculation formula is:

[0085]

[0086] in, and In the data buffer N The mean and variance of the measurement data;

[0087] When satisfied and When the power consumption equipment is started, it is determined that the power consumption equipment is started and the active power waveform of the power consumption equipment is recorded. 、 and is the set value, and ;

[0088] When satisfied again When the power consumption equipment is stopped, it is determined that the power consumption equipment has stopped, and the recording of the active power waveform of the power consumption equipment is stopped.

[0089] The waveform dataset is represented as ,in Indicates the j Active power waveform of each electrical device;

[0090] The expression of Gaussian mixture model is:

[0091]

[0092] in, is the probability density function, and Respectively i The mean vector and covariance matrix of the Gaussian mixture components, is the mixing coefficient, satisfying and , is the number of Gaussian mixture components, that is, the number of cluster centers, is the Gaussian mixture model i A probability distribution model.

[0093] Step 33) includes:

[0094] Based on the waveform data set The parameters in the Gaussian mixture model 、 and

[0095] Perform iterative calculations until convergence or the maximum number of iterations is reached. The iterative calculation formula is:

[0096]

[0097]

[0098]

[0099] in, is the posterior distribution probability. According to Bayes’ theorem, The calculation formula is:

[0100]

[0101] After the iterative calculation is completed, the Gaussian mixture model will transform the waveform data set Divided into K Cluster , each cluster is labeled Determine as follows:

[0102]

[0103] A working state type label is set for each cluster to obtain the working state classification result of the electrical equipment with respect to active power.

[0104] Data preprocessing includes filling missing values, removing outliers and denoising;

[0105] The process of filling missing values includes: filling in deleted values and missing values. For discrete missing values, the data of the previous moment before the missing value is used to fill in the missing values.

[0106] The process of removing outliers includes: setting the maximum load value of the user , will exceed 1.2 The user load data of the selected items are deleted as outliers.

[0107] Taking the Canadian public dataset AMPds2 as an example, AMPds2 covers more than 20 types of electrical equipment, including total energy consumption, metering time series data of electrical equipment, and meteorological data. This example selects equipment with a large energy consumption, including refrigerators, washing machines, dryers, dishwashers, televisions, lighting, and heat pumps. Hourly meteorological data for the same year is collected. Meteorological data includes temperature, wind speed, and air pressure.

[0108] When the sliding window algorithm is used to obtain the start time and stop time of the electric device, the window size, working state threshold, and non-working state threshold of the specific electric device are shown in Table 1:

[0109] Table 1 Sliding window size and event judgment threshold table (unit: W)

[0110]

[0111] The load decomposition method proposed in this embodiment is compared with the existing load decomposition method in two aspects. On the one hand, the load state feature extraction is compared with the commonly used K-Means algorithm. Figure 2 This is the classification result of dishwasher equipment using Gaussian mixture model clustering algorithm. Figure 3 The K-Means algorithm is used to classify dishwashers. It can be seen that the Gaussian mixture model clustering algorithm obtains more classifications, that is, the decomposition accuracy is higher. On the other hand, the load decomposition model is compared with the conventional CNN, LSTM, and the newer CNN-GRU model. The evaluation index is the mean absolute error. , the calculation formula is as follows:

[0112]

[0113] in, for t The estimated active power at time , for t The actual value of active power at the moment, The smaller the load decomposition accuracy, the higher the load decomposition accuracy. Taking the dishwasher as an example, the comparison results of the influence of different clustering algorithms on the decomposition accuracy are shown in Table 2:

[0114] Table 2 Different clustering algorithms Comparison Table

[0115]

[0116] It can be seen from Table 2 that the embodiment fully utilizes the "soft classification" and cluster flexibility advantages of the Gaussian mixture model clustering algorithm, so that the situations where the power consumption devices are similar but the status are different during operation can be distinguished.

[0117] The comparison of the results of different load decomposition models is shown in Table 3:

[0118] Table 3 Different load decomposition models Comparison Table

[0119]

[0120] From Table 3, it can be seen that except for the heat pump, which is slightly lower than the conventional neural network, the accuracy of other electrical equipment has been improved. Therefore, it can be seen that the load decomposition method proposed in this embodiment can accurately and effectively achieve non-invasive load decomposition.

[0121] This embodiment proposes a load decomposition method based on the Inception neural network. To address the large amount of redundant data in the active power data of electrical equipment, a sliding window algorithm is designed to eliminate redundancy. The "soft classification" and flexible clustering advantages of the Gaussian mixture model clustering algorithm are utilized to fully explore load state characteristics, improve the classification accuracy of the electrical equipment working status, and thus improve the accuracy of active power estimation. User total energy consumption and meteorological data are used as inputs to the Inception neural network. Multiple sub-neural networks are used to deeply explore the characteristics of the input data, significantly improving the accuracy of total energy consumption decomposition.

[0122] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.

Claims

1. A load decomposition method based on inception neural network, characterized in that: include: Obtain reference load data for several electrical devices, including the reference total energy consumption data of all electrical devices and reference meteorological data for the corresponding time period. Input the reference load data into the trained Inception neural network to obtain load decomposition results, including the operating status type of each electrical device and the corresponding active power estimate. Analyze the energy consumption and electricity usage patterns of all electrical devices based on the load decomposition results. The training process of the inception neural network includes: 1) Obtain the historical total energy consumption data and historical metering time series data of all electrical equipment, as well as the historical meteorological data for the corresponding time period; 2) Perform data preprocessing; 3) Based on the historical metering time series data of all electrical equipment, the sliding window algorithm and Gaussian mixture model clustering algorithm are combined to obtain the classification results of the working status of electrical equipment with respect to active power; 4) The historical total energy consumption data, historical metering time series data, working status classification results and historical meteorological data of the corresponding time period constitute a training set, and the training set inception neural network is used; Step 3) includes: 31) Based on the historical metering time series data of the electrical equipment, the start and stop times of the electrical equipment are obtained through the sliding window algorithm; 32) Based on historical metering time series data, extract the active power waveform of the electrical equipment between the start time and the stop time. The active power waveforms of all electrical equipment constitute a waveform data set; 33) Based on the waveform data set, cluster the working status of the electrical equipment using the Gaussian mixture model clustering algorithm to obtain the classification results of the working status of the electrical equipment with respect to active power; Step 31) includes: Create a data buffer and store it in sequence N Each time a new measurement data is collected, the earliest stored measurement data will be deleted; calculate t Working status characteristics at all times , the calculation formula is: in, and In the data buffer N The mean and variance of the measurement data; When satisfied and When the power consumption equipment is started, it is determined that the power consumption equipment is started and the active power waveform of the power consumption equipment is recorded. 、 and is the set value, and ; When satisfied again When the power consumption equipment is stopped, it is determined that the power consumption equipment is stopped, and the recording of the active power waveform of the power consumption equipment is stopped; Step 33) includes: Based on the waveform data set The parameters in the Gaussian mixture model 、 and Perform iterative calculations until convergence or the maximum number of iterations is reached. The iterative calculation formula is: in, is the posterior distribution probability, Indicates the j The active power waveform of each electrical device is calculated as follows: in, and Respectively i The mean vector and covariance matrix of the Gaussian mixture components, is the mixing coefficient, satisfying and , is the number of Gaussian mixture components, that is, the number of cluster centers, is the Gaussian mixture model i A probability distribution model; After the iterative calculation is completed, the Gaussian mixture model will transform the waveform data set Divided into K Cluster , each cluster is labeled Determine as follows: A working state type label is set for each cluster to obtain the working state classification result of the electrical equipment with respect to active power.

2. A load decomposition method based on inception neural network according to claim 1, characterized in that: The inception neural network includes a convolutional neural network, a GRU neural network and a BP neural network. The reference load data is input into each sub-neural network respectively, and the load decomposition results of each sub-neural network are fused to obtain the final load decomposition result.

3. The load decomposition method based on the inception neural network according to claim 1 is characterized in that: The waveform data set is represented as ,in Indicates the j Active power waveform of each electrical device; The expression of the Gaussian mixture model is: in, is the probability density function, and Respectively i The mean vector and covariance matrix of the Gaussian mixture components, is the mixing coefficient, satisfying and , is the number of Gaussian mixture components, that is, the number of cluster centers, is the Gaussian mixture model i A probability distribution model.

4. The load decomposition method based on the inception neural network according to claim 1 is characterized in that: The data preprocessing includes filling missing values, removing outliers and denoising.

5. The load decomposition method based on the inception neural network according to claim 4 is characterized in that: The missing value filling process includes: Fill in the deleted and missing values. For discrete missing values, use the data of the previous moment to fill in the missing values.

6. The load decomposition method based on the inception neural network according to claim 4 is characterized in that: The process of removing outliers includes: Set the maximum user load value , will exceed 1.2 The user load data of the selected items are deleted as outliers.

7. The load decomposition method based on the Inception neural network according to claim 1 is characterized in that: The meteorological data includes temperature, wind speed and air pressure.

Images