User typical air conditioner load operation mode division method and system based on deep clustering

Abstract

The invention discloses a user typical air conditioner load operation mode division method and system based on deep clustering, and the method comprises the steps: carrying out the detection of a collected total power signal, and obtaining a sudden change point in the power signal; forming a load event sample based on the elements of the abrupt change point, judging whether a fluctuation event exists or not, and if yes, removing the fluctuation event which is mistakenly detected due to load fluctuation, and updating a load event detection result; performing normalization processing on the updated power sequence of the load event; constructing and training a low-dimensional feature extraction model to obtain an optimized low-dimensional feature extraction model; performing feature extraction based on the optimized low-dimensional feature extraction model and the normalized power sequence of the load event; and the low-dimensional features are clustered based on a DBSCAN clustering algorithm to obtain clusters, and then the air conditioner load modes which the clusters belong to are screened out. According to the method, template library support is provided for template-based air conditioner load mode recognition, and the accuracy and convenience of detection are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of deep learning, and relates to a method and system for dividing user typical air-conditioning load operation modes based on deep clustering. Background Art

[0002] Monitoring the operating status of air conditioners can be summarized as an electrical load monitoring task. The main process of monitoring the load status of specific electrical appliances is to extract load features from the measured data, and effectively identify the components of the power load by combining the pre-built load feature library. The current load identification algorithms are divided into two types: mathematical optimization and pattern recognition. The load identification method based on mathematical optimization is to find the optimal solution of the optimization problem in the feature space. Such methods usually require the load features to be additive; pattern recognition is to learn the characteristic patterns of the load features of each power load to achieve load identification. It has the advantages of simplicity and practicality and is widely used. However, the existing technology focuses more on estimating the energy consumption of air conditioners by establishing a relationship between air conditioners and accessible external information. These methods are simple and effective for estimating the total air conditioning load in a specific area, but the performance is difficult to guarantee for individual users with different usage habits. Other heuristic rule-based algorithms produce good results for the target scenario. However, these rules may also lead to the risk of reducing the universality of unsupervised algorithms. Some template matching-based technologies have good application prospects, but the monitoring effect of such technologies depends on the quality of the template library. Existing research finds it difficult to extract a complete and high-quality air-conditioning load pattern template library in an unsupervised manner, which makes the template matching-based technology perform poorly. Summary of the invention

[0003] The purpose of the present invention is to solve the problem that it is difficult to estimate the energy consumption of air conditioners for individual users with different usage habits in the prior art, and that it is difficult for existing methods to extract a complete and high-quality air conditioner load pattern template library in an unsupervised manner, resulting in poor performance of template matching-based technologies. A method and system for dividing user typical air conditioner load operation modes based on deep clustering are provided.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions:

[0005] The method for dividing the user's typical air conditioning load operation mode based on deep clustering includes:

[0006] Collect the total power signal at the user bus;

[0007] Detect the collected total power signal to obtain the mutation point in the power signal;

[0008] Based on the elements of the mutation point, a load event sample is formed. Based on the load event sample, it is determined whether there is a fluctuation event. If so, the fluctuation event that is misdetected due to load fluctuation is removed, and the load event detection result is updated.

[0009] Normalizing the power sequence of the updated load event;

[0010] Build and train a low-dimensional feature extraction model to obtain the most optimized low-dimensional feature extraction model;

[0011] Feature extraction is performed based on the optimized low-dimensional feature extraction model and the power sequence of the load event after normalization;

[0012] Based on the DBSCAN clustering algorithm, low-dimensional features are clustered to obtain clusters, and then the air-conditioning load patterns to which the clusters belong are screened out.

[0013] A further improvement of the present invention is:

[0014] Furthermore, the collected total power signal is detected to obtain the mutation point in the power signal, specifically:

[0015] A sliding window with a fixed length is set to scan the total power signal, and several sampling points are set on the total power signal. When the center of the sliding window is located at the i-th sampling point P i When the sliding window scans, the sequence is recorded as SW i :

[0016] SW i ={SW i pre ,P i ,SW i post} (1)

[0017] Among them, SW i pre represents the power sample sequence in the front sliding window, SW i post Represents the power sample sequence in the rear sliding window; SW i pre and SW i post The length of SW is w, then i pre and SW i post The definition of is as follows:

[0018]

[0019] When SW i pre and SWi post The mean of satisfies formula (3), then P i For the mutation point:

[0020] μ(SW i post )-μ(SW i pre )>P th_low (3)

[0021] In the formula, μ() represents the mean calculation, P th_low is the low power threshold used to determine the mutation point.

[0022] Furthermore, based on the mutation point, a load event sample is formed, specifically:

[0023] If the i-th sampling point P i and the jth sampling point P j Both are mutation points. It is determined whether the two mutation points belong to the same load event, as shown in formula (4):

[0024] ji-2w<T th (4)

[0025] Among them, T th is the duration threshold; if formula (4) is satisfied, then if the mutation point P i and mutation point P j For the same load event;

[0026] The coordinate sequence of the load event l is denoted as a closed set D event (l):

[0027] D event (l)=[L(l),R(l)] (5)

[0028] Where L(l) and R(l) are the left and right coordinate boundaries of the event, respectively, and are defined as in formula (6):

[0029]

[0030] Among them, w means the front and rear sliding windows SW i pre and SW i post Length.

[0031] Furthermore, based on the load event samples, it is determined whether there is a fluctuation event, specifically:

[0032] For load event l, if it satisfies formula (7), it is judged to be a fluctuation event:

[0033]

[0034] Among them, P th_high To set a high power threshold for identifying fluctuation events; P k is the kth sampling point.

[0035] Furthermore, the power sequence of the updated load event is normalized, specifically:

[0036]

[0037] Among them, P event (n) represents the actual value of the power sequence of the nth load event, min(·) and max(·) represent the minimum and maximum operations in the power sequence, respectively. norm (n) is the normalized power sequence of the nth load event.

[0038] Furthermore, the low-dimensional feature extraction network consists of 5 fully connected layers, the number of units in the input layer and the output layer is L, the number of units in the remaining hidden layers is 256, and the activation function of each layer is g = tanh, specifically:

[0039]

[0040] Among them, the number of units in the input layer of the low-dimensional feature extraction network is L, and the input data X∈R L ; The weight of the first hidden layer is W 1 ∈R L*256 ; Bias is b 1 ∈R 256 ; The activation function is: g(x) = tanh(x); The output is H 1 =tanh(W 1 X+b 1 ); the weight of the second hidden layer is W 2 ∈R 256*256 , with a bias of b 2 ∈R 256 ; The activation function is: g(x) = tanh(x); The output is H 2 =tanh(W 2 H 1 +b 2 ); the weight of the third hidden layer is W 3 ∈R 256*256 , with a bias of b 3 ∈R 256 ; The activation function is: g(x) = tanh(x); The output is H 3 =tanh(W 3 H 2 +b 3 ); the number of units in the output layer is L, and the weight is W4 ∈R 256*L ; Bias is b 4 ∈R L ; The activation function is Softmax; The output is Y = W 4 H 3 +b 4 .

[0041] Further, the low-dimensional feature extraction model is trained to obtain the optimized low-dimensional feature extraction model, specifically: norm (n) is a zero element added to the end to make its length the same as the original longest load event power sequence length L, and the power waveform feature x of load event n is obtained. n The input of the low-dimensional feature extraction network is the set X=[x 1 ,...x N ], the output is a low-dimensional feature set H = [h 1 ,...,h N ]; train the low-dimensional feature extraction network in an unsupervised manner, and the loss function is defined as formula (10):

[0042]

[0043] Among them, H 3 and H 5 They represent the output of the input X after passing through the 3rd and 5th fully connected layers, respectively, and W (m) and b (m) denote the weight and bias of the mth fully connected layer, w 1 and w 2 denote weight factors, respectively, ||·|| F represents the F-norm, ||·|| 2 Indicates the 2-norm, c is a sparse matrix, and is obtained by formula (11):

[0044]

[0045] Among them, c i represents the i-th column of c, c ii Then it means c i The i-th element of .

[0046] Furthermore, based on the DBSCAN clustering algorithm, low-dimensional features are clustered to obtain clusters, and then the air conditioning load patterns to which the clusters belong are screened out, specifically:

[0047] Cluster low-dimensional feature data based on the DBSCAN algorithm to obtain multiple clusters;

[0048] Based on the visualization tool of clustering results and typical values ​​of load characteristics, clusters not related to air conditioning are excluded;

[0049] Analyze the remaining clusters and identify the clusters that meet the characteristics of air conditioning load;

[0050] Through manual inspection, the air conditioning load pattern to which the cluster belongs is determined.

[0051] Furthermore, typical values ​​of load characteristics include: power range and time distribution; the remaining clusters are analyzed to identify clusters that meet the air-conditioning load characteristics, specifically: based on the typical load characteristics of the air conditioner, determine whether the remaining clusters contain relevant load characteristics; the typical load characteristics of the air conditioner include: high power peak when the air conditioner is started and periodic start-stop mode.

[0052] The user typical air conditioning load operation mode classification system based on deep clustering includes:

[0053] A collection module, wherein the collection module collects a total power signal at a user bus;

[0054] A detection module, wherein the detection module detects the collected total power signal to obtain a mutation point in the power signal;

[0055] A determination module, wherein the determination module forms a load event sample based on the elements of the mutation point, determines whether a fluctuation event exists based on the load event sample, and if so, removes the fluctuation event that is misdetected due to the load fluctuation, and updates the load event detection result;

[0056] A normalization processing module, wherein the normalization processing module performs normalization processing on the power sequence of the updated load event;

[0057] A construction module, wherein the construction module constructs and trains a low-dimensional feature extraction model to obtain an optimized low-dimensional feature extraction model;

[0058] A feature extraction module, wherein the feature extraction module performs feature extraction based on an optimized low-dimensional feature extraction model and a power sequence of a load event after normalization;

[0059] The clustering module clusters the low-dimensional features based on the DBSCAN clustering algorithm to obtain clusters, and then screens out the air-conditioning load mode to which the clusters belong.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] The present invention accurately detects load events that mark the conversion of electrical appliance load modes based on the total power signal, and extracts low-dimensional spatial features based on the power signal sequence of the load event by designing a low-dimensional feature extraction network, and then combines the clustering algorithm to mine the load mode set belonging to the air conditioner in a completely unsupervised manner, providing template library support for template-based air conditioner load mode recognition, thereby improving the accuracy and convenience of detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0063] Figure 1 It is a flow chart of a method for classifying typical user air-conditioning load operation modes based on deep clustering of the present invention;

[0064] Figure 2 It is a structural schematic diagram of a user typical air conditioning load operation mode classification system based on deep clustering of the present invention;

[0065] Figure 3 Schematic diagram of the low-dimensional feature extraction network structure. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0067] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0068] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0069] In the description of the embodiments of the present invention, it should be noted that if the terms "upper", "lower", "horizontal", "inner", etc. indicate an orientation or positional relationship based on the orientation or positional relationship shown in the drawings, or the orientation or positional relationship in which the product of the invention is usually placed when in use, it is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. In addition, the terms "first", "second", etc. are only used to distinguish the description, and cannot be understood as indicating or implying relative importance.

[0070] In addition, if the term "horizontal" appears, it does not mean that the component must be absolutely horizontal, but can be slightly tilted. For example, "horizontal" only means that its direction is more horizontal than "vertical", which does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0071] In the description of the embodiments of the present invention, it is also necessary to explain that, unless otherwise clearly specified and limited, the terms "set", "install", "connect", and "connect" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal connection of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0072] The present invention is further described in detail below in conjunction with the accompanying drawings:

[0073] See also Figure 1 The present invention discloses a method for dividing user typical air conditioning load operation modes based on deep clustering, comprising:

[0074] S101, collecting the total power signal at the user bus;

[0075] S102, detecting the collected total power signal to obtain a mutation point in the power signal;

[0076] A sliding window with a fixed length is set to scan the total power signal, and several sampling points are set on the total power signal. When the center of the sliding window is located at the i-th sampling point P i When the sliding window scans, the sequence is recorded as SW i :

[0077] SW i ={SW i pre ,P i ,SW i post} (1)

[0078] Among them, SW i pre represents the power sample sequence in the front sliding window, SW i post Represents the power sample sequence in the rear sliding window; SW i pre and SW i post The length of SW is w, then i pre and SW i post The definition of is as follows:

[0079]

[0080] When SW i pre and SW i post The mean of satisfies formula (3), then P i For the mutation point:

[0081] μ(SW i post )-μ(SW i pre )>P th_low (3)

[0082] In the formula, μ() represents the mean calculation, P th_low is the low power threshold used to determine the mutation point.

[0083] S103, forming a load event sample based on the elements of the mutation point, and determining whether there is a fluctuation event based on the load event sample. If there is, removing the fluctuation event that is misdetected due to the load fluctuation, and updating the load event detection result;

[0084] Based on the mutation point, a load event sample is formed, specifically:

[0085] If the i-th sampling point P i and the jth sampling point P j Both are mutation points. It is determined whether the two mutation points belong to the same load event, as shown in formula (4):

[0086] ji-2w<T th (4)

[0087] Among them, T th is the duration threshold; if formula (4) is satisfied, then if the mutation point P i and mutation point P j For the same load event;

[0088] The coordinate sequence of the load event l is denoted as a closed set D event (l):

[0089] D event (l)=[L(l),R(l)] (5)

[0090] Where L(l) and R(l) are the left and right coordinate boundaries of the event, respectively, and are defined as in formula (6):

[0091]

[0092] Among them, w means the front and rear sliding windows SW i pre and SW i post Length.

[0093] Based on the load event samples, determine whether there is a fluctuation event, specifically:

[0094] For load event l, if it satisfies formula (7), it is judged to be a fluctuation event:

[0095]

[0096] Among them, P th_high To set a high power threshold for identifying fluctuation events; P k is the kth sampling point.

[0097] S104, normalizing the updated power sequence of the load event;

[0098]

[0099] Among them, P event (n) represents the actual value of the power sequence of the nth load event, min(·) and max(·) represent the minimum and maximum operations in the power sequence, respectively. norm (n) is the normalized power sequence of the nth load event.

[0100] S105, constructing and training a low-dimensional feature extraction model to obtain an optimized low-dimensional feature extraction model;

[0101] The low-dimensional feature extraction network consists of 5 fully connected layers, and the number of units in the input layer and output layer is L.

[0102] The number of units in the remaining hidden layers is 256, and the activation function of each layer is g = tanh, specifically:

[0103]

[0104] Among them, the number of units in the input layer of the low-dimensional feature extraction network is L, and the input data X∈R L ; The weight of the first hidden layer is W 1 ∈R L*256 ; Bias is b 1 ∈R 256 ; The activation function is: g(x) = tanh(x); The output is H 1 =tanh(W 1 X+b 1 ); the weight of the second hidden layer is W 2 ∈R 256*256 , with a bias of b 2 ∈R 256 ; The activation function is: g(x) = tanh(x); The output is H 2 =tanh(W 2 H 1 +b 2 ); the weight of the third hidden layer is W 3 ∈R 256*256 , with a bias of b 3 ∈R 256 ; The activation function is: g(x) = tanh(x); The output is H 3 =tanh(W 3 H 2 +b 3 ); the number of units in the output layer is L, and the weight is W 4 ∈R 256*L ; Bias is b 4 ∈R L ; The activation function is Softmax; The output is Y = W 4 H 3 +b 4 .

[0105] Train the low-dimensional feature extraction model to obtain the optimized low-dimensional feature extraction model, specifically: norm (n) is a zero element added to the end to make its length the same as the original longest load event power sequence length L, and the power waveform feature x of load event n is obtained. n The input of the low-dimensional feature extraction network is the set X=[x 1 ,…x N ], the output is a low-dimensional feature set H = [h 1 ,…,h N ]; train the low-dimensional feature extraction network in an unsupervised manner, and the loss function is defined as formula (10):

[0106]

[0107] Among them, H 3 and H 5 They represent the output of the input X after passing through the 3rd and 5th fully connected layers, respectively, and W (m) and b (m) denote the weight and bias of the mth fully connected layer, w 1 and w 2 denote weight factors, respectively, ||·|| F represents the F-norm, ||·|| 2 Indicates the 2-norm, c is a sparse matrix, and is obtained by formula (11):

[0108]

[0109] Among them, c i represents the i-th column of c, c ii Then it means c i The i-th element of .

[0110] S106, performing feature extraction based on the optimized low-dimensional feature extraction model and the power sequence of the load event after normalization;

[0111] S107, clustering the low-dimensional features based on the DBSCAN clustering algorithm to obtain clusters, and then screening out the air conditioning load pattern to which the clusters belong.

[0112] Cluster low-dimensional feature data based on the DBSCAN algorithm to obtain multiple clusters;

[0113] Based on the visualization tool of clustering results and typical values ​​of load characteristics, clusters not related to air conditioning are excluded;

[0114] Analyze the remaining clusters and identify the clusters that meet the characteristics of air conditioning load;

[0115] Through manual inspection, the air conditioning load pattern to which the cluster belongs is determined.

[0116] Typical values ​​for load characteristics include: power range and time distribution;

[0117] The remaining clusters are analyzed to identify clusters that meet the air conditioning load characteristics, specifically: based on the typical load characteristics of the air conditioner, it is determined whether the remaining clusters contain relevant load characteristics; the typical load characteristics of the air conditioner include: a high power peak when the air conditioner is started and a periodic start-stop mode.

[0118] See also Figure 2 The present invention discloses a user typical air conditioning load operation mode classification system based on deep clustering, comprising:

[0119] A collection module, wherein the collection module collects a total power signal at a user bus;

[0120] A detection module, wherein the detection module detects the collected total power signal to obtain a mutation point in the power signal;

[0121] A determination module, wherein the determination module forms a load event sample based on the elements of the mutation point, determines whether a fluctuation event exists based on the load event sample, and if so, removes the fluctuation event that is misdetected due to the load fluctuation, and updates the load event detection result;

[0122] A normalization processing module, wherein the normalization processing module performs normalization processing on the power sequence of the updated load event;

[0123] A construction module, wherein the construction module constructs and trains a low-dimensional feature extraction model to obtain an optimized low-dimensional feature extraction model;

[0124] A feature extraction module, wherein the feature extraction module performs feature extraction based on an optimized low-dimensional feature extraction model and a power sequence of a load event after normalization;

[0125] The clustering module clusters the low-dimensional features based on the DBSCAN clustering algorithm to obtain clusters, and then screens out the air-conditioning load mode to which the clusters belong.

[0126] Example:

[0127] The present invention proposes a method for dividing user typical air conditioning load operation modes based on deep clustering, comprising the steps of:

[0128] Step 1: Using an event detection method based on multi-threshold settings, according to the total power signal collected from the user's main power outlet, detect load events that mark changes in the operating status of electrical appliances;

[0129] Step 2: Build and train the low-dimensional feature extraction model of this method to extract the low-dimensional features of the load events, and finally use the clustering algorithm to divide the air-conditioning load pattern, that is, the set of load events that occur multiple times.

[0130] Further, the specific contents of the method of the present invention are as follows:

[0131] Step 1 includes: collecting the total power signal from the user bus, and detecting the load event of the electrical appliance behavior that occurs when the electrical appliance load state switches, as follows:

[0132] First, detect the mutation point in the power signal. Set a fixed-length sliding window to scan the acquired 1Hz total power signal. When the center of the sliding window is located at the i-th sampling point P i When the sliding window scans the sequence SWi It can be written as:

[0133] SW i ={SW i pre ,P i ,SW i post} (12)

[0134] Where SW i pre represents the power sample sequence in the front sliding window, SW i post Represents the power sample sequence in the back sliding window. SW i pre and SW i post If the length of both is w, then the two can be defined as follows:

[0135]

[0136] When SW i pre and SW i post If equation (14) is satisfied, then P i For the mutation point:

[0137] μ(SW i post )-μ(SW i pre )>P th_low (14)

[0138] In the formula, μ(·) represents the mean calculation, P th_low is the low power threshold used to determine the mutation point.

[0139] Then, the elements determined as mutation points are connected to form a load event sample. i and P j Both are judged as mutation points. If they satisfy formula (15), they are judged to belong to the same load event:

[0140] ji-2w<T th (15)

[0141] Where, T th is the duration threshold. The coordinate sequence of the load event l is denoted as a closed set D event (l):

[0142] D event (l)=[L(l),R(l)] (16)

[0143] Where L(l) and R(l) are the left and right coordinate boundaries of the event, respectively, and their definitions are as shown in formula (17):

[0144]

[0145] Identify and remove the fluctuation events that are misdetected due to load fluctuation. For an event l, if it satisfies formula (18), it is judged as a fluctuation event:

[0146]

[0147] Where P th_high To set a high power threshold for identifying fluctuation events, remove samples that are judged to be fluctuation events and update the load event detection result.

[0148] Step 2 includes: based on the deep clustering analysis method, the load event samples detected in step 1 are divided into load pattern sets belonging to different electrical appliances, as follows:

[0149] The power sequence of the detected load event is normalized according to formula (19):

[0150]

[0151] Where P event (n) represents the actual value of the power sequence of the nth load event, min(·) and max(·) represent the minimum and maximum operations in the power sequence, respectively. norm (n) is the power sequence after normalization of the nth load event. norm (n) is a zero element added to the end to make its length the same as the original longest load event power sequence length L, and the power waveform feature x of load event n is obtained. n .

[0152] The low-dimensional feature extraction network described in step 2 is as shown in the attached Figure 3 , its structure consists of 5 fully connected layers, the number of units in the input layer and the output layer is L, the number of units in the remaining hidden layers is 256, and the activation function of each layer is g = tanh, defined as formula (20):

[0153]

[0154] The input of the low-dimensional feature extraction network is a set X consisting of power waveform features of all N load events = [x 1 ,...x N ], the output is a low-dimensional feature set H = [h 1 ,...,h N]. The network is trained in an unsupervised manner, and the loss function is defined as follows:

[0155]

[0156] In the formula, H 3 and H 5 They represent the output of the input X after passing through the 3rd and 5th fully connected layers, respectively, and W (m) and b (m) denote the weight and bias of the mth fully connected layer, w 1 and w 2 denote weight factors, respectively, ||·|| F represents the F-norm, ||·|| 2 Indicates the 2-norm, C is a sparse matrix, which can be obtained by solving equation (22):

[0157]

[0158] Among them, c i represents the i-th column of C, c ii Then it means c i The ith element of .

[0159] The DBSCAN clustering algorithm is used to cluster the low-dimensional features output by the trained low-dimensional feature extraction network. The clusters mined correspond to the load patterns composed of multiple load events from the same electrical appliance. Finally, the load pattern belonging to the air conditioner is selected by manual screening.

[0160] The key parameters in the present invention are: in step 1, SW i pre and SW i post The length w is 4 seconds, and the low power threshold P th_low The value is 5W, and the duration threshold is T th The value is 8 seconds, and the high power threshold P th_high The value is 20W; in step 2, the parameter density radius of the DBSCAN algorithm is set to 0.3, and the minimum number of data points is set to 2.

[0161] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for dividing user typical air conditioning load operation modes based on deep clustering, characterized in that: include: Collect the total power signal at the user bus; Detect the collected total power signal to obtain the mutation point in the power signal; Based on the elements of the mutation point, a load event sample is formed. Based on the load event sample, it is determined whether there is a fluctuation event. If so, the fluctuation event that is misdetected due to load fluctuation is removed, and the load event detection result is updated. Normalizing the power sequence of the updated load event; Build and train a low-dimensional feature extraction model to obtain the most optimized low-dimensional feature extraction model; Feature extraction is performed based on the optimized low-dimensional feature extraction model and the power sequence of the load event after normalization; Based on the DBSCAN clustering algorithm, low-dimensional features are clustered to obtain clusters, and then the air-conditioning load patterns to which the clusters belong are screened out.

2. The method for classifying user typical air conditioning load operation modes based on deep clustering according to claim 1 is characterized in that: The method of detecting the collected total power signal to obtain a mutation point in the power signal is as follows: A sliding window with a fixed length is set to scan the total power signal, and several sampling points are set on the total power signal. When the center of the sliding window is located at the i-th sampling point P i When the sliding window scans, the sequence is recorded as SW i : Among them, SW i pre represents the power sample sequence in the front sliding window, SW i post Represents the power sample sequence in the rear sliding window; SW i pre and SW i post The length of SW is w, then i pre and SW i post The definition of is as follows: When SW i pre and SW i post The mean of satisfies formula (3), then P i For the mutation point: μ(SW i post )-μ(SW i pre )>P th_low (3) In the formula, μ() represents the mean calculation, P th_low is the low power threshold used to determine the mutation point.

3. The method for classifying user typical air conditioning load operation modes based on deep clustering according to claim 2 is characterized in that: The load event samples are formed based on the mutation points, specifically: If the i-th sampling point P i and the jth sampling point P j Both are mutation points. It is determined whether the two mutation points belong to the same load event, as shown in formula (4): j-i-2w<T th (4) Among them, T th is the duration threshold; if formula (4) is satisfied, then if the mutation point P i and mutation point P j For the same load event; The coordinate sequence of the load event l is denoted as a closed set D event (l): D event (l)=[L(l),R(l)] (5) Where L(l) and R(l) are the left and right coordinate boundaries of the event, respectively, and are defined as in formula (6): Among them, w means the front and rear sliding windows SW i pre and SW i post Length.

4. The method for classifying user typical air conditioning load operation modes based on deep clustering according to claim 3 is characterized in that: The determination of whether there is a fluctuation event based on the load event sample is specifically as follows: For load event l, if it satisfies formula (7), it is judged to be a fluctuation event: Among them, P th_high To set a high power threshold for identifying fluctuation events; P k is the kth sampling point.

5. The method for classifying user typical air conditioning load operation modes based on deep clustering according to claim 4 is characterized in that: The normalization process of the power sequence of the updated load event is specifically as follows: Among them, P event (n) represents the actual value of the power sequence of the nth load event, min(·) and max(·) represent the minimum and maximum operations in the power sequence, respectively. norm (n) is the normalized power sequence of the nth load event.

6. The method for classifying user typical air conditioning load operation modes based on deep clustering according to claim 5 is characterized in that: The low-dimensional feature extraction network consists of 5 fully connected layers. The number of units in the input layer and the output layer is L, the number of units in the remaining hidden layers is 256, and the activation function of each layer is g=tanh, specifically: Among them, the number of units in the input layer of the low-dimensional feature extraction network is L, and the input data X∈R L ; The weight of the first hidden layer is W1∈R L*256 ; Bias is b1∈R 256 ; The activation function is: g(x) = tanh(x); The output is H1 = tanh(W1X+b1); The weight of the second hidden layer is W2∈R 256*256 , with the bias b2∈R 256 ; The activation function is: g(x) = tanh(x); The output is H2 = tanh(W2H1+b2); The weight of the third hidden layer is W3∈R 256*256 , with the bias b3∈R 256 ; The activation function is: g(x) = tanh(x); The output is H3 = tanh(W3H2 + b3); The number of units in the output layer is L, and the weight W4∈R 256*L ; Bias is b4∈R L ; The activation function is Softmax; The output is Y=W4H3+b4.

7. The method for classifying user typical air conditioning load operation modes based on deep clustering according to claim 6 is characterized in that: The training of the low-dimensional feature extraction model to obtain the optimized low-dimensional feature extraction model is specifically as follows: norm (n) is a zero element added to the end to make its length the same as the original longest load event power sequence length L, and the power waveform feature x of load event n is obtained. n ; The input of the low-dimensional feature extraction network is the set x=[x1,…x N ], the output is a low-dimensional feature set H = [h1, ..., h N ]; train the low-dimensional feature extraction network in an unsupervised manner, and the loss function is defined as formula (10): Among them, H 3 and H 5 They represent the output of the input x after passing through the 3rd and 5th fully connected layers, respectively. (m) and b (m) denote the weight and bias of the mth fully connected layer, w1 and w2 denote the weight factors, ||·|| F represents the F-norm, ||·||2 represents the 2-norm, c is a sparse matrix, and is obtained by formula (11): Among them, c i represents the i-th column of c, c ii Then it means c i The ith element of .

8. The method for classifying user typical air conditioning load operation modes based on deep clustering according to claim 7 is characterized in that: The DBSCAN clustering algorithm is used to cluster low-dimensional features to obtain clusters, and then the air conditioning load mode to which the clusters belong is screened out, specifically: Cluster low-dimensional feature data based on the DBSCAN algorithm to obtain multiple clusters; Based on the visualization tool of clustering results and typical values ​​of load characteristics, clusters not related to air conditioning are excluded; Analyze the remaining clusters and identify the clusters that meet the characteristics of air conditioning load; Through manual inspection, the air conditioning load pattern to which the cluster belongs is determined.

9. The method for classifying user typical air conditioning load operation modes based on deep clustering according to claim 8, characterized in that: The typical values ​​of the load characteristics include: power range and time distribution; the remaining clusters are analyzed to identify the clusters that meet the air-conditioning load characteristics, specifically: based on the typical load characteristics of the air conditioner, determine whether the remaining clusters contain relevant load characteristics; the typical load characteristics of the air conditioner include: high power peak when the air conditioner is started and periodic start-stop mode.

10. A user typical air conditioning load operation mode classification system based on deep clustering, characterized in that: include: A collection module, wherein the collection module collects a total power signal at a user bus; A detection module, wherein the detection module detects the collected total power signal to obtain a mutation point in the power signal; A determination module, wherein the determination module forms a load event sample based on the elements of the mutation point, determines whether a fluctuation event exists based on the load event sample, and if so, removes the fluctuation event that is misdetected due to the load fluctuation, and updates the load event detection result; A normalization processing module, wherein the normalization processing module performs normalization processing on the power sequence of the updated load event; A construction module, wherein the construction module constructs and trains a low-dimensional feature extraction model to obtain an optimized low-dimensional feature extraction model; A feature extraction module, wherein the feature extraction module performs feature extraction based on an optimized low-dimensional feature extraction model and a power sequence of a load event after normalization; The clustering module clusters the low-dimensional features based on the DBSCAN clustering algorithm to obtain clusters, and then screens out the air-conditioning load mode to which the clusters belong.

Images