A fine-grained load segment prediction method

Abstract

The application discloses a fine-grained load segmentation prediction method, and the steps are as follows: based on the daily load curve, dimension reduction indexes are extracted, and a user power consumption mode clustering model is established; a multi-task group load prediction model is constructed; different categories of user daily load curve segmentation is completed; the fluctuation rate, fluctuation standard deviation, fluctuation degree, high energy consumption proportion and low energy consumption proportion indexes of different fluctuation segments are analyzed, and the type of the fluctuation segment is identified; a group prediction result direct correction prediction model is established for the rising and falling trend load fluctuation segment; a cyclic classification correction prediction model is established for the high energy consumption fluctuation, low energy consumption fluctuation, sharp peak fluctuation and oscillation fluctuation segment; and a fine-grained user complete load curve prediction model is constructed. The application is aimed at fine-grained load which is difficult to predict, improves the prediction accuracy, ensures that the power grid can accurately master the fine-grained load curve change situation, matches the power demand situation of the end user at any time, and helps the rapid development of intelligent energy consumption and point-to-point power transaction.

Description

Technical Field

[0001] The present invention relates to a load forecasting method, and in particular to a fine-grained load curve segmentation forecasting method. Background Art

[0002] With the continuous development of the power system, the grid's demand for smart energy use has become more urgent. In addition, in modern power markets, peer-to-peer (P2P) power trading has gradually emerged. This trading method requires matching the power supply and demand of end users at all times. Accurately grasping the changes in fine-grained load curves is of great significance to the power system. However, fine-grained loads are highly volatile and have poor cyclical regularity, making accurate prediction difficult. Existing load forecasting research mainly focuses on the establishment of models at the system level, sub-level, feeder level, and building level. A universal forecasting model is obtained by uniformly training large-scale loads. This method is not targeted enough and cannot fully meet the current power sector's pursuit of load forecasting accuracy. Summary of the Invention

[0003] Purpose of the invention: In view of the problems existing in the prior art, the purpose of the present invention is to provide a method that can improve the accuracy of fine-grained load forecasting, so that the terminal load can meet the controllable requirements of the power system and participate in the system scheduling.

[0004] Technical solution: The fine-grained load segmentation prediction method of the present invention includes the following steps:

[0005] (1) Extract dimensionality reduction indicators based on daily load curves and establish a user electricity consumption pattern clustering model;

[0006] (2) Construct a multi-task group load forecasting model;

[0007] (3) Complete the segmentation of daily load curves for different types of users;

[0008] (4) Analyze the volatility K, standard deviation σ, volatility R, high energy consumption ratio Ph and low energy consumption ratio P1 of different fluctuation segments to complete the type identification of the fluctuation segment;

[0009] (5) Establishing a group prediction result to directly revise the prediction model for the rising and falling trend fluctuation segments;

[0010] (6) Establish a cyclic classification correction prediction model for high energy consumption fluctuation, low energy consumption fluctuation, peak fluctuation and oscillation fluctuation segments;

[0011] (7) Construct a fine-grained user complete load curve prediction model.

[0012] Furthermore, in step (1), the dimensionality reduction index is extracted based on the daily load curve, including:

[0013] The daily load curve was analyzed and seven daily load characteristic indicators including load rate, daily peak-valley difference rate, maximum utilization hour rate, maximum load occurrence time, minimum load occurrence time, peak load rate and valley load rate were extracted as dimensionality reduction indicators of the residents' daily load curve; among them,

[0014] Among them, the load rate k1 is:

[0015] k1=P ave / P max ;

[0016] Where, P ave is the daily average load; P max is the maximum daily load;

[0017] The maximum utilization hour rate k2 is:

[0018]

[0019] Where, P(t)dt is the daily electricity load; P(t) is the residential load at time t; t is the electricity consumption period;

[0020] The daily peak-to-valley difference rate k3 is:

[0021] k3=(P max -P min ) / P max ;

[0022] Where, P min is the minimum daily load;

[0023] The peak load rate k4 and valley load rate k5 are:

[0024] k4=P ave.peak / P ave ;

[0025] k5=P ave.low / P ave ;

[0026] Where, P ave.peak and P ave.low are the average load during peak period and the average load during valley period respectively; based on the fine-grained load power consumption characteristics and relevant electricity price policies, the peak periods are stipulated as 8:00-12:00 and 14:00-21:00; the valley periods are stipulated as 21:00-24:00 and 0:00-8:00; P ave The average daily load.

[0027] Furthermore, in step (1), a user electricity consumption pattern clustering model is established, including:

[0028] Forming a clustering index matrix from daily load characteristic indicators

[0029] Among them, x i，j The element in the i-th row and j-th column of the matrix represents the j-th daily load characteristic index corresponding to the i-th user, j = 1, 2, ..., 7; n is the number of all fine-grained users within a certain range;

[0030] The matrix is ​​used as the input of K-means algorithm to cluster all fine-grained user electricity consumption patterns within a certain range; K-means algorithm will cluster each user's electricity consumption pattern to each category cluster center C = {C1, C2, C3, ..., C K}Euclidean distance d j，t , divide n users into K categories; Euclidean distance d j，t for:

[0031]

[0032] Among them, y j，t represents the value corresponding to the tth indicator of the jth user; c i，t Indicates the corresponding value of the tth indicator of the i-th cluster center;

[0033]

[0034] Among them, l i is the total number of users of category i; Y n is the set of all types of users; compare the calculation results of the Euclidean distance, traverse each fine-grained user data sample, and divide it into the category to which the cluster center with the smallest Euclidean distance belongs, and obtain the set of K clusters S = {S1, S2, S3, ..., S K}.

[0035] Furthermore, the specific contents of constructing the multi-task group load forecasting model in step (2) include:

[0036] According to the division result of step (1), K categories of users are obtained. All fine-grained user loads in each category are summed and averaged to obtain K-category group load data. These K-category group loads and historical climate data are standardized:

[0037]

[0038] Where x represents a feature in the sample, x* represents the standardized feature value obtained using the max-min method, min and max represent the minimum and maximum values ​​of the sample feature set, respectively;

[0039] The standardized data is used as the input of the fully connected deep neural network to establish a multi-task group load forecasting model. The Adam algorithm is used as the optimization method for the model parameters to complete the parameter adjustment work. The parameter update formula is as follows:

[0040]

[0041] Where θ t represents the parameters of the network, g t represents the gradient, Indicates the gradient solution of the neuron parameters of each layer of the neural network; m t and Represent the gradient g before and after correction respectively t The first moment estimate of v t and They represent the gradient g before and after correction respectively t The default settings of the other parameters are: a=0.001,β1=0.9,β2=0.999,ε=10 -8 .

[0042] Furthermore, the step (3) of using the SDA algorithm to segment the daily load curves of K types of users includes:

[0043] (3.1) Select the time window width ε. The value of ε reflects the sensitivity of the SDA algorithm to noise in the same trend fluctuation segment.

[0044] (3.2) If the starting point is A, then the lengths of BA and CA are the time window width ε. Once points B and C are determined, they will not change.

[0045] (3.3) Find the next load point D on the load curve. The upper and lower windows BA and CA swing towards point D at the same time. Then the upper and lower windows are updated to BD and CD.

[0046] (3.4) Repeat step (3.3) and continuously update the window until the inflection point H is reached. At the same time, to ensure that all load points between A and H are within the range of the upper and lower windows, the upper window BG will not swing to point H, while the lower window continues to swing to point H, and the extension lines of the upper and lower windows still intersect.

[0047] (3.5) The upper window BG remains unchanged, and the lower window swings to point I. At this point, the upper and lower windows are parallel, and the iteration ends;

[0048] (3.6) Update the starting point of the fluctuation segment to I and repeat the above steps for a new round of iteration;

[0049] The calculation formulas for the upper window and lower window are:

[0050]

[0051] Among them, D up Indicates upper window, D down represents the lower window, p0 represents the fine-grained load value at the starting point of the fluctuation segment, p(t) represents the fine-grained load value at time t, and T represents the total number of sample data points;

[0052] During the iteration process, the condition for the end of a fluctuation segment division is:

[0053]

[0054] Among them, t m =mimt refers to t m Take D up ≥D down The minimum time when the conditions are met, the t value at this moment is the end time of the fluctuation segment;

[0055] According to the above steps, the K-type group load curve obtained in step (1) is segmented and finally obtained A fluctuation segment.

[0056] Furthermore, in step (4) The fluctuation rate K, fluctuation standard deviation σ, fluctuation degree R, high energy consumption ratio Ph and low energy consumption ratio Pl of each fluctuation segment, the above five types of load fluctuation characteristic index parameter values ​​are as follows:

[0057] The volatility K is:

[0058]

[0059] Among them, P O Indicates the starting power value of the fluctuation segment; P E Indicates the end power value of the fluctuation segment; T indicates the number of sample points of the daily load curve;

[0060] The standard deviation of fluctuation σ is:

[0061]

[0062] Among them, P v Indicates the power mean of the fluctuation segment; P i Indicates the power value of each sample point on the fluctuation segment; N indicates the number of sample points included in the fluctuation segment;

[0063] The volatility R is:

[0064]

[0065] Among them, |P i -P i+1| represents the power difference between adjacent moments in the fluctuation segment;

[0066] The calculation formulas for high energy consumption ratio Ph and low energy consumption ratio Pl are:

[0067] P h =T H / T;

[0068] P l =T L / T;

[0069] Among them, T H and T L They represent the time when the power value is above 90% of the daily load peak value and the time when the load value is below 110% of the daily load valley value within the fluctuation range respectively;

[0070] The above five types of load fluctuation characteristic index parameter values ​​are used as clustering indicators, and the K-means algorithm is used to classify the load fluctuation characteristic index parameters. The fluctuation segments are clustered and analyzed; the average values ​​of the five index parameters of all fluctuation segments in each category obtained by clustering are calculated; the type of fluctuation segment is determined according to the average value of the parameters; and finally K user load curves divided according to different types of fluctuation segments are obtained.

[0071] Furthermore, in step (5), a direct correction prediction model based on the group prediction results is established:

[0072] The cosine distance of the first-order difference vector of the load is used to reflect the degree of difference in the shape changes of the two load curves and to extract the task-specific representation of each single task. The specific calculation steps of the first-order difference cosine distance are as follows:

[0073] The calculation formula for the first-order difference value of the historical load (average load per household) of various resident groups is:

[0074] z clu，i =x clu，(i+1) -x clu，i i∈[1,n-1];

[0075] The calculation formula for the average first-order difference value of the historical load of a fine-grained user belonging to this category is:

[0076] Z resi，i =x resi，(i+1) -x resi，i i∈[1,n-1];

[0077] Where z clu，i and z resi，i are the ith dimension value of the historical load difference vector of each household in the resident group and the ith dimension value of the historical load difference vector of a fine-grained user in this category, respectively. The total dimension of one day is n = 96, xclu，i , x resi，i are the i-th dimension value after normalization of the historical daily load of each household in the resident group and the i-th dimension value after normalization of the historical load of fine-grained users;

[0078] The calculation formula of the differential cosine distance of the load curve is:

[0079] d c =1-c resi，clu ;

[0080]

[0081] Where, d c is the cosine distance between the difference vector of group load and fine-grained user load; c resi，clu is the cosine similarity of the two load curves; z resi , z clu are the difference vectors of the two daily load curves formed by the first-order difference values ​​of each dimension;

[0082] The load value of the i-th dimension of the fine-grained user load obtained after correction under standardization is defined as:

[0083] x′ resi，i+1 =x′ clu，i +z′ resi，i i∈[1,n-1];

[0084] Where x′ clu，i Represents the i-th dimension value of the group load prediction result under the class standardization of the prediction object, i.e., the fine-grained user; z′ resi，i is the i-th dimension value of the first-order difference vector of fine-grained user load calculated by the following formula:

[0085]

[0086] Among them, z′ clu，i is the i-th dimension value of the standardized first-order difference vector calculated based on the group load output by the FDNN model;

[0087] Furthermore, in the cyclic classification correction prediction method of step (6), the specific steps of a single iteration are:

[0088] (6.1) Based on the prediction results of the tth iteration, the first-order difference cosine distance between the fine-grained load curve of each household and the load curve of each group prediction output (average per household) is calculated, and the reciprocal of this value is As the fitness of the mth user for the kth fine-grained load forecasting model, Refers to the adaptability of the mth household to the kth load forecasting model;

[0089] (6.2) After traversing all fine-grained users and K types of training models, the fitness matrix E(M×K) is formed;

[0090] (6.3) The larger the fitness value, the higher the degree of adaptability of the fine-grained user to the load forecasting model of the corresponding category. Therefore, based on the fitness matrix E, the mth household is reclassified into the category corresponding to the element with the largest fitness value in the mth row of matrix E.

[0091] (6.4) After completing the redivision of all residents, proceed to the prediction phase of this cycle, and the prediction algorithm is the same as step (2);

[0092] (6.5) Repeat multiple cycles until the pre-set cycle end condition is met. After traversing all residents and all categories of training models, the obtained fitness matrix E(M×K) can be expressed as:

[0093]

[0094] The condition for ending the loop is set as follows: the similarity of the categories of each household in the clustering correction results of the t-1th and tth times reaches 65%; or the error of the prediction results of the t-1th and tth times satisfies:

[0095] MAPE(t-1)-MAPE(t)≤0.1%;

[0096] It can be considered that the cycle process has reached stability, and when t = T times, the cycle result is optimal; where MAPE(t) refers to the prediction standard deviation at time t;

[0097] After the cycle ends, the historical load data of each fine-grained user is used as the test set for the prediction model input, and the prediction training model of its category is used as the load predictor for the fine-grained user. The predictor output is the final result of the load prediction for each fine-grained user.

[0098] Furthermore, in step (7), a fine-grained user complete load curve prediction model is constructed, and the steps are as follows:

[0099] Use the group prediction results established in step (5) to directly modify the prediction model and predict the rising and falling trend fluctuation segments;

[0100] The remaining fluctuation segments are predicted and analyzed using the cyclic classification correction prediction method;

[0101] For a certain fine-grained user, the prediction results of all fluctuation segments on the daily load curve to be predicted are obtained, and a simple sum is performed to obtain the complete daily load forecast result of the fine-grained user.

[0102] Beneficial effects: Compared with the existing technology, the present invention has the following significant advantages: 1. It can improve the accuracy of fine-grained load forecasting and accurately grasp the changes in fine-grained load curves; 2. It ensures that the power grid can match the power demand of end users at any time, and promotes the rapid development of smart energy use and point-to-point electricity trading. BRIEF DESCRIPTION OF THE DRAWINGS

[0103] Figure 1 is a flow chart of the present invention;

[0104] Figure 2 Schematic diagram of the fully connected deep neural network structure;

[0105] Figure 3 The load segmentation principle diagram based on the SDA algorithm is shown in Figure 2.

[0106] Figure 4 This is a flow chart for determining the type of fluctuation segment;

[0107] Figure 5 Flowchart of a fine-grained load forecasting method based on correction of group load forecast output curve;

[0108] Figure 6 The flowchart of the cycle classification correction prediction method based on the group load training model is shown;

[0109] Figure 7 This is a flow chart of the fine-grained load segmentation prediction method;

[0110] Figure 8 (a), (b), (c), (d), (e) and (f) are the clustering results of residential users’ electricity consumption patterns and the cluster centers of each category;

[0111] Figure 9 (a), (b), (c), (d), (e) and (f) are the SDA segmentation results of daily load curves for various types of residents;

[0112] Figure 10 (a), (b), (c), (d), (e) and (f) are the classification results of daily load curve types for various types of residents. DETAILED DESCRIPTION

[0113] The technical solution of the present invention is further illustrated below with reference to the accompanying drawings and specific embodiments.

[0114] like Figure 1 As shown, the fine-grained load segmentation prediction method of the present invention includes the following steps:

[0115] Step 1: Analyze the daily load curve and extract dimensionality reduction indicators:

[0116] By analyzing the daily load curve, we extracted seven characteristic indicators: load factor, daily peak-to-valley difference, maximum utilization hour rate, maximum load occurrence time, minimum load occurrence time, peak load factor, and valley load factor. These indicators are used as dimensionality reduction indicators for the residential daily load curve. The definitions and physical meanings of each indicator are shown in Table 1. In this example, combining fine-grained load characteristics and relevant electricity pricing policies, the peak hours are defined as 8:00-12:00 and 14:00-21:00; and the valley hours are defined as 21:00-24:00 and 12:00-8:00.

[0117] Table 1: Characteristic indicators based on daily load curve

[0118]

[0119] In the above table, the calculation formula for load rate is:

[0120] k1=P ave / P max (1-1)

[0121] Among them, P ave is the daily average load; P max The maximum daily load.

[0122] The formula for calculating the maximum utilization hour rate is:

[0123]

[0124] The calculation formula for the daily peak-to-valley difference rate is:

[0125] k3=(P max -P min ) / P max (1-3)

[0126] Among them, P min The minimum daily load.

[0127] The calculation formulas for peak load rate and valley load rate are:

[0128] k4=P ave.peak / P ave (1-4)

[0129] k5=P ave.low / P ave (1-5)

[0130] Among them, P ave.peak and P ave.low are the average load during peak period and the average load during valley period, respectively.

[0131] Step 2: Based on the calculation results of the daily load dimension reduction index of fine-grained users, establish a user power consumption pattern clustering model:

[0132] Forming a clustering index matrix from daily load characteristic indicators The matrix is ​​used as the input of K-means algorithm to cluster the electricity consumption patterns of all fine-grained users within a certain range. Where n is the number of all fine-grained users within a certain range. K-means algorithm will cluster the electricity consumption patterns of each user into cluster centers C = {C1, C2, C3, ..., C K}, and divide n users into K categories. The calculation formula of Euclidean distance is:

[0133]

[0134] Among them, y j，t represents the value corresponding to the tth indicator of the jth user; c i，t The corresponding value of the t-th indicator of the i-th cluster center is calculated as follows:

[0135]

[0136] Among them, l i is the total number of users in the i-th category. Compare the calculation results of the Euclidean distance, traverse each fine-grained user data sample, and divide it into the category of the cluster center with the smallest Euclidean distance, and obtain the set of K clusters S = {S1, S2, S3, ..., S K It should be noted that the number of clusters K of the K-means algorithm needs to be determined in advance using the elbow method and the silhouette coefficient method. The basic principle of the elbow method is that as the number of categories increases, the sum of the squared errors (SSE) of each element in each category will continue to decrease, and the rate of decrease will also decrease gradually; when a certain inflection point is reached, the broken line will tend to be flat as the K value continues to increase, and the value corresponding to that inflection point will be selected as the final number of clusters. The calculation formula of SSE is:

[0137]

[0138] SSE represents the clustering error of all samples and reflects the clustering effect. When the total number of clusters K selected is less than the actual number of clusters, increasing the set K value can significantly improve the intra-cluster aggregation degree of each cluster, which is manifested by a significant decrease in SSE. However, when K gradually increases to approach the actual number of clusters, the improvement in clustering effect is no longer significant, and the SSE curve tends to flatten.

[0139] The formula for calculating the silhouette coefficient of a sample is as follows:

[0140]

[0141] Where a(i) represents the difference between samples within each cluster, which is the average Euclidean distance from the jth sample to the rest of the samples within the cluster. b(i) represents the difference between samples in different clusters, which is the minimum Euclidean distance from the jth sample to the rest of the samples outside the cluster. The average silhouette coefficient of all samples is the silhouette coefficient of the clustering result, which ranges from -1 to 1. A larger value indicates a higher similarity between samples within the same category, and a lower similarity between samples in different categories, indicating a better clustering effect.

[0142] Step 3: Based on the fine-grained user clustering results, a multi-task group load forecasting model for each type of user is constructed:

[0143] According to the clustering results, all fine-grained user loads in each category are summed and averaged to obtain K-category group load data. Then, a single-hidden-layer FDNN is used to construct a multi-task group load prediction model for each type of user. The specific structure of the fully connected deep neural network model with a single hidden layer is as follows: Figure 2 shown.

[0144] The FDNN multi-task load forecasting model aims to minimize the loss function J(w, b) generated by forward propagation and iteratively optimizes the model's parameters w and b at each layer. Considering the Adam learning algorithm's fast computational speed and low memory usage, which can produce a more efficient deep learning model, the Adam algorithm is used as the model parameter optimization method to complete the parameter adjustment work. The parameter update formula of the Adam algorithm is as follows:

[0145]

[0146] Where θ t represents the parameters of the network, g t represents the gradient, m t and Represent the gradient g before and after correction respectively t The first moment estimate of v t and They represent the gradient g before and after correction respectively t The default settings of the other parameters are: a=0.001,β1=0.9,β2=0.999,ε=10 -8 In addition, this embodiment selects the ReLU function as the activation function of the FDNN to further increase the nonlinearity of the model. In order to further improve the prediction effect of the FDNN model and accelerate the convergence speed of the neural network, the min-max normalization method is used to linearly change the original input data:

[0147]

[0148] Among them, min and max represent the minimum and maximum values ​​of the sample feature set, respectively.

[0149] Step 4: Use the SDA algorithm to segment the daily load curves of K-type users:

[0150] Considering the strong randomness of fine-grained loads, a fine-grained load curve segmentation method based on the swinging door algorithm (SDA) is introduced. The SDA algorithm is a relatively simple linear segmentation method with a variable time window. Compared with algorithms with fixed window width, this method can obtain more accurate fine-grained load curve segmentation results. The specific steps of the SDA algorithm are as follows:

[0151] Step 1: Select the time window width ε. The value of ε reflects the sensitivity of the SDA algorithm to noise in the same trend fluctuation segment. It is necessary to select ε based on the actual situation of the example.

[0152] Step 2: If Figure 3 As shown, the starting point is A, then the length of BA and CA is the time window width ε, and points B and C will not change after they are determined;

[0153] Step 3: Find the next load point D on the load curve. The upper and lower windows BA and CA swing toward point D at the same time. The upper and lower windows are then updated to BD and CD.

[0154] Step 4: Repeat step 3 and continue the window update process until the inflection point H is reached. At the same time, in order to ensure that all load points between A and H are within the range of the upper and lower windows, the upper window BG will not swing to point H, while the lower window continues to swing to point H, and the extension lines of the upper and lower windows still intersect;

[0155] Step 5: The upper window BG remains unchanged, and the lower window swings to point I. At this point, the upper and lower windows are parallel, and the iteration ends;

[0156] Step 6: Update the starting point of the fluctuation segment to I and repeat the above steps for a new round of iteration.

[0157] The principle of dividing the daily load curve of residents based on the SDA algorithm is shown in formula (1-12):

[0158]

[0159] Among them, D up Indicates upper window, D down represents the lower window, p0 represents the fine-grained load value at the starting point of the fluctuation segment, p(t) represents the fine-grained load value at time t, and T represents the total number of sample data points, 96.

[0160] During the iterative process, when the upper and lower windows meet the constraints of formula (1-13), the division of this fluctuation segment is completed.

[0161]

[0162] Among them, t m =mint refers to t m Take D up ≥D down The minimum time when the conditions are met, the t value at this moment is the end time of the fluctuation segment.

[0163] Taking into account the commonality of electricity consumption of elements within a class and the differences in electricity consumption characteristics of elements between classes in the clustering results, this embodiment will segment the load curves of K groups respectively, and finally obtain Among them, since the SDA algorithm is a variable window algorithm and the power load characteristics of different categories of users are different, the number of fluctuation segments obtained by dividing the load curve of each group is N k They are also different.

[0164] Step 5: Analyze the five indicators of volatility K, standard deviation σ, volatility R, high energy consumption ratio Ph and low energy consumption ratio P1 in different fluctuation segments to complete the type identification of the fluctuation segment:

[0165] Calculate The five parameter values ​​of the volatility K, standard deviation σ, volatility R, high energy consumption ratio Ph and low energy consumption ratio Pl of each fluctuation segment are shown in Table 2.

[0166] Table 2: Characteristic indicators based on daily load curve

[0167]

[0168] In Table 2, the calculation formula for volatility K is:

[0169]

[0170] Among them, P O Indicates the starting power value of the fluctuation segment; P E Indicates the end power value of the fluctuation segment; T indicates the number of sample points of the daily load curve.

[0171] The calculation formula for the standard deviation of fluctuation σ is:

[0172]

[0173] Among them, P v Indicates the power mean of the fluctuation segment; P i Indicates the power value of each sample point on the fluctuation segment; N indicates the number of sample points contained in the fluctuation segment.

[0174] The calculation formula for volatility R is:

[0175]

[0176] Among them, |P i -P i+1 | represents the power difference between adjacent moments in the fluctuation segment.

[0177] The calculation formulas for high energy consumption ratio Ph and low energy consumption ratio Pl are:

[0178] P h =T H / T (1-17) P l =T L / T (1-18)

[0179] Among them, T H and T L They respectively represent the time when the power value is above 90% of the daily load peak value and the time when the load value is below 110% of the daily load valley value within the fluctuation period.

[0180] The five types of load fluctuation characteristic index parameter values ​​are used as clustering indicators, and the K-means algorithm is used to classify the load fluctuation characteristic parameters. The cluster analysis is performed on each fluctuation segment. The average values ​​of the five index parameters of all fluctuation segments in each category obtained by clustering are calculated, and the type of fluctuation segment is determined based on the average value of the parameters. The specific process of determination is as follows: Figure 4 , and finally K user daily load curves divided into different types of fluctuation segments will be obtained.

[0181] Step 6: Establish a fine-grained load correction prediction model based on the group load prediction output curve:

[0182] The difference between the daily load curve for a single fine-grained user and the average daily load curve for all users in their category lies primarily in their curve shape. Therefore, this embodiment modifies the prediction results based on these different load curve morphological characteristics, ensuring the uniqueness of each single task. This practical application of the principle of "using a sufficiently large dataset to extract common patterns among fine-grained users and using a small dataset specific to each individual to modify each individual model" is achieved.

[0183] It is known that the first-order difference vector of a data set can effectively extract the power change vector that reflects the load's morphological changes, such as increase, decrease, and stability. In addition, the cosine distance between two vectors can be used to characterize the consistency of the different vector directions. Therefore, the cosine distance of the first-order difference vector of the load is introduced to reflect the degree of difference in the morphological changes of the two load curves and to extract the task-specific representation of each single task. The specific calculation steps for the first-order difference cosine distance are as follows:

[0184] The calculation formula for the first-order difference value of the historical load of each group (average load per household) is:

[0185] z clu，i =x clu，(i+1) -x clu，i i∈[1,n-1] (1-19)

[0186] The calculation formula for the average first-order difference value of the historical load of a fine-grained user belonging to this category is:

[0187] z resi，i =x resi，(i+1) -x resi，i i∈[1,n-1] (1-20)

[0188] Where z clu，i and z resi，i are the i-th dimension value of a class group load difference vector and the i-th dimension value of a fine-grained user load difference vector in the class, n=96, x clu，i , x resi，i They are the i-th dimension value after normalization of the group average daily load and the i-th dimension value after normalization of the single fine-grained user load, respectively.

[0189] The calculation formula of the differential cosine distance of the load curve is:

[0190] d c =1-c resi，clu (1-21)

[0191]

[0192] Where, d c is the cosine distance between the difference vector of group load and fine-grained load; c resi，clu is the cosine similarity of the two load curves; z resi , z clu are the difference vectors of the two daily load curves formed by the first-order difference values ​​of all dimensions calculated according to formulas (1-19) and (1-20), respectively.

[0193] Based on the above ideas, the i-th dimension load value of the single fine-grained load obtained after correction under standardization is defined as shown in formula (1-23).

[0194] x′ resi，i+1 =x′ clu，i +z′ resi，i i∈[1,n-1] (1-23)

[0195] Where x′ clu，iRepresents the i-th dimension value of the group load prediction result under the normalization of the class of the prediction object (i.e., a single fine-grained user). resi，i is the i-th dimension value of the single fine-grained load first-order difference vector calculated by formula (1-24):

[0196]

[0197] Among them, z′ clu，i It is the i-th dimension value of the standardized first-order difference vector calculated based on the group load output by the FDNN model.

[0198] Therefore, x′ resi The vector obtained after denormalization is the fine-grained load forecast result. In summary, the process of the fine-grained load correction forecast method based on the group load forecast output curve is as follows: Figure 5 shown.

[0199] Step 7: Establish a cycle classification correction prediction model based on the group load training model:

[0200] While the ultimate goal of the present invention is to achieve a highly accurate, fine-grained load forecast output, the goal of clustering in step two is to achieve a resident classification result with high similarity between intra-class elements and significant differences between inter-class elements. By ensuring that fine-grained user classification is performed into categories with smaller differences in the corresponding historical load curve morphology, thereby enhancing the structural cohesion between the clustering and prediction stages, a fine-grained load forecast model with even higher adaptability and fit can be obtained. Therefore, this step, based on the principle of a loop, proposes a cyclic classification and correction prediction method based on a group load training model.

[0201] The idea of ​​a loop is to repeatedly execute the loop body content when certain conditions are met. Set the total number of loops to T. During each loop, a category division correction and a prediction model training for each K-category group load are performed. After each prediction is completed, the next category division correction is performed based on the calculated fitness matrix. The specific loop body content is: During the t-th iteration, the classification correction is performed based on the fitness matrix calculated in the previous iteration to obtain a new K-category user cluster, and a new round of prediction model training is performed using the updated K-category user historical load data. Each category of users will obtain a prediction model suitable for that type of fine-grained users through training. The specific steps of a single iteration are:

[0202] 1) According to the prediction results of the tth iteration, calculate the first-order difference cosine distance between each user's load curve and the load curve of each group prediction output (average per household), and use the reciprocal of this value As the fitness of the mth user (a total of M = 600 users) for the kth fine-grained load forecasting model, where Refers to the fitness of the mth user and the kth load forecasting model;

[0203] 2) After traversing all users and K types of training models, a fitness matrix E(M×K) is formed;

[0204] 3) The larger the fitness value, the higher the degree of adaptability of the fine-grained user to the load forecasting model of the corresponding category; therefore, based on the fitness matrix E, the mth household is reclassified into the category corresponding to the element with the largest fitness value in the mth row of the matrix E;

[0205] 4) After completing the redivision of all residents, proceed to the prediction phase of this cycle;

[0206] 5) Repeat the loop multiple times until the pre-set loop end condition is met. After traversing all residents and all categories of training models, the obtained fitness matrix E(M×K) can be expressed as:

[0207]

[0208] The loop ends when: the similarity of the categories of each fine-grained user reaches 65% in the clustering correction results before and after the t-1th time; or the error of the prediction results before and after the t-1th time satisfies:

[0209] MAPE(t-1)-MAPE(t)≤0.1% (1-26)

[0210] It can be considered that the cycle process has reached stability, that is, when t = T times, the cycle result is optimal.

[0211] After the cycle is completed, the historical load data of each fine-grained user is used as the test set of the prediction model input, and the prediction training model of its category is used as the load predictor of the fine-grained user. The output of the predictor is the final result of the load prediction of each fine-grained user. In summary, the process of this method is as follows Figure 6 shown.

[0212] Step 8: Establish a fine-grained user complete load curve prediction model:

[0213] When using traditional analysis methods, the prediction results of the upward and downward trend fluctuation segments are ahead or behind; therefore, the group prediction results established in step six are used to directly correct the prediction model and predict the upward and downward trend fluctuation segments.

[0214] Since the cyclic classification and correction prediction model established in step 7 continuously strengthens the structural correlation between the classification and prediction stages through cyclic correction, it generally has the best prediction effect. Therefore, the cyclic classification and correction prediction method is used to predict and analyze the remaining fluctuation segments.

[0215] For a certain fine-grained user, the prediction results of all fluctuation segments on the daily load curve to be predicted are obtained, and by simple addition, the complete daily load prediction result of the fine-grained user can be obtained.

[0216] In order to verify the feasibility and correctness of the present invention, the following examples are given:

[0217] The daily load and temperature data for 600 households in a residential complex in Jiangsu Province were obtained as the data basis for this example analysis. The load rate, daily peak-to-valley difference rate, maximum utilization hour rate, maximum load occurrence time, minimum load occurrence time, peak load rate, and valley load rate parameter values ​​of the 600 households were calculated to form a 600×7 clustering indicator matrix. K-means clustering was performed based on this matrix, and the residential users in the complex were divided into 6 categories. The clustering results are shown below: Figure 8 shown.

[0218] The SDA algorithm is used to segment the daily load curves of the six types of residents. The segmentation results are as follows: Figure 9 As shown in Figure 3, the six types of residential load curves are divided into 109 load fluctuation segments, and the number of fluctuation segments corresponding to each type of residential load curve is [16, 14, 14, 26, 21, 18]. Then, by calculating the fluctuation characteristic index, the fluctuation characteristics of each fluctuation segment are quantitatively described. Based on the calculated values ​​of these five types of load fluctuation characteristic indicators, a clustering index matrix is ​​established, and the K-means algorithm is used to perform cluster analysis on the 109 fluctuation segments. The average values ​​of the fluctuation characteristic indicators of each cluster fluctuation segment obtained by clustering are shown in Table 3. Among them, the last column of the table is based on Figure 4 The type determination results of various types of fluctuation segments completed by the process shown.

[0219] Table 3: Average values ​​of fluctuation characteristic parameters of each cluster fluctuation segment and their determination types

[0220]

[0221] Note: The bold data in the table are the maximum values ​​of the column.

[0222] According to the type determination results of each fluctuation segment, the classification results of the six types of residential daily load curves are obtained as follows: Figure 10 As shown in Figure 2. Different types of fluctuation segments have the following characteristics:

[0223] 1) Downward trend fluctuation: In this type of fluctuation period, the overall residential load shows a downward trend, accompanied by a certain degree of fluctuation;

[0224] 2) Upward trend fluctuation: In this type of fluctuation period, the residential load generally shows an upward trend, accompanied by a certain degree of fluctuation;

[0225] 3) High energy consumption fluctuation: In this type of fluctuation segment, the residential load is large, and the fluctuation degree of this type of fluctuation segment is smaller than that of other fluctuation segments;

[0226] 4) Low energy consumption fluctuation: In this type of fluctuation segment, the residential load is relatively small, and the fluctuation degree of this type of fluctuation segment is also relatively small compared with other fluctuation segments;

[0227] 5) Peak fluctuation: caused by the switching behavior of high-power appliances in the home or the switching behavior of multiple loads at the same time, the load curve shows obvious peaks and the fluctuation is very severe;

[0228] 6) Oscillation: This is caused by the electricity consumption of small-power electrical appliances in the home or the unstable electricity consumption of some electrical appliances, and manifests as random vibration of the residents' load.

[0229] Finally, a fine-grained residential load correction method based on the group load forecast output curve is used to predict and analyze upward and downward trend fluctuations. Simultaneously, a cyclic classification correction prediction method based on the group load training model is used to predict and analyze high energy consumption fluctuations, low energy consumption fluctuations, spike fluctuations, and oscillation fluctuations. Finally, the prediction results for all fluctuation segments are combined to obtain a complete daily load forecast curve.

[0230] Based on the present invention and two other fine-grained residential load forecasting methods, the residential load in a certain community was analyzed and compared with the prediction results of traditional methods. The average MAPE values ​​output by different methods are shown in Table 4.

[0231] Table 4: Comparison of the effects of different methods

[0232]

[0233] Obviously, among the above methods, the segmented prediction method proposed in the present invention can obtain the best prediction effect by using different modified prediction methods to conduct segmented prediction research based on the load characteristics of different fluctuation segments, with a MAPE value of only 34.96%.

Claims

1. A fine-grained load segmentation prediction method, characterized in that: The following steps are involved: (1) Extract dimensionality reduction indicators based on daily load curves and establish a user electricity consumption pattern clustering model; (2) Construct a multi-task group load forecasting model; (3) Complete the segmentation of daily load curves for different types of users; (4) Analyze the volatility K, standard deviation σ, volatility R, high energy consumption ratio Ph and low energy consumption ratio Pl of different fluctuation segments to complete the type identification of the fluctuation segment; (5) Establishing a group prediction result to directly revise the prediction model for the rising and falling trend fluctuation segments; (6) Establish a cyclic classification correction prediction model for high energy consumption fluctuation, low energy consumption fluctuation, peak fluctuation and oscillation fluctuation segments; (7) Construct a fine-grained user complete load curve prediction model; In step (1), the dimensionality reduction index is extracted based on the daily load curve, including: The daily load curve was analyzed and seven daily load characteristic indicators were extracted, including load rate, daily peak-to-valley difference rate, maximum utilization hour rate, maximum load occurrence time, minimum load occurrence time, peak load rate and valley load rate, as dimensionality reduction indicators of the residents' daily load curve; In step (1), a user electricity consumption pattern clustering model is established, including: The clustering index matrix X is formed from the daily load characteristic indicators. This matrix is ​​used as the input of the K-means algorithm to cluster the power consumption patterns of all fine-grained users within a certain range. The K-means algorithm will divide n users into K categories based on the Euclidean distance between each user's power consumption pattern and the cluster center of each category. Compare the calculation results of the Euclidean distance, traverse each fine-grained user data sample, and divide it into the category corresponding to the cluster center with the smallest Euclidean distance, and obtain a set S of K clusters; The specific contents of building a multi-task group load forecasting model in step (2) include: According to the division result of step (1), K categories of users are obtained. All fine-grained user loads in each category are summed and averaged to obtain K-category group load data; these K-category group loads and historical climate data are standardized; Using the standardized data as the input of a fully connected deep neural network, a multi-task group load forecasting model was established. The Adam algorithm was used as the optimization method for the model parameters to complete the parameter adjustment. In step (5), a direct correction prediction model based on the group prediction results is established: The cosine distance of the first-order difference vector of the load is used to reflect the degree of difference in the shape changes of the two load curves and to extract the task-specific representation of each single task. The specific calculation steps of the first-order difference cosine distance are as follows: The calculation formula for the first-order difference value of the historical load (average load per household) of various resident groups is: z clu,i =x clu,(i+1) -x clu,i i∈[1,n-1]; The calculation formula for the average first-order difference value of the historical load of a fine-grained user belonging to this category is: z resi,i =x resi,(i+1) -x resi,i i∈[1,n-1]; Where z clu,i and z resi,i are the ith dimension value of the historical load difference vector of each household in the resident group and the ith dimension value of the historical load difference vector of a fine-grained user in this category, respectively. The total dimension of one day is n = 96, x clu,i ,x resi,i are the i-th dimension value after normalization of the historical daily load of each household in the resident group and the i-th dimension value after normalization of the historical load of fine-grained users; The calculation formula of the differential cosine distance of the load curve is: d c =1-c resi,clu ; Where, d c is the cosine distance between the difference vector of group load and fine-grained user load; c resi,clu is the cosine similarity of the two load curves; z resi ,z clu are the difference vectors of the two daily load curves formed by the first-order difference values ​​of each dimension; The load value of the i-th dimension of the fine-grained user load obtained after correction under standardization is defined as: x' resi,i+1 =x' clu,i +z' resi,i i∈[1,n-1]; Among them, x' clu,i Represents the i-th dimension value of the group load forecast result under the standardization of the class to which the forecast object belongs, i.e., the fine-grained user; z' resi,i is the i-th dimension value of the first-order difference vector of fine-grained user load calculated by the following formula: Among them, z' clu,i It is the i-th dimension value of the standardized first-order difference vector calculated based on the group load output by the FDNN model.

2. The fine-grained load segmentation prediction method according to claim 1 is characterized in that: In step (1), when extracting the dimensionality reduction index based on the daily load curve, the load rate k1 is: k1=P ave / P max ; Where, P ave is the daily average load; P max is the maximum daily load; The maximum utilization hour rate k2 is: Where, is the daily electricity load; P(t) is the resident load at time t; t is the electricity consumption period; The daily peak-to-valley difference rate k3 is: k3=(P max -P min ) / P max ; Where, P min is the minimum daily load; The peak load rate k4 and valley load rate k5 are: k4=P ave.peak / P ave ; k5=P ave.low / P ave ; Where, P ave.peak and P ave.low They are the average load during peak period and the average load during valley period respectively; combined with the fine-grained load power consumption characteristics and relevant electricity price policies, the peak periods are stipulated as: 8:00-12:00 and 14:00-21:00; the valley periods are: 21:00-24:00 and 0:00-8:00; P ave The average daily load.

3. The fine-grained load segmentation prediction method according to claim 2 is characterized in that: In step (1), Forming a clustering index matrix from daily load characteristic indicators Among them, x i,j The element in the i-th row and j-th column of the matrix represents the j-th daily load characteristic index corresponding to the i-th user, j = 1, 2, ..., 7; n is the number of all fine-grained users within a certain range; The matrix is ​​used as the input of K-means algorithm to cluster all fine-grained user electricity consumption patterns within a certain range; K-means algorithm will cluster each user's electricity consumption pattern to each category cluster center C = {C1, C2, C3, ..., C K }Euclidean distance d j,t , divide n users into K categories; Euclidean distance d j,t for: Among them, y j,t represents the value corresponding to the tth indicator of the jth user; c i,t Indicates the corresponding value of the tth indicator of the i-th cluster center; Among them, l i is the total number of users of category i; Y n is the set of all types of users; compare the calculation results of the Euclidean distance, traverse each fine-grained user data sample, and divide it into the category of the cluster center with the smallest Euclidean distance, and obtain the set of K clusters S = {S1, S2, S3, ..., S K }.

4. The fine-grained load segmentation prediction method according to claim 3 is characterized in that: In step (2), when constructing the multi-task group load forecasting model, The standardized data is used as the input of the fully connected deep neural network to establish a multi-task group load forecasting model. The Adam algorithm is used as the optimization method for the model parameters to complete the parameter adjustment work. The parameter update formula is as follows: Where θ t represents the parameters of the network, g t represents the gradient, Indicates the gradient solution of the neuron parameters of each layer of the neural network; m t and Represent the gradient g before and after correction respectively t The first moment estimate of v t and They represent the gradient g before and after correction respectively t The default settings of the other parameters are: a=0.001,β1=0.9,β2=0.999,ε=10 -8 .

5. The fine-grained load segmentation prediction method according to claim 3 is characterized in that: Step (3) of using the SDA algorithm to segment the daily load curves of K-type users includes: (3.1) Select the time window width ε. The value of ε reflects the sensitivity of the SDA algorithm to noise in the same trend fluctuation segment. (3.2) If the starting point is A, then the lengths of BA and CA are the time window width ε. Once points B and C are determined, they will not change. (3.3) Find the next load point D on the load curve. The upper and lower windows BA and CA swing towards point D at the same time. Then the upper and lower windows are updated to BD and CD. (3.4) Repeat step (3.3) and continuously update the window until the inflection point H is reached. At the same time, to ensure that all load points between A and H are within the range of the upper and lower windows, the upper window BG will not swing to point H, while the lower window continues to swing to point H, and the extension lines of the upper and lower windows still intersect. (3.5) The upper window BG remains unchanged, and the lower window swings to point I. At this point, the upper and lower windows are parallel, and the iteration ends; (3.6) Update the starting point of the fluctuation segment to I and repeat the above steps for a new round of iteration; The calculation formulas for the upper window and lower window are: Among them, D up Indicates upper window, D down represents the lower window, p0 represents the fine-grained load value at the starting point of the fluctuation segment, p(t) represents the fine-grained load value at time t, and T represents the total number of sample data points; During the iteration process, the condition for the end of a fluctuation segment division is: Among them, t m =mint refers to t m Take D up ≥D down The minimum time when the conditions are met, the t value at this moment is the end time of the fluctuation segment; According to the above steps, the K-type group load curve obtained in step (1) is segmented and finally obtained A fluctuation segment.

6. The fine-grained load segmentation prediction method according to claim 5 is characterized in that: In step (4) The fluctuation rate K, fluctuation standard deviation σ, fluctuation degree R, high energy consumption ratio Ph and low energy consumption ratio Pl of each fluctuation segment, the above five types of load fluctuation characteristic index parameter values ​​are as follows: The volatility K is: Among them, P O Indicates the starting power value of the fluctuation segment; P E Indicates the end power value of the fluctuation segment; T indicates the number of sample points of the daily load curve; The standard deviation of fluctuation σ is: Among them, P v Indicates the power mean of the fluctuation segment; P i Indicates the power value of each sample point on the fluctuation segment; N indicates the number of sample points included in the fluctuation segment; The volatility R is: Among them, |P i -P i+1 | represents the power difference between adjacent moments in the fluctuation segment; The calculation formulas for high energy consumption ratio Ph and low energy consumption ratio Pl are: P h =T H / T; P l =T L / T; Among them, T H and T L They represent the time when the power value is above 90% of the daily load peak value and the time when the load value is below 110% of the daily load valley value within the fluctuation range respectively; The above five types of load fluctuation characteristic index parameter values ​​are used as clustering indicators, and the K-means algorithm is used to classify the load fluctuation characteristic index parameters. The fluctuation segments are clustered and analyzed; the average values ​​of the five index parameters of all fluctuation segments in each category obtained by clustering are calculated; the type of fluctuation segment is determined according to the average value of the parameters; and finally K user load curves divided according to different types of fluctuation segments are obtained.

7. The fine-grained load segmentation prediction method according to claim 6 is characterized in that: In step (6) of the cyclic classification correction prediction method, the specific steps of a single iteration are: (6.1) Based on the prediction results of the tth iteration, the first-order difference cosine distance between the fine-grained load curve of each household and the load curve of each group prediction output (average per household) is calculated, and the reciprocal of this value is As the fitness of the mth user for the kth fine-grained load forecasting model, Refers to the adaptability of the mth household to the kth load forecasting model; (6.2) After traversing all fine-grained users and K types of training models, the fitness matrix E(M×K) is formed; (6.3) The larger the fitness value, the higher the degree of adaptability of the fine-grained user to the load forecasting model of the corresponding category. Therefore, based on the fitness matrix E, the mth household is reclassified into the category corresponding to the element with the largest fitness value in the mth row of matrix E. (6.4) After completing the redivision of all residents, proceed to the prediction phase of this cycle, and the prediction algorithm is the same as step (2); (6.5) Repeat multiple cycles until the pre-set cycle end condition is met; after traversing all residents and all categories of training models, the obtained fitness matrix E(M×K) is: The condition for ending the loop is set as follows: the similarity of the categories of each household in the clustering correction results of the t-1th and tth times reaches 65%; or the error of the prediction results of the t-1th and tth times satisfies: MAPE(t-1)-MAPE(t)≤0.1%; It can be considered that the cycle process has reached stability, and when t = T times, the cycle result is optimal; where MAPE(t) refers to the prediction standard deviation at time t; After the cycle ends, the historical load data of each fine-grained user is used as the test set for the prediction model input, and the prediction training model of its category is used as the load predictor for the fine-grained user. The predictor output is the final result of the load prediction for each fine-grained user.

8. The fine-grained load segmentation prediction method according to claim 7 is characterized in that: In step (7), a fine-grained user complete load curve prediction model is constructed as follows: Use the group prediction results established in step (5) to directly modify the prediction model and predict the rising and falling trend fluctuation segments; The remaining fluctuation segments are predicted and analyzed using the cyclic classification correction prediction method; For a certain fine-grained user, the prediction results of all fluctuation segments on the daily load curve to be predicted are obtained, and a simple sum is performed to obtain the complete daily load forecast result of the fine-grained user.

Images