CN111091247A - Power load prediction method and device based on deep neural network model fusion - Google Patents
Power load prediction method and device based on deep neural network model fusion Download PDFInfo
- Publication number
- CN111091247A CN111091247A CN201911346947.7A CN201911346947A CN111091247A CN 111091247 A CN111091247 A CN 111091247A CN 201911346947 A CN201911346947 A CN 201911346947A CN 111091247 A CN111091247 A CN 111091247A
- Authority
- CN
- China
- Prior art keywords
- neural network
- fusion
- day
- load data
- layer
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 230000004927 fusion Effects 0.000 title claims abstract description 88
- 238000000034 method Methods 0.000 title claims abstract description 42
- 238000003062 neural network model Methods 0.000 title claims abstract description 23
- 238000013528 artificial neural network Methods 0.000 claims abstract description 109
- 238000012549 training Methods 0.000 claims abstract description 44
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 22
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 20
- 230000008014 freezing Effects 0.000 claims abstract description 7
- 238000007710 freezing Methods 0.000 claims abstract description 7
- 238000012545 processing Methods 0.000 claims description 13
- 238000010276 construction Methods 0.000 claims description 9
- 238000013135 deep learning Methods 0.000 claims description 6
- 238000004590 computer program Methods 0.000 claims description 3
- 230000000694 effects Effects 0.000 description 6
- 238000007637 random forest analysis Methods 0.000 description 5
- 230000008569 process Effects 0.000 description 4
- 238000000605 extraction Methods 0.000 description 3
- 239000011159 matrix material Substances 0.000 description 3
- 238000012795 verification Methods 0.000 description 3
- 208000025174 PANDAS Diseases 0.000 description 2
- 208000021155 Paediatric autoimmune neuropsychiatric disorders associated with streptococcal infection Diseases 0.000 description 2
- 240000004718 Panda Species 0.000 description 2
- 235000016496 Panda oleosa Nutrition 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000006870 function Effects 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 238000012706 support-vector machine Methods 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 230000003321 amplification Effects 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000005611 electricity Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000003199 nucleic acid amplification method Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 239000013307 optical fiber Substances 0.000 description 1
- 238000007500 overflow downdraw method Methods 0.000 description 1
- 230000008092 positive effect Effects 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
- 230000003442 weekly effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
- G06Q10/063—Operations research, analysis or management
- G06Q10/0639—Performance analysis of employees; Performance analysis of enterprise or organisation operations
- G06Q10/06393—Score-carding, benchmarking or key performance indicator [KPI] analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/20—Administration of product repair or maintenance
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
- G06Q50/06—Electricity, gas or water supply
Abstract
The invention relates to a day-ahead power load prediction method based on neural network model fusion, which comprises the following steps: acquiring load data of the electric meter before the power load date to be predicted, and clustering and grouping the load data by an AP (access point) clustering algorithm; respectively constructing a CNN neural network, an LSTM neural network and a CNN-LSTM combined neural network, and training the neural networks; discarding the output layers of the three deep neural networks, taking the output of the hidden layer at the previous layer of the output layer as a high-level feature, and combining the output of the hidden layers into a fused high-level feature; freezing the network weights of the trained three networks, adding a multi-channel convolution layer as a model fusion layer, and training parameters of the fusion layer; respectively inputting load data of the previous N days of the date to be predicted into the fusion model to obtain final day-ahead power load data; the invention uses the fused network model to predict the day-ahead load, thereby improving the precision of the day-ahead load prediction.
Description
Technical Field
The invention belongs to the technical field of intelligent power distribution and utilization, and particularly relates to a power load prediction method and device based on deep neural network model fusion.
Background
The important significance of the day-ahead load prediction is that the power company can be guided to optimize the start-stop and maintenance plan of the generator set, so that the running economy of the power system is improved. Currently, with the development of smart grids, the load prediction problem meets new challenges. The wide access of the intelligent electric meter makes it possible to improve the accuracy of the load prediction before the day through the data of the intelligent electric meter. By the deep learning method, a large amount of data information of the intelligent electric meter is obtained based on collection, and the information and the data are analyzed, processed and fed back, so that multisource heterogeneous mass data can be fully utilized, and a new solution idea is provided for the problem that the traditional physical modeling method is difficult to process. The application of the artificial intelligence technology in load prediction is developed, the combination of the big data of the intelligent electric meter and the artificial intelligence technology is promoted, the new demand of intelligent power distribution and utilization on load prediction is met, and the method has important significance for development and construction of intelligent power distribution and utilization.
The method has the advantages that the method carries out day-ahead load prediction based on data of the intelligent electric meter, and has the problems that the sampling frequency of the intelligent electric meter is high, the acquired load curve is more precise, the load curve characteristics formed by all users are complex, the change trend of the load curve is difficult to fully learn and predict by the existing load prediction method based on a single deep neural network, and the neural networks with different characteristics have advantages and disadvantages, and the effect of the neural networks with different depths cannot be fully exerted by predicting by using a single neural network; in addition, the electricity utilization characteristics of different users are different, the same deep neural network is used for modeling, and a consistent prediction result is difficult to obtain.
Therefore, in view of the problems, the invention provides a day-ahead load prediction method based on intelligent electric meter grouping prediction and deep neural network model fusion.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provides a day-ahead load prediction method based on intelligent electric meter grouping prediction and deep neural network model fusion. Firstly, grouping users through an AP clustering algorithm, and then training a plurality of applicable deep neural networks for prediction. The output layer of the deep neural network is cut off, the high-level characteristics of the neural networks with different depths are combined through multi-channel convolution, the training process is repeated, and fusion of different network models is achieved, so that the precision of the load prediction in the day-ahead is improved.
The technical problem to be solved by the invention is realized by adopting the following technical scheme:
the method for predicting the day-ahead power load based on neural network model fusion comprises the following steps:
s1, acquiring load data of the intelligent electric meter M days before the power load date to be predicted, clustering and grouping the load data through an AP clustering algorithm, and distinguishing users with different power utilization modes to enable the users in each group to have similar power utilization modes;
s2, respectively constructing a CNN neural network, an LSTM neural network and a CNN-LSTM combined neural network, and training the neural networks by using the packet data obtained in the step S1;
s3, discarding the output layers of the three deep neural networks trained in the step S2, taking the output of the hidden layer at the layer before the output layer as a high-level feature, and combining the hidden layer outputs of the three deep neural networks into a new feature vector to obtain the high-level feature after the three deep neural networks are fused;
s4, freezing the network weights of the three networks trained in the step S2, adding a multi-channel convolutional layer as a model fusion layer, and training parameters of the fusion layer by using the packet data obtained in the step S1, namely completing the fusion of the three neural networks to form a new fusion model;
s5, respectively inputting the load data of each group of intelligent electric meters in N days before the date to be predicted into the fusion model trained in the step S4 according to each group of user load data obtained in the step S1, predicting each group of load data by the fusion model, and adding each group of predicted values to obtain the final day-ahead power load data;
in the steps S1 and S5, M and N are natural numbers, M is greater than or equal to 150, and N is greater than or equal to 3 and less than or equal to 10.
Further, the neural network combining the CNN neural network, the LSTM neural network and the CNN-LSTM in step S2 is constructed and trained using the keras deep learning toolkit in python programming language.
Further, the training method of the neural network combining the CNN neural network, the LSTM neural network and the CNN-LSTM in step S2 includes:
the intelligent ammeter group clustered in the step S1 is segmented into a sample set used for training, specifically: and (4) dividing each group of the clustered intelligent electric meters in the step (S1) into a plurality of samples, wherein each sample consists of N +1 days, using the previous N days as the input of the network, using the next day as the label of the network and the reference value when evaluating the model, and respectively training the three deep neural networks constructed in the step (S2) on each sample set to obtain corresponding deep neural network models.
Further, the training method of the fusion layer parameters in step S4 includes:
the intelligent ammeter group clustered in the step S1 is segmented into a sample set used for training, specifically: and (4) dividing each group of the intelligent electric meters subjected to clustering in the step (S1) into a plurality of samples, wherein each sample consists of N +1 days, using the previous N days as the input of the network, and using the next day as the label of the network and the reference value when evaluating the model, and respectively training the fusion layer parameters in the step (S4) on each sample set to obtain the corresponding fusion model.
Further, the load data of each group of smart meters in step S5 refers to: and adding the load data of all the users in each group to obtain the load data of each group of intelligent electric meters.
A day-ahead power load prediction device based on neural network model fusion comprises:
the intelligent electric meter load data acquiring and grouping module is used for acquiring load data of the intelligent electric meter M days before the day-ahead power load date to be predicted, clustering and grouping the load data through an AP clustering algorithm, and distinguishing users with different power utilization modes, so that each group of users have similar power utilization modes;
the deep neural network construction and training module is used for respectively constructing a neural network combining a CNN neural network, an LSTM neural network and a CNN-LSTM neural network, and training the neural network based on the grouped data obtained in the intelligent electric meter load data acquisition grouping module;
the advanced feature acquisition module is used for discarding the output layers of the three deep neural networks trained in the deep neural network construction and training module, taking the output of the hidden layer at the previous layer of the output layer as an advanced feature, and combining the hidden layer outputs of the three deep neural networks into a new feature vector to obtain the advanced feature after the three deep neural networks are fused;
the deep neural network fusion and training module is used for freezing the network weights of the three networks trained by the advanced characteristic acquisition module, adding a multi-channel convolution layer as a model fusion layer, and training parameters of the fusion layer by using the packet data obtained in the intelligent electric meter load data acquisition grouping module, namely completing the fusion of the three neural networks so as to form a new fusion model;
and the day-ahead power load data prediction acquisition module is used for inputting the load data of each group of intelligent electric meters in the N days before the date to be predicted into a trained fusion model in the deep neural network fusion and training module, predicting each group of load data by using the fusion model, and adding the predicted values to obtain the final day-ahead power load data.
A computing device, comprising:
one or more processing units;
a storage unit for storing one or more programs,
wherein the one or more programs, when executed by the one or more processing units, cause the one or more processing units to perform the method for predicting a power load based on neural network model fusion as described above.
A computer-readable storage medium having non-volatile program code executable by a processor, the computer program when executed by the processor implementing the steps of the neural network model fusion based day-ahead power load prediction method described above.
The invention has the advantages and positive effects that:
according to the method, through a load curve of an ammeter, clustering grouping is carried out through an AP clustering algorithm, then a plurality of applicable deep neural networks are trained through user load data after clustering grouping, an output layer of each deep neural network is cut off, advanced features of the neural networks with different depths are combined through multi-channel convolution, trained network weights are frozen, the training process is repeated, accordingly fusion of different network models is achieved, day-ahead load prediction is carried out through the fused network models, and the precision of the day-ahead load prediction is improved.
Drawings
The technical solutions of the present invention will be described in further detail below with reference to the accompanying drawings and examples, but it should be understood that these drawings are designed for illustrative purposes only and thus do not limit the scope of the present invention. Furthermore, unless otherwise indicated, the drawings are intended to be illustrative of the structural configurations described herein and are not necessarily drawn to scale.
Fig. 1 is a load curve of a first group after clustering smart meters by using AP clustering according to an embodiment of the present invention;
fig. 2 is a second group of load curves after grouping smart meter clusters by using AP clustering according to an embodiment of the present invention;
fig. 3 is a load curve of a third group obtained by grouping smart meter clusters by using AP clustering according to an embodiment of the present invention;
fig. 4 is a fourth group of load curves obtained by grouping clusters of smart meters by using AP clustering according to the embodiment of the present invention;
fig. 5 is a load curve of a fifth group obtained by clustering and grouping the smart meters by using AP clustering according to the embodiment of the present invention;
fig. 6 is a comparison graph of the average absolute error MAE indexes obtained by performing direct prediction and block prediction on different algorithms provided in the embodiment of the present invention and the fusion model of the embodiment;
FIG. 7 is a comparison graph of MAPE indexes of mean absolute percentage errors obtained by direct prediction and block prediction of different algorithms and fusion models of this embodiment;
fig. 8 is a comparison graph of mean square error MSE indexes obtained by direct prediction and block prediction performed by different algorithms and the fusion model of the embodiment of the present invention.
In fig. 6, 7 and 8, RF-Direct indicates that the prediction is directly performed by using an RF algorithm; SVM-Direct means that an SVM algorithm is used for directly predicting; LSTM-group represents that group prediction is carried out by using LSTM; LSTM-Direct indicates that prediction is directly performed by using LSTM; LSTM-CNN-Grouped represents that LSTM-CNN is used for grouping prediction; LSTM-CNN-Direct indicates that LSTM-CNN is used for Direct prediction; CNN-group represents that the CNN is used for grouping prediction; CNN-Direct indicates that CNN is used for Direct prediction; model Fusion-Direct indicates that the Fusion Model of the invention is used for Direct prediction; model Fusion-Grouped represents that the Fusion Model of the invention is used for carrying out grouping prediction;
Detailed Description
First, it should be noted that the specific structures, features, advantages, etc. of the present invention will be specifically described below by way of example, but all the descriptions are for illustrative purposes only and should not be construed as limiting the present invention in any way. Furthermore, any individual technical features described or implicit in the embodiments mentioned herein may still be continued in any combination or subtraction between these technical features (or their equivalents) to obtain still further embodiments of the invention that may not be mentioned directly herein.
It should be noted that the embodiments and features of the embodiments in the present application may be combined with each other without conflict.
The present invention will be described in detail with reference to fig. 1 to 8.
Example 1
As shown in fig. 1 to 8, in this embodiment, by taking the day-ahead load prediction of 1097 smart meter users in a certain area as an example, the area totally contains load data of all the users 2013-01-01 to 2014-01-01 for one year, and the day-ahead power load prediction is performed by using the day-ahead power load prediction method of the present invention, which includes the following steps:
s1, clustering and grouping the users through an AP clustering algorithm, and distinguishing the users with different power utilization modes to ensure that each group of users have similar power utilization modes;
compared with the traditional methods such as k-means and DBSCAN, the clustering number of the AP clustering does not need to be firstly distributed, random initialization is not needed, and when the AP clustering algorithm is executed for multiple times, the result is completely the same, which means that the power consumption group obtained from clustering is more stable. The most representative point in the cluster is called a class representative point in the AP clustering algorithm. Unlike the cluster centers in other conventional algorithms, the class representative points are precise data points in the original data, not the cluster centers obtained by averaging a plurality of data points. This means that we can obtain a true typical power consumer in each group, which facilitates the following studies on consumer power consumption behavior.
In this embodiment, the main process of grouping the smart meter clusters by using AP clustering is as follows:
step 1: assuming there are N users in the total load, the weekly average power curve for each user in the year is obtained and denoted Xi(i is 1, …, N), obtaining the similarity (i.e. euclidean distance) between every two users by calculating the sum of the power difference values of the points of the two power curves, and obtaining an N × N similarity matrix S;
step 2: assume that there are two key matrices: attraction information matrix (R ═ R (i, k)]N×N) And attribution information matrix (a ═ a (i, k)]N×N) (ii) a r (i, k) describes the degree to which data object k fits as the cluster center for data object i; a (i, k) describes how well data object i chooses data object k as its cluster center; all elements in R and a are first initialized to 0 and then calculated by the following formulas:
r(i，j)＝s(i，j)-max{a(i，k)+s(i，k)}，k∈1，2，...，N and k≠j
and step 3: updating A and R alternately by iteration through the formula, calculating attribution information of each point after each iteration (namely calculating a clustering center of the point), finishing clustering when the attribution information of each point does not change after at least 15 continuous iterations, wherein the calculation formula of the attribution information of each point is as follows:
k＝argmaxk{a(i,k)+r(i,k)}，k∈1,2,…N
in the formula: if i ≠ k, then point i itself is the cluster center, and if i ≠ k, then k is the cluster center of point i. After clustering is finished, the number of clustering centers is the grouping number, the grouping number is determined by an algorithm without self-designation, and the k value is the id of the typical user corresponding to each group.
The intelligent electric meters are grouped through AP clustering, 5 groups of users are automatically determined by the algorithm, the power utilization curves of all groups of users are drawn, as shown in figures 1, 2, 3, 4 and 5, it can be seen that the AP clustering effectively distinguishes the users with different power utilization modes, all groups of users have similar power utilization modes, and the load curves of all groups of users are added, so that the load curve of each group of users can be obtained;
s2, respectively constructing a CNN neural network, an LSTM neural network and a CNN-LSTM combined neural network, and training the neural networks;
in this embodiment, a neural network, a pure CNN neural network and a pure LSTM neural network which are combined with CNN-LSTM are respectively constructed, the neural network construction and training both use a keras deep learning toolkit in python programming language, and the network structures are shown in tables 1, 2 and 3; it should be noted that the final output of the neural network is consistent with the number of sampling points of the power curve of the next day.
TABLE 1 neural network architecture for CNN-LSTM
TABLE 2 neural network architecture for CNN
TABLE 3 neural network architecture for LSTM
And respectively making a sample set for each group of intelligent electric meter data obtained after clustering by using a pandas data processing toolkit of python programming language, wherein 5 sample sets are formed. Each sample set comprises a plurality of samples, each sample comprises a power curve of 8 days, a 0 point in each night is used as a reference point, the power curve of 7 days before the time point is used as an input of the neural network, and the power curve of 1 day after the time point is used as a label of the neural network. Since 365 days of data a year are shared, each sample set can divide the data of a year into 365-7-358 samples.
It should be noted that, for comparison, the pandas data processing kit in python programming language is used to make the total data of the non-clustered electric meters into a sample set, which is compared with the sample set of the clustered 5 groups of smart electric meters, and 80 percent of the samples in the 6 sample sets are specified to train the neural network, and the rest 20 percent of the samples are used to perform the effect verification. Respectively training three constructed deep neural networks on each sample set to obtain corresponding deep neural network models;
through comparison, the following results are found: by utilizing three deep neural networks after 5 groups of intelligent electric meters are trained through clustering, more internal details can be obtained through grouping prediction for networks with strong feature extraction functions such as CNN and CNN-LSTM, and therefore higher prediction precision is achieved.
S3, discarding the output layers of the three deep neural networks trained in the step S2, taking the output of the hidden layer at the layer before the output layer as a high-level feature, and combining the hidden layer outputs of the three deep neural networks into a new feature vector to obtain the high-level feature after the three deep neural networks are fused;
s4, freezing the network weights of the three networks trained in the step S2, adding a multi-channel convolutional layer as a model fusion layer, and training parameters of the fusion layer by using the sample set of the 5 groups of clustered intelligent electric meters in the step S2, namely completing the fusion of the three neural networks to form a new fusion model;
the invention adopts the multi-channel convolution pair as the fusion layer, and the reason is that the characteristics acquired by the neural networks with different depths are advanced characteristics, the number of the layers of the network is continuously deepened, which inevitably brings the loss of the characteristics, and the multi-channel convolution respectively adopts different convolution cores to convolute the same object, which is equivalent to observing the same object from different angles, so the fusion of the advanced characteristics is more effective, and the network structure of the fusion layer is shown in table 4:
TABLE 4 fusion layer network architecture
S5, inputting the load data of each group of smart meters 7 days before the date to be predicted into the fusion model trained in step S4, wherein the load data of each group of smart meters is: adding the load data of all users in each group to obtain the load data of each group of intelligent electric meters; and predicting each group of load data by the fusion model, and then adding the predicted values of each group to obtain the final day-ahead power load data.
For example, in this embodiment, a verification set of 20 percent in each sample set is used to perform effect verification on the three constructed neural networks and the fusion model, a power curve of the previous seven days in each sample is used as an input of each prediction model, the model predicts the input to obtain a predicted value of a power curve of the next day, the predicted value is compared with a real power curve of the day, and three evaluation indexes, namely, a Mean Absolute Error (MAE), a Mean Absolute Percentage Error (MAPE), and a Mean Square Error (MSE), are calculated, where the following are detailed experimental results:
1. fusion effect analysis
In step S1, we divide the smart meters into 5 groups by AP clustering, and for each sample set, the experimental results before and after fusion of the three neural networks are shown in table 5:
TABLE 5 Pre-and post-fusion error results
2. Packet prediction effect analysis
Load prediction was performed for this example using CNN, LSTM, CNN-LSTM and model fusion methods. By contrast, we have also added some traditional algorithms, such as Random Forest (RF) algorithms and Support Vector Machines (SVMs). The direct prediction means that a total power curve formed by all users is directly input to each model for prediction, and the grouping prediction means that the users are firstly grouped by using the steps of the example S1-S5, then each group of users is individually predicted, and finally the results are added to obtain the final prediction result. The results are shown in FIGS. 6, 7 and 8:
the results show that the conventional methods (e.g., SVM and RF) are significantly worse than the deep learning method. In both direct prediction and group prediction, model fusion has higher accuracy than other methods. Particularly, after the grouping prediction, the error of model fusion is further reduced, and the fact that the grouping prediction and the model fusion algorithm can work cooperatively is proved, so that the performance of the intelligent electric meter is further improved. For other deep learning methods, the effect of group prediction is not prominent. For networks with powerful feature extraction functions, such as CNN and CNN-LSTM, more internal details can be obtained by packet prediction, thereby realizing higher prediction accuracy. However, for networks with poor feature extraction (e.g., LSTM), packet prediction can lead to false amplification and reduced accuracy.
A day-ahead power load prediction device based on neural network model fusion comprises:
the intelligent electric meter load data acquiring and grouping module is used for acquiring load data of the intelligent electric meter M days before the day-ahead power load date to be predicted, clustering and grouping the load data through an AP clustering algorithm, and distinguishing users with different power utilization modes, so that each group of users have similar power utilization modes;
the deep neural network construction and training module is used for respectively constructing a neural network combining a CNN neural network, an LSTM neural network and a CNN-LSTM neural network, and training the neural network based on the grouped data obtained in the intelligent electric meter load data acquisition grouping module;
the advanced feature acquisition module is used for discarding the output layers of the three deep neural networks trained in the deep neural network construction and training module, taking the output of the hidden layer at the previous layer of the output layer as an advanced feature, and combining the hidden layer outputs of the three deep neural networks into a new feature vector to obtain the advanced feature after the three deep neural networks are fused;
the deep neural network fusion and training module is used for freezing the network weights of the three networks trained by the advanced characteristic acquisition module, adding a multi-channel convolution layer as a model fusion layer, and training parameters of the fusion layer by using the packet data obtained in the intelligent electric meter load data acquisition grouping module, namely completing the fusion of the three neural networks so as to form a new fusion model;
and the day-ahead power load data prediction acquisition module is used for inputting the load data of each group of intelligent electric meters in the N days before the date to be predicted into a trained fusion model in the deep neural network fusion and training module, predicting each group of load data by using the fusion model, and adding the predicted values to obtain the final day-ahead power load data.
A computing device, comprising:
one or more processing units;
a storage unit for storing one or more programs,
wherein the one or more programs, when executed by the one or more processing units, cause the one or more processing units to perform the method for predicting a power load based on neural network model fusion as described above; it is noted that the computing device may include, but is not limited to, a processing unit, a storage unit; those skilled in the art will appreciate that the computing device including the processing unit, the memory unit do not constitute a limitation of the computing device, may include more components, or combine certain components, or different components, for example, the computing device may also include input output devices, network access devices, buses, etc.
A computer-readable storage medium having non-volatile program code executable by a processor, the computer program when executed by the processor implementing the steps of the neural network model fusion based day-ahead power load prediction method described above; it should be noted that the readable storage medium can be, for example but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof; the program embodied on the readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing. For example, program code for carrying out operations of the present invention may be written in any combination of one or more programming languages, including an object oriented programming language such as Java, C + + or the like and conventional procedural programming languages, such as the C programming language or similar programming languages. The program code may execute entirely on the user's computing device, partly on the user's device, as a stand-alone software package, or entirely on a remote computing device or server. In situations involving remote computing devices, the remote computing devices may be connected to the user computing device through any kind of network, including a Local Area Network (LAN) or a Wide Area Network (WAN), or may be connected to external computing devices (e.g., through the internet using an internet service provider).
The present invention has been described in detail with reference to the above examples, but the description is only for the preferred examples of the present invention and should not be construed as limiting the scope of the present invention. All equivalent changes and modifications made within the scope of the present invention shall fall within the scope of the present invention.
Claims (9)
1. The day-ahead power load prediction method based on neural network model fusion is characterized by comprising the following steps of: the day-ahead power load prediction method comprises the following steps:
s1, acquiring load data of the intelligent electric meter M days before the power load date to be predicted, clustering and grouping the load data through an AP clustering algorithm, and distinguishing users with different power utilization modes to enable the users in each group to have similar power utilization modes;
s2, respectively constructing a CNN neural network, an LSTM neural network and a CNN-LSTM combined neural network, and training the neural networks by using the packet data obtained in the step S1;
s3, discarding the output layers of the three deep neural networks trained in the step S2, taking the output of the hidden layer at the layer before the output layer as a high-level feature, and combining the hidden layer outputs of the three deep neural networks into a new feature vector to obtain the high-level feature after the three deep neural networks are fused;
s4, freezing the network weights of the three networks trained in the step S2, adding a multi-channel convolutional layer as a model fusion layer, and training parameters of the fusion layer by using the packet data obtained in the step S1, namely completing the fusion of the three neural networks to form a new fusion model;
s5, respectively inputting the load data of each group of intelligent electric meters in N days before the date to be predicted into the fusion model trained in the step S4 according to each group of user load data obtained in the step S1, predicting each group of load data by the fusion model, and adding each group of predicted values to obtain the final day-ahead power load data;
in steps S1 and S5, M and N are both natural numbers, and M > N. .
2. The neural network model fusion-based day-ahead power load prediction method according to claim 1, characterized in that: the construction and training of the CNN neural network, the LSTM neural network and the CNN-LSTM combined neural network in the step S2 all use a keras deep learning toolkit in python programming language.
3. The neural network model fusion-based day-ahead power load prediction method according to claim 1, characterized in that: the training method of the CNN neural network, the LSTM neural network and the CNN-LSTM combined neural network in step S2 includes:
the intelligent ammeter group clustered in the step S1 is segmented into a sample set used for training, specifically: and (4) dividing each group of the clustered intelligent electric meters in the step (S1) into a plurality of samples, wherein each sample consists of N +1 days, using the previous N days as the input of the network, using the next day as the label of the network and the reference value when evaluating the model, and respectively training the three deep neural networks constructed in the step (S2) on each sample set to obtain corresponding deep neural network models.
4. The neural network model fusion-based day-ahead power load prediction method according to claim 1, characterized in that: the training method of the fusion layer parameters in the step S4 includes:
the intelligent ammeter group clustered in the step S1 is segmented into a sample set used for training, specifically: and (4) dividing each group of the intelligent electric meters subjected to clustering in the step (S1) into a plurality of samples, wherein each sample consists of N +1 days, using the previous N days as the input of the network, and using the next day as the label of the network and the reference value when evaluating the model, and respectively training the fusion layer parameters in the step (S4) on each sample set to obtain the corresponding fusion model.
5. The neural network model fusion-based day-ahead power load prediction method according to claim 1, characterized in that: in step S5, the load data of each group of smart meters refers to: and adding the load data of all the users in each group to obtain the load data of each group of intelligent electric meters.
6. The neural network model fusion-based day-ahead power load prediction method according to claim 1, characterized in that: in the steps S1 and S5, M and N are natural numbers, M is more than or equal to 150, and N is more than or equal to 3 and less than or equal to 10.
7. A day-ahead power load prediction device based on neural network model fusion is characterized in that: the method comprises the following steps:
the intelligent electric meter load data acquiring and grouping module is used for acquiring load data of the intelligent electric meter M days before the day-ahead power load date to be predicted, clustering and grouping the load data through an AP clustering algorithm, and distinguishing users with different power utilization modes, so that each group of users have similar power utilization modes;
the deep neural network construction and training module is used for respectively constructing a neural network combining a CNN neural network, an LSTM neural network and a CNN-LSTM neural network, and training the neural network based on the grouped data obtained in the intelligent electric meter load data acquisition grouping module;
the advanced feature acquisition module is used for discarding the output layers of the three deep neural networks trained in the deep neural network construction and training module, taking the output of the hidden layer at the previous layer of the output layer as an advanced feature, and combining the hidden layer outputs of the three deep neural networks into a new feature vector to obtain the advanced feature after the three deep neural networks are fused;
the deep neural network fusion and training module is used for freezing the network weights of the three networks trained by the advanced characteristic acquisition module, adding a multi-channel convolution layer as a model fusion layer, and training parameters of the fusion layer by using the packet data obtained in the intelligent electric meter load data acquisition grouping module, namely completing the fusion of the three neural networks so as to form a new fusion model;
and the day-ahead power load data prediction acquisition module is used for inputting the load data of each group of intelligent electric meters in the N days before the date to be predicted into a trained fusion model in the deep neural network fusion and training module, predicting each group of load data by using the fusion model, and adding the predicted values to obtain the final day-ahead power load data.
8. A computing device, characterized by: the method comprises the following steps:
one or more processing units;
a storage unit for storing one or more programs,
wherein the one or more programs, when executed by the one or more processing units, cause the one or more processing units to perform the method of any of claims 1-6.
9. A computer-readable storage medium with non-volatile program code executable by a processor, characterized in that the computer program realizes the steps of the method according to any one of claims 1 to 6 when executed by the processor.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911346947.7A CN111091247A (en) | 2019-12-24 | 2019-12-24 | Power load prediction method and device based on deep neural network model fusion |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911346947.7A CN111091247A (en) | 2019-12-24 | 2019-12-24 | Power load prediction method and device based on deep neural network model fusion |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN111091247A true CN111091247A (en) | 2020-05-01 |
Family
ID=70396781
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911346947.7A Pending CN111091247A (en) | 2019-12-24 | 2019-12-24 | Power load prediction method and device based on deep neural network model fusion |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111091247A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112001486A (en) * | 2020-08-28 | 2020-11-27 | 河北工业大学 | Load decomposition method based on deep learning |
| CN112308342A (en) * | 2020-11-25 | 2021-02-02 | 广西电网有限责任公司北海供电局 | Daily load prediction method based on deep time decoupling and application |
| CN112329997A (en) * | 2020-10-26 | 2021-02-05 | 国网河北省电力有限公司雄安新区供电公司 | Power demand load prediction method and system, electronic device, and storage medium |
| CN112653241A (en) * | 2020-12-16 | 2021-04-13 | 南方电网深圳数字电网研究院有限公司 | Power grid monitoring method and device and Internet of things metering unit |
| CN112800685A (en) * | 2021-03-26 | 2021-05-14 | 湖南大学 | High-precision motor multi-physical field coupling simulation calculation method |
| CN112926645A (en) * | 2021-02-22 | 2021-06-08 | 国网四川省电力公司营销服务中心 | Electricity stealing detection method based on edge calculation |
| CN113837227A (en) * | 2021-08-26 | 2021-12-24 | 北京智芯微电子科技有限公司 | Load prediction method, device, chip, electronic device and storage medium |
| CN117557304A (en) * | 2024-01-11 | 2024-02-13 | 国网浙江省电力有限公司 | Electric quantity and electricity price level fusion prediction method based on modal decomposition and neural network |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109740420A (en) * | 2018-11-22 | 2019-05-10 | 深圳市华尊科技股份有限公司 | Vehicle illegal recognition methods and Related product |
| CN109934375A (en) * | 2018-11-27 | 2019-06-25 | 电子科技大学中山学院 | Power load prediction method |
| WO2019141040A1 (en) * | 2018-01-22 | 2019-07-25 | 佛山科学技术学院 | Short term electrical load predication method |
| US20190379589A1 (en) * | 2018-06-12 | 2019-12-12 | Ciena Corporation | Pattern detection in time-series data |
-
2019
- 2019-12-24 CN CN201911346947.7A patent/CN111091247A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2019141040A1 (en) * | 2018-01-22 | 2019-07-25 | 佛山科学技术学院 | Short term electrical load predication method |
| US20190379589A1 (en) * | 2018-06-12 | 2019-12-12 | Ciena Corporation | Pattern detection in time-series data |
| CN109740420A (en) * | 2018-11-22 | 2019-05-10 | 深圳市华尊科技股份有限公司 | Vehicle illegal recognition methods and Related product |
| CN109934375A (en) * | 2018-11-27 | 2019-06-25 | 电子科技大学中山学院 | Power load prediction method |
Non-Patent Citations (1)
| Title |
|---|
| HAIWEN CHEN: "Day-ahead aggregated load forecasting based on two-terminal sparse coding and deep neural network fusion", ELSEVIER * |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112001486A (en) * | 2020-08-28 | 2020-11-27 | 河北工业大学 | Load decomposition method based on deep learning |
| CN112001486B (en) * | 2020-08-28 | 2022-07-22 | 河北工业大学 | Load decomposition method based on deep learning |
| CN112329997A (en) * | 2020-10-26 | 2021-02-05 | 国网河北省电力有限公司雄安新区供电公司 | Power demand load prediction method and system, electronic device, and storage medium |
| CN112308342A (en) * | 2020-11-25 | 2021-02-02 | 广西电网有限责任公司北海供电局 | Daily load prediction method based on deep time decoupling and application |
| CN112653241A (en) * | 2020-12-16 | 2021-04-13 | 南方电网深圳数字电网研究院有限公司 | Power grid monitoring method and device and Internet of things metering unit |
| CN112926645A (en) * | 2021-02-22 | 2021-06-08 | 国网四川省电力公司营销服务中心 | Electricity stealing detection method based on edge calculation |
| CN112800685A (en) * | 2021-03-26 | 2021-05-14 | 湖南大学 | High-precision motor multi-physical field coupling simulation calculation method |
| CN113837227A (en) * | 2021-08-26 | 2021-12-24 | 北京智芯微电子科技有限公司 | Load prediction method, device, chip, electronic device and storage medium |
| CN113837227B (en) * | 2021-08-26 | 2024-02-02 | 北京智芯微电子科技有限公司 | Load prediction method, load prediction device, chip, electronic equipment and storage medium |
| CN117557304A (en) * | 2024-01-11 | 2024-02-13 | 国网浙江省电力有限公司 | Electric quantity and electricity price level fusion prediction method based on modal decomposition and neural network |
| CN117557304B (en) * | 2024-01-11 | 2024-03-29 | 国网浙江省电力有限公司 | Electric quantity and electricity price level fusion prediction method based on modal decomposition and neural network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111091247A (en) | Power load prediction method and device based on deep neural network model fusion | |
| Zhang et al. | Probabilistic solar irradiation forecasting based on variational Bayesian inference with secure federated learning | |
| Jiang et al. | Inbound tourism demand forecasting framework based on fuzzy time series and advanced optimization algorithm | |
| CN103488662A (en) | Clustering method and system of parallelized self-organizing mapping neural network based on graphic processing unit | |
| Li et al. | An efficient manifold regularized sparse non-negative matrix factorization model for large-scale recommender systems on GPUs | |
| CN110751318A (en) | IPSO-LSTM-based ultra-short-term power load prediction method | |
| CN114169442B (en) | Remote sensing image small sample scene classification method based on double prototype network | |
| Feng et al. | Design and simulation of human resource allocation model based on double-cycle neural network | |
| CN110990718A (en) | Social network model building module of company image improving system | |
| CN116187835A (en) | Data-driven-based method and system for estimating theoretical line loss interval of transformer area | |
| CN113591368A (en) | Comprehensive energy system multi-energy load prediction method and system | |
| CN116112563A (en) | Dual-strategy self-adaptive cache replacement method based on popularity prediction | |
| CN116169670A (en) | Short-term non-resident load prediction method and system based on improved neural network | |
| Liu et al. | POI recommendation method using deep learning in location-based social networks | |
| CN110570041A (en) | AP clustering-based prospective year typical daily load prediction method | |
| Kim et al. | Tutorial on time series prediction using 1D-CNN and BiLSTM: A case example of peak electricity demand and system marginal price prediction | |
| CN114078024A (en) | Advertisement inventory estimation method, device, medium and electronic equipment | |
| CN116777646A (en) | Artificial intelligence-based risk identification method, apparatus, device and storage medium | |
| Li et al. | An improved genetic-XGBoost classifier for customer consumption behavior prediction | |
| Majidpour | Time series prediction for electric vehicle charging load and solar power generation in the context of smart grid | |
| CN113780679B (en) | Load prediction method and device based on ubiquitous power Internet of things | |
| CN113762591B (en) | Short-term electric quantity prediction method and system based on GRU and multi-core SVM countermeasure learning | |
| Zhang et al. | Building Energy Consumption Prediction Based on Temporal-Aware Attention and Energy Consumption States | |
| CN115081551A (en) | RVM line loss model building method and system based on K-Means clustering and optimization | |
| Zhang et al. | Intelligent travelling visitor estimation model with big data mining |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20200501 |
|
| WD01 | Invention patent application deemed withdrawn after publication |