CN109754113B - Load prediction method based on dynamic time warping and long-and-short time memory - Google Patents
Load prediction method based on dynamic time warping and long-and-short time memory Download PDFInfo
- Publication number
- CN109754113B CN109754113B CN201811439799.9A CN201811439799A CN109754113B CN 109754113 B CN109754113 B CN 109754113B CN 201811439799 A CN201811439799 A CN 201811439799A CN 109754113 B CN109754113 B CN 109754113B
- Authority
- CN
- China
- Prior art keywords
- data
- load
- short
- long
- users
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 74
- 230000015654 memory Effects 0.000 title claims abstract description 26
- 238000012549 training Methods 0.000 claims abstract description 55
- 238000013528 artificial neural network Methods 0.000 claims abstract description 27
- 238000012545 processing Methods 0.000 claims abstract description 14
- 238000011176 pooling Methods 0.000 claims abstract description 12
- 230000006399 behavior Effects 0.000 claims abstract description 11
- 230000005611 electricity Effects 0.000 claims abstract description 8
- 125000004122 cyclic group Chemical group 0.000 claims abstract description 7
- 230000006403 short-term memory Effects 0.000 claims abstract description 7
- 238000007781 pre-processing Methods 0.000 claims abstract description 4
- 239000010410 layer Substances 0.000 claims description 38
- 238000004422 calculation algorithm Methods 0.000 claims description 22
- 238000012360 testing method Methods 0.000 claims description 18
- 210000002569 neuron Anatomy 0.000 claims description 16
- 230000000306 recurrent effect Effects 0.000 claims description 13
- 239000011159 matrix material Substances 0.000 claims description 11
- 238000005457 optimization Methods 0.000 claims description 11
- 238000004364 calculation method Methods 0.000 claims description 10
- 230000006870 function Effects 0.000 claims description 10
- 230000008569 process Effects 0.000 claims description 9
- 230000002354 daily effect Effects 0.000 claims description 5
- 230000000694 effects Effects 0.000 claims description 5
- 238000010606 normalization Methods 0.000 claims description 5
- 230000003203 everyday effect Effects 0.000 claims description 4
- 230000007613 environmental effect Effects 0.000 claims description 3
- 238000011156 evaluation Methods 0.000 claims description 3
- 238000005259 measurement Methods 0.000 claims description 3
- 230000001537 neural effect Effects 0.000 claims description 3
- 239000002356 single layer Substances 0.000 claims description 3
- 230000003442 weekly effect Effects 0.000 claims description 3
- 230000007787 long-term memory Effects 0.000 abstract description 5
- 210000004027 cell Anatomy 0.000 description 7
- 238000011160 research Methods 0.000 description 5
- 238000013135 deep learning Methods 0.000 description 4
- 230000007774 longterm Effects 0.000 description 3
- YHXISWVBGDMDLQ-UHFFFAOYSA-N moclobemide Chemical compound C1=CC(Cl)=CC=C1C(=O)NCCN1CCOCC1 YHXISWVBGDMDLQ-UHFFFAOYSA-N 0.000 description 3
- 238000013459 approach Methods 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 239000002245 particle Substances 0.000 description 2
- 230000000737 periodic effect Effects 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 101001095088 Homo sapiens Melanoma antigen preferentially expressed in tumors Proteins 0.000 description 1
- 102100037020 Melanoma antigen preferentially expressed in tumors Human genes 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 230000002457 bidirectional effect Effects 0.000 description 1
- 230000001413 cellular effect Effects 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 238000005265 energy consumption Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000003064 k means clustering Methods 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 230000002688 persistence Effects 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
- 238000005406 washing Methods 0.000 description 1
Images
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
- Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
- Y04S10/00—Systems supporting electrical power generation, transmission or distribution
- Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a load prediction method based on dynamic time warping and long-and-short time memory, which comprises the following steps: s1, acquiring basic data required by user short-term load prediction from the power system: s2, clustering users with similar electricity consumption behaviors by using a dynamic time warping method according to historical load data of the users; s3, pooling the user data of the same category; s4, selecting training data, preprocessing the training data and using the preprocessed training data as input; s5, constructing a short-term load prediction method based on the deep long-term and short-term memory cyclic neural network, and verifying effectiveness. According to the method and the device, the users with similar electricity utilization behaviors are clustered aiming at the characteristic that the base number of the users to be predicted is large, and the prediction efficiency is improved. Meanwhile, the data in the same category are subjected to pooling processing, so that the diversity of training data is increased, the short-term load prediction precision is improved, and the method has certain engineering application significance.
Description
Technical Field
The invention relates to a method for predicting short-term load of residents in an electric power system, in particular to a method for predicting short-term load of residents in the electric power system, which is used for predicting load of the electric power system and belongs to the technical field of pattern recognition and image processing.
Background
The load prediction of residents in the power system is based on historical load change rules and combined with factors such as weather and economy to scientifically predict loads in the next days or hours. Accurate load prediction is an important decision basis for scheduling power production and equipment maintenance plans. Therefore, it is necessary to research new methods and techniques for resident load prediction to improve the accuracy and reliability of load prediction and meet the engineering requirements.
In recent years, with the continuous development of deep learning technology, some deep learning models are gradually applied to the research of time series data. Among them, the Recurrent Neural Network (RNN) is a neural network with a self-recurrent structure, allows persistence of information flow of time-series data on a network layer, and is ideally suited to processing time-series data. As research progresses, RNNs have produced numerous variants, such as bidirectional recurrent neural networks, long-short term memory recurrent neural networks (LSTM), gated recurrent neural networks (GRU), and the like. Among numerous RNN variants, the LSTM network effectively solves the problems of gradient disappearance, gradient explosion, insufficient long-term memory and the like of the RNN, so that the recurrent neural network can really and effectively utilize long-distance time sequence information. Based on the technical foundation, how to provide a short-term household power load prediction method based on the LSTM for solving the problems of single dimension and strong data randomness of short-term household power demand data and ensuring that the short-term household power load prediction method has good prediction performance is the key point of research of people in the industry at present.
In addition, in order to further improve the accuracy of the existing short-time household power load prediction method, a combination prediction model is proposed by many scholars. The most typical way is to perform clustering according to different user load curves. The load profile is a 24 hour record of home energy consumption, and while consumer behavior indicates a device usage pattern, the exact time of device usage may vary, resulting in a load profile of a different shape. Thus, the daily 24 hour load curve for a household may show a large variation, whether on different days of the week or on a given day of the week. In order to properly classify households by behavior, the classification scheme must take these changes into account in time and focus on the basic structure of the device usage, which can also be considered as the basis for energy usage by the household.
For this situation, a method of comparing and classifying the home load curves by using Dynamic Time Warping (DTW) has also been developed. Under this metric, the time axis is stretched or shrunk to find the best match between two load curves and to find a canonical "shape" to represent a set of load curves. Under this measurement, two load curves classified as similar do not necessarily have similar values at each time point, but have similar shapes under the DTW measurement.
In summary, how to provide a novel method for predicting the short-term load of residents in an electrical power system based on the prior art, which takes the advantages of the prior art as a basis and gives full play to the advantages of the prior art, improves the accuracy of prediction and reduces the complexity of an algorithm, is a problem to be solved by the technical personnel in the field.
Disclosure of Invention
In view of the above-mentioned drawbacks of the prior art, the present invention provides a load prediction method based on dynamic time warping and long-and-short time memory, which includes the following steps:
s1, acquiring basic data required by user short-term load prediction from the power system;
s2, clustering users with similar electricity consumption behaviors by using a dynamic time warping method according to historical load data of the users;
s3, pooling the user data of the same category;
s4, selecting load data, meteorological data and forecast date type data of the forecast previous day, week and month, preprocessing the data and inputting the preprocessed data;
s5, constructing a short-term load prediction method based on the deep long short-term memory cyclic neural network, and verifying the effectiveness of the method by predicting the scene 24h in advance.
Preferably, the basic data in S1 includes historical load data and meteorological data, the historical load data is load data of 1h interval every day of the historical day, and the meteorological data at least includes the environmental temperature at the predicted time and the predicted day date type.
Preferably, S2 specifically includes the following steps:
s21, assuming that the load sequences of the two users are X, Y, where X ═ X 1 ,x 2 ,...x N }、Y＝{y 1 ,y 2 ,...y N Define a regular path matrix M [24 x 24 ]]The matrix elements (m, n) represent x m And y n Distance between two points, the value of matrix element (m, n) is d (x) m ,y n )＝(x m -y n ) 2 ；
S22, defining dynamic warping path sequence P ═ (P) 1 ,p 2 ,...p k ) Wherein p is k ＝(m k ,n k ) And max (m, n) is more than or equal to K and less than m + n-1;
s23 total cost of dynamic warping between load sequences X and Y
Dynamic time warping distance DTW (X, Y) ═ c, defining X and Y P (X, Y), wherein,
P * ＝argminc P (X, Y), the weekly average DTW distance of the two users is used as the clustering distance in the lower clustering algorithm to obtain
Wherein t represents the day of the week, t is 0 to 6, X represents the period from Monday to Sunday t And Y t The load curves of the t days of the two users are respectively shown.
S24, a series of daily load curves X and the cluster number D of the users are given, the users are clustered into D classes according to the DTW distance of the load curves, and the class D and the attribution function C are obtained d And a clustering center λ d So as to make cluster interior and (S) c ) At minimum, the calculation formula is
Then, randomly selecting a clustering core, and performing iterative updating by adopting a K-medoid-based clustering algorithm.
Preferably, the dynamic warping path sequence in S22 satisfies the following condition:
boundary condition, p 1 Is ═ 1, and p K ＝(N,N)；
Monotonicity Condition, m 1 ≤m 2 ≤...≤m K And n is 1 ≤n 2 ≤...≤n K ；
Step size limiting condition, if p k-1 ＝(a',b')，The next point p of the path k The requirement for (a, b) is that (a-a ') is 1 or less and (b-b') is 1 or less.
Preferably, S3 specifically includes the following steps:
s31, adding an ID tag to the user in a virtual variable mode;
s32, dividing user data into a training set and a test set;
and S33, combining all training data to construct a training pool, and then constructing a test pool through the same process.
Preferably, S4 specifically includes the following steps:
s41, normalizing the resident user load data set L by the normalization formula
Wherein,
normalized data value for a variable, x (i) raw data for the variable, x max 、x min Respectively the maximum value and the minimum value of the original data;
s42, conducting sparse processing on the meteorological factor data set W and the date type I of the user by using one-hot codes, and obtaining an input data set of the user, wherein the input data set is L, W and I;
and S43, pre-training the historical data of the previous day, the previous week and the previous month, then using the pre-training as the input of the prediction model, specifically processing the data of three days in a single-layer full-connection network mode, and finally outputting the data with the same dimension as the input data set of a single day.
Preferably, the method for constructing a short-term load prediction method based on a deep long-short term memory cycle neural network in S5, wherein the structure of the deep long-short term memory cycle neural network comprises: an input layer, a hidden layer and an output layer; the input layer is composed of neurons representing inputs; the hidden layer consists of neurons representing intermediate variables, and the hidden layer and the next hidden layer are directly connected by the neurons; the output layer is composed of neurons representing output results.
Preferably, the method for constructing the short-term load prediction based on the deep long-short-term memory recurrent neural network described in S5 specifically includes the following steps:
s51, carrying out batch division on the user training data in the S4 so as to facilitate calculation;
s52, initializing various parameters of the long-term and short-term memory cyclic neural training network;
s53, inputting training data to perform forward propagation training;
and S54, determining a loss function and an optimization method according to the predicted value and the true value to perform back propagation optimization.
Preferably, in S5, the average absolute percentage error and the mean square error are used as the evaluation indexes of the model prediction effect, and the calculation formulas are respectively,
wherein n is the number of predicted points, y i For the load true value of the ith predicted point,
and predicting the value for the ith predicted point.
Compared with the prior art, the invention has the advantages that:
according to the load prediction method based on dynamic time warping and long-and-short time memory, aiming at the characteristic that the base number of the user to be predicted is large, the users with similar electricity utilization behaviors are clustered based on the DTW similarity, the prediction efficiency is improved, the clustering accuracy of the users is increased, the number of clustering categories is reduced, and the complexity of a clustering algorithm is reduced.
Meanwhile, aiming at the problem that the LSTM load prediction method based on deep learning is easy to over-fit, the diversity of training data is increased and the generalization capability of the model is improved by performing pooling processing on data in the same category, so that the short-term load prediction precision is improved, and the method has certain engineering application significance.
In addition, the invention also provides reference for other related problems in the same field, can be expanded and extended on the basis of the reference, is applied to the technical scheme of other load prediction methods in the same field, and has very wide application prospect.
The following detailed description of the embodiments of the present invention is provided in connection with the accompanying drawings for the purpose of facilitating understanding and understanding of the technical solutions of the present invention.
Drawings
FIG. 1 is a flow chart of a prediction method of the present invention;
FIG. 2 is a load prediction network architecture model based on LSTM;
FIG. 3 is a diagram of the internal structure of an LSTM neuron;
FIG. 4 is a clustering result of original user data;
fig. 5 is a graph showing the predicted effect on 8 months and 8 days.
Detailed Description
As shown in fig. 1, the invention discloses a load prediction method based on dynamic time warping and long-and-short time memory, and the idea of the invention is to cluster power users based on dynamic time warping similarity, and cluster users with similar electricity utilization habits into one class. And then carrying out random pooling operation on users of the same category, and increasing the scale and diversity of training data so as to increase the generalization capability of the prediction model. And then, respectively establishing a load prediction model based on long-time and short-time memory for the user data sets after the pooling, wherein the proposed method has better engineering adaptability.
Because the load curve of the user has the characteristics of multilaterality and time delay, certain deviation exists in clustering the load users based on the Euclidean distance, and judgment cannot be carried out. Dynamic Time Warping (DTW, Dynamic Time Warping) is a method for measuring the similarity between two Time sequences with different lengths. The method has wide application, and is mainly used in template matching, such as isolated word speech recognition (whether two sections of speech represent the same word or not), gesture recognition, data mining, information retrieval and the like. Therefore, the invention carries out K-medoid-based clustering by calculating the DTW similarity of two user loads.
Specifically, the load prediction method based on dynamic time warping and long-and-short time memory comprises the following steps:
and S1, acquiring basic data required by user short-term load prediction from the power system.
The basic data comprises historical load data and meteorological data, the historical load data is load data with an interval of 1h every day in historical days, and the meteorological data at least comprises the environmental temperature at the prediction moment and the date type of the prediction day.
And S2, clustering the users with similar electricity utilization behaviors by using a dynamic time warping method according to the historical load data of the users.
The step S2 may be detailed as a clustering method based on DTW similarity according to the historical load data of the users, that is, the DTW similarity of each user load is calculated first, and then K-medoid-based clustering is performed on each user according to the DTW distance, specifically including the following steps:
s21, since the load data collected every day are all equal in length, it is assumed that the load sequences of two users are X, Y, where X is { X ═ X { (X) 1 ,x 2 ,...x N }、Y＝{y 1 ,y 2 ,...y N Define a regular path matrix M [24 x 24 ]]The matrix elements (m, n) represent x m And y n The distance between two points, here we mean the Euclidean distance, the value of the matrix element (m, n) is d (x) m ,y n )＝(x m -y n ) 2 。
S22, defining dynamic canonical path sequence P ═ (P) 1 ,p 2 ,...p k ) Wherein p is k ＝(m k ,n k ) And max (m, n) is more than or equal to K and less than m + n-1.
The dynamic warping path sequence satisfies the following conditions:
(1) boundary condition, p 1 Is equal to (1,1), andp K ＝(N,N)。
(2) monotonicity Condition, m 1 ≤m 2 ≤...≤m K And n is 1 ≤n 2 ≤...≤n K 。
(3) Step size limiting condition, if p k-1 (a ', b') the next point p of the path k The requirement for (a, b) is that (a-a ') is 1 or less and (b-b') is 1 or less. That is, a certain point can only be aligned with its adjacent point, and the alignment step cannot exceed 1.
S23, total cost of dynamic regulation between load sequences X and Y
Next, an optimal path P with the minimum cost needs to be found from the possible paths * . Thus we define the dynamic time warping distance of X and Y
Wherein, P * ＝arg min c P (X, Y), however, since the load curve has the characteristic of high volatility, we take the weekly average DTW distance of two users as the clustering distance in the lower clustering algorithm to obtain
Wherein t represents the day of the week, t is 0 to 6, and represents the period from Monday to Sunday, X t And Y t The load curves of the t days of the two users are respectively shown.
S24, a series of daily load curves X and the clustering number D of the users are given, the users are clustered into D classes according to the DTW distance of the load curves, namely the target is to find the class D and the attribution function C d And a clustering center λ d So as to make cluster interior and (S) c ) At minimum, the calculation formula is
Then, randomly selecting a clustering core, and performing iterative updating by adopting a K-medoid-based clustering algorithm.
The DTW similarity-based K-medoid clustering method comprises the following specific algorithm steps:
1. d data are randomly selected as particles (reference points) Oj (j is 1,2, …, D) in the user daily load curve sample X.
2. Repeatedly assigning the remaining sample points to the D cluster classes;
3. randomly selecting a non-particle sample Orandom; calculating the reference points of the exchange objects Orandom and O1, repeating the operation in the step 2, generating a new group of cluster classes, and calculating the target function S c If S is c If the value is less than 0, Orandom and O1 are exchanged, a new cluster class is reserved, and otherwise, the original central point and the cluster are reserved. This step is repeated until the D center points are no longer changed.
Because the clustered users in the same category have similar power utilization behaviors, the unified load prediction model is established by pooling the user data in the same category, so that the generalization capability of the model is enhanced, and the accuracy of the prediction model is improved.
And S3, performing pooling processing on the user data of the same type. In order to prevent overfitting of a load prediction algorithm based on deep learning and improve the generalization capability of a prediction model, pooling of users among classes completed by clustering in S2 is needed in S3, and overfitting is inhibited by increasing the scale of a training set, wherein S3 specifically comprises the following steps:
s31, adding I D tags to the user in the form of virtual variables.
And S32, dividing the user data into a training set and a testing set.
And S33, combining all training data to construct a training pool, and then constructing a test pool through the same process.
And S4, selecting load data, meteorological data and forecast date type data of the forecast previous day, week and month, and preprocessing the data to be used as input.
S4 specifically includes the following steps:
s41, carrying out normalization processing on the resident user load data set L, wherein the normalization formula is
Wherein,
normalized data value for a variable, x (i) raw data for the variable, x max 、x min The maximum and minimum values of the raw data, respectively.
And S42, performing sparse processing on the meteorological factor data set W and the date type I of the user by using one-hot codes to obtain an input data set of the user, wherein the input data set is L, W and I.
S43, because the load of the user presents obvious periodic characteristics, based on the consideration, historical data of the predicted day before, week before and month before are pre-trained, then the historical data are used as the input of a prediction model, the data of three days are processed in a single-layer full-connection network mode, and the final output is the same as the dimension of the input data set of a single day.
S5, constructing a short-term load prediction method based on the deep long short-term memory cyclic neural network, and verifying the effectiveness of the method by predicting the scene 24h in advance.
The structure of the deep long-short term memory circulation neural network comprises: an input layer, a hidden layer, and an output layer. The input layer is composed of neurons representing inputs. The hidden layer is composed of neurons representing intermediate variables, and the hidden layer and the next hidden layer are directly connected by the neurons. The output layer is composed of neurons representing output results.
The method for constructing the short-term load prediction based on the deep long-term and short-term memory cyclic neural network in the S5 specifically comprises the following steps:
and S51, dividing the user training data in the S4 into batches for calculation.
And S52, initializing various parameters of the long-term and short-term memory cyclic neural training network.
And S53, inputting training data to perform forward propagation training.
And S54, determining a loss function and an optimization method according to the predicted value and the true value to perform back propagation optimization.
In S5, for the clustered users, the data are divided into training sets and test sets by a random pooling method, and prediction model training based on LSTM is performed on the training sets and the test sets, respectively. The average absolute percentage error (MAPE) and the mean square error (RMSE) are adopted as model prediction effect evaluation indexes, the calculation formulas are respectively,
wherein n is the number of predicted points, y i For the load true value of the ith predicted point,
is the predicted value of the ith predicted point.
Specifically, the load prediction structure model based on the LSTM method is shown in fig. 2, and the input layer includes load history data, meteorological factors, and date type.
A Recurrent Neural Network (RNN) is a Neural network with a feedback structure whose inputs are related not only to the current inputs and the weights of the network, but also to previous network inputs, and therefore, in theory, RNNs are well suited for processing sequence data. However, when the RNN learning is long-term dependent, the gradient disappears or the gradient explodes, which results in the model being untrained. To overcome this problem, Hochreiter et al proposed a long-term memory-cycling neural network (LSTM), which introduced the updating of cell states and three gate structure control information on cell states, enabling the long-term flow of information on the network.
Internal structure of LSTM neural network As shown in FIG. 3, LSTM defines and maintains an internal memory cell state, cell state C, throughout the cycle in order to establish a temporal connection t Then through the forgetting door f t And input gate i t And an output gate o t Three gate structure for updating and maintainingProtect or delete information within the cellular state. The forward calculation process is as follows:
f t ＝σ(W f g[h t-1 ,x t ]+b f )，
i t ＝σ(W f g[h t-1 ,x t ]+b i )，
o t ＝σ(W o g[h t-1 ,x t ]+b o )，
h t ＝o t gtanh(C t )，
wherein: c t 、C t-1 Respectively representing the cell states at the current time and the last time,
representing candidate states of input, f t 、i t 、o t Respectively showing a forgetting gate, an input gate, an output gate, W f 、W i 、W C 、W o 、b f 、b i 、b C 、b o Sigmoid and hyperbolic tangent activation functions are represented, respectively. First, hide the layer output h at the last moment t-1 And the current output x t Calculating coefficients of the forgetting gate, the input gate and the output gate through the formula; then the hidden layer output h at the last moment t-1 And the current output x t Obtaining the candidate state of the current neuron by a formula
Then the last cell state C is determined by the forgetting gate and the input gate t-1 And candidate state at the current time
The ratio of the hidden layer to the current cell state is calculated by a formula t 。
And in the LSTM network training process, an error term of the output value and the true value of each LSTM neuron is reversely calculated by adopting a time back propagation algorithm, the gradient of each weight is calculated according to the corresponding error term, and the weight is updated by applying a gradient optimization algorithm.
The load prediction model based on the LSTM comprises an input layer, a hidden layer, an output layer, a network training module and an optimization module. The LSTM constructed in the present invention contains 3 hidden layers, 1 output layer. The LSTM network predicts the power load value at the ith time point, and needs to input the power load value at the previous L points as input, which is called as the sequence length. Because the power load data has the periodic characteristic, the invention aims to take the historical load data of the load prediction day, week and month before the load prediction day, week and month after being weighted by a full-connection network as the training data set T of the LSTM prediction network.
This stage of the method includes model training and model testing based on pooled load prediction:
in the training part, the deep recurrent neural network is trained by load data batches randomly taken from a user load pool, so that the LSTM network learns not only the load characteristics of a single user but also the common characteristics and uncertainty of loads in the same category.
In the test section, the test load curve is fed into the trained LSTM network. Assume that the load curve data set after data washing is Ψ 1 and the test households are listed in the set D ═ D1, D2. Next, the network configuration parameters of LSTM need to be determined, and we use L and H to represent the network depth (layer number) and the number of hidden units, respectively. With these parameters, the training and testing process can proceed as follows:
1) initializing an LSTM prediction network, and establishing network configuration parameters, namely network depth L, hidden layer number H and batch size C.
2) Network training iteration after the network is started, the program then runs a training iteration period until the network is well trained and the loss function of the neural network is the mean square error (RMSE) loss.
In each training period, training data is randomly selected from a training data pool and then input into a feedforward neural network for network training. Each training batch is two matrices of fixed size, namely an input matrix of size C × I and an output matrix of size C × O. The time cost and the number of iterations of the training process depend on the size J of the fed-back data sequence, the selected optimization method, the size of the network (M, I) and the size of the scale of one training (batch) C. To achieve a good balance between training efficiency and efficacy, the training batch size C is variable during the training process:
1) in the early period, in order to quickly approach the optimum point, C is set to a small value.
2) Then C is gradually increased towards better training performance but at the expense of time cost.
Test iteration and performance benchmark
Individual families were then tested for a trained deep recurrent neural network by predicting the neural network as a feed forward. In the testing process, load prediction is carried out on testing families one by one so as to determine whether the proposed method can singly realize the performance improvement of the load prediction. At each iteration, performance comparisons were made with other load prediction methods, including ARIMA, SVR, RNN and deep RNN, which were trained using only load data from the test home.
The following provides additional description of the above embodiments and verification of the advantages of the above embodiments by way of specific embodiments.
The invention adopts 23254 real-time load values of 1057 users provided by a certain urban power grid company as research objects. The sampling frequency of the load was collected every 1 hour for a period of 2016, 7 months and 19 days, to 2016, 8 months and 9 days, during which a 22-day load curve was recorded for each user.
According to the invention, firstly, DTW-based clustering is carried out on all users, and the users with similar behaviors are classified into one class. A good clustering algorithm should be such that the total distance (Sc) of the samples after clustering to the same class is as small as possible and the distance (WB) between the centers of the different classes is as large as possible. Therefore, we use the ratio γ of the two transverse criteria as a reference to evaluate the performance of the algorithm. Smaller γ indicates better clustering performance.
Wherein λ is i Representing the center of the ith category. Fig. 4 shows the performance comparison between the DTW clustering algorithm and the K-means clustering and the EM algorithm, and it is obvious that the performance of the DTW clustering algorithm is always better than the other two regardless of the change of the number K of the clusters.
The effectiveness of the method provided by the invention is verified by predicting the actual load value of a certain power grid. And respectively establishing LSTM prediction models for the clustered data, and predicting 168 load values from 8 months, 2 days and 1 to 8 months, 9 days and 24 hours in advance.
The invention adopts an LSTM network structure with three hidden layers, the training batch size is 96, the number of neurons in the hidden layers is 24, the optimization algorithm is Adam optimization, the learning rate is 0.002, and the loss function is RMSE. The invention respectively establishes four load prediction models of ARIMA, SVR, LSTM and DTW-LSTM, and compares the prediction performances. FIG. 5 is a diagram of the prediction effect of 8 months and 8 days, and it can be known that the DTW-LSTM short-term load prediction model can better approach the true value and has better prediction accuracy.
Table I compares the proposed DTW-LSTM prediction model against the other three classical prediction algorithms in terms of MARE and RSE. All indices shown in the table are taken as average values for all test households. The LSTM algorithm has better performance than other conventional algorithms. And after the DTW clustering is introduced, the prediction precision is further improved. Specifically, the proposed DTW-LSTM increases MARE and RSE by 6.45% and 6.96%, respectively, compared to the conventional LSTM. Compared with ARIMA, MARE and RSE of DTW-LSTM are respectively improved by 19.46% and 16.28%, and the performance is improved more remarkably.
TABLE 1 comparison of load prediction Performance
According to the load prediction method based on dynamic time warping and long-and-short time memory, aiming at the characteristic that the base number of the user to be predicted is large, the users with similar electricity utilization behaviors are clustered based on the DTW similarity, the prediction efficiency is improved, the clustering accuracy of the users is increased, the number of clustering categories is reduced, and the complexity of a clustering algorithm is reduced.
Meanwhile, aiming at the problem that the LSTM load prediction method based on deep learning is easy to over-fit, the diversity of training data is increased and the generalization capability of the model is improved by performing pooling processing on data in the same category, so that the short-term load prediction precision is improved, and the method has certain engineering application significance.
In addition, the invention also provides reference for other related problems in the same field, can be expanded and extended on the basis of the reference, is applied to the technical scheme of other load prediction methods in the same field, and has very wide application prospect.
It will be evident to those skilled in the art that the invention is not limited to the details of the foregoing illustrative embodiments, and that the present invention may be embodied in other specific forms without departing from the spirit or essential attributes thereof. The present embodiments are therefore to be considered in all respects as illustrative and not restrictive, the scope of the invention being indicated by the appended claims rather than by the foregoing description, and all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein, and any reference signs in the claims are not intended to be construed as limiting the claim concerned.
Furthermore, it should be understood that although the present description refers to embodiments, not every embodiment may contain only a single embodiment, and such description is for clarity only, and those skilled in the art should integrate the description, and the embodiments may be combined as appropriate to form other embodiments understood by those skilled in the art.
Claims (7)
1. A load prediction method based on dynamic time warping and long-and-short time memory is characterized by comprising the following steps:
s1, acquiring basic data required by user short-term load prediction from the power system;
the basic data comprises historical load data and meteorological data, the historical load data is load data with an interval of 1h every day in historical days, and the meteorological data at least comprises an environmental temperature at a prediction moment and a prediction day and day type;
s2, clustering the users with similar electricity consumption behaviors by using a dynamic time warping method according to the historical load data of the users;
s3, pooling the user data of the same category, specifically comprising the following steps:
s31, adding an ID tag to the user in a virtual variable mode;
s32, dividing user data into a training set and a test set;
s33, combining all training data to construct a training pool, and then constructing a test pool through the same process;
s4, selecting load data, meteorological data and forecast date type data of the forecast day before, week before and month before, and preprocessing the data to be used as input;
s5, constructing a short-term load prediction method based on the deep long short-term memory cyclic neural network, and verifying the effectiveness of the method by predicting the scene 24h in advance.
2. The load prediction method based on dynamic time warping and long-and-short time memory as claimed in claim 1, wherein S2 specifically comprises the following steps:
s21, assuming that the load sequences of the two users are X, Y, where X ═ X 1 ,x 2 ,...x N }、Y＝{y 1 ,y 2 ,...y N Define a regular path matrix M [24 x 24 ]]The matrix elements (m, n) represent x m And y n Distance between two points, the value of matrix element (m, n) is d (x) m ,y n )＝(x m -y n ) 2 ；
S22, defining dynamic canonical path sequence P ═ (P) 1 ,p 2 ,...p k ) Wherein p is k ＝(m k ,n k ) And max (m, n) is more than or equal to K and less than m + n-1;
s23 total cost of dynamic warping between load sequences X and Y
Defining dynamic time warping distances for X and Y
Wherein, P * ＝arg min c P (X, Y), the weekly average DTW distance of the two users is used as the clustering distance in the lower clustering algorithm to obtain
Wherein t represents the day of the week, t is 0 to 6, and represents the period from Monday to Sunday, X t And Y t Respectively representing the load curves of the t days of the two users;
s24, a series of daily load curves X and the clustering number D of the users are given, the users are clustered into D classes according to the DTW distance of the load curves, and the class D and the attribution function C are obtained d And a clustering center λ d So as to make cluster interior and (S) c ) At minimum, the calculation formula is
Then, randomly selecting a clustering core, and performing iterative updating by adopting a K-medoid-based clustering algorithm.
3. The load prediction method based on dynamic time warping and long-and-short time memory as claimed in claim 2, wherein said dynamic warping path sequence in S22 satisfies the following condition:
boundary condition, p 1 Is ═ 1,1), and p K ＝(N,N)；
Monotonicity Condition, m 1 ≤m 2 ≤...≤m K And n is 1 ≤n 2 ≤...≤n K ；
Step size limiting condition, if p k-1 (a ', b') the next point p of the path k The requirement for (a, b) is that (a-a ') is 1 or less and (b-b') is 1 or less.
4. The load prediction method based on dynamic time warping and long-and-short time memory as claimed in claim 1, wherein S4 specifically comprises the following steps:
s41, carrying out normalization processing on the resident user load data set L, wherein the normalization formula is
Wherein,
normalized data value for a variable, x (i) raw data for the variable, x max 、x min Respectively the maximum value and the minimum value of the original data;
s42, conducting sparse processing on the meteorological factor data set W and the date type I of the user by using one-hot codes, and obtaining an input data set of the user, wherein the input data set is L, W and I;
and S43, pre-training the load data, meteorological data and predicted date and type data of the previous day, week and month, and then serving as the input of a prediction model, specifically processing the data of three days in a single-layer full-connection network mode, wherein the final output is the same as the dimension of the input data set of a single day.
5. The method for load prediction based on dynamic time warping and long-and-short term memory as claimed in claim 1, wherein the step of constructing a short-term load prediction method based on a deep long-and-short term memory recurrent neural network in S5, the structure of the deep long-and-short term memory recurrent neural network comprises: an input layer, a hidden layer and an output layer; the input layer is composed of neurons representing inputs; the hidden layer consists of neurons representing intermediate variables, and the hidden layer and the next hidden layer are directly connected by the neurons; the output layer is composed of neurons representing output results.
6. The load prediction method based on dynamic time warping and long-and-short-term memory as claimed in claim 1, wherein the step of constructing the short-term load prediction method based on the deep long-and-short-term memory recurrent neural network in S5 includes the following steps:
s51, dividing the user training data in the S4 into batches for calculation;
s52, initializing each parameter of the long-short term memory recurrent neural training network;
s53, inputting training data to perform forward propagation training;
and S54, determining a loss function and an optimization method according to the predicted value and the true value to perform back propagation optimization.
7. The load prediction method based on dynamic time warping and long-and-short time memory as claimed in claim 1, wherein: in S5, the mean absolute percentage error and the mean square error are used as the evaluation indexes of the model prediction effect, and the calculation formulas are respectively,
wherein n is preNumber of measurement points, y i For the true load value of the ith predicted point,
is the predicted value of the ith predicted point.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811439799.9A CN109754113B (en) | 2018-11-29 | 2018-11-29 | Load prediction method based on dynamic time warping and long-and-short time memory |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811439799.9A CN109754113B (en) | 2018-11-29 | 2018-11-29 | Load prediction method based on dynamic time warping and long-and-short time memory |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109754113A CN109754113A (en) | 2019-05-14 |
| CN109754113B true CN109754113B (en) | 2022-08-30 |
Family
ID=66403408
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811439799.9A Active CN109754113B (en) | 2018-11-29 | 2018-11-29 | Load prediction method based on dynamic time warping and long-and-short time memory |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109754113B (en) |
Families Citing this family (43)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110210993B (en) * | 2019-05-22 | 2023-04-07 | 重庆大学 | Urban short-term gas load prediction method based on cyclic neural network model |
| CN110210677B (en) * | 2019-06-06 | 2022-03-04 | 国网山东省电力公司莱芜供电公司 | Bus short-term daily load prediction method and device combining clustering and deep learning algorithm |
| CN110543942A (en) * | 2019-08-28 | 2019-12-06 | 广西大学 | Multi-space-time long and short memory depth network accurate prediction method |
| CN110569925B (en) * | 2019-09-18 | 2023-05-26 | 南京领智数据科技有限公司 | LSTM-based time sequence abnormality detection method applied to power equipment operation detection |
| CN110674993A (en) * | 2019-09-26 | 2020-01-10 | 广东电网有限责任公司 | User load short-term prediction method and device |
| CN112686483B (en) * | 2019-10-17 | 2024-08-13 | 中国移动通信集团陕西有限公司 | Early warning area identification method, early warning area identification device, computing equipment and computer storage medium |
| CN110956306A (en) * | 2019-10-23 | 2020-04-03 | 广东电网有限责任公司 | Load prediction method based on load clustering |
| CN111082419B (en) * | 2019-12-06 | 2022-06-24 | 国网河北省电力有限公司电力科学研究院 | Intelligent power distribution network user access scheme management method and system based on big data technology |
| CN111222688B (en) * | 2019-12-06 | 2024-01-12 | 北京国电通网络技术有限公司 | Daily load prediction method for commercial building |
| CN111027776A (en) * | 2019-12-13 | 2020-04-17 | 北京华展汇元信息技术有限公司 | Sewage treatment water quality prediction method based on improved long-short term memory LSTM neural network |
| CN111369078A (en) * | 2020-02-14 | 2020-07-03 | 迈拓仪表股份有限公司 | Water supply quality prediction method based on long-term and short-term memory neural network |
| CN111210091A (en) * | 2020-02-25 | 2020-05-29 | 上海积成能源科技有限公司 | System model for predicting short-term power load based on long and short-term memory model of recurrent neural network |
| CN111461400B (en) * | 2020-02-28 | 2023-06-23 | 国网浙江省电力有限公司 | Kmeans and T-LSTM-based load data completion method |
| CN111445078A (en) * | 2020-03-31 | 2020-07-24 | 国网河北省电力有限公司 | Comprehensive energy system multi-element load prediction method based on long-term and short-term memory neural network |
| CN111476298B (en) * | 2020-04-07 | 2022-04-26 | 杭州国彪超声设备有限公司 | Power load state identification method in home and office environment |
| CN111582548A (en) * | 2020-04-14 | 2020-08-25 | 广东卓维网络有限公司 | Power load prediction method based on multivariate user behavior portrait |
| CN111583064B (en) * | 2020-05-11 | 2022-09-09 | 国网四川省电力公司电力科学研究院 | Load production time interval detection method based on dynamic time warping and storage medium |
| CN111723873A (en) * | 2020-06-29 | 2020-09-29 | 南方电网科学研究院有限责任公司 | Power sequence data classification method and device |
| CN111783684A (en) * | 2020-07-03 | 2020-10-16 | 湖南大学 | Bi-GRU-SA-based household electrical load decomposition method and system |
| CN111913575B (en) * | 2020-07-24 | 2021-06-11 | 合肥工业大学 | Method for recognizing hand-language words |
| CN112348070A (en) * | 2020-10-30 | 2021-02-09 | 广东电网有限责任公司电力调度控制中心 | Method and system for forecasting medium and short term loads of smart power grid |
| CN112118143B (en) * | 2020-11-18 | 2021-02-19 | 迈普通信技术股份有限公司 | Traffic prediction model training method, traffic prediction method, device, equipment and medium |
| CN112508259A (en) * | 2020-12-02 | 2021-03-16 | 重庆大学 | Medium-and-long-term power consumption prediction method and system for independent user categories |
| CN112418342A (en) * | 2020-12-07 | 2021-02-26 | 中山大学 | Mobile equipment calculation force prediction method and device under federated learning scene |
| CN112365098A (en) * | 2020-12-07 | 2021-02-12 | 国网冀北电力有限公司承德供电公司 | Power load prediction method, device, equipment and storage medium |
| CN112580260B (en) * | 2020-12-22 | 2024-09-17 | 广州杰赛科技股份有限公司 | Pipe network water flow prediction method and device and computer readable storage medium |
| WO2022133832A1 (en) * | 2020-12-23 | 2022-06-30 | Huawei Technologies Co., Ltd. | Apparatus and method for predictive computer modelling |
| CN112613666B (en) * | 2020-12-26 | 2023-02-14 | 福建维力能源科技有限公司 | Power grid load prediction method based on graph convolution neural network and transfer learning |
| CN112929445B (en) * | 2021-02-20 | 2022-06-07 | 山东英信计算机技术有限公司 | Recommendation system-oriented link prediction method, system and medium |
| CN113033898A (en) * | 2021-03-26 | 2021-06-25 | 国核电力规划设计研究院有限公司 | Electrical load prediction method and system based on K-means clustering and BI-LSTM neural network |
| CN113125037B (en) * | 2021-04-06 | 2024-06-07 | 红塔烟草(集团)有限责任公司 | Cable conductor temperature estimation method based on distributed optical fiber on-line temperature measurement system |
| CN113255900A (en) * | 2021-06-23 | 2021-08-13 | 河北工业大学 | Impulse load prediction method considering improved spectral clustering and Bi-LSTM neural network |
| CN113448808B (en) * | 2021-08-30 | 2021-11-12 | 北京必示科技有限公司 | Method, system and storage medium for predicting single task time in batch processing task |
| CN113821574B (en) * | 2021-08-31 | 2024-07-30 | 北京达佳互联信息技术有限公司 | User behavior classification method and device and storage medium |
| CN113987910A (en) * | 2021-09-18 | 2022-01-28 | 国网江苏省电力有限公司信息通信分公司 | Method and device for identifying load of residents by coupling neural network and dynamic time planning |
| CN113822366A (en) * | 2021-09-29 | 2021-12-21 | 平安医疗健康管理股份有限公司 | Service index abnormality detection method and device, electronic equipment and storage medium |
| CN113988436B (en) * | 2021-11-01 | 2023-04-28 | 广西电网有限责任公司 | Power consumption prediction method based on LSTM neural network and hierarchical relation correction |
| CN114781759A (en) * | 2022-06-16 | 2022-07-22 | 国网天津市电力公司经济技术研究院 | Resident load prediction method and device based on neural network and dynamic mirror image reduction |
| CN115481788B (en) * | 2022-08-31 | 2023-08-25 | 北京建筑大学 | Phase change energy storage system load prediction method and system |
| CN115423301B (en) * | 2022-09-01 | 2023-04-25 | 杭州达中科技有限公司 | Intelligent electric power energy management and control method, device and system based on Internet of things |
| CN115454009A (en) * | 2022-11-11 | 2022-12-09 | 常熟理工学院 | Component distribution model predictive control method for chemical production |
| CN116029201B (en) * | 2022-12-23 | 2023-10-27 | 浙江苍南仪表集团股份有限公司 | Gas flow prediction method and system based on clustering and cyclic neural network |
| CN117232097B (en) * | 2023-11-09 | 2024-02-20 | 上海轻环能源科技有限公司 | Central air conditioner refrigerating station optimal control method and system based on self-learning fusion model |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105608512A (en) * | 2016-03-24 | 2016-05-25 | 东南大学 | Short-term load forecasting method |
| CN108416690A (en) * | 2018-01-19 | 2018-08-17 | 中国矿业大学 | Load Forecasting based on depth LSTM neural networks |
-
2018
- 2018-11-29 CN CN201811439799.9A patent/CN109754113B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105608512A (en) * | 2016-03-24 | 2016-05-25 | 东南大学 | Short-term load forecasting method |
| CN108416690A (en) * | 2018-01-19 | 2018-08-17 | 中国矿业大学 | Load Forecasting based on depth LSTM neural networks |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109754113A (en) | 2019-05-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109754113B (en) | Load prediction method based on dynamic time warping and long-and-short time memory | |
| CN109063911B (en) | Load aggregation grouping prediction method based on gated cycle unit network | |
| Han et al. | Short-term forecasting of individual residential load based on deep learning and K-means clustering | |
| CN109492748B (en) | Method for establishing medium-and-long-term load prediction model of power system based on convolutional neural network | |
| CN111861013A (en) | Power load prediction method and device | |
| CN116721537A (en) | Urban short-time traffic flow prediction method based on GCN-IPSO-LSTM combination model | |
| Tian et al. | Daily power demand prediction for buildings at a large scale using a hybrid of physics-based model and generative adversarial network | |
| CN113344288B (en) | Cascade hydropower station group water level prediction method and device and computer readable storage medium | |
| CN111915092A (en) | Ultra-short-term wind power prediction method based on long-time and short-time memory neural network | |
| CN113298318A (en) | Novel overload prediction method for distribution transformer | |
| Akpinar et al. | Forecasting natural gas consumption with hybrid neural networks—Artificial bee colony | |
| CN111428766A (en) | Power consumption mode classification method for high-dimensional mass measurement data | |
| CN115759458A (en) | Load prediction method based on comprehensive energy data processing and multi-task deep learning | |
| CN115481788B (en) | Phase change energy storage system load prediction method and system | |
| CN113962454A (en) | LSTM energy consumption prediction method based on dual feature selection and particle swarm optimization | |
| Liu et al. | Multi-stage residual life prediction of aero-engine based on real-time clustering and combined prediction model | |
| Wei et al. | An instance based multi-source transfer learning strategy for building’s short-term electricity loads prediction under sparse data scenarios | |
| CN113762591B (en) | Short-term electric quantity prediction method and system based on GRU and multi-core SVM countermeasure learning | |
| Sun et al. | Load-forecasting method for IES based on LSTM and dynamic similar days with multi-features | |
| CN113537539A (en) | Multi-time-step heat and gas consumption prediction model based on attention mechanism | |
| CN115310355A (en) | Multi-energy coupling-considered multi-load prediction method and system for comprehensive energy system | |
| CN117390550A (en) | Low-carbon park carbon emission dynamic prediction method and system considering emission training set | |
| CN110852628A (en) | Rural medium and long term load prediction method considering development mode influence | |
| CN115796327A (en) | Wind power interval prediction method based on VMD (vertical vector decomposition) and IWOA-F-GRU (empirical mode decomposition) -based models | |
| Wang et al. | Prediction of air pollution based on FCM-HMM Multi-model |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |