CN111160620B - Short-term wind power prediction method based on end-to-end memory network - Google Patents

Short-term wind power prediction method based on end-to-end memory network Download PDF

Info

Publication number
CN111160620B
CN111160620B CN201911247976.8A CN201911247976A CN111160620B CN 111160620 B CN111160620 B CN 111160620B CN 201911247976 A CN201911247976 A CN 201911247976A CN 111160620 B CN111160620 B CN 111160620B
Authority
CN
China
Prior art keywords
data
vector
node
memory network
historical data
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
Application number
CN201911247976.8A
Other languages
Chinese (zh)
Other versions
CN111160620A (en
Inventor
祝永晋
查满霞
严佳欣
谢林枫
马吉科
龙玲莉
李同哲
司加胜
周德宇
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Southeast University
Jiangsu Fangtian Power Technology Co Ltd
Original Assignee
Southeast University
Jiangsu Fangtian Power Technology Co Ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Southeast University, Jiangsu Fangtian Power Technology Co Ltd filed Critical Southeast University
Priority to CN201911247976.8A priority Critical patent/CN111160620B/en
Publication of CN111160620A publication Critical patent/CN111160620A/en
Application granted granted Critical
Publication of CN111160620B publication Critical patent/CN111160620B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

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
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • 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—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/06—Energy or water supply

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • Economics (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Strategic Management (AREA)
  • Human Resources & Organizations (AREA)
  • Software Systems (AREA)
  • Tourism & Hospitality (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Mathematical Physics (AREA)
  • Data Mining & Analysis (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • General Business, Economics & Management (AREA)
  • Molecular Biology (AREA)
  • Artificial Intelligence (AREA)
  • Marketing (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Primary Health Care (AREA)
  • Water Supply & Treatment (AREA)
  • Public Health (AREA)
  • Development Economics (AREA)
  • Game Theory and Decision Science (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a short-term wind power prediction method based on an end-to-end memory network, which comprises the following steps: collecting data of a wind power plant; dividing the collected wind power plant data into historical data, predicted multi-node weather data and power data, and carrying out normalization processing on the historical data and the predicted multi-node weather data; encoding historical data by using a multi-head self-attention mechanism model, and storing the encoded feature vectors into a memory pool of an end-to-end memory network; encoding the predicted multi-node weather data by adopting an attention mechanism, and taking an encoding result as an input vector of an end-to-end memory network; taking the power data as an output vector of the end-to-end memory network to train the end-to-end memory network; and predicting the output power of the fan by using the trained end-to-end memory network. Compared with the conventional short-term wind power prediction method, the method disclosed by the invention can be used for paying attention to the information implicit in the historical data, and has higher prediction accuracy and stability.

Description

Short-term wind power prediction method based on end-to-end memory network
Technical Field
The invention relates to the field of wind power prediction, in particular to a short-term wind power prediction method.
Background
The short-term wind power prediction means that the power output by a wind power place within 1-4 hours in the future is predicted according to NWP data of wind speed, temperature, wind direction, radiation, precipitation, cloud cover and the like and historical data of a wind power plant. The method has the advantages that the generated power of the wind power plant is accurately predicted in a short period, so that the power dispatching distribution can be timely adjusted by a national power grid, the power quality and the economic benefit are improved, and the safe and stable operation of a power system is guaranteed.
The wind power prediction method mainly comprises three types: physical methods, statistical methods, and artificial intelligence methods. The physical method is characterized in that a physical model is built according to the existing knowledge to describe the conversion process of wind energy-kinetic energy, and the output power of the wind turbine is calculated by combining a wind turbine power curve, but the physical method has the problem of low prediction accuracy because model parameters are difficult to accurately select. The statistical method is characterized in that a non-linear mapping relation between data samples is established by mining a statistical rule among historical data of the wind power plant, and the non-linear mapping relation comprises an autoregressive integral moving average method, a persistence method, a Kalman filtering method and the like.
The artificial intelligence method combines models such as a support vector machine and an artificial neural network, and defines a wind power prediction task as a time series modeling problem by carrying out nonlinear analysis on a large amount of historical data. In the existing wind power short-term prediction methods, some prediction methods based on BP neural networks initialize weight thresholds of the BP neural networks through an empire competition algorithm and a particle swarm optimization algorithm, so that better effects than those of general BP neural networks are obtained, but the method still does not consider the relevance of wind power on a time sequence; in other methods based on the LSTM network, a principal component analysis method is adopted to screen out factors which have large influence on power from weather data, and then the LSTM is used for modeling a weather data time sequence.
Disclosure of Invention
The purpose of the invention is as follows: in order to overcome the defects in the prior art, the invention provides a method for short-term prediction of wind power, which is based on numerical weather forecast data and data acquired by an SCADA (supervisory control and data acquisition) system, and comprises the steps of firstly obtaining historical data feature codes by using a multi-head self-Attention mechanism (Transformer), then obtaining feature codes of predicted multi-node weather data by using an Attention mechanism, and finally predicting the power by combining the obtained feature codes through an end-to-end memory network (MemN2N) model.
The technical scheme is as follows: in order to achieve the purpose, the invention adopts the technical scheme that:
a short-term wind power prediction method based on an end-to-end memory network is characterized by comprising the following steps:
collecting data of a wind power plant;
dividing the collected wind power plant data into historical data, predicted multi-node weather data and power data, and carrying out normalization processing on the historical data and the predicted multi-node weather data;
encoding historical data by using a Transformer model, and storing the encoded characteristic vectors into a memory pool of MemN 2N; encoding the predicted multi-node weather data by adopting an attention mechanism, and taking an encoding result as an input vector of MemN 2N; performing MemN2N training with the power data as an output vector of MemN 2N;
and (4) predicting the output power of the fan by using the trained MemN2N network.
The method for acquiring the wind power plant data comprises the following steps of: judging the data of each node, and if the proportion of empty data in the node is more than or equal to 35%, deleting the node; otherwise, zero element filling processing is carried out on the obtained null data in the node.
The historical data is constructed by splicing numerical weather forecast data and SCADA system collected data, the numerical weather forecast data comprises wind speed and wind direction, air temperature, air pressure, radiation, precipitation and cloud cover amount at different altitudes, and the SCADA system collected data comprises a fan number and wind speed, wind direction, temperature and power measured by a fan sensor.
The method for dividing and normalizing the data specifically comprises the following steps: for each time node, taking numerical weather forecast data of N nodes in front of the time node and SCADA system collected data as historical data, namely X ═ { X ═ X1，x2，...，xN}; taking the numerical weather forecast data of the M nodes behind the numerical weather forecast data as predicted multi-node weather data, namely B ═ B1，b2，...，bM}; taking the actual power data of the M nodes behind the model as power data, namely the output of the model; the historical data and predicted multinode weather data are then normalized to [0,1 ]]In the interval, the output of the model is not normalized.
A method of encoding historical data using a Transformer model, comprising:
(1) a transform-based multi-head mechanism divides an encoder into T heads, and randomly initializes three matrixes in each head
Figure BDA0002305923500000021
Where X ∈ X represents any one of the historical data nodes, | X | represents the number of data in X,
Figure BDA0002305923500000022
representing a subspace coding length;
(2) for each history data xnBelongs to X, N is more than or equal to 1 and less than or equal to N, respectively passes through with WQ、WK、WVMultiplying to obtain a problem vector QnKey vector KnValue vector Vn；
(3) For current historical data node xiUsing problem vector QiKey vector K with all history data, respectively1，K2，...，KNMultiplying by a point to calculate each historical data xnAt xiScore S ═ S in (1)1，s2，...，sN}；
(4) For a score snOperating with epsilon S to obtain
Figure BDA0002305923500000031
Figure BDA0002305923500000032
Then all the
Figure BDA0002305923500000033
Respectively corresponding value vectors VnMultiply and sum to obtain xiSingle-headed precoding vector e of1；
(5) For xiPerforming (2) - (4) in all heads to obtain xiPrecoding vectors in all headers e1，e2，...，eTWill { e }1，e2，...，eTAll the items in the Chinese character are spliced to obtain xiPre-coding of (2); x is to beiPrecoding and xiAre summed to obtain
Figure BDA0002305923500000034
Then will be
Figure BDA0002305923500000035
Obtained by a feed-forward neural network
Figure BDA0002305923500000036
To pair
Figure BDA0002305923500000037
And
Figure BDA0002305923500000038
summing to obtain xiSingle layer coding result of (2)
Figure BDA0002305923500000039
(6) For all historical dataNode xnE, executing (2) - (5) for X to obtain single-layer coding results of all historical data;
(7) for the L-layer encoder, the single-layer encoding result in (6) is used as the input of the next-layer encoder, and (1) - (6) are executed, so that the final encoding H ═ H of all historical data can be obtained1，h2，...，hN}；
Wherein, T, N, dkL is a hyper-parameter, T is the number of headers in each encoder, N is the number of historical data nodes, dkFor scaling, it is generally set to the dimension of the key vector K, and L is the number of layers of the encoder.
The operation of encoding the predicted multi-node weather data by adopting the attention mechanism specifically comprises the following steps: for predicted multi-node weather data B ═ B1，b2，...，bMFor each node biE.g. B is assigned a weight αiI is more than or equal to 1 and less than or equal to M, multiplying each node by the corresponding weight and summing to obtain a result u as the input of the MemN2N model; wherein M is a hyper-parameter and is the number of the predicted nodes.
The method for predicting the power by using the MemN2N model comprises the following steps:
(1) introducing the characteristic vector H into a memory pool of a long-term memory component of the MemN2N model;
(2) random initialization of two embedded matrices
Figure BDA00023059235000000310
And
Figure BDA00023059235000000311
wherein X belongs to X and represents any historical data node, | X | represents the number of data in X, and len represents the coding length. For code vector h in memory pooliE is H, i is more than or equal to 1 and less than or equal to N, HiMultiplying respectively with A and C to obtain hiMapping m in two different feature spacesiAnd ci(ii) a The same operation is executed to all the code vectors in the memory pool to obtain two feature matrixes M1＝{m1，m2，...，mNAnd M2＝{c1，c2，...，cN}；
(3) The obtained input vector u and a first feature matrix M1Multiplying, and calculating an attention vector P ═ P by using the obtained result through softmax operation1，p2，...，pN]Wherein p isiFor the input vector u to the i-th feature vector hiThe attention of (1);
(4) for the feature matrix M2Each feature vector c iniIt is associated with the corresponding attention P in the vector PiMultiplying, summing all weighted eigenvectors to obtain a context vector o, and adding the context vector o to the input u to obtain a result r1As input to the next layer;
(5) for G layer MemN2N network, repeating the steps (2) - (4), and outputting vector r of the last layerGRespectively performing point multiplication on the M different vectors, wherein the result of each point multiplication is a predicted power; wherein G is a hyper-parameter and is the number of layers of the MemN2N network.
According to the short-term wind power prediction method based on the end-to-end memory network, the long-distance dependency information in the historical data is mined by using the Transformer, and the dependency information is introduced into the memory pool of the MemN2N network, so that compared with the LSTM network, the MemN2N network has obvious advantages in memory storage, and therefore the prediction accuracy is higher; in the multi-step prediction process, the input component and the output component of the MemN2N network are further improved based on an attention mechanism, and the correlation between continuous multi-step weather data is respectively modeled according to different step sizes, so that the stability is higher in the multi-step prediction process.
Drawings
FIG. 1 is a flow chart of a method of an embodiment of the present invention.
Detailed Description
The invention is further illustrated by the following examples in connection with specific embodiments thereof, it is to be understood that these examples are intended in an illustrative rather than in a limiting sense and that various equivalent modifications of the invention as described herein will occur to those skilled in the art upon reading the present disclosure and are intended to be covered by the appended claims.
A short-term wind power prediction method based on an end-to-end memory network is disclosed, and as shown in figure 1, the method comprises two steps of preprocessing original data and training a model.
The problem can be described as follows: with X ═ X1，x2，...，xNDenotes history data, with Q ═ Q1，q2，...，qMAnd expressing weather data of predicted multiple nodes, wherein the task of short-term wind power prediction is to predict the output power of the wind power plant at M future time points based on historical data X and weather data Q at M future time points.
Firstly, preprocessing original data
Step 1, for each node Z in the original data set ZiJudging, and if the proportion of the null data in the node is more than or equal to 35%, deleting the node; otherwise, zero element filling processing is carried out on the null data in the node.
Step 2, dividing the processed Z into historical data, predicted multi-node weather data and power data, specifically: for each node Z in ZiTaking the weather data and the actual power data of the previous N nodes as historical data, namely X ═ X1，x2，...，xNN is the number of the historical data nodes; will be followed by M nodes (including z)i) As predicted multi-node weather data, i.e., B ═ B1，b2，...，bMM is the number of predicted multi-nodes; taking the actual power data of the M nodes behind the model as power data, namely the output of the model; then, the historical data X and the predicted multi-node weather data B are processed to [0,1 ]]And normalization in the interval is carried out, and normalization processing is not carried out on the output of the model.
Secondly, training an end-to-end memory network prediction model
Step 3, in the stage of training the encoder of the Transformer, dividing each encoder into T heads, and then, in each head, dividing each encoder into T headsMachine initialization three | x | xR matrices WQ、WK、WVWhere x represents any one of the historical data nodes, | x | represents the number of data in x,
Figure BDA0002305923500000051
representing the subspace coding length. For each historical data node xnBelongs to X (1 is less than or equal to N is less than or equal to N), respectively passes through with WQ、WK、WVMultiplying to obtain a problem vector QnKey vector KnValue vector Vn。
For current historical data node xiUsing problem vector QiKey vector K with all history data, respectively1，K2，...，KNMultiplying by the points, and calculating each historical data node xnAt xiScore S ═ S in (1)1，s2，...，sN}; each score snIs epsilon S (1. ltoreq. n.ltoreq.N) divided by
Figure BDA0002305923500000052
Then performing softmax operation to obtain
Figure BDA0002305923500000053
Then all the
Figure BDA0002305923500000054
Respectively corresponding value vectors VnMultiply and sum to obtain xiSingle-headed precoding vector e of1. Wherein d iskTo scale, the dimension of the key vector K is typically set.
For x in all headsiPerforming the same operation, x can be obtainediPrecoding vectors in all headers e1，e2，...，eTWill { e }1，e2，...，eTAll the items in the Chinese character are spliced to be xiPre-coding of (2); then, for xiIs pre-coded with xiThe result of the summation is
Figure BDA0002305923500000055
Will be provided with
Figure BDA0002305923500000056
Obtained by a feed-forward neural network
Figure BDA0002305923500000057
To pair
Figure BDA0002305923500000058
Medicine for curing cancer
Figure BDA0002305923500000059
Summing to obtain xiSingle layer coding result of
Figure BDA00023059235000000510
The same operation is carried out on other historical data nodes, and a single-layer coding result of all historical data can be obtained.
For an L-layer encoder, the single-layer encoding result is used as the input of the next-layer encoder, and the same operation is performed on the encoder, so that the final encoding H ═ H of all the historical data can be obtained1，h2，...，hN}. And N is the number of the historical data nodes.
Step 4, for predicted multi-node weather data B ═ B1，b2，...，bMFor each node biE.g. B is assigned a weight αi(1 ≦ i ≦ M), multiplying and summing each node with its corresponding weight, and taking the result u as the input of the MemN2N model.
Step 5, in the parameter training stage of the MemN2N network, firstly, introducing the coding result H in the step 3 into a memory pool of a long-term memory component of the MemN2N model; then, for a single-layer MemN2N network, two | x | × len embedding matrices A and C are randomly initialized, and for the coding vector h in the memory pooliE is H (i is more than or equal to 1 and less than or equal to N), H isiMultiplying respectively with A and C to obtain hiMapping m in two different feature spacesiAnd ci(ii) a Performing the same for all code vectors in the memory poolOperate to obtain two feature matrices M1＝{m1，m2，...，mNAnd M2＝{c1，c2，...，cN}; the input vector u obtained in the step 4 and the first feature matrix M are combined1Multiplying, and calculating an attention vector P ═ P by using the obtained result through softmax operation1，p2，...，pN]Wherein p isi(1. ltoreq. p. ltoreq.N) represents the attention of the input vector u to the ith feature vector; for the feature matrix M2Each feature vector c iniIt is associated with the corresponding attention P in the vector PiMultiplying, summing all weighted eigenvectors to obtain a context vector o, and adding the context vector o to the input u to obtain a result r1As input for the next layer. Where len denotes a code length.
Finally, for the G layer MemN2N network, the operation of the single layer network is repeatedly executed, and the output vector r of the last layer isGAnd respectively carrying out point multiplication on the M different vectors, wherein the result of each point multiplication is a predicted power.
Thirdly, testing the trained model
For weather data node to be predicted
Figure BDA0002305923500000061
And corresponding historical data node
Figure BDA0002305923500000062
Will be provided with
Figure BDA0002305923500000063
Inputting into a Transformer, using an attention mechanism
Figure BDA0002305923500000064
And (4) performing calculation, inputting the calculation result into MemN2N, and outputting the model, namely the predicted power.
The short-term wind power short-term prediction algorithm flow based on the end-to-end memory network is as follows:
Figure BDA0002305923500000065
Figure BDA0002305923500000071
Figure BDA0002305923500000081
in the experimental process, the parameters are set as follows: the number of layers L of the transform encoder is 6, the number of headers T in each encoder is 4, the number of layers G of the MemN2N network is 3, and the encoding length Len is 100. The short-term wind power short-term prediction method based on the end-to-end memory network provided by the invention is characterized in that a test is carried out on an actually measured data set of a certain power company in Jiangsu, the root-mean-square errors in the steps 2, 4, 8 and 16 are respectively 401KW, 408KW, 418KW and 435KW, and the root-mean-square errors of an LSTM model on the data set are respectively 571KW, 549KW, 522KW and 596 KW. The error of the method provided by the invention fluctuates about 72% of the LSTM model, and the prediction effect is superior to that of the LSTM model.
The short-term wind power prediction method based on the end-to-end memory network not only can mine the relevance of the wind power on a time sequence, but also solves the problem of information loss in the long-time sequence analysis process, and is beneficial to timely adjustment of the national power grid for power dispatching.
The above description is only of the preferred embodiments of the present invention, and it should be noted that: it will be apparent to those skilled in the art that various modifications and adaptations can be made without departing from the principles of the invention and these are intended to be within the scope of the invention.

Claims (7)

1. A short-term wind power prediction method based on an end-to-end memory network is characterized by comprising the following steps:
collecting data of a wind power plant;
dividing the collected wind power plant data into historical data, predicted multi-node weather data and power data, and carrying out normalization processing on the historical data and the predicted multi-node weather data;
encoding historical data by using a multi-head self-attention mechanism model, and storing the encoded feature vectors into a memory pool of an end-to-end memory network; encoding the predicted multi-node weather data by adopting an attention mechanism, and taking an encoding result as an input vector of an end-to-end memory network; taking the power data as an output vector of the end-to-end memory network to train the end-to-end memory network;
and predicting the output power of the fan by using the trained end-to-end memory network.
2. The end-to-end memory network-based short-term wind power prediction method according to claim 1, wherein the acquiring of the wind farm data further comprises preprocessing the acquired wind farm data, and the preprocessing method specifically comprises: judging the data of each node, and if the proportion of empty data in the node is more than or equal to 35%, deleting the node; otherwise, the zero element filling processing is carried out on the empty data in the node.
3. The short-term wind power prediction method based on the end-to-end memory network as claimed in claim 1, characterized in that the historical data is constructed by splicing numerical weather forecast data and SCADA system collected data, wherein the numerical weather forecast data includes wind speed and direction, temperature, air pressure, radiation, precipitation and cloud cover at different altitudes, and the SCADA system collected data includes fan number, wind speed, direction, temperature and power measured by fan sensor.
4. The end-to-end memory network-based short-term wind power prediction method according to claim 1, wherein the method for dividing and normalizing data specifically comprises the following steps: for each time node, the numerical weather forecast data and the SCADA system data of the N nodes before the time node are collectedCollecting data as historical data, i.e. X ═ X1,x2,…,xN}; taking the numerical weather forecast data of the M nodes behind the numerical weather forecast data as predicted multi-node weather data, namely B ═ B1,b2,…,bM}; taking the actual power data of the M nodes behind the model as power data, namely the output of the model; the historical data and predicted multi-node weather data are then normalized to [0,1 ]]In the interval, the output of the model is not normalized.
5. The end-to-end memory network-based short-term wind power prediction method according to claim 1, wherein the method for encoding historical data by using a multi-head self-attention mechanism model comprises the following steps:
(1) a multi-head mechanism based on a multi-head self-attention mechanism divides an encoder into T heads, and randomly initializes three matrixes in each head
Figure FDA0003562009570000011
Where X ∈ X represents any one of the historical data nodes, | X | represents the number of data in X,
Figure FDA0003562009570000012
representing a subspace coding length;
(2) for each history data xnBelongs to X, N is more than or equal to 1 and less than or equal to N, respectively passes through with WQ、WK、WVMultiplying to obtain a problem vector QnKey vector KnValue vector Vn；
(3) For current historical data node xiUsing problem vector QiKey vector K with all history data, respectively1,K2,…,KNMultiplying by a point to calculate each historical data xnAt xiScore S ═ S1,s2,…,sN}；
(4) For a score snOperating with epsilon S to obtain
Figure FDA0003562009570000021
Figure FDA0003562009570000022
In the formula, e is an exponential operator; dkRepresents a zoom scale set to a key vector KnDimension (d); then all the
Figure FDA0003562009570000023
Respectively corresponding value vectors VnMultiply and sum to obtain xiSingle-headed precoding vector e of1；
(5) For xiOperations (2) - (4) are performed in all headers, resulting in xiPrecoding vectors in all headers e1,e2,…,eTWill { e }, will { e1,e2,…,eTConcatenating all the items in the } to get xiPre-coding of (2); x is to beiPrecoding and xiAre summed to obtain
Figure FDA0003562009570000024
Then will be
Figure FDA0003562009570000025
Obtained by a feed-forward neural network
Figure FDA0003562009570000026
To pair
Figure FDA0003562009570000027
And
Figure FDA0003562009570000028
summing to obtain xiSingle layer coding result of
Figure FDA0003562009570000029
(6) For theAll historical data nodes xnE, executing (2) - (5) for X to obtain single-layer coding results of all historical data;
(7) for the L-layer encoder, the single-layer encoding result in (6) is used as the input of the next-layer encoder, and (1) - (6) are executed, so that the final encoding H ═ H of all historical data can be obtained1,h2,…,hN}；
Wherein, T, N, dkL is a hyper-parameter, T is the number of heads in each encoder, N is the number of historical data nodes, dkFor scaling, the dimension of the key vector K is set, and L is the number of layers of the encoder.
6. The end-to-end memory network-based short-term wind power prediction method according to claim 5, wherein the operation of encoding the predicted multi-node weather data by using an attention mechanism specifically comprises: for predicted multi-node weather data B ═ B1,b2,…,bMFor each node b }iE.g. B is assigned a weight αiI is more than or equal to 1 and less than or equal to M, multiplying and summing each node and the corresponding weight thereof, and taking the obtained result u as the input of the end-to-end memory network model; wherein M is a hyper-parameter and is the number of the predicted nodes.
7. The end-to-end memory network-based short-term wind power prediction method according to claim 6, wherein the method for performing power prediction by using an end-to-end memory network model comprises the following steps:
(1) introducing the characteristic vector H into a memory pool of a long-term memory component of the end-to-end memory network model;
(2) random initialization of two embedded matrices
Figure FDA00035620095700000210
And
Figure FDA00035620095700000211
wherein X ∈ X represents any one historical data node, and | X | represents data in XNumber, len, represents the code length; for code vector h in memory pooliE is H, i is more than or equal to 1 and less than or equal to N, HiMultiplying respectively with A and C to obtain hiMapping m in two different feature spacesiAnd ci(ii) a The same operation is executed to all the code vectors in the memory pool to obtain two feature matrixes M1＝{m1,m2,…,mN} and M2＝{c1,c2,…,cN}；
(3) The obtained input vector u and a first feature matrix M are combined1Multiplying, and calculating attention vector P ═ P by softmax operation according to the obtained result1,p2,…,pN]Wherein p isiFor input vector u to ith feature vector hiThe attention of (1);
(4) for the feature matrix M2Each feature vector c iniIt is associated with the corresponding attention P in the vector PiMultiplying, summing all weighted eigenvectors to obtain a context vector o, and adding the context vector o to the input u to obtain a result r1As input for the next layer;
(5) for the G layer end-to-end memory network, repeating the steps (2) - (4), and outputting the vector r of the last layerGRespectively performing point multiplication on the M different vectors, wherein the result of each point multiplication is a predicted power; wherein G is a hyper-parameter, which is the number of layers of the end-to-end memory network.
CN201911247976.8A 2019-12-06 2019-12-06 Short-term wind power prediction method based on end-to-end memory network Active CN111160620B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911247976.8A CN111160620B (en) 2019-12-06 2019-12-06 Short-term wind power prediction method based on end-to-end memory network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911247976.8A CN111160620B (en) 2019-12-06 2019-12-06 Short-term wind power prediction method based on end-to-end memory network

Publications (2)

Publication Number Publication Date
CN111160620A CN111160620A (en) 2020-05-15
CN111160620B true CN111160620B (en) 2022-06-17

Family

ID=70556524

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911247976.8A Active CN111160620B (en) 2019-12-06 2019-12-06 Short-term wind power prediction method based on end-to-end memory network

Country Status (1)

Country Link
CN (1) CN111160620B (en)

Families Citing this family (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112653142B (en) * 2020-12-18 2022-09-02 武汉大学 Wind power prediction method and system for optimizing depth transform network
CN112991090B (en) * 2021-02-05 2023-10-20 江南大学 Photovoltaic power prediction method based on transducer model
CN113011673A (en) * 2021-03-31 2021-06-22 新奥数能科技有限公司 Method and device for monitoring and early warning water level of cooling tower
CN112949945B (en) * 2021-04-15 2022-09-02 河海大学 Wind power ultra-short-term prediction method for improving bidirectional long-term and short-term memory network
CN113408785B (en) * 2021-05-20 2023-04-07 上海晨翘智能科技有限公司 Method, device, equipment and storage medium for predicting optical power
CN113988449B (en) * 2021-11-05 2024-04-12 国家电网有限公司西北分部 Wind power prediction method based on transducer model
CN114239971A (en) * 2021-12-20 2022-03-25 浙江大学 Daily precipitation prediction method based on Transformer attention mechanism
CN114091361B (en) * 2022-01-24 2022-05-17 中汽数据(天津)有限公司 Weather event-based transform model construction method

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108711847A (en) * 2018-05-07 2018-10-26 国网山东省电力公司电力科学研究院 A kind of short-term wind power forecast method based on coding and decoding shot and long term memory network
CN109214575A (en) * 2018-09-12 2019-01-15 河海大学 A kind of super short-period wind power prediction technique based on small wavelength short-term memory network
CN109802430A (en) * 2018-12-29 2019-05-24 上海电力学院 A kind of wind-powered electricity generation power grid control method based on LSTM-Attention network

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108711847A (en) * 2018-05-07 2018-10-26 国网山东省电力公司电力科学研究院 A kind of short-term wind power forecast method based on coding and decoding shot and long term memory network
CN109214575A (en) * 2018-09-12 2019-01-15 河海大学 A kind of super short-period wind power prediction technique based on small wavelength short-term memory network
CN109802430A (en) * 2018-12-29 2019-05-24 上海电力学院 A kind of wind-powered electricity generation power grid control method based on LSTM-Attention network

Also Published As

Publication number Publication date
CN111160620A (en) 2020-05-15

Similar Documents

Publication Publication Date Title
CN111160620B (en) Short-term wind power prediction method based on end-to-end memory network
CN109214592B (en) Multi-model-fused deep learning air quality prediction method
CN110414788B (en) Electric energy quality prediction method based on similar days and improved LSTM
CN112712209B (en) Reservoir warehousing flow prediction method and device, computer equipment and storage medium
CN113554466B (en) Short-term electricity consumption prediction model construction method, prediction method and device
CN110751318A (en) IPSO-LSTM-based ultra-short-term power load prediction method
CN113988449B (en) Wind power prediction method based on transducer model
CN113344288B (en) Cascade hydropower station group water level prediction method and device and computer readable storage medium
CN115688579A (en) Basin multi-point water level prediction early warning method based on generation of countermeasure network
CN114462718A (en) CNN-GRU wind power prediction method based on time sliding window
CN111178585A (en) Fault reporting amount prediction method based on multi-algorithm model fusion
CN115271186B (en) Reservoir water level prediction and early warning method based on delay factor and PSO RNN Attention model
CN113822418A (en) Wind power plant power prediction method, system, device and storage medium
CN114997578A (en) Smart wind power plant real-time state evaluation method based on deep learning
CN114021836A (en) Multivariable reservoir water inflow amount prediction system based on different-angle fusion, training method and application
CN116014722A (en) Sub-solar photovoltaic power generation prediction method and system based on seasonal decomposition and convolution network
CN117748495A (en) Photovoltaic power station short-term generation power prediction method based on residual neural network
CN112836876B (en) Power distribution network line load prediction method based on deep learning
CN110738363A (en) photovoltaic power generation power prediction model and construction method and application thereof
Vogt et al. Wind power forecasting based on deep neural networks and transfer learning
CN113762591A (en) Short-term electric quantity prediction method and system based on GRU and multi-core SVM counterstudy
CN112232570A (en) Forward active total electric quantity prediction method and device and readable storage medium
CN117154680A (en) Wind power prediction method based on non-stationary transducer model
CN116307139A (en) Wind power ultra-short-term prediction method for optimizing and improving extreme learning machine
CN113962432A (en) Wind power prediction method and system integrating three-dimensional convolution and light-weight convolution threshold unit

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