CN115275991A - Active power distribution network operation situation prediction method based on IEMD-TA-LSTM model - Google Patents

Active power distribution network operation situation prediction method based on IEMD-TA-LSTM model Download PDF

Info

Publication number
CN115275991A
CN115275991A CN202210906679.5A CN202210906679A CN115275991A CN 115275991 A CN115275991 A CN 115275991A CN 202210906679 A CN202210906679 A CN 202210906679A CN 115275991 A CN115275991 A CN 115275991A
Authority
CN
China
Prior art keywords
data
attention
time
unit
formula
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202210906679.5A
Other languages
Chinese (zh)
Inventor
郑真
朱峰
黄晨宏
马小丽
时珊珊
姚尚坤
杜洋
颜华敏
肖远兵
肖文渊
李建宁
牟锴
张冠花
汪笃红
马晔晖
邢海军
黄一楠
田书欣
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.)
State Grid Shanghai Electric Power Co Ltd
Original Assignee
State Grid Shanghai Electric Power 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 State Grid Shanghai Electric Power Co Ltd filed Critical State Grid Shanghai Electric Power Co Ltd
Priority to CN202210906679.5A priority Critical patent/CN115275991A/en
Publication of CN115275991A publication Critical patent/CN115275991A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • H—ELECTRICITY
    • H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
    • H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
    • H02J3/00—Circuit arrangements for ac mains or ac distribution networks
    • H02J3/003—Load forecast, e.g. methods or systems for forecasting future load demand
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—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
    • 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
    • H—ELECTRICITY
    • H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
    • H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
    • H02J2203/00—Indexing scheme relating to details of circuit arrangements for AC mains or AC distribution networks
    • H02J2203/20—Simulating, e g planning, reliability check, modelling or computer assisted design [CAD]

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Economics (AREA)
  • Theoretical Computer Science (AREA)
  • Strategic Management (AREA)
  • Data Mining & Analysis (AREA)
  • Human Resources & Organizations (AREA)
  • Pure & Applied Mathematics (AREA)
  • Computational Mathematics (AREA)
  • Marketing (AREA)
  • Tourism & Hospitality (AREA)
  • Mathematical Physics (AREA)
  • General Business, Economics & Management (AREA)
  • Operations Research (AREA)
  • Mathematical Optimization (AREA)
  • Mathematical Analysis (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Algebra (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Primary Health Care (AREA)
  • Public Health (AREA)
  • Water Supply & Treatment (AREA)
  • Probability & Statistics with Applications (AREA)
  • Power Engineering (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Databases & Information Systems (AREA)
  • Software Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Development Economics (AREA)
  • Game Theory and Decision Science (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Quality & Reliability (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses an IEMD-TA-LSTM model-based active power distribution network operation situation prediction method, which comprises an input unit, a data processing unit, a situation prediction unit and an output unit, wherein the input unit is used for inputting a state of a power distribution network; the input unit constructs an input data set of the combined model; the data processing unit decomposes the original time sequence data of the distributed power supply output and the load power into a plurality of time sequence components with different characteristics and optimizes the network hyper-parameters; the situation prediction unit induces the time-space characteristics of the exogenous data and the time sequence information by using a TA-LSTM neural network which integrates a residual convolution attention, a space attention and a time attention triple attention mechanism; the output unit merges the prediction results of all the components, and then a situation evaluation index is established from the node voltage and branch load angles to obtain a final situation prediction result. The method and the device can optimize the prediction level of the operation situation of the power distribution network and improve the prediction precision.

Description

Active power distribution network operation situation prediction method based on IEMD-TA-LSTM model
Technical Field
The invention relates to an active power distribution network operation situation prediction method based on an IEMD-TA-LSTM model, which is used in the field of power distribution network operation prediction monitoring.
Background
The countryside is vast in China and has various abundant wind power and light energy resources, so that the renewable energy system has extremely high application prospect in China. However, instability, fluctuation and counterregulation of new energy power generation can bring great impact to a power system, and block safe and stable operation of a power grid, so that further development of an active power distribution network is blocked. However, if the uncertainty of the new energy output and the load can be grasped, and the accurate prediction of the operation situation of the active power distribution network can be realized, the problems can be alleviated to a certain extent. Therefore, a method for predicting the operation situation of the active power distribution network with high precision is needed to further improve the safe and reliable operation level of the active power distribution network.
The existing documents, fan Panpan, yuan Yiping, sun Wenlei, etc., published in the computer integrated manufacturing system (2021,27 (7): 1993-2004) 'wind turbine generator risk situation prediction fusing multi-period SCADA data' uses SCADA data to perform situation prediction based on LSTM neural network, and residual error is predicted through active power to quantify the severity of risk state. In the risk evaluation process, a unit risk state severity outlier model is constructed by using a fuzzy C-means algorithm to divide a state model. Zhang Qun, tang Zhenhao, wang Gong and the like, which are published in solar science and newspaper (2021,42 (10): 275-281), reconstruct wind field data according to chaos analysis results, obtain input required by a prediction model by combining with a classification prediction tree, establish a wind power prediction model by adopting a method based on a long-short term memory artificial neural network, and finally further improve the model precision by utilizing an error correction strategy based on historical prediction errors. 6253 BiLSTM multi-wind turbine generator set ultra-short term power prediction based on time mode attention mechanism, published in "high voltage technology" (2022,48 (5): 1884-1892) by Wang Yugong, shi Yunxiang, zhou Xu, etc., firstly obtains different modal components of a fan original power signal by using ensemble empirical mode decomposition, extracts complex relations among multiple fans from hidden row vectors obtained from a BiLSTM network based on a TPA mechanism, and finally applies the model to regional fan power prediction of decentralized distribution. Yang Yuqing and Zhang Yi short-term wind power prediction based on mRMR and VMD-AM-LSTM, published in control engineering (2022,29 (1): 10-17), uses a variational modal decomposition algorithm to decompose a wind power sequence into several components with different central frequencies, then establishes an attention mechanism and a long-term short-term memory hybrid prediction model for each component in combination with the meteorological features screened out by maximum correlation-minimum redundancy, and finally superposes the prediction results of each component to obtain the final wind power. Zhao Dongmei, du Gang, liu Xin and so on, which are published in modern power (2022,39 (1): 9-19), 'wind power combination prediction model based on time sequence decomposition and machine learning', respectively constructs a prediction model combining an empirical mode decomposition technology and a long-short term memory neural network, a prediction model combining a variational mode decomposition technology, a simulated annealing algorithm and a depth confidence network as a basic prediction model, and utilizes the result obtained by constructing and processing the basic model by using an extreme gradient lifting algorithm to obtain a final prediction result. Zhang Shuqing, yang Zhenning, jiang Anqi and the like, which are published in the solar science and newspaper (2022,43 (6): 204-211), short-term wind power prediction based on EN-SKPCA reduction and FPA optimization LSTMNN, perform dimension reduction on meteorological data by using an elastic network sparse kernel principal component analysis reduction method, and optimize a long-term and short-term memory neural network prediction model for wind power output prediction by using a flower pollination algorithm. The above documents only take the application of external source data such as wind direction, wind speed, weather and the like in wind power into consideration, and the phenomenon of characteristic change after random time sequence information decomposition is neglected in part of methods. How to improve the accuracy and reliability of the prediction and monitoring of the operation situation of the power grid is a main target of technicians.
Disclosure of Invention
The invention aims to overcome the defects of the prior art, and provides an active power distribution network operation situation prediction method based on an IEMD-TA-LSTM model, which can optimize the power distribution network operation situation prediction level and improve the prediction precision.
One technical scheme for achieving the above purpose is as follows: an active power distribution network operation situation prediction method based on an IEMD-TA-LSTM model is characterized by comprising an input unit, a data processing unit, a situation prediction unit and an output unit;
the method comprises the steps that an input unit obtains distributed energy output time sequence data, internal element data of distribution network load data, meteorological data and external source time sequence data of electric power market data, and an input data set of a combined model is jointly constructed;
the data processing unit decomposes the original time sequence data of the distributed power supply output and the load power into a plurality of time sequence components with different characteristics by using a fully adaptive noise set empirical mode decomposition (IEMD) algorithm according to the data provided by the input unit, and introduces an improved manta ray optimization IMRFO algorithm to optimize network hyper-parameters;
the situation prediction unit induces the time-space characteristics of the exogenous data and the time sequence information by using a TA-LSTM neural network which integrates a residual convolution attention, a space attention and a time attention triple attention mechanism based on the time sequence data and the network hyper-parameters output by the data processing unit;
the output unit merges the prediction results of all the components, and then a situation evaluation index is established from the node voltage and branch load angles to obtain a final situation prediction result.
Further, the input unit refers to internal element data of history and actually measured time sequence data of distributed energy output, history and actually measured data of power distribution network load, and history and actually measured external time sequence data of meteorological data and electric power market data.
Furthermore, the data processing unit is composed of an IEMD decomposition unit and an IMRFO optimization unit;
the IEMD decomposition unit decomposes the wind power output and load power original time sequence components into a plurality of components with single and obvious characteristics by using an IEMD algorithm, and the flow is as follows:
step 1, adding Gaussian white noise into an original time sequence to obtain an ith to-be-decomposed signal xi(t)；
xi(t)＝x(t)+ε0E1(ωi(t))，i＝1,2,...,n
In the formula, omegai(t) is zero mean unit variance Gaussian white noise, ε0As magnitude of noise, E1(ωi(t)) is ωi(t) a first EMD component;
step 2, calculating x by EMD decompositioni(t) local mean to obtain the residual r of the first decomposition1(t) and a first modal component
Figure BDA0003772731890000031
Figure BDA0003772731890000032
In the formula, M (') is a local mean operator;
step 3, calculating a second residual error r2(t) and a second modal component
Figure BDA0003772731890000041
Figure BDA0003772731890000042
Step 4, for each of the remaining stages, i.e., k =1,2.
Figure BDA0003772731890000043
And 5, repeating the step 4 until the residual signal can not be decomposed any more, and finally obtaining n modal components and a final residual signal Re (t). The original signal sequence x (t) can be expressed as:
Figure BDA0003772731890000044
the IMRFO optimization unit is used for carrying out optimization execution strategy on the neural network hyper-parameter by utilizing an improved manta ray optimization algorithm according to different time sequence component characteristics obtained by the IEMD decomposition unit;
for the chain type foraging process of the bat ray, the mathematical expression of the position update is as follows:
Figure BDA0003772731890000045
Figure BDA0003772731890000046
in the formula (I), the compound is shown in the specification,
Figure BDA0003772731890000047
and
Figure BDA0003772731890000048
the position of the ith individual and the optimal individual in the d-dimensional space of the tth generation; r is [0,1]Uniformly distributed random numbers; alpha is a weight factor;
for the spiral feeding searching process of the manta ray, the mathematical expression of the position update is as follows:
when T/T > rand
Figure BDA0003772731890000051
Figure BDA0003772731890000052
Wherein, beta is a weight factor; r is1Is [0,1]Uniformly distributed random numbers; t is the maximum iteration number;
when T/T is less than or equal to rand
Figure BDA0003772731890000053
Figure BDA0003772731890000054
In the formula (I), the compound is shown in the specification,
Figure BDA0003772731890000055
for a new position, x, generated randomlyuAnd xlFor the upper and lower bounds of the search space, rjIs [0,1]Uniformly distributed random numbers;
for the somersault type food searching strategy, the position updating expression is as follows:
Figure BDA0003772731890000056
in the formula, S is a somersault factor; r isj2And rj3Is [0,1]Uniformly distributed random numbers;
the bat ray optimization algorithm cannot ensure that the optimal solution is globally optimal and the diversity of the population is poor, and the improvement strategy is as follows:
step 1, obtaining an initial population based on an improved Tent mapping method;
setting the initial position of the bat ray population by Tent mapping, wherein the expression is as follows:
Figure BDA0003772731890000057
because the chaos sequence generation iteration process has the defects of short period and unstable period, a random equation method is introduced to improve the chaos sequence generation iteration process, namely when the chaos sequence generation iteration process is finished
Figure BDA0003772731890000058
A small period cycle falling within the unstable period point or 5 is then improved by:
Figure BDA0003772731890000061
in the formula, rtIs [0,1]Uniformly distributed random numbers;
step 2, self-adaptive nonlinear descending weight coefficient;
in the chain type searching and optimizing iterative process of MRFO algorithm, r is divided intojThe method is changed into the following form:
Figure BDA0003772731890000062
wherein r is increased with tjThe optimization algorithm is gradually reduced, and the local optimization capability in the early stage of algorithm iteration is enhanced;
in the process of searching for food by the muscle turning bucket, the muscle turning bucket factor S is a certain value, and in the later iteration stage, the algorithm needs smaller search step length, and an overlarge value can weaken the search capability of the algorithm, and similarly, the S can be changed into the following form:
Figure BDA0003772731890000063
in the formula, SmaxAnd SminFor the upper and lower limits of the somersault factor S, the S is gradually reduced along with the increase of t, and the local search and optimization capability in the later period of iteration is enhanced.
Further, the situation prediction unit is composed of a residual attention subunit, a spatial attention subunit, a temporal attention subunit and an LSTM prediction subunit;
the residual error attention subunit consists of a structure formed by connecting a plurality of convolution layer residual errors;
the 1 st layer is an input layer, the input phasor is exogenous time sequence data with given length, and K is the exogenous data quantity:
X＝(x1,x2,…,xK)
wherein, the specific expression of the exogenous data is as follows:
Xi＝(xi,t-d+1,xi,t-d+2,…,xi,t)T,i＝1,2…K
in the formula, d is a time step;
layers 2 and 3 are two-dimensional convolution modules, and the convolution layers do not change the dimension and length of input data:
H1＝(ωc1*X+bc1)
H2＝(ωc2*X+bc2)+X
in the formula, omegac1And omegac2Are all convolution weights; bc1And bc2Are all bias terms;
the space attention subunit performs space feature extraction on the preliminary feature extraction data provided by the last unit by using a space attention mechanism to further obtain an encoded time sequence feature sequence;
take the t-th time step as an example, a single-step input feature vector containing K features
Figure BDA0003772731890000071
Figure BDA0003772731890000072
Attention weight vector etThe calculation formula is as follows:
Figure BDA0003772731890000073
in the formula (I), the compound is shown in the specification,
Figure BDA0003772731890000074
distributing attention weight corresponding to each input feature at the time t; v. ofe、weAnd beA weight matrix and a bias vector that are attention weights; sigma represents a sigmoid activation function;
by pair etUsing Softmax function to normalize and obtain the kth characteristic attention score
Figure BDA0003772731890000075
Inputting a feature vector XtRecalculated as a weighted vector
Figure BDA0003772731890000076
Figure BDA0003772731890000077
Figure BDA0003772731890000078
The time attention subunit extracts time sequence data from the time sequence characteristics of the time sequence data by using a time attention mechanism, and decodes the coding information provided by the time attention subunit for final prediction;
xtobtaining the hidden layer state h of the LSTM after the encoding of the feature attention modulet：
Figure BDA0003772731890000079
In the formula (f)LSTMRepresents an LSTM cell;
through time attention module pair htPerforming decoding operation to give different weight scores to the hidden state output by the LSTM unit
Figure BDA00037727318900000710
And weighting with the hidden layer state at the corresponding historical moment to obtain the comprehensive time sequence information state rtThe calculation process is shown as the following formula:
Figure BDA00037727318900000711
Figure BDA00037727318900000712
Figure BDA00037727318900000713
in the formula (I), the compound is shown in the specification,
Figure BDA0003772731890000081
distributing time attention weights corresponding to various historical moments at the time t, wherein tau is the length of an input sequence time window; v. ofd、wdAnd bdA weight matrix and a bias vector that are attention weights;
finally, a prediction result y of the future n steps is obtained through an output layerτ+n：
yτ+n＝σ(wrrt+br)
In the formula, wrAnd brRespectively a weight matrix and an offset vector of the full connection layer;
finally, obtaining a prediction result y of the future n steps through an output layerτ+n：
yτ+n＝σ(wrrt+br)
In the formula, wrAnd brRespectively a weight matrix and an offset vector of the full connection layer;
the LSTM prediction subunit performs high-precision time sequence prediction on the wind power output and the load power by using complete feature extraction information provided by the multi-attention mechanism unit; the concrete structure is as follows:
forgetting gate FG controls last time unit state ct-1The information to be saved is saved in the current unit state ctIn (1), the input gate IG controls the input x at the current timetHow much information is saved to the current cell state ctThe output gate OG controls the current cell state ctThe information amount in the previous step is saved to the current output state htThe preparation method comprises the following steps of (1) performing; the LSTM algorithm is updated and calculated as follows:
ft＝σ(wf·[ht-1,xt]+bf)
it＝σ(wi·[ht-1,xt]+bi)
ot＝σ(wo·[ht-1,xt]+bo)
Figure BDA0003772731890000082
Figure BDA0003772731890000083
Figure BDA0003772731890000084
in the formula (f)t、it、ot、ctAnd htAre respectively FG、IG、OGA state matrix of a memory cell and an output cell; sigma represents a sigmoid activation function;
Figure BDA0003772731890000085
is a candidate value vector of the current memory cell state; w is af、wi、woAnd wcAre respectively FG、IG、OGAnd a weight matrix of the memory cells; bf、bi、boAnd bcAre respectively FG、IG、OGAnd biasing of the memory cell; tanh is the activation function;
Figure BDA0003772731890000086
representing the hadamard product.
Further, the output unit comprises situation evaluation indexes of node voltage out-of-limit margin, branch load severity and voltage/current fluctuation coefficients, and the operation situation of the active power distribution network is quantitatively analyzed through calculation results of the indexes;
aiming at the problem of voltage out-of-limit possibly caused by node voltage change, a node voltage out-of-limit margin is adopted
Figure BDA00037727318900000916
The future change trend of the voltage of each node of the power distribution network is reflected:
Figure BDA0003772731890000091
in the formula (I), the compound is shown in the specification,
Figure BDA0003772731890000092
the voltage per unit value of the node i to be predicted at the future time t is obtained;
Figure BDA0003772731890000093
the mean value of the upper limit and the lower limit of the ith node voltage is obtained;
aiming at the overload safety problem of branch current, the load severity of the branch is adopted
Figure BDA0003772731890000094
Reflecting the future change trend of each branch:
Figure BDA0003772731890000095
Figure BDA0003772731890000096
in the formula, LRl,tThe load rate of the first line at the future time t;
Figure BDA0003772731890000097
the current of the first branch to be predicted at the future time t; i isl,NThe rated current of the first line;
in order to quantify the fluctuation degree of the node voltage and the branch current and simultaneously evaluate the strength of the node voltage/branch current in the active power distribution network influenced by new energy and load change, a node voltage fluctuation coefficient is provided
Figure BDA0003772731890000098
And branch current fluctuation coefficient
Figure BDA0003772731890000099
Indexes are respectively expressed as follows:
Figure BDA00037727318900000910
Figure BDA00037727318900000911
in the formula (I), the compound is shown in the specification,
Figure BDA00037727318900000912
and
Figure BDA00037727318900000913
the voltage value of a node i to be predicted at the future time t and the current value of a branch l to be predicted at the future time t are obtained;
Figure BDA00037727318900000914
and
Figure BDA00037727318900000915
and the predicted voltage average value of the node i to be predicted in the future D period and the predicted current average value of the branch l to be predicted in the future D period are obtained.
The active power distribution network operation situation prediction method based on the IEMD-TA-LSTM model can fully utilize the active power distribution network operation situation combined prediction method of internal element data and exogenous meteorological data, and introduces an intelligent optimization algorithm and a multiple attention mechanism to optimize model parameters by taking different time sequence characteristics after random data decomposition as targets, so that the prediction precision of the combined model is improved.
Drawings
FIG. 1 is a schematic flow diagram of an operation situation prediction method of an active power distribution network based on an IEMD-TA-LSTM model according to the present invention;
FIG. 2 is a diagram of the DA-LSTM model architecture;
FIG. 3 is a schematic diagram of a residual attention mechanism;
FIG. 4 is a diagram of the structure of an LSTM unit;
fig. 5 is a schematic diagram of an improved IEEE-33 node active power distribution network.
Detailed Description
In order to better understand the technical solution of the present invention, the following detailed description is made by specific examples:
as shown in fig. 1, an IEMD-TA-LSTM combined prediction model for an active power distribution network operation situation of a fused Improved fully adaptive noise ensemble empirical mode decomposition (IEMD), a Triple-stage association (TA), and a Long short-term memory neural network (LSTM) includes an input unit 1, a data processing unit 2, a situation prediction unit 3, and an output unit 4, which are connected in sequence.
The input unit 1 acquires internal element data and exogenous time sequence data of distributed energy output time sequence data, distribution network load data and meteorological data and electric power market data, and an input data set of a combined model is constructed together.
The data processing unit 2 decomposes the distributed power output and load power original time sequence data into a plurality of time sequence components with different characteristics by using an improved fully adaptive noise set empirical mode decomposition (IEMD) algorithm according to the data provided by the input unit 1, and introduces an improved manta ray optimization IMRFO algorithm to optimize network hyper-parameters.
The situation prediction unit 3 fully excavates the time-space characteristics of the external source data and the time sequence information by using a TA-LSTM neural network integrating triple attention mechanisms of residual convolution attention, spatial attention and temporal attention based on the time sequence data and the network hyper-parameters output by the data processing unit 2.
The output unit 4 combines the prediction results of the components, and then establishes a situation evaluation index from the node voltage and branch load angles to obtain a final situation prediction result.
The input unit 1 refers to internal element data of history and actually measured time sequence data of distributed energy output, history and actually measured data of power distribution network load, and history and actually measured external time sequence data of meteorological data and electric power market data.
The data processing unit 2 comprises an IEMD decomposition unit 21 and an IMRFO optimization unit 22.
The IEMD decomposition unit 21 mainly uses an IEMD algorithm, which adds a certain amount of white gaussian noise and averages the generated signal components in the iterative process of decomposing the output of the distributed power source represented by wind power and the load power signal.
The IEMD algorithm can decompose the wind power output and load power original time sequence components into a plurality of components with single and obvious characteristics, and the simple flow is as follows:
step 1, adding Gaussian white noise into an original time sequence to obtain an ith to-be-decomposed signal xi(t)；
xi(t)＝x(t)+ε0E1(ωi(t))，i＝1,2,...,n
In the formula, ωi(t) is zero mean unit variance Gaussian white noise, ε0As the magnitude of noise, E1(ωi(t)) is ωi(t) a first EMD component of (t).
Step 2, calculating x by EMD decompositioni(t) local mean to obtain the residual r of the first decomposition1(t) and a first modal component
Figure BDA0003772731890000111
Figure BDA0003772731890000112
In the formula, M (") is a local mean operator.
Step 3, calculating a second residual error r2(t) and a second modal component
Figure BDA0003772731890000113
Figure BDA0003772731890000114
Step 4, for each of the remaining stages, i.e., k =1,2.
Figure BDA0003772731890000121
And 5, repeating the step 4 until the residual signal can not be decomposed, and finally obtaining n modal components and a final residual signal Re (t). The original signal sequence x (t) can be expressed as:
Figure BDA0003772731890000122
the IMRFO optimizing unit 22 performs an optimization execution strategy on the neural network hyper-parameter by using an improved mantray optimization (IMRFO) algorithm according to different time sequence component characteristics obtained by the IEMD decomposing unit 21, so as to avoid manually adjusting the neural network hyper-parameter corresponding to each component each time.
A Manta ray optimization (MRFO) algorithm is three kinds of feeding strategies of a population of Manta ray, which are provided by observing feeding behaviors of the Manta ray, and the three kinds of feeding strategies are respectively chain feeding, whirlwind feeding and somersault feeding. And exploring and developing the search space through different food searching strategies to update the individual position.
For the chain type foraging process of the bat ray, the mathematical expression of the position update is as follows:
Figure BDA0003772731890000123
Figure BDA0003772731890000124
in the formula (I), the compound is shown in the specification,
Figure BDA0003772731890000125
and
Figure BDA0003772731890000126
the position of the ith individual and the optimal individual in the d-dimensional space of the tth generation; r is [0,1]Uniformly distributed random numbers; alpha is a weighting factor.
For the spiral type grazing process of the bat ray, the mathematical expression of the position update is as follows:
when T/T > rand
Figure BDA0003772731890000131
Figure BDA0003772731890000132
Wherein, beta is a weight factor; r is1Is [0,1]Uniformly distributed random numbers; and T is the maximum iteration number.
When T/T is less than or equal to rand
Figure BDA0003772731890000133
Figure BDA0003772731890000134
In the formula (I), the compound is shown in the specification,
Figure BDA0003772731890000135
for a new position, x, generated randomlyuAnd xlFor the upper and lower bounds of the search space, rjIs [0,1]Uniformly distributed random numbers.
For the somersault type food searching strategy, the position updating expression is as follows:
Figure BDA0003772731890000136
in the formula, S is a somersault factor; r is a radical of hydrogenj2And rj3Is [0,1]Uniformly distributed random numbers.
The algorithm needs few iterations and is high in efficiency, but in practical application, the defects that the optimal solution cannot be guaranteed to be globally optimal, the population diversity is poor and the like are also discovered. Therefore, an improved MRFO algorithm is provided for solving the problem that the TA-LSTM network has difficulty in determining the hyper-parameters of the TA-LSTM network during training and learning. The improvement strategy is specifically as follows:
step 1, obtaining an initial population based on an improved Tent mapping method;
the initial population of random production will affect the speed and accuracy of the algorithm solution. The characteristics of randomness, regularity, ergodicity and the like of the chaotic sequence enable the generated initial solution to be complex and diverse, so that the initial solution generated by the chaotic sequence is easy to converge. Compared with other mappings, tent mapping can generate more uniformly distributed sequences, so the patent adopts Tent mapping to set the initial position of the bat ray population, and the expression is as follows:
Figure BDA0003772731890000145
because the chaos sequence generation iteration process has the defects of short period and unstable period, a random equation method is introduced to improve the chaos sequence generation iteration process, namely when the chaos sequence generation iteration process is finished
Figure BDA0003772731890000141
A small period cycle falling within the unstable period point or 5 is then improved by:
Figure BDA0003772731890000142
in the formula, rtIs [0,1]Uniformly distributed random numbers.
Step 2, self-adaptive nonlinear descending weight coefficient;
in the chain type searching and optimizing iterative process of MRFO algorithm, r isjIn [0,1]Internally generated randomly, which makes the later searching step length the same as the earlier one, thereby influencing the speed of force searching, so r can be usedjThe method is changed into the following form:
Figure BDA0003772731890000143
wherein r is increased with tjThe local optimization capability in the early stage of algorithm iteration is enhanced.
In the process of searching for food by the muscle turning bucket, the muscle turning bucket factor S is a certain value. In the later period of iteration, the algorithm needs smaller search step length, and the search capability of the algorithm is weakened by an overlarge value. Similarly, S can be modified to the following form:
Figure BDA0003772731890000144
in the formula, SmaxAnd SminFor the upper and lower limits of the somersault factor S, the S is gradually reduced along with the increase of t, and the local search and optimization capability in the later period of iteration is enhanced.
The situation prediction unit 3 is composed of a residual attention subunit 31, a spatial attention subunit 32, a temporal attention subunit 33, and an LSTM prediction subunit 34, and the network structure thereof is shown in fig. 2. Wherein, the residual attention mechanism strengthens the influence of external source data on the extraction of the internal element data characteristics; the space attention mechanism is used for calculating the contribution degree of the input features to the target features; time is paid attention to a mechanism for calculating the contribution of historical time to current time prediction, and the prediction accuracy of the prediction model is improved through the distribution of differential dynamic weights.
The residual attention subunit 31 is formed by a structure in which a plurality of convolution layer residuals are connected, the structure performs feature extraction on input external source data, and excessive information is not lost in the feature extraction process by means of a residual block structure. The network structure is shown in fig. 3.
The 1 st layer is an input layer, the input phasor is exogenous time sequence data with given length, and K is the exogenous data quantity:
X＝(x1,x2,…,xK)
wherein, the specific expression of the exogenous data is as follows:
Xi＝(xi,t-d+1,xi,t-d+2,…,xi,t)T,i＝1,2…K
wherein d is a time step.
Layers 2 and 3 are two-dimensional convolution modules, and the convolution layers do not change the dimension and length of input data:
H1＝(ωc1*X+bc1)
H2＝(ωc2*X+bc2)+X
in the formula, ωc1And ωc2Are all convolution weights; bc1And bc2Are all bias terms.
Attention can be paid to the fact that correlation between the feature sequence and the target feature can be mined, and dynamic weight distribution is conducted on the features with different influence degrees through a weight probability distribution mechanism. Taking the t-th time step as an example, the single-step input feature vector containing K features
Figure BDA0003772731890000151
Attention weight vector etThe calculation formula is as follows:
Figure BDA0003772731890000152
in the formula (I), the compound is shown in the specification,
Figure BDA0003772731890000153
distributing attention weight corresponding to each input feature at the time t; v. ofe、weAnd beA weight matrix and a bias vector that are attention weights; σ denotes the sigmoid activation function.
By pair etNormalization processing is carried out by using Softmax function to obtainKth feature attention score
Figure BDA0003772731890000154
Inputting a feature vector XtRecalculated as a weighted vector
Figure BDA0003772731890000155
Figure BDA0003772731890000156
Figure BDA0003772731890000157
The time attention subunit 33 performs time series data extraction on the time series characteristics of the time series data by using a time attention mechanism, and decodes the encoded information provided by the time attention subunit for final prediction.
The time attention mechanism distinguishes the influence of the time sequence information carried by each historical moment of the input sequence on the prediction output of the current moment by distributing attention weight to the time sequence information. x is the number oftObtaining the hidden layer state h of the LSTM after the encoding of the feature attention modulet：
Figure BDA0003772731890000161
In the formula (f)LSTMIndicating an LSTM unit.
Through time attention module pair htPerforming decoding operation to give different weight scores to the hidden state output by the LSTM unit
Figure BDA0003772731890000162
And weighting with the hidden layer state at the corresponding historical moment to obtain the comprehensive time sequence information state rtThe calculation process is shown as the following formula:
Figure BDA0003772731890000163
Figure BDA0003772731890000164
Figure BDA0003772731890000165
in the formula (I), the compound is shown in the specification,
Figure BDA0003772731890000166
distributing time attention weights corresponding to various historical moments at the time t, wherein tau is the length of an input sequence time window; v. ofd、wdAnd bdA weight matrix and a bias vector which are attention weights.
Finally, a prediction result y of the future n steps is obtained through an output layerτ+n：
yτ+n＝σ(wrrt+br)
In the formula, wrAnd brRespectively, the weight matrix and the offset vector of the fully-connected layer.
The LSTM prediction subunit 34 performs high-precision time sequence prediction of the wind power output and the load power by using the complete feature extraction information provided by the multiple attention mechanism unit.
The LSTM adds a Forgetting Gate (FG), an Input Gate (IG) and an Output Gate (OG) on the basis of a circulating neural network, and can introduce a memory function module for storing a data state in the structure, thereby effectively solving the problem that the RNN can not construct a long-time sequence prediction model due to gradient surge and disappearance. The specific cell structure of LSTM is shown in fig. 4.
FG controls last moment cell state ct-1The information needed to be saved is saved in the current unit state ctIn the middle, IG controls the input x at the current timetHow much information is saved to the current cell state ctIn the OG control, the OG controls the currentCell state ctThe information amount in the previous step is saved to the current output state htIn (1). The LSTM algorithm is updated and calculated as follows:
ft＝σ(wf·[ht-1,xt]+bf)
it＝σ(wi·[ht-1,xt]+bi)
ot＝σ(wo·[ht-1,xt]+bo)
Figure BDA0003772731890000171
Figure BDA0003772731890000172
Figure BDA0003772731890000173
in the formula (f)t、it、ot、ctAnd htAre respectively FG、IG、OGA state matrix of a memory cell and an output cell; sigma represents a sigmoid activation function;
Figure BDA0003772731890000174
a candidate vector for the current cell state; w is af、wi、woAnd wcAre respectively FG、IG、OGAnd a weight matrix of the memory cells; bf、bi、boAnd bcAre respectively FG、IG、OGAnd biasing of the memory cell; tanh is the activation function;
Figure BDA0003772731890000175
representing the hadamard product.
The output unit 4 mainly comprises situation evaluation indexes such as node voltage out-of-limit margin, branch load severity and voltage/current fluctuation coefficients, and the operation situation of the active power distribution network is quantitatively analyzed through calculation results of the indexes.
Aiming at the problem of voltage out-of-limit possibly caused by node voltage change, a node voltage out-of-limit margin is adopted
Figure BDA0003772731890000176
The future change trend of the voltage of each node of the power distribution network is reflected:
Figure BDA0003772731890000177
in the formula (I), the compound is shown in the specification,
Figure BDA0003772731890000178
the voltage per unit value of the node i to be predicted at the future time t is obtained;
Figure BDA0003772731890000179
is the average value of the upper limit and the lower limit of the voltage of the ith node.
Aiming at the overload safety problem of branch current, the severity of branch load is adopted
Figure BDA00037727318900001710
Reflecting the future change trend of each branch:
Figure BDA00037727318900001711
Figure BDA0003772731890000181
in the formula, LRl,tThe load factor of the first line at the future time t;
Figure BDA0003772731890000182
the current of the first branch to be predicted at the future time t; I.C. Al,NThe rated current of the l line.
In order to quantify the fluctuation degree of node voltage and branch current and simultaneously evaluate the strength of the influence degree of the node voltage/branch current in the active power distribution network by new energy and load change, a node voltage fluctuation coefficient is provided
Figure BDA0003772731890000183
And branch current fluctuation coefficient
Figure BDA0003772731890000184
Indexes are respectively expressed as follows:
Figure BDA0003772731890000185
Figure BDA0003772731890000186
in the formula (I), the compound is shown in the specification,
Figure BDA0003772731890000187
and
Figure BDA0003772731890000188
the voltage value of a node i to be predicted at the future time t and the current value of a branch l to be predicted at the future time t are obtained;
Figure BDA0003772731890000189
and
Figure BDA00037727318900001810
and the predicted voltage average value of the node i to be predicted in the future D period and the predicted current average value of the branch l to be predicted in the future D period are obtained.
Examples
The example takes a modified IEEE-33 node active power distribution network as a research object, and the topological structure of the example is shown in figure 5. The system comprises 33 nodes and 32 branches, and wind power generators are connected to the nodes 16, 20 and 31, and the capacity of the wind power generators is 500kW, 500kW and 1000kW respectively. Load power data and environment data of a certain area from 1 month and 1 day in 2017 to 12 months and 31 days in 2017 are selected as historical data, actual measurement fan historical data of an actual wind power plant of the certain area in the same time period are combined, sampling intervals are all 15min, and all the used data are processed by abnormal data. The IEMD-TA-LSTM algorithm is adopted to predict the change trend of the loads of all nodes and the wind power output in the system in 24 hours in the future, the operation situation of the power distribution network is obtained through the trend calculation of the active power distribution network, and evaluation indexes of the situation of the active power distribution network are output.
The standard deviation of the IEMD algorithm for adding the Gaussian white noise is 0.2, and the adding times are 200. Because too many input samples are limited by giving the IEMD decomposition results of the node 16 wind power output data (rated power 1000 kW) and the node 7 load data (rated power 200 kW) of one week only in space, the original data are decomposed through the IEMD algorithm, the non-stationarity and irregularity of the original data can be reduced, and the accuracy of the prediction of the operation elements of the power distribution network is improved.
IMRFO algorithm settings: the population size is 10 and the upper limit of the number of iterations is 20. A decomposition sequence of load power data of the node 7: (
Figure BDA0003772731890000191
And Re) and corresponding daily environmental data as inputs to the TA-LSTM prediction model, with the first 70% of the load history data as the training set and the last 30% as the test set. And (4) predicting the load of 24 hours in the future with the point 24 of 11, 27 and 27 in 2017 as the starting point, and obtaining a load change prediction result at 0-24 days on 28 days. Similarly, taking the wind power output at the node 16 as an example, the wind power change trend at 0-24 time in the future is predicted by the same method.
To verify the prediction performance of the IEMD-TA-LSTM model proposed in this patent, root Mean Square Error (RMSE), mean Absolute Error (MAE) and goodness-of-fit coefficient (R-Square, R-LSTM) are used2) The index evaluates the predicted result. Load prediction results and electric power prediction results of LSTM, TA-LSTM, IMRFO-TA-LSTM and the proposed IEMD-TA-LSTM prediction models are contrastively analyzed, and the calculation results of the prediction performance indexes of the models are respectively shown in tables 1 and 2.
TABLE 1 evaluation index comparison of load prediction results
Prediction model RMSE/kW MAE/kW R2
IEMD-TA-LSTM 0.8026 0.6461 0.9720
IMRFO-TA-LSTM 2.0181 1.4509 0.8230
TA-LSTM 2.1214 1.5907 0.8044
LSTM 3.0114 2.4548 0.6058
TABLE 2 evaluation index comparison of wind power prediction results
Figure BDA0003772731890000192
Figure BDA0003772731890000201
As can be seen from table 1, the simple LSTM model has a poor prediction effect on load prediction. Compared with IMRFO-TA-LSTM, TA-LSTM and LSTM, the IEMD-TA-LSTM prediction model reduces 60.23%, 62.16% and 73.35% in RMSE aspect; the MAE index is reduced by 55.47%, 59.08% and 73.68% respectively; r2The values are respectively improved by 18.10%, 20.84% and 60.45%, which indicates that the model has excellent prediction performance. Comparing TA-LSTM and single LSTM, the RMSE and MAE indexes of the prediction result of the TA-LSTM model are respectively reduced by 29.55 percent and 35.20 percent, and R is reduced by2The value is improved by 32.78%, which shows that by introducing the residual convolution processing and the multiple attention mechanism of space and time, the characteristic information, the time sequence information and the correlation of the characteristic information and the time sequence information can be effectively mined, and the prediction performance of the LSTM is improved; the IMRFO-TA-LSTM is better than the TA-LSTM, which shows that the IMRFO algorithm can find the optimal hyper-parameter combination, so that the model prediction performance is improved; the IEMD-TA-LSTM model provided by the patent has better performance than the IMRFO-TA-LSTM algorithm, shows that through sequence decomposition, the regularity of the decomposed sequence is better, and the prediction precision of the algorithm is effectively improved.
In order to obtain the short-term operation situation of the active power distribution network, firstly, the prediction results of all load nodes and wind power and electric power are obtained according to the IEMD-TA-LSTM model. After the change trend of the load and the wind power of the IEEE33 node system in the future 24 hours is obtained, node voltage out-of-limit margin indexes of each node of the system in the future 24 hours and branch load severity indexes of each branch are obtained through load flow calculation analysis of multiple time discontinuities respectively.
It can be seen that the node voltage out-of-limit margin indexes based on the predicted values of the load and the wind power are all in the range of [0.95,1], the node voltage is out-of-limit condition when no node voltage exists, and each node voltage is in a normal state.
The branch load severity indexes based on the predicted values of the load and the wind power are all in the value of [0,0.7]However, it can also be seen that the branch load severity indicator of the branches 22 and 23 is close to 0.7 for a long time, and there is a potential heavy-duty operation risk. And in order to evaluate the accuracy of the node voltage out-of-limit margin index and the branch load severity index operation situation prediction result, carrying out quantitative analysis by using the MAE index. And obtaining the voltage out-of-limit margin index of each node and the MAE of the branch load severity index of each branch on each time section based on the actual value and the predicted value of the load and the wind power. MAE indexes of node voltage out-of-limit margin indexes of all nodes are lower than 1.3 multiplied by 10-3p.u., the MAE index of the voltage phase angle of each node is lower than 1.5 x 10-2And p.u., the prediction result is more accurate. In addition, as for the fluctuation coefficient indexes of each node voltage and each branch current, the fluctuation degree index obtained based on the real value of the load and the wind power basically matches with the fluctuation degree index obtained based on the predicted value of the load and the wind power, and the error is small. Therefore, the IEMD-TA-LSTM combined prediction model can accurately track the node voltage out-of-limit margin of each node and each branch of the whole network, the branch load severity index and other operation situations.
It should be understood by those skilled in the art that the above embodiments are only for illustrating the present invention and are not to be used as a limitation of the present invention, and that the changes and modifications of the above embodiments are within the scope of the appended claims as long as they are within the true spirit of the present invention.

Claims (5)

1. An IEMD-TA-LSTM model-based active power distribution network operation situation prediction method is characterized by comprising an input unit, a data processing unit, a situation prediction unit and an output unit;
the method comprises the steps that an input unit obtains distributed energy output time sequence data, internal element data and meteorological data of distribution network load data and external source time sequence data of electric power market data, and an input data set of a combined model is constructed together;
the data processing unit decomposes the original time sequence data of the distributed power output and the load power into a plurality of time sequence components with different characteristics by using a fully adaptive noise set empirical mode decomposition (IEMD) algorithm according to the data provided by the input unit, and introduces an improved manta ray optimization IMRFO algorithm to optimize network hyper-parameters;
the situation prediction unit induces the time-space characteristics of the exogenous data and the time sequence information by using a TA-LSTM neural network which integrates a residual convolution attention, a space attention and a time attention triple attention mechanism based on the time sequence data and the network hyper-parameters output by the data processing unit;
the output unit merges the prediction results of all the components, and then a situation evaluation index is established from the node voltage and branch load angles to obtain a final situation prediction result.
2. The method of claim 1, wherein the input unit refers to internal element data of historical and measured time series data of distributed energy output, historical and measured data of distribution network load, and historical and measured external time series data of meteorological data and electric power market data.
3. The method according to claim 1, wherein the data processing unit comprises an IEMD decomposition unit and an IMRFO optimization unit;
the IEMD decomposition unit decomposes the wind power output and load power original time sequence components into a plurality of components with single and obvious characteristics by using an IEMD algorithm, and the flow is as follows:
step 1, adding Gaussian white noise into an original time sequence to obtain an ith to-be-decomposed signal xi(t)；
xi(t)＝x(t)+ε0E1(ωi(t))，i＝1,2,...,n
In the formula, ωi(t) is zero mean unit variance Gaussian whiteNoise, epsilon0As the magnitude of noise, E1(ωi(t)) is ωiA first EMD component of (t);
step 2, calculating x by EMD decompositioni(t) local mean to obtain the residual r of the first decomposition1(t) and a first modal component
Figure FDA0003772731880000011
Figure FDA0003772731880000012
In the formula, M (') is a local mean operator;
step 3, calculating a second residual error r2(t) and a second modal component
Figure FDA0003772731880000021
Figure FDA0003772731880000022
Step 4, for each of the remaining phases, i.e., k =1,2,..., n, calculating the kth residual signal and the kth modal component according to the above steps, resulting in:
Figure FDA0003772731880000023
and 5, repeating the step 4 until the residual signal can not be decomposed any more, and finally obtaining n modal components and a final residual signal Re (t). The original signal sequence x (t) can be expressed as:
Figure FDA0003772731880000024
the IMRFO optimization unit is used for carrying out optimization execution strategy on the neural network hyper-parameters by utilizing an improved manta ray optimization algorithm according to different time sequence component characteristics obtained by the IEMD decomposition unit;
for the chain type foraging process of the bat ray, the mathematical expression of the position update is as follows:
Figure FDA0003772731880000025
Figure FDA0003772731880000026
in the formula (I), the compound is shown in the specification,
Figure FDA0003772731880000027
and
Figure FDA0003772731880000028
the position of the ith individual and the optimal individual in the d-dimensional space of the tth generation; r is [0,1]Uniformly distributed random numbers; alpha is a weight factor;
for the spiral type grazing process of the bat ray, the mathematical expression of the position update is as follows:
when T/T > rand
Figure FDA0003772731880000029
Figure FDA00037727318800000210
Wherein, beta is a weight factor; r is1Is [0,1]Uniformly distributed random numbers; t is the maximum iteration number;
when T/T is less than or equal to rand
Figure FDA00037727318800000211
Figure FDA0003772731880000031
In the formula (I), the compound is shown in the specification,
Figure FDA0003772731880000032
for a new position, x, generated randomlyuAnd xlFor the upper and lower bounds of the search space, rjIs [0,1]Uniformly distributed random numbers;
for the somersault type food searching strategy, the position updating expression is as follows:
Figure FDA0003772731880000033
in the formula, S is a somersault factor; r is a radical of hydrogenj2And rj3Is [0,1]Uniformly distributed random numbers;
the manta ray optimization algorithm cannot ensure that the optimal solution is global optimal and the diversity of the population is poor, and the improvement strategy is as follows:
step 1, obtaining an initial population based on an improved Tent mapping method;
setting the initial position of the bat ray population by Tent mapping, wherein the expression is as follows:
Figure FDA0003772731880000034
because the chaos sequence generation iteration process has the defects of short period and unstable period, a random equation method is introduced to improve the chaos sequence generation iteration process, namely when the chaos sequence generation iteration process is finished
Figure FDA0003772731880000035
A small period cycle falling within the unstable period point or 5 is then improved by:
Figure FDA0003772731880000036
in the formula, rtIs [0,1]Uniformly distributed random numbers;
step 2, self-adaptive nonlinear descending weight coefficient;
in the chain type searching and optimizing iterative process of MRFO algorithm, r is setjThe method is changed into the following form:
Figure FDA0003772731880000037
wherein r is increased with tjThe optimization algorithm is gradually reduced, and the local optimization capability in the early stage of algorithm iteration is enhanced;
in the process of searching for food by the muscle turning bucket, the muscle turning bucket factor S is a certain value, and in the later iteration stage, the algorithm needs smaller search step length, and an overlarge value can weaken the search capability of the algorithm, and similarly, the S can be changed into the following form:
Figure FDA0003772731880000038
in the formula, SmaxAnd SminFor the upper and lower limits of the somersaulting factor S, the S can be gradually reduced along with the increase of t, so that the local searching and optimizing capacity in the later iteration stage is enhanced.
4. The method according to claim 1, wherein the IEMD-TA-LSTM model-based method for predicting the operation situation of the active distribution network,
the situation prediction unit consists of a residual attention subunit, a spatial attention subunit, a temporal attention subunit and an LSTM prediction subunit;
the residual error attention subunit consists of a structure formed by connecting a plurality of convolution layer residual errors;
the 1 st layer is an input layer, the input phasor is exogenous time sequence data with given length, and K is the exogenous data quantity:
X＝(x1,x2,…,xK)
wherein, the specific expression of the exogenous data is as follows:
Xi＝(xi,t-d+1,xi,t-d+2,…,xi,t)T,i＝1,2…K
in the formula, d is a time step;
layers 2 and 3 are two-dimensional convolution modules, and the convolution layers do not change the dimension and length of input data:
H1＝(ωc1*X+bc1)
H2＝(ωc2*X+bc2)+X
in the formula, ωc1And ωc2Are all convolution weights; b is a mixture ofc1And bc2Are all bias terms;
the space attention subunit performs space feature extraction on the preliminary feature extraction data provided by the last unit by using a space attention mechanism to further obtain an encoded time sequence feature sequence;
take the t-th time step as an example, a single-step input feature vector containing K features
Figure FDA0003772731880000041
Attention weight vector etThe calculation formula is as follows:
Figure FDA0003772731880000042
in the formula (I), the compound is shown in the specification,
Figure FDA0003772731880000043
distributing attention weight corresponding to each input feature at the time t; v. ofe、weAnd beA weight matrix and a bias vector that are attention weights; sigma represents a sigmoid activation function;
by pair etNormalizing by using a Softmax function to obtain a kth characteristic attention score
Figure FDA0003772731880000044
Inputting a feature vector XtRecalculated as a weighted vector
Figure FDA0003772731880000045
Figure FDA0003772731880000046
Figure FDA0003772731880000047
The time attention subunit extracts time sequence data from the time sequence characteristics of the time sequence data by using a time attention mechanism, and decodes the coding information provided by the time attention subunit for final prediction;
xtobtaining the hidden layer state h of the LSTM after the characteristic attention module codingt：
Figure FDA0003772731880000051
In the formula (f)LSTMRepresents an LSTM cell;
through time attention module pair htPerforming decoding operation to give different weight scores to hidden state output by LSTM unit
Figure FDA0003772731880000052
And weighting with the hidden layer state at the corresponding historical moment to obtain the comprehensive time sequence information state rtThe calculation process is shown as the following formula:
Figure FDA0003772731880000053
Figure FDA0003772731880000054
Figure FDA0003772731880000055
in the formula (I), the compound is shown in the specification,
Figure FDA0003772731880000056
distributing time attention weights corresponding to various historical moments at the time t, wherein tau is the length of an input sequence time window; v. ofd、wdAnd bdA weight matrix and a bias vector that are attention weights;
finally, obtaining a prediction result y of the future n steps through an output layerτ+n：
yτ+n＝σ(wrrt+br)
In the formula, wrAnd brRespectively a weight matrix and a bias vector of the full connection layer;
finally, obtaining a prediction result y of the future n steps through an output layerτ+n：
yτ+n＝σ(wrrt+br)
In the formula, wrAnd brRespectively a weight matrix and an offset vector of the full connection layer;
the LSTM prediction subunit utilizes complete feature extraction information provided by the multiple attention mechanism units to perform high-precision time sequence prediction on the wind power output and the load power; the concrete structure is as follows:
forgetting gate FG controls last time unit state ct-1The information to be saved is saved in the current unit state ctIn (1), the input gate IG controls the input x at the current timetHow much information is saved to the current cell state ctThe output gate OG controls the current cell state ctThe information amount in the previous step is saved to the current output state htPerforming the following steps; updating of LSTM algorithmsAnd calculated as follows:
ft＝σ(wf·[ht-1,xt]+bf)
it＝σ(wi·[ht-1,xt]+bi)
ot＝σ(wo·[ht-1,xt]+bo)
Figure FDA0003772731880000057
Figure FDA0003772731880000058
Figure FDA0003772731880000061
in the formula, ft、it、ot、ctAnd htAre respectively FG、IG、OGA state matrix of a memory cell and an output cell; sigma represents a sigmoid activation function;
Figure FDA0003772731880000062
is a candidate value vector of the current memory cell state; w is af、wi、woAnd wcAre respectively FG、IG、OGAnd a weight matrix of the memory cells; bf、bi、boAnd bcAre respectively FG、IG、OGAnd biasing of the memory cell; tanh is the activation function;
Figure FDA0003772731880000063
representing the hadamard product.
5. The method for predicting the operating situation of the active power distribution network based on the IEMD-TA-LSTM model according to claim 1, wherein the output unit includes situation evaluation indexes of node voltage out-of-limit margin, branch load severity, and voltage/current fluctuation coefficient, and the operating situation of the active power distribution network is quantitatively analyzed through calculation results of the indexes;
aiming at the problem of voltage out-of-limit possibly caused by node voltage change, a node voltage out-of-limit margin is adopted
Figure FDA0003772731880000064
The future change trend of the voltage of each node of the power distribution network is reflected:
Figure FDA0003772731880000065
in the formula (I), the compound is shown in the specification,
Figure FDA0003772731880000066
the voltage per unit value of the node i to be predicted at the future time t is obtained;
Figure FDA0003772731880000067
the mean value of the upper limit and the lower limit of the ith node voltage is obtained;
aiming at the overload safety problem of branch current, the load severity of the branch is adopted
Figure FDA0003772731880000068
Reflecting the future change trend of each branch:
Figure FDA0003772731880000069
Figure FDA00037727318800000610
in the formula, LRl,tThe load factor of the first line at the future time t;
Figure FDA00037727318800000611
the current of the first branch to be predicted at the future time t; i isl,NRated current of the first line;
in order to quantify the fluctuation degree of node voltage and branch current and simultaneously evaluate the strength of the influence degree of the node voltage/branch current in the active power distribution network by new energy and load change, a node voltage fluctuation coefficient is provided
Figure FDA00037727318800000612
And branch current fluctuation coefficient
Figure FDA00037727318800000613
Indexes are respectively expressed as follows:
Figure FDA00037727318800000614
Figure FDA0003772731880000071
in the formula (I), the compound is shown in the specification,
Figure FDA0003772731880000072
and
Figure FDA0003772731880000073
the voltage value of a node i to be predicted at the future time t and the current value of a branch l to be predicted at the future time t are obtained;
Figure FDA0003772731880000074
and
Figure FDA0003772731880000075
for the predicted voltage average value of the node i to be predicted in the future D period and the predicted current level of the branch l to be predicted in the future D periodAnd (4) average value.
CN202210906679.5A 2022-07-29 2022-07-29 Active power distribution network operation situation prediction method based on IEMD-TA-LSTM model Pending CN115275991A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210906679.5A CN115275991A (en) 2022-07-29 2022-07-29 Active power distribution network operation situation prediction method based on IEMD-TA-LSTM model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210906679.5A CN115275991A (en) 2022-07-29 2022-07-29 Active power distribution network operation situation prediction method based on IEMD-TA-LSTM model

Publications (1)

Publication Number Publication Date
CN115275991A true CN115275991A (en) 2022-11-01

Family

ID=83771833

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210906679.5A Pending CN115275991A (en) 2022-07-29 2022-07-29 Active power distribution network operation situation prediction method based on IEMD-TA-LSTM model

Country Status (1)

Country Link
CN (1) CN115275991A (en)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116304846A (en) * 2023-05-24 2023-06-23 国网江西省电力有限公司电力科学研究院 CVT internal insulation abnormality online assessment method based on self-supervision learning
CN116703249A (en) * 2023-08-07 2023-09-05 南京师范大学 Reliability analysis method based on CKL wind power capacity prediction
CN116881704A (en) * 2023-09-06 2023-10-13 北京新亚盛创电气技术有限公司 Early warning method and system for power grid running state
CN117079736A (en) * 2023-10-17 2023-11-17 河北金锁安防工程股份有限公司 Gas concentration prediction method and system for intelligent gas sensing
CN118520304A (en) * 2024-07-19 2024-08-20 国网山西省电力公司营销服务中心 Deep learning multilayer active power distribution network situation awareness and assessment method and system

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116304846A (en) * 2023-05-24 2023-06-23 国网江西省电力有限公司电力科学研究院 CVT internal insulation abnormality online assessment method based on self-supervision learning
CN116304846B (en) * 2023-05-24 2023-09-12 国网江西省电力有限公司电力科学研究院 CVT internal insulation abnormality online assessment method based on self-supervision learning
CN116703249A (en) * 2023-08-07 2023-09-05 南京师范大学 Reliability analysis method based on CKL wind power capacity prediction
CN116703249B (en) * 2023-08-07 2024-01-19 南京师范大学 Reliability analysis method based on CKL wind power capacity prediction
CN116881704A (en) * 2023-09-06 2023-10-13 北京新亚盛创电气技术有限公司 Early warning method and system for power grid running state
CN116881704B (en) * 2023-09-06 2023-11-14 北京新亚盛创电气技术有限公司 Early warning method and system for power grid running state
CN117079736A (en) * 2023-10-17 2023-11-17 河北金锁安防工程股份有限公司 Gas concentration prediction method and system for intelligent gas sensing
CN117079736B (en) * 2023-10-17 2024-02-06 河北金锁安防工程股份有限公司 Gas concentration prediction method and system for intelligent gas sensing
CN118520304A (en) * 2024-07-19 2024-08-20 国网山西省电力公司营销服务中心 Deep learning multilayer active power distribution network situation awareness and assessment method and system

Similar Documents

Publication Publication Date Title
CN115275991A (en) Active power distribution network operation situation prediction method based on IEMD-TA-LSTM model
Lipu et al. Artificial intelligence based hybrid forecasting approaches for wind power generation: Progress, challenges and prospects
CN108304623B (en) Probability load flow online calculation method based on stack noise reduction automatic encoder
CN112348271A (en) Short-term photovoltaic power prediction method based on VMD-IPSO-GRU
CN109615146B (en) Ultra-short-term wind power prediction method based on deep learning
CN110942205B (en) Short-term photovoltaic power generation power prediction method based on HIMVO-SVM
Lu et al. A hybrid approach for day-ahead forecast of PV power generation
Zhang et al. Wind speed prediction research considering wind speed ramp and residual distribution
CN112149883A (en) Photovoltaic power prediction method based on FWA-BP neural network
CN114021483A (en) Ultra-short-term wind power prediction method based on time domain characteristics and XGboost
Chen et al. Research on wind power prediction method based on convolutional neural network and genetic algorithm
CN111091141B (en) Photovoltaic backboard fault diagnosis method based on layered Softmax
Xing et al. Comparative study of reformed neural network based short‐term wind power forecasting models
Wang et al. Accurate solar cell modeling via genetic Neural network-based Meta-Heuristic algorithms
CN114742298A (en) Ultra-short-term power prediction method for wind power generation
Munir et al. Artificial neural network based simplified one day ahead forecasting of solar photovoltaic power generation
Xu et al. Short-term electricity consumption forecasting method for residential users based on cluster classification and backpropagation neural network
CN108694475A (en) Short-term time scale photovoltaic cell capable of generating power amount prediction technique based on mixed model
CN108038518A (en) A kind of photovoltaic generation power based on meteorological data determines method and system
CN115456286A (en) Short-term photovoltaic power prediction method
CN115796327A (en) Wind power interval prediction method based on VMD (vertical vector decomposition) and IWOA-F-GRU (empirical mode decomposition) -based models
Zhao et al. Short-term photovoltaic power prediction based on DE-GWO-LSTM
Zhang et al. Wind farm wind power prediction method based on CEEMDAN and DE optimized DNN neural network
Wu et al. Short-term PV Power Prediction based on VMD-SE-TCAN Model
Tian et al. Wind Speed Prediction based on Long Short-Term Memory network and Policy Suggestions for Wind Power Industry Development

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