An active power distribution network operation situation prediction method based on an IEMD-TA-LSTM model

Abstract

The application discloses a kind of active distribution network operation situation prediction methods based on IEMD-TA-LSTM model, including input unit, data processing unit, situation prediction unit and output unit;Input unit constructs the input dataset of combination model;Data processing unit decomposes the original time series data of distributed power output and load power into several characteristic different time series components, and optimizes network hyperparameters;Situation prediction unit uses TA-LSTM neural network of triple attention mechanism of fusion residual convolution attention, spatial attention, time attention, inducts the space-time characteristics of exogenous data and time series information;The output unit merges the prediction results of each component, and then establishes situation evaluation index from node voltage, branch load angle, obtains the final situation prediction result.The application can optimize the distribution network operation situation prediction level, improve the precision of prediction.

Description

TECHNICAL FIELD

[0001] The application relates to an active power distribution network operation situation prediction method based on an IEMD-TA-LSTM model in the field of power distribution network operation prediction monitoring. BACKGROUND

[0002] China has vast territory and rich wind and light energy resources, so the renewable energy system has a high application prospect in China. However, the instability, volatility and counter-regulation of new energy power generation itself will bring a great impact on the power system, hinder the safe and stable operation of the power grid, and further hinder the further development of the active power distribution network. However, if the uncertainty of new energy output and load can be grasped, the active power distribution network operation situation prediction can be realized, and the above problems can be alleviated to a certain extent. Therefore, an active power distribution network operation situation prediction method with high precision is urgently needed to further improve the safe and reliable operation level of the active power distribution network.

[0003] After searching the existing literature, it was found that in the existing literature, Fan Panpan, Yuan Yiping, Sun Wenlei, etc. published in Computer Integrated Manufacturing Systems (2021, 27(7): 1993-2004) Fusion of multi-period SCADA data for wind turbine risk situation prediction Use SCADA data for LSTM neural network-based situation prediction, and quantify risk state severity through active power prediction residual. In the risk assessment process, the fuzzy C-means algorithm is used to construct the unit risk state severity outlier model to divide the state model. Zhang Qun, Tang Zhenhao, Wang Gong, etc. published in Solar Energy (2021, 42(10): 275-281) Short-term wind power prediction model based on long and short-term memory network According to the results of chaotic analysis, the wind field data is reconstructed, and the input required by the prediction model is obtained by combining the classification prediction tree, and the long and short-term memory artificial neural network method is used to establish the wind power prediction model, and finally the error correction strategy based on historical prediction error is used to further improve the model accuracy. Wang Yuhong, Shi Yunxiang, Zhou Xu, etc. published in High Voltage Technology (2022, 48(5): 1884-1892) BiLSTM multi-wind turbine short-term power prediction based on time pattern attention mechanism First, use ensemble empirical mode decomposition to obtain different modal components of the original wind turbine power signal, and based on the TPA mechanism, extract the complex relationship between multiple wind turbines from the hidden row vector obtained by the BiLSTM network, and finally apply this model to the regional wind turbine power prediction of scattered distribution. Yang Yuqing, Zhang Yi published in Control Engineering (2022, 29(1): 10-17) Short-term wind power prediction based on mRMR and VMD-AM-LSTM Use the variational mode decomposition algorithm to decompose the wind power sequence into several components with different central frequencies, then combine the weather features selected by the maximum correlation-minimum redundancy to establish an attention mechanism and a long and short-term memory hybrid prediction model, and finally superimpose the prediction results of each component to obtain the final wind power. Zhao Dongmei, Du Gang, Liu Xin, etc. published in Modern Electric Power (2022, 39(1): 9-19) Wind power combination prediction model based on time series decomposition and machine learning Respectively construct an empirical mode decomposition technology and long and short-term memory neural network combined prediction model and a variational mode decomposition technology, simulated annealing algorithm and deep belief network combined prediction model as the basic prediction model, and use the extreme gradient boosting algorithm to process the results obtained by the basic model to obtain the final prediction result. Zhang Shuqing, Yang Zhenning, Jiang Anqi, etc. published in Solar Energy (2022, 43(6): 204-211) Short-term wind power prediction based on EN-SKPCA dimensionality reduction and FPA optimized LSTMNN Use the elastic net sparse kernel principal component analysis reduction method to reduce the dimensionality of the meteorological data, and use the flower pollination algorithm to optimize the long and short-term memory neural network prediction model for wind power prediction.The above documents rarely consider the application of wind direction, wind speed, meteorological and other exogenous data in wind power, and some methods ignore the change in characteristics after decomposition of random time series information. How to improve the accuracy and reliability of grid operation situation prediction monitoring is the main goal of technical personnel. SUMMARY

[0004] The purpose of the present application is to overcome the shortcomings of the prior art, provide an active distribution network operation situation prediction method based on IEMD-TA-LSTM model, which can optimize the distribution network operation situation prediction level and improve the prediction accuracy.

[0005] A technical solution to achieve the above-mentioned purpose is: an active distribution network operation situation prediction method based on IEMD-TA-LSTM model, characterized by comprising an input unit, a data processing unit, a situation prediction unit and an output unit;

[0006] The input unit acquires distributed energy output time series data, internal element data of distribution network load data and meteorological data, and exogenous time series data of power market data, to jointly construct the input data set of the combined model;

[0007] The data processing unit uses the completely self-adaptive noise ensemble empirical mode decomposition IEMD algorithm to decompose the distributed power output and load power original time series data into several characteristic different time series components according to the data provided by the input unit, and introduces the improved manta ray optimization IMRFO algorithm to optimize the network hyperparameters;

[0008] The situation prediction unit uses the TA-LSTM neural network based on the time series data and network hyperparameters output by the data processing unit, which fuses residual convolution attention, spatial attention and temporal attention triple attention mechanism, to induce the spatiotemporal characteristics of exogenous data and time series information;

[0009] The output unit combines the prediction results of each component, and then establishes a situation evaluation index from the node voltage and branch load angle to obtain the final situation prediction result.

[0010] Further, the input unit refers to the historical and measured time series data of distributed energy output, the internal element data of historical and measured data of distribution network load, and the historical and measured exogenous time series data of meteorological data and power market data.

[0011] Further, the data processing unit is composed of an IEMD decomposition unit and an IMRFO optimization unit;

[0012] The IEMD decomposition unit uses the IEMD algorithm to decompose the wind power output and load power original time series components into several components with single and obvious characteristics, and the process is as follows:

[0013] Step 1, add Gaussian white noise to the original time series to obtain the i-th signal to be decomposed x i (t);

[0014] x i (t) = x(t) + ε0E1(ω i (t)), i = 1, 2,..., n

[0015] where ω i (t) is a zero-mean unit-variance Gaussian white noise, ε0is the noise size, E1(ω i (t)) is the first EMD component of ω i (t);

[0016] Step 2, calculate the local mean value of x i (t) by EMD decomposition to obtain the first decomposition residual r1(t) and the first modal component

[0017]

[0018] where M(〃) is the local mean operator;

[0019] Step 3, calculate the second residual r2(t) and the second modal component

[0020]

[0021] Step 4, for each stage, i.e. k = 1, 2,..., n, according to the above steps, calculate the k-th residual signal and the k-th modal component, to obtain:

[0022]

[0023] Step 5, repeat step 4 until the residual signal cannot be decomposed any more, and finally obtain n modal components and the final residual signal Re(t). The original signal sequence x(t) can be represented as:

[0024]

[0025] The IMRFO optimization unit is based on the characteristics of different time series components obtained by the IEMD decomposition unit, and uses the improved manta ray optimization algorithm to optimize the execution strategy of the neural network hyperparameters;

[0026] For the chain foraging process of manta ray, the mathematical expression of its position update is:

[0027]

[0028]

[0029] where, and are the positions of the ith individual of the tth generation and the optimal individual in d-dimensional space; r is a random number uniformly distributed in [0, 1]; and a is a weight factor;

[0030] For the spiral foraging process of manta ray, the mathematical expression of position update is:

[0031] When t / T > rand

[0032]

[0033]

[0034] where, b is a weight factor; r1 is a random number uniformly distributed in [0, 1]; and T is the maximum number of iterations;

[0035] When t / T ≤ rand

[0036]

[0037]

[0038] where, is a new position randomly generated, x u and x l are the upper and lower limits of the search space, r j is a random number uniformly distributed in [0, 1];

[0039] For the somersault foraging strategy, the expression of position update is:

[0040]

[0041] where, S is the somersault factor; r j2 and r j3 are random numbers uniformly distributed in [0, 1];

[0042] The manta ray optimization algorithm cannot guarantee that the optimal solution is the global optimum, and the population diversity is poor. The specific improvement strategies are as follows:

[0043] Step 1, obtain the initial population based on the improved Tent mapping method;

[0044] The Tent mapping is used to set the initial position of the manta ray population, and the expression is as follows:

[0045]

[0046] Since there are short cycle and unstable cycle in the chaotic sequence generation iteration process, the random equation method is introduced to improve it, that is, when falling into unstable cycle point or small cycle within 5, the improvement is made through the following formula:

[0047]

[0048] In the formula, r t is a random number uniformly distributed in [0, 1];

[0049] Step 2, adaptive nonlinear decreasing weight coefficient;

[0050] In the chain foraging optimization iteration process of MRFO algorithm, r j is changed to the following form:

[0051]

[0052] In the formula, with the increase of t, r j will gradually decrease, which enhances the local optimization ability of the algorithm in the early iteration stage;

[0053] In the somersault foraging process, the somersault factor S is a certain value, and in the later iteration stage, the algorithm needs smaller search step, and too large value will weaken the search ability of the algorithm, and S can be changed to the following form:

[0054]

[0055] In the formula, S max and S min are the upper and lower limits of the somersault factor S, and with the increase of t, S will gradually decrease, which enhances the local search and optimization ability in the later iteration stage.

[0056] Further, the situation prediction unit is composed of a residual attention subunit, a spatial attention subunit, a time attention subunit and an LSTM prediction subunit;

[0057] The residual attention subunit is composed of a structure in which a plurality of convolution layers are connected in series;

[0058] The first layer is an input layer, the input phasor is a given length of exogenous time series data, and K is the number of exogenous data:

[0059] X = (x1, x2, …, x K )

[0060] Wherein, the specific expression of exogenous data is:

[0061] X i = (x i,t-d+1 , xi,t-d+2 ,…,x i,t ) T ,i=1,2…K

[0062] where d is the time step;

[0063] The 2nd and 3rd layers are two-dimensional convolution modules, and the convolution layer does not change the dimension and length of the input data:

[0064] H1=(ω c1 *X+b c1 )

[0065] H2=(ω c2 *X+b c2 )+X

[0066] where ω c1 and ω c2 are convolution weights; b c1 and b c2 are bias terms;

[0067] The spatial attention subunit uses a spatial attention mechanism to perform spatial feature extraction on the preliminary feature extraction data provided by the previous unit, and further obtains an encoded time sequence feature sequence;

[0068] Taking the tth time step as an example, the single-step input feature vector containing K features is The attention weight vector e t is calculated as follows:

[0069]

[0070] where is the attention weight distribution corresponding to each input feature at time t; v e , w e and b e are the weight matrix and bias vector of the attention weight; and σ represents a sigmoid activation function.

[0071] By normalizing e t using a Softmax function, the kth feature attention score is obtained The input feature vector X t is recalculated as a weighted vector

[0072]

[0073]

[0074] The time attention subunit extracts time series features of the time series data by using a time attention mechanism, decodes the encoded information provided by the spatial attention subunit, and is used for final prediction;

[0075] x t After encoding by the feature attention module, the hidden layer state h of the LSTM is obtained t :

[0076]

[0077] In the formula, f LSTM represents the LSTM unit;

[0078] The h t is decoded by the time attention module, and different weight scores are given to the hidden states output by the LSTM unit and weighted with the corresponding historical time hidden layer state to obtain the comprehensive time sequence information state r t The calculation process is shown in the following formula:

[0079]

[0080]

[0081]

[0082] In the formula, is the time attention weight distribution corresponding to each historical time at t time, and τ is the input sequence time window length; v d , w d and b d are the weight matrix and bias vector of the attention weight;

[0083] Finally, the prediction result y τ+n of the future n steps is obtained through the output layer:

[0084] y τ+n = σ (w r r t +b r )

[0085] In the formula, w r and b r are the weight matrix and bias vector of the full connection layer, respectively;

[0086] Finally, the prediction result y τ+n of the future n steps is obtained through the output layer:

[0087] y τ+n = σ (w r r t +b r)

[0088] where w r and b r are the weight matrix and bias vector of the full connection layer, respectively;

[0089] The LSTM prediction subunit uses the complete feature extraction information provided by the above multiple attention mechanism unit to perform high-precision time series prediction of wind power output and load power; the specific structure is as follows:

[0090] The forget gate FG controls the information in the previous time unit state c t-1 that needs to be saved and saves it to the current unit state c t , the input gate IG controls how much information in the current time input x t is saved to the current unit state c t , and the output gate OG controls the amount of information in the current unit state c t , which is saved to the current output state h t ; the update and calculation of the LSTM algorithm are as follows:

[0091] f t =σ(w f ·[h t-1 ,x t ]+b f )

[0092] i t =σ(w i ·[h t-1 ,x t ]+b i )

[0093] o t =σ(w o ·[h t-1 ,x t ]+b o )

[0094]

[0095]

[0096]

[0097] where f t , i t , o t , c t and h t are F G , I G , O G, the state matrix of memory cell and output unit; σ represents sigmoid activation function; is the candidate value vector of current memory cell state; w f i o c are the weight matrix of F G , I G , O G and memory cell respectively; b f , b i , b o and b c are the bias of F G , I G , O G and memory cell respectively; tanh is activation function; represents Hadamard product.

[0098] Further, the output unit includes node voltage out-of-limit margin, branch load severity, voltage / current fluctuation coefficient, and situation evaluation index, and the operation situation of active power distribution network is quantitatively analyzed by the calculation results of these indexes;

[0099] In view of the voltage out-of-limit problem caused by node voltage change, the node voltage out-of-limit margin is used to reflect the future change trend of each node voltage of the power distribution network:

[0100]

[0101] In the formula, is the voltage per unit of the node i to be predicted at future time t; is the average value of the upper and lower limits of the voltage of the i th node;

[0102] In view of the overload safety problem of branch current, the branch load severity is used to reflect the future change trend of each branch:

[0103]

[0104]

[0105] In the formula, LR l,t is the load rate of the l th line at future time t; is the current of the l th branch to be predicted at future time t; I l,N is the rated current of the l th line;

[0106] is the fluctuation degree of the node voltage and branch current, and the strength and weakness of the influence of new energy and load change on the node voltage / branch current in the active power distribution network are evaluated, and the node voltage fluctuation coefficient​​​ and branch current fluctuation coefficient indicators, the expressions of which are respectively:

[0107]

[0108]

[0109] wherein, and are the voltage value of the node i to be predicted at the future time t and the current value of the branch l to be predicted at the future time t; and are 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.

[0110] 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 take the different time sequence characteristics after random data decomposition as the target, introduce intelligent optimization algorithm and multiple attention mechanism to optimize the model parameters, and improve the prediction accuracy of the combined model. BRIEF DESCRIPTION OF DRAWINGS

[0111] Figure 1 is a flowchart of the active power distribution network operation situation prediction method based on the IEMD-TA-LSTM model of the present application;

[0112] Figure 2 is a DA-LSTM model structure diagram;

[0113] Figure 3 is a residual attention mechanism diagram;

[0114] Figure 4 is an LSTM unit structure diagram;

[0115] Figure 5 is an improved IEEE-33 node active power distribution network diagram. DETAILED DESCRIPTION

[0116] In order to better understand the technical solutions of the present application, the following will be described in detail through specific examples:

[0117] As Figure 1As shown, an active power distribution network operation situation IEMD-TA-LSTM combined prediction model fusing an improved complete ensemble empirical mode decomposition (IEMD), a triple-stage attention (TA), and a long short-term memory (LSTM) is provided, which comprises an input unit 1, a data processing unit 2, a situation prediction unit 3, and an output unit 4 connected in sequence.

[0118] The input unit 1 acquires distributed energy output time series data, internal element data of power distribution network load data, and external source time series data of meteorological data and power market data, and jointly constructs an input data set of the combined model.

[0119] The data processing unit 2 decomposes the original time series data of the distributed power output and the load power into several characteristic time series components according to the data provided by the input unit 1, and optimizes the network hyperparameters by introducing an improved manta ray optimization (IMRFO) algorithm.

[0120] The situation prediction unit 3 uses a TA-LSTM neural network fusing residual convolution attention, spatial attention, and temporal attention triple attention mechanisms based on the time series data and network hyperparameters output by the data processing unit 2, and fully excavates the spatiotemporal characteristics of the external source data and the time series information.

[0121] The output unit 4 combines the prediction results of each component, then establishes situation evaluation indexes from the perspectives of node voltage and branch load, and obtains the final situation prediction result.

[0122] The input unit 1 refers to the historical and measured time series data of the distributed energy output, the internal element data of the historical and measured data of the power distribution network load, and the historical and measured external source time series data of the meteorological data and the power market data.

[0123] The data processing unit 2 comprises an IEMD decomposition unit 21 and an IMRFO optimization unit 22.

[0124] The IEMD decomposition unit 21 mainly uses the IEMD algorithm, which adds a certain amount of Gaussian white noise in the decomposition iteration process of the distributed power output and the load power signal represented by wind power, and calculates the mean value of the generated signal component.

[0125] The IEMD algorithm can decompose the original time-series components of wind power output and load power into several components with single and obvious characteristics. Its simplified process is as follows:

[0126] Step 1: Add Gaussian white noise to the original time series to obtain the i-th signal to be decomposed, x. i (t);

[0127] x i (t)=x(t)+ε0E1(ω i (t)), i = 1, 2, ..., n

[0128] In the formula, ω i (t) represents zero-mean, unit-variance Gaussian white noise, ε0 represents the noise level, and E1(ω) represents the noise level. i (t) is ω i The first EMD component of (t).

[0129] Step 2, calculate x by EMD decomposition. i The local average value of (t) is used to obtain the residual r1(t) of the first decomposition and the first modal component.

[0130]

[0131] In the formula, M(〃) is the local mean operator.

[0132] Step 3: Calculate the second residual r2(t) and the second modal component.

[0133]

[0134] Step 4: For each of the remaining stages, i.e., k = 1, 2, ..., n, calculate the k-th residual signal and the k-th modal component according to the above steps, and obtain:

[0135]

[0136] Step 5: Repeat step 4 until the residual signal can no longer be decomposed, ultimately obtaining n modal components and the final residual signal Re(t). The original signal sequence x(t) can be represented as:

[0137]

[0138] The IMRFO optimization unit 22 optimizes the neural network hyperparameters according to the characteristics of the different time sequence components obtained by the IEMD decomposition unit 21, and uses an improved manta ray foraging optimization (IMRFO) algorithm to perform optimization on the execution strategy of the neural network hyperparameters, avoiding manual adjustment of the neural network hyperparameters corresponding to each component each time.

[0139] The manta ray foraging optimization (MRFO) algorithm is a population optimization algorithm proposed by observing the foraging behavior of manta rays. The foraging strategies of manta ray populations include chain foraging, cyclone foraging, and somersault foraging. Through different foraging strategies, the search space is explored and developed to update the position of individuals.

[0140] For the chain foraging process of manta rays, the mathematical expression for updating the position is:

[0141]

[0142]

[0143] wherein, and are the positions of the ith individual and the optimal individual in the d-dimensional space in the tth generation; r is a random number uniformly distributed in [0, 1]; and a is a weight factor.

[0144] For the spiral foraging process of manta rays, the mathematical expression for updating the position is:

[0145] When t / T > rand

[0146]

[0147]

[0148] wherein, b is a weight factor; r1 is a random number uniformly distributed in [0, 1]; and T is the maximum number of iterations.

[0149] When t / T ≤ rand

[0150]

[0151]

[0152] wherein, is a newly generated random position, x u and x l are the upper and lower limits of the search space, r j is a random number uniformly distributed in [0, 1].

[0153] For the somersault foraging strategy, the position update expression is:

[0154]

[0155] In the formula, S is the somersault factor; r j2 and r j3 are random numbers uniformly distributed in [0, 1].

[0156] The algorithm requires fewer iterations and is efficient, but in practical applications, it is also found that it cannot guarantee that the optimal solution is globally optimal, and the population diversity is poor. Therefore, aiming at the problem that the TA-LSTM network has its own hyperparameters difficult to determine when training and learning, an improved MRFO algorithm is proposed. The improvement strategy is as follows:

[0157] Step 1, obtain the initial population based on the improved Tent mapping method;

[0158] The initial population produced at random will affect the speed and accuracy of the algorithm. The randomness, regularity and ergodicity of chaotic sequences make the initial solution generated by chaotic sequences complex and diverse, so that the initial solution generated by chaotic sequences is easy to converge. Compared with other mappings, Tent mapping can generate more uniformly distributed sequences, so this patent uses Tent mapping to set the initial position of the manta ray population, and its expression is as follows:

[0159]

[0160] Because there are short cycles and unstable cycles in the chaotic sequence generation iteration process, a random equation method is introduced to improve it, that is, when falls into an unstable cycle point or a small cycle within 5, it is improved by the following formula:

[0161]

[0162] In the formula, r t is a random number uniformly distributed in [0, 1].

[0163] Step 2, adaptive nonlinear decreasing weight coefficient;

[0164] In the chain foraging optimization iteration process of the MRFO algorithm, r j is randomly generated in [0, 1], which makes the search step in the later period the same as in the early period, thereby affecting the speed of force search, so r j can be changed to the following form:

[0165]

[0166] Wherein, with the increase of t, r j will gradually decrease, enhancing the local optimization ability of the algorithm in the early iteration.

[0167] In the somersault foraging process, the somersault factor S is a certain value. In the later iteration, the algorithm needs a smaller search step, and a too large value will weaken the search ability of the algorithm. Similarly, S can be changed to the following form:

[0168]

[0169] Wherein, S max and S min are the upper and lower limits of the somersault factor S, which will gradually decrease with the increase of t, enhancing the local search and optimization ability in the later iteration.

[0170] The situation prediction unit 3 is composed of a residual attention subunit 31, a spatial attention subunit 32, a time attention subunit 33 and an LSTM prediction subunit 34, and the network structure is as shown in Figure 2 The residual attention mechanism strengthens the influence of exogenous data on the extraction of internal element data features; the spatial attention mechanism is used to calculate the contribution of input features to target features; the time attention mechanism is used to calculate the contribution of historical time to the prediction of the current time, and the prediction accuracy of the prediction model is improved through differentiated dynamic weight allocation.

[0171] The residual attention subunit 31 is composed of a structure in which a plurality of convolution layers are connected in series, which extracts features from the input exogenous data, and the residual block structure makes the feature extraction process not lose too much information. The network structure is as shown in Figure 3 .

[0172] The first layer is an input layer, and the input phasor is a given length of exogenous time series data, and K is the number of exogenous data:

[0173] X = (x1, x2, …, x K )

[0174] Wherein, the specific expression of the exogenous data is:

[0175] X i = (x i,t-d+1 ,x i,t-d+2 ,…,x i,t ) T ,i = 1, 2…K

[0176] Wherein, d is the time step.

[0177] The second and third layers are two-dimensional convolution modules, and the convolution layer does not change the dimension and length of the input data:

[0178] H1 = (ω c1 *X + b c1 )

[0179] H2 = (ω c2 *X + b c2 )+X

[0180] where ω c1 and ω c2 are convolution weights; b c1 and b c2 are bias terms.

[0181] Attention can mine the correlation between the feature sequence and the target feature, and through its weight probability distribution mechanism, it dynamically allocates weights to different influence degree features. Taking the t-th time step as an example, the single-step input feature vector containing K features is The attention weight vector e t is calculated as follows:

[0182]

[0183] where e is the attention weight distribution corresponding to each input feature at time t; v e , w e and b e are the weight matrix and bias vector of the attention weight; and σ represents the sigmoid activation function.

[0184] By normalizing e t with the Softmax function, the k-th feature attention score is obtained The input feature vector X t is recalculated as the weighted vector

[0185]

[0186]

[0187] The time attention subunit 33 uses the time attention mechanism to extract time series features from the time series data, and decodes the encoding information provided by the spatial attention subunit for the final prediction.

[0188] The time attention mechanism assigns attention weights to the time series information carried by each historical time of the input sequence to distinguish its influence on the prediction output at the current time. t After encoding by the feature attention module, the hidden layer state h t of the LSTM is obtained:

[0189]

[0190] where f LSTM represents the LSTM unit.

[0191] The h t is processed by the time attention module, and the hidden state output by the LSTM unit is given different weight scores and weighted with the corresponding historical time hidden layer state to obtain the comprehensive time sequence information state r t The calculation process is as shown in the following formula:

[0192]

[0193]

[0194]

[0195] wherein, is the time attention weight distribution of each historical time at t time, and τ is the input sequence time window length; v d , w d and b d are the weight matrix and bias vector of the attention weight.

[0196] Finally, the prediction result y τ+n of the future n steps is obtained through the output layer:

[0197] y τ+n = σ(w r r t +b r )

[0198] wherein, w r and b r are the weight matrix and bias vector of the full connection layer.

[0199] The LSTM prediction sub-unit 34 uses the complete feature extraction information provided by the above multiple attention mechanism unit to perform high-precision time sequence prediction of wind power output and load power.

[0200] The LSTM network adds a forget gate (FG), an input gate (IG), and an output gate (OG) on the basis of the recurrent neural network, and can introduce a "memory function" module for storing data state in the structure, effectively solving the problem that the RNN cannot construct a long time sequence prediction model due to the gradient explosion and disappearance. The specific unit structure of the LSTM is as shown in Figure 4 .

[0201] The FG controls the unit state c t-1The information in the current cell state c is saved t The IG controls the input x at the current time t The amount of information saved in the current cell state c t The OG controls the current cell state c t The amount of information in the current output state h is saved t The LSTM algorithm is updated and calculated as follows:

[0202] f t = σ(w f · [h t-1 , x t ] + b f )

[0203] i t = σ(w i · [h t-1 , x t ] + b i )

[0204] o t = σ(w o · [h t-1 , x t ] + b o )

[0205]

[0206]

[0207]

[0208] where f t , i t , o t , c t and h t are the state matrices of the F G , I G , O G , memory cell and output cell, respectively; σ represents the sigmoid activation function; is the candidate value vector of the current memory cell state; w f , w i , w o and w c are the weight matrices of the F G , I G , O G and memory cell, respectively; b f , b i , b o and b c are the bias vectors of the F GI G O G and the bias of the memory cell; tanh is the activation function; denotes the Hadamard product.

[0209] The output unit 4 mainly includes node voltage overrun margin, branch load severity, voltage / current fluctuation coefficient and other state evaluation indexes, and the operation state of the active distribution network is quantitatively analyzed through the calculation results of these indexes.

[0210] In view of the voltage overrun problem caused by the change of node voltage, the node voltage overrun margin is used to reflect the future change trend of the voltage of each node in the distribution network:

[0211]

[0212] In the formula, is the voltage per unit value of the node i to be predicted at the future time t; is the average value of the upper and lower limits of the voltage of the i th node.

[0213] In view of the overload safety problem of branch current, the branch load severity is used to reflect the future change trend of each branch:

[0214]

[0215]

[0216] In the formula, LR l,t is the load rate of the l th line at the future time t; is the current of the l th branch to be predicted at the future time t; I l,N is the rated current of the l th line.

[0217] The voltage fluctuation coefficient and the current fluctuation coefficient are indexes for quantifying the fluctuation degree of node voltage and branch current, and evaluating the strength of the influence of new energy and load change on the node voltage / branch current in the active distribution network. The expressions are respectively:

[0218]

[0219]

[0220] In the formula, and are the voltage value of the node i to be predicted at the future time t and the current value of the branch l to be predicted at the future time t; and Let be the average predicted voltage of node i in the future D time period and the average predicted current of branch l in the future D time period.

[0221] Example

[0222] This example uses an improved IEEE-33 node active distribution network as the research object, and its topology is as follows: Figure 5 As shown, the system comprises 33 nodes and 32 branches, with wind turbines connected at nodes 16, 20, and 31, with capacities of 500kW, 500kW, and 1000kW, respectively. Load power data and environmental data for a certain region from January 1, 2017 to December 31, 2017 were selected as historical data, combined with measured historical data of wind turbines from actual wind farms in the region during the same period. The sampling interval was 15 minutes, and all data had undergone anomaly processing. The proposed IEMD-TA-LSTM algorithm was used to predict the load and wind power output trends of each node in the system over the next 24 hours. The distribution network operation status was obtained through active distribution network power flow calculation, and active distribution network status assessment indicators were output.

[0223] The IEMD algorithm adds Gaussian white noise with a standard deviation of 0.2 and 200 additions. Due to the large number of input samples, and limited by space, only the IEMD decomposition results for one week's wind power output data (rated power 1000kW) at node 16 and load data (rated power 200kW) at node 7 are presented. Decomposing the original data using the IEMD algorithm can reduce the non-stationarity and irregularity of the original data, which is beneficial to improving the accuracy of distribution network operation factor prediction.

[0224] IMRFO algorithm settings: population size 10, maximum number of iterations 20. The load power data decomposed from node 7 is then used as a sequence ( The load (and Re) and corresponding daily environmental data are used as inputs to the TA-LSTM prediction model. The first 70% of the historical load data is used as the training set, and the last 30% is used as the test set. The load for the next 24 hours, starting from 24:00 on November 27, 2017, is predicted, and the load change prediction results for 0:00-24:00 on the 28th are obtained. Similarly, taking the wind power output at node 16 as an example, the same method is used to predict the wind power change trend for the next 0:00-24:00.

[0225] To verify the predictive performance of the IEMD-TA-LSTM model proposed in this patent, the root mean square error (RMSE), mean absolute error (MAE), and goodness-of-fit coefficient (R-Square) were used. 2The indicators evaluate the prediction results. The prediction results of the load and the wind power of the LSTM, TA-LSTM, IMRFO-TA-LSTM and the proposed IEMD-TA-LSTM prediction models are compared and analyzed, and the calculation results of the performance indicators of each model are shown in Tables 1 and 2, respectively.

[0226] Table 1 Comparison of evaluation indicators of load prediction results

[0227] Prediction model RMSE / kW MAE / kW [R 2 ]]> 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

[0228] Table 2 Comparison of evaluation indicators of wind power prediction results

[0229]

[0230]

[0231] It can be seen from Table 1 that for the prediction of the load, the pure LSTM model has poor prediction effect. The IEMD-TA-LSTM prediction model reduces the RMSE by 60.23%, 62.16% and 73.35% respectively, the MAE by 55.47%, 59.08% and 73.68% respectively, and the R 2 value by 18.10%, 20.84% and 60.45% respectively compared with the IMRFO-TA-LSTM, TA-LSTM and LSTM, which shows that the proposed model has excellent prediction performance. Compared with the TA-LSTM and the single LSTM, the RMSE and MAE indicators of the prediction results of the TA-LSTM model are reduced by 29.55% and 35.20% respectively, and the R 2 value is increased by 32.78%, which shows that by introducing the residual convolution processing, the spatial and temporal multiple attention mechanism, the feature information and the time sequence information and their correlation can be effectively mined, and the prediction performance of the LSTM is improved; the IMRFO-TA-LSTM performs better than the TA-LSTM, which shows that the IMRFO algorithm can find the optimal combination of hyperparameters, so that the prediction performance of the model is improved; the performance of the IEMD-TA-LSTM model proposed in the patent is better than that of the IMRFO-TA-LSTM algorithm, which shows that by sequence decomposition, the decomposition sequence has better regularity, and the prediction accuracy of the algorithm is effectively improved.

[0232] In order to obtain the short-term operation situation of the active distribution network, first of all, the prediction results of all load nodes and wind power are obtained according to the proposed IEMD-TA-LSTM model. After obtaining the change trend of the load and the wind power of the IEEE33 node system in the future 24 hours, through the multi-time section power flow calculation and analysis, the node voltage out-of-limit margin indicators of each node and the branch load severity indicators of each branch in the system in the future 24 hours are obtained.

[0233] It can be seen that the node voltage out-of-limit margin index based on the predicted values of load and wind power is in the range of [0.95, 1], and there is no node voltage out-of-limit situation, and each node voltage is in a normal state.

[0234] The branch load severity index based on the predicted values of load and wind power is in the normal operation range of [0, 0.7], but it can also be seen that the branch load severity index of branches 22 and 23 is close to 0.7 for a long time, and there is a potential risk of heavy load operation. In order to evaluate the accuracy of the operation trend prediction results of the node voltage out-of-limit margin index and the branch load severity index, the MAE index is used for quantitative analysis. Based on the true values and predicted values of load and wind power, the MAE of the node voltage out-of-limit margin index and the branch load severity index of each node at each time section is obtained. The MAE index of the node voltage out-of-limit margin index of each node is less than 1.3x10 -3 p.u., and the MAE index of the voltage phase angle of each node is less than 1.5x10 -2 p.u., and the prediction result is relatively accurate. In addition, for the fluctuation coefficient index of each node voltage and each branch current, it can be seen that the fluctuation degree index based on the true values of load and wind power and the fluctuation degree index based on the predicted values of load and wind power are basically consistent, and the error is very small. Therefore, the IEMD-TA-LSTM combined prediction model can accurately track the operation trend of the node voltage out-of-limit margin and the branch load severity index of each node and each branch in the whole network.

[0235] Those skilled in the art of the technical field should realize that the above embodiments are only used to illustrate the present application, and are not used as a limitation of the present application, as long as the changes and modifications of the above described embodiments are within the scope of the spirit of the present application.

Claims

1. An active power distribution network operation situation prediction method based on an IEMD-TA-LSTM model, characterized in that, The input unit, the data processing unit, the situation prediction unit and the output unit are included. The input unit acquires distributed energy output time series data, internal element data of distribution network load data and meteorological data, and external source time series data of power market data, and jointly constructs an input data set of a combination model. The data processing unit decomposes the distributed power output and the load power original time series data into several characteristic time series components by using the completely self-adaptive noise ensemble empirical mode decomposition (IEMD) algorithm according to the data provided by the input unit, and introduces the improved manta ray optimization (IMRFO) algorithm to optimize the network hyperparameters. The situation prediction unit uses a TA-LSTM neural network that fuses residual convolution attention, spatial attention and temporal attention triple attention mechanisms to induce the spatio-temporal characteristics of external data and time series information based on the time series data and network hyperparameters output by the data processing unit. The output unit combines the prediction results of each component, and then establishes a situation evaluation index from the node voltage and branch load angle to obtain the final situation prediction result.

2. The active power distribution network operation situation prediction method based on the IEMD-TA-LSTM model according to claim 1, characterized in that, The input unit refers to the historical and measured time series data of distributed energy output, the internal element data of historical and measured data of distribution network load, and the historical and measured external source time series data of meteorological data and power market data.

3. The active distribution network operation situation prediction method based on the IEMD-TA-LSTM model according to claim 1, characterized in that, 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 attention subunit is composed of a structure in which a plurality of convolution layers are connected in series. The first layer is an input layer, and the input phasor is a given length of external source time series data, and K is the number of external source data: X = (x1, x2,..., x K ) Wherein, d is the time step. X i = (x i,t-d+1 , x i,t-d+2 , …, x i,t ) T i = 1, 2…K The spatial attention subunit uses a spatial attention mechanism to extract spatial features from the preliminary feature extraction data provided by the previous unit, and further obtains an encoded time series feature sequence. The 2nd and 3rd layers are two-dimensional convolution modules, and the convolution layer does not change the dimension and length of the input data: H1=(ω c1 *X+b c1 ) H2 = (ω c2 *X + b c2 )+ X where ω c1 and ω c2 are convolution weights; b c1 and b c2 are bias terms; The temporal attention subunit uses a temporal attention mechanism to extract time series data of the time series features of the time series data, and decodes the encoded information provided by the spatial attention subunit for final prediction. Take the t-th time step as an example, a single-step input feature vector containing K features Attention weight vector e t The calculation formula is as follows: In the formula, is the attention weight distribution corresponding to each input feature at time t; v e , w e and b e are the weight matrix and bias vector of the attention weight; and σ represents a sigmoid activation function. By e t The kth feature attention score is obtained by normalizing processing with a Softmax function The input feature vector X t Recalculated as a weighted vector The LSTM prediction subunit uses the complete feature extraction information provided by the above multiple attention mechanism units to perform high-precision time series prediction of wind power output and load power; the specific structure is as follows: x t h after encoding by the characteristic attention module t : In the formula, f LSTM denotes an LSTM unit; h t The decoding operation is performed by the time attention module, and different weight scores are given to the hidden states of the LSTM unit output and the corresponding historical moment hidden layer state is weighted to obtain the comprehensive time sequence information state r t The calculation process is as follows: In the formula, is the time attention weight distribution corresponding to each historical time at time t, τ is the length of the input sequence time window; v d , w d and b d are the weight matrix and bias vector of the attention weight. Finally, the prediction result y of future n steps is obtained through the output layer τ+n : y τ+n = σ(w r r t + b r ) where w r and b r are the weight matrix and bias vector of the fully connected layer, respectively. Finally, the prediction result y of future n steps is obtained through the output layer τ+n : y τ+n = σ(w r r t + b r ) where w r and b r are the weight matrix and bias vector of the fully connected layer, respectively. The output unit includes situation evaluation indexes of node voltage overrun margin, branch load severity and voltage / current fluctuation coefficient, and quantitatively analyzes the active distribution network operation situation through the index calculation results. The forget gate FG controls how much of the previous cell state c is forgotten t-1 The input gate IG controls how much information to save from the input x t at the current time step into the current cell state c t The output gate OG controls how much information from the current cell state c t is allowed to be saved into the current output state h t The amount of information in the current cell state c t is updated and computed as follows: f t = σ(w f · [h t-1 , x t ]+b f ) i t = σ(w i · [h t-1 , x t ]+b i ) o t = σ(w o · [h t-1 , x t ]+b o ) where f t , i t , o t , c t , and h t are the state matrices of the F G , I G , O G , memory cells, and output cell, respectively; σ denotes the sigmoid activation function; is the candidate value vector for the current memory cell state; w f , w i , w o , and w c are the weight matrices of the F G , I G , O G , and memory cells, respectively; b f , b i , b o , and b c are the biases of the F G , I G , O G , and memory cells, respectively; tanh is the activation function; denotes the Hadamard product.

4. The active power distribution network operation situation prediction method based on the IEMD-TA-LSTM model according to claim 1, characterized in that, ​ To address the potential voltage over-limit issue caused by node voltage variations, a node voltage over-limit margin is adopted. To reflect the future voltage trends at various nodes of the distribution network: wherein is the voltage of the node i at time t; is the average of the upper and lower limits of the voltage of the i-th node. For the overload security problem of branch current, the branch load severity Future trends of each branch of the reaction: wherein LR l,t is the load rate of the lth line at future time t; is the current of the lth branch to be predicted at future time t; I l,N is the rated current of the lth line; To quantify the fluctuation degree of node voltage and branch current, and evaluate the influence degree of new energy and load change on node voltage / branch current in active power distribution network, the node voltage fluctuation coefficient and branch current fluctuation coefficient indices are proposed, whose expressions are respectively: wherein, and are the voltage value of the node i to be predicted at future time t and the current value of the branch l to be predicted at future time t; and are the predicted voltage average value of the node i to be predicted at future D period and the predicted current average value of the branch l to be predicted at future D period.

Images