CN105512808A - Power system transient stability assessment method based on big data - Google Patents

Abstract

The invention discloses a power system transient stability assessment method based on big data, and the method comprises two stages: sample training and result evaluation. The sample training stage comprises the steps: carrying out the centralized training and learning of ELMs through employing big data of a power system, and also completing the confirmation of each transient stability characteristic quantity weight parameter. The result evaluation stage comprises the steps: firstly building a transient stability key characteristic library; secondly extracting key characteristics in current real-time data, and transmitting the key characteristics to a trained EMLs; finally carrying out the comprehensive analysis of the output result of the EMLs through a decision evaluation module, and giving a transient stability assessment result and the confidence of the current result. The method takes big data as the background, is based on the reasonable dimension reduction and complete use of mass data of a power grid, is based on a learning machine cluster, and meets the requirements of complete and real-time safety early warning of an intelligent power grid.

Description

A kind of transient stability evaluation in power system method towards large data

Technical field

The invention belongs to Transient Security for Power Systems assessment technology field, relate to a kind of electric power system transient stability fast evaluation method towards large data.

Background technology

Continuous progress along with electric power information, the wide area measurement technology (WAMS) based on PMU develop rapidly, the state measurement of intelligent grid electric system to high sampling rate, on a large scale, the future development of fast continuously record and mass memory.The typical large data characteristicses such as it is large that current power system data has data volume, and data type is complicated, and data publication scope is wide, data acquisition transmission speed is fast, the development of intelligent grid makes electric system enter large data age.

The electrical energy production that electric system is made up of multi-layer networks such as generating, transmission of electricity, distribution and electricity consumptions and feed system, there is process energy large, cover the features such as region is wide, cladding element is many, dynamic process is complicated, strong nonlinearity, one of the most complicated man-made system, a transient security inherently high-dimensional nonlinear problem of electric system.

Method at present for Transient Security for Power Systems assessment has two kinds.The first utilizes off-line analysis, and the mode of manual decision is realized with experience by dispatcher, and this method is very poor to the method for operation adaptability that electrical network is changeable; Second method is the correlation technique utilizing modern control theory; extract the steady temporarily feature of current electric grid; then the transient stability of the instrument of neural network, support vector machine etc. to electrical network is utilized to judge; this method is except being easy to be absorbed in except the problems such as local minimum trap; usually also need the training grown very much and processing time, be difficult to the requirement of real-time meeting the large data of electric system.

Although the large packet of power train contains all information that can characterize current electric grid transient security, and how to find these information of these information, Appropriate application in mass data, making assessment to electrical network transient security, is when previous problem demanding prompt solution.

Summary of the invention

In order to solve the problems of the technologies described above, the present invention proposes a kind of transient stability evaluation in power system method towards large data, while solution traditional neural network algorithm is easy to be absorbed in the problems such as local minimum trap, meets the requirement of real-time of large data.

The technical solution adopted in the present invention is: a kind of transient stability evaluation in power system method towards large data, is characterized in that, comprise the following steps:

Step 1: according to practical power systems feature, determines the criterion of current power power system transient stability;

Step 2: setting training and correlation parameter initial value;

Step 3: determine transient stability foundation characteristic storehouse;

Step 4: utilize the large data of electric system history to complete sample training;

Step 5: extract transient state stable key syndrome data in current power system, and be sent in the multiple limits unit (ELMs) trained;

Step 6: carry out decision-making of comprehensive evaluation to the Output rusults of multiple limits unit (ELMs), provides the steady temporarily assessment result of current power system and the degree of confidence of current results.

As preferably, the criterion of electric power system transient stability described in step 1 to characterize the whether stable parameter value of current power system transient modelling, when parameter value crosses certain threshold value, just assert that current system can not meet transient stability requirement.

As preferably, training and correlation parameter initial value described in step 2 comprise the limit unit ELM sum E, the input feature vector number f of limit unit ELM, hidden layer node number interval [j_min, j_max], maximum frequency of training i_max.

As preferably, utilize the large data of electric system history to complete sample training described in step 4, its specific implementation comprises following sub-step:

Step 4.1: concentration training study is carried out to multiple limits unit (ELMs);

Step 4.2: step 4.2: in iterative computation foundation characteristic storehouse, each feature is to the weight coefficient of transient stability, determines key feature group.

As preferably, carry out concentration training study to multiple limits unit (ELMs) described in step 4.1, its idiographic flow is as follows:

(1) first time circulation is performed: the maximum frequency of training i_max of i=1to;

(2) second time circulation is performed: e=1to limit unit ELM sum E;

(3) training sample is determined at random;

(4) determine at random current by training limit unit ELM node in hidden layer j ∈ [j _min, j _max];

(5) in a foundation characteristic storehouse random choose f feature as current input limits unit ELM, wherein f is the input feature vector number of limit unit ELM;

(6) input weight vector wj and bj is determined at random;

(7) β=H is calculated ^-1t, wherein: H is neural network hidden layer output matrix, and T is sample matrix;

(8) judge whether second time circulation circulates complete, if then terminate second time circulation;

(9) judge whether first time circulation circulates complete, if then terminate first time circulation.

As preferably, steady characteristic quantity weight coefficient temporarily described in step 4.2, characterizes different electrical parameter to the separating capacity between sample; The feature that weight coefficient is large, little to the discrimination of sample Output rusults of the same type, namely limit unit EML output valve difference is little; The feature that weight coefficient is large, comparatively large to dissimilar sample area calibration, namely limit unit EML output valve difference is large; And the feature that weight coefficient is little, its output valve is to dissimilar sample and insensitive described sample of the same type refers to that all transient stability result of determination are for safety or unsafe all sample sets; Described dissimilar sample refers to not at other samples of sample place set.

As preferably, steady characteristic quantity weight coefficient temporarily described in step 4.2, computing method comprise the following steps:

(1) circulation is performed: the maximum frequency of training i_max of i=1to;

$\left(2\right)---W{\[X\]}^{i+1}=W{\[X\]}^i-\frac{dis(X,R_i,H)}{N}+\frac{dis(X,R_i,M)}{N};$

$\left(3\right)---dis(X,R,R^{\′})=\frac{|val(X,R)-val(X,R^{\′})|}{max\left(X\right)-\min \left(X\right)};$

(4) judge whether circulation circulates complete, if then end loop;

Wherein: X represents certain steady feature temporarily; the weight coefficient that representation feature X is final; R _iit is the sample of i-th training; H is from R in similar sample _inearest sample; M is sample nearest from Ri in inhomogeneity sample; Val (X, R) is the value of feature X in R sample; Function dis (X, R, R ') calculate feature X between sample R and R ' relatively to difference; N is this example total sample number, and its existence makes

As preferably, provide the steady temporarily assessment result of current power system and the degree of confidence of current results described in step 6, the design standard of its fiducial interval is:

If current power system is in transient stability state, result of determination=1; Current power system can not maintain transient stability state, result of determination=-1, then for the Output rusults Y of each the study unit in ELMs _e∈ [-1,1]:

If Y _e∈ [0.7,1.3] or Y _e∈ [-1.3 ,-0.7], current ELM study unit exports credible: u=u+1;

If Y _e∈ [-0.7,0.7], current ELM study unit exports insincere: v=v+1.

As preferably, provide the temporary steady assessment result of current power system and the degree of confidence of current results described in step 6, its temporary steady decision criteria is:

If u is characterized in E ELM and learns, in unit, to export the ELM number of reliable result; V characterizes insincere number of results;

If u<80%*E, then the credible output of current ELMs does not reach threshold requirement, and current results is insincere, judges unsuccessfully;

If u>80%*E, and then current system transient stability;

If u>80%*E, and then current system not transient stability;

The present invention mainly comprises sample training and two stages of evaluation of result.The sample training stage is divided into again the focusing study of ELMs and extraction two parts of electrical network transient state key feature, and the former surely judges to provide accurate and reliable data supporting for follow-up temporarily; The latter requires input variable science dimensionality reduction for the large real-time property of intelligent grid.The evaluation of result stage provides final power system transient stability safe condition.Compared with conventional method, the effect that the present invention is useful is:

(1) devise a kind of evaluating system being applicable to large data, rational dimensionality reduction is carried out to the large data of electrical network, realize assessing the safety case of electrical network comprehensively, reasonably;

(2) utilize the feature that extreme learning machine is own, adopt the method parallel training and the multiple EMLs of use concentrated, compared with traditional neural network, it has significant rapidity; Compared with single EML, there is again better stability and high precision;

(3) Decision-Making Evaluation module learns the output of unit in conjunction with each ELM, while providing precise evaluation result, also gives the degree of confidence of evaluation result from statistical angle in conjunction with reasonably judge rule.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of the embodiment of the present invention;

Fig. 2 is the sample training process flow diagram of the embodiment of the present invention;

Fig. 3 be in the embodiment of the present invention each feature by the descending sort figure of weight coefficient.

Embodiment

Below in conjunction with the accompanying drawing in example of the present invention, carry out clear, complete description to the technical scheme in example of the present invention, obviously, described embodiment is only one embodiment of the present of invention, instead of whole embodiments.Based on the embodiment in invention, all embodiments that those of ordinary skill in the art obtain under the prerequisite not making creative work, all belong to the scope of protection of the invention.

Utilize the large data of electric system history to carry out focusing study training to multiple limits unit (EXTREMELEARNINGMACHINES, ELMs), in mass data, extract the key feature that can characterize current power power system transient stability situation; Then key feature group is input in the ELMs trained, finally utilizes Decision-Making Evaluation criterion to draw final transient stability evaluation result and the degree of confidence of result.The while that system having high-precision, also meet the rapidity requirement of the large data of intelligent grid.

Limit unit (ELM), it is a kind of single hidden layer feedforward neural network be made up of input layer, hidden layer and output layer, due to its hidden layer structure relatively simple (only having one deck), the speed making it train and to judge is all very fast, can meet the requirement of real-time of large data structure; Because its sample training process has general randomization concept, make to which overcome the shortcomings that traditional neural network is easy to be absorbed in minimum value trap etc.The training method of single threshold unit (ELM) has had and has related in other documents, therefore only briefly introduces herein, if training sample number is n, with sigmoid function as activation function, the training method of single ELM is as follows:

Output function is:

$f_l\left(x_i\right)={\mathrm{\&Sigma;}}_{j=1}^l{\mathrm{\&beta;}}_{j}sigmoid(w_j{x}_i+b_j)=t_i,i=1,2,...n---\left(1\right)$

According to the correlation theory of ELM, (1) formula can be abbreviated as:

Hβ＝T(2)

Wherein: H=H (w ₁, w ₂, w ₃... w _l, b ₁, b ₂, b ₃... b _l, x ₁, x ₂x ₃x _n), H is neural network hidden layer output matrix; T=[t ₁, t ₂, t ₃.t _n] ^t, T is sample matrix; L is neuron number; N is sample number.

For single sample training process, random given implicit nodes, input weight vector w _jand b _j, then weight matrix β=H is exported ^-1t, wherein H ^-1it is the generalized inverse matrix of H.

The present embodiment for research object, judges its transient stability situation under various operating mode with IEEE57 node power system.Method mainly comprises training and decision-making two parts.Sample training part comprises again the concentration training of ELMs and electrical network transient state key feature extracts two parts.Decision part, first at large extracting data key feature, is then sent in the cluster EMLs trained, and after last its output of comprehensive analysis, provides the degree of confidence of steady assessment result and current results temporarily.

Ask for an interview Fig. 1, a kind of transient stability evaluation in power system method towards large data provided by the invention, comprises the following steps:

Step 1: according to practical power systems feature, determines the criterion of current power power system transient stability; The criterion of electric power system transient stability to characterize the whether stable parameter value of current power system transient modelling, when parameter value crosses certain threshold value, just assert that current system can not meet transient stability requirement.

Step 2: setting training and correlation parameter initial value, comprises the limit unit ELM sum E, the input feature vector number f of limit unit ELM, hidden layer node number interval [j_min, j_max], maximum frequency of training i_max.

In the present embodiment training process, the type of correlation parameter and initial value are as following table 1.

The type of table 1 correlation parameter and initial value

Parameter name	Value	Parameter name	Value
				Total sample number	6525	Training sample sum N	5220
Transient stability sample	4010	Test sample book sum	1305
				The unstable sample of transient state	2515	Single ELM input feature vector number f	35
ELM sum	300	Hidden layer node h number interval	[150 350]
				Frequency of training i	0	Sample training number of times	3000

Step 3: determine transient stability foundation characteristic storehouse;

Step 4: utilize the large data of electric system history to complete sample training, its specific implementation comprises following sub-step:

Step 4.1: concentration training study is carried out to multiple limits unit (ELMs); Ask for an interview Fig. 2, its specific implementation comprises following sub-step:

(1) 6525 groups of samples in electric system historical data and simulation software generation Sample Storehouse are utilized, wherein 4010 groups of transient stability samples, 2515 groups of unstable samples.Randomly draw 5220 (80% of sum) group sample as training sample, remain 1305 groups of samples and do test.

(2) to choose based on all parameters that can characterize electrical power system transient feature in feature database, feature database based on 410 parameters such as choose each busbar voltage amplitude, each busbar voltage phase place in this example, each load is meritorious, each reactive load, generated power, generator reactive, entire system are meritorious, entire system is idle.

(3) carry out i-th sample training, training comprises the focusing study training of ELMs and the iterative computation of current signature weight parameter.

The focusing study training of ELMs and the iterative algorithm of current signature weight parameter as follows:

(1) first time circulation is performed: FORi=1to3000 (frequency of training is set);

(2) second time circulation is performed: FORe=1to300 (selecting by the ELM trained);

(3) training sample is determined at random;

(4) determine at random current by training ELM node in hidden layer j ∈ [150350];

(5) in foundation characteristic storehouse random choose 30 features as input quantity current input ELM;

(6) input weight vector w is determined at random _jand b _j;

(7) β=H is calculated ^-1t;

(8) $W{\&lsqb;X\&rsqb;}^{i+1}=W{\&lsqb;X\&rsqb;}^i-\frac{dis(X,R_i,H)}{5220}+\frac{dis(X,R_i,M)}{5220}$ (upgrading current signature weight coefficient);

(9) judge whether second time circulation circulates complete, if then terminate second time circulation;

(10) judge whether first time circulation circulates complete, if then terminate first time circulation.

Step 4.2: in iterative computation foundation characteristic storehouse, each feature is to the weight coefficient of transient stability, determines key feature group;

Wherein temporary steady characteristic quantity weight coefficient, characterizes different electrical parameter to the separating capacity between sample; The feature that weight coefficient is large, little to the discrimination of sample Output rusults of the same type, namely limit unit EML output valve difference is little; The feature that weight coefficient is large, comparatively large to dissimilar sample area calibration, namely limit unit EML output valve difference is large; And the feature that weight coefficient is little, its output valve is insensitive to dissimilar sample.

Sample of the same type refers to that all transient stability result of determination are all sample sets of safety (dangerous); Inhomogeneity sample, refers to not at other samples of sample place set.

Based on this, in sample training process by continuous iteration, deduct sample nearest in similar sample set, add sample nearest in inhomogeneity sample set, finally draw the weight coefficient of each feature, concrete grammar is as follows:

(1) circulation is performed: the maximum frequency of training i_max of i=1to;

(2) $W{\&lsqb;X\&rsqb;}^{i+1}=W{\&lsqb;X\&rsqb;}^i-\frac{dis(X,R_i,H)}{N}+\frac{dis(X,R_i,M)}{N};$

(3) $dis(X,R,R^{\&prime;})=\frac{|val(X,R)-val(X,R^{\&prime;})|}{\max \left(X\right)-\min \left(X\right)};$

(4) judge whether circulation circulates complete, if then end loop;

In formula: X represents certain steady feature temporarily; the weight coefficient that representation feature X is final; R _iit is the sample of i-th training; H is from R in similar sample _inearest sample; M is from R in inhomogeneity sample _inearest sample; Val (X, R) is the value of feature X in R sample; Function dis (X, R, R ') calculate feature X between sample R and R ' relatively to difference; N is this example total sample number, and its existence makes

And, by all features in transient stability foundation characteristic storehouse, according to finally obtaining weight coefficient difference, arrange from big to small, select the forward 10% the most key feature of rank.

Step 5: extract transient state stable key syndrome data in current power system, and be sent in the multiple limits unit (ELMs) trained;

Step 6: carry out decision-making of comprehensive evaluation to the Output rusults of multiple limits unit (ELMs), provides the steady temporarily assessment result of current power system and the degree of confidence of current results.

Its specific implementation process is:

(1) as shown in Figure 3, the feature of 400 in foundation characteristic storehouse is pressed the height descending sort of weight coefficient;

(2) front 40 features composition key feature storehouse comprising No. 7 motors outputs and gain merit is chosen;

(3) 40 key features of current electric grid are extracted;

(4) 40 key features are sent in the EMLs trained;

(5) Output rusults of each ELM of Decision Evaluation module collection, the transient stability situation of objective evaluation current electric grid, and provide the degree of confidence of evaluation result.

The method to set up of the concrete fiducial interval of the present embodiment is:

If current power system is in transient stability state, result of determination=1; Current power system can not maintain transient stability state, result of determination=-1, then for the Output rusults Y of each the study unit in ELMs _e∈ [-1,1]:

Y _e∈ [0.7,1.3] or Y _e∈ [-1.3 ,-0.7]: current ELM study unit exports credible;

Y _e∈ [-0.7,0.7]: current ELM study unit exports insincere;

The steady temporarily decision criteria of the electric power system transient stability of the present embodiment is:

Learn in unit if u is 300 ELM, export the ELM sum of reliable result; V characterizes insincere result sum.If u<80%*E, illustrate that the credible output of current ELMs does not reach threshold requirement, current results is insincere, judges unsuccessfully;

If u>80%*E, and result of determination: current system transient stability; If u>80%*E, and result of determination: current system is transient stability not;

Embodiment final testing result is as following table 2:

Table 2 final testing result

Specific embodiment described herein is only to the explanation for example of the present invention's spirit.Those skilled in the art can make various amendment or supplement or adopt similar mode to substitute to described specific embodiment, but can't depart from spirit of the present invention or surmount the scope that appended claims defines.

Claims (9)

1., towards a transient stability evaluation in power system method for large data, it is characterized in that, comprise the following steps:

Step 1: according to practical power systems feature, determines the criterion of current power power system transient stability;

Step 2: setting training and correlation parameter initial value;

Step 3: determine transient stability foundation characteristic storehouse;

Step 4: utilize the large data of electric system history to complete sample training;

Step 5: extract transient state stable key syndrome data in current power system, and be sent in the multiple limits unit (ELMs) trained;

Step 6: carry out decision-making of comprehensive evaluation to the Output rusults of multiple limits unit (ELMs), provides the steady temporarily assessment result of current power system and the degree of confidence of current results.

2. the transient stability evaluation in power system method towards large data according to claim 1, it is characterized in that: the criterion of electric power system transient stability described in step 1, to characterize the whether stable parameter value of current power system transient modelling, when parameter value crosses certain threshold value, just assert that current system can not meet transient stability requirement.

3. the transient stability evaluation in power system method towards large data according to claim 1, it is characterized in that: training and correlation parameter initial value described in step 2 comprise the limit unit ELM sum E, the input feature vector number f of limit unit ELM, hidden layer node number interval [j_min, j_max], maximum frequency of training i_max.

4. the transient stability evaluation in power system method towards large data according to claim 1, is characterized in that: utilize the large data of electric system history to complete sample training described in step 4, its specific implementation comprises following sub-step:

Step 4.1: concentration training study is carried out to multiple limits unit (ELMs);

Step 4.2: in iterative computation foundation characteristic storehouse, each feature is to the weight coefficient of transient stability, determines key feature group.

5. the transient stability evaluation in power system method towards large data according to claim 4, is characterized in that, carry out concentration training study to multiple limits unit (ELMs) described in step 4.1, its idiographic flow is as follows:

(1) first time circulation is performed: the maximum frequency of training i_max of i=1to;

(2) second time circulation is performed: e=1to limit unit ELM sum E;

(3) training sample is determined at random;

(4) determine at random current by training limit unit ELM node in hidden layer j ∈ [j _min, j _max];

(5) in a foundation characteristic storehouse random choose f feature as current input limits unit ELM, wherein f is the input feature vector number of limit unit ELM;

(6) input weight vector w is determined at random _jand b _j;

(7) β=H is calculated ^-1t, wherein: H is neural network hidden layer output matrix, and T is sample matrix;

(8) judge whether second time circulation circulates complete, if then terminate second time circulation;

(9) judge whether first time circulation circulates complete, if then terminate first time circulation.

6. the transient stability evaluation in power system method towards large data according to claim 4, is characterized in that, described in step 4.2, temporary steady characteristic quantity weight coefficient, characterizes different electrical parameter to the separating capacity between sample; The feature that weight coefficient is large, little to the discrimination of sample Output rusults of the same type, namely limit unit EML output valve difference is little; The feature that weight coefficient is large, comparatively large to dissimilar sample area calibration, namely limit unit EML output valve difference is large; And the feature that weight coefficient is little, its output valve is insensitive to dissimilar sample; Described sample of the same type refers to that all transient stability result of determination are for safety or unsafe all sample sets; Described dissimilar sample refers to not at other samples of sample place set.

7. the transient stability evaluation in power system method towards large data according to claim 4 or 6, is characterized in that, temporary steady characteristic quantity weight coefficient described in step 4.2, and computing method comprise the following steps:

(1) circulation is performed: the maximum frequency of training i_max of i=1to;

(2) $W{\&lsqb;X\&rsqb;}^{i+1}=W{\&lsqb;X\&rsqb;}^i-\frac{dis(X,R_i,H)}{N}+\frac{dis(X,R_i,M)}{N};$

(3) $dis(X,R,R^{\&prime;})=\frac{|val(X,R)-val(X,R^{\&prime;})|}{\max \left(X\right)-\min \left(X\right)};$

(4) judge whether circulation circulates complete, if then end loop;

Wherein: X represents certain steady feature temporarily; the weight coefficient that representation feature X is final; R _iit is the sample of i-th training; H is from R in similar sample _inearest sample; M is from R in inhomogeneity sample _inearest sample; Val (X, R) is the value of feature X in R sample; Function dis (X, R, R ') calculate feature X between sample R and R ' relatively to difference; N is this example total sample number, and its existence makes

8. the transient stability evaluation in power system method towards large data according to claim 1, is characterized in that, provide the steady temporarily assessment result of current power system and the degree of confidence of current results described in step 6, the design standard of its fiducial interval is:

If current power system is in transient stability state, result of determination=1; Current power system can not maintain transient stability state, result of determination=-1, then for the Output rusults Y of each the study unit in ELMs _e∈ [-1,1]:

If Y _e∈ [0.7,1.3] or Y _e∈ [-1.3 ,-0.7], current ELM study unit exports credible: u=u+1;

If Y _e∈ [-0.7,0.7], current ELM study unit exports insincere: v=v+1.

9. the transient stability evaluation in power system method towards large data according to claim 8, is characterized in that, provide the steady temporarily assessment result of current power system and the degree of confidence of current results described in step 6, its temporary steady decision criteria is:

If u is characterized in E ELM and learns, in unit, to export the ELM number of reliable result; V characterizes insincere number of results;

If u<80%*E, then the credible output of current ELMs does not reach threshold requirement, and current results is insincere, judges unsuccessfully;

If u>80%*E, and then current system transient stability;

If u>80%*E, and then current system not transient stability;