CN105429134A - Grid voltage stability prediction method based on power big data - Google Patents

Abstract

The invention discloses a grid voltage stability prediction method based on power big data. According to the invention, voltage stability analysis is firstly carried out to a power grid in combination with the bifurcation theory, which shows that a saddle-node bifurcation point has a very close relation with the voltage stability of the power grid, and therefore the saddle-node bifurcation point is taken as one of indicators used for judging the voltage stability of the power grid; then, based on the big data technology, effective data mining is carried out to data obtained from the power grid; a nonlinear mapping relation between power data and the maximum load parameter is built by using an improved BP neural network based on the particle swarm optimization, thus achieving a prediction effect, i.e. the maximum load parameter can be predicated according to the current operating status of the power grid; in combination with the current load parameter of the power grid, a voltage stability margin can be calculated and the stability degree of the power grid can be judged; further, according to the prediction result, an adjusting control scheme for grid voltage can be made to increase the stability margin, enhance voltage stability and finally achieve the objective of long-term safe and stable operation of the power grid.

Description

A kind of Network Voltage Stability Forecasting Methodology based on the large data of electric power

Technical field

Invention belongs to electric power project engineering field, particularly a kind of Network Voltage Stability Forecasting Methodology based on the large data of electric power.

Background technology

The industries such as large data technique takes the lead in the Internet, telecommunications, finance occur, are often referred to the data volume of more than 10TB scale.If large data operated according to the ability that rational cost and time limit caught, managed and processed these data, its scale or complexity are beyond common technology.Investigative technique for large data rises to will of the state, and country has the scale of data and the ability of maintenance data will become the important component part of overall national strength, occupying and controlling also will become the contention focus between country and between enterprise data.The multinational giants such as IBM, Microsoft, Google, Amazon obtain stronger competitiveness by the development of large data technique.IBM invests 16,000,000,000 dollars and carries out more than 30 purchase relevant to large data, makes the stable rapid growth of achievement; EBay accurately calculates by data mining the return that in advertisement, each keyword brings, and since 2007, advertising expense reduces 99%.In October, 2012, set up that the first academic consultative and advisory body specializing in large market demand and development---the large data craft committee is learned by China's Telecommunication.At present, domestic correlation technique mainly concentrates on data mining related algorithm, practical application and the research about theoretical side.In the related, that more representative is the Hadoop of Apache Software Foundation exploitation and the MapReduce of Google's exploitation.The intelligent level of modern power systems is more and more higher, along with increasing of kind of sensor in electrical network and adding of smart machine, the various power equipments being in running status are often automatically measured once through a time interval, and measurement result is pooled to rapidly in a database through data transmission network.After distance, the size of database often can reach the scale of TB rank, no matter the kind from data or the scale from data, these electric power datas have belonged to the category of large data.These data class recorded in real time from electrical network are assorted many, and flowing velocity is fast, and value density is low, and conventional data processing means can not effectively excavate it, therefore need the large data technique with reference to current trend to carry out mining analysis to it.

Power system voltage instability/voltage collapse accident is an importance of electric power system loses stability.At present there is multiple angles and method is studied Network Voltage Stability, according to the time domain of voltage stabilization, voltage stabilization can be divided into Transient Voltage Stability, the stable and long-term voltage stability of mid-life voltage.The time range of Transient Voltage Stability is 0 ~ 10 second, main research angle be after induction motor and Voltage Instability, particularly short circuit caused by HVDC (high voltage direct current) on line access weak pattern system induction motor because of the Voltage Instability problem of the asynchronous machine step-out of accelerating the unstability that causes and cause due to the weak contact of network.The stable time domain of mid-life voltage is 1 ~ 5 minute, comprises OLTC, the effect of voltage regulating device and generator overexcitation limiter in power distribution network.The time domain of long-term voltage stability is 20 ~ 30 minutes, its factor of being mainly correlated with is that load rapid, high volume increases, transmission-line power increases in a large number, permanently can lose asynchronism etc. because of low-voltage by the load that causes of load and homothermal control load owing to existing in system.What also do not have at present that a kind of method can be long-term predicts the stability of line voltage, thus under ensureing the state that line voltage is in steady operation always.

Summary of the invention

Goal of the invention: the object of the invention is to for the deficiencies in the prior art, provide a kind of can accurately and rapidly to the Network Voltage Stability Forecasting Methodology based on the large data of electric power that the stability of line voltage is predicted.

Technical scheme: the invention provides a kind of Network Voltage Stability Forecasting Methodology based on the large data of electric power, comprise the following steps:

Step 1: in conjunction with bifurcation theory, sets up Network Voltage Stability judge index, and wherein Network Voltage Stability judge index is voltage stability margin wherein, λ _maxprepresent the peak load parameter using BP neural network prediction system out current; λ represents the load parameter of current actual measurement;

Step 2: garbled data kind in electric power system, selecting to associate close data with voltage stability and carries out gathering and be normalized preliminary treatment to the data collected, being made into the original sample for training BP neural network model;

Step 3: use the sample data that step 2 obtains, particle cluster algorithm (hereinafter referred PSO algorithm) is utilized to carry out optimal selection to the connection weights and threshold in BP neural net, and the BP neural net through particle cluster algorithm optimization is trained, obtain the neural network model after training;

Step 4: the relevant real time data input step 3 of current for electrical network operation is trained the BP neural network model drawn, under the Nonlinear Mapping rule determined, obtain the peak load parameter lambda that electrical network is current _maxp; According to formula calculating voltage stability margin μ;

Step 5: the voltage stability margin μ obtained according to step 4, judges the stability of line voltage, carry out regulating and controlling to line voltage.

Further, in described step 1 load parameter according to formula calculate, wherein P is the actual measurement active power of electrical network current loads; P ₀for the ground state active power of network load actual measurement.

Further, be normalized pretreated method to collecting data be in described step 2: in the parameter value of each parameter, select maximum x _maxwith minimum value x _min, according to formula initial data is all transformed to the number in interval [-1,1].The data due to individual species can be avoided so excessive or too smallly to have a negative impact to follow-up neural network training process, avoid making training result converge on local minimum prematurely.

Further, that selects in described step 2 associates close data from voltage stability and comprises electrical network interdependent node voltage magnitude and phase angle under different steady operational status, flow to the active power and reactive power that flow out this node, and the corresponding peak load parameter of electrical network using Continuation Method to calculate.

Further, in described step 5 Network Voltage Stability judge standard be: when μ ∈ (0.5,1] time, line voltage is in degree of stability; When μ ∈ (0.2,0.5] time, line voltage is in warning degree; When μ ∈ (0,0.2] time, line voltage is in unstable degree.

Further, the measure in described step 5, regulating and controlling being carried out to line voltage comprise adopt automatic voltage adjustor of power generator to regulate, adopt on-load tap-changing transformer to regulate, adopt BIFURCATION CONTROL device to regulate, adopt electrical network automatic voltage control system to regulate, increase static synchronous compensating device, increase static passive compensation device or shunt capacitor, load rejection.

Operation principle: first the present invention carries out Voltage Stability Analysis in conjunction with bifurcation theory to electric power system, finds that saddle-node bifurcation point has very close contact for the stability affecting line voltage, and using this as one of index judging Network Voltage Stability degree.Use for reference current large data technique afterwards, effective data mining is carried out to the data obtained from electrical network.Use the Nonlinear Mapping relation between BP neural network electric power data (such as relevant Power Flow Information etc.) and peak load parameter improved based on PSO algorithm, thus a kind of effect of prediction can be reached, namely current according to electrical network running status dopes peak load parameter point, i.e. saddle-node bifurcation point.In conjunction with current load parameter, calculate voltage stability margin, judge the degree of stability of line voltage.Further according to predicting the outcome, determining regulable control scheme, increasing stability margin, strengthen the stability of voltage, reach the object of electrical network long-term safety stable operation.

Beneficial effect: compared with prior art, the present invention can predict line voltage for a long time and monitor, and ensures that line voltage is in stable state, maintaining system safety sustainable operation for a long time; And it is more accurate and quick to the prediction of Network Voltage Stability.

Accompanying drawing explanation

Fig. 1 is flow chart of the present invention;

Fig. 2 is the IEEE14 node system model schematic set up in the embodiment of the present invention;

Fig. 3 is the λ-V curve chart of the node 4 that the IEEE14 node system in the embodiment of the present invention under variable load model calculates through Continuation Method;

Fig. 4 is optimal adaptation degree change curve when using PSO algorithms selection BP neural net optimized parameter in the embodiment of the present invention;

Fig. 5 is the matched curve figure of peak load parameter prediction value and the actual value using the BP neural net through PSO algorithm optimization to obtain in the embodiment of the present invention.

Embodiment

Below technical solution of the present invention is described in detail, but protection scope of the present invention is not limited to described embodiment.

As shown in Figure 1, the invention provides a kind of Network Voltage Stability Forecasting Methodology based on the large data of electric power, specifically comprise the following steps:

Step 1: in conjunction with bifurcation theory, sets up voltage stability judge index;

Set up power system simulation model, as shown in Figure 2, in the present embodiment, set up IEEE14 node system simulation model.

Use Continuation Method to system model carry out diverging calculate after obtain on the λ-V balance point curve of system and curve saddle-node bifurcation point; Load parameter when being occurred by saddle-node bifurcation point is considered as peak load parameter lambda _max.As shown in Figure 3, display be the λ-V curve of IEEE14 node system interior joint 4, the flex point being arranged in right side is exactly the saddle-node bifurcation point of above-mentioned Bifurcation, and system cloud gray model has peak load parameter lambda under this dotted state _max, these data are most important for the present invention.In follow-up forecasting process, the peak load parameter lambda of system _maxpbe the Output rusults of BP neural network prediction.

Getting voltage stability margin μ is:

$\mathrm{\μ}=|1-\frac{\mathrm{\λ}}{{\mathrm{\λ}}_{\max p}}|---\left(1\right)$

When μ ∈ (0.5,1] time, definition line voltage be in degree of stability; When μ ∈ (0.2,0.5] time, definition line voltage be in warning degree; When μ ∈ (0,0.2] time, definition line voltage be in unstable degree.

Step 2: in all data class measured in electric power system, screening associates close data with Network Voltage Stability and gathers.To investigate the voltage stability of IEEE14 node system interior joint 4, determine that the data class collected is node 4 voltage magnitude of IEEE14 node system when 121 kinds of steady operations and phase angle, the each circuit be connected with node 4 and transformer branch flow to and flow out active power and the reactive power of node 4, as the input variable of BP neural net, totally 12 inputs; Under these 121 kinds of stable states, use Continuation Method to calculate the corresponding peak load parameter of electrical network, as the output variable of BP neural net, totally 1 output.Be normalized preliminary treatment stored in after database to these data, pretreated method selects maximum x in the parameter value of each parameter _maxwith minimum value x _min, according to formula (2)

$y=-1+2\frac{x-x_{\min }}{x_{max}-x_{\min }}---\left(2\right)$

Initial data is all transformed to the number in interval [-1,1], wherein, x represents initial data, and y represents that initial data is according to the value after formula (2) conversion.Value y after being changed by the initial data of each parameter constitutes original sample.The data due to individual species can be avoided so excessive or too smallly to have a negative impact to follow-up neural network training process, avoid making training result converge on local minimum prematurely.Pretreated data are as the original sample of follow-up neural network training.

Step 3: the data in the original sample using step 2 to obtain, PSO algorithm is utilized to carry out optimal selection to the connection weights and threshold in BP neural net, and the BP neural net through particle cluster algorithm optimization is trained, obtain the neural network model after training;

First set up the BP neural network model of one 12 dimension input and 1 dimension output, adopt classical 3 layer network structures, input layer number is 12; The nodes of middle hidden layer is rule of thumb 2 extraordinarily 1 of input dimension, namely 23; Output layer nodes is 1, thus determines the basic structure of BP neural net.

In PSO algorithm, entity is regarded as particle, and the position of particle is exactly required solution.For the present embodiment, now get connection weights in BP neural net and threshold value as two-dimensional particles, and the fitness function for evaluating particle in PSO algorithm just adopts the mean square error function evaluating BP neural net performance, the parameter making neural net mean square error minimum by search performance fast finding in global scope that PSO algorithm is powerful.Jth dimension value for i-th particle upgrades by following two formula:

$v_{ij}^{k+1}={\mathrm{wv}}_{ij}^k+c_1\mathrm{\ξ}(p_{ij}^k-x_{ij}^k)+c_2\mathrm{\η}(p_{gj}^k-x_{ij}^k)---\left(3\right)$

$x_{ij}^{k+1}=x_{ij}^k+{\mathrm{\γv}}_{ij}^{k+1}---\left(4\right)$

In formula (3), formula (4), v _ijand x _ijrepresent jth dimension velocity information and the positional information of i-th particle in solution space respectively; p _ijrepresent the jth dimension history optimal location value that i-th particle oneself searches, p _gjrepresent the jth dimension optimal location value that all particles searches; W is the coefficient keeping the original speed of particle, is called inertia weight; c ₁be the weight coefficient of self optimal value of particle tracking in an iterative process, default value is 2; c ₂be the weight coefficient of particle tracking colony optimal value in an iterative process, default value is 2; γ is when upgrading particle position, and the coefficient added before speed, is referred to as " constraint factor ", and default value is 1; ξ and η is positioned at [0,1] interval random number.PSO algorithm is the optimized algorithm based on iteration pattern, the respective value calculated after representing kth time iteration, the respective value calculated after representing kth+1 iteration.

Each new particle is exactly connection weights and threshold new in BP neural net.Each particle is substituted in BP neural net and draws corresponding mean square error, when mean square error reach be satisfied with index or iterations reach the number of times of regulation time, BP neural net just can obtain optimum connection weights and threshold, and under this parameter combinations, BP neural net has optimum performance.

Write the program of PSO algorithm optimization BP neural net, complete the initialization of all PSO algorithm relevant parameters; Wherein, the maximal rate of particle is 0.5, judges that the index of adaptive value be minimal error is 0.001, and weights between 0.1 ~ 0.9, and are linear from 0.9 declines; The random value of initial position between-1 ~ 1 of particle; The initial velocity of particle is a random value between-2 ~ 2; Total number of particles for searching for is 45.Data in original sample step 2 obtained substitute into program and are optimized calculating, and as shown in Figure 4, this time parameter optimization is when iterations reaches 135 times, particle fitness tends to be steady, finally be stabilized in a value, thus obtain optimal particle, be i.e. optimum connection weights and threshold.

After this, optimum is connected weights and threshold value is assigned to BP neural net, data training BP neural net in the original sample using step 2 to obtain, finally obtains the BP neural network model meeting least mean-square error requirement, thus ensures the accuracy of subsequent prediction data.

Step 4: train the neural network model drawn by the relevant real time data input step 3 of current for electrical network operation, prediction show that the current peak load parameter of system is 3.9315p.u., the system peak load parameter now using Continuation Method to calculate is 3.9252p.u., can be considered actual value, the relative error of predicted value and actual value is only 0.1605%.

When use is predicted through the BP neural net of PSO algorithm optimization, if sample data is not through normalization preliminary treatment, the peak load parameter prediction value then obtained is 3.9324p.u., be 0.1834% with the relative error of actual value, be greater than to use and carry out through pretreated sample data the error predicted.

Predict without the BP neural net of PSO algorithm optimization if use, even if carried out normalization preliminary treatment to sample data, obtaining peak load parameter prediction value is 3.9359p.u., reach 0.2700% with the relative error of actual value, be greater than the predicated error used through the BP neural net of PSO algorithm optimization.

The peak load parameter prediction value using the BP neural net through PSO algorithm optimization to obtain and the matched curve of actual value, as shown in Figure 5, because sample data is all through preliminary treatment, figure can find out the point much doped overlaps substantially with original point, the estimated performance as seen through the BP neural net of PSO algorithm optimization is very superior.

After obtaining peak load parameter prediction value, the current stability margin of node 4 voltage can be calculated by formula (1), thus judge current voltage stabilization degree, for step 5 provides reference.Such as system when ground state is run, current loads parameter lambda=1p.u., and peak load parameter prediction value λ _maxp=3.9315p.u., then by formula (1) calculate the current stability margin μ of node 4 voltage be 0.7456, μ ∈ (0.5,1], therefore in the current state of operation, node 4 voltage is stable.

Step 5: the voltage stability margin μ obtained according to step 4, judges the stability of line voltage, carry out regulating and controlling to line voltage.In conjunction with the Network Voltage Stability degree of prediction, formulate regulation scheme fast and effectively.Conventional voltage-regulation measure has employing automatic voltage adjustor of power generator (AVR) to regulate, adopts on-load tap-changing transformer (OLTC) to regulate, adopts BIFURCATION CONTROL device to regulate, adopts electrical network automatic voltage control system (AVC) to regulate, increases static synchronous compensating device (STATCOM), increases static passive compensation device (SVC) or shunt capacitor, load rejection etc.By the adjustment of these measures, system running state should be able to away from saddle-node bifurcation point, and like this, voltage stability domain degree must increase, and the stability of line voltage accesses enhancing to a certain degree.

Claims (6)

1., based on a Network Voltage Stability Forecasting Methodology for the large data of electric power, it is characterized in that: comprise the following steps:

Step 1: in conjunction with bifurcation theory, sets up Network Voltage Stability judge index, and wherein Network Voltage Stability judge index is voltage stability margin μ; wherein, λ _maxprepresent the peak load parameter using BP neural network prediction electrical network out current; λ represents the load parameter of current actual measurement;

Step 2: garbled data kind in electric power system, selecting to associate close data with voltage stability and carries out gathering and be normalized preliminary treatment to the data collected, being made into the original sample for training BP neural network model;

Step 3: the data in the original sample using step 2 to obtain, particle cluster algorithm is utilized to carry out optimal selection to the connection weights and threshold in BP neural net, and the BP neural net through particle cluster algorithm optimization is trained, obtain the neural network model after training;

Step 4: the relevant real time data input step 3 of current for electrical network operation is trained the BP neural network model drawn, under the Nonlinear Mapping rule determined, obtain the peak load parameter lambda that electrical network is current _maxp; According to formula calculating voltage stability margin μ;

Step 5: the voltage stability margin μ obtained according to step 4, judges the stability of line voltage, carry out regulating and controlling to line voltage.

2. the Network Voltage Stability Forecasting Methodology based on the large data of electric power according to claim 1, is characterized in that: in described step 1, load parameter is according to formula calculate, wherein P is the actual measurement active power of electrical network current loads; P ₀for the ground state active power of network load actual measurement.

3. the Network Voltage Stability Forecasting Methodology based on the large data of electric power according to claim 1, is characterized in that: being normalized pretreated method to the data collected in described step 2 is: in the parameter value of each parameter, select maximum x _maxwith minimum value x _min, according to formula initial data is all transformed to the number in interval [-1,1].

4. the Network Voltage Stability Forecasting Methodology based on the large data of electric power according to claim 1, it is characterized in that: that selects in described step 2 associates close data with voltage stability, comprise the interdependent node voltage magnitude of electrical network under different steady operational status and phase angle, flow to the active power and reactive power that flow out this node, and the corresponding peak load parameter of electrical network using Continuation Method to calculate.

5. the Network Voltage Stability Forecasting Methodology based on the large data of electric power according to claim 1, is characterized in that: in described step 5 Network Voltage Stability judge standard be: when μ ∈ (0.5,1] time, line voltage is in degree of stability; When μ ∈ (0.2,0.5] time, line voltage is in warning degree; When μ ∈ (0,0.2] time, line voltage is in unstable degree.

6. the Network Voltage Stability Forecasting Methodology based on the large data of electric power according to claim 1, is characterized in that: the measure in described step 5, regulating and controlling being carried out to line voltage comprise adopt automatic voltage adjustor of power generator to regulate, adopt on-load tap-changing transformer to regulate, adopt BIFURCATION CONTROL device to regulate, adopt electrical network automatic voltage control system to regulate, increase static synchronous compensating device, increase static passive compensation device or shunt capacitor, load rejection.