CN106662846A - Method for estimating status of ac networks and subsequent adaptive control - Google Patents

Method for estimating status of ac networks and subsequent adaptive control Download PDF

Info

Publication number
CN106662846A
CN106662846A CN201480081138.7A CN201480081138A CN106662846A CN 106662846 A CN106662846 A CN 106662846A CN 201480081138 A CN201480081138 A CN 201480081138A CN 106662846 A CN106662846 A CN 106662846A
Authority
CN
China
Prior art keywords
networks
network
state
measurement
training
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN201480081138.7A
Other languages
Chinese (zh)
Inventor
A.J.赫尔南德斯曼乔拉
F.舍特勒
J.洛特斯
M.斯蒂格
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Siemens AG
Original Assignee
Siemens AG
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Siemens AG filed Critical Siemens AG
Publication of CN106662846A publication Critical patent/CN106662846A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B13/00—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion
    • G05B13/02—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric
    • G05B13/0265—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric the criterion being a learning criterion
    • G05B13/027—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric the criterion being a learning criterion using neural networks only
    • H—ELECTRICITY
    • H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
    • H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
    • H02J13/00—Circuit arrangements for providing remote indication of network conditions, e.g. an instantaneous record of the open or closed condition of each circuitbreaker in the network; Circuit arrangements for providing remote control of switching means in a power distribution network, e.g. switching in and out of current consumers by using a pulse code signal carried by the network
    • H02J13/00002—Circuit arrangements for providing remote indication of network conditions, e.g. an instantaneous record of the open or closed condition of each circuitbreaker in the network; Circuit arrangements for providing remote control of switching means in a power distribution network, e.g. switching in and out of current consumers by using a pulse code signal carried by the network characterised by monitoring
    • H—ELECTRICITY
    • H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
    • H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
    • H02J3/00—Circuit arrangements for ac mains or ac distribution networks
    • H02J3/001—Methods to deal with contingencies, e.g. abnormalities, faults or failures
    • H02J3/0012—Contingency detection
    • H—ELECTRICITY
    • H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
    • H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
    • H02J2203/00—Indexing scheme relating to details of circuit arrangements for AC mains or AC distribution networks
    • H02J2203/20—Simulating, e g planning, reliability check, modelling or computer assisted design [CAD]
    • H—ELECTRICITY
    • H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
    • H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
    • H02J3/00—Circuit arrangements for ac mains or ac distribution networks
    • H02J3/36—Arrangements for transfer of electric power between ac networks via a high-tension dc link
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02E—REDUCTION OF GREENHOUSE GAS [GHG] EMISSIONS, RELATED TO ENERGY GENERATION, TRANSMISSION OR DISTRIBUTION
    • Y02E60/00—Enabling technologies; Technologies with a potential or indirect contribution to GHG emissions mitigation
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02E—REDUCTION OF GREENHOUSE GAS [GHG] EMISSIONS, RELATED TO ENERGY GENERATION, TRANSMISSION OR DISTRIBUTION
    • Y02E60/00—Enabling technologies; Technologies with a potential or indirect contribution to GHG emissions mitigation
    • Y02E60/60—Arrangements for transfer of electric power between AC networks or generators via a high voltage DC link [HVCD]
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
    • Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
    • Y04S10/00—Systems supporting electrical power generation, transmission or distribution
    • Y04S10/30—State monitoring, e.g. fault, temperature monitoring, insulator monitoring, corona discharge
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
    • Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
    • Y04S40/00—Systems for electrical power generation, transmission, distribution or end-user application management characterised by the use of communication or information technologies, or communication or information technology specific aspects supporting them
    • Y04S40/20—Information technology specific aspects, e.g. CAD, simulation, modelling, system security

Abstract

A method for estimating the status of an AC network is disclosed. The method comprises a step of providing a training dataset comprising a set of training examples (X=[X0 X1 X2 ...Xn]) and a corresponding output of the status of the AC network (y), a step of training a hypothesis function (h[Theta](x)) based on the training dataset and finally a step of estimating the status of the AC network using the hypothesis function (h[Theta](x)). Based on the estimated status of the AC network the controller parameters of the power converter are adjusted to achieve an optimum performance behavior in the power converter. In one embodiment of the disclosed method, the status of the AC network is one of an islanded network condition or a non-islanded network condition, in another embodiment the status of the AC network is the loss of the last feeder and in yet another embodiment the status of the AC network is the strength of the AC network and the amount of nearby voltage control present.

Description

For the method for estimating the state of AC network and subsequent Self Adaptive Control
Technical field
The present invention relates to a kind of method for estimating the state of AC networks, is used to estimate AC more particularly, to a kind of The state of network and the method that Self Adaptive Control is realized based on the state of AC networks.
Background technology
Estimation with the state with regard to AC networks is for the safety and stable operation and general electricity of high power conversion device Force system is extremely important.In this case, the state of AC networks means there is isolated island or non-island network condition, load feedback The loss of line or the amount of the intensity of AC networks and neighbouring voltage controller.Once the state of known AC networks, it is possible to set Adaptive control process is put, such as in can adjusting the controller parameter being associated with power converter to realize power converter Improved performance behavior.For example, for the weak AC network with weak voltage control quantity nearby, control will in one way be adjusted Device parameter, and for the strong AC networks with strong voltage control quantity nearby, controller parameter is adjusted in another way.
Under the conditions of island network, the sub-fraction of system is isolated with the remainder of AC networks.This is likely due to line The contingency that road is tripped or transformer fault is similar with other.Island network condition is potential dangerous situation, its Severe stress may be caused on the electrical equipment of AC networks, caused systematic jitters and cut-out or power-off may be caused.
The loss of load feeder line is applied to Power electronic converter, and due to system-contingent, its is completely isolated or from AC nets Network disconnects.System-contingent can be the unexpected tripping operation of such as AC circuits or transformer.The loss of load feeder line is potential danger Dangerous situation condition, it may cause the stress of danger to the electrical equipment in Power electronic converter.
Not used for the quick and reliable system of the state for estimating AC networks, in several milliseconds that it can be after generation The isolated island of identification load feeder line or loss, the reason for this is probably the limit stress on electrical equipment.Additionally, can not accurately estimate The intensity of meter AC networks, and nearby voltage-controlled amount may cause the work(controlled by such as SVC, HVDC or wind turbine Rate electronic equipment take less than optimum control action, cause systematic jitters and interruption in power.
Up to the present, in the case of rapid dynamic power electronic controller (such as LCC and VSC HVDC), need to come The outer triggering signal of automatic network operator recognizing the network condition of isolated island, to take in the presence of island-grid condition Appropriate control action.In do not know how in the prior art to occur at it 50 to 100 milliseconds quickly accurately Prediction isolated island or non-island network condition.
Estimation with the state with regard to AC networks, can be according to the presence of the intensity of system and neighbouring voltage controller To adjust controller parameter, such as by adjusting gain, time constant etc., and therefore improvement controller adapts to new system condition Ability.
Automatic gain adjustment for Power electronic converter is a more and more important problem, and it is only partly solved Certainly.It allows converter, and such as static var compensator or STATCOM adjust its controller parameter, with system dynamic course Middle offer better performance.
Based on the measurement started during limit, automatic gain adjustment is had been realized in so far.Gain test is basic It is upper to include measuring injection or absorption due to a certain amount of reactive power (dQ) from converter and the electricity in caused bus Buckling (dV).Then use the voltage change (dv/dq) of the change relative to reactive power to estimate to be connected to converter The intensity of bus.The intensity of bus is by it to voltage and frequency during idle and active power are injected or absorbed in that bus The sensitivity of change is limited.Compared with weak bus, voltage and frequency change are less likely to occur strong bus.
Do not exist from power converter power injection or absorb in the case of, another kind of gain test method is to cut AC wave filters, shnt capacitor or reactor are changed or removed, and measures rate of change of the voltage relative to reactive power again (dv/dq)。
With the increase of the quantity of the voltage control power electronic converter in AC networks, ultimate challenge up to the present One of be to select with appropriate controller parameter, such as gain that takes into account depositing for neighbouring " other " voltage control apparatus .
The reliable of amount without system strength and neighbouring voltage controller estimates the reason for being less than optimum performance behavior, Such as power electronic drives rise time, stabilization time maximum overshoot, not enough damped oscillation in converter etc..
Up to the present, gain test as above is performed to estimate the intensity of the bus that converter is connected to, but If there is neighbouring voltage control apparatus, such as other quick dynamic compensations, then the result of gain test will mislead.Example Such as, the very weak bus with a large amount of voltage controllers nearby may look like the strong bus of gain test, cause completely Unsuitable controller parameter adjustment.
Therefore, a kind of method of the state for estimating AC networks exactly is had increasing need for.Once have estimated the shape of AC networks State, it is possible to correspondingly change the controller parameter of power converter using identical information, to realize more preferable systematicness Energy.
The content of the invention
It is an object of the invention to the state of fast prediction AC network, for example, if there is isolated island or non-island network Condition, if there occurs load feeder line loss, and the intensity of AC networks and exist neighbouring voltage-controlled amount be it is how many, And therefore controller parameter is adjusted to realize optimal system performance.The purpose is connected to power conversion by a kind of for estimation Realizing, the method includes the method for the state of the AC networks of device：The step of providing training dataset, the training dataset includes Training sample set (X=[X0 X1 X2 ... Xn]) and AC networks state correspondence output (y)；Assembled for training based on training data Practice and assume function (hθ(x)) the step of；And it is last using hypothesis function (hθ(x)) estimate AC networks state the step of.
Hypothesis function (the h of trainingθ(x)) use be supervised learning method form.Supervised learning is the instruction from mark Practice the machine learning task of inferred from input data function.Training data is made up of training example.Training example includes that input vector is (" special Levy X ") and its associated correct output valve (y).The task of supervised learning algorithm is analyzing and training data and proposes to assume function (hθ(x)), it can be used for mapping new example, i.e., for the h of all training examplesθ(x)≈y。
In the present invention, the task of supervised learning method is based on by the input vector of selected " feature X " and its right The training set that the output y (i.e. correct option) for answering is constituted, to create function (hθ(x)) quickly and accurately estimating AC networks State, be island shape network condition be also non-island shape network condition, if there occurs the loss of load feeder line, and AC networks Intensity and neighbouring voltage controller amount.
One of the state of AC networks and the challenge of accurate estimation of network strength are quickly obtained using intelligent algorithm It is the input for recognizing appropriate " feature X " for use as algorithm.For the present invention, it is proposed that calculating model of neural networks.
Neutral net is imitated in some sense the mankind and is learned through experience from training dataset popularization and the ability for learning Ability.Neutral net is used for prediction and estimation problem.For the problem solved using neutral net, what needs were well understood by Input.Need to be best understood by which function for predicting that correct output is critically important.Such input can be readily available, But how to combine them to accurately be estimated unclear.The next one requires it is have the output that is well understood by, i.e., with regard to Expect the information of the species of the output for being estimated, predict or modeling.Next factor is to use available experience.For training god Jing networks, the sample with the training set obtained by experience.Concentrate in these sample datas, input, characteristic vector (" feature X ") It is the known case for training neutral net with output (y).
The method also includes adjusting the step of the controller parameter of power converter based on the state of estimated AC networks Suddenly.For example, for island network condition, controller parameter is adjusted in one way, for non-island network condition, with another kind Mode adjusts controller parameter.Similarly, for what voltage and frequency were relatively easy to change with weak voltage control quantity nearby is weak AC networks, adjust in one way controller parameter, and for voltage and frequency are less prone to as strong voltage nearby is controlled The strong AC networks measured and change, adjust in another way controller parameter.
For example, the voltage controller gain for the SVC or STATCOM of strong AC networks needs the voltage higher than weak AC network Controller gain.
In one embodiment, the state of AC networks is one of isolated island or non-island network condition, the loss of load feeder line Or the intensity of the AC networks with neighbouring voltage control quantity.Island network condition or load feedback are recognized in several milliseconds after occurring The loss of line can prevent the limit stress in electrical equipment.It may also help in power electronic equipment and takes Optimal Control to move Make, occur so as to preventing systematic jitters and having a power failure.Even if in such as HVDC LCC converters, VSC HVDC, wind turbine In the presence of the nonlinearity power electronic equipment of machine and STATCOM/SVC, the supervised learning method proposed in the present invention AC network states fast and accurately are allowed to recognize.
As it was previously stated, under the conditions of island network, the sub-fraction of system is isolated with the remainder of AC networks.This may Occur due to the incident similar with other of the tripping operation of circuit or the failure of transformer.Island network condition is Potential dangerous situation, it may cause severe stress on the electrical equipment of AC networks, cause systematic jitters and can Power cut or power-off can be caused.
The loss of load feeder line is applied to Power electronic converter, and due to system-contingent, its is completely isolated or from AC nets Network disconnects.System-contingent can be tripping unintentionally for such as AC circuits or transformer.The loss of load feeder line is potential danger Dangerous situation condition, it may cause the stress of danger to the electrical equipment in Power electronic converter.
In another embodiment, from voltage measurement (V) and the reactive power exchange (Q of network measureex) and network survey The active power of amount exchanges (Pex) and AC networks network measurement (Freq) in one or more derive training sample sets Close (X=[X0 X1 X2 ... Xn])。
For this application of neutral net, i.e., for estimating the state of AC networks, the present invention proposes for example above-mentioned measurement As the feature of the input that be used as neutral net.Any other suitable measurement can also be used.In the mother needed for network Voltage measurement is carried out at line.It is appropriate with the voltage measurement identical voltage measurement for transducer voltage controller.Positive sequence Fundamental frequency voltages measurement can serve as substituting.Also measurement measures (Q with the reactive power exchange of networkex), the active power with network Exchange measurement (Pex) and network measurement.These measurements can be at the bus needed for network or in the power line of electrical network Place is carried out.
In another embodiment, training sample set (X=was measured before, during and/or after system incident [X0 X1 X2 ... Xn]).Therefore, for the system incident to be detected, some form of fault detect is needed so that one Denier failure is eliminated, then the measurement being input into needed for " feature X ", i.e. training example can start.A kind of such fault detect side Method can be that the event of failure or accident in AC networks is marked using under-voltage detection, and trigger neutral net AC network State estimation procedure.For example, the under-voltage of 0.8V can serve as trigger.Can be used for the other factorses of detection failure can be Voltage, overcurrent, undercurrent or other it is similar can monitoring condition.Under-voltage condition provides a kind of failure simply and readily Detection method.
In one embodiment, system incident is fault clearance event.In another embodiment, the accidental thing of system Part is gain test, and wherein reactive power is injected or absorbed in AC networks from power converter.It is also possible, however, to use any Other gain test methods, for example, do not exist from power converter power injection or absorb in the case of, AC wave filters Or reactor can be switched on or disconnect, and measure change (dv/dq) of the voltage relative to power.
In one embodiment, the first time period measurement voltage measurement before and after, during system incident (V) AC networks are introduced in and the step of second time period.In another embodiment of the method, in the accidental thing of system Before and after, during part first time period measurement network measurement (Freq) and the step of second time period in by its Introduce AC networks.In yet another embodiment, the first time period measurement before and after, during system incident is active Power Exchange measures (Pex), and the step of second time period in be introduced into AC networks.In another embodiment, in system First time period measurement reactive power exchange measurement (Q before and after, during incidentex) and in the step of second time period AC networks are introduced in rapid.
, corresponding to the time period for measuring, second time period is corresponding to time step every time between measurement for first time period It is long.
In the exemplary embodiment, when system incident is fault clearance event, first time period crosses over 50ms, the Two time periods were 5ms.In first 50ms after fault clearance event occurs, start to measure on required bus, and with 5ms step-lengths are measured.
In a further exemplary embodiment, when system incident is gain test, first time period crosses over 200ms, Second time period is 1ms.Measurement starts on the bus needed for converter, continues the span of 200ms, and it is covered in gain survey Duration before and after, during examination, step-length is 1ms.
The property of the feature proposed due to this application for neutral net, the time step between measurement is for example built View 5ms or 1ms, and the length of overall measurement, for example, advise 50ms or 200ms, affects the input " feature X " of neutral net Quantity.
For the length of the measurement of " feature " definition, i.e. first time period, the precision estimated is affected.The measurement for using is got over Long, it is more accurate to predict, because there are algorithm more information to come provides basis for its prediction.If other time of measuring show Better performance, then can use other time of measuring, it is important that the accuracy of prediction is acceptable, and the prediction exists May be used to take effective control action on time.
For the time step of " feature " definition, i.e. second time period, estimated accuracy is affected.If other times step shows Show better performance, it is possible to use other times step-length, it is important that the accuracy of prediction is acceptable.For example, with 5ms or The small step length of 1ms is measured and will be provided for training hypothesis function hθThe more preferable and finer training dataset of (x).
Table 1 below illustrates the exemplary training data collection of " feature X ", it is assumed that with by input vector or training example (X=[X0 X1 X2 ... Xn]) and AC networks state correspondence output (y) constitute ' m' training sample, can be used for Hypothesis function h is utilized using calculating model of neural networksθX method that () estimates the state of AC networks.In this case, to Amount X is by " feature X "=(X=[X0 X1 X2 ... Xn])∈R40+1The input variable for being given." feature X " for being proposed is related to The measurement carried out after incident, such as fault clearance event.
Table 1
First row has the number (m) of the training sample for using.Next group of row are comprising from the idle of voltage (V) and network Power Exchange measures (Qex) measurement (P is exchanged with the active power of networkex) and AC networks network measurement (Freq) lead The input variable " feature X " for going out.After fault clearance or incident occur, all measurements are measured with the step-length of 5ms, are continued 50ms.Last row gives the correct output state (y) of AC networks.For the exemplary training data collection shown in table 1, The state of AC networks is estimated as island network condition or non-island network condition.These results ' 0' or ' 1' in binary form Represent, wherein ' 0' be intended as the island network condition of AC network states, and ' 0' means non-island network condition.
In this example, table 1 is merely provided for training hypothesis function hθOne sample training data set of (x).Upper table is only It is an example data set.However, for different networks, it is possible to use different samples is forming corresponding data set.
Table 2
Table 2 below shows another exemplary training dataset.First row has the number of the training sample for using (m).Next group of row measure (Q comprising the reactive power exchange from voltage (V) and networkex) exchange survey with the active power of network Amount (Pex) and input variable " feature X " derived from network measurement (Freq) of AC networks.In fault clearance or incident After generation, all measurements are measured with the step-length of 5ms, continue 50ms.Last row provides the correct output state (y) of AC networks, I.e., if there is load feeder line is lost or yet suffers from the connection with AC networks.For the exemplary instruction shown in table 2 Practice example data set, the state of AC networks is estimated as loading the loss of feeder line.These results ' 0' or ' 1' in binary form Represent, wherein ' 1' mean to load the loss of feeder line or be fully disconnected with AC networks, and ' 0' means to yet suffer from With AC network connections.
Another exemplary training dataset is table 3 below shows, wherein incident is gain test, i.e. it is related to It is used as training sample " feature X " (X=[X during staying in gain test0 X1 X2 ... Xn]) measurement.
Table 3
Above-mentioned " feature X " that proposed is related to the measurement during gain test, it is therefore desirable to which certain gain test is opened Begin mark so that for the measurement of input " feature X " can start.For example, user can send defeated when gain test starts Enter as triggering.
In table 3, as in table 1, first row has the individual training sample for using of quantity (m).Next group of row include from Input variable " feature X " derived from voltage (V) and the reactive power exchange measurement with network obtained at the bus of converter (Qex).All measurements are surveyed in the 200ms before and after, during the gain test as emergency event with the step-length of 1ms Amount.Last row gives the correct output state (y) of AC networks.For the exemplary training data collection shown in table 3, AC The state of network is estimated as one in 15 State- outputs.Multicategory classification problem is proposed for the example, its Middle neutral net exports hθX () will predict one in many different classifications.For the task, made using 15 different classes For example.The class of given 15 propositions described below, also discloses the correct option vector for training neutral net “y”。
15 different classes are only an exemplary embodiment, however, as requested or user need, it is also possible to make With the grader of any other higher or lower quantity.First kind output condition is noted as y=[1 0000000 000000 0], the state representation of AC networks is the very strong AC networks with strong voltage control quantity nearby by it.The Two class output conditions are noted as y=[0 1000000000000 0], and it represents that the state of AC networks is tool There are the very strong AC networks of medium voltage control quantity nearby.3rd class output condition is noted as y=[0 010000 0000000 0], the state representation of AC networks is the very strong AC networks with weak voltage control quantity nearby by it. Similarly, for strong, normal intensity, weak and very weak AC networks and it is corresponding strong, neutralize weak voltage control quantity nearby, deposit In the output condition of 12 other classes.
In another embodiment, derived using calculating model of neural networks and assume function hθ(x).Include training sample given In the case of the training dataset of the set of this (X) and observation output (y), neutral net is used for prediction and estimation problem.This Training dataset contributes to setting up a function that can be applicable to future condition, with pre- in the case of given input condition set Survey most probable output.In disclosed method, training dataset, i.e. training sample (X=[XC X1 X2 ... Xn]) and Correspondence output (y) of the state of AC networks is known for training neutral net to assume function hθThe situation of (x).Artificial neuron Network allows multidimensional nonlinear pattern-recognition, as a result provides and predicts more rapidly and accurately.
In one embodiment of the method, calculating model of neural networks includes input layer, at least one hidden layer and defeated Go out layer.Input layer includes input feature value (X=[X0 X1 X2 ... Xn]), hidden layer includes activation unit (a0, a1, ...ak), output layer includes the correspondence output (h of the state of AC networksθ(x)).In calculating model of neural networks, the first weight square Battle array (θ(1)) control input layer input feature value (X=[X0 X1 X2 ... Xn]) to hidden layer activation unit (a0, a1... ak) Function Mapping, and the second weight matrix (θ(2)) control hidden layer activation unit (a0, a1... ak) arrive defeated Go out the correspondence output (h of the state of the AC networks of layerθ(x)) Function Mapping.The target of neutral net is by AC nets with high accuracy The state estimation of network is the weight matrix (θ of neutral net(1), θ(2)) function, i.e. hθ(x) ≈ y, it is adaptable to which all training are shown Example.
There can be multiple hidden layers.In such neutral net, there will be more than two weight matrix (θ(j)).Use Following rule can be readily available the weight matrix (θ of each layer(j)) size.If network has S in layer jjIndividual unit is simultaneously And with S in layer j+1kIndividual unit, then weight matrix θ(j)Will be with size [Sk x Sj+1]。
The result of training neutral net is to find the Function Mapping for controlling the layer j to layer j+1 from distribution over the entire network Inside weight matrix (θ(j)).If neutral net is trained to, these weight matrix (θ(j)) be used to become using input Amount is the state (y) of the AC networks that " feature X " carrys out forecasting system.
In the first example of the preferred embodiment of the method, the structure of neutral net is 40 lists used in input layer Unit (X=[X0 X1 X2 ... X40]) three-layer neural network, 8 unit (a0, a1... a8) in hidden layer, one defeated In going out layer (y).When the state of AC networks is predicted to be isolated island or non-island network condition or load feeder loss, the embodiment It is useful.
In the second example of the preferred embodiment of the method, the structure of neutral net is 400 used in input layer Unit (X=[X0 X1 X2 ... X400]) three-layer neural network, 25 unit (a0, a1... a25) in hidden layer, 15 Unit is in output layer (y).When the state of AC networks is predicted to be one in foregoing 15 classifications, the enforcement Example is useful, including intensity and the different changes of the neighbouring voltage control quantity for existing of prompting AC networks.
Element number in the quantity and hidden layer of layer can change.This by train neutral net needed for precision and Computing capability produces impact.In above preferred embodiment propose structure prove be to a great extent it is accurate, without Substantial amounts of computing capability is wanted to train.Other neural network structures can be used to improve precision.
In another embodiment, the method also includes using Sigmoid functionsTo derive hypothesis function (hθ(x)) the step of, wherein hθ(x)=g [θ(2){g(θ(1)X) }], wherein iteratively adjusting weight matrix (θ(1), θ(2)) come Determine weight matrix (θ(1), θ(2)), so that cost function (J (θ)) is minimized.Sigmoid functions (g (z)) are used as in hidden layer Activation primitive, it is used for the calculating of the output of neutral net, i.e. (hθ(x)).Any other activation function is also possible.Make Formed as activation primitive with Sigmoid functions (g (z)) and assume function (hθ(x)), wherein hθ(x)=g [θ(2){g(θ(1)x)}]。 Given one group of input, i.e. input variable " feature X "=[X0 X1 X2 ... Xn]∈Rn+1, and for each layer (θ(1), θ(2)) Given weight matrix (θ(j)), the estimated state (h of the output of neutral net, i.e. AC networksθ(x)), it is possible to use hθ(x)=g [θ(2){g(θ(1)X) }] calculating.Cost function (J (θ)) is given：
Wherein ' m' be training sample number, ' λ ' is the regularization parameter for avoiding over-fitting, ' L' is layer in network Sum, s1Be layer ' in 1' unit number, if ' 1' tools have three layers, ' 1' can be 1,2 or 3 wherein the layer of neutral net.
In yet another embodiment, weight matrix (θ is iteratively adjusted by using back-propagation algorithm(1), θ(2)) come Determine weight matrix (θ(1), θ(2)), to minimize cost function, J (θ).Once neutral net is trained to, then these weight squares Battle array θ(j)Be used to only predict the state of AC networks using input " feature X ".Enabled the method to using back-propagation algorithm Rapidly and easyly obtain weight matrix θ(j)Optimal possible estimation.However, there are other calculations that can be used for identical purpose Method.Another example of this algorithm is conjugate gradient algorithms.
In another embodiment, cost function J (θ) is in the AC networks for estimating to be obtained from learning algorithm (h θ (x)) The tolerance of the error in state, the error is that output (y) corresponding with the state of AC networks is compared.Cost function J (θ) accurately assume (h to obtain by training learning algorithmθ(x)), contribute to making method more accurately predict the defeated of AC networks Do well.
Description of the drawings
Above and other feature of the present invention is solved referring now to the accompanying drawing of the present invention.Illustrated embodiment is intended to explanation And the unrestricted present invention.Accompanying drawing includes the following drawings, wherein identical reference represents phase in entire disclosure and accompanying drawing Same part.
Fig. 1 shows the schematic diagram of the HVDC transmission system with the converter station equipped with power converter.
Fig. 2 shows the schematic diagram of non-island network condition.
Fig. 3 is shown comprising the flow chart according to the step of disclosed method.
Fig. 4 depicts the block diagram of the controller of power converter and power converter.
Fig. 5 shows the Exemplary neural network computing model according to disclosed method.
Fig. 6 shows the another exemplary calculating model of neural networks according to disclosed method.
Specific embodiment
As shown in figure 1, AC networks or electrical network 4 are connected to converter station 6 via transformer 5.Converter bus 3 is by converter Stand and 6 be connected to its corresponding AC network 4, and AC networks 4 are connected to bus 11 transformer 5 of sending side.Converter station 6 exists Sending side is connected to the converter station 6 of receiving side by DC transmission lines 7.Power be transmitted through DC transmission lines 7 greatly away from From above carrying out.Converter station 6 includes the (not shown in figure 1) of electric power converter 1.
In the embodiment shown, obtaining from the bus 11 being connected to needed for AC networks 4 will be used as training example (X=[X0 X1 X2...Xn]) one of input voltage measurement (V).From the voltage measurement and electric current that for example carry out at required bus 11 Measurement, further derives and such as measures (Q with the reactive power exchange of AC networksex) exchange measurement with the active power of AC networks (Pex) etc other input.
Fig. 2 is the schematic diagram of the state of the AC networks 4 for being shown as the non-island network condition for HVDC systems.
According to Fig. 2, the big AC being made up of the high-voltage transmission network including generating and middle voltage and low voltage distribution network Network 8 is connected to AC generator units 9 by power line 10.Big AC networks 8 form AC nets together with AC generator units 9 in sending side Network or grid 4.The AC networks 4 are connected to converter station 6 via transformer 5.In receiving side, there is inverter electrical network 12.
Converter station 6 is connected to AC networks 4 by converter bus 3, and AC networks 4 are connected to sending side by bus 11 Transformer 5.Sending side be system from its produce AC power with by DC transmission lines 7 in the side of DC transmission over networks.Send The converter station 6 of side is connected to the converter station 6 of receiving side via DC transmission lines 7.Required bus 11 is by big AC electrical networks 8 and AC Generator unit 9 is connected to the transformer 5 of sending side.The DC electric power of transmission is from AC generator units 9 and big AC networks 8.
For example, when the sub-fraction of system and the remainder of AC networks are disconnected, there is island network condition.Root According to Fig. 2, little shielding system is HVDC rectifier converters station 6 and transformer 5, converter bus 3 and generator unit 9.
In the case of island network condition, power line 10 trips and little AC systems, i.e. HVDC rectifier converters station 5th, 3,6 and generator unit 9 disconnect from big AC networks 8 and isolate with the remainder of AC networks.After power line 10 trips, The AC electric power for carrying out arrogant AC networks 8 is no longer present, so HVDC systems must be adjusted quickly from sending side is transferred to receiving side DC electric power, to match the power of the generation of three AC generator units 9.
Such case only shows the exemplary form how isolated island network condition may occur, but is formed Many other different modes of island shape network condition are possible.
Fig. 3 shows the step of estimation is connected to the state of the AC networks 4 of power converter 1 (referring to Fig. 1) to be performed. Method 100 includes providing the first step 101 of training dataset, and the training dataset includes training sample set (X=[X0 X1 X2 ... Xn]) and AC networks 4 state correspondence output (y).Method 100 includes training hypothesis letter based on training dataset Number (hθ(x)) second step 102.Hereafter, perform using hypothesis function (hθ(x)) estimating the 3rd step of the state of AC networks 4 Rapid 103.
The method 100 presented in the present invention allows quickly and correctly to estimate the state of AC networks 4.
Fig. 4 shows the block diagram of power converter 1 and its controller 2.Each power converter 1 has its corresponding control Device 2, by the controller power converter 1 is controlled.The estimated state of AC networks output is (hθ(x)), for changing or optimizing The parameter of controller 2, to realize better performance during system dynamic, for example, rise time, stabilization time, maximum overshoot Deng.
Including one group of training sample (X=[X0 X1 X2 ... Xn]) training dataset be from voltage measurement (V) and/or Reactive power exchange measures (Qex) and other it is previously mentioned to AC networks 4 before incident, during or after carry out Measurement.And these measurements are obtained from the bus 11 being connected to needed for power converter 1.
Based on training dataset, it is assumed that function (hθ(x)) learn to predict the state of AC networks 4 exactly.System busbar The intensity of change for example after the generation of island network condition, may negatively affect immediately and be after system incident The performance of system.The method 100 will ensure that system adapts to new system condition and or even provides after system incident and change The dynamic property entered.
Fig. 5 shows the Exemplary neural network computing model 13 according to disclosed method.In the exemplary model 13 In, the structure of neutral net is three-layer neural network.Ground floor is to include n+1 unit (X=[X0 X1 X2 ... Xn]) Input layer 14, including bias unit X0.The second layer is to use k+1 unit (a0, a1... ak) hidden layer 15, including basis The bias unit a of Fig. 50.The unit of hidden layer 15 also referred to as activates unit.And third layer is with least one unit (hθ (x)) output layer 16.For the particular exemplary calculating model of neural networks, output layer 16 only has 1 list according to Fig. 5 Unit.
(the θ of first weight matrix 17(1)) control input layer 14 input variable set (X=[X0 X1 X2 ... Xn]) arrive Activation unit (a of hidden layer 150, a1... ak) Function Mapping, (the θ of the second weight matrix 18(2)) control swashing for hidden layer 15 Unit (a living0, a1... ak) to output layer 16 AC networks corresponding states (hθ(x)) Function Mapping.Fig. 5 illustrate only god The example of Jing network models 13, but there are other neural computing moulds of different number of unit and the different numbers of plies in per layer Type 13 can also be created for identical purpose.
Fig. 6 illustrates another the such Exemplary neural network computing model 13 according to disclosed method.Show at this In example property model 13, the structure of neutral net is also three-layer neural network.According to Fig. 6, it is proposed that the structure conduct of neutral net The 400+1 unit (X=[X used in input layer 140 X1 X2...X400]) three-layer neural network model, 25+1 unit (a0, a1... a25) in hidden layer 15, and 15 unit (C1 C2...C15) in output layer 16.
Although the invention has been described with reference to specific embodiments, but the description is not meant to solve in a limiting sense Release.When with reference to description of the invention, the various modifications of the disclosed embodiments and the alternate embodiment of the present invention are for this Art personnel will become clear from.Therefore, it is contemplated that can be in the situation without departing from defined embodiments of the invention Under modify.
Reference numerals list
1 power converter
2 controllers
3 buses
4 AC networks
5 transformers
6 converter station
7 DC transmission lines
8 big AC networks
9 AC generator units
10 power lines
Bus needed for 11
13 calculating model of neural networks
14 input layers
15 hidden layers
16 output layers
17 first weight matrix θ(1)
18 second weight matrix θ(2)
100 methods
The step of 101 offer training dataset
Function (h is assumed in 102 trainingθ(x)) the step of
The step of state of 103 estimation AC networks

Claims (15)

1. one kind is connected to the method (100) of the state of the AC networks (4) of power converter (1), methods described for estimation (100) include：
The step of-offer training dataset (101), the training dataset includes training sample set (X=[X0 X1 X2 ... Xn]) and the AC networks (4) state correspondence output (y)；
- hypothesis function (h is trained based on the training datasetθ(x)) the step of (102)；
- use the hypothesis function (hθ(x)) come (103) the step of the state for estimating the AC networks (4).
2. method (100) according to claim 1, wherein, methods described (100) is also included based on the estimated AC The step of state of network (4) is come controller (2) parameter for adjusting the power converter (1).
3. method (100) according to any one of the claims, wherein, the state of the AC networks (4) is one Or the intensity of multiple isolated islands or non-island network condition, load feeder loss or the AC networks with neighbouring voltage control quantity.
4. method (100) according to any one of the claims, wherein, the training sample set (X=[X0 X1 X2 ... Xn]) one or more derivation from following：The reactive power exchange measurement of voltage measurement (V) and AC networks (4) (Qex) exchange measurement (P with the active power of AC networks (4)ex) and AC networks (4) network measurement (Freq).
5. method (100) according to claim 4, wherein, measured before, during and/or after system incident Training sample set (X=[the X0 X1 X2 ... Xn])。
6. method (100) according to claim 5, wherein the system incident is fault clearance event or gain surveying One in examination.
7. method (100) according to any one of the claims 5 to 6, wherein, before system incident, the phase Between and afterwards first time period measurement voltage measurement (V), the step of second time period in by the voltage measurement introduce AC Network (4).
8. method (100) according to any one of claim 5 to 7, wherein, before the system incident, the phase Between and first time period afterwards measure the network and measure (Freq), the step of second time period in by the network Frequency measurement introduces AC networks.
9. method (100) according to any one of claim 5 to 8, wherein, before the system incident, the phase Between and afterwards first time period measurement exchange measurement (P with the active power of the AC networks (4)ex), and in the second time AC networks (4) are introduced in the step of section.
10. method (100) according to any one of claim 5 to 9, wherein, before system incident, period and First time period afterwards is measured and measures (Q with the reactive power exchange of the AC networks (4)ex), and in second time period AC networks (4) are introduced in step.
11. methods (100) according to any one of the claims, wherein, using calculating model of neural networks (13) Derive the hypothesis function (hθ(x))。
12. methods (100) according to the next item up claim, wherein, the calculating model of neural networks (13) includes：
- input layer (14), including the training dataset (X=[X0 X1 X2 ... Xn])；
- at least one hidden layer (15), including activation unit (a0, a1... ak)；With
- output layer (16), it includes the correspondence output (h of the state of the AC networksθ(x))；
Wherein the first (17) weight matrix (θ(1)) the control input layer (14) training dataset (X=[X0 X1 X2 ... Xn]) to the hidden layer (15) activation unit (a0, a1... ak) Function Mapping, and the second (18) weight matrix (θ(2)) the control hidden layer (15) activation unit (a0, a1... ak) to the output layer (16) AC networks state Correspondence output (hθ(x)) Function Mapping.
13. methods (100) according to the next item up claim, wherein, methods described (100) also includes using Sigmoid Function
g ( z ) = 1 1 + e - z
To derive the hypothesis function (hθ(x)) the step of, wherein
hθ(x)=g [θ(2){g(θ(1)X) }],
Wherein iteratively adjust weight matrix (θ (1), θ(2)) (17,18) determining the weight matrix (θ(1), θ(2)) (17, 18) so that cost function (J (θ)) is minimized.
14. methods (100) according to the next item up claim, wherein, iteratively adjust by using back-propagation algorithm Weight matrix (the θ(1), θ(2)) (17,18) determining the weight matrix (θ(1), θ(2)) (17,18).
15. methods (100) according to any one of claim 10 to 12, wherein, the cost function (J (θ)) is to estimate Count from the hypothesis function (hθ(x)) obtain AC networks (4) state error tolerance, the error is compared to AC nets For correspondence output (y) of the state of network (4).
CN201480081138.7A 2014-08-12 2014-08-12 Method for estimating status of ac networks and subsequent adaptive control Pending CN106662846A (en)

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/EP2014/067256 WO2016023579A1 (en) 2014-08-12 2014-08-12 Method for estimating status of ac networks and subsequent adaptive control

Publications (1)

Publication Number Publication Date
CN106662846A true CN106662846A (en) 2017-05-10

Family

ID=51429253

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201480081138.7A Pending CN106662846A (en) 2014-08-12 2014-08-12 Method for estimating status of ac networks and subsequent adaptive control

Country Status (2)

Country Link
CN (1) CN106662846A (en)
WO (1) WO2016023579A1 (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113872198A (en) * 2021-09-29 2021-12-31 电子科技大学 Active power distribution network fault recovery method based on reinforcement learning method

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3588724A1 (en) * 2018-06-26 2020-01-01 CSEM Centre Suisse D'electronique Et De Microtechnique SA Method of operating a power distribution system
CN114705947A (en) * 2022-03-16 2022-07-05 广东电网有限责任公司 Island detection model training method, device, equipment and medium

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102237691A (en) * 2011-07-06 2011-11-09 东北大学 Wind energy and solar energy grid-connected generation system and control method thereof
CN102549872A (en) * 2009-06-18 2012-07-04 Abb技术有限公司 Controlling an inverter device of a high voltage DC system for supporting an AC system
CN102611140A (en) * 2012-03-23 2012-07-25 合肥工业大学 Grid-connected inverter island detection method based on wavelet transform and neural network
US20130274946A1 (en) * 2012-04-13 2013-10-17 Owen Jannis Schelenz Methods and systems for controlling a power plant

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102549872A (en) * 2009-06-18 2012-07-04 Abb技术有限公司 Controlling an inverter device of a high voltage DC system for supporting an AC system
CN102237691A (en) * 2011-07-06 2011-11-09 东北大学 Wind energy and solar energy grid-connected generation system and control method thereof
CN102611140A (en) * 2012-03-23 2012-07-25 合肥工业大学 Grid-connected inverter island detection method based on wavelet transform and neural network
US20130274946A1 (en) * 2012-04-13 2013-10-17 Owen Jannis Schelenz Methods and systems for controlling a power plant

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
OMAR N. FAQHRULDIN，等: "A universal islanding detection technique for distributed generation using pattern recognition", 《IEEE TRANSACTIONS ON SMART GRID》 *
SHUHUI LI，等: "Artificial neural networks for control of a grid-connected rectifier/inverter under disturbance,dynamic and power converter switching conditions", 《IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS》 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113872198A (en) * 2021-09-29 2021-12-31 电子科技大学 Active power distribution network fault recovery method based on reinforcement learning method
CN113872198B (en) * 2021-09-29 2022-09-13 电子科技大学 Active power distribution network fault recovery method based on reinforcement learning method

Also Published As

Publication number Publication date
WO2016023579A1 (en) 2016-02-18

Similar Documents

Publication Publication Date Title
CN103795058B (en) The air extract analysis of electric power system and system failure sort method
CN101373905B (en) Visualization method for voltage stability domain of electric power system
Tayeb Faults detection in power systems using artificial neural network
CN106909989A (en) A kind of grid disturbance Forecasting Methodology and device
Hussein et al. Robust artificial neural network-based models for accurate surface temperature estimation of batteries
CN106383315A (en) New energy automobile battery state of charge (SOC) prediction method
Kamalasadan et al. A new intelligent algorithm for online voltage stability assessment and monitoring
CN106384186A (en) Distributed new energy grid-connected power quality monitoring method and system
CN104638671A (en) Island detection method based on introducing particle swarm optimization and phase disturbance
CN106662846A (en) Method for estimating status of ac networks and subsequent adaptive control
CN102930408B (en) A kind of 750kV electric grid secondary equipment state appraisal procedure based on information fusion
Krishna et al. Transient stability assessment using artificial neural networks
Zhukov et al. A hybrid artificial neural network for voltage security evaluation in a power system
Gupta et al. PMU-ANN based real time monitoring of power system electromechanical oscillations
Jadidi et al. Design of an intelligent hybrid diagnosis scheme for cyber-physical PV systems at the microgrid level
CN102087311B (en) Method for improving measurement accuracy of power mutual inductor
Naderi et al. Detection of false data injection cyberattacks: Experimental validation on a lab-scale microgrid
Mahto et al. Data driven approach for optimal power flow in distribution network
CN105242177B (en) A kind of island detection method based on System Discrimination
CN104578108B (en) Reactive power optimization fine on-line control method with load fluctuation characteristics considered
CN114111892A (en) Distribution transformer monitoring and early warning device and method with uncertain information fusion
CN105424224A (en) Isolating switch state monitoring method and device
Liang et al. Standalone condition diagnosing of fuel cell in microgrid composed of wind turbine/fuel cell/combined heat & power using Variational Mode Decomposition analysis model
Belkacemi et al. Real-time intelligent control and cascading failure prevention in microgrid systems based on neural network algorithm: an experimental approach
KR20220082234A (en) Method and apparatus for diagnosing battery status through soc estimation

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20170510