WO2016023579A1

WO2016023579A1 - Method for estimating status of ac networks and subsequent adaptive control

Info

Publication number: WO2016023579A1
Application number: PCT/EP2014/067256
Authority: WO
Inventors: Alvaro Jose Hernandez Manchola; Frank Schettler; Jürgen LOTTES; Manuel STEGER
Original assignee: Siemens Aktiengesellschaft
Priority date: 2014-08-12
Filing date: 2014-08-12
Publication date: 2016-02-18
Also published as: CN106662846A

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 ··· X_n] ) and a corresponding output of the status of the AC network (y), a step of training a hypothesis function (h_θ(x)) based on the training dataset and finally a step of estimating the status of the AC network using the hypothesis function (h_θ(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

Method for estimating status of AC networks and subsequent adaptive control

The present invention relates to a method for estimating the status of AC networks and more particularly a method for es^¬ timating the status of AC networks and based on the status of the AC network, enabling an adaptive control.

Having an estimate about the status of an AC network is very important for the safe and stable operation of large power converters and for the power system in general. The status of the AC network means, in this case, existence of either an islanded or a non-islanded network condition, the loss of the last feeder or the strength of the AC network and the amount of nearby voltage controllers. Once the status of the AC net^¬ work is known, an adaptive control process can be set off, for example the controller parameters associated with the power converter can be adjusted to achieve improved perfor^¬ mance behavior in the power converter. For example, for a weak AC network with weak amount of nearby voltage control the controller parameters are to be adjusted in one way and for a strong AC network with strong amount of nearby voltage control the controller parameters are to be adjusted in an^¬ other way.

In an islanded network condition a small part of the system gets isolated from the rest of the AC network. This can hap- pen due to tripping of a line, or failure of a transformer and other such similar contingencies. The occurrence of an islanded network condition, is a potentially critical situa^¬ tion which can cause dangerous stresses on the electrical equipment of the AC network, cause system instability and can potentially lead to power outages or black outs.

The loss of the last feeder is applicable to power electronic converters that, due to a system contingency, are completely isolated or disconnected from the AC network. The system con^¬ tingency can be, for example, an inadvertent tripping of an AC line or a transformer. The loss of the last feeder is a potentially critical situation which can cause dangerous stresses on the electrical equipment within the power elec^¬ tronic converter.

Not having a fast and a reliable system for estimating the status of the AC network which is able to identify an island or loss of the last feeder condition within some milliseconds after its occurrence can be the cause of critical stresses on electrical equipment. Moreover not being able to accurately estimate the strength of the AC network and the amount of nearby voltage control can result in below optimal control actions taken by power electronic devices, such as SVC, HVDC, or wind turbine controls, leading to system instability and power outages.

Until now, in case of fast dynamic power electronic control- lers like LCC and VSC HVDC, an external trigger signal is re^¬ quired from the network operator identifying the islanded network condition, in order to take the appropriate control actions in presence of an islanded electrical network condi^¬ tions. Accurately predicting an islanded or non-islanded net- work condition very quickly, say within 50 to 100 milliseconds of its occurrence, is not known in the state of the art.

Having an estimate about the status of the AC network, one can adjust the controller parameters, for example by adjust- ing gains, time constants, etc, depending on the strength of the system and the presence of nearby voltage controllers and therefore improve the controller' s ability to adapt to new system conditions. Automatic gain adjustment for power electronic converters is an increasingly important topic which has been only partially addressed. It allows converters, e.g. Static Var compensator or STATCOM, to adjust their controller parameters to provide better performance during system dynamics.

Automatic gain adjustment has so far been implemented based on measurements started during a steady state condition. A gain test basically consists of measuring the voltage change (dV) in a bus bar due to the injection or absorption of a certain amount of reactive power (dQ) from the converter. The change in voltage with respect to the change of reactive pow- er (dv/dq) is then used to estimate the strength of the bus bar connected to the converter. The strength of the bus bar is defined by its sensitivity to voltage and frequency chang^¬ es when reactive and active power are injected or absorbed in that bus bar. A strong bus bar is less prone to voltage and frequency changes compared to a weak bus bar.

Another gain test method, in the absence of power injection or absorption from the power converter, is to switch in or out AC filters, shunt capacitors or reactors, and measure again the ratio of change in voltage over reactive power (dv/dq) .

With an increasing amount of voltage controlling power electronic converters in AC networks, one of the big challenges so far is to have a proper controller parameter selection, e.g. gain, which takes into account the presence of "other" voltage controlling equipment in the vicinity.

Not having a reliable estimate of the system strength and the amount of nearby voltage controllers is the cause of below optimal performance behavior, e.g. rise time, settling time maximum overshoot, insufficiently damped oscillations etc., in power electronic driven converters. Until now, gain tests as described above are performed to es^¬ timate the strength of the bus bar the converter is connected to, however if there are nearby voltage controlling equip^¬ ment, e.g. other fast dynamic compensation, the result of the gain test will be misleading. For example, a very weak bus bar, with a big amount of nearby voltage controllers might seem like a strong bus bar to the gain test, resulting in completely inappropriate controller parameter adjustments.

Therefore, there is a growing need for a method that accu^¬ rately estimates the status of the AC network. Once the sta^¬ tus of the AC network is estimated the same information can be used for accordingly changing the controller parameters of the power converter to achieve better system performance.

It is an object of the present invention to quickly predict the status of the AC network, e.g. whether an islanded or a non-islanded network condition exists, whether a loss of the last feeder has occurred and what is the strength of the AC network and the amount of nearby voltage control present, and accordingly adjust the controller parameters to achieve opti^¬ mum system performance. The object is achieved by a method for estimating status of an AC network connected to a power converter comprising a step of providing a training dataset comprising a set of training examples (X=[Xo Xi X2 ··· Xn ] ) and a corresponding output of the status of the AC network (y) , a step of training a hypothesis function (h_e(x)) based on the training dataset and finally a step of estimating the status of the AC network using the hypothesis function (h_e(x)) .

The use of a trained hypothesis function (h_e(x)) is a form of a supervised learning method. Supervised learning is a ma^¬ chine learning task of inferring a function from labeled training data. The training data consists of training examples. The training examples consist of an input vector ("fea^¬ tures X") and its associated correct output value (y) . The task of a supervised learning algorithm is to analyze the training data and come up with a hypothesis function h_e (x) , which can be used for mapping new examples, i.e h_e (x) ~y for all training examples. In this invention the task of the supervised learning method is, based on a training set made up of a selected input vec^¬ tor of "features X" and their corresponding output y, i.e. correct answer, to create a function h_e (x) that quickly and accurately estimates the status of the AC network, whether it is an islanded network condition or a non-islanded network condition, whether a loss of the last feeder has taken place, and the strength of the AC network as well as the amount of voltage controllers in the vicinity.

One of the challenges of employing artificial intelligence algorithms to quickly obtain an accurate estimate of the sta^¬ tus of the AC network and the network strength is to identify the appropriate "features X" to use as an input to the algo- rithms. For this invention a neural network computational model is proposed.

Neural network's ability to generalize and learn from the training dataset mimics, in some sense, human's ability to learn from experience. Neural networks are used for predic^¬ tion and estimation problems. For a problem to be solved us^¬ ing neural networks, inputs that are well understood are needed. A good idea of which features are important for pre^¬ dicting the correct output is required. Such inputs may be easily available, but how to combine them to obtain an accu^¬ rate estimation is not clear. The next requirement is to have outputs that are well understood too, i.e. information about the kind of output that is desired to be estimated, predicted or modeled. The next factor is the use of the experience that is available. For training the neural network we have samples of training set which have been obtained by experience. In these sample datasets both the inputs, feature vectors ("fea^¬ tures X") and the outputs (y) , are known cases that are used to train the neural network.

The method further comprises a step of adjusting controller parameters of the power converter based on the status of the AC network estimated. For example, for an islanded network condition the controller parameters are to be adjusted in one way and for a non-islanded network condition the controller parameters are to be adjusted in another way. Similarly, for a weak AC network which is more prone to voltage and frequen- cy changes with weak amount of nearby voltage control the controller parameters are to be adjusted in one way and for a strong AC network which is less prone to voltage and frequency changes with strong amount of nearby voltage control the controller parameters are to be adjusted in another way.

For example, the voltage controller gain of an SVC or statcom for a strong AC network needs to be higher than the voltage controller gain of a weak AC network. In an embodiment the status of the AC network is one of an islanded or a non-islanded network condition, loss of last feeder or strength of the AC network with amount of nearby voltage control. Recognizing an islanded network condition or loss of last feeder within some milliseconds after its occur- rence can prevent critical stresses in electrical equipment. It can also help power electronic devices to take optimal control actions and thereby prevent system instability and power outages from occurring. The supervised learning method presented in this invention, allows for fast and accurate AC network status recognition, even in presence of highly nonlinear power electronic equipments like HVDC LCC converters, VSC HVDC, wind turbines and STATCOM/SVC.

As mentioned earlier, in an islanded network condition a small part of the system gets isolated from the rest of the AC network. This can happen due to tripping of a line, or failure of a transformer and other such similar contingencies. The occurrence of an islanded network condition, is a potentially critical situation which can cause dangerous stresses on the electrical equipment of the AC network, cause system instability and can potentially lead to power outages or black outs. The loss of the last feeder is applicable to power elec^¬ tronic converters that, due to a system contingency, are com^¬ pletely isolated or disconnected from the AC network. The system contingency can be, for example, an inadvertent trip- ping of an AC line or a transformer. The loss of the last feeder is a potentially critical situation which can cause dangerous stresses on the electrical equipment within the power electronic converter. In another embodiment the set of training examples (X=[Xo Xi X2 ... X_n] ) is derived from one or more of voltage measurements (V) , reactive power exchange with network measurements (Q_ex) , active power exchange with network measurements (P_ex) and network frequency measurements (Freq) of the AC network.

For this application of neural networks, i.e. for estimation of the status of the AC network, this invention proposes for example the above mentioned measurements as features to be used as an input to the neural network. Any other suitable measurements may also be used. The voltage measurements are taken at a desired bus bar of the network. The same voltage measurement as the one used for the converter voltage con^¬ troller is appropriate. Positive sequence fundamental fre^¬ quency voltage measurement may be used as an alternative. The reactive power exchange with the network measurements (Q_ex) , the active power exchange with the network measurements (P_ex) as well as the network frequency measurements are also meas^¬ ured. These measurements can be taken at the desired bus bar of the network or at the grid's power line.

In a further embodiment the set of training examples (X=[X0 XI X2 ... Xn] ) are measured before, during and/or after a system contingency event. Thus, for the system contingency event to be detected some form of fault detection is required so that once the fault is cleared the measurements required for the input "features X", i.e. the training examples, can start. One such fault detection method can be the use of un- der-voltage detection to flag a fault incident or contingency event in the AC network and trigger the neural network Ac network status estimation process. For instance an under- voltage of 0.8 V may be used as a trigger. Other factors that can be used for detecting a fault can be over-voltage, over- current, under-current, or other similar monitorable condi^¬ tions. Under-voltage condition provides for a simple and easy fault detection method.

In an embodiment the system contingency event is a fault clearing event. In yet another embodiment the system contingency event is a gain test, wherein reactive power is inject^¬ ed or absorbed into the AC network from the power converter. However, any other methods of gain test can also be used, for example, in the absence of power injection or absorption from the power converter, AC filters or reactors can be switched in or out and the corresponding change in voltage over power (dv/dq) is measured.

In an embodiment the voltage measurements (V) are measured for a first time period spanning before, during and after the system contingency event, induced into the AC network, in steps of a second time period. In another embodiment of the method the network frequency measurements (Freq) are measured for a first time period spanning before, during and after the system contingency event, induced into the AC network, in steps of a second time period. In yet another embodiment the active power exchange measurements (P_ex) are measured for a first time period spanning before, during and after the sys^¬ tem contingency event, induced into the AC network, in steps of a second time period. In another embodiment the reactive power exchange measurements (Q_ex) are measured for a first time period spanning before, during and after the system contingency event, induced into the AC network, in steps of a second time period.

The first time period corresponds to the time period for which the measurements are taken and the second time period corresponds to the time step length between each of the meas^¬ urements .

In an exemplary embodiment, when the system contingency event is a fault clearing event, the first time period spans for 50ms and the second time period is 5ms. The measurements are started at the desired bus bar in the first 50ms after the fault clearing event occurs and measurements are taken in 5ms steps .

In another exemplary embodiment, when the system contingency event is a gain test, the first time period spans for 200ms and the second time period is 1ms. The measurements are started at the desired bus bar of the converter for a span of 200ms, which covers the duration before, during and after the gain test is done, in steps of 1ms.

Due to the nature of the features proposed for this applica^¬ tion of neural networks, the time step length between meas- urements, e.g. proposed 5ms or 1ms, and the length of the overall measurement, e.g. proposed 50ms or 200ms, influence the number of inputs "features X" of the neural network.

The length of the measurements, i.e. the first time period, used for the "features" definition influences the accuracy of the estimation. The longer the measurements used the more ac^¬ curate is the prediction because the algorithm has more in^¬ formation to base its prediction on. Other measuring times can be used if they show better performance, what is im- portant is that the accuracy of the prediction is acceptable, and that the prediction is available in time to take effec^¬ tive control actions.

The time steps used for the "features" definition, i.e. the second time period, influences the estimation accuracy. Other time steps can be used if they show better performance, what is important is that the accuracy of the predictions is ac^¬ ceptable. Taking measurements in small steps of 5ms or 1ms, for example, will provide better and a finer training dataset for training the hypothesis function h_e(x) .

Table 1 below shows an exemplary training dataset of "fea- tures X" assuming we have ^xm' training examples each of them made up of an input vector or training examples (X=[Xo Xi X2 ··· X_n] ) and the corresponding output of the status of the AC network (y) available for the method using neural network computation model for estimation of the status of the AC net- work using the hypothesis function (h_e(x)) . In this case the vector X is input variables given by "features X"=[Xo Xi X2 ··· X40] e R⁴⁰⁺¹. The proposed "features X" involves measurements taken after the contingency event, for example a fault clear^¬ ing event .

TABLE 1

The first column has the number (m) of training samples used. The next set of columns contain the input variables "features X" which are derived from voltage (V) , reactive power ex^¬ change with network measurements (Q_ex) , active power exchange with network measurements (P_ex) and network frequency meas^¬ urements (Freq) of the AC network. All measurements are meas^¬ ured in steps of 5ms for 50ms after the clearance of the fault or occurrence of the contingency event. The last column gives the correct output status of an AC network (y) . For this exemplary training dataset shown in the Table 1, the status of the AC network is estimated as an islanded network condition or a non-islanded network condition. These outcomes are represented in a binary form of ^λ0' or ^Λ1' where ^Λ1' im^¬ plies an islanded network condition and a ^λ0' implies a non- islanded network condition as the status of the AC network.

In this example the Table 1 provides only one sample training dataset used for training the hypothesis function h_e(x) . The above table is just a sample dataset. However, for different networks different samples can be used for forming the corre^¬ sponding dataset.

TABLE 2

Table 2 below shows yet another exemplary training dataset. The first column has the number (m) of training samples used. The next set of columns contain the input variables "features X" which are derived from voltage (V) , reactive power ex^¬ change with network measurements (Q_ex) , active power exchange with network measurements (P_ex) and network frequency meas^¬ urements (Freq) of the AC network. All measurements are meas^¬ ured in steps of 5ms for 50ms after the clearance of the fault or occurrence of the contingency event. The last column gives the correct output status of an AC network (y) , i.e. whether there is a loss of the last feeder or is a connection to the AC network still existing. For this exemplary training example dataset shown in the Table 2, the status of the AC network is estimated as the loss of the last feeder. These outcomes are represented in a binary form of ^λ0' or ^Λ1' where Λ1' implies a loss of the last feeder, or complete disconnec- tion from the AC network and a ^λ0' implies that there still exists a connection with the AC network.

Table 3 below shows yet another exemplary training dataset where the contingency event is a gain test, i.e. it involves measurements to be used as training examples "features X" (X=[Xo Xi X2 ··· Xn ] ) taken during a gain test.

TABLE 3

The proposed "features X" described above involve measure^¬ ments during the gain test, therefore some kind of gain test start flag is required so that the measurements for the input "features X" can start. For example, a user can send an input as a trigger at the start of a gain test.

In Table 3, just as in Table 1, the first column has the num^¬ ber (m) of training samples used. The next set of columns contain the input variables "features X" which are derived from voltage (V) and reactive power exchange with network measurements (Q_ex) taken at the bus bar of the converter. All measurements are measured in steps of 1ms for 200ms spanning before, during and after the gain test as the contingency event. The last column gives the correct output status of an AC network (y) . For this exemplary training dataset shown in the Table 3, the status of the AC network is estimated as one from fifteen status outputs. A multi-class classification problem is proposed for this example, wherein the neural net^¬ work output, he (x) , will predict one of many different clas- ses. For this task fifteen different classes are used as an example. Given the fifteen proposed classes described hereaf^¬ ter, the correct answer vector "y" used to train the neural network is also disclosed. The fifteen different classes is just one exemplary embodi^¬ ment, however any other number of classifiers, higher or lower, can also be used depending on requirements or user needs. The first class of output condition is annotated as y= [ 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0] which represents the status of the AC network as a very strong AC network with strong amount of nearby voltage control. The second class of output condition is annotated as y= [0 1 0 0 0 0 0 0 0 0 0 0 0 0 0] which rep^¬ resents the status of the AC network as a very strong AC net^¬ work with medium amount of nearby voltage control. The third class of output condition is annotated as y= [ 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0] which represents the status of the AC network as a very strong AC network with weak amount of nearby voltage control. Similarly there are twelve other classes of out^¬ put condition for strong, normal strength, weak, and very weak AC networks and corresponding strong, medium and weak amount of nearby voltage control.

In a further embodiment the hypothesis function (h_e(x)) is deduced using a neural network computational model. Neural networks are used for prediction and estimation problems giv^¬ en a set of training dataset comprising a set of training examples (X) and observed outputs (y) . This training

datasethelps in the establishment of a function that can be applied to future conditions to predict the most probable output given the set of input conditions. In the disclosed method, the training dataset, i.e. the training examples (X=[Xo Xi X2 ··· Xn ] ) and the corresponding output of the status of the AC network (y) are known cases that are used to train the neural network hypothesis function (h_e(x)). Artificial Neural Networks allow for a multi-dimensional non-linear pat^¬ tern recognition which as a result provides faster and accurate predictions. In an embodiment of the method, the neural network computa^¬ tional model comprises an input layer, at least one hidden layer and an output layer. The input layer comprises input feature vector (X=[Xo Xi X2 ··· n ] ) the hidden layer comprises activation units (ao, ai, ... a_k) and the output layer comprises the corresponding output of the status of the AC network h_e (x) . In the neural network computational model a first weight matrix (θ⁽¹⁾) controls function mapping of the input feature vector (X=[X₀ Xi X2 ··· X_n ] ) of the input layer to the ac- tivation units (ao, ai, ... a_k) of the hidden layer and a second weight matrix (θ^<2)) controls function mapping of the activa^¬ tion units (ao, ai, ... a_k) of the hidden layer to the corre^¬ sponding output of the status of the AC network (h_e (x) ) of the output layer. The goal of the neural network is to esti- mate the status of the AC network as a function of the neural network's weight matrices (θ⁽¹⁾, θ^<2)) with high accuracy, i.e h_e (x) ¾y, for all training examples.

There can be more than one hidden layer. In such a neural network there would be more than two weight matrices (θ^<3)) . The dimension of the matrix of weights (θ^<3)) for each layer can be easily obtained, using the following rule. If a net^¬ work has S units in layer j and S_k units in layer j+1, then the weight matrix 6^<3) will be of dimension [S_k x S_j+i] .

The result of training the neural network is finding the in^¬ ternal weights matrices (θ^<3)) which control function mapping from layer j to layer j+1 distributed throughout the network. Once the neural network is trained these weights matrices (9^<j)) are used to predict the status of the AC network (y) of the system using the input variable, i.e. "features X".

In a first example of a preferred embodiment of the method, the structure of the neural network is a three layer neural network using 40 units (X=[Xo Xi X2 ··· X40] ) in the input layer, 8 units (ao, ai, ... as) in the hidden layer and one output lay^¬ er (y) . This embodiment is useful when the status of the AC network is predicted as either an islanded or a non-islanded network condition or the loss of the last feeder.

In a second example of a preferred embodiment of the method, the structure of the neural network is a three layer neural network using 400 units (X=[Xo Xi X2 ··· X400] ) in the input lay^¬ er, 25 units (ao, ai, ... a₂s) in the hidden layer and 15 units in the output layer (y) . This embodiment is useful when the status of the AC network is predicted as one of the fifteen classes as described earlier, comprising different variations suggesting the strength of the AC network and the amount of nearby voltage control present.

The number of layers can vary as well as the number of units in the hidden layer. This will have an effect on the accuracy and computing power needed to train the neural network. The structure proposed in the above-mentioned preferred embodi^¬ ment proves to be accurate to a large extent, while not re^¬ quiring a lot of computing power to train. Other neural net- work structures may be used for improved accuracy.

In a further embodiment, the method further comprises a step of using Sigmoid function g{z) = _1+e-_z for deducing the hypoth^¬ esis function (h_e(x)), where h_e(x) = g[6⁽-²'⁾{g(6⁽-¹½)}] , wherein the weight matrices (θ⁽¹⁾, 6^<2)) are determined by iteratively ad^¬ justing the weight matrices (θ⁽¹⁾, 6^<2)) to minimize a cost function (J(6)). The Sigmoid function (g(z)) is used as the activation function in the hidden layer, which is used for the calculation of the output of the neural network i.e.

h₉ (x) . Any other activation functions are also possible. The hypothesis function (h_e(x)) is formed using the Sigmoid func^¬ tion (g(z)) as an activation function where

h_e(x) = g[6⁽-²'⁾{g(6⁽-¹½)}] . Given a set of inputs, i.e. input varia^¬ bles "features X"=[X₀ Xi X₂ - XJ e Rⁿ⁺¹, and given the weight matrix (6^<j)) for each of the layers (6⁽¹⁾, 6^<2)), the output of the neural network, i.e. the estimated status of the AC net^¬ work (h_e(x)), can be calculated using h_e(x) = g[6⁽-²'⁾{g(6⁽-¹½)}] . The above mentioned cost function (J(6)) is given by: e (x⁽ⁱ⁾))+ (l-y⁽ⁱ⁾) .ln(l- h_e (x⁽ⁱ⁾))] +

where ^xm' is the number of training examples, ^λλ' is a regu- larization parameter to avoid over-fitting, ^XL' is the total number of layers in the network and s_± is the number of units in layer ^Λ1' , wherein if the layer ^Λ1' of the neural network has 3 layers, ^Λ1' can be 1,2 or 3.

In yet another embodiment the weight matrices (θ⁽¹⁾, θ^<2)) are determined by iteratively adjusting the weight matrices (θ⁽¹⁾, θ^<2)) using a back-propagation algorithm, in order to minimize the cost function, J(9) . Once the Neural network is trained these weights matrices 6^<3) are used to predict the status of the AC network using only the input "features X". Use of back-propagation algorithm enables the method to quickly and easily arrive at the best possible estimates of the weight matrices (θ^<3))· However, there are other algorithms which can be used for the same purpose. Another example of such an al^¬ gorithm is the conjugate gradient algorithm.

In another embodiment the cost function (J(6)) is a measure of error in estimating the status of the AC network obtained from the learning algorithm (ηθ(χ) ) compared to the corre^¬ sponding output of the status of the AC network (y) . The cost function (J(6)) helps in aligning the method to predict the output status of the AC network more accurately by training the learning algorithm to obtain an accurate hypothesis

(he(x) ) .

The above-mentioned and other features of the invention will now be addressed with reference to the accompanying drawings of the present invention. The illustrated embodiments are in^¬ tended to illustrate, but not limit the invention. The draw^¬ ings contain the following figures, in which like numbers re^¬ fer to like parts, throughout the description and drawing. FIG.l is a schematic diagram showing an HVDC transmission system having converter stations fitted with power converters .

FIG.2 is a schematic diagram showing a non-islanded network condition .

FIG.3 shows a flowchart containing the steps according to the disclosed method.

FIG.4 depicts a block diagram of a power converter and a controller of the power converter. FIG.5 shows an exemplary neural network computational model according to the disclosed method.

FIG.6 shows another exemplary neural network computational model according to the disclosed method.

As seen in FIG.l, the AC networks or grids 4 are connected to the converter stations 6 via transformers 5. The converter bus bar 3 connects the converter stations 6 to their corre^¬ sponding AC network 4 and the bus bar 11 connects the AC net- work 4 to the transformer 5 at the transmission side. The converter station 6 at the transmission side is connected to the converter station 6 at the receiving side via DC transmission line 7. The transmission of power takes place over large distances through the DC transmission line 7. The con- verter station 6 comprises of power converters 1 (not shown in FIG.1) .

The voltage measurements (V) to be used as one of the inputs for the training examples (X=[Xo Xi X2 ··· Xn ] ) are taken, in the shown embodiment, from the desired bus bar 11 connected to the AC network 4. Other inputs like the reactive power ex^¬ change with the AC network measurements (Q_ex) , active power exchange with the AC network measurements (P_ex) are further derived from the voltage measurements and current measure^¬ ments taken, for example, at the desired bus bar 11.

FIG.2 is a schematic diagram showing the status of the AC network 4 as a non-islanded network condition for an HVDC system.

According to FIG.2, the big AC network 8, made up of high voltage transmission network including generation as well as middle voltage and low voltage distribution networks, is con^¬ nected through a powerline 10 to the AC power generating units 9. The big AC network 8 along with the AC power generating units 9 together forms the AC network or grids 4 on the transmission side. This AC network 4 is connected to the con- verter station 6 via transformer 5. At the receiving side there is an inverter grid 12.

The converter bus bar 3 connects the converter station 6 to the AC network 4 and the bus bar 11 connects the AC network 4 to the transformer 5 at the transmission side. The transmis^¬ sion side is the side of the system from where the AC power is generated for transmission over the DC network though the DC transmission line 7. The converter station 6 at the transmission side is connected to the converter station 6 at the receiving side via DC transmission line 7. The desired bus bar 11 connects the big AC network 8 and the AC power gener^¬ ating units 9 to the transformer 5 at the transmission side. The transmitted DC power comes from the AC power generating units 9 as well as the big AC network 8.

An islanded network condition occurs, for example, when a small part of the system gets disconnected from rest of the AC network. According to FIG.2, the small isolated system is the HVDC rectifier converter station 6 together with the transformer 5, the converter bus bar 3 and the generating units 9. In the event of an islanded network condition the powerline 10 trips and the small AC system, i.e. the HVDC rectifier converter station 5, 3, 6 and the generating units 9, gets disconnected from the big AC network 8 and gets isolated from the rest of the AC network. After the tripping of the power- line 10 the AC power from the big AC network 8 is no longer present so the HVDC system has to quickly adjust the DC power transmitted from the transmission to the receiving side in order to match the generating power of the three AC power generating units 9.

This situation displays only one exemplary form of how an isolated islanded network condition might occur, however many other different ways of formation of an islanded network con- ditions are possible.

FIG.3 displays the steps to be carried out for estimating the status of the AC network 4 connected to the power converter 1 (See FIG.l) . The method 100 comprises a first step 101 of providing a training dataset comprising a set of training examples (X=[Xo Xi X2 ··· X_n] ) and a corresponding output (y) of the status of the AC network 4. The method 100 comprises a second step 102 of training a hypothesis function (h_e(x)) based on the training dataset. Thereafter a third step 103 of estimating the status of the AC network 4 using the hypothe^¬ sis function (h_e(x)) is carried out.

The method 100 presented in this invention allows for fast and accurate estimation of the status of the AC network 4.

FIG.4 displays a block diagram of a power converter 1 and its controller 2. Each power converter 1 has its corresponding controller 2 through which the power converter 1 is controlled. The estimated status of the AC network output is (h_e(x)) useful to change or optimize the controller 2 parame^¬ ters in order to achieve a better performance during system dynamics, eg. rise time, settling time, maximum overshoot etc . The training dataset comprising the set of training examples (X=[Xo Xi X2 ··· Xn ] ) is derived from voltage measurements (V) and/or reactive power exchange measurements (Q_ex) and other previously mentioned measurements with the AC network 4 be^¬ fore, during or after the contingency event. And these meas^¬ urements are taken from the desired bus bar 11 connected to the power converter 1. Based on the training dataset, the hypothesis function

(h_e(x)) learns to accurately predict the status of the AC network 4. The changed strength of system bus bars immediate^¬ ly after a system contingency event, for example, after the occurrence of an islanded network condition, can affect the performance of the system adversely. This method 100 will en^¬ sure that the system adapts to the new system conditions and gives improved dynamic performance even post a system contin^¬ gency event . FIG.5 shows an exemplary neural network computational model

13 according to the disclosed method. In this exemplary model 13 the structure of the neural network is a three layer neu^¬ ral network. The first layer is the input layer 14 comprising n+1 units (X=[Xo Xi X2 ··· Xn ] ) , including a bias unit Xo . The second layer is the hidden layer 15 which uses k+1 units (ao, ai, ... a_¾) , including a bias unit ao according to FIG.5. The units of the hidden layer 15 are also called as activation units. And the third layer is the output layer 16 having at least 1 unit ( h_e(x)) . For this particular exemplary neural network computational model the output layer 16 has only 1 unit according to FIG.5.

A first weight matrix 17 (θ⁽¹⁾) controls function mapping of the input variable set (X=[Xo Xi X2 ··· Xn ] ) of the input layer 14 to the activation units (ao, ai, ... a_k) of the hidden layer 15 and a second weight matrix 18 (θ^<2)) controls function map^¬ ping of the activation units (ao, ai, ... ajj of the hidden lay^¬ er 15 to the corresponding status of the AC network (h_e(x))) of the output layer 16. FIG.5 shows only an example of a neu^¬ ral network model 13, other neural network computational mod^¬ els 13 with different number of units in each layer and dif^¬ ferent number of layers can also be created for the same pur- pose.

FIG. 6 shows one other such exemplary neural network computa^¬ tional model 13 according to the disclosed method. In this exemplary model 13 the structure of the neural network is again a three layer neural network. According to FIG.6 the structure of the neural network is proposed as a three layer neural network model using 400+1 units (X=[Xo Xi X2 ··· X400] ) in the input layer 14, 25+1 units (ao, ai, ... a₂s) in the hidden layer 15 and 15 units in the output layer 16 (Ci C_{2 ...} C15) .

Although the invention has been described with reference to specific embodiments, this description is not meant to be construed in a limiting sense. Various modifications of the disclosed embodiments, as well as alternate embodiments of the invention, will become apparent to persons skilled in the art upon reference to the description of the invention. It is therefore contemplated that such modifications can be made without departing from the embodiments of the present inven^¬ tion as defined.

List of Reference Numbers

1 power converter

2 controller

3 bus bar

4 AC network

5 transformer

6 converter station

7 DC transmission line

8 big AC network

9 AC power generating units

10 powerline

11 desired bus bar

13 neural network computational model

14 input layer

15 hidden layer

16 output layer

17 first weight matrix θ⁽¹⁾

18 second weight matrix θ^<2)

100 method

101 a step of providing a training dataset

102 a step of training a hypothesis function (h_e(x))

103 a step of estimating the status of the AC network

Claims

Patent claims

1. A method (100) for estimating status of an AC network (4) connected to a power converter (1), the method (100) comprising:

- a step (101) of providing a training dataset compris^¬ ing a set of training examples (X=[Xo Xi X2 ··· Xn ] ) and a corresponding output (y) of the status of the AC network (4) ;

- a step (102) of training a hypothesis function (h_e(x)) based on the training dataset;

- a step (103) of estimating the status of the AC net^¬ work (4) using the hypothesis function (h_e(x)) .

2. The method (100) according to claim 1, wherein the method (100) further comprises a step of adjusting con^¬ troller (2) parameters of the power converter (1) based on the status of the AC network (4) estimated.

3. The method (100) according to any one of the above claims, wherein the status of the AC network (4) is one or more of an islanded or a non-islanded network condi^¬ tion, loss of last feeder or strength of the AC network with amount of nearby voltage control.

4. The method (100) according to any one of the above claims, wherein the set of training examples (X=[Xo Xi X₂ ... X_n] ) is derived from one or more of voltage measure^¬ ments (V) , reactive power exchange with the AC network (4) measurements (Q_ex) , active power exchange with the

AC network (4) measurements (P_ex) and network frequency measurements (Freq) of the AC network (4) .

5. The method (100) according to claim 4, wherein the set of training examples (X=[Xo Xi X2 ··· Xn ] ) are measured before, during and/or after a system contingency event.

6. The method (100) according to claim 5, wherein the system contingency event is one of a fault clearing event or a gain test.

7. The method (100) according to any one of the above claims 5 to 6, wherein the voltage measurements (V) are measured for a first time period spanning before, during and after the system contingency event, induced into the AC network (4), in steps of a second time period.

8. The method (100) according to any one of the claims 5 to 7, wherein the network frequency measurements (Freq) are measured for a first time period spanning before, during and after the system contingency event, induced into the AC network, in steps of a second time period.

9. The method (100) according to any of the claims 5 to 8, wherein the active power exchange with the AC network (4) measurements (P_ex) are measured for a first time pe- riod spanning before, during and after the system contingency event induced into the AC network (4), in steps of a second time period.

10. The method (100) according to any of the claims 5 to 9, wherein the reactive power exchange with the AC net^¬ work (4) measurements (Q_ex) are measured for a first time period spanning before, during and after the system contingency event induced into the AC network (4), in steps of a second time period.

11. The method (100) according to any one of the above claims, wherein the hypothesis function (h_e(x)) is de^¬ duced using a neural network computational model (13) .

12. The method (100) according to the above claim, wherein the neural network computational model (13) com^¬ prises : - an input layer (14) comprising the training dataset

(x=[Xo Xi x₂ ... x_n] ) ;

- at least one hidden layer (15) comprising activation units (ao, ai, ... a¾) ; and

- an output layer (16) comprising the corresponding output of the status of the AC network (η_θ(χ));

wherein a first (17) weight matrix (θ⁽¹⁾) controls func^¬ tion mapping of the training data set (X=[Xo Xi X2 ··· Xn ] ) of the input layer (14) to the activation units (ao, ai, ... a_¾) of the hidden layer (15) and a second (18) weight matrix (θ^<2)) controls function mapping of the activation units (ao, ai, ... a_k) of the hidden layer (15) to the cor^¬ responding output of the status of the AC network

(h_e(x)) of the output layer (16) .

13. The method (100) according to the above claim, wherein the method (100) further comprises a step of us^¬ ing Sigmoid function for deducing the hypothesis function (h_e(x)), where h_e(x)= g[e«{g(eWx)}],

wherein the weight matrices (θ⁽¹⁾, θ^<2)) (17, 18) are de^¬ termined by iteratively adjusting the weight matrices (θ⁽¹⁾, θ^<2)) (17, 18) to minimize a cost function (J (Θ) ) .

14. The method (100) according to the above claim, wherein the weight matrices (θ⁽¹⁾, θ^<2)) (17, 18) are de^¬ termined by iteratively adjusting the weight matrices (θ⁽¹⁾, θ^<2)) (17, 18) using a back-propagation algorithm.

15. The method (100) according to any one of the claims

10 to 12, wherein the cost function (J(6)) is a measure of error in estimating the status of the AC network (4) obtained from the hypothesis function (h_e(x)) compared to the corresponding output (y) of the status of the AC network ( 4 ) .