DE19518030C1 - Arrangement and method for determining the operating states of an electrical power supply network - Google Patents

Abstract

A process and system are disclosed for determining the states of operation of an electric energy supply network that transports, distributes and makes available electric energy to consumers. A number of neural networks is provided that corresponds to the number of characterising electrical properties of the nodes of an electric network. The neural networks have the same structure as the electric network and have neurones that correspond to the electric network nodes which are interconnected by synapses in the same way as the electric network nodes are interconnected. Their connections pass through dipoles and the synaptic weights may be set from the outside of the network at 0 or 1, depending on the corresponding positions of the dipoles. The scalar products of weighting vectors and input vectors of the neurones are compared to a threshold value equal to 1 and depending on whether the result of the comparison is higher than, equal to or less than 1, the neurones generate a 1 or a 0. The weights of the outputs of corresponding neurones in the individual neural networks are superimposed on each other depending on the priority of the characterising electrical properties.

Description

The invention relates to an arrangement for determining the Operating states of an electrical energy network with which electrical Energy is transported, distributed and made available to consumers and on a procedure for exercising with the order.

The operating states of an electrical power supply network who the u. a. found in switching requirements to before Execution of the switching operation based on the respective network status check whether the ge desired switching action is permitted. Is known for. B. a method for switching error-protected actuation of the switchgear of a switchgear location where the topological structure of the switchgear and the topolo structure independent switching fault protection interlocking rules be working. From the current position signal from the Schaltge councils and a basic set of topological elements with which all possible operating states of switchgear assemblies can be specified are the current operating state of the switchgear det. The current operating status is saved by means of a saved switching error Protection locking rules to release or block Switching action request commands processed by starting from Start and end nodes and the type of switchgear Baumentwick lungs are carried out with which the incl status of the switchgear galvanically connected network areas determined are determined for the dominant states according to the following priority are: voltage, load, earth, isolated node or undefined Ge offers. The dominant states of the areas are compared with switching error protection interlocking rule n enable or block tion of the switching action request command is determined (DE 38 12 072 C3).  

The calculation time for the electrical potential distribution of the networks depends on the network size and degree of meshing of the respective network. Based on the data A procedural algo has so far also been a moderate description of an electrical network rithmic method used to calculate the potential distribution. Starting from Formative network elements such as generators and feeders were closed along Switch wanted ways through the network. That from a formative network element accessible objects along the viable path then have the same potential as the defining element. With very large networks, e.g. B. about 100,000 or more Betriebsmit tel, the method is no longer capable of real distribution of the potential to provide time. A parallel calculation can only be carried out with great effort to lead. A reduction in computing time is only possible by selecting special cases, e.g. B. Subnets and replacement networks, possible with complicated special programming treatments must be calculated. The method works essentially incrementally, i. H. to a new route search must be carried out for each individual switch position change. This means that in the event of a fault in which many switch positions change, a high Re is necessary before the final result is available. There is still a time time-consuming initialization phase due to the incremental behavior required.

An arrangement for determining the operating states of an electrical device is known Energy network with which energy is transported, distributed and made available to consumers using a neural network. The state variables of the energy network who entered into a state memory to which the neural network is connected, the Carries out stability calculations for the energy network at high speed (JP 07-107667 A).

Finally, it is known for optimizing the load distribution of an electrical energy network as a simulation model to use a Hopfield network (recurrent neural network) (JP 02-136034 A).

The invention is based on the problem for determining the operating states of a electrical power network an arrangement and a method for performing with the An order to develop the potential distribution even with large energy networks can be obtained quickly.

The problem is solved for the arrangement according to the invention in that one of the An number of defining electrical properties of the nodes of a network dependent number of neural networks with the same structure is provided, each new for the network node have rons, each corresponding to the connections between the network nodes are synaptically connected to each other in such a way that in the case of a connection running over bipoles the synaptic weights according to the respective position of the two poles of outside in the network to be set to zero or one that the dot products from weight vector and  Compare the input vector of the neurons with a threshold value of one that, depending on the result, greater than or equal to or less than one the neurons each generate a one or a zero and that the weights of the exits of each other in the individual neurons corresponding neurons depending on the Priority of the formative properties are superimposed.

In the arrangement described above, each network node replaced by a binary neuron. The paths between the nodes will be in accordance with the network topology as synaptic connections built that are bidirectional. The connections run often between the network nodes via bipoles, which are Lei circuit breakers, disconnectors or transformers. The Bipoles are simulated by synaptic weights, one open two-pole the weight zero and a closed two-pole the weight equals one. The energy network is thus through a recurrent neural network with symmetrical weight matrix replicated. The synaptic weights are corresponding to that respective network status set. Therefore, there is no training phase for the neural networks required. These symmetrical Hopfield networks solve the problem of property propagation. Fires a neuron, like this the corresponding node has the property under consideration, otherwise not. As a feedback network, it converges iteratively against the correct one Solution.

Formative network nodes can be simulated by formative neurons that output the value one.

In a preferred embodiment, instead of embossing Neurons in the neural networks weight vectors of one and on introduce weight vectors of zero from non-formative neurons see.

It is useful if the neural networks instead of edge neurons NEN each have additional neurons that via a coupling of the Edge neurons corresponding synaptic weight to that in the neuronal The neuron adjacent to the edge neuron are coupled forward, the via a synaptic corresponding to the coupling of the edge neuron  Weight is coupled to a downstream neuron with the state of the edge neuron is output. Through the first one Based on the measures described, a neural network with we less recurrence. A neuron is closed under the edge neuron understand this is just a bi-directional connection to a neighbor ron has.

In a further preferred embodiment, instead of one Transfer neurons each have an additional neuron over one of the synaptic ones Coupling the transfer neuron corresponding weight to one of each neighboring ones that are not synaptically coupled within the framework of the network structure Neurons of the neural network are coupled forward, the neighboring bearded neurons with each other through a bidirectional synaptic Coupling according to a Ge calculated by an auxiliary neuron important connected by the synaptic weights of the trans ferneurons is determined, and being the neighboring neurons over each a Ge corresponding to the synaptic coupling of the transfer neuron important to a downstream neuron that the Output state of the transfer neuron. Under transfer neuron is here to understand a neuron that is two-way, synaptic verb to neighbors. The recurrent network does not contain more the transfer neuron or the transfer neurons. Here is to be take into account that already the replacement of part of the transfer neurons a significant reduction in the determination of the network status means necessary time. The entire network is at the above be described embodiment expanded by two additional neurons, where the before and after the respective rest of the recurrent network stored subnetworks are linear.

An additional neuron is preferably used instead of a transfer neuron via a synaptic coupling of the transfer neuron appropriate weight to one of the neighboring neurons coupled forward, which are bidirectional with each other by one of one Auxiliary neuron calculated synaptic weight coupled by the auxiliary neuron depending on the weight of the synaptic neurons and calculated from the weight between the two neighboring neurons the neighboring neurons each via synaptic weights, which correspond to the synaptic weights of the transfer neuron  a downstream neuron that is coupled to the state of the transfer neuron. In this embodiment, too bidirectional coupling of one or more transfer neurons the subnets and auxiliary networks are replaceable, which are linear, whereby a shortening of the necessary for determining the condition of the networks Time.

Are there neurons in the respective neural network that have more connected to neighboring neurons as two synaptic connections then every synaptic connection becomes a neighboring neuron replaced in the manner described in connection with transfer neurons. Neurons that use more than two connections to neighboring Neu rons are referred to as distribution neurons.

If the transformations described above are completed, this results in a feed-forward linear network that is divided into three Subnetwork is structured, namely an upstream network, an auxiliary ge weight network, the additional synaptic weights from known weights ten of the recurrent network, and from a downstream one Network on whose neurons the result of the energy network is available. The transformed neural network is completely determined from the To pology of the electrical network. The weights are given or he result from the two-pole positions generated by messages and determine the setting of the weights. The network needs no training phase, it shows no initialization behavior and does not work incrementally. It works without recursivity and determines the propagation of a formative property in an optimally short time Time. The result is exact at all times.

Improvements in the calculation of the potential distribution of energy network can already be transformed with a not completely mated neural network. He can make such improvements be enough by z. B. only the edge neurons or transfer neurons or distribution neurons in whole or in part in the manner described above be replaced.

A method of performing with one of the above Orders according to the invention is that the synaptic Ge  important to be adjusted to the binary states of the bipoles and that the states of the energy network from the states of the neurons can be determined by assigning physical quantities. Depending on The states can depend on the type of network and the degree of transformation of the energy network only at the downstream neurons be determined.

The neural networks described above can be used as hardware networks or be trained in software. The Er stands for hardware networks result in one or more changes in the state of the bipoles available faster.

The invention is illustrated below with reference to a drawing described embodiment described in more detail, from which there are Details, features and advantages emerge.

Show it:

Fig. 1 is a diagram of an electrical power supply system;

Fig. 2 is an illustration of the power supply network shown in Figure 1 by a node-path graph.

Figure 3 is a binary feedback neural network as a model for the power supply network.

FIG. 4a, b shows a diagram for illustrating the transformation of bidi-directionally connected to the edge of neurons of the network in a feed-forward network structure;

Fig. 5a, b is a scheme to illustrate the transformation of bidirectionally coupled with non-coupled neighboring neurons th transfer neurons of the neural network acc. Fig. 3 in a feedforward network structure;

Fig. 6a, b shows a diagram to illustrate the transformation of bidirectionally coupled neighboring neurons coupled transfer neurons of the neural network acc. Fig. 3 in a feedforward network structure;

Fig. 7 is a neural network in the schema;

Fig. 8 is a symbolic network node which may comprise a series of formative properties that through paths other network nodes affect;

Fig. 9 is a schematic representation of two network nodes that are connected by a path and

Fig. 10 is a built up of multiple subnets neural network.

Fig. 1 shows an example of an electric power supply system having a supply 1 in form of a generator, of a power switch 3 is arranged in series with a transformer 2 and. At the circuit breaker 3 , a busbar 4 is connected, from which two branches, not specified, which have circuit breakers 5 , 6 . The circuit breakers 5 , 6 are connected on their ends facing away from the busbar 4 together to a circuit breaker 7 . A further circuit breaker 8 , 9 is arranged in series with each circuit breaker 5 , 6 . The power switches 8 , 9 , consumers 10 , 11 are connected downstream.

An electrical power supply network, such as. B. Darge in Fig. 1 is modeled by the elements nodes and bipoles. Two poles are resources that can change the course of potential, e.g. B. transformers, circuit breakers, disconnectors, earth electrodes, etc. Nodes are the respective edge poles of the two-pole system and the connections between the two-pole system, for. B. lines and busbars. Nodes are loaded with potential. There are formative nodes that bring a specific property "grounded", "loaded", "supplied" etc. into the network. Other nodes obtain the properties described above by connecting them via a two-pole circuit with pre nodes are connected. The above-mentioned properties are distributed over the closed bipoles to the nodes of the network and thus determine the potentials at the various points in the network.

The energy supply network shown in Fig. 1 can be represented schematically in the form of a node-path graph. The node K0 corresponds to the feed 1 and is connected via a path W0, which corresponds to the transformer 2 , to the node K1. The node K1 symbolizes the connecting line between the transformer 2 and the circuit breaker 3 . Path W1 corresponds to circuit breaker 3 . The busbar 4 is symbolized by the node K2. The two circuit breakers 5 , 6 each have the paths W2, W3 arranged which end in the node K3, K4, which are the lines between the circuit breakers 5 , 8 and 6 , 9 , respectively. These lines are shown in FIG. 2 connected to each other through the path W4. The nodes K3, K4 are each connected through the paths W5, W6, each of which symbolizes one of the power switches 8 , 9 , to the nodes K5, K6, to which the consumers 10 , 11 correspond. One way Wx, e.g. B. one of the paths W0 to W6, always acts bidirectionally between the two adjacent nodes. The path Wx is feasible if the corresponding dipole transmits the property of an attached node to the other. Otherwise it is not feasible.

Different bipoles behave towards the different egg properties that enter the electrical network through formative nodes gen, very different. So z. B. a transformer the egg property "loaded", but not the property "grounded" equally eat.

The electrical network is therefore replicated as often as it is under there are various distinctive knot properties. For the individual pre The properties of the computation are reduced Sharing on the net for a yes / no statement for each property on everyone Node. The resulting potential curve in the network is determined by overlay The different yes / no statements taking Domi into account financial criteria determined, which are given for the defining properties will. The electrical network is used for each defining characteristic simulated a neural network.

In the following, the electrical network is considered for only one property. The binary network is formed in the following manner, which was explained in connection with FIG. 2. Each of the nodes K0 to K6 is replaced by a binary neuron N0 to N6. Every path between two nodes is established as a synaptic connection. The paths G0 to G6 correspond to the paths W0 to W6. Formative nodes are simulated by formative neurons 1.

A corresponding neural network is shown in FIG. 3, the circles symbolizing neurons, the rectangles, synaptic weights and the oval surfaces constant inputs. Each path between two nodes is set up as a synaptic connection, which is labeled G0 to G6 in FIG. 3. Because of the bidirectionality of the paths, this results in a recurrent neural network with a symmetrical weight matrix. The current values of the synaptic weights correspond to the positions of the way bipoles, whereby the following assignment exists:

Path not feasible <= <two-pole open <= <weight = 0
Way feasible <= <two-pole closed <= <weight = 1.

Since the synaptic weights correspond to the respective two-pole positions training period is omitted, so the neural network is through this construction ideally on the underlying electrical network trained. Every binary neuron forms the dot product of weight vector and input vector and compares it to the threshold "1". Is if the dot product is greater than or equal to this threshold, this will fire Neuron, therefore outputs a "1", otherwise a "0". Every formative new ron (equivalent of a formative node) of the eigen to be considered shaft always fires, so it has an additional synapti connection with weight "1" to a constant input "1".

This symmetrical, implicitly pre-trained Hopfield network solves the problem property propagation. If a neuron fires, it corresponds the relevant property, otherwise not. Because of the Re The following characteristics result from the competition: As a feedback network it iteratively converges to the right solution, for which large electrical networks require a certain period of time, d. H. the Be the calculation time depends on the degree of meshing of the electrical network zes.

According to the invention, the Hopfield network is transformed into a feedforward neural network in the manner described below:
First, the neurons of the Hopfield network are classified according to the number of bidirectional synaptic connections to neighboring neurons.

Number = 0 = <isolated neuron.
Number = 1 = <edge neuron.
Number = 2 = <transfer neuron.
Number <= 3 = <distribution neuron.

All neurons in the network are sorted in the order of smallest ver number of connections to the largest number of connections in a transformation  subject to procedures that successively reduce the recurrent network and creates a feedforward network. The process ends when the recurrent network is reduced to zero.

If the number of synaptic connections from a neuron to its neighbor neurons is zero, the neuron can either form or not be enough. If the neuron is formative, it always fires and can therefore be removed from the recurrent network without retroactive effect. A not formative neuron never fires and can therefore also from the recur pension network are removed.

If the number of bidirectional synaptic connections from a neuron to neighboring neurons is one, the neural network is converted. This conversion is explained in more detail with reference to FIGS. 4a and 4b. The neuron denoted N _x , a so-called edge neuron, has only a bidirectional connection with the weight G _x to a neighboring neuron N _n , which is connected in any way to the recurrent remaining part of the network N _r .

The original neuron N _x is gem. Fig. 4b by an additional neuron n _x replaced, which is fed forward on the same weight G _x at the only neighbor neuron N _n. The result of the recurrent residual network is also fed forward via the same weight G _x to a downstream neuron N _x '. This creates a linear binary neural network upstream and downstream of the recurrent residual network. The states of the original neuron N _{x are} available on the neuron N _x ′. The solution properties are retained, the computing time is reduced by reducing the recurrent network.

With two synaptic connections of a neuron to neighboring neurons there are two different cases. In the first case, the neighbors have new ones no direct synaptic connection between them. In the second In this case there is a direct synaptic connection between the neighbors neurons.

The first case is described below with reference to FIGS. 5a and 5b. It is a neuron N _y , a so-called transfer neuron, connected via the two bi-directional synaptic connections with the weights G _a and G _b to a neighboring neuron N _a and N _b . The two neighboring neurons N _a , N _b are connected to the recurrent residual network N _r 'in a manner not shown.

During the conversion, as can be seen from FIG. 5b, the original neuron N _{y is} replaced by an additional neuron n _y , which is _coupled forward to the two neighboring neurons N _a , N _b via the same weights G _a and G _b . The two neighboring neurons N _a and N _b , which belong to the recurrent residual network N _r , are also coupled to a neuron N _y 'via the same weights G _a and G _b . They also receive a new bi-directional synaptic connection, the weight of which is calculated by an auxiliary neuron H _y . The recurrent residual network no longer contains the neuron N _y . The overall network was expanded by two additional neurons n _y and H _y . The upstream and downstream subnets and the auxiliary network are linear.

The auxiliary neuron calculates the weight G _y based on the weights G _a and G _b , which are each evaluated in half. At the neuron N _y 'the states of the original neuron N _{y can be} tapped.

The transformation of neurons with two synaptic connections to neighboring neurons is described with reference to FIGS. 6a, b, which also have direct synaptic connections with one another.

A neuron labeled N _z , a so-called transfer neuron according to Fig. 6a, is bidirectionally connected via synaptic weights G _a ', G _b ' with neighboring neurons N _a ', N _b ', which are directly connected to each other bidirectionally synaptically accordingly the weight G _from . The connection of the post bar neurons N _a ', N _b ' with the recurrent residual network N _r '' is retained in the corresponding manner.

According to. Fig. 6b, the original neuron N _z replaced by an additional neuron n _z , which is coupled over the same Ge weights G _a 'and G _b ' to the two neighboring neurons N _a ', N _b '. The two neighboring neurons N _a 'and N _b ', which belong to the re-competing residual network N _r '', are also coupled via the same weights G _a 'and G _b ' to a neuron N _z '. In addition, the bidirectional weight G _ab between neuron N _a 'and N _b ' is replaced by the weight G _z , which is calculated by an auxiliary neuron H _z . The recurrent residual network no longer contains the neuron N _z . The overall network was expanded by two additional neurons n _z and H _z . The upstream and downstream subnets and the auxiliary network are linear.

The auxiliary neuron H _z calculates the synaptic weight G _z from the syn aptic weights G _ab, which is multiplied by one, and from the weights G _a ′, G _b ′, which are each multiplied by 1/2.

If a neuron has more than two synaptic connections to its neighbor has a transformation based on each of the neurons Transformation of two synaptic connections described above executed. Every synaptic connection becomes a neighbor neuron treated or transformed separately, d. H. z. B. at three connections each performed three transformations.

The transformations described above can, but do not have to, for all edge, transfer and distribution neurons are performed. With a complete transformation you get a forward cop peltes binary neural network, which is structured into the following three subnets is:

• 1. An auxiliary weight network, the additional synaptic Ge weights calculated from known weights.
• 2. An upstream network.
• 3. A downstream network, on whose neurons the whole result can be decreased.

This transformed network is completely determined from the topology of the electrical network. Such a neural network is shown in FIG. 7 and is designated by 12 . The weights of the network 12 are calculated or result from the two-pole positions, which determine the weights of the synaptic inputs of neurons as inputs. The inputs are generally designated 13 in FIG. 7. The outputs 14 generally designated in FIG. 7 are made available by the neurons of the downstream network and represent the current potentials of the network nodes.

The neural network 14 does not require a training phase, it shows no initialization behavior and does not work incrementally. It does not require any recursiveness and determines the propagation of a formative property in an optimally short time. The result is exact at all times.

The neural network 14 can be implemented in hardware or software. The achievable advantages are mentioned again below:

• - The calculation time is almost linear from the network size dependent.
• - The solution method (neural network) is parallel sizable.
• - The neural network determines the result non-incrementally tell. In the event of a malfunction, many switch positions can change collected and then worked together.
• - The initialization phase is omitted.
• - The neural network also determines results in real time for very large electrical networks.
• - The neural network can be based on a special neural Accelerate hardware "as you like".

It has already been mentioned above that the network is reproduced as often as there are different characteristics. FIG. 8 shows a topological network node 15 which can have a number of properties which are associated with E0, E1, E2, E3. . .En are designated. The properties E0 to E3 are also already described above. These properties are z. B. spread on two routes 16 , 17 , which are connected to the network node 15 ver.

In Fig. 9 two nodes 18 , 19 are shown, which are connected via a path 20 , which can have different defining properties E0 to En. The potential of an electrical network node depends on its connection to a formative network node.

So z. B. a network node simultaneously with one

• - Formative node: feed by generator 1 (20 kV)
• - Formative node: feed through generator 2 (20 kV)
• - formative knot: load

be connected.  

This electrical network node then has the properties

• - Potential 20 kV, fed by generator 1 (Property E2)
• - Potential 20 kV, powered by generation 2 (Property E3)
• - charged (property E1)

on.

This becomes the overall message:

• - Safe supplied load at potential 20 kV -

combined.

The fact whether a certain knot has a formative knot connected or not connected to a certain property is un depending on whether there are also connections to other formative There are nodes with different properties.

It is therefore possible to kno the distribution of all characteristics defining ten in the electrical network by determining the distribution each individual property in the network is determined independently of one another and then combined into one overall statement, which is the partial statements combined and weighted according to priority.

So there is a yes / no statement for each network node and each property to determine, which is carried out with the help of binary neurons.

Therefore, the electrical network is modeled by N neural networks that determine yes / no statements for each node independently of each other for each electrical property. Here N is the number of different properties. Such a neural network 21 is shown in FIG. 10. The independently ascertained results of the N binary neurons presenting a node are then combined to form a total statement which, as mentioned above, depends on the priority of the properties, which in turn is determined by the physical conditions of the respective property. The priority is therefore given.

Claims (6)

1. Arrangement for the determination of the operating states of an electrical power network's, with which electrical energy is transported, distributed and made available to consumers, characterized in that a number of neural networks with the same structure provided depending on the number of defining electrical properties of the nodes of a network which each have neurons (N0, N1, N2, N3, N4) for the network nodes, which are each synaptically connected to one another in accordance with the connections between the network nodes in such a way that the synaptic weights (G0, G1, G2, G3, G4, G5, G6) according to the respective positions of the bipoles from the outside in the network to zero or one, so that the scalar products of the weight vector and input vector of the neurons (N0 ... N4) each have a threshold value of one who compared that, depending on the comparison result greater than or equal to or less than one the neurons (N0. . .N4) generate a one or zero and that the weights of the outputs of the neurons corresponding to each other in the individual neural networks are superimposed on one another depending on the priority of the defining properties. 2. Arrangement according to claim 1, characterized, that instead of formative network nodes in the neural networks with neurons Weight vectors of one and new instead of non-defining network nodes rones with zero weight vectors are provided. 3. Arrangement according to claim 1 or 2, characterized in that the neural networks instead of edge neurons each have additional neurons (n _x ), via a coupling of the respective edge neurons (N _x ) corresponding synaptic weight (G _x ) to the in the neural network the edge neuron (N _x) adjacent neuron (N _n) forward are coupled, the corresponding via a coupling of the peripheral neuron (N _x) synaptic weight (G _x) is to a subsequent neuron (N _x ') fed forward to the the state of the edge neuron (N _x ) is available. 4. Arrangement according to one or more of the preceding claims, characterized in that instead of a transfer neuron (N _y ) each an additional neuron (n _y ) via a synaptic coupling of the transfer neuron (N _y ) corresponding weight (G _a , G _b ) the disclosed Benach to each of a non-directly synaptically bidi-directionally coupled within the network structure neurons (N _a, N _b) is coupled to forward that the adjacent neurons (N _a, N _b) to each other by a bidirectional synaptic coupling according to one of a Auxiliary neuron (H _y ) calculated weight are connected, which is determined by the synaptic weights (G _a , G _b ) of the transfer neuron, and that the neighboring neurons (N _a , N _b ) each via one of the synaptic coupling of the transfer neuron (N _y ) corre sponding weight (G _a , G _b ) are coupled to a downstream neuron (N _y ′), which indicates the state of the transfer neuron (N _y ). 5. Arrangement according to one or more of the preceding claims, characterized in that instead of a transfer neuron (N _z ) each have an additional neuron (n _z ) via a synaptic coupling of the transfer neuron (N _z ) corresponding weight (G _a ', G _b ') Is coupled forward to one of the neighboring neurons (N _a ', N _b '), which are bidirectionally coupled to one another by a synaptic weight (G _z ) calculated by an auxiliary neuron (H _z ), which is generated by the auxiliary neuron (H _z ) depending on the weight (G _ab , G _a ′, G _b ′) of the synaptic couplings between the transfer neuron (N _z ) and neighboring neurons and the weight between the two neighboring neurons (N _a ′, N _b ′) is calculated, and that the adjacent neurons (N _a ', N _b'), depending on synaptic weights (G _a ', G _b'), which the correspond to the synaptic weights transfer neuron _{(N), and} forward to a subsequent neuron (N _z ') are coupled, the state d of the transfer neuron (N _z ) indicates. 6. Method for determining the operating conditions of an electrical Energy network with a device according to one or more of the preceding claims, characterized, that the synaptic weights correspond to the binary states that of the bipoles are set and that the states of En energy network from the states of the neurons under assignment to formative physical quantities can be determined.