CH690637A5 - voltage control and reactive power of a system generating electricity. - Google Patents

Description

5

10

15

20

25

30

35

40

45

50

55

60

65

CH 690 637 A5

Description

The present invention relates to a device for controlling voltage and reactive power of the system producing electricity, using a learning algorithm to optimally control a controlled object whose characteristics are difficult to identify.

Description of the prior art:

Figs. 1 (a) and 1 (b) are flowcharts of learning algorithms adopted by the conventional network which is described, for example, in an article written by RJ Williams et al., Entitled "A Learning Algorithm for Continually Running Fully Recurrent Neural Networks ”from Neural Computation Vol. 1, No. 2, 1989.

The learning algorithm is implemented as follows. As shown in fig. 1 (a), the learning algorithm begins with a step ST1 at which target values ck (t) are introduced as output values yk (t) intended to be output by a neural network for inputs supplied to the neural network , where k is the number of neural elements constituting the neural network.

We then go to a step ST2 to find the errors ek (t), which are the differences between the target values dk (t) and the actual output values yk (t) actually produced by the neural network when certain inputs are supplied to the neural network.

Then, we go to a step ST3 to find the value of an evaluation function J (t) conforming to equation (1) which expresses the evaluation function J (t) as the sum of the squares of the errors ek (t) as follows:

J (t) = 12 {(and (t)) 2} (1)

2k

Then, we go to step ST4 to compare the value of the evaluation function J (t) with a predetermined value e fixed in advance. If in step ST4 it is found that the value of the evaluation function J (t) is less than the predetermined value e, the learning process is stopped. On the other hand, if it is found in step ST4 that the value of the evaluation function J (t) is greater than the predetermined value e, the process continues until steps ST5 and ST6 to modify the weights of the connectivity among the neural elements i (or the connectivity coefficients wy between the neural elements i and the inputs j), in accordance with equation 2 given below:

Wii (t) = Wij (t) + ÀWii (t) (2)

where AWij (t) are time-dependent correction quantities for the connectivity coefficients Wij. The time-dependent correction quantities AWi] (t) are found using the steepest descent method, illustrated in figs. 1 (b) and 8 (a), according to the following equations:

AWi / t) = oc 2 ek (t) pku (t) (3)

k £ v pk “(t + i) = f '(& (t)) [2 WuP'iXt) + SikZKt)] (4)

i & u p \ (tc) = 0 (5)

1 if i = k

Sik = {(6)

0 in other cases in which:

f is a transfer function of the neural elements,

pkij is an auxiliary function expressed by equation (4) used to simplify the equations,

Zk represents the inputs provided to the neural elements and

Sk is a sum of the weighted entries.

The weights of the connectivity are modified so that the value of the evaluation function J (t) becomes smaller than the predetermined value e.

2

5

10

15

20

25

30

35

40

45

50

55

60

65

CH 690 637 A5

Ori then returns to step ST2, passing through the calculation step ST3, the comparison step ST4 and the modification steps ST5 and ST6. This loop is repeated until the value of the evaluation function J (t) becomes smaller than the predetermined value e, in which case the learning process is stopped.

Fig. 2 is a diagram of the configuration of the conventional voltage and reactive power control of the system producing electricity. Reference 1 in the figure is a system producing electricity to be controlled. Reference 2 indicates a neuro-system simulator, also called a neural network simulator of the system producing electricity. The neurosystem simulator 2 is a neural network which has completed a first learning phase, in advance, so that its input / output relationships correspond to those of the system producing electricity 1, serving as a control object. Reference 3 is a VQ control device for controlling the upward and downward movement of a socket used in a tap switch transformer, as a function of the magnitude of the load thereof, in accordance with the control diagram shown in the fig. 4 or switching on / off a capacitive capacitor - shunt reactor. The VQ 3 control unit is also known under the name of voltage and reactive power control. Reference 4 indicates a neuroVQC, a neural network which has completed the first learning phase in advance, so that its input / output relationships correspond to those of the VQ control unit 3. NeuroVQC 4 is also known as the name of the neural network voltage and reactive power command.

Fig. 3 shows a typical configuration of the neuro-system simulator 2 and of the neuroVQC 4 in a second learning phase. In the second learning phase, the inputs / outputs of the neuroVQC 4 are connected to the inputs / outputs of the neuro-system simulator 2 as shown in the figure. In the second learning phase, the neural network of neuroVQC 4 receives the instruction to carry out a learning process so that the outputs of neuro-system simulator 2 correspond to the inputs of neuroVQC 4.

Now, we will explain the operating principle of the voltage and reactive power control of the system producing electricity.

First of all, the first learning phases are carried out, in advance, by the neural network of the neuro-system simulator 2, so that its input / output relationships correspond to those of the system producing electricity 1 and by the neural network of neuroVQC 4, so that its input / output relationships correspond to those of the control device VQ 3 at each point on the time axis. The neuro-system 2 and the neuroVQC 4 are produced to provide in the first stages of learning characteristics equivalent respectively to those of the system producing electricity 1 and of the control device VQ 3.

The information input / output of the system producing electricity 1 such as the vertical position of the socket and that concerning the ON / OFF states of Se obtained in the first learning phase does not only concern its own system producing electricity. electricity 1, but also includes data obtained from power plants located on other sites by, among other means, cable connections or the like.

Then, the second learning phase is carried out by connecting the inputs / outputs of the neuroVQC 4 to the inputs / outputs of the neuro-system simulator 2 as shown in the figure. As described above, the neural network of neuroVQC 4 receives the instruction to undergo the second learning process so that the outputs of neuro-system simulator 2 correspond to the inputs of neuroVQC 4.

Finally, the actual control operation in which the neuroVQC 4 controls the electricity producing system 1 is based on the results of the first and second learning phases.

The conventional neural network learning algorithm is used as described above. Under these conditions, in a process of learning the dynamic characteristics of a command object or an command, the target values can change suddenly. In this case, problems arise because, among other things, it takes a long time to complete the learning process. This is because with the conventional learning algorithm, the weights of the connectivity among the neural elements are changed by considering only the errors between the output values of the neural network and the target values.

In addition, the conventional voltage and reactive power control of the electricity generating system has a configuration described above. In such a configuration, in order to carry out a first learning phase, the input / output information of power stations located on other sites must be transmitted to its own power station by cables or the like. It inevitably follows the problems that installing cables and the like is inconvenient and that it takes a long time for the neural network to complete a learning process, to mention just a few.

In view of the above, an object of the present invention is to provide a learning algorithm for a neural network which makes it possible to complete the learning process thereof more quickly, even when the target values change suddenly. .

Another object of the present invention is to provide voltage and power control

3

5

10

15

20

25

30

35

40

45

50

55

60

65

CH 690 637 A5

reactive of the system producing electricity which can obtain input / output information from power plants located on other sites for learning purposes, without cables or the like,

Another object of the present invention is to provide a voltage and reactive power control of the electricity producing system which can adjust the characteristics of a neural network voltage and reactive power control to optimal values for the characteristics. of a system producing controlled electricity, without interrupting the control of the system producing electricity, even if the characteristics of the system producing electricity change.

According to the first aspect of the present invention, to achieve the above-mentioned objectives, a learning algorithm for a neural network is provided, by means of which the errors between the neural network output values and the target values are first determined, then we calculate a sum of the squares of the errors and the squares of the changes of errors, and finally we modify the weights of the connectivity among the neural elements of the neural network in a direction to decrease the sum only if it is found that the sum is greater than a predetermined value.

According to the second aspect of the present invention, a learning algorithm for a neural network is provided, with which the errors between the neural network output values and the target values are first determined, then a sum of the squares is calculated. errors and finally the weights of the connectivity among the neural elements of the neural network are modified using a first order search technique in one direction to decrease the sum only if it is found that the sum is greater than a predetermined value.

According to the third aspect of the present invention, a voltage and reactive power control of the electricity generating system is provided, in which a neural network of a neural network simulator of the electricity generating system learns the input relationships. / output of systems producing electricity other than its own power plant from data estimated on the basis of the reactive power passing through the transformer primary and the input / output relationships of its own power plant, in advance, during a process of learning the input / output relationships of the electricity producing system.

According to the fourth aspect of the present invention, there is provided a voltage and reactive power control of the electricity generating system equipped with a relearning evaluation unit, in which the control of the electricity producing system is switched of a neural network voltage and reactive power control on a neural network voltage and reactive power command and on a neural network voltage and reactive power control provided for training and on a neural network simulator of the producing system electricity also provided for learning, as well as for learning the new input / output relationships of the electricity producing system and recording the learning results respectively in the neural network voltage and reactive power control and in the neural network simulator of the system producing electricity in the event of detectio n error between the input / output relationships of the neural network simulator of the electricity producing system and the input / output relationships of the electricity producing system.

As mentioned above, the number of iterations required in the learning process can be decreased with the learning algorithm for a neural network used in accordance with the first aspect of the present invention, since errors are first determined between the neural network output values and the target values, then we compute a sum of the squares of the errors and the squares of the error changes, and finally we modify the weights of the connectivity among the neural elements of the neural network in one direction to decrease the sum only if it is found that the sum is greater than a predetermined value.

Furthermore, the number of iterations required in the learning process can be reduced with the learning algorithm for a neural network produced in accordance with the second aspect of the present invention, since the errors between the values are first determined. of the neural network output and the target values, then we calculate a sum of the squares of the errors and finally we modify the weights of the connectivity among the neural elements of the neural network using the first order search technique in one direction to decrease the sum only if the sum is found to be greater than a predetermined value.

In addition, it is not necessary to install cables or the like, because in the control of voltage and reactive power of the system producing electricity implemented according to the third aspect of the present invention, the neural network of neural network simulator of the electricity producing system learns the input / output relationships of electricity producing systems other than its own power plant from data estimated on the basis of the reactive power passing through the transformer primary and input / output relationships of one's own power plant, in advance, during a process of learning the input / output relationships of systems producing electricity.

In addition, the characteristics of the neural network reactive power and voltage control can be adjusted to values which are optimal for the characteristics of the electricity generating system, without interrupting the control of the electricity generating system even when the characteristics of the electricity producing system change, since the voltage and reactive power control of the electricity producing system carried out according to a fourth aspect of the present invention is equipped with a relearning evaluation unit, allowing the order from

4

5

10

15

20

25

30

35

40

45

50

55

60

65

CH 690 637 A5

system producing electricity to switch from voltage and reactive power control to neural network to voltage and reactive power control and to voltage and reactive power control to neural network intended for learning and on neural network simulator of the electricity producing system also intended for learning, as well as to learn new input / output relationships of the electricity producing system and to record the learning results respectively in the voltage control and of reactive power to the neural network and to the neural network simulator of the electricity-producing system, in the event of detection of an error between the input / output relations of the neural network simulator of the electricity-producing system and the input relations / exit from the system producing electricity.

The above new objects and characteristic features of the present invention, as well as others, will become more clearly apparent from the detailed description which follows, when taken in connection with the attached drawings. It should be understood, however, that the drawings are for illustrative purposes only and are not intended to define the limits of the present invention.

Brief description of the drawings:

Fig. 1 (a) is a flow diagram of a learning algorithm used in the conventional neural network;

fig. 1 (b) is a flowchart showing in more detail a step ST5 of the flowchart shown in FIG. 1 (a);

fig. 2 is a diagram showing the configuration of a conventional voltage and reactive power control of the system producing electricity;

fig. 3 shows a typical configuration of a neuro-system simulator 2 and a neuroVQC 4 in a second learning phase;

fig. 4 is a conceptual diagram showing an operational algorithm of a voltage and reactive power control;

fig. 5 (a) is a flow diagram of a learning algorithm used in a neural network adopted in an embodiment carried out in accordance with a first aspect of the present invention;

fig. 5 (b) is a flowchart showing in more detail a step ST5 of the flowchart shown in FIG. 5 (a);

fig. 6 is a conceptual diagram describing an evaluation function;

fig. 7 is a learning flowchart used to modify the weights of connectivity among the neural elements used in a neural network adopted in an embodiment carried out in accordance with a second aspect of the present invention;

fig. 8 is a conceptual diagram describing a steepest descent method and an optimum gradient technique;

fig. 9 is a diagram showing the configuration of a voltage and reactive power control of the electricity producing system produced according to a third aspect of the present invention;

fig. 10 shows a typical configuration of a neuro-system simulator 5 and a neuroVQC 4 in a second learning phase;

fig. 11 is a systematic diagram showing a current supply 1;

fig. 12 is a diagram showing the estimated data of the input / output relationships of a power plant located on another site;

fig. 13 is a graph showing a process for estimating a power outlet operation carried out by a power station located on another site;

fig. 14 is a diagram showing the configuration of a voltage and reactive power control of the electricity producing system produced according to a fourth aspect of the present invention;

fig. 15 is a graph showing typical changes in power consumption; fig. 16 is a graph showing a portion of training data;

fig. 17 is a graphical diagram showing the extrapolation of an existing learning pattern; fig. 18 is a diagram showing typical changes in system characteristics; and fig. 19 is a diagram showing the configuration of a neural network of the type known under the designation of recurrent neural network.

Detailed description of preferred embodiments

The preferred embodiments of the present invention will now be described with reference to the accompanying drawings. The descriptions of the common component parts are omitted here to avoid unnecessary repetitions.

First form of execution

We will now describe a first embodiment of the present invention with reference to the diagrams in the appendix.

Fig. 5 (a) is a flowchart of a learning algorithm used in a neural network

5

5

10

15

20

25

30

35

40

45

50

55

60

65

CH 690 637 A5

adopted in a first embodiment according to a first aspect of the present invention, while FIG. 5 (b) is a flow diagram showing in more detail a step ST5 of the flow diagram shown in FIG. 5 (a). In the figure, the same notations as those in FIG. 1 are used to designate the same component parts or equivalent component parts. Next, the operations of the flowcharts are described.

The comparison of the flowchart shown in fig. 5 (a) and that of FIG. 1 (a) highlights the difference between step ST11 of the first and step ST3 of the second. The two stages differ from each other in that the equations used to calculate the evaluation function J (t) are different. The formula used in step ST11 is given by equations (7) to (9) as follows:

J (t) def 1 E {(and (t)) 2 + yCAet (t)) 2} (7) 2 t e »

dk (t) - yk (t) if k € = T (t)

ek (t) def {(8)

- 0 in other cases in which:

y is a constant greater than zero and

T (t) is a fixed number of neural elements for which target values exist.

Aet (t) = e> (t) - ek (t-i) (9)

By comparing equations (1) and (7), we find that they are different from each other, in that the square values {Aek (t)} 2 of the changes in errors are added to the right side of equation (7).

In this way, the changes in errors are taken into account in the weights modifying the connectivity among the neural elements or connectivity coefficient w, j between the neural elements i and the inputs j in steps ST5 and ST6. Under these conditions, one can get the information of whether the changes in errors caused by modified weights of connectivity are changed in an increasing or decreasing direction. We can thus find the appropriate quantities for the correction AWjj (t). It follows that the learning algorithm can be performed with a shorter time required to complete the learning process than with the conventional algorithm.

The quantities suitable for the Awy correction (t) are given by equations (10) to (13) as follows:

AWi, <t) = aZ {ek (t) + Aek (t)} pko (t) (10)

k t u pkiKt + l) = f '(Sk (t)) [Z WuPSiCt) + SikZi (t)] (11)

k £ u pkiXto) = 0 (12)

1 when i = k

8 = {(13)

0 in other cases in which:

f is a transfer function of the neural elements,

pky is an auxiliary function expressed by equation (11) to simplify the equations,

Zk is an input supplied to the neural elements and Sk is a sum of weighted inputs.

It should be noted that in the case of an evaluation function J (t) taking into account changes in errors by assigning y the value 1, the time required to complete the learning process is reduced to one quarter of that required for an evaluation function J (t) which does not take into account changes in errors by assigning y the value 0, provided that the other conditions are the same.

Second form of execution:

Fig. 7 is a flowchart of a learning algorithm used to modify the weights of the

6

5

10

15

20

25

30

35

40

45

50

55

60

65

CH 690 637 A5

connectivity among the neural elements used in a neural network adopted in a second embodiment carried out in accordance with a second aspect of the present invention.

The implementation of the learning algorithm is described below.

The difference between the present invention and the previous one is the algorithm used to modify the weights of connectivity among the neural elements Wij (t). To be more precise, in the previous invention, the quantities for the correction AWij (t) were found in accordance with a flow diagram of FIG. 1 (b) using the steepest descent method illustrated in fig. 8 (a). The quantities for the correction AWij (t) are then used to correct the weights of the connectivity wij (t) in steps ST5 and ST6 of the flowchart shown in FIG. 1 (a). On the other hand, in the case of the present invention, the quantities for the correction AWij (t) are found in accordance with a flow diagram of FIG. 7 using a first order search technique such as the optimum gradient technique shown in FIG. 8 (b) which is also known as a second order interpolation technique. The quantities for the correction AWi] (t) are then used to correct the weights of the connectivity Wij (t).

The optimum gradient technique is characterized by the fact that the learning process can be managed with fewer interactions to find a target function. In this embodiment, this technique is used to find the value of a appearing on the right side of equation 3 which optimizes the evaluation function J (t).

When referring to fig. 7, the learning algorithm is explained as follows:

1. If 1-1 (1)> H (0), find H (ce) by giving a to a series of values starting with 1/2 to 1/4 - etc., until H (a) <H (0). Then set a = 0, b = a and c = 2a.

It should be noted that the modified coefficients of connectivity among the neural elements wij (n + 1) and the evaluation function H (a) are defined respectively by equations (14) and (15) as follows

WiXn + l) = WiXn) + aAWij (14)

H (a) = J (v / ij: Wij + aAWij) (15)

2. If 1-1 (1) <H (0), find a, b and c that satisfy H (b) <H (c). In other words, continuously increase the value of a until H (ct) becomes larger than H (a) immediately before. Then record the values of a, b and c for which H (c), the current H (a), is greater than H (b), the H (a) immediately before.

3. Calculate ae according to equation (16) given as follows!

at. = JL H (a) (c'-b2) + H (b) (a2-c2) + H (c) (b'-a2) (16)

2 H (a) (c-b) + H (bXa-c) + H (cXb-a)

4. If H (ae) <H (b), use ae as a found optimum. Otherwise, use b as a found optimum a.

The value of a found in the manner described above is introduced into equation (3) to calculate the correction amounts AWjj (t) and modify the weights of the connectivity among the neural elements.

With the learning algorithm described above, the time required to complete the learning process can be reduced to 1/3 that of the conventional steepest descent method.

Third form of execution:

Fig. 9 is a diagram showing the configuration of a voltage and reactive power control of the electricity generating system performed in accordance with the third aspect of the present invention. In the figure, the same notations are used as those used previously to indicate the same or equivalent components, and no description will be given.

Reference 5 in the figure is a neural network simulator, also known as a neural network simulator of the electricity producing system which learns the input / output relationships of electricity producing systems other than its own power plant. from data estimated on the basis of the reactive power passing through the transformer primary Qi and the input / output relationships (power condenser Se) of its own power station during a process of learning the input / output relationships of a system producing electricity 1, in advance.

Next, the operating principle of the voltage and reactive power control of the electricity producing system is described as follows.

When the neural network simulator 5 learns, in advance, the input / output relationships of the system producing electricity 1, the information on its own power plant such as the primary and secondary transformer voltages Vi and V2, the active powers passing through the primary and secondary transformer Pi and P2, the reactive powers passing through the primary and se7

5

10

15

20

25

30

35

40

45

50

55

60

65

CH 690 637 A5

transformer transformer Qi and Q2, a switchgear with charged sockets (socket) and the power capacitor (Se) can be obtained easily, since these pieces of information are local data as shown in fig. 11. In this voltage and reactive power control of the system producing electricity, these elements of information are used to estimate the input / output relationships of a power plant located on another site (or the capacitor Se of the other site which is called here to simplify "other Se") in order to eliminate bulky circuits, installation of cables and the like.

If a more detailed description is given, the reactive power passing through the transformer primary Qi of its own power plant, which can be obtained locally on its own site, is a reflection, among other things, of the operations of all power plants, including included its own power plant. Under these conditions, the difference between the reactive power passing through the transformer primary Qi and the value of the power capacitor Se of its own power plant is a reflection of the variations in reactive power of power plants other than its own.

The factors of variation of the reactive power of power plants other than its own = Qi + Se.

In other words, the fact of incorporating a power capacitor in its own power plant decreases the reactive power passing through the transformer primary Qi by the reactive power for incorporating the power capacitor Se. By eliminating the stationary elements corresponding to the variations in reactive power consumed by the load, for example, one can obtain the other Se for the control operation of power plants other than one's own. It should be noted that by "elements in the stationary state" is meant elements with a lower degree of variation. In reality, however, the elements are eliminated by a filter or the like, since the variations in stationary state caused, inter alia, by the power consumed, actually exist (see fig. 12).

In a second learning phase where the neural network simulator 5 is connected to a neuroVQC 4, the other Se for the operation of control of power plants other than his is set to zero. Under these conditions, the neural network simulator can learn only the effects provided by its own power plant.

A typical technique for estimating a power outlet operation from a power plant located on another site is described as follows. Fig. 13 is a diagram showing a method for estimating a power outlet operation carried out by a power station located on another site. In the figure, the points representing the relationships of the values of the primary and secondary transformer voltages Vi and V2 have been plotted at different times. The figure shows that the resulting curve is not continuous at certain points and that the slopes of the curve do not form a continuous line at certain points. The first discontinuities are caused by the power capacitor Se from its own power station or another power capacitor Se from another power station, while the second discontinuities are due to a power outlet operation at its own power station or to another power plant.

Information on the power outlet operation can be obtained from your own power plant. Indeed, when the power outlet of its own power plant is not made, the second discontinuities can be used to estimate a power outlet operation of a power plant located on another site, assuming that the outlet current from its own power plant was carried out.

Fourth form of execution:

The diagram in fig. 14 shows the configuration of a voltage and reactive power control of the electricity producing system produced in accordance with a fourth aspect of the present invention. The reference 11 shown in the figure is a relearning evaluation unit used in a voltage and reactive power control of the electricity producing system, where the control of a power station 1 is switched from a neuroVQC 4 to a control unit VQ 3 and on a neuroVQC 12 intended for learning and on a neuro-system simulator 13 also provided for learning, as well as for learning the new input / output relationships of the system producing electricity 1 and record the learning results in neuroVQC 4 and in a neuro-system simulator 2 in the event of detection of an error between the input / output relations of the neuro-system simulator 2 and the input / output relations of the producing system of electricity 1. Reference 14 is a storage unit for preserving the input / output relationships of the system producing electricity 1.

Next, we will explain the operating principle.

When the voltage and reactive power control of the electricity-producing system is actually put into operation, the neuro-system simulator 2 and the neuroVQC 4 first undergo a learning process at the power plant and then the operations of the electricity generating system 1 are entered as input / output information.

Fig. 15 shows an example of a current request. For such a current demand, the information elements such as the primary and secondary voltages of transformer Vi and V2, the active powers passing through the primary and secondary of transformer Pi and P2 and the 8

5

10

15

20

25

30

35

40

45

50

55

60

65

CH 690 637 A5

Reactive elements passing through the primary and secondary transformer Ch and Q2 are sampled successively with a typical sampling period of 10 seconds, as shown in the figure. The pieces of information sampled for a period ranging from several minutes to several tens of minutes are learned as a reason.

Then neuroVQC 4 controls the electricity producing system 1 using the learning results as the basis. However, the characteristics of the system producing electricity 1 during the learning period may differ from its real characteristics. Consequently, the relearning evaluation unit 11 between the outputs of the neuro-system simulator 2 and those of the system producing electricity 1 to detect any error, a disagreement between the outputs.

If an error is detected, the relearning evaluation unit 11 decides that it is necessary for the neuro-system simulator 2 and the neuroVQC to undergo a relearning process.

When it is decided that a relearning process is necessary, the relearning evaluation unit 11 temporarily switches the control of the electricity producing system 1 of the neuroVQC 4 to the control device VQ 3 , because the detected error indicates that the neural network of neuroVQC 4 is not adjusted to an optimum state.

As the most recent information on the input / output relationships of the electricity-producing system 1 is recorded in the storage unit 14, the relearning evaluation unit 11 requests the neuro-system simulator 13 used to for learning purposes and neuroVQC 12 also used for learning purposes to undergo a relearning process based on the most recent information.

In response to the request made by the relearning evaluation unit 11, the neurosystem simulator 13 and the neuroVQC 12 carry out the relearning process. At the end of the relearning process, the neuro-system simulator 13 in learning and the neuroVQC 12 in learning copy the results respectively to the neuro-system simulator 2 and to neuroVQC 4. A series of operations in the process of learning is finally carried out by returning the control of the system producing electricity 1, of the control unit VQ 3, on the neuroVQC 4.

As mentioned above, the learning algorithm for a neural network used in accordance with the first aspect of the present invention has the effect of allowing the learning process to be completed in a short period of time, even if the target values are changed abruptly, because with this learning algorithm we first find the errors between the output values of the neural network and the target values, then we calculate a sum of the squares of the errors and the squares of the changes of error and finally we modify the weights of the connectivity among the neural elements of the neural network in a direction to decrease the sum only if it is found that the sum is greater than a predetermined value.

Furthermore, the learning algorithm for a neural network used in accordance with the second aspect of the present invention allows the learning process to be completed in a short period of time even when the target values change abruptly, because with this algorithm, we first determine the errors between the output values of the neural network and the target values, then we calculate a sum of the squares of the errors and finally we modify the weights of the connectivity among the neural elements of the neural network using the first order search technique in one direction to decrease the sum only if it is found that the sum is greater than a predetermined value.

In addition, the voltage and reactive power control of the system producing electricity has the effect of making it possible to obtain and learn information on the input / output relationships of power stations located on other sites without 'it is necessary to install cables or the like, because in the voltage and reactive power control of the electricity producing system produced according to the third aspect of the present invention, the neural network of the neural network simulator of the producing system of electricity learns the input / output relationships of systems producing electricity other than its own power plant from data estimated on the basis of the reactive power passing through the transformer primary and the input / output relationships of its own power plant, in advance, during a process of learning the input / output relationships of systems producing electricity you.

In addition, the voltage and reactive power control of the electricity producing system has the effect of allowing the characteristics of the neural network voltage and reactive power control to be adjusted to values which are optimal for the characteristics. of the system producing electricity, without interrupting the control of the system producing electricity even if the characteristics of the system producing electricity change, because the voltage and reactive power control of the system producing electricity, realized according to the fourth embodiment of the present invention is equipped with a relearning evaluation unit, allowing the control of a system producing electricity to be switched from the voltage and reactive power control to the network neuronal on voltage and reactive power control and on voltage and reactive power control e with neural network intended for learning and on the neural network simulator of the system producing electricity also provided for

9

5

10

15

20

25

30

35

40

45

50

55

60

65

CH 690 637 A5

learning, as well as to learn the new input / output relationships of the electricity producing system and record the learning results in the neural network voltage and reactive power control and in the neural network simulator respectively system producing electricity in the event of an error being detected between the input / output relationships of the neural network simulator of the system producing electricity and the input / output relationships of the system producing electricity.

Claims (1)

Claim 1. Device for controlling voltage and reactive power of a system producing electricity from a power plant, comprising: a voltage and reactive power control with a neural network comprising a first neural network capable of performing, in advance , a learning process of matching the input / output relationships thereof to the input / output relationships of a voltage and reactive power command, and a neural network simulator of the electricity producing system comprising a neural network capable of carrying out, in advance, a learning process consisting in matching the input / output relationships thereof to the input / output relationships of the system producing electricity under control, where said inputs and outputs of said neural network voltage and reactive power control are connected to the inputs and outputs of said neural network simulator of the s a system producing electricity and said neural network reactive power and voltage control able to use the results of a relearning process intended to match the inputs of said neural network voltage and reactive power control to the outputs of said neural network neural network simulator of the electricity producing system for controlling said electricity producing system, said voltage and reactive power control of the electricity producing system capable of estimating the input / output relationships of electricity producing systems electricity other than its own power plant, on the basis of the information on the reactive power passing through the transformer primary and the input / output relationships of its own power plant during said learning process carried out in advance by said network of said neural network simulator of the system producing electricity to learn the input / output relationships of said electricity producing system. 10