JP2547833B2 - Computer control system - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Advantages of the Invention] (Field of Industrial Application) The present invention relates to a computer control system for controlling a large-scale object such as a power system or an industrial plant.

(Prior Art) As a computer control system for controlling an electric power system, an industrial plant, etc., there is one as shown in FIG. As shown in the figure, this calculation control system includes a control target 1, a control unit 2, state quantity transmission units 3a, 3b, 3c, ..., Parameter transmission units 4a, 4b, 4c ,. It has a man-machine interface section 7.

The controlled object 1 is a controlled object such as an electric power system or an industrial plant. The control unit 2 is a generator control device or the like if the control target 1 is a power system, and a furnace or motor control device or the like if the control target 1 is a steel mill system.

The state quantity transmitters 3a, 3b, 3c ... Transmit the state quantity indicating the state of the controlled object 1. The parameter transmission units 4a, 4b, 4c ... Transmit parameters that affect the operation and operation method of the controlled object 1. The receiving unit 5 receives the information transmitted from the state quantity transmitting units 3a, 3b, 3c ... And the parameter transmitting units 4a, 4b, 4c.

The computer 6 generates a control signal to the control unit 2 based on the state quantity collected by the receiving unit 5 or monitors the presence or absence of an abnormality in the control target 1, and if the abnormality is evaluated, operates the result. It has functions such as notifying employees.

The man-machine interface unit 7 is a device for displaying the above-mentioned evaluation results, or an assembly of operating devices for the operator to send signals to the computer and the control unit.

Such a computer control system has a function of evaluating information about a control target in order to determine the control content.For example, in the overload elimination control of the transmission line in the power system control, the current flowing through the transmission line is The state quantity is sent to the computer 6, and the computer 6 compares it with the maximum allowable current of each transmission line, and if there is a transmission line that exceeds it, evaluates it as an overload occurrence as a system state and eliminates it. I do.

As in this example, if the judgment condition for performing the state evaluation necessary to perform a certain control is simple, the state evaluation is easy, but the scale of the controlled object and the required control become complicated. As a result, it becomes difficult to formulate the judgment conditions for state evaluation, and the need to use empirical knowledge is increasing.

Further, in the above example, it is sufficient to make a determination for each current of each transmission line, and it is not necessary to comprehensively grasp the current distribution as one pattern, but depending on the content of control, it may be comprehensively determined as one pattern of states. It is necessary to analyze and evaluate the state of the controlled object based on the result. For example, when considering the situation where the stationary stability and the voltage stability gradually decrease due to heavy load in the power system, it is important for the operator such as the power supply dispatching center to operate to cope with this. However, for that purpose, the computer must supply information to the operator so that the operator can correctly grasp the state of the system. Then, for such purpose, it is necessary to comprehensively evaluate the state quantities such as the voltage of the node, the phase difference angle, and the passing power of the transmission line as a pattern. As an evaluation method for this, for example, when a state quantity pattern that is a system is observed, the pattern is represented as a representative pattern from the degree of similarity to some representative state quantity patterns for which evaluation results have already been given. One of the methods is to classify it into any of the above and regard it as the evaluation of the observed pattern based on the evaluation contents of the representative pattern.

However, a function for comprehensively evaluating the state of the controlled object based on empirical knowledge from data such as state quantities given as a pattern is required, but with the current computer system configuration, for example, in the computer of an expert system Considering the introduction of software based on such a knowledge engineering method, some problems arise in practice.

That is, in order to detect the feature of the state quantity pattern using the conventional knowledge engineering method, it is necessary to perform a combination of a large number of judgment conditions, but in the case of such a method, the pattern has a predetermined judgment condition. If any one of them does not match, the pattern is excluded even if the entire pattern roughly matches the feature corresponding to the determination condition. That is, the combination of the judgment conditions is too strict, and there is a possibility that the classification cannot be roughly performed. In addition, it requires a considerable amount of labor to rigorously develop empirical knowledge for feature detection into a specific technical style such as a production rule to create a combination of determination conditions. Further, in an actual application, what was considered to be the representative pattern of the state quantity of the control target at the time of creating the rule may be slightly different from the representative pattern in the actual system. In this case, after introducing the expert system to the actual system,
Although it becomes necessary to adjust the rules, a group of rules for extracting a certain feature as a pattern have interrelationships, and thus there are various problems if they cannot be changed easily.

(Problems to be Solved by the Invention) As described above, in the conventional computer control system, the state of the controlled object is comprehensively evaluated from the global characteristics of the pattern of the controlled object state quantity and the patterns of other various quantities. I couldn't. In addition, there are various problems such as the need to develop the empirical knowledge for the evaluation into a strict description format, and the adjustment after introduction into the actual system is not easy.

The present invention has been made in view of such problems,
The purpose is that the pattern of the state quantity of the controlled object can comprehensively evaluate the state of the controlled object from the global characteristics of the patterns of other various quantities, and the empirical knowledge for the evaluation can be strictly evaluated. An object of the present invention is to provide a computer control system that does not need to be expanded to a description format and that can be easily adjusted after being introduced into an actual system.

[Structure of the Invention] (Means for Solving the Problems) In order to achieve the above-mentioned object, the present invention is composed of a layered neural network on which a predetermined learning is performed, and a pattern of the state quantity to be controlled from the input layer. And an evaluation unit that outputs evaluation results of a plurality of evaluation items for the control target from the output layer, and a unit that controls the control target based on the evaluation result output from the evaluation unit. .

(Operation) In the present invention, by adding the evaluation unit using the neural network, for example, the state quantity of the power system, that is, the voltage of each node, the pattern of the power passing through the transmission line, and other quantities, such as the temperature and the thundercloud. It is possible to comprehensively evaluate the state in which the power system is placed, for example, whether or not the system is in a dangerous state when considering the stability, from the pattern of the amount indicating the occurrence state. The empirical knowledge for this evaluation does not need to be developed into a strict description style such as production rules. In addition, adjustment after introduction into the system becomes easy.

(Embodiment) An embodiment of the present invention will be described in detail below with reference to the drawings. However, since the contents of control performed by a computer control system for operating a power system are diverse,
Particularly, voltage control will be described as an example.

FIG. 1 is a block diagram showing the configuration of a computer control system according to an embodiment of the present invention, and elements having the same functions as those of the conventional example shown in FIG. 9 are given the same numbers. This embodiment is characterized in that the evaluation section 8 is added.

In the case of the present embodiment, the controlled object 1 is a power system. In the case of voltage control, the control unit 2 is a terminal voltage adjusting device of a generator, a reactive power compensating capacitor, a reactor and the like.

The state quantity transmitters 3a, 3b, 3c ... Send the state quantities measured by the sensors in the power system to the receiver 5. This state quantity is the voltage value of each node (substation) in this embodiment. For example, the state quantity transmitter 3a has a bus voltage of a substation A (not shown), and the transmitter 3b has a bus voltage of a substation B (not shown). Sent to 5.

The parameter transmission units 4a, 4b, 4c, etc. transmit to the reception unit 5 parameters other than the state quantity that affect the operation and operation method of the power system, such as the amount of power demand and the amount of thundercloud occurrence. However, in the case of the voltage control according to the present embodiment, these are not used and the evaluation unit 8 performs only the pattern of the voltage value which is the state quantity.

In the case of the present embodiment, the voltages V ₁ , V ₂ , V ₃ ... Of the 25 main substations in the power transmission system are used, and these voltages are sent directly to the computer 6 which is the arithmetic unit, and logically Although it is also used for processing, it is directly applied to the neural network of the evaluation unit 8 at the same time. The evaluation unit 8 performs a neural network operation (neural network) from the voltage pattern, evaluates the state of the system, displays the evaluation result on the man-machine interface 7 and notifies the operator at the same time. A control signal is sent to the control unit 2 based on the evaluation result.

In the evaluation unit 8, three layers s as shown in FIG.
A neural network of t and u is used. This neural network imitates the neural network of a living body, and is a collection of nodes corresponding to neurons (nerve cells). Nodes are also called units. Unit j has an input / output characteristic fj (uj), but in most cases an S-shaped function is used. The output vj of this unit j is determined by the output of other units and this function, and is represented by the following equation.

vj = fj (Σ w ji vi−θj) (1) where vj is the output of another unit i, w ji is a weighting coefficient indicating the degree of influence of the output of unit i on unit j, and θj is It is a threshold.

In this embodiment, the voltage pattern is input to the input layer neuron s consisting of 25 nodes, and the output value is output from the 12 neurons of the output layer u, and the output is 12 shown in FIG.
It corresponds to each evaluation item.

The evaluation unit 8 evaluates the input voltage pattern based on empirical knowledge, and 14 patterns are used as typical patterns observed in the power system targeted in this embodiment.

4 and 5 show the pattern. That is, the horizontal axis represents the node numbers of 25 substations, and the vertical axis represents the voltage value sent from the substations. 4 and 5 show two of the 14 patterns.

FIG. 3 shows the evaluations for these 14 patterns, which were obtained empirically. In this figure, items with circles are evaluation items that are satisfied for the pattern.

In the conventional expert system, the relationship between each pattern and the evaluation item must be developed into a strict description form such as a rule, but in the present invention, the need is eliminated by using a neural network, and the pattern and its pattern are learned. Can be associated with the evaluation items that are satisfied.

When a certain representative pattern is input to the input layer of the learned neural network, the neuron in the output layer corresponding to the evaluation item established for the representative pattern outputs a large output. In this embodiment, a back propagation method which is widely used as a learning method is adopted. According to this method, it is not necessary to associate the representative patterns of FIGS. 4 and 5 with the evaluation items of FIG. 3 in a strict description format of rules as in the conventional expert system. And a pattern of evaluation items corresponding to it, for example, an output layer neuron corresponding to a satisfied evaluation item is set to 1.0, and a pattern corresponding to a failure evaluation item is set to 0.0. It suffices to modify the many existing weighting factors little by little. The weighting coefficient indicates w ij in the equation (1), that is, the strength of the connection indicated by the link of the neural network in FIG. The repetition for this learning took about 20,000 times in this example.

Next, the calculation result in this embodiment will be described. Sixth
As an example, the output of the output layer unit when the pattern 5 is input to the learned neural network is shown as an example, and the unit corresponding to the establishment evaluation item of this pattern outputs a large output, that is, an output close to 1.0.

FIG. 7 shows the output of the output layer unit when the pattern 5 mixed with about 20% of the heterogeneous pattern 6 is input to the learned neural network.
As long as the overall and global characteristics of the pattern are retained to some extent, the unit corresponding to the evaluation result of the pattern shows a large output even if the pattern is considerably altered in this way, and the evaluation unit We are doing a comprehensive evaluation of the pattern.

The lack of the ability to deal with the inconsistency between the representative pattern assumed in advance and the representative pattern actually observed, which has been mentioned as a problem of the conventional technique, can be solved by adding an evaluation unit using a neural network. To show this, it is assumed that the 14 representative patterns assumed and the representative patterns actually observed have the same number of patterns but the shapes of the patterns are slightly different. FIG. 8 shows an example of new representative patterns corresponding to the old pattern 5. Without repeating the learning for this new pattern from the beginning, the learning was performed by giving the pair of the new pattern and the evaluation item to the already learned neural network for the old pattern. A new representative pattern can be associated with an evaluation item by learning a fraction of the number of times.

The present invention is not limited to the embodiments described above and shown in the drawings, and various modifications can be made without departing from the scope of the invention. It also realizes the characteristics of a neural network whose structure is such that the output of a neuron is determined by the input given as the weighted sum of the outputs of other neurons and the threshold value associated with that neuron. If so, the evaluation unit may be an analog electronic circuit, or may be simulated by a mathematical formula in a dedicated computer, and various embodiments can be adopted according to the application field, and the present invention is not limited to a specific implementation method. . Further, the present invention is not limited to the neural network in which the weighting coefficient is determined by the backpropagation method, and includes neural networks by various other learning methods.

[Effect of the Invention] As described in detail above, according to the present invention, by adding a state evaluation unit using a neural network, a strict description format of knowledge for evaluating a state is not required, In addition, a computer control system that can comprehensively and globally evaluate the characteristics of patterns and can easily deal with inconsistencies in representative patterns that are actually observed with representative patterns that have been assumed in advance. Can be provided.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of a computer control system according to an embodiment of the present invention, FIG. 2 is a diagram showing a neural network used in the evaluation unit 8, and FIG. 3 is a diagram showing evaluation items for voltage patterns. , Fig. 4 and Fig. 5 are diagrams showing voltage patterns observed in the target electric power system,
6 and 7 are diagrams showing the output of the neural network, FIG. 8 is a diagram showing an actual voltage pattern that is slightly different from a voltage pattern assumed in advance, and FIG. 9 is a block diagram showing the configuration of a conventional computer control system. It is a figure. 1 ... Control object 2 ... Control unit 3a, 3b, 3c ... State quantity transmitting unit 4a, 4b, 4c ... Parameter transmitting unit 5 ... Receiving unit 6 ... Computer 7 ... Man-machine interface unit 8 ... Evaluation department

Claims (1)

(57) [Claims] 1. A layered neural network that has been subjected to predetermined learning, in which a pattern of state quantities of a control target is input from an input layer, and evaluation results of a plurality of evaluation items to be controlled targets are output from an output layer. And a means for controlling the controlled object based on the evaluation result output from the evaluation section.