JPH0439277A - Elevator control device - Google Patents

Abstract

PURPOSE:To improve arrival time predictive precision by converting traffic state data containing a cage position, an operation direction, and a call being to respond to a state usable for a neural net to input it to the neural net, and computing the arrival predictive time of a cage. CONSTITUTION:A data exchange means 10C converts traffic state data, such as a cage position, an operation direction, and a call being to respond, to data usable for a neural net to input it to the neural net of an arrival predictive time computing means 10D. The arrival predictive time computing means 10D computes an arrival predictive time by learning and correcting input data by means of a correcting means 10G by referring to data of a data-for-learning producing means 10F. A computing result is converted to a form usable for operation of a given control purpose by means of a data converting means 10C to output the form. This constitution enables a computing predictive time to approach an actual arrival time and improves group management performance.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] This invention relates to an elevator control device that can accurately predict the time required for an elevator car to arrive at each floor.

[Prior Art] Conventionally, in an elevator system in which a plurality of cars are installed together, group management operation is usually performed, and one example of such group management operation is, for example, an assignment method. The allocation method means that as soon as a hall call is registered, an evaluation value is calculated for each car, and the car with the best evaluation value is selected as the allocated car to be serviced, and only the allocated car is used for the above hall call. This is intended to improve operational efficiency and shorten waiting time by making the system respond.

At this time, the estimated waiting time of the hall call is generally used to calculate the evaluation value.
In the elevator group management device described in Publication No. 64, when a hall call is registered, the sum of the square values of the predicted waiting times of all hall calls when the hall call is temporarily assigned to each car is calculated. The evaluation value is obtained, and the car with the minimum evaluation value is selected as the allocated car.

In this case, the predicted waiting time consists of the duration of the hall call (the time that has passed since the hall call was registered) and the expected arrival time (the time from the current position until the car arrives at the floor of the hall call mentioned above). (predicted value of the time required).

By using the evaluation value obtained in this way, it is possible to reduce the waiting time of hall calls (in particular, to reduce the number of long-waiting calls of one minute or more).

However, if the accuracy of the expected arrival time is lost, the evaluation value has no meaning as a reference value for selecting an assigned car, and ultimately it becomes impossible to reduce the waiting time for hall calls. Therefore, the accuracy of the expected arrival time has a large impact on the performance of group management.

Next, a conventional method of calculating an expected arrival time will be specifically explained.

The expected arrival time is calculated as shown in (A) below, assuming that the car runs back and forth between both roadside floors.

(A) Calculate the time required to travel (traveling time) from the distance between car position 1 and the target floor, calculate the time required to stop (stopping time) from the number of stops at floors along the way, and then The estimated arrival time is obtained by adding the time of
(See Japanese Patent Publication No. 20742 and Japanese Patent Publication No. 54-34978).

Further, in order to improve the prediction accuracy of the stop F time on the floor where the car is located or the floor where the car is scheduled to stop, prediction methods as shown in (B) to (E) below have been proposed.

(B) Correct the expected arrival time according to the car condition on the floor where the car is located (decelerating, door opening, door opening, door closing, running, etc.) (Special Publication No. 57-40074) (see publication).

(C) Detect the number of people getting on or getting off at the scheduled stop floor using a detection device or a prediction device, and correct the estimated arrival time according to these numbers of people (Japanese Patent Publication No. 57-40072 and Japanese Patent Application Laid-Open No. 58-162472).

(D) The expected arrival time is corrected, taking into account that the boarding and alighting times differ depending on whether the scheduled floor is a car call response or a hall call response (see Japanese Patent Publication No. 57-40072).

(E) Each floor is based on statistical data of the actual stop time (door opening time, boarding/alighting time, door closing time) for each floor, and the door opening time determined by simulation and built into the group control device. Predicting stoppage time for each floor (Unexamined Japanese Patent Publication No. 1
(See Japanese Patent Laid-Open No. 59-138579).

In addition, when considering the possibility that a future call will be registered on a floor where the car is not scheduled to stop and the car will stop, the methods shown in (F) to (H) below can be used to improve the accuracy of arrival prediction. is proposed.

(F) The number of car calls caused by stopping in response to a landing call on an intermediate floor is predicted based on statistical data regarding the number of passengers in the past, and further based on the statistical probability distribution of car calls that have occurred in the past. , the predicted number of car calls is distributed to the floors in front of it, and the stop time due to derived car calls is predicted (see Japanese Patent Publication No. 63-34111).

(G) Calculate the probability that the car will stop for each floor and direction based on the number of times the car reverses direction and the number of people getting on and off by direction in the past, and correct the expected arrival time based on the calculation results ( (Refer to Japanese Patent Application Laid-Open No. 5926872).

() () Predict the stop time due to car calls at each floor based on the alighting rate for each floor calculated for each floor direction (Special Publications Publication No. 63
(Refer to Publication No.-64383).

In addition, cars often operate by reversing direction at the floor midway depending on the highest call or lowest call, but in order to prevent an error between the expected arrival time and the actual arrival time, the following (I) and (Jo), prediction methods have also been proposed in which the direction is reversed at an intermediate floor before reaching the terminal floor.

(I) Calculate the expected arrival time by determining the travel time to the floor with the farthest call in front of the car in the direction of travel, and the travel time from that floor to the floor with the call in the opposite direction.
(See Japanese Patent Publication No. 54-16293).

(J) Empty Cars For cars for which the running direction has not been set, the expected arrival time is calculated on the assumption that they will go directly to each floor (see Japanese Patent Publication No. 8621/1986).

In this case, the upper reversal floor (the highest call reversal floor) is usually set to the highest call floor, and the lower reversal floor (lowest call reversal floor) is usually set to the lowest call floor. be done. However, for example, even if an upward reversal floor is set, if there is a landing call in the up direction on a floor halfway, it is necessary to predict the occurrence of a new car call, and the upward reversal floor can be accurately set. Similarly, it is difficult to accurately set a downward inversion floor. As a result, the number of error factors increases because the other condition of an inverted floor is predicted and calculated.

Furthermore, as described in Japanese Patent Application Laid-Open No. 1-275381, a group management control device has been proposed which selects a car to be assigned to a hall call based on calculations using a neural network that does not correspond to neurons in the human brain. However, it does not consider improving the calculation accuracy of the expected arrival time or the expected degree of congestion in the car [Problem to be solved by the invention] As described above, conventional elevator control devices In order to accurately calculate the time, various factors are used: the current car status, the predicted number of passengers getting on and off at the stopping floor, the type of the current call to be answered, the prediction of the occurrence of car calls, and the allocation for new hall calls. , the prediction of reversing floors, the current traffic conditions on each floor, etc., are considered, and each of these is calculated as one element of the calculation formula. However, if we try to make predictions by taking into account all of these factors and accurately calculate traffic conditions that are constantly changing in complex ways,
The calculation formula for the expected arrival time becomes even more complex, and given the limitations of human ability, it becomes difficult to develop new calculation formulas with the aim of improving calculation accuracy. On the other hand, if detailed prediction calculations are performed, the calculation time increases, and there is a problem that it is impossible to realize the function of determining the assigned car and predicting the expected arrival time at the same time as registering the hall call. Ta.

This invention was made to solve the above-mentioned problems, and by making flexible predictions that approximate actual traffic conditions and traffic volume, it is possible to predict accurate arrival times close to actual arrival times. [Means for Solving the Problems] An elevator control device according to the present invention inputs traffic condition data including a car position, a running direction, and a call to be answered into a neural network. There is an input data conversion means that converts it into a form that can be used as data, an input layer that takes in the input data, an output layer 5 that outputs data corresponding to the expected arrival time, and a weighting coefficient between the input layer and the output layer. The apparatus includes an expected arrival time calculating means that includes a set intermediate layer and constitutes a neural network, and an output data converting means that converts output data into a form that can be used for a predetermined control operation.

Moreover, an elevator control device according to another invention of the present invention includes:
At a predetermined time while the elevator is in operation, it stores the expected arrival time at a predetermined landing and the input data at that time, and counts the time elapsed until the car stops at or passes the predetermined landing. Stored as actual arrival time,
learning data creation means for outputting the stored input data, expected arrival time, and actual arrival time as a set of learning data; and correction means for modifying the weighting coefficient of the expected arrival time calculation means using the learning data. It is further equipped with the following.

[Operation] In this invention, traffic condition data is input into a neural network, an expected arrival time is calculated by calculations that approximate the actual car arrival time, and this expected arrival time is used to perform elevator operation in accordance with a predetermined purpose. control.

Further, in another invention of the present invention, learning data is created based on the calculated prediction result, traffic condition data at that time, and actual measurement data, and based on the learning data, a neural network is applied to the expected arrival time calculation means. By automatically modifying the weighting coefficients in the net), flexible prediction calculations that approximate actual traffic conditions and traffic demand are performed.

[Example] Hereinafter, an example of the present invention will be described with reference to the drawings. 1st
The figure is a functional block diagram showing the overall configuration of an embodiment of the present invention, and FIG. 2 is a block diagram showing the schematic configuration of the group management device in FIG. 1.

In FIG. 1, the group management device t<10) is functionally composed of the following means (IOA) to (1, (ID), (IOF Controls the car control devices (11) and (12) of the car (for No. 2 car).

The hall call registration means (IOA) registers and cancels the hall calls for each floor (upward and downward hall calls), and records the elapsed time since the hall call was registered (i.e.,
duration).

An allocating means (IOB) that selects and allocates the best car to service a hall call may, for example, predict and calculate the waiting time for each car to respond to a hall call on each floor, and calculate the square value of the waiting time for each car to respond to a hall call on each floor. Allocate the car that minimizes the sum of .

The data conversion means (IOC) is input data that converts traffic condition data such as car position, direction of operation, and calls to be answered (car calls or assigned platform calls) into a form that can be used as input data for the neural network. and an output data conversion means for converting the output data of the neural network (data corresponding to the expected arrival time) into a form that can be used for a predetermined control purpose operation (for example, calculating a predicted waiting time). .

The expected arrival time calculation means (IOD) that calculates the expected arrival time of each car according to the time zone has an input layer that takes in input data and an output that outputs data corresponding to the expected arrival time, as described later. It includes a neural network consisting of a layer and an intermediate layer between the input layer and the output layer in which weighting coefficients are set.

The learning data creation means (IOF) includes the expected arrival time of each car and the input data at that time (traffic condition data),
Subsequent actual measurement data (teacher data) regarding the arrival time of each car are stored, and these are output as learning data.

The modification unit (IOG) uses the learning data to learn and modify the function of the neural network in the expected time of arrival calculation unit (IOD).

Car control device (II) and (1) for Unit 1 and Unit 2
2) have the same configuration. For example, the car control device (11) for the first car uses the well-known means (I) as follows.
It consists of IA) to (IIE).

The hall call cancellation means (IIA) outputs a hall call cancellation signal for the hall call of each floor. Car call registration means (
IIB> registers the car call for each floor.

The arrival forecast light control means (IIC) controls the lighting of arrival forecast lights (not shown) on each floor. Operation control means (IID
) controls the running and stopping of the car in order to determine the running direction of the car and to respond to the hall calls assigned to the car. The door control means (11, E) controls opening/closing of the door at the entrance/exit of the car.

Further, in FIG. 2, the group management device f (10) is composed of a well-known microcomputer, and includes an MPU (microprocessing unit) or a CPU (101),
It is composed of a ROM (102), a RAM (103), an input circuit (104), and an output circuit (105).

The input circuit (104) has a six-way terminal that receives the hall button signal (14) from the hall button of each floor and the status signals of the first car and the second car from the car control devices (11) and (12). ,
The output circuit (105) outputs a hall button light signal (15) to the hall button light built into each hall button, and a car control device (11).
) and a command signal to (12) are output.

FIG. 3 is a functional block diagram specifically showing the relationship between the data conversion means <10C) and the expected arrival time calculation means (IOD> in FIG. 1).

In FIG. 3, the input data conversion means, that is, the input data conversion subunit (10C^), and the output data conversion means, that is, the output data conversion subunit (IOCB) are the first
It constitutes the data conversion means (IOC) in the figure. or,
The expected arrival time calculation unit (10D^) inserted between the input data conversion subunit (IOCA) and the output data conversion subunit (IOCB>) consists of a neural network and is the expected arrival time calculation means in FIG. (IOD
) constitutes a prediction calculation subroutine used in

The input data conversion subunit (10C^) inputs information such as vehicle location, direction of travel, calls to be answered (i.e. vehicle calls and assigned platform calls), statistical characteristics of traffic flow (number of passengers per 5 minutes, 5 Converts traffic condition data such as the number of people getting off the train per minute into a form that can be used as input data for the neural network (10D^).

The output data conversion subunit (IOCB) converts the output data (data corresponding to the expected arrival time) of the neural network (10D^) into a form that can be used to calculate the evaluation value of the hall call allocation operation.

Estimated arrival time calculation unit consisting of neural network (
IODA) is an input data conversion subunit (IOCA).
), an output layer (10D^3) that uses data corresponding to the expected arrival time as output data, an input layer (100^1) and an output layer (10D ^3) and an intermediate layer (10D^2) in which weighting coefficients are set.

Each of these layers (IOCA1) to (IOCA3) are connected to each other via a network, and are each composed of a plurality of nodes.

Here, input layer (10D^1), intermediate layer (IOCA2)
and the number of nodes in the output layer (10D^3), respectively, are N
1, N2, and N3, the number of nodes N3 in the output layer (10D^3) is N5 = 2 (FL-1> where FL is the number of floors in the building, and the number of nodes in the input layer (10D^1 ) and intermediate layer (10D
The numbers N1 and N2 of nodes in ^2) are determined by the number of floors PL of the building, the type of input data to be used, the number of cars, etc., respectively.

Also, the variables i, j, and k are set as i=1.2. ..., N1 j=1.2. ..., NZ k=1.2. ..., N3, the input value and output value of the i-th node of the input layer (100^1) are xal(i) and yal(iL intermediate layer (1
The input value and output value of the j-th node of 0D^2) are xa2(
j) and ya2(j), the input value and output value of the node of the output layer (10D^3) are xa3(k) and ya3(k
).

Also, the i-th node of the input layer (10D^1) and the intermediate layer (10
wal(i,
j), the weighting coefficient between the intermediate layer (10DA2) j/-F and the output layer (10DA3) j/-
(j, k), the relationship between the input value and output value of each node is yal(i)=1/[1+expf-xal(iN]
・＋■x a2(j)−Σfwal(i,j)X
yal(i)) ...■ (summation formula using i=1 to N1) % formula% (summation formula using j=1 to N2) ya3(k>=1/rl+exp(-xa3(k))1
It is represented by +++■. However, 0≦wal(i, j)≦1 and 0≦wa2(j, k)≦1.

Fig. 4 is a flowchart diagram schematically showing the group management program stored in the ROM (102) in the group management device (10), and Fig. 5 is the arrival time prediction when the plate for No. 1 in Fig. 4 is allocated. A flowchart diagram specifically showing the program,
FIG. 6 is a flowchart specifically showing the learning data creation program in FIG. 4, and FIG. 7 is a flowchart specifically showing the modification program in FIG. 4.

Hereinafter, with reference to FIG. 4, the group management operation of the embodiment of the present invention shown in FIGS. 1 to 3 will be described.

First, the group management device (10) uses a well-known input program (
According to step 31), the hall button signal (14) and the status signals from the car controllers (11) and (12) are taken in. The status signals input here include the car position, running direction, stop or running status. , door open/close status, car load, car call, cancellation signal for hall call, etc.

Next, the well-known hall call registration program (step 32)
Accordingly, it is determined whether the hall call is registered or canceled, whether the hall hall lights are turned on or off, and the duration of the hall call is calculated.

Next, it is determined whether a new hall call C has been registered (step 33), and if it has been registered, the arrival time prediction program at the time of provisional allocation for car No. 1 is executed (step 34).
), the expected arrival time Ta1(k) for each landing of car No. 1 when a new landing call C is temporarily assigned to car No. 1.
Calculate.

Similarly, the arrival time prediction program (
In step 35), the expected arrival time Ta2 for each landing of the No. 2 car when the landing call C is provisionally assigned to the No. 2 car.
(k) is calculated.

In addition, the arrival time prediction program (steps 36 and 37) for non-tentative allocation in which the new hall call C is ignored and is not allocated to either car 1 or 2 is executed, and Expected arrival time Tb1(k) at the boarding point
and Tb2(k) are calculated.

Next, by the allocation program (step 38), the expected arrival time TaHk) calculated in steps 34 to 37, T
Waiting time evaluation values W1 and W2 are calculated based on a2(k), Tb1(k), and Tb2(k>), and the car with the minimum evaluation value is selected as the regular allocated car. An assignment command and a forecast command corresponding to the hall call C are set in the car.For the calculation method of the waiting time evaluation value W and W2, for example,
It is described in Publication No. 8484.

Next, the output program (step 39) sends the hall magic lantern signal (15) set as described above to the hall, and also sends the assignment signal, forecast signal, etc. to the car control device (
11) and (12).

In addition, in the learning data creation program (step 40), converted traffic condition data and
The predicted arrival time at each landing and the actual measurement data of the subsequent arrival time of each car are stored, and these are output as learning data.

In addition, in the modification program (step 41), using the learning data, the expected arrival time calculation means (IOD>
Modify the weight coefficients of the network.

In this way, the group management device (10) performs steps 31 to 4.
1 is repeatedly executed to perform group management control of a plurality of elevator cars.

Next, the operation of the arrival time prediction program in steps 34 to 37 will be specifically explained with reference to FIG. 5, taking step 34 as an example.

First, a new hall call C is temporarily assigned to car No. 1, and assigned hall call data is created to be input to the input data conversion subunit (10C^) (step 50).

Incidentally, in step 35, the assigned hall call data is created by temporarily allocating to car No. 2, and in steps 36 and 37, the assigned hall call data in the case where no provisional assignment is made is used as is for input as the assigned hall call data.

Next, among the input traffic condition data, data regarding the car (car position,
The expected arrival time calculation unit (10D^ ) input data for each node of the input layer (10D^1) x aHl) ~ x al
(N 1) (step 51).

Here, the number of floors FL of the building is 12 floors, and for the landing number f, f=1.2. ..., 11 are 1, 2, respectively.
...Represents the uphill landing on the 11th floor, f = 12.1
3.・22 are 12 and 11 respectively. . . , represents the downbound landing on the second floor. For example, the car status "car position floor is f, operating direction is up" is xal(f)=1xaHi)=0 (i=1. 2.・ , 22, 1 ≠ f ),
It is expressed as a normalized value between 0 and 1.

In addition, the car calls for the 1st to 12th floors xal (23) to xal (
34) is represented by rl if it is registered, and "0" if it is not registered, and the assigned landing calls xal (35) to If it is assigned, it is expressed as "1", if it is not assigned, it is expressed as rQ, and the assigned hall call xal in the down direction from the 12th floor to the 2nd floor
(46) to xal(56) are "1" if assigned
, if it is not allocated, it is represented as "0".

Also, the number of passengers for 5 minutes in the upward direction from the 1st floor to 11111 x
al (57) to xal (67) are values between 0 and 1 obtained by dividing the number of passengers per 5 minutes obtained from past traffic statistics by the maximum possible value NNmax (for example, 100 people). Normalize to Similarly, the number of passengers xal(68) to xal(
78), Number of people getting off in 5 minutes in the upward direction from 1st floor to 11th floor x
al (79) to xal (89), and the number of people getting off in the down direction from the 12th floor to the 2nd floor for 5 minutes xal (90) to xa
l (100) is also normalized by dividing by the maximum value NNmax.

Note that the method for normalizing input data is not limited to the above method, and there are many ways to normalize input data! and the direction of travel can also be expressed separately. For example, when the car position floor is f, the input value xal(1) of the first node representing the first floor of the car position is set as input value xa
l(2), "+IJ" in the up direction, "-1" in the down direction
, no direction may be represented as "0".

In this way, when the input data for the input layer (10D^1) is set in step 51, the following step 52
56, network calculations are performed to predict the arrival time when a new hall call C is provisionally assigned to car No. 1.

First, using the input data xaHi), the output value yal(i) of the input layer (10D^1) is calculated from equation (2) (step 52).

Next, the output value yal(i) obtained by the formula (■) is multiplied by the weighting coefficient waHi,j), and summed for i=1 to N1, and from the formula (■), the intermediate layer (10D^2 ) input value x
a2(j) is calculated (step 53).

Next, using the input value xa2(i) obtained from formula ■,
From formula 0, the output value ya2(j) of the middle layer (10D^2)
is calculated (step 54).

Next, the output value ya2(j) obtained by formula 0 is multiplied by the weighting coefficient wa2(j, k), and summed for j-1 to N2. ) is calculated (step 55).

Then, using the input value xa3(k) obtained from formula ■,
From formula 0, the output value of the output layer (10D^3) y a3(k
) is calculated (step 56).

As described above, when the network calculation of the expected arrival time is completed, the output data conversion subunit (IO
CB), the output value 3' a30) ~ y a3(k)
The final expected arrival time is determined by converting the form (step 57).

At this time, each node of the output layer (10D^3) corresponds to a landing for each direction, and the output value y of the 1st to 11th nodes
a3(1) to ya3(11) are respectively 1.2.・
..., is used to determine the calculated value of the expected arrival time of the upbound landing on the 11th floor, and the output value of the 12th to 22nd nodes y a
3(12) to y a3(22) are each used to determine the calculated value of the expected arrival time of the downbound landing area.

That is, first, the output value ya3(k) of the node is converted into the expected arrival time T(k>) at the landing, and this expected arrival time T(k) is expressed as T(k)=ya3(k)XNTmax−
−−■ It is expressed as follows. However, NTmax is a constant value representing the maximum possible value of the expected arrival time. Here, since the output value ya3(k) of the node is normalized to the range of 0 to 1, the maximum value NTs+
By multiplying by ax, the expected arrival time T (k) is converted so that it can be used for calculating the evaluation value of hall call assignment.

In this way, the arrival time prediction program (steps 34-
In 37), the causal relationship between traffic conditions and expected arrival time was expressed using a network, and the traffic condition data was fed into a neural network to calculate the expected arrival time, so it was possible to achieve accuracy that could not be achieved with conventional methods. , it is possible to obtain an expected arrival time that is close to the actual arrival time. Furthermore, since the car assigned to the hall call is selected based on this expected arrival time, the waiting time for the hall call can be shortened.

However, this network is a neural network (IO
Since it changes depending on the weighting coefficient −aHi, j) and wa2(j, k) connecting each node in D^), the weighting coefficient w
By appropriately changing and modifying aHi,j) and wa2(j,k) through learning, a more appropriate expected arrival time can be determined.

Next, while referring to FIG. 6 and FIG. 7, the learning data creation program (step 40) and the modification program (step 41) were executed by the learning data creation means (IOF) and the modification means (IOG). Another embodiment of this invention will be described below.

Note that learning (network modification) in this case is efficiently performed using the pack propagation method.

The pack propagation method is a method that uses the error between the output data of the network and the desired output data (teacher data) created from actual measurement data, control target values, etc. to modify the weighting coefficients that connect the network. .

In FIG. 6, which shows the learning data creation program (step 40) in detail, first, permission to create new learning data has been generated (set), and a new hall call C has been assigned. It is determined whether it is immediately after that (step 61).

If the permission to create learning data is set and the hall call C is assigned, the traffic state data xal(1) to xal(N 1
) and output data y a3 (1) to y a3 (N 3) corresponding to the expected arrival time of each landing at this time are stored as part of the m-th learning data (teacher data) (step 62 ).

Subsequently, permission to create new learning data is reset, and an actual measurement command for the actual arrival time is set to start counting the actual arrival time (step 63).

As a result, in step 61 of the next calculation cycle,
Since it is determined that permission to create new learning data is not set, the process advances to step 64.

Also, in step 64, it is determined whether or not the 111 arrival time actual measurement command is set, but in step 63
Since the actual measurement command is set in step 6,
5, it is determined whether the assigned car has responded to hall call C or not.

If the car has not stopped at the hall of hall call C, the process proceeds to step 66, where it is determined whether the car position f of the assigned car has changed.

When a change in the car position Wf is detected in a calculation cycle several times later, the process proceeds from step 66 to step 67, and the actual arrival time at this time is stored as part of the m-th learning data. This is the original training data and is expressed as the actual arrival time TA(f) of the hall call C at the hall.

Further, in step 65 of a calculation cycle several times later, when a decision to stop at the hall for hall call C is detected, step 6
8, the actual arrival time at this time is stored as part of the m-th learning data (actual arrival time T A (C)).

Then, the actual measurement command for the actual arrival time is reset to finish counting the actual arrival time, and the learning data number m
is incremented, and permission to create new learning data is set again (step 69).

In this way, according to the time when the landing calls were allocated,
The input data and output data regarding the assigned car, as well as the actual arrival times for each landing on the intermediate floor that the assigned car stopped at or passed through until the assigned car responded to the landing call C, are used as learning data. It is created repeatedly and memorized.

Next, the modification means (1, 0G> uses the learning data in the modification program (step 41) in FIG. 4 to modify the network of the neural network (10D^).

This correction operation will be explained in more detail below with reference to FIG.

First, it is determined whether or not it is time to modify the network (step 71). If it is time to modify the network, the following steps 72 to 78 are executed.

Here, the number m of currently stored learning data sets is
When the number becomes 8 (for example, 500) or more, it is determined that it is time to modify the network. Note that the number S of criteria for learning data can be arbitrarily set depending on the scale of the network, such as the number of installed elevators, the number of floors FL of a building, and the number of hall calls.

If it is determined in step 71 that the number m of learning data sets is 8 or more, the learning data counter number n
After initially setting TA(k) to "1" (step 72), extract the actual arrival time TA(k) from the n-th learning data, and calculate the values of the nodes corresponding to these landings, that is, the teacher data d a( k) (k・1,2.-N3), d a (k)-T A (k)/N Tmax
- Determine from ■ (step 73).

Next, the output value y a3 (1) - y a3 (N
3) and the teacher data d a(1) to d a(N3), square the difference between the two and sum k=1 to N3, Ea=ΣNda(k) ya3(k) )'
]/2, −■ (k=1 to N3). Then, using the error Ea obtained from Equation 0, the weighting coefficient w a 2 (j, k) (j・1, 2 . −・・
, N2, k=1.2. -, N3) as follows (step 74).

First, by differentiating the error Ea of equation 0 with wa2(j,k) and rearranging it using the above-mentioned equations ① to ②, we get the weighting coefficient wa2(
The amount of change Δwa2(j, k) in j, k) is Δwa2(j
, k)−−a (θEa/θwa2(j, k))=−α
・da2(k)・y a2(j) ... Represented by ■. However, α is a parameter representing the learning speed,
Any value within the range of 0 to 1 may be selected. Also, in equation 0, da2(k) = fya3(k)-da(k)lya
3(k) fl-ya3(k)1. In this way, when the amount of change Δwa2(j, k) of the weighting coefficient wa2(j, k) is calculated, the weighting coefficient wa2(j, k) is calculated using the following [phase] formula.
j, k) are modified.

wa2 (j, k) ← wa2 (j, k) + Δwa2 (j,
k) −'HE Also, similarly, the weighting coefficient wal(i j
) (i4.2.-, Nl, jl, 2.-N2) according to the following formulas 0 and 0 (step 75),
.

First, the change IΔwal(i,
j), △wal(i, j)-1α・δal(j)・yal(i
＞ ...Determine from ■. However, in equation 0, δa
l(j) is the following summation formula with k=1 to N3, δal(j)=Σ(da2(k)・wa2(j,k)
・ya2(j)X [1-y a2(j)]fl Represented. The amount of change Δwal(i, j
), the weighting coefficient wal(i,
j) is corrected.

wal(i,j)←wal(i,j)+ΔwaHi,j
)...@In addition, in steps 74 and 75 above,
Only the weighting coefficients related to the landings where training data exists are modified. In other words, as explained in the learning data creation program (Fig. 6), the actual arrival time is used as training data only for landings on floors halfway between the car position at the time of allocation and the landing of landing call C. Therefore, the weighting coefficients for other landings are not modified.

In this way, correction step 7 using the nth learning data
When steps 3 to 75 are performed, the learning data number n is incremented (step 76), and in step 77 it is determined that all the learning data have been modified (02).
The processing of steps 73 to 76 is repeated until the number of steps 73 to 76 is reached.

Then, when all the learning data are corrected, the weight coefficients wal (i, j) and wa2 that have been corrected are
(j, k) is registered in the expected arrival time calculation means (IOD>) (step 78).

At this time, all the learning data used for correction are cleared and the learning data number m is initialized to "1" so that the latest learning data can be stored again. In this way, network modification (learning) of the neural network (10D^) is completed.

In this way, we create learning data based on actual measured values,
Using these learning data, the expected arrival time calculation means (IO
Since the weighting coefficients waHi,j) and wa2(j,k) of D> are modified, it is possible to automatically respond to changes in the traffic flow within the building.

In addition, as the input data representing the characteristics of traffic flow, we used the number of people boarding and disembarking in 5 minutes on the landing side, which was calculated in the past, so that the car position, direction of operation, Furthermore, compared to the case where only the call to be answered is used as input data, a more flexible and accurate prediction calculation can be realized.

In the above embodiment, the input data conversion means converts the car position, the direction of travel, and the call to be answered as input data, but the traffic condition data used as input data is limited to these. For example, car status (decelerating, door opening, door opening, door closing, waiting to close door, running, etc.), duration of hall call, duration of car call , car load, number of cars under group management, etc. can be used as input data. In addition, by using not only current traffic condition data but also traffic condition data in the near future (history of car movements, history of call response conditions, etc.) as input data, -
It becomes possible to calculate the expected arrival time accurately.

In addition, when a landing call is assigned, the learning data creation means (IOF) records the expected arrival time of the assigned car at each landing and the input data at that time, as well as the information that the assigned car receives at the landing call after that. Although the actual arrival times at the landings stopped or passed before responding are stored as a set of learning data, the timing at which learning data is created is not limited to this. For example, the previous input data The learning data creation time may be set to the time when the time elapsed since the time of memorization exceeds a predetermined time (for example, 1 minute).
For example, the learning data may be created every minute (for example, every minute). Also, the more learning data under various conditions is collected, the better the learning conditions will be. Even if you decide in advance a typical possible state, such as when the car is in a predetermined state (decelerating, stopping, etc.), and create learning data when that state is detected. good.

In addition, the learning data creation means (IOF) stores only the actual arrival time at the landing where the assigned car has stopped or passed at the stop where the assigned car responds to the assigned landing call as training data, and the correction means (IOF) When modifying the weighting coefficients using IOG), only the weighting coefficients related to the stored training data are modified, but the method of extracting the training data is as follows.
For example, the expected arrival times for all landings and the actual arrival times that could be measured during the operation of the car are stored, and weights related to the landings for which training data exists are stored. Only the coefficient may be corrected.Here, a landing where the actual arrival time could not be measured is, for example, when the direction of the car reverses on the floor midway,
If the car becomes an empty car or a car that does not have an assigned call on an intermediate floor (corresponding to a landing far away from the reversing floor), the input data and the landing far from the floor where the car became empty are stored. This corresponds to the landing behind the floor where the car is currently located (for example, during upward movement, the landing is below the current position).

Further, although the expected arrival time calculating means (IOD) corrects the weighting coefficient every time the number of stored learning data reaches a predetermined number, the timing of correcting the weighting coefficient is not limited to this. . For example, at a predetermined time (e.g.
The weighting coefficients may be corrected using the learning data stored up to that time (every hour), and when traffic becomes quiet, the expected arrival time calculation means (IOC) may calculate the expected arrival time more frequently. It is also possible to modify the weighting coefficient when the number decreases.

Furthermore, the weighting coefficient modification step is performed multiple times (e.g., 50
500 times for 0 data) #i may be repeated to converge the weighting coefficients so as to obtain a desired approximate output.

[Effects of the Invention] As described above, according to the present invention, there is provided an input data conversion means for converting traffic condition data including car positions, running directions, and calls to be answered into a form that can be used as input data for a neural network. , an input layer that takes in input data, an output layer that outputs data corresponding to the expected arrival time, and an intermediate layer that is located between the input layer and the output layer and has weighting coefficients set, and constitutes a neural network. Equipped with an expected arrival time calculation means and an output data conversion means for converting the output data into a form that can be used for a predetermined control purpose, the traffic condition data is input into the neural network and the time required for the car to arrive at the landing is calculated. Since it is calculated as the expected arrival time, it is possible to calculate the expected arrival time by a calculation that is close to the actual arrival time, and it is also possible to improve the performance of group management based on this accurate expected arrival time. This has the effect of providing a highly efficient elevator control device.

According to another invention of the present invention, at a predetermined time while the elevator is in operation, the expected arrival time of a predetermined car and the input data at that time, as well as the actual arrival time of the predetermined car are stored. and further comprising a learning data creation means for outputting these as learning data of a − set, and a modification means for modifying the weighting coefficient of the predicted arrival time calculation means using the learning data, and the calculated prediction result and Since the weighting coefficients in the neural network are automatically corrected based on the traffic condition data and actual measurement data at that time, it is possible to automatically respond to changes in the actual traffic flow within the building. This has the effect of providing an elevator control device with high time prediction accuracy.

[Brief explanation of the drawing]

FIG. 1 is a functional block diagram showing the overall configuration of an embodiment of this invention and another invention of the present invention, FIG. 2 is a block diagram showing a schematic configuration of the group management device in FIG. 1, and FIG. FIG. 1 is a block diagram specifically showing the data conversion means and expected arrival time calculation means; FIG. 4 is a flowchart schematically showing the group management program stored in the ROM in FIG. 2; and FIG. is a flowchart diagram specifically showing the arrival time prediction calculation program at the time of provisional allocation for the No. 1 aircraft in FIG. 4,
FIG. 6 is a flowchart specifically showing the learning data creation program in FIG. 4, and FIG. 7 is a flowchart specifically showing the modification program in FIG. 4. (10C...Data conversion means (10C^)...Input data conversion subunit (IOC)
B...Output data conversion subunit (10D^)...
・Neural net (10D^)...Input layer (10D^2)...
・Middle layer (10D^3)...Output layer (IOD)...Estimated arrival time calculation means (IOF>・
. . . Learning data creation means (IOC) . . . Modification means wal (i, j), wa 2 (j, k) - weighting coefficient section, in the drawings, the same reference numerals indicate the same or equivalent parts. Figure 2

Claims (2)

[Claims] (1) In an elevator control device that predicts and calculates the time required for an elevator car to arrive at a landing as an expected arrival time, and controls the operation of the car using the expected arrival time, the position and operating direction of the car and input data conversion means for converting traffic condition data including calls to be answered into a form that can be used as input data of a neural network; an input layer that takes in the input data; and an intermediate layer between the input layer and the output layer in which a weighting coefficient is set, and an expected arrival time calculation means constituting the neural network; An elevator control device comprising: output data conversion means for converting into a usable form. (2) At a predetermined time while the elevator is in operation, the expected arrival time at a predetermined landing and the input data at that time are stored, and the time elapsed until the car stops at or passes the predetermined landing. learning data creation means for counting and storing as an actual arrival time, and outputting the stored input data, the expected arrival time, and the actual arrival time as a set of learning data; 2. The elevator control device according to claim 1, further comprising: modifying means for modifying the weighting coefficient of the expected arrival time calculating means.