JP2006268221A - Predicted-data generation method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To generate predicted data by predicting data in a predicting period from data immediately prior to the prediction, based on the result of learning of a neural network using learning data which are selected according to different kinds of conditions, by means of a predicted data generation system with the neural network which constitutes a monitoring device. <P>SOLUTION: Predicted data are generated by predicting data in a predicting period from data immediately prior to the prediction, based on the result of learning of a neural network using collected condition data, as learning data, similar to the predicting period, by means of a predicted data generation system with the neural network which constitutes a monitoring device. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a data generation method for predicting and generating near-future data that continues to current data in monitoring data collection for water supply and sewerage facilities and the like.

As water treatment plants and sewerage facilities in water and sewage facilities become more sophisticated and diversified, and the automation to reduce the burden on operators has progressed, the monitoring equipment for water treatment plants and sewage facilities simply monitors the state of the equipment. Trend charts and databases that provide information on graphs of monitoring data for sewage facilities, water purification plants that have evolved from the functions that occupy, have become a major position.

The monitoring device stores the collected monitoring data in a database and simultaneously displays a trend graph of the specified data, and visually monitors based on the years of experience and know-how of skilled operators. Or, if the condition becomes true due to judgment conditions such as the data stored sequentially in the database exceeding or falling below a preset value, an alarm, phone call, etc. is automatically issued to the supervisor and appropriate measures are taken This is done. (For example, see Patent Document 1)
JP 2004-187260 A

The conventional techniques described above have the following problems.
(1) The trend graph displayed by the monitoring device displays collected data and visually monitors it, but uses similar data under similar conditions in the past and is not useful for monitoring.
(2) There is also a method of displaying past trend graphs by simply specifying the date and time, but this method predicts near-future data by comparing it with data currently in progress, and issues warnings and various operation controls. It is not possible to take appropriate actions quickly.

  The present invention has been made in consideration of the above-described problems of the prior art, and the object of the present invention is to select learning data from collected performance data according to various conditions, learn a coupling coefficient of a neural network, It is an object of the present invention to create a neural network with excellent generalization ability and to provide a prediction data generation method based on this neural network.

In other words, the means to achieve that purpose is
(1) In claim 1,
In monitoring data collection, when predicting and generating near-future data that continues with current data, the data during the prediction period is used as the learning data of the neural network using the existing data, and the trained neural network This is a prediction data generation method in which data during a prediction period is estimated and created by a network.

(2) In claim 2,
Specify and select conditions such as period, time, day of the week, weather, temperature, etc. from the monitoring data already acquired for the neural network learning data, and set some learning data according to the conditions, the output unit of the neural network is The monitoring data value (partition point) corresponding to the time when the prediction period is divided at regular intervals, and the input unit of the neural network corresponds to the time when the data corresponding to the monitoring data immediately before the prediction period is divided at regular intervals This is a method of generating prediction data using a neural network that learns and guesses as a value.

(3) In claim 3,
Giving the data immediately before the prediction period to the learning result of the neural network, obtaining the guess data value during the forecast period, converting each guess data value into partition points, linearly interpolating the partition points, and further by random numbers, This is a prediction data generation method using a neural network approximated to the waveform of the original data.

Next, the operation will be described.
In the first to third aspects of the present invention, as in the above (1), (2), and (3), the learning data selected according to various conditions is used by the prediction data generation method by the neural network constituting the monitoring device. Since the data during the prediction period is estimated and generated from the data immediately before the prediction from the learning result of the neural network, a highly accurate prediction data generation method can be obtained.

As described above, the prediction data generation method by the neural network constituting the monitoring apparatus of the present invention, the data during the prediction period is estimated from the data immediately before the prediction from the learning result of the neural network using the learning data selected according to various conditions. Therefore, a highly accurate prediction data generation method is realized and is extremely effective in practical use.

  FIG. 1 is an overall configuration diagram of an embodiment of a monitoring apparatus including a prediction data generation method according to an embodiment of the present invention, where 12 is a monitoring apparatus, 1 is a learning data control unit, 2 is a neural network learning control unit, 3 Is a neural network (coupling coefficient), 4 is a neural network estimation control unit, 5 is a working memory (learning data control unit 1, neural network learning control unit 2, neural network estimation control unit 4, input control unit 6, output control unit 7 (Including a place to store temporarily input or calculated data used in the above and a learning data selection condition for setting various conditions), 6 is an input control unit, 7 is an output control unit, and 8 is actual data , 9 is learning data, 10 is an input device, 11 is an output / display device (or output to another device), and 13 is a function processing unit of each monitoring device.

When the collected performance data and various set values are input to the input device 10 in the monitoring device 12, each performance data is stored in the working memory 5 via the input control unit 6.
Each input value is input to the learning data control unit 1, where learning is performed according to the actual data on the working memory 5, various setting values in the learning data selection conditions, that is, the prediction period, time, day of the week, weather, temperature, and the like. The range of data is selected, and the corresponding performance data is output to the learning data 9.

  The learning data creation result can be output / displayed on the output / display device 11 via the output control unit 7.

When various setting values are input to the input device 10 in the monitoring device 12, the input values are stored in the working memory 5 via the input control unit 6. The learning data 9 is stored in the working memory 5 by the neural network learning control unit 2.
Each input value is input to the neural network learning control unit 2, where the neural network learning control unit 2 performs learning based on the learning data on the working memory 5 and various setting values in the learning data selection conditions. The coupling coefficient between the input layer and the intermediate layer and the coupling coefficient between the intermediate layer and the output layer are updated.
The coupling coefficient is stored inside the neural network (coupling coefficient) 3. The learning result can be output / displayed on the output / display device 11 via the output control unit 7.

  When the monitoring device 12 inputs monitoring data immediately before the prediction period from the input device 10, the input value is stored in the working memory 5 via the input control unit 6. The neural network estimation control unit 4 calculates an output value based on a normalized value of the coupling coefficient of the neural network (coupling coefficient) 3 and the input value, and further inversely calculates the value, which is used as an estimated value. The result is output / displayed on the output / display device 11 via the output control unit 7.

  FIG. 2 is a diagram illustrating an example of the prediction period. The data collection interval (sampling interval) of the monitoring device is assumed to be t seconds. The prediction period is n minutes. In this example, the interval for dividing the prediction period at a constant interval is 1 minute.

  FIG. 3 is a diagram illustrating an example of learning data selection conditions. In the selection condition and designation example of FIG. 3, the period is a period corresponding to the prediction period, a period corresponding to twice the prediction period immediately before the prediction, and the time is data based on the same time as the start time of the prediction period. The day of the week is not specified. (If specified, the same day of the week, weekdays, holidays, etc. are specified.) The weather is not specified. There are three categories: 1) clear, 2) cloudy, and 3) rain. ) For the air temperature, data on the day when the prediction is generated and the temperature within 10 degrees of the temperature difference are used. The number of learning data is 100 data, and the selection order is selected from data having a new time. In the classification method, data within a period is classified into points at intervals of 1 minute, and the divided points are used for learning.

  FIG. 4 is a diagram illustrating an example of learning data. This is the learning data selected according to the learning data selection condition of FIG. 3, and in this example, the period three times the prediction period is used as the learning data.

FIG. 5 is a conceptual diagram showing an example of a neural network in the monitoring apparatus. As a structure and learning means of the neural network, there is a typical error back propagation method (back propagation method). (For example, “Neural Network Model and Connectionism”, written by Shunichi Amari, published by the University of Tokyo Press.)
The input / output relationship of the unit i other than the input layer, that is, the intermediate layer and the output layer is expressed by equations (1), (2), and (3). The input to the unit i is represented by Oj (j = 1 to N), the output is represented by Yi, and the coupling coefficient for each Oj is represented by Wij.
Product sum of inputs Xi = Σ WijOj (1)
(1) The expression is applied to the function f (Xi) and converted. In general, a sigmoid function such as the following equation (2) is often used as a function.
f (Xi) = 1 / {1 + exp (−Xi)} (2)
Output Yi = f (Xi) (3)
Here, the value of Yi is a number between 0 and 1.
On the other hand, the input layer unit uses the input value as it is as the output value. The input layer normalizes the division points of the period corresponding to the immediately preceding prediction period in the learning data shown in FIG. 4 to correspond to the input unit. In this example, the number of units in the intermediate layer is half the number of output units. The output layer normalizes the segment points of the period corresponding to the prediction period in the learning data shown in FIG. 4 and corresponds to the output unit.
In addition, the maximum and minimum values are obtained for each learning data so that the input and output numbers are in the range between 0 and 1, and the inverse of the width of the maximum and minimum values is used as the scale, and each value is the minimum value. Normalize by multiplying with a pull scale. When the output numerical value is actually used, it is reversely calculated and used as each dividing point.

  FIG. 6 is a diagram illustrating an example of a random number generation method. After linearly interpolating between the segment points, a random value within a specified range is obtained for each sampling point between the segment points and added to the sampling points. The designated range is the range of the maximum value and the minimum value between the sections of the original collected data of the learning data.

It is a block diagram of the system of the Example of this invention. It is a figure which shows the prediction period which is one Example of this invention. It is a figure which shows the learning data selection conditions which are one Example of this invention. It is a figure which shows the learning data which is one Example of this invention. It is a conceptual diagram which shows the neural network which is one Example of this invention. It is a figure which shows the generation method by the random number which is one Example of this invention.

Explanation of symbols

1 Learning data control unit 2 Neural network learning control unit 3 Neural network (coupling coefficient)
4 Neural network guess control part 5 Working memory (learning data selection condition)
6 Input Control Unit 7 Output Control Unit 8 Actual Data 9 Learning Data 10 Input Device 11 Device that Performs Output / Display 12 Monitoring Device 13 Each Function Processing Unit of Monitoring Device

Claims (3)

In monitoring data collection, when predicting and generating near-future data that continues with current data, the data during the prediction period is used as the learning data of the neural network using the existing data, and the trained neural network Prediction data generation method that estimates and creates data during the prediction period by the network. Specify and select conditions such as period, time, day of the week, weather, temperature, etc. from the monitoring data already acquired for the neural network learning data, and set some learning data according to the conditions, the output unit of the neural network is The monitoring data value (partition point) corresponding to the time obtained by dividing the prediction period at regular intervals, and the input unit of the neural network corresponds to the time obtained by dividing the data corresponding to the monitoring data immediately before the prediction period at regular intervals. A method of generating prediction data using a neural network that learns and guesses as a value. Giving the data immediately before the prediction period to the learning result of the neural network, obtaining the guess data value during the forecast period, converting each guess data value into partition points, linearly interpolating the partition points, and further by random numbers, A prediction data generation method using a neural network approximated to the waveform of the original data.

Images