JPH04220758A - Time sequence data prediction and prediction recognition method - Google Patents

Abstract

PURPOSE:To effectively realize the prediction of time sequence data and the recognition of the prediction result even when an object generating the time sequence data cannot be theoreically analized and cannot be expressed with an expression by the time sequence data. CONSTITUTION:The data in an arbitrary time zone of preliminarily prepared time sequence data for learning are inputted in a neural network 1 consisting of a plurality of neurons and the network is made to learn so that the time sequence data of a prescribed time after can be outputted. data for each time unit by means of the neural network 1, a prediction value for a prescribed time unit after can be obtained. By separately providing another neural network 4 for recognition and inputting the prediction value from the neural network 1 and the time sequence data, the prediction of the time sequence data can be recognized.

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method of predicting future values of time series data or pattern data using a neural network, and a method of pattern recognition from the results of the prediction.

[0002] Time-series data is widely used on a daily basis, such as financial data such as stock prices, exchange rates, and futures transactions, sensor signals in factories and plants, power consumption, and road traffic. And from those time series data (close)
Techniques for predicting future values (states) are needed. For example, when controlling a plant using signals from sensors installed in plant equipment, if the (near) future state of the plant can be predicted, control can be performed in advance, and if there is an abnormality in the equipment, control can be performed in advance. However, early detection is possible.

0003

[Background Art] Conventionally, when trying to predict a state that will occur in the near future using time series data obtained from a target object (for example, a "control target"), a person who is familiar with the handling of the target object (for example, a "control target") Forecasting can be done by relying on the experience of the individual, or by analyzing and predicting using statistical or probabilistic methods if possible, or by linearly approximating the obtained time series data using an appropriate approximation function. method was used.

0004

[Problems to be Solved by the Invention] In general, prediction of time series data often cannot be described using theoretical formulas. for example,
In the case of plant control, it is often difficult to measure the parameters and time constants of plant equipment, and the disturbances are large, so it is often necessary to rely on control by experienced personnel, which leads to problems in terms of reliability. There were many problems.

The present invention has been made in view of the above-mentioned problems, and even when it is not possible to theoretically analyze or mathematically describe an object that generates time-series data, it is possible to effectively calculate the future value of the time-series data. The purpose is to provide a method for predicting and recognizing information.

[0006]

According to the invention, the above-mentioned objects are achieved by the means set out in the appended claims.

That is, in the invention as claimed in claim 1, a neural network is given learning time series data and output data for the time series data to perform learning in advance, and the neural network is used to perform learning. , is a time series data prediction method that predicts data at an arbitrary time point ahead of input time series data.

[0008] Furthermore, regarding the invention as claimed in claim 2, using a plurality of neural networks as claimed in claim 1,
This is a time-series data prediction method that outputs data for a plurality of different unit times ahead of input time-series data to predict data for any future time period.

Furthermore, regarding the invention as claimed in claim 3,
The neural network according to claim 1 is a neural network that predicts data one unit time ahead, and a plurality of neural networks are used, and each neural network predicts the predicted value of the neural network that predicts data one unit time ahead. It is a time series data prediction method that uses time series data to predict the next time's data.

[0010] Regarding the invention described in claim 4, claim 1
This is a time-series data learning method that increases the content of learning pattern data as the amount of time-series data increases when the described neural network performs learning.

[0011] Regarding the invention described in claim 5, claim 1
Using the method described in , 2 or 3, a recognition network is used that predicts data at an arbitrary time point in the time series data, and also causes a neural network to perform pattern recognition prediction of the time series data using the predicted value. This is a predictive recognition method for time series data.

[0012] Regarding the invention described in claim 6, claim 5
This is a time-series data predictive recognition method that uses a plurality of the described recognition networks to perform pattern recognition and prediction of an arbitrary time period in the future for inputted time-series data.

[0013] Regarding the invention described in claim 7, claim 6
This is a predictive recognition method for time-series data in which the recognition prediction results obtained by the described method are further subjected to pattern recognition using a neural network for comprehensive determination to obtain a final determination result.

The means used in the present invention will be explained below with reference to the drawings.

That is, FIG. 1 shows a first principle configuration diagram of the present invention, and corresponds to the invention recited in claim 1, in which time-series data is input to a neural network 1 that has completed learning in advance. It inputs data and outputs a predicted value at an arbitrary time point.

[0016] In this case, if there is other valid data that influences the predicted value, it can be provided.

Further, FIG. 2 shows a second principle configuration diagram of the present invention, which corresponds to the invention recited in claim 2 of the claims, and which corresponds to the second principle of the present invention. Multiple neural networks (11 to 1n
) and make predictions for any time period.

FIG. 3 is a diagram showing the third principle of the present invention.
Corresponding to the invention recited in claim 3 of the claims, the neural networks 21 to 2n predict data one unit time ahead of the input time series data, and each neural network predicts the data of the input time series data one unit time ahead. Use predictions to make data predictions.

FIG. 4 shows a fourth principle configuration diagram of the present invention, which corresponds to the fourth principle of the invention,
With the increase in time series data, learning pattern data 31 ~
3m (pattern set ■ to m) is increased and the neural network is made to perform learning.

FIG. 5 shows a fifth principle configuration diagram of the present invention, which corresponds to the fifth aspect of the invention,
The predicted values and time series data from the neural network 1 for prediction are input to the neural network 4 for recognition to perform predictive recognition of the time series data.

FIG. 6 shows a sixth principle configuration diagram of the present invention, and corresponds to the sixth aspect of the invention,
Using predicted values of time series data obtained by a plurality of neural networks 11 to 1n that perform predictions i (i = 1 to n) time ahead, a plurality of neural networks for recognition 50 to 5m are used to predict an arbitrary time period. Obtain recognition results about.

FIG. 7 shows a configuration diagram of the seventh principle of the present invention, which corresponds to the invention set forth in claim 7 of the claims, and the individual recognition prediction results obtained in FIG. Further, the neural network 6 is configured to perform recognition and obtain a comprehensive judgment result.

[0023]

[Operation] In the present invention, prediction of time series data is performed using a prediction neural network. Furthermore, by causing the neural network to perform pattern recognition using the prediction results, it is possible to perform early abnormality detection and failure prediction in, for example, plant control.

[0024] The operation of each invention described in the claims will be explained below.

FIG. 8 shows a first operational principle explanatory diagram of the present invention, which corresponds to the invention recited in claim 1, and is a neural network composed of a plurality of neurons 10. (In this example, the 3-layer structure) is input with data from any time period of time series data prepared in advance for learning, and learning is performed so that time series data from any time ahead is output. .

[0026] Then, by actually inputting the sampling values of the time series data for each unit time to the neural network, a predicted value for a predetermined unit time ahead is obtained. In addition,
As mentioned above, in addition to time series data, the input of the neural network is based on the accuracy of predicted values.

FIG. 9 shows a second operational principle explanatory diagram of the present invention, which corresponds to the second aspect of the invention and is a neural network for predicting time-series data at an arbitrary time point. It is possible to prepare a plurality of neural networks (in the example shown in the figure, six neural networks 11 to 16) and specify an arbitrary time period to predict data.

[0028] In this case, the number of neural networks is equal to the number of units of time constituting the required time period, and each neural network is assigned one unit of time ahead, two units of time ahead, and n units of time ahead. Predict future data. Therefore, by preparing n neural networks, it is possible to predict data in a time period equal to "n x unit time".

FIG. 10 is a diagram illustrating the third operating principle of the present invention.
FIG. 11 shows a fourth diagram illustrating the operating principle of the present invention. Each of these corresponds to the invention recited in claim 3 of the claims.

In the example shown in FIG. 9, a plurality of neural networks for predicting 1 to n units of time ahead are prepared, but in the example shown in FIG.
Now, n neural networks for predicting one unit time ahead are prepared (in the example of this figure, six neural networks 21 to 26).

[0031] First, time-series data is input to the first neural network 21, and a predicted value for one unit time ahead is obtained. Next, the time-series data and the predicted value of the first neural network are input to the next neural network 22 with a lag of one unit time to obtain the predicted value of the next time.

Thereafter, similar processing is performed to predict data for n time periods. Note that in this case, only one type of neural network is required for learning.

On the other hand, in the case of FIG. 11, n neural networks are prepared that have different numbers of input units (input time periods) and predict one unit time ahead (in the example of this figure, neural networks 71 to 71) are prepared. 76 pieces).

First, the first neural network 71
It is the same as the case of FIG. 10 that time-series data is inputted into , and prediction is made for one unit time ahead. However, in the next neural network 72 , the data one unit time ahead of the first neural network 71 is used as input data. The predicted value of is also added to the input data.

[0035] The same process is repeated thereafter. In this case, it is necessary to prepare n neural networks with different numbers of input units.

[0036] In addition, regarding the invention recited in claim 4, in normal neural network learning (back propagation method), there are two methods: a method of learning one pattern at a time, and a method of learning all patterns at once. However, the present invention uses an additive pattern learning method that gradually increases the number of patterns to be learned. In this learning method, learning is started with one pattern or a small number of patterns, and then, as learning progresses, learning patterns are gradually increased (added).

Generally, similar patterns often appear in time-series data, and according to the additional pattern learning method of the present invention, if the added learning pattern data is similar to a pattern learned in the past, learning is not performed. It can be omitted, and patterns that are close to patterns learned in the past can be learned without requiring much learning time. Furthermore, even when learning data for a neural network is added, learning can be easily performed using this additional pattern learning method.

As explained above, the major feature of the time series data prediction method according to the present invention is that the time series data prediction is performed using a neural network.
There is no need to describe the problem using mathematical formulas, and even for problems that cannot or are difficult to express using mathematical formulas, it is possible to predict time-series data through neural network learning. Furthermore, the additional pattern learning method facilitates learning of time series data.

Next, the operation of the time-series data predictive recognition method according to the present invention will be explained.

FIG. 12 is a fifth diagram illustrating the operating principle of the present invention, and corresponds to claims 5 and 6 of the claims.

That is, in this example, the above-mentioned claims 1 and 2
Alternatively, according to the invention described in 3, after predicting time series data, the time series data including the predicted value is inputted to the recognition prediction neural network 50 to 5n that sequentially performs recognition, thereby making recognition prediction for any time period. is being carried out.

FIG. 13 is a sixth explanatory diagram of the operating principle of the present invention, which corresponds to the seventh aspect of the invention.

That is, in this example, the recognition result in FIG. 12 is further processed by the comprehensive judgment neural network 6, thereby making it possible to improve the recognition rate.

As described above, in the time-series data predictive recognition method of the present invention, the time-series data up to now and the predicted values by the multi-neural network are input to the multi-neural network that performs recognition prediction to obtain the recognition prediction result. At the same time, it is also possible to further input each recognition prediction result to a comprehensive judgment neural network to obtain a final judgment result.

[0045] Therefore, by performing detailed division of labor in the neural network, each neural network can be made relatively small-scale, and a system configuration that is easy to understand for a designer can be achieved.

[0046]

[Embodiment] As an embodiment of the present invention, a stock price prediction system will be explained.

FIG. 14 is a diagram showing the hardware configuration of the stock price prediction system, which includes a personal computer body 11, a monitor 12, a keyboard 13, a mouse 14,
and a telephone line and its modem 15 for collecting stock price data.
It consists of

FIG. 15 is a diagram showing the menu structure that can be selected in this stock price prediction system. The configuration of the menu will be explained below.

- Stock Selection: Select the stock (company) of the stock price data collected from the telephone line.・Initialization: Set the predicted time.・Update: Start learning after constructing a neural network and specifying the learning range.・Result display: Displays the value predicted by the neural network during learning.・Recognition: Specify a prediction range (period) and make predictions within that period.・Buy and sell: Perform a stock buy and sell simulation based on the prediction results.・Evaluation: Statistically analyze the difference between the predicted value of the neural network and the actual result within the prediction range.・Exit: End the stock price prediction system. The present invention is applied to the "buying and selling" menu.

[0050] As an example, a recognition prediction network is used to learn to "buy" a stock when it increases at a certain rate, and "sell" when it declines at a certain rate over a certain period of time, and then a comprehensive judgment neural The network makes current buy/sell decisions based on the buy/sell prediction results.

FIG. 16 is a diagram showing an example of the recognition results of this embodiment. The broken line portion indicated by number 16 in the figure indicates the predicted data, and it is shown that the predicted data matches well with the actual data.

[0052]

[Effects of the Invention] According to the time-series data prediction method of the present invention, time-series data can be predicted using a neural network, and it is applicable to problems (objects) that are difficult to express mathematically. Predictions can be made for any time period using multi-network prediction methods. Further, by using the additional pattern learning method, learning of time series data can be easily performed.

Furthermore, according to the method for predictive recognition of time series data of the present invention, recognition can be performed based on the results of prediction of time series data using a neural network, so early recognition or detection is possible. moreover,
By dividing the labor of neural networks, it is possible to downsize the system, which is expected to shorten learning time.

[Brief explanation of the drawing]

FIG. 1 is a first principle configuration diagram of the present invention.

FIG. 2 is a second principle configuration diagram of the present invention.

FIG. 3 is a configuration diagram of the third principle of the present invention.

FIG. 4 is a fourth principle configuration diagram of the present invention.

FIG. 5 is a fifth principle configuration diagram of the present invention.

FIG. 6 is a configuration diagram of the sixth principle of the present invention.

FIG. 7 is a seventh principle configuration diagram of the present invention.

FIG. 8 is a diagram illustrating the first operating principle of the present invention.

FIG. 9 is a second diagram illustrating the principle of operation of the present invention.

FIG. 10 is a diagram illustrating the third operating principle of the present invention.

FIG. 11 is a diagram illustrating the fourth principle of operation of the present invention.

FIG. 12 is a diagram illustrating the fifth operating principle of the present invention.

FIG. 13 is a diagram illustrating the sixth principle of operation of the present invention.

FIG. 14 is a diagram showing the hardware configuration of a stock price prediction system that is an embodiment of the present invention.

FIG. 15 is a diagram showing a menu structure that can be selected in the stock price prediction system.

FIG. 16 is a diagram showing an example of recognition prediction results of the stock price prediction system.

[Explanation of symbols]

1 Neural network for prediction 11 ~1n Neural network 21 that predicts 1 to n hours ahead ~2n Neural network 31 that predicts 1 unit time ahead ~3m Learning pattern data 4 Neural network for recognition 51 ~5m 1~m time ahead Neural network for recognition 6 Neural network for comprehensive judgment 71 to 7
Neural network 10 that predicts one unit time ahead with a different number of input units Neuron 11 Personal computer 12 Monitor 13 Keyboard 14 Mouse 15 Modem 16 Prediction data

Claims (7)

[Claims] Claim 1: A neural network is trained in advance by giving time-series data for learning and output data for the time-series data, and the neural network is used to perform learning on arbitrary input time-series data. A time series data prediction method characterized by predicting data for a time ahead. 2. The present invention is characterized in that a plurality of neural networks according to claim 1 are used to output data for a plurality of different unit times ahead of input time series data, thereby predicting data for an arbitrary time period in the future. A method for predicting time series data. 3. The neural network according to claim 1 is a neural network that predicts data one unit time ahead, and a plurality of neural networks are used, and each neural network is a neural network that predicts data one unit time ahead. A method for predicting time series data characterized by using predicted values of a network to predict data for the next time. 4. A method for learning time-series data, which comprises increasing the content of learning pattern data as the amount of time-series data increases when the neural network according to claim 1 performs learning. 5. The method according to claim 1, 2, or 3 is used to predict data at an arbitrary time point in the time series data, and the predicted value is used to send the time series data to a neural network for recognition. A predictive recognition method for time series data, characterized by performing pattern recognition prediction. 6. Prediction of time series data, characterized in that a plurality of neural networks for recognition according to claim 5 are used to perform pattern recognition prediction of an arbitrary time period in the future for input time series data. Recognition method. 7. A time series characterized in that the recognition prediction result obtained by the method according to claim 6 is further subjected to pattern recognition using a neural network for comprehensive judgment to obtain a final judgment result. Predictive recognition methods for data.