JPH06193400A - Air pollution-degree predicting device - Google Patents

Abstract

PURPOSE:To predict the degree of air pollution for an arbitrary time by learning the relationship with time of the degree of air pollution and a status value having an effect on the degree of air pollution by using a recurrent type network. CONSTITUTION:The degree of air pollution in a tunnel and a status value having an effect on the degree of air pollution are measured in a value with time by using a VI-value measuring means 1, a passage-number measuring means 2 and an air-velocity measuring means 3, and stored previously in an experience storage means 4. A recurrent type network is used as a learning means 5, the value with time is learned to the learning network 6A until the error of a VI value is reduced sufficiently by employing the experience value, and the weight of the network is corrected. A predicting means 7 predicts the next VI value by using a predicting network 6B, to which learning is completed previously, and a VI predicting value 8 displaying the degree of air pollution is obtained.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an air pollution degree predicting apparatus for predicting the degree of air pollution by utilizing a recurrent network in order to ventilate the inside of a tunnel or the like at an appropriate timing.

[0002]

2. Description of the Related Art Ventilation control in a tunnel has become an extremely important problem from the viewpoint of improving safety by securing a field of view and improving comfort for a driver. As this ventilation control system, various systems have been provided so far, and feedback control, regulator control introducing an evaluation function, predictive control, fuzzy theory applied control, etc. have been tried.

[0003]

Here, the road tunnel ventilation system is generally a distributed constant system having a strong non-linearity and a large delay element. For this reason, the situation of a long time in the past has affected the haze transmittance (hereinafter referred to as VI value), which is a typical index indicating the degree of pollution of air, and this cannot be expressed by a mathematical model. There is a problem that it is difficult to predict the VI value as a result of control and confirm the validity of the range.

The present invention has been made to solve the above-mentioned problems, and its purpose is to accurately predict the degree of air pollution including the VI value and to perform appropriate ventilation according to the situation. An object of the present invention is to provide a possible air pollution degree prediction device.

[0005]

In order to achieve the above object, the present invention has been made by paying attention to the property that a recurrent network can learn a temporal order relation.
That is, the first aspect of the invention is to measure an air pollution degree, a means to measure a situation value affecting the air pollution degree, a means to store the pollution degree and the situation value as an actual value, an input layer, an intermediate layer. Corrects network weights over time using a recurrent network so that the error between a single recurrent network consisting of layers, context layers, and output layers and the actual values stored over time is reduced. Means for learning the relationship between the situation value and the pollution degree, and means for predicting the pollution degree at an arbitrary time using a recurrent network based on the already learned relationship between the situation value and the pollution degree It is equipped with and.

A second aspect of the present invention comprises means for measuring the degree of air pollution, means for measuring a situation value affecting the degree of air pollution, means for storing the degree of pollution and the situation value as actual values, an input layer, The error is reduced between a plurality of recurrent networks each of which includes an intermediate layer, a context layer, and an output layer and which are provided for each condition that affects the degree of air pollution, and the actual value stored over time. Means for learning the relationship between the situation value and the pollution degree by modifying the weight of the network over time using a recurrent network selected according to the condition, and the situation value and the pollution degree already learned. And a means for predicting the degree of contamination at an arbitrary time using a recurrent network selected according to the above conditions.

[0007]

The various factors in the past, such as wind speed and the number of vehicles passing through, are related to the pollution of the internal air of a tunnel or the like. In order to learn this relationship, the present invention uses a recurrent network. .

The recurrent network is a network having a context layer in addition to a hierarchical structure including an input layer, one or more intermediate layers, and an output layer. Since the past history is retained in this context layer, temporal order relations can be learned. There are several types of recurrent networks, but the typical one is Elman's network, which has one middle layer and the number of units in the context layer is equal to the number of units in the middle layer. Is a network with a connection to.

The learning of the temporal order relation in the present invention is performed as follows. First, time t = 1, 2, ..., T
At the wind speed W
(t), the number of passing vehicles C (t), and the VI value V (t) as an index indicating the degree of air pollution are stored as actual values. Set the value of the context layer to 0. Wind speed W (1) with t = 1
And the number of passing vehicles C (1) are input to the input layer to obtain an output value.
To reduce the error from the actual VI value V (1),
The backpropagation method is used to modify the weights of the network while tracing from the output layer to the input layer in the opposite direction.

The value of the middle layer in the above is copied to the context layer, and when t = 2, the wind speed W (2) and the number of passing vehicles C
Input (2) into the input layer and obtain the output value. Actual VI value V
The network weight is corrected using the back propagation method so that the error between (2) and (2) is reduced. Similarly to the above, t = 3, 4, ..., T is sequentially repeated, and the network weight is corrected by using the back propagation method.

It is judged whether or not the total error is smaller than a predetermined value, and if it is smaller, the learning is ended, but if not, the process returns to that and the learning is continued. At the time point when the learning is completed, the temporal relationship among the wind speed, the number of passing vehicles, and the VI value learned by the network is stored.

Therefore, the VI value V (n) at t = n is obtained by the following procedure. A) Set the value of the context layer to 0. Wind speed W with t = 1
(1) and the number of passing vehicles C (1) are input to the input layer. B) The value of the intermediate layer in A) is copied to the context layer, and the wind speed W (2) and the number of passing vehicles C (2) are input to the input layer with t = 2. C) Similarly, if t = 3, 4, ..., T is sequentially repeated, the output value at t = n is V (n), and the VI predicted value V (n) at any time n is obtained. .

In the present invention, as described above, the recurrent network is made to learn the relationship between the situation value and the air pollution degree. In the first and second inventions, when the pollution degree is predicted, the situation does not become exactly the same as that learned, but the neural network has generalization ability,
If it is similar to the learned situation, it is possible to predict a pollution degree close to the learned situation. In particular, in the second invention described in claim 2, since a plurality of recurrent networks are trained according to conditions, at the time of prediction, the most suitable network is selected according to the conditions at that time, and its value is output. As a result, the prediction accuracy can be further improved.

[0014]

Embodiments of the present invention will be described below with reference to the drawings. First, FIG. 1 shows an apparatus configuration of an embodiment of the first invention, in which one network is provided. In FIG. 1, V as an index showing the degree of air pollution in the tunnel
A VI value measuring means 1 for measuring an I value, a passing vehicle number measuring means 2 for measuring a passing amount of vehicles as a situation value that affects the degree of air pollution, and a wind speed measuring means 3 for similarly measuring a wind speed in a tunnel. By using it, the air pollution degree in the tunnel and the situation value affecting it are measured over time. These collected data are stored in the performance storage means 4 as performance values.

As will be described in detail later, the learning means 5 causes the learning network 6A to learn the relationship between the number of passing vehicles, the wind speed, and the VI value using the above-mentioned actual values until the error in the VI value becomes sufficiently small. The prediction means 7 predicts the next VI value using the prediction network 6B that has already been learned, and obtains the VI predicted value 8 indicating the air pollution degree. Here, since the contents of the prediction network 6B are copied from the learning network 6A at regular time intervals, the contents of both are substantially the same and can be said to be a single network. However, as a result of learning, incorrect data may be stored, so the prediction network 6B
It may be checked by a human before copying to.

Next, FIG. 2 shows a device configuration of an embodiment of the second invention. In this embodiment, the actual values to be learned by the network are classified into three types, that is, fine at daytime, rainy at daytime, and nighttime,
Different networks are made to learn about these different conditions of time and weather. As shown in the figure, in this embodiment, unlike the embodiment of FIG. 1, the CO value (carbon monoxide concentration) is used as an index indicating the pollution degree of air instead of the VI value, but other indexes may be used. Alternatively, it is also possible to use other indexes, for example, the VI value and the like together.

As the status value, the number of passing vehicles predicted by the passing number predicting means 12 and the atmospheric pressure measured by the atmospheric pressure measuring means 13 are used, and these and the CO value measured by the CO value measuring means 11 are actual values. Is stored in the result storage means 14. Note that other types of values can be used for the above situation value.

The learning means 15 selects the learning network according to the condition of the actual value from the learning network selecting means 19 among 161A, 162A, 163A, and the predicting means 17 selects the prediction networks 161B, 162B, 1
The predictive network selecting means 20 selects one of 63B that best matches the current time and weather conditions, and sets the output value as the predicted CO value 18 indicating the air pollution degree.
Similar to the embodiment of FIG. 1, the contents of the prediction networks 161B, 162B, 163B are copies of the learning networks 161A, 162A, 163A at fixed time intervals.

FIG. 3 shows a schematic configuration of the recurrent network used in each of the above embodiments. This network has a context layer 34 in addition to a hierarchical structure including an input layer 31, an intermediate layer 32, and an output layer 33. Middle layer 32
Is one layer, the number of units (the number of neurons) of the context layer 34 is equal to the number of units of the intermediate layer 32, and there is a connection from the context layer 34 to the intermediate layer 32. Here, the middle layer 3
Although the case where 2 is one layer is shown, the number of layers may be two or more, and accordingly, the context layer 34 may be two or more layers.

FIG. 4 shows a specific configuration of this recurrent network. In the input layer 31, the number of large vehicles currently passing, the number of large vehicles predicted to pass after Δt, the wind speed of the road, the natural wind speed, the number of jet fans operating, the dust collector operating rate, and the current VI value are input. Has a total of 7 units. The intermediate layer 32 connected to the input layer 31 has a total of 10 units, and the context layer 34 also has a total of 10 units. The output layer 33 connected to the intermediate layer 32 has Vt after Δt.
It has only one unit corresponding to the I predicted value.

FIG. 5 shows a flowchart of learning by the recurrent network. Time t = 1, 2, ...
..., up to T, the number of large vehicles L (t), road wind speed Wr (t),
Natural wind speed Wn (t), jet fan operating number J (t), dust collector operating rate D (t), VI value V (t), and time t = T +
It is assumed that the number of large vehicles L (T + 1) and VI value V (T + 1) in 1 are stored as actual values. Here, the unit of time (time interval Δt) is 5 minutes.

In FIG. 5, the value of the context layer 34 is set to 0 and t = 1 is set (S1). Next, the number of large vehicles L (t),
L (t + 1), roadway speed Wr (t), natural wind speed Wn (t), jet fan operating number J (t), dust collector operating rate D (t), V
The I value V (t) is input to the input layer 31, and the output value is obtained from the output layer 33. The network weight is corrected by using the back propagation method so that the output error between the actual value and the VI value V (t + 1) becomes small (S2).

Next, the value of the intermediate layer 32 in step S2 is copied to the context layer 34 (S3). The above process is repeated until t = T. That is, when t = T, the process proceeds to the next step S5, otherwise t is incremented by 1 and the process returns to step S2 (S4). It is determined whether or not the evaluation value of the overall error is smaller than a predetermined value, and if it is smaller, the learning ends there. If not, the process returns to step S1 to continue learning (S5).

When the learning is completed in this manner, the temporal relationship among the wind speed, the number of large vehicles passing through and the VI value learned by the learning network is stored.

Next, FIG. 6 shows a flow chart of VI value prediction by the recurrent network. Time t =
1, 2, ..., N, the number of large vehicles L (t), road wind speed Wr (t), natural wind speed Wn (t), jet fan operating number J (t), dust collector operating rate D (t), It is assumed that the VI value V (t) is stored as the actual value, and the number of large vehicles L (n + 1) at time t = n + 1 is known as the predicted value.

In FIG. 6, the value of the context layer 34 is set to 0 and t = 1 is set (S11). Next, the number of large vehicles L
(t), L (t + 1), roadway wind speed Wr (t), natural wind speed Wn
(t), jet fan operating number J (t), dust collector operating rate D
(t) and VI value V (t) are input to the input layer 31, and the output layer 33
The output value is obtained from (S12).

Next, the value of the intermediate layer 32 in step S12 is copied to the context layer 34 (S13). The above process is repeated until t = n. That is, when t = n, the process proceeds to the next step S15, otherwise t
Is incremented by 1 and the process returns to step S12 (S14). The output value at this time becomes the VI predicted value V (n + 1) as the air pollution degree at an arbitrary time t = n (S15).

Next, when performing prediction at t = n + 1,
By saving the state of the network at t = n, iterative calculations from t = 1 to n can be omitted, and the VI value is continuously predicted by immediately proceeding from step S12 to step S15. It is possible.

[0029]

As described above, the first or second invention is
A recurrent network is used to learn the relationship between the air pollution degree and the situation value affecting it over time, and the learning result predicts the air pollution degree at any time.

Therefore, even if the actual VI value or the like shows the influence of the past time, it is possible to predict the pollution degree with high accuracy by reflecting them, and the ventilation of the inside of the tunnel or the like can be made more appropriate. It can be done efficiently. Further, since the present invention is not particularly limited in the type of situation value, it can be applied not only to predicting the degree of pollution of air in tunnels but also to predicting the degree of pollution in various facilities that require periodic ventilation. Is.

Particularly, in the second aspect of the invention, the prediction accuracy can be further improved by learning and predicting according to conditions such as time and weather.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of the first invention.

FIG. 2 is a block diagram showing the configuration of an embodiment of the second invention.

FIG. 3 is a diagram showing a schematic configuration of a recurrent network used in each example.

FIG. 4 is a diagram showing a specific configuration of a recurrent network used in each example.

FIG. 5 is a flowchart of learning by the recurrent network of each embodiment.

FIG. 6 is a flowchart of prediction by the recurrent network of each embodiment.

[Explanation of symbols]

1 VI value measuring means 2 Passing vehicle number measuring means 3 Wind speed measuring means 4,14 Result storage means 5,15 Learning means 6A, 161A, 162A, 163A Learning network 6B, 161B, 162B, 163B Prediction network 7,17 Prediction means 8 VI Prediction value 11 CO value measurement means 12 Passing vehicle number prediction means 13 Atmospheric pressure measurement means 18 CO prediction value 19 Learning network selection means 20 Prediction network selection means 31 Input layer 32 Intermediate layer 33 Output layer 34 Context layer

 ─────────────────────────────────────────────────── ─── Continuation of front page (72) Inventor Kazuteru Shingai 1-1 Tanabe Nitta, Kawasaki-ku, Kawasaki-shi, Kanagawa Fuji Electric Co., Ltd. (72) Kazuo Ono 1 Tanabe Nitta, Kawasaki-ku, Kawasaki-shi, Kanagawa No. 1 within Fuji Electric Co., Ltd.

Claims (2)

[Claims] 1. A means for measuring the degree of air pollution, a means for measuring a situation value affecting the degree of air pollution, a means for storing the pollution degree and the situation value as actual values, an input layer, an intermediate layer, and a context. By modifying the weight of the network over time using a recurrent network so that the error between a single recurrent network consisting of a layer and an output layer and the actual value stored over time becomes small. A means for learning the relationship between the situation value and the pollution degree, and means for predicting the pollution degree at an arbitrary time using a recurrent network based on the relationship between the situation value and the pollution degree that have already been learned. An air pollution degree prediction device characterized in that 2. A means for measuring an air pollution degree, a means for measuring a situation value that affects the air pollution degree, a means for storing the pollution degree and the situation value as actual values, an input layer, an intermediate layer, and a context. Layer and output layer, and a plurality of recurrent networks provided for each condition that affects the degree of air pollution, and an error between the actual value stored over time is reduced according to the condition. Means for learning the relationship between the situation value and the pollution degree by modifying the weight of the network over time using the selected recurrent network, and based on the relationship between the situation value and the pollution degree that have already been learned. An air pollution degree prediction device comprising: a means for predicting a pollution degree at an arbitrary time using a recurrent network selected according to the condition.