JP6977605B2 - Sewer pipe culvert water level prediction device and sewer pipe culvert water level prediction method - Google Patents

Description

The present invention relates to a sewer pipe in-ditch water level predicting device and a sewer pipe in-drain water level predicting method.

In recent years, as a countermeasure against flood damage caused by torrential rain, etc., the development of technology for predicting the water level in sewer pipes has been promoted.

Patent Document 1, which is an example of the prior art, discloses a real inundation map system capable of predicting an inundation situation and an inundation damage situation and controlling the management of forced drainage. The real inundation map system disclosed in Patent Document 1 predicts the water level in the sewer pipe using at least the past and present rainfall data of the target area and the topographical data of the target area.

Further, Non-Patent Document 1 which is an example of the prior art discloses a dam inflow prediction system using a neural network and fuzzy theory. The dam inflow prediction system disclosed in Non-Patent Document 1 fuses the output values of two neural networks learned in a steady state with a small inflow and a flood with a large inflow by fuzzy inference to obtain an outflow. Predict. The training data used for training the two neural networks is divided into steady time and flood time, and the index of division is determined by the value of the dam inflow amount determined in advance according to the magnitude of the dam inflow amount. ..

Japanese Unexamined Patent Publication No. 2002-29863

Tatsuya Iizaka, Tetsuro Matsui, Yoshiteru Ueki, "Development of Dam Inflow Prediction System by Neuro-Fuzzy", IEEJ Transactions on Electrical Engineers, 1999, Vol. 119, No. 10, p. 1020-1025

However, according to the technique disclosed in Patent Document 1, various data including topographical data of the target area are required, and the amount of information to be set is enormous, so that a large amount of time and cost are required. There was a problem.

Further, in the technique disclosed in Non-Patent Document 1, since the training data is divided by a predetermined value, the training data may be small or absent in each divided range. In such a case, the training data may be small or absent. There was a problem that the reliability of the prediction was low.

The present invention has been made in view of the above, and an object of the present invention is to predict the water level in a sewer pipe in a short time and at low cost while ensuring the reliability of the prediction.

The present invention that solves the above-mentioned problems and achieves an object is a sewer pipe water level predicting device that predicts the water level in the sewer pipe at the predicted target time in the target area, and is in the sewer pipe in the target area. Data of the data storage unit, a data storage unit that stores water level data and mesh precipitation data, a model learning unit that learns a plurality of linear regression models using the data of the data storage unit, and determines parameters. A threshold generation unit that generates a threshold for dividing the water level in the sewer pipe, a learning parameter storage unit that stores the parameters of the plurality of linear regression models, and the thresholds. It is a sewer pipe water level prediction device provided with a water level prediction value calculation unit that calculates a water level prediction value in the sewer pipe using the data of the data storage unit and the parameter stored in the learning parameter storage unit. ..

In the sewer pipe water level prediction device having the above configuration, the threshold value generation unit can generate the threshold value by clustering.

Alternatively, in the sewer pipe water level prediction device having the above configuration, the threshold value generation unit may generate the threshold value by a neighborhood search method.

In the sewer pipe water level prediction device that generates the threshold value of the above configuration by the neighborhood search method, it is preferable that the objective function of the neighborhood search method includes the number of training data.

Alternatively, the present invention is a method for predicting the water level in the sewer pipe in the target area at the predicted target time, using the water level data in the sewer pipe and the mesh precipitation data in the target area. Then, by training a plurality of linear regression models to determine parameters, and using the sewerage pipe water level data and mesh precipitation data, a threshold for dividing the water level in the sewerage pipe is generated. That, the parameters and the thresholds of the plurality of linear regression models are stored, the water level data in the sewer pipe and the mesh precipitation data, and the parameters are used to calculate the predicted value of the water level in the sewer pipe. Including that.

According to the present invention, it is possible to predict the water level in the sewer pipe in a short time and at low cost while ensuring the reliability of the prediction.

It is a block diagram which shows the structure of the water level prediction device in a sewer pipe which concerns on Embodiment 1. FIG. It is a flowchart which shows the operation of the model learning part shown in FIG. It is a flowchart which shows the operation of the water level prediction value calculation part shown in FIG. It is a figure which shows the time-dependent change of the water level in a sewer pipe with the horizontal axis as time, and the vertical axis as the water level in a sewer pipe. It is a block diagram which shows the structure of the threshold value generation part provided in the water level prediction device in a sewer pipe which concerns on Embodiment 2. FIG. It is a flowchart which shows the operation of the threshold value generation part provided in the water level prediction apparatus in a sewer pipe which concerns on Embodiment 2. FIG. It is a flowchart which shows the operation of the threshold value generation part provided in the water level prediction apparatus in a sewer pipe which concerns on Embodiment 3. FIG.

Hereinafter, a mode for implementing the sewer pipe in-ditch water level prediction device and the sewer pipe in-drain water level prediction method of the present invention will be described with reference to the attached drawings. However, the present invention is not limited to the description of the following embodiments.

<Embodiment 1>
FIG. 1 is a block diagram showing a configuration of a water level prediction device in a sewer pipe according to the present embodiment. The sewer pipe water level prediction device 100 shown in FIG. 1 includes a data storage unit 101, a model learning unit 102, a threshold value generation unit 103, a learning parameter storage unit 104, and a water level prediction value calculation unit 105. ..

The data storage unit 101 acquires and stores mesh precipitation data and sewer pipe water level data in real time. The mesh precipitation data includes the actual mesh precipitation value and the predicted mesh precipitation value in a certain range including the target point. In addition, the water level data in the sewer pipe includes the water level in the sewer pipe around the target point. Here, the mesh precipitation actual value may be used as the mesh precipitation past the present time, and the mesh precipitation predicted value may be used as the mesh precipitation before the present time. The mesh precipitation data and the water level data in the sewer pipe accumulated in the data storage unit 101 are supplied to the model learning unit 102, the threshold value generation unit 103, and the water level prediction value calculation unit 105. The data storage unit 101 can be realized by a recording medium such as a semiconductor memory and a magnetic disk.

Model learning unit 102, the water level data in a mesh precipitation data and sewerage sewer stored in the data storage unit 101, and acquires the number of clusters K and initial threshold [delta] _k ^init is input by the user, these Multiple linear regression models are constructed using the data from, and training is performed to determine the parameters. The model learning unit 102 generates a linear regression model having the same number as the number of regions (number of clusters) of the water level in the sewer pipe divided by the _{threshold value ^ δ k} generated by the threshold value generation unit 103. Here, "^ δ _k " is an alternative notation for characters in which "^ (circumflex)" is placed on top of _{"δ k".} The number of regions (number of clusters) of the water level in the divided sewer pipe is the number obtained by adding 1 to the number of threshold values generated by the threshold value generation unit 103. The model learning unit 102 can be realized by a processor such as an MPU (Micro-Processing Unit) and a CPU (Central Processing Unit).

Threshold generator unit 103 acquires the accumulated mesh precipitation data and sewerage sewer in water level data in the data storage unit 101 and the initial threshold value [delta] _k ^init, sewer pipe culvert these data by clustering Generate a _{threshold ^ δ k} that divides the water level. A plurality of linear regression models generated by the model learning unit 102 are assigned to each cluster by _{the threshold value ^ δ k} of the water level in the sewer pipe generated by the threshold value generation unit 103. The threshold value generation unit 103 can be realized by a processor such as an MPU and a CPU.

The learning parameter storage unit 104 stores the parameters of the plurality of linear regression models determined by the model learning unit 102 and the threshold value ^ δ _{k of the} cluster generated by the threshold value generation unit 103. The learning parameter storage unit 104 can be realized by a recording medium such as a semiconductor memory and a magnetic disk.

The water level prediction value calculation unit 105 includes mesh precipitation data and sewer pipe water level data stored in the data storage unit 101, a threshold value ^ δ _k stored in the learning parameter storage unit 104, and a plurality of linear regression models. And the parameter of is acquired, and the predicted value of the water level in the sewer pipe at the predicted target time is calculated. Here, when calculating the predicted value of the water level in the sewer pipe, among the plurality of linear regression models constructed by the model learning unit 102, the linear regression of the region to which the current water level value in the sewer pipe obtained from the data storage unit 101 applies. Select a model. The water level prediction value calculation unit 105 can be realized by a processor such as an MPU and a CPU.

It should be noted that among the configurations included in the above-mentioned sewer pipe water level prediction device 100, those realized by the processor may be realized by the same one processor. Further, among the configurations included in the above-mentioned sewer pipe water level prediction device 100, those realized by the recording medium may be realized by the same one recording medium.

<Operation of model learning unit 102>
FIG. 2 is a flowchart showing the operation of the model learning unit 102 shown in FIG. First, when the process is started, the model learning unit 102 acquires the mesh precipitation data and the water level data in the sewer pipe from the data storage unit 101 (S101: data acquisition step).

Next, the model learning unit 102, based on user input, to set the number of clusters K, a K-1 pieces of initial threshold δ _k ^init (S102: data input step). The number of clusters K is the number of divisions of the water level data in the sewer pipe, and is a value determined and input by the user using an input device (not shown). Initial threshold [delta] _k ^init, depending on the sewer pipe culvert in water level overlooking the prediction, a set value input is determined by the user. As an example, if it is desired to predict the water level in the sewer pipe in the range of 50% to 80% of the full pipe, the number of clusters K = 3, 2 initial thresholds δ ₁ ^init = 50%, δ ₂ ^init. = 80% is entered by the user. The model learning unit 102 sets the number of clusters K = 3 and the initial threshold values δ ₁ ^init = 50% and δ ₂ ^init = 80%. Initial threshold [delta] _k ^init is supplied to the threshold generator 103, a threshold generation unit 103, as described below, to generate a threshold ^ [delta] _k from the initial threshold value [delta] _k ^init.

Next, the model learning unit 102 _{acquires the threshold value ^ δ k} from the threshold value generation unit 103 (S103: threshold value acquisition step). These S102 to S103 are referred to as clustering steps. Here, k is an integer from 1 to K-1.

Next, the model learning unit 102 divides the training data into clusters according to the water level in the sewer pipe using the _{threshold value ^ δ k acquired in S103, and trains the linear regression model in each cluster (S104).} : Linear regression model learning step).

Next, the model learning unit 102 outputs the threshold value ^ δ _k acquired in S103 and the parameters of the plurality of linear regression models obtained by learning in S104 to the learning parameter storage unit 104, and ends the process (. S105: Learning result output step).

<Operation of threshold generation unit 103>
Threshold generator 103, the model learning unit 102 obtains the initial threshold [delta] _k ^init as the initial value, the clustering method generates a threshold ^ [delta] _k based on the distribution of the training data. Here, as a clustering method, a k-means method and a mixed normal distribution can be exemplified.

<Operation of water level prediction value calculation unit 105>
FIG. 3 is a flowchart showing the operation of the water level prediction value calculation unit 105 shown in FIG.

First, when the process is started, the water level prediction value calculation unit 105 has the threshold value ^ δ _k generated by the threshold value generation unit 103 from the learning parameter storage unit 104, and a plurality of determinations determined by the model learning unit 102. Acquire the parameters of the linear regression model (S201: model parameter acquisition step).

Next, the water level prediction value calculation unit 105 acquires the latest mesh precipitation data and sewer pipe water level data at the current time from the data storage unit 101 (S202: current data acquisition step).

Next, the water level prediction value calculation unit 105 selects the cluster to which the water level in the sewer pipe at the current time belongs by comparing the water level in the sewer pipe at the current time acquired in S202 with the threshold value ^ δ _k. (S203: Cluster selection step).

Next, the water level prediction value calculation unit 105 predicts the water level in the sewer pipe at a time before the current time by using the linear regression model of the cluster determined in S203 (S204: water level prediction step).

Next, the water level prediction value calculation unit 105 calculates the water level in the sewer pipe pipe predicted in S204 as the water level prediction value in the sewer pipe pipe (S205: water level prediction value output step in the sewer pipe pipe). Here, if the current time has reached the predicted end time, the process ends. If the current time has not reached the predicted end time, it returns to S202, acquires the mesh precipitation data and the water level data in the sewer pipe at a later time from the data storage unit 101, and S202 until the predicted end time is reached. The process of ~ S205 is repeated.

As the output destination of the predicted water level in the sewer pipe in S205, a display and a printer, which are user interfaces, can be exemplified.

FIG. 4 is a diagram showing the time course of the water level in the pipe, with the horizontal axis representing the time and the vertical axis representing the water level in the pipe. Here, as shown in FIG. 4, the _{water level in the conduit is divided into three regions by the threshold values δ 1} and δ ₂ , and the first linear regression model is set in the first cluster, and the first linear regression model is set. A second linear regression model is set in the two clusters, and a third linear regression model is set in the third cluster. In FIG. 4, the water level in the pipe at a predetermined time earlier than the current time is represented by a black circle.

According to the present embodiment described above, by using a plurality of linear regression models without using the topographical data of the target area, etc., the water level prediction is highly accurate by the mesh precipitation data in a certain range and the water level data in the sewer pipe. Can be done. Further, according to the present embodiment, since the threshold value for dividing the data is generated by clustering, it is possible to perform learning in consideration of the distribution and number of learning data. Furthermore, since a linear regression model is used, high-speed learning is possible.

<Embodiment 2>
In the first embodiment, a sewer pipe water level prediction device including a threshold value generation unit that generates a threshold value by clustering has been described, but the present invention is not limited thereto. In this embodiment, a sewer pipe water level prediction device including a threshold value generation unit that generates a threshold value by a local search method will be described.

The sewer pipe water level prediction device according to the present embodiment differs from the sewer pipe water level prediction device shown in FIG. 1 only in the configuration and operation of the threshold generation unit, and the other configurations are shown in FIG. Since it is the same as that of the above, the description of the first embodiment is incorporated.

First, as a premise of this embodiment, if the acquisition period of the training data is short and the reliability of the predicted value by cross-validation of the training data is low, if the global optimization is performed, the reliability of the predicted value is significantly lowered. There is. On the other hand, in the local optimization by the local search method, since the user having domain knowledge determines the initial threshold value, it is possible to avoid a large deviation from the initial threshold value, and the acquisition period of training data is short. In some cases, the reliability of the predicted value can be guaranteed. The threshold value generation unit in the present embodiment performs cross-validation on the training data and performs a local search method using the obtained prediction accuracy as an objective function to generate a threshold value with high prediction accuracy.

<Configuration of threshold generation unit 103A>
FIG. 5 is a block diagram showing the configuration of the threshold value generation unit 103A included in the water level prediction device in the sewer pipe according to the present embodiment. The threshold value generation unit 103A shown in FIG. 5 includes a current threshold value storage unit 131, an end determination unit 132, a neighborhood threshold value generation unit 133, a cluster division unit 134, a cross-validation unit 135, and a threshold. A value update unit 136 is provided.

The current threshold storage unit 131 stores the current threshold. Currently, the threshold storage unit 131 can be realized by a recording medium such as a semiconductor memory and a magnetic disk.

The end determination unit 132 determines the end according to the end condition given according to the algorithm of the local search method. The end determination unit 132 can be realized by a processor such as an MPU and a CPU.

Near threshold generator 133, according to the algorithm of local search method, and generates a proximity threshold ^¯ [delta] _k ⁿ of the current threshold [delta] _k. Here, " ^￣ δ _k ⁿ " is an alternative notation for characters in which ^"￣ (overline)" is placed on top of " _{δ k} ^n". Further, n is an integer from 1 to N, and N is an element of a set of positive integers. The neighborhood threshold generation unit 133 can be realized by a processor such as an MPU and a CPU.

Clustering unit 134 divides the vicinity threshold ^¯ [delta] _k ⁿ the water level in the tube culvert into K clusters. The cluster division unit 134 can be realized by a processor such as an MPU and a CPU.

The cross-validation unit 135 performs cross-validation on the training data and calculates the obtained prediction accuracy F as an evaluation value. The cross-validation unit 135 can be realized by a processor such as an MPU and a CPU.

Threshold updating unit 136, based on the prediction accuracy F is an evaluation value in the vicinity of the threshold ^¯ [delta] _k ^n, to update the current threshold [delta] _k in accordance with the algorithm of local search. The threshold value update unit 136 can be realized by a processor such as an MPU and a CPU.

<Operation of threshold generation unit 103A>
FIG. 6 is a flowchart showing the operation of the threshold value generation unit 103A included in the water level prediction device in the sewer pipe according to the present embodiment. First, when starting the process, the threshold value generating unit 103A, the initial threshold [delta] _k ^init with the current threshold value [delta] _k (S301: initial threshold determination step).

Next, the end determination unit 132 determines whether or not the end condition is satisfied (S302: end determination step). Here, the end condition is given according to the algorithm of the local search method. When the end condition is satisfied (S302: end), the threshold generation unit 103A outputs the current threshold value δ _k as the threshold value ^{^} δ _k (S307: threshold value output step), and performs processing. finish.

If the end condition is not satisfied: in (S302 continuation) is near the threshold value generating unit 133, according to the algorithm of local search method, and generates a proximity threshold ^¯ [delta] _k ⁿ of the current threshold [delta] _k ( S303: Neighborhood threshold generation step).

Then, the cluster dividing unit 134 divides the vicinity threshold ^¯ [delta] _k ⁿ the water level in the tube culvert into K clusters (S304: Clustering step).

Next, the cross-validation unit 135 performs cross-validation on the learning data and calculates the obtained prediction accuracy F as an evaluation value (S305: evaluation step based on the prediction accuracy F). Here, in cross-validation, the training data is divided into simulated learning data and simulated verification data, model learning is performed using the simulated learning data, and water level prediction is performed using the simulated verification data to obtain a prediction accuracy F. Then, the generalization performance is evaluated.

Next, the threshold updating unit 136, based on the prediction accuracy F is an evaluation value in the vicinity of the threshold ^¯ [delta] _k ^n, to update the current threshold [delta] _k in accordance with the algorithm of local search (S306 : Threshold update step). After the threshold value update unit 136 stores the updated _{current threshold value δ k} in the current threshold value storage unit 131, the end determination unit 132 determines whether or not the end condition is satisfied (S302: When the end condition is satisfied (S302: end), the threshold generation unit 103A _{outputs the current threshold value δ k} (S307: threshold value output step), and ends the process. ..

Of the configurations included in the above-mentioned threshold value generation unit 103A, those realized by the processor may be realized by the same one processor, and the configuration included in the above-mentioned threshold value generation unit 103A may be realized. Of these, those realized by the recording medium may be realized by the same one recording medium.

According to the present embodiment described above, since the prediction accuracy can be evaluated, the accuracy of the water level prediction value in the pipe can be further improved as compared with the first embodiment.

<Embodiment 3>
In the first embodiment, a water level prediction device in a sewer pipe including a threshold value generation unit that generates a threshold value by clustering will be described, and in the second embodiment, the objective function of the local search method is simply defined as the prediction accuracy by cross-validation. Although the water level prediction device in the sewer pipe including the threshold value generation unit that generates the threshold value by the local search method of objective optimization has been described, the present invention is not limited thereto. In this embodiment, in addition to the prediction accuracy by cross-validation, the threshold generation unit that generates a threshold by a local search method of multi-objective optimization that includes the number of training data held by each divided cluster in the objective function. The water level prediction device in the sewer pipe culvert equipped with the above will be described.

The sewer pipe water level prediction device according to the present embodiment differs from the sewer pipe water level prediction device shown in FIG. 1 only in the configuration and operation of the threshold generation unit, and the other configurations are as shown in FIG. Since they are the same, the description of the first embodiment is incorporated. Further, the threshold value generation unit included in the water level prediction device in the sewer pipe according to the present embodiment differs only in the operation from the threshold value generation unit shown in FIG. 5, and the configuration is the same as that shown in FIG. Therefore, the description of the second embodiment is used.

<Operation of threshold generation unit 103A>
FIG. 7 is a flowchart showing the operation of the threshold value generation unit included in the water level prediction device in the sewer pipe according to the third embodiment. In FIG. 7, only S305 in FIG. 6 is changed to S305A, and the other steps are the same as in FIG. 6, so the description of the second embodiment is incorporated.

Threshold generator unit 103A calculates the prediction accuracy F obtained in cross-validation in the training data, the training data number M ^k held by each cluster divided by the proximity threshold ^¯ [delta] _k ^n, the prediction accuracy Using F, the prediction accuracy is evaluated by the evaluation function of the following equation (1) (S305A: evaluation step based on the ^{prediction accuracy F and the number of training data M k).}

The multi-objective function of the above equation (1) can be exemplified by using the function f as a weighted sum function and calculating the evaluation value by the following equation (2) as a general solution. Note that w ₁ and ω _k are weight constants.

In the above formula (2), it is possible to set the priority to the number of learning data for each cluster by appropriately setting the weight omega _k on learning data number M ^k.

Alternatively, in the multipurpose function of the above equation (1), the evaluation value is only the prediction accuracy F, and ^{the maximization of the number of training data Mk} is reduced to the constraint condition of the following equation (3) by the ε constraint method. It can be exemplified as a solution method.

In the above equation (3), the priority can be set by increasing _{the ε k of the cluster having a high priority.}

As a result, in addition to guaranteeing the number of training data for the cluster in the range where the number of training data tends to decrease (for example, high water level data), the data that the user having domain knowledge wants to predict preferentially can be obtained. It is also possible to make adjustments such as increasing the priority of the cluster to which it belongs. That is, it is possible to increase the degree of freedom in the prediction accuracy of each cluster while ensuring the overall prediction accuracy F.

As an example, a case where the number of clusters is K = 2 and the water level is divided into a high water level range and a low water level range will be described. Here, the number of training data in the low water level range is much larger than the number of training data in the high water level range, and the number of training data in the low water level range is sufficiently larger than the number of training data in the high water level range, which is 100 times or more. And.

At this time, the overall prediction accuracy F is higher when fitting to the data in the low water level range where the number of training data is large than when fitting to the data in the high water level range where the number of training data is small. Therefore, the threshold value δ ₁ is designed so that the number of training data in the low water level range is large, that is, the threshold value δ ₁ is high.

However, it is especially important to predict the high water level range as a countermeasure against flood damage caused by torrential rain. _{Therefore, if the weight ω 2} of the number of data in the high water level range is increased, the number of data in the high water level range is secured, that is, the threshold value δ ₁ is designed to be low, and the overall prediction accuracy F is determined. While maintaining high, the prediction accuracy of data in the high water level range can be improved.

According to the present embodiment described above, in the local search method for determining the threshold value of the water level in the conduit, the prediction accuracy as a whole is improved by including the number of training data in the evaluation value in addition to the prediction accuracy. Even when the number of training data of a cluster of high importance is small, the prediction accuracy of the cluster can be improved.

The present invention is not limited to the above-described embodiment, and includes various modifications in which components are added, deleted, or converted from the above-mentioned configuration.

100 Sewer pipe water level prediction device 101 Data storage unit 102 Model learning unit 103, 103A Threshold generation unit 104 Learning parameter storage unit 105 Water level prediction value calculation unit 131 Current threshold value storage unit 132 End judgment unit 133 Near threshold Value generation unit 134 Cluster division unit 135 Cross-validation unit 136 Threshold update unit

Claims (5)

It is a sewer pipe water level prediction device that predicts the water level in the sewer pipe at the prediction target time in the target area.
A data storage unit that stores water level data in sewer pipes and mesh precipitation data in the target area,
A model learning unit that learns a plurality of linear regression models and determines parameters using the data of the data storage unit, and a model learning unit.
A threshold generation unit that generates a threshold value for dividing the water level in the sewer pipe using the data of the data storage unit, and
A learning parameter storage unit that stores the parameters and thresholds of the plurality of linear regression models, and
A sewer pipe water level prediction device including a water level prediction value calculation unit for calculating a water level prediction value in a sewer pipe using the data of the data storage unit and the parameters stored in the learning parameter storage unit. The water level prediction device in a sewer pipe according to claim 1, wherein the threshold value generation unit generates the threshold value by clustering. The water level prediction device in a sewer pipe according to claim 1, wherein the threshold value generation unit generates the threshold value by a local search method. The water level prediction device in a sewer pipe according to claim 3, wherein the objective function of the neighborhood search method includes the number of learning data. It is a method of predicting the water level in the sewer pipe in the sewer pipe at the prediction target time in the target area.
The model learning unit learns a plurality of linear regression models using the water level data in the sewer pipe and the mesh precipitation data in the target area to determine the parameters.
The threshold value generation unit uses the sewerage pipe water level data and the mesh precipitation data to generate a threshold value for dividing the water level in the sewerage pipe.
The learning parameter storage unit stores the parameters and the threshold values of the plurality of linear regression models.
A method for predicting a water level in a sewer pipe, which comprises calculating a water level predicted value in the sewer pipe using the water level data in the sewer pipe, mesh precipitation data, and the parameters.

Images