JP2022015250A - Effluent quality prediction device and effluent quality prediction method - Google Patents

Abstract

To eliminate abnormal data included in time-series data from a plurality of sensors at a water treatment site.SOLUTION: In an effluent quality prediction device (100), a pre-processing unit (103a) performs filtering to a learning dataset, which is a portion of learning data for a first prescribed time in learning data. A model generation unit (103c) generates a prediction model for predicting effluent quality data a prescribed time after the learning dataset. Meanwhile, an effluent quality prediction unit (105) applies the prediction model to the latest data stored in a latest data database (104) and predicts effluent quality data.SELECTED DRAWING: Figure 1

Description

The present invention relates to a technique for predicting the quality of discharged water based on time-series data from a plurality of sensors in a sewage treatment plant.

As a technique for predicting the quality of discharged water from a sewage treatment plant, there is an invention disclosed in Patent Document 1 below. This Patent Document 1 relates to a technique for predicting a monitored amount of plant equipment such as water treatment equipment.

The primary predictor uses the data group, the reference annual average model and the reference seasonal average model to generate an annual average primary prediction model group including a neural model, a multiple regression model, and a clustering model, and a seasonal average primary prediction model group. do. Then, the predicted value of the monitored amount of each prediction model is calculated as the prediction result data using the annual average primary prediction model group and the seasonal average primary prediction model group.

The secondary forecasting unit generates a standard secondary forecasting model using the operation performance data (annual performance data and seasonal performance data). Then, using the reference secondary prediction model, the prediction result data, and the operation record data, a secondary prediction model for predicting the annual and seasonal average value of the monitored quantity is generated.

Finally, the secondary prediction unit assigns weights to the monitored predicted values (prediction result data), and uses a neural network with a plurality of weighted monitored targets as nodes to determine the predicted values of the monitored quantities. calculate.

Japanese Unexamined Patent Publication No. 2013-161336 (published on August 19, 2013)

However, Patent Document 1 does not describe a technique for removing abnormal data included in time-series data from a plurality of sensors in a sewage treatment plant.

One aspect of the present invention is to provide a effluent quality prediction device capable of removing anomalous data included in time-series data from a plurality of sensors in a sewage treatment plant.

In order to solve the above problems, the discharged water quality prediction device according to one aspect of the present invention includes a sensor information acquisition unit that acquires time-series data of detected values from a plurality of sensors installed in a sewage treatment plant, and a sensor. A learning data database that stores the time-series data acquired by the information acquisition unit as training data, a latest data database that stores the time-series data acquired by the sensor information acquisition unit as the latest data, and the first of the training data. The training data for a predetermined time is used as the training data set, and the preprocessing unit that filters the training data set and the prediction model for predicting the value of the discharged water quality data after the second predetermined time are preprocessed. It includes a model generation unit generated from the training data set after being filtered by the unit, and a discharge water quality prediction unit that predicts the value of the discharge water quality data by applying a prediction model to the latest data.

Further, in order to solve the above-mentioned problems, the discharged water quality prediction method according to one aspect of the present invention includes a step of acquiring time-series data of detected values from a plurality of sensors installed in a sewage treatment plant. The step of storing the acquired time-series data as training data, the step of storing the acquired time-series data as the latest data, and the training data for the first predetermined time in the training data as the training data set, and the training data. A step to filter the set, a step to generate a prediction model for predicting the value of the discharged water quality data after the second predetermined time from the training data set after being filtered, and the latest data. It includes the step of predicting the value of the discharged water quality data by applying the prediction model.

According to one aspect of the present invention, it is possible to remove anomalous data included in time-series data from a plurality of sensors in a sewage treatment plant.

It is a functional block diagram which shows the schematic structure of the effluent water quality prediction apparatus which concerns on one Embodiment of this invention. It is a figure which shows an example of the training data after being processed by a pre-processing part. It is a figure for demonstrating the steady state data and the non-steady state data in the discharged water quality data. It is a flowchart for demonstrating the processing procedure of the prediction model generation part. It is a figure which shows the adjustment item example of the hyperparameter of RNN used by the model generation part. It is a figure which shows the generation of the prediction model schematically. It is a figure which shows the case where the non-stationary data cannot be predicted correctly. It is a figure which shows the case where the non-stationary data can be predicted correctly. It is a flowchart for demonstrating the processing procedure of the pre-processing part. It is a flowchart for demonstrating the optimization of each parameter of a filter. It is a figure which shows the optimization of each parameter of a filter schematically. It is a flowchart for demonstrating the processing procedure of the non-stationary data change part. It is a figure for demonstrating sampling of training data. It is a figure for demonstrating the method of increasing the number of non-stationary data. It is a figure for demonstrating another method for increasing the number of non-stationary data. It is a figure which shows the initial extraction data of the non-stationary data. It is a figure for demonstrating the method of reducing the number of non-stationary data. It is a figure which shows typically the process of the non-stationary data change part shown in FIG. It is a flowchart for demonstrating the processing procedure of the effluent quality prediction unit. It is a figure which shows the prediction of the discharged water quality data schematically. It is a graph which shows the measured value of COD when the number of non-stationary data is not increased or decreased, and the predicted value of COD predicted by the discharged water quality prediction unit. It is a graph which shows the measured value of COD when the number of non-stationary data is increased or decreased, and the predicted value of COD predicted by the effluent quality prediction unit. It is a figure which shows the configuration example of the computer for realizing the effluent quality prediction apparatus.

Hereinafter, one embodiment of the present invention will be described in detail. For convenience of explanation, the same members are designated by the same reference numerals, and their names and functions are also the same. Therefore, their detailed description will not be repeated.

(Configuration of effluent quality prediction device 100)
FIG. 1 is a functional block diagram showing a schematic configuration of a effluent water quality prediction device 100 according to an embodiment of the present invention. The discharged water quality prediction device 100 includes a sensor information acquisition unit 101, a learning data database 102, a prediction model generation unit 103, a recent data database 104, a discharged water quality prediction unit 105, and an output unit 106.

Although not shown, the sewage treatment plant has pump pumping volume, return flow rate, air supply volume, mixed suspended matter concentration (MLSS (Mixed Liquor Suspended Solid)), dissolved oxygen concentration (DO (Dissolved Oxygen)), and COD. Sensors for measuring data (detection values) such as (Chemical Oxygen Demand), TN (Total Nitrogen) (total nitrogen), and TP (Total Phosphorus) (total phosphorus) are installed. The data from these sensors are sent to the discharged water quality prediction device 100 one by one.

COD, TN and TP are known as indicators of the quality of effluent from sewage treatment plants. Therefore, hereinafter, the data sent from the three sensors that measure each of COD, TN, and TP is referred to as discharged water quality data. On the other hand, data sent from sensors other than the above three sensors will be referred to as sensor information data.

The sensor information acquisition unit 101 acquires sensor information data and discharged water quality data, and adds time information data (year, month, day, day / holiday, and time) to these data to generate learning data. In the following description, the sensor information acquisition unit 101 acquires sensor information data and discharged water quality data every minute, but the present invention is not limited to this.

The learning data database 102 sequentially stores the learning data generated by the sensor information acquisition unit 101 for a period required for learning, for example, a period of one year or more.

The prediction model generation unit 103 includes a preprocessing unit 103a, a non-stationary data change unit 103b, a model generation unit 103c, a model evaluation unit 103d, and a model change unit 103e. The preprocessing unit 103a processes the learning data stored in the learning data database 102 so that it can be learned.

FIG. 2 is a diagram showing an example of training data after being processed by the preprocessing unit 103a. Each learning data consists of 5 items of time information data (year, month, day, day / holiday, and time) and 15 items of sensor information data (pump pumping amount, return flow rate, air supply amount, mixed suspended matter). Concentration, dissolved oxygen concentration, etc.) and three items of discharged water quality data (COD, TN, and TP).

FIG. 3 is a diagram for explaining steady-state data and non-steady-time data in the discharged water quality data. The alarm level is a threshold value at which the discharged water quality data deteriorates and it is necessary to take measures such as changing the operation. The constant data is the data of the time when the discharged water quality data is below the alarm level. The non-stationary data is data for the time when the discharged water quality data exceeds the warning level.

When the balance between the steady-state data and the non-stationary data of the discharged water quality is good (the ratio of the non-stationary data to the discharged water quality data for the whole period is large to some extent), the steady-time data and the non-stationary data of the discharged water quality are used. Can be predicted with high accuracy. However, when the ratio of the non-stationary data is small, the prediction accuracy of the stationary data is high, but the prediction accuracy of the non-stationary data is low.

The non-stationary data changing unit 103b extracts non-stationary data from the learning data generated by the preprocessing unit 103a, and increases or decreases the number of non-stationary data according to the ratio of the non-stationary data. Change the training data. The details of the non-stationary data changing unit 103b will be described later.

The model generation unit 103c relearns at predetermined update cycles using the learning data changed by the non-stationary data change unit 103b, and generates a prediction model corresponding to the operation change of the sewage treatment plant. The details of this model generation unit 103c will be described later.

The model evaluation unit 103d evaluates the prediction model generated by the model generation unit 103c, and if the evaluation error is equal to or less than the target error, it is determined that the prediction model is sufficiently learned. The details of this model evaluation unit 103d will be described later.

The model change unit 103e determines whether or not the discharged water quality prediction unit 105 is executing the prediction process. If the prediction process is not being executed, the model change unit 103e updates the prediction model generated by the model generation unit 103c as the latest prediction model.

The latest data database 104 sequentially stores the learning data generated by the sensor information acquisition unit 101 only for a period required for prediction, for example, for a period of one hour or more. That is, the latest data database 104 always stores the latest learning data for one hour or more as the latest data.

The discharged water quality prediction unit 105 includes a pretreatment unit 105a, a prediction processing unit 105b, and a post-treatment unit 105c. The preprocessing unit 105a processes the latest data stored in the latest data database 104 into a predictable format.

The prediction processing unit 105b predicts the discharged water quality after a predetermined time (eg, after several hours) using the latest data generated by the pretreatment unit 105a and the prediction model updated by the model change unit 103e. Details will be described later. In the following, a case where the predetermined time is 2.5 hours is mainly illustrated. In the present specification, this predetermined time is also referred to as a "second predetermined time".

The post-treatment unit 105c converts the discharged water quality predicted by the prediction processing unit 105b into a format that can be output.

The output unit 106 outputs various information generated by the discharged water quality prediction unit 105 to a control device, a display device, etc. (not shown) of the sewage treatment plant.

(Processing procedure of prediction model generation unit 103)
FIG. 4 is a flowchart for explaining the processing procedure of the prediction model generation unit 103. First, the preprocessing unit 103a acquires the learning data stored in the learning data database 102 (step S1). Then, the preprocessing unit 103a interpolates the missing value of the training data and removes the abnormal value (step S2).

Here, the missing value means a value when the measured value is not output correctly due to inspection of the sensor or the like. Further, the abnormal value means a value that is abnormal due to a sensor failure or the like. These are determined, for example, by whether or not the value is out of a certain range. The details of the process of this step S2 will be described later.

Next, the non-stationary data changing unit 103b calculates the ratio of the non-stationary data in the training data after the preprocessing is performed by the preprocessing unit 103a. Then, the non-stationary data changing unit 103b changes the learning data by increasing or decreasing the number of non-stationary data so that the ratio of the non-stationary data becomes a predetermined value (step S3). The details of the process of this step S3 will be described later.

Next, the model generation unit 103c creates a plurality of training data sets using the training data after being changed by the non-stationary data change unit 103b. As shown in FIG. 2, the training data for 24 hours (1440 points) surrounded by the thick line is regarded as one training data set. Then, the training data for 24 hours shifted down by one point (one line) is used as the next training data set. In this way, the learning data sets shifted downward by one point (one line) are sequentially created. In the present specification, the time of this training data set is also referred to as a "first predetermined time".

Hereinafter, the learning data for 24 hours (1440 points) will be described as one learning data set, but this is an example, and the learning data for other times (points) may be used as the learning data set.

Then, the model generation unit 103c creates a prediction model. In the present embodiment, it is decided to use a learning method of a recurrent neural network suitable for processing time series data, for example, a recurrent neural network (hereinafter referred to as RNN), and the current hyper. Create a prediction model using parameters.

FIG. 5 is a diagram showing an example of adjustment items of RNN hyperparameters used by the model generation unit 103c. Hyperparameters include the time of the training dataset, the number of epochs, the optimization function, the weight attenuation coefficient, the multiplier of the training rate, the update interval of the training rate, and the number of nodes. The model generation unit 103c generates a prediction model while sequentially adjusting the values of these hyperparameters.

The adjustment range of each item of hyperparameters is determined as shown on the right side of FIG. 5, and it is possible to change to these values. Therefore, it is possible to create a predictive model with all combinations of adjustment ranges of each item of hyperparameters. The hyperparameters are not limited to the items shown in FIG. 5, and may include the number of layers, the number of latent variables, the activation function, and the like.

The model generation unit 103c adjusts the parameters by selecting the values of each item of the hyperparameters shown in FIG. 5 and applying them to the RNN structure. Then, an RNN machine learning method, for example, a BPTT (Back Propagation Through Time) method is applied, and a prediction model is created using the adjusted hyperparameters (step S4).

Further, it is known that it is not always easy to reflect the past intermediate layer output in the RNN. Therefore, in order to improve this problem, it is also possible to use a network called LSTM (Long Short Term Memory). Since the concept of the length of the storage period is introduced in the LSTM for the intermediate layer output of the RNN, the influence of the output in the distant past can be retained.

It is also possible to use a method called GRU (Gated Recurrent Unit). In GRU, the input gate and the oblivion gate are integrated into one gate as an update gate. According to the GRU, it is easier to maintain a memory of the characteristics of the event before a long step.

Next, the model generation unit 103c divides the plurality of training data sets into training data, validation data, and evaluation data. Then, the model generation unit 103c uses the training data and the verification data to perform learning and verification with the prediction model created in step S4, and adjusts the network parameters so that the predicted values approach the actually measured values of the training data. (Step S5).

This learning and verification will be described in detail with reference to FIG. 2 again. The model generation unit 103c uses the training data set shown by the thick line in FIG. 2 as training data, and the COD value after 2.5 hours as verification data. Then, the model generation unit 103c sets the training data in the input layer of the prediction model created in step S4, and sets the value of COD after 2.5 hours as the verification data of the output layer. The model generation unit 103c updates the prediction model by performing re-learning using a predetermined learning method (BPTT method, LSTM method, GRU method, etc.). The verification data is the value of COD 2.5 hours after the training data at the latest time in the training data set.

FIG. 6 is a diagram schematically showing the generation of a prediction model. As training data 201, 23 items (5 items of time information data + 15 items of sensor information data + 3 items of discharged water quality data) of each row of the training data set are input to the input layer of the RNN layer, and the value of COD 2.5 hours later. Is set in the output layer as verification data. It should be noted that only 5 items of time information data + 15 items of sensor information data, for a total of 20 items, may be input to the input layer.

The RNN layer 202 is a recurrent neural network, and the passage of time of the intermediate layer must be taken into consideration. Therefore, the error back propagation (BP) method cannot be applied to the RNN layer 202 as it is. Therefore, the fully connected layer 203 is created by expanding the RNN layer in the time direction via the intermediate layer output, and the prediction model 204 in a format to which the BP method can be applied is created. The so-called BPTT method can be applied. Further, a learning method such as the above-mentioned LSTM method or GRU method may be used.

The items in each row of the training data 201 are sequentially input to this prediction model 204, the prediction value is calculated, and the BPTT method or the like is used so that the prediction value approaches the verification data (measured value of COD after 2.5 hours). Learning is performed by adjusting the parameters (weight, bias). Then, the prediction model 204 is updated by sequentially retraining while shifting the rows of the training data set.

In the example of FIG. 6, the case of learning COD is described. However, when learning the TN value, the measured value of TN after 2.5 hours may be learned as verification data. Alternatively, in the case of learning TP, the measured value of TP after 2.5 hours may be learned as verification data. Further, if the above three items (COD, TN, and TP) of the discharged water quality data are to be learned at once, the measured values of these three items may be learned as verification data.

Further, in the above example, the case where the predetermined time is "2.5 hours" is illustrated. However, the predetermined time is not limited to this, and may be any time.

Next, the model evaluation unit 103d uses the evaluation data and makes a prediction using the prediction model adjusted in step S5. The model evaluation unit 103d evaluates the predicted value and the measured value (step S6). Here, the model evaluation unit 103d calculates the difference (absolute value) between the predicted value and the measured value as an error, and saves the maximum error as an evaluation error.

Further, when the measured value is the non-stationary data, the model evaluation unit 103d determines whether or not the predicted value is correctly the non-stationary data. Then, the value obtained by dividing the number of points that can be accurately predicted as non-stationary data by the number of points of all non-stationary data is calculated as the predicted correct answer rate.

FIG. 7 is a diagram showing a case where the non-stationary data cannot be predicted correctly. FIG. 7 shows a case where the non-stationary data at one location includes the non-stationary data at four points. The model evaluation unit 103d determines that the answer is incorrect because all the predicted data are steady-state data. In addition, 〇 indicates a part that can be predicted correctly, and × indicates a part that cannot be predicted correctly.

FIG. 8 is a diagram showing a case where the non-stationary data can be correctly predicted. Similarly, FIG. 8 shows a case where the non-stationary data at one place includes the non-stationary data at four points. Since the model evaluation unit 103d predicts that one or more points of the prediction data are non-stationary data, it determines that the answer is correct.

In this way, the model evaluation unit 103d regards continuous non-stationary data as one non-stationary data, determines whether or not the predicted data is correct, and determines whether the predicted data is correct or not, and the non-stationary data that was the correct answer. Calculate the number of locations. Then, the value obtained by dividing the number of points of the non-stationary data that was the correct answer by the number of points of all the non-stationary data is defined as the predicted correct answer rate.

As the evaluation data, all the input items of the 24-hour learning data set may be used. However, input items for several hours may be extracted from the latest time in the training data set and used as evaluation data.

Next, the model evaluation unit 103d determines whether or not the evaluation error is equal to or less than the target error (step S7). The target error is a predetermined value, and is a value for determining whether or not the prediction model 204 is sufficiently trained.

If the evaluation error is larger than the target error (steps S7 and No), the process proceeds to step S9. If the evaluation error is equal to or less than the target error (S7, Yes), the model evaluation unit 103d determines whether or not the predicted correct answer rate is equal to or greater than the target correct answer rate (step S8). If the predicted correct answer rate is smaller than the target correct answer rate (S8, No), the process proceeds to step S9.

In step S9, the model evaluation unit 103d determines whether or not the number of updates of the prediction processing unit 105b is the second or more (step S9). If the number of updates of the prediction processing unit 105b is the second or more (S9, Yes), the model evaluation unit 103d determines whether or not the predetermined update time has been exceeded (step S10). This predetermined update time is a time for ending the process if the evaluation error does not become small even after a sufficient time has elapsed, and is set as, for example, a time of about several days.

If the predetermined update time is exceeded (step S10, Yes), the model evaluation unit 103d determines that the prediction model 204 that is less than or equal to the target error and is greater than or equal to the target correct answer rate with the current learning data cannot be created. Then, the process ends without updating the prediction model 204 (step S11).

Further, if the predetermined update time is not exceeded (steps S10 and No), the model generation unit 103c and the model evaluation unit 103d return to step S4 and repeat the processes of S4 to S8. Here, the model generation unit 103c performs the processing after step S3 after waiting for the first set time. This first set time is the time required for one learning, and is set as, for example, half a day or one day, although it varies depending on the amount of data.

By repeating steps S4 to S8, the model generation unit 103c and the model evaluation unit 103d sequentially create a prediction model 204 of hyperparameters of various combinations shown in FIG. 5, and use the created prediction model 204 to create a training data set. Learning will be performed sequentially, and the prediction model 204 will be updated sequentially.

If the predicted correct answer rate is equal to or higher than the target correct answer rate (step S8, Yes), the model evaluation unit 103d stores the predicted model 204 in the model change unit 103e (step S12). It can be said that this prediction model 204 is a prediction model that is close to the optimum among the prediction models generated by various combinations of hyperparameters.

Next, the model change unit 103e determines whether or not the effluent water quality prediction unit 105 is executing the prediction process (step S13). If the effluent water quality prediction unit 105 is executing the prediction process (steps S13, Yes), the model change unit 103e confirms the state of the effluent water quality prediction unit 105 every second set time. The second set time is the time required for the discharged water quality prediction unit 105 to carry out one prediction process, and is set as, for example, a time of about several minutes.

If the effluent quality prediction unit 105 is not executing the prediction process (steps S13, No), the model change unit 103e updates the prediction model 204 used by the prediction processing unit 105b to the latest prediction model 204 (step S14). , End the process.

Arbitrary values can be set for the predetermined update time, the first set time, and the second set time, such as changing the number of sensors installed in the sewage treatment plant and changing the amount of training data. Appropriate values shall be set according to.

FIG. 9 is a flowchart for explaining the processing procedure of the preprocessing unit 103a (step S2 in FIG. 4). This process removes abnormal data contained in the training data by filtering the training data set.

First, when there is a missing value in the training data, the preprocessing unit 103a interpolates the missing value by using the value of the same item in the training data of the previous and next times. Alternatively, the preprocessing unit 103a may interpolate the missing value by using the value of the same item at the same time on the same day of the previous week (S61). Then, the preprocessing unit 103a adds the time information data to the interpolated learning data (S62).

Next, the preprocessing unit 103a acquires the number N of the parameters of the filter to be optimized (S63). The number N of the parameters varies depending on the digital filter to be applied, but in the present embodiment, a Butterworth filter, which is one of the low-pass filters, is used. The four parameters of the Butterworth filter are the passband frequency F _p , its gain G _p , the blocking band frequency F _s , and its gain G _s (N = 4). Hereinafter, the case of the Butterworth filter will be described, but other digital filters may be used.

Next, the preprocessing unit 103a acquires the maximum value (Max _i ) and the minimum value (Min _i ) of each parameter of the filter to be optimized in advance (S64). Here, i = 1, 2, ..., N, and in the case of a Butterworth filter, N = 4.

Next, the preprocessing unit 103a sets N + 1 random combinations of numerical values within the range of the maximum value (Max _i ) and the minimum value (Min _i ) for each parameter of the filter to be optimized (S65). The combination of this random parameter is P ₁ = (F _p1 , G _p1 , F _s1 , G _s1 ), P ₂ = (F _p2 , G _p2 , F _s2 , G _s2 ), ..., P ₅ = (F _p5 ). , G _p5 , F _s5 , G _s5 ).

Next, the preprocessing unit 103a filters the training data set using the parameter set P _j (j = 1, 2, ..., N + 1) to create five training data sets. Then, the model generation unit 103c trains the prediction model created by using the hyperparameters using each of the five training data sets, and creates five prediction models (S66). The learning data set used first is, for example, training data for 24 hours, as described above. Further, the model generation unit 103c shall create a prediction model using the optimum hyperparameters created in advance.

Next, the model evaluation unit 103d inputs evaluation data for each of the five prediction models learned by the model generation unit 103c, and calculates the absolute value (error) of the difference between the calculated predicted value and the measured value. Calculated for each of the five prediction models. Then, the model evaluation unit 103d obtains the maximum value of the error for each of the five prediction models. Here, the function for obtaining the maximum value of the error of each prediction model is defined as the objective function f. Further, the measured value is, for example, the value of COD (TN, TP) after 2.5 hours, as described above.

The preprocessing unit 103a ranks a set of five parameters using the maximum error values f (P ₁ ), f (P ₂ ), ..., F (PN + 1) obtained by the model evaluation unit 103d (P _{N + 1} ). S67). Here, the maximum value of the error of each parameter set after being ordered is f (P ₁ ) ≤ f (P ₂ ) ≤ f (P ₃ ) ≤ f (P ₄ ) ≤ f (P ₅ ). Become. Assuming that m is the number of repetitions (1 ≦ m) when optimizing, the set of parameters after the m-th order is P ₁ ^(m) , P ₂ ^(m) , P ₃ ^(m). , P ₄ ^(m) , P ₅ ^(m) . The set of these parameters is called a point, and the point P ₁ ^(m) at this time is the best point B ^(m) , the point P ₄ ^(m) is the second worst point V ^(m) , and the point P ₅ ⁽ point P 5). ^{Let m} ) be called the worst point W ^(m) .

Finally, the pretreatment unit 103a optimizes each parameter using the best point B ^(m) , the second worst point V ^(m) , and the worst point W ^(m) (S68). In the present embodiment, for example, the Nelder-Mead method is used as the optimization method, but the method is not limited to this, and a steepest descent method, a genetic algorithm, or the like may be used.

FIG. 10 is a flowchart for explaining the processing procedure of optimizing each parameter of the filter (step S68 in FIG. 9). The initial value of m will be described as 0. First, the preprocessing unit 103a calculates m = m + 1, and obtains the center of gravity ^(m) using four points P _j ^(m) (j = 1, 2, 3, 4) other than the point W ^(m) . (S71).

Next, the preprocessing unit 103a obtains a point R ^(m) that externally divides the line segment W ^(m) G ^(m) by 2: 1 (S72).

R ^(m) = 2G ^(m) -W ^(m) ... (2)
Next, the preprocessing unit 103a creates a set of parameters corresponding to each of R ^(m) and B ^(m) . Then, the model generation unit 103c creates a prediction model corresponding to each of R ^(m) and B ^(m) . Then, the model evaluation unit 103d evaluates the prediction model corresponding to each of R ^(m) and B ^(m) , so that the maximum values f (R ^(m) ) and f (B ^(m) of the respective errors are evaluated. ) Is calculated. Hereinafter, this series of processes will be simply described as "calculating the value of the objective function f".

The preprocessing unit 103a determines whether or not f (R ^(m) ) is f (B ^(m) ) or less (S73). If f (R ^(m) ) is f (B ^(m) ) or less (S73, Yes), find the point E ^(m) that divides the line segment W ^(m) R ^(m) into 3: 2. (S74).

E ^(m) = 3G ^(m) -2W ^(m) ... (3)
Next, the value of f (E ^(m) ) is calculated, and the preprocessing unit 103a determines whether or not f (E ^(m) ) is f (R ^(m) ) or less (S75). If f (E ^(m) ) is f (R ^(m) ) or less (S75, Yes), the point W ^(m) is deleted and the point E ^(m) is added. Then, W ^(m ) is updated to V ^(m) (S76), and the process proceeds to step S84. If f (E ^(m) ) is larger than f (R ^(m) ) (S75, No), the process proceeds to step S77.

If f (R ^(m) ) is larger than f (B ^(m) ), the value of (S73, No), f (V ^(m) ) is calculated, and the preprocessing unit 103a determines f (R ⁽ m)). It is determined whether or not ^m) ) is f (V ^(m) ) or less (S78). If f (R ^(m) ) is f (V ^(m) ) or less (S78, Yes), the process proceeds to step S77.

In step S77, the pretreatment unit 103a deletes the point W ^{(m ) and adds the point R (m)} . Then, W ^(m ) is updated to V ^(m) , and the process proceeds to step S84.

Further, if f (R ^(m) ) is larger than f (V ^(m) ) (S78, No), the preprocessing unit 103a is the midpoint S ^(m ) of the line segment W ^(m) G ^(m ). (S79).

S ^(m) = 1 / 2G ^(m) + 1 / 2W ^(m) ... (4)
Then, the values of f (S ^(m) ) and f (W ^(m) ) are calculated, and the preprocessing unit 103a determines whether or not f (S ^(m) ) is f (W ^(m) ) or less. Is determined (S80). If f (S ^(m) ) is f (W ^(m) ) or less (S80, Yes), the preprocessing unit 103a deletes the point W ^(m ), adds the point S ^(m) , and steps. Processing proceeds to S83.

Further, if f (S ^(m) ) is larger than f (W ^(m) ) (S80, No), the preprocessing unit 103a is at the midpoint P _j ⁽ m) of each line segment B ^(m) P _j ^(m) . ^m)' is obtained (j = 2, 3, 4, 5), each point P _j ^(m) is updated to P _j ^(m)' (S82), and the process proceeds to step S83.

P _j ^(m)' = 1 / 2P _j ^(m) + 1 / 2B ^(m) ... (5)
In step S83, the preprocessing unit 103a ranks each point using the objective function f, and the process proceeds to step S84.

In step S84, the preprocessing unit 103a determines whether or not the setting determination formula is satisfied. As this setting determination formula, for example, whether the number of optimization repetitions m exceeds the preset upper limit q, whether the optimization time exceeds the preset upper limit, and the best point B ^(m) fluctuate. Whether or not the number of optimization repetitions that were not performed exceeds the preset upper limit value, whether or not f (B (m)) is below the preset threshold value, and whether or not the following equation (6) presets the threshold value. Whether or not it is less than the above, and so on.

If the setting determination formula is not satisfied (S84, No), the process returns to step S71 and the subsequent processing is repeated. Further, if the setting determination formula is satisfied (S84, Yes), the preprocessing unit 103a adopts the combination of the parameters of the point B ^(m) (S85). That is, the preprocessing unit 103a filters the first training data set using the combination of the parameters of the points B ^(m) , and uses the filtered training data set after step S3 in FIG. Processing will be performed.

If the non-stationary data change process (step S3 in FIG. 4) by the non-stationary data change unit 103b is unnecessary, the process of step S4 may be performed without performing the process of step S3.

FIG. 11 is a diagram schematically showing the optimization of each parameter of the filter. In FIG. 11, for the sake of simplicity, the case of N = 2 (two-dimensional) is shown. G11 indicates the best point B ⁽¹⁾ , the second worst point V ⁽¹⁾ , and the worst point W ⁽¹⁾ when m = 1.

G12 shows the case where m = 2, and the point added by the process shown in FIG. 10 is the best point B ⁽²⁾ , and the best point B ⁽¹⁾ shown in G11 is the second worst point V ^(2). The second worst point V ⁽¹⁾ shown in G11 is updated to the worst point W ⁽²⁾ , and the worst point W ⁽¹⁾ shown in G11 is deleted. Then, the same process shown in FIG. 10 is repeated.

G13 indicates the best point B ^(q-1) , the second worst point V ^(q-1) , and the worst point W ^(q-1) when m = q-1. q is the upper limit set value of m.

G14 indicates the best point B ^(q) , the second worst point V ^(q) , and the worst point W ^(q) when m = q. In this way, by repeating the process shown in FIG. 10 up to the upper limit set value q, the error is gradually reduced and the optimum parameter can be obtained.

FIG. 12 is a flowchart for explaining the processing procedure of the non-stationary data changing unit 103b (step S3 in FIG. 4). First, the non-stationary data changing unit 103b samples the learning data every minute after being processed by the preprocessing unit 103a, and creates learning data every 10 minutes, for example.

The non-stationary data change unit 103b extracts non-stationary data whose sensor data exceeds the alarm level from the data every 10 minutes. Then, the non-stationary data changing unit 103b obtains the total number of points of the extracted non-stationary data, and obtains the ratio P of the total number to all the data (step S31).

FIG. 13 is a diagram for explaining sampling of training data. In the following description, the case of non-stationary data of COD, which is one of the discharged water quality data, will be described, but the same applies to the non-stationary data of other discharged water quality data (TN, PN).

FIG. 13 shows the sensor data of COD every minute, and the non-stationary data changing unit 103b extracts the sensor data of the fifth point among the sensor data of ten points. In FIG. 13, only the non-stationary data of the COD exceeding the alarm level is shown, but the sensor data of the fifth point among the 10 sensor data is similarly extracted from the steady state data of the COD. And generate data every 10 minutes. In FIG. 13, the sensor data of COD every 1 minute is described as sampling data, and the sensor data of COD every 10 minutes is described as initial extraction data.

Here, the preset items used by the non-stationary data changing unit 103b to increase / decrease the non-stationary data will be described. _PMAX indicates the maximum percentage of non-stationary data, and P _MIN indicates the minimum percentage of non-stationary data. Further, μ indicates the adjustment coefficient of the increase / decrease range of the ratio of the non-stationary data, and N _MIN indicates the minimum increase / decrease range of the ratio of the non-stationary data.

Further, assuming that P0 ^(k) is the minimum ratio of the non-stationary data at the time of the kth selection and P3 ^(k) is the maximum ratio of the non-stationary data at the time of the kth selection, the kth non-stationary data. The increase / decrease range N ^(k) of the ratio of is μ × (P3 ^(k) −P0 ^(k) ).

Non-linear data using an optimization method (nonlinear programming) based on the idea that the optimal percentage of non-stationary data exists between the preset minimum percentage P _MIN and maximum percentage P _MAX . Find the optimal ratio of.

The non-stationary data change unit 103b substitutes 1 for the number of calculations k (step S32), substitutes _PMIN for the first minimum ratio P0 ^(k) , and _PMAX for the first maximum ratio P3 ^(k) . (Step S33), and the value of μ × (P3 ^(k) −P0 ^(k) ) is substituted into the increase / decrease range N ^(k) of the kth ratio (step S34).

Next, the non-stationary data changing unit 103b substitutes P0 ^(k) + N ^(k) for the ratio P1 ^(k ) and substitutes P3 ^(k) -N ^(k) for the ratio P2 ^(k) (step). S35). Then, the non-stationary data changing unit 103b creates a learning data set D1 by increasing the non-stationary data so that the ratio becomes P1 ^(k) , and the non-stationary data so that the ratio becomes P2 ^(k) . Is reduced to create the training data set D2 (steps S36 and S37). At this time, the non-stationary data changing unit 103b shall increase or decrease the non-stationary data based on the ratio P calculated in step S31 to create the learning data sets D1 and D2.

FIG. 14 is a diagram for explaining a method of increasing the number of non-stationary data. As shown in FIG. 14, since the sampling data around the initial extraction data often includes non-stationary data, additional non-stationary data can be extracted. FIG. 14 shows an example in which the non-stationary data is tripled by adding the non-stationary data at the second point and the non-stationary data at the eighth point among the ten points.

FIG. 15 is a diagram for explaining another method for increasing the number of non-stationary data. As shown in FIG. 15, non-stationary data whose amplitude is increased by 1.05 times with respect to the amplitude of the initial extracted data is created, and the non-stationary data is added to the training data. In addition, non-stationary data whose amplitude is reduced by 0.95 times with respect to the amplitude of the initial extracted data is created, and the non-stationary data is added to the training data. In this way, the number of non-stationary data can be tripled.

FIG. 16 is a diagram showing initial extraction data of non-stationary data. In FIG. 16, a case where the sampling interval of the initial extraction data is 1 minute will be described, but the sampling interval may be longer than 1 minute.

FIG. 17 is a diagram for explaining a method of reducing the number of non-stationary data. In FIG. 17, the sampling interval is set to 5 minutes, and the second point of the five non-stationary time data (initial extraction data) is extracted. As a result, the number of non-stationary data can be reduced to 0.2 times the initial extracted data.

For example, if the number of initial extracted data is 180,000 and the number of non-stationary data is 6,000, the number of non-stationary data is used to correctly predict the non-stationary data. Need to increase. If the number of non-stationary data is added by 20,000 using the above-mentioned method of increasing non-stationary data, the total number of added data will be 200,000, and the added non-stationary data will be 26. It will be a thousand points. As a result, the ratio of the non-stationary data becomes 13%, and the non-stationary data can be predicted correctly.

Next, the model generation unit 103c creates a prediction model M1 using the training data set D1 created by the non-stationary data change unit 103b, and creates a prediction model M2 using the training data set D2 (step S38). ). Here, the model generation unit 103c shall create a prediction model using hyperparameters considered to be the optimum combination among the hyperparameter adjustment items shown in FIG. 5, and predict by learning and verifying the prediction model. Update models M1 and M2.

Next, the model evaluation unit 103d uses the evaluation data to make predictions using the prediction models M1 and M2 created in step S38. The model evaluation unit 103d evaluates the predicted values and the actually measured values of the predicted models M1 and M2 (step S39). Here, the model evaluation unit 103d calculates the difference between the predicted value and the measured value of the predicted models M1 and M2 as an error, and saves the maximum error as an evaluation error.

Next, the non-stationary data changing unit 103b determines whether or not the evaluation error of the prediction model M1 is equal to or less than the evaluation error of the prediction model M2 (step S40). If the evaluation error of the prediction model M1 is equal to or less than the evaluation error of the prediction model M2 (S40, Yes), the non-stationary data change unit 103b temporarily stores the training data set D1 (step S41).

Next, the non-stationary data changing unit 103b substitutes P0 ^(k) for the minimum ratio P0 ^{(k + 1)} at the k + 1st time, and substitutes P2 ^(k) for the maximum ratio P3 ^{(k + 1)} at the k + 1th time (step S42). ). Then, P1 ^(k) is substituted into the k + 1th ratio P1 ^{(k + 1)} , and the calculated value of P2 ^(k) −N ^(k ) is substituted into the k + 1th ratio P2 ^{(k + 1)} (step S43).

If the evaluation error of the prediction model M1 is larger than the evaluation error of the prediction model M2 (S40, No), the non-stationary data change unit 103b temporarily stores the training data set D2 (step S44).

Next, the non-stationary data changing unit 103b substitutes P1 ^(k) for the minimum ratio P0 ^{(k + 1)} at the k + 1st time, and substitutes P3 ^(k) for the maximum ratio P3 ^{(k + 1)} at the k + 1th time (step S45). ). Then, the calculated value of P1 ^(k) + N ^(k) is substituted into the k + 1st ratio P1 ^{(k + 1)} , and P2 ( ^k ) is substituted into the k + 1st ratio P2 ^{(k + 1)} (step S46).

Next, the non-stationary data changing unit 103b substitutes the calculated value of μ × (P3 ^{(k + 1)} −P0 ^{(k + 1)} ) into N ⁽ k + 1) and updates it (step S47), and increases / decreases the range N ^{(k + 1).} Is determined whether or not is equal to or less than the minimum increase / decrease width _NMIN (step S48).

If the increase / decrease width N ^{(k + 1)} is larger than the minimum increase / decrease width N _MIN (S48, No), the non-stationary data change unit 103b substitutes the value of k + 1 for the number of calculations k (step S49), and in step S36. After returning, the subsequent processing is repeated. Further, if the increase / decrease width N ^{(k + 1)} is equal to or less than the minimum increase / decrease width N _MIN (S48, Yes), the non-stationary data change unit 103b adopts the learning data set primaryly stored (step S50). The process proceeds to step S4 in FIG.

FIG. 18 is a diagram schematically showing the processing of the non-stationary data changing unit 103b (step S3 in FIG. 4) shown in FIG. In FIG. 18, the horizontal axis represents the ratio of non-stationary data, and the vertical axis represents the evaluation error E in a graph format.

The selection range of the optimum ratio for the first time is from the minimum ratio P0 ⁽¹⁾ ( _PMIN ) to the maximum ratio P3 ⁽¹⁾ ( _PMAX ), and the first increase / decrease range N ⁽¹⁾ is μ × (P3 ⁽ P3). ¹⁾ Substitute the calculated value of -P0 ⁽¹⁾ ), and let the ratio P1 ⁽¹⁾ be P0 ⁽¹⁾ + N ⁽¹⁾ , and the ratio P2 ⁽¹⁾ be P3 ⁽¹⁾ -N ⁽¹⁾ . At this time, since the evaluation error E2 of the prediction model M2 is larger than the evaluation error E1 of the prediction model M1, the ratio P2 ⁽¹⁾ is substituted for the maximum ratio P3 ⁽²⁾ , and the ratio P2 ⁽²⁾ is P2 ⁽¹ ). ⁾ -N Substitute the calculated value of ⁽¹⁾ . As a result, the maximum ratio P3 ⁽²⁾ and the ratio P2 ⁽²⁾ move to the center (left side) by N ⁽¹⁾ (G1).

Next, the selection range of the optimum ratio for the second time is set from the minimum ratio P0 ⁽²⁾ to the maximum ratio P3 ⁽²⁾ , and the second increase / decrease range N ⁽²⁾ is μ × (P3 ⁽²⁾ -P0 ⁽ . ²⁾ Substitute the calculated value in). At this time, since the evaluation error E2 of the prediction model M2 is smaller than the evaluation error E1 of the prediction model M1, the ratio P1 ⁽²⁾ is substituted into the minimum ratio P0 ⁽³⁾ , and the ratio P1 ⁽³⁾ is P1 ⁽² ). ⁾ + N Substitute the calculated value of ⁽²⁾ . As a result, the minimum ratio P0 ⁽³⁾ and the ratio P1 ⁽³⁾ move to the center (right side) by N ⁽²⁾ (G2).

Next, the selection range of the optimum ratio for the third time is set from the minimum ratio P0 ⁽³⁾ to the maximum ratio P3 ⁽³⁾ , and the increase / decrease range N ⁽³⁾ for the third time is μ × (P3 ⁽³⁾ −P0 ⁽ . ³⁾ Substitute the calculated value in). At this time, since the evaluation error E2 of the prediction model M2 is larger than the evaluation error E1 of the prediction model M1, the ratio P2 ⁽³⁾ is substituted for the maximum ratio P3 ⁽⁴⁾ , and the ratio P2 ⁽⁴⁾ is P2 ⁽³ ). ⁾ -N Substitute the calculated value of ⁽³⁾ . As a result, the maximum ratio P3 ⁽⁴⁾ and the ratio P2 ⁽⁴⁾ move to the center (left side) by N ⁽³⁾ (G3).

The same process is repeated, the selection range of the optimum ratio for the T time is set from the minimum ratio P0 ^(T) to the maximum ratio P3 ^(T) , and the increase / decrease range N ^(T) for the T time is μ × (P3 ^(T)). -Substitute the calculated value of P0 ^(T) ). At this time, since the increase / decrease width N ^(T) of the Tth time is smaller than the minimum increase / decrease width N _MIN , the primary stored learning data set is adopted.

(Treatment procedure of effluent quality prediction unit 105)
FIG. 19 is a flowchart for explaining the processing procedure of the discharged water quality prediction unit 105. First, the preprocessing unit 105a acquires the latest data stored in the latest data database 104 at predetermined prediction cycles (step S21). The preprocessing unit 103a performs preprocessing of learning data. More specifically, when there is a missing value or an abnormal value in the training data, the preprocessing unit 103a interpolates the missing value or removes the abnormal value (step S22). The missing values, the abnormal values, the interpolation method and the removal method of those values are the same as those described in the preprocessing unit 103a of the prediction model generation unit 103.

The process shown in FIG. 19 is performed again at predetermined prediction cycles. This predetermined prediction cycle is the time required for the discharged water quality prediction unit 105 to carry out one prediction process, and is set as, for example, a time of about several minutes. Any value can be set for this predetermined prediction cycle, and an appropriate value shall be set according to a change in the amount of training data or the like.

Next, the prediction processing unit 105b applies the latest data after preprocessing by the pretreatment unit 105a to the prediction model 204 created by the prediction model generation unit 103, and the discharged water quality data after 2.5 hours. Predict the value of (COD, TN, TP) (step S23).

FIG. 20 is a diagram schematically showing the prediction of the discharged water quality data. As the latest data 205, 23 items (5 items of time information data + 15 items of sensor information data + 3 items of discharged water quality data) of each row of the training data set are input to the input layer of the prediction model 204. The prediction model 204 predicts the value of COD after 2.5 hours and outputs it as output data 206.

Similarly, it is also possible to predict the value of TN after 2.5 hours by using the prediction model 204 created for the prediction of TN. It is also possible to predict the TP after 2.5 hours by using the prediction model 204 created for the prediction of the TP.

Finally, the post-treatment unit 105c performs post-treatment on the discharged water quality data (prediction result) predicted by the prediction processing unit 105b. Specifically, the post-processing unit 105c converts the prediction result into a format that can be output by the output unit 106 (step S24), and ends the processing.

(Example of prediction result)
The quality of discharged water was predicted at a sewage treatment plant. The number of data was 1 minute data (about 2.45 million data) for 4 years and 8 months, the input data type (including time information data) was 20 items, and the output data type was 3 items (COD, TN and TP). .. The learning data set was 23 items (input data type + output data type) x 12 hours. The LSTM method was used as the machine learning method.

FIG. 21 is a graph showing the measured value of COD and the predicted value of COD predicted by the effluent quality prediction unit 105 when the number of non-stationary data is not increased or decreased. The vertical axis shows the value of COD, and the horizontal axis shows the time. This graph shows the measured and predicted values of COD for one day, and shows that the non-stationary data cannot be predicted correctly.

FIG. 22 is a graph showing the measured value of COD when the number of non-stationary data is increased or decreased and the predicted value of COD predicted by the discharged water quality prediction unit 105. This graph shows that the non-stationary data can be predicted correctly.

(effect)
As described above, according to the discharged water quality prediction device 100 according to the present embodiment, the pretreatment unit 103a uses a digital filter to remove abnormal data and noise included in the learning data, so that the quality is high. It was possible to generate training data and improve the prediction accuracy.

In addition, since RNN was used to predict the discharged water quality data and learning methods such as the BPTT method, LSTM method, and GRU method were applied, it became possible to predict the discharged water quality data with one prediction model.

In addition, since RNN is used to predict the discharged water quality data, even if the operation of the sewage treatment plant is changed, it can be dealt with simply by changing the hyperparameters and creating a prediction model, which reduces the prediction accuracy. It became possible to prevent it.

Further, since the preprocessing unit 103a optimizes the parameters of the digital filter, it is possible to generate higher quality training data and improve the prediction accuracy.

In addition, the model generation unit 103c creates a prediction model using various hyperparameters, and predicts the discharged water quality data using the prediction model with less evaluation error, so that the discharged water quality data is accurate. It became possible to predict.

[Example of implementation by software]
The control block (particularly the sensor information acquisition unit 101, the prediction model generation unit 103, the discharge water quality prediction unit 105 and the output unit 106) of the discharge water quality prediction device 100 is a logic circuit (hardware) formed in an integrated circuit (IC chip) or the like. ) Or by software.

In the latter case, the effluent water quality prediction device 100 includes a computer that executes instructions of a program that is software that realizes each function. The computer includes, for example, one or more processors and a computer-readable recording medium that stores the program. Then, in the computer, the processor reads the program from the recording medium and executes the program, thereby achieving the object of the present invention. As the processor, for example, a CPU (Central Processing Unit) can be used. As the recording medium, a "non-temporary tangible medium", for example, a ROM (Read Only Memory) or the like, a tape, a disk, a card, a semiconductor memory, a programmable logic circuit, or the like can be used. Further, a RAM (Random Access Memory) for expanding the above program may be further provided. Further, the program may be supplied to the computer via any transmission medium (communication network, broadcast wave, etc.) capable of transmitting the program. It should be noted that one aspect of the present invention can also be realized in the form of a data signal embedded in a carrier wave, in which the above program is embodied by electronic transmission.

[Computer configuration example]
FIG. 23 is a diagram showing a configuration example of a computer for realizing a effluent water quality prediction device (eg, effluent water quality prediction device 100) according to one aspect of the present invention. The computer includes a computer body 300, a display 401, a keyboard 402, and a mouse 403. A program for allowing the computer body 300 to function as a effluent water quality predictor by operating the keyboard 402 and the mouse 403 while looking at the screen displayed on the display 401 (hereinafter referred to as a effluent water quality prediction program). To start.

The computer body 300 includes a CPU 301, a GPGPU (General-Purpose computing on Graphics Processing Units) 302, a ROM 303 for storing a boot program, a RAM 304 for storing a program to be executed, work data, etc., a hard disk 305, and a DVD ( It includes a Digital Versatile Disc) drive 306, a network I / F (Interface) 307, a memory port 308, and an RTC (Real Time Clock) 309, each of which is connected to an internal bus 310.

The GPU PPU 302 is an application of GPU computing resources for purposes other than image processing, and has high processing performance, so it has a wide range of fields such as matrix operations such as neural networks, optimization problems, cryptanalysis, and voice processing. Used in.

The discharged water quality prediction program is recorded on a recording medium such as a DVD 406 and a removable memory 405, and is installed on the hard disk 305 via the DVD drive 306 or the memory port 308 under the control of the CPU 301. Further, the discharged water quality prediction program may be downloaded via the public line 404 and the network I / F 307 and installed on the hard disk 305.

The CPU 301 executes a discharge water quality program stored in the RAM 304, receives sensor information data and discharge water quality data from a plurality of sensors installed in the sewage treatment plant via the network I / F 307 and the public line 404, and receives them. The sensor information acquisition unit 101 is realized by adding the time information data acquired from the RTC 309 to the data of the above and generating the learning data.

Further, the CPU 301 executes the discharged water quality program stored in the RAM 304, and sequentially stores the learning data generated by the sensor information acquisition unit 101 in the hard disk 305 or the like, thereby realizing the learning data database 102 and the latest data database 104. do.

Further, the CPU 301 executes the effluent water quality program stored in the RAM 304, learns the RNN together with the GPGPU 302, and generates the prediction model 204, thereby realizing the prediction model generation unit 103.

Similarly, the CPU 301 executes the effluent water quality prediction program stored in the RAM 304, performs an operation of applying the latest data 205 to the prediction model 204 together with the GPGPU 302, and generates the prediction data, thereby realizing the effluent water quality prediction unit 105. do.

Further, the CPU 301 executes a effluent water quality prediction program stored in the RAM 304, and controls various information generated by the effluent water quality prediction unit 105 via the network I / F 307 and the public line 404, which is not shown in the sewage treatment plant. The output unit 106 is realized by transmitting to a device, a display device, or the like.

〔summary〕
The discharged water quality prediction device according to the first aspect of the present invention has a sensor information acquisition unit that acquires time-series data of detected values from a plurality of sensors installed in a sewage treatment plant, and a time-series acquired by the sensor information acquisition unit. A training data database that stores data as training data, a recent data database that stores time-series data acquired by the sensor information acquisition unit as the latest data, and training data for the first predetermined time in the training data are learned. Learning after the pre-processing unit filters the pre-processing unit that forms the data set and filters the training data set, and the prediction model for predicting the value of the discharged water quality data after the second predetermined time. It is provided with a model generation unit generated from a data set and a discharge water quality prediction unit that predicts the value of the discharge water quality data by applying a prediction model to the latest data.

In the discharge water quality prediction device according to the second aspect of the present invention, in the above aspect 1, the model generation unit uses the training data set as training data and the measured value of the discharged water quality data after the second predetermined time as verification data. It may be configured to generate a predictive model by training a recurrent neural network using training data and verification data.

In the discharge water quality prediction device according to the third aspect of the present invention, in the above aspect 2, the model generation unit generates a prediction model while changing the hyper parameters of the recurrent neural network, and the discharge water quality prediction device further obtains a training data set. Model evaluation in which at least a part of the training data is used as evaluation data, and a prediction model in which the error between the prediction value obtained by inputting the evaluation data into the prediction model and the verification data is less than a predetermined value is set in the discharged water quality prediction unit. It may have a structure including a part.

In the discharge water quality prediction device according to the fourth aspect of the present invention, in the third aspect, the pretreatment unit generates a set of a plurality of parameters used in the filtering process and filters the set of the plurality of parameters into the training data set. Processing is performed to create multiple training data sets, the model generation unit generates multiple prediction models using the multiple training data sets, and the model evaluation unit inputs the evaluation data into the multiple prediction models. The maximum error between the predicted value obtained by It may be provided with a configuration for extracting the optimum set of parameters for filtering by performing the above.

The discharged water quality prediction method according to the fifth aspect of the present invention includes a step of acquiring time-series data of detected values from a plurality of sensors installed in a sewage treatment plant and a step of storing the acquired time-series data as training data. A step of storing the acquired time-series data as the latest data, a step of using the training data for the first predetermined time in the training data as the training data set, and a step of filtering the training data set, and a second step. By applying the prediction model to the step of generating a prediction model for predicting the value of the discharged water quality data after a predetermined time from the training data set after filtering, and to the latest data, the discharged water quality data. Includes a step to predict the value of.

[Additional notes]
The present invention is not limited to the above-described embodiments, and various modifications can be made within the scope of the claims, and the embodiments obtained by appropriately combining the technical means disclosed in the different embodiments. Is also included in the technical scope of the present invention.

100 Discharge water quality prediction device 101 Sensor information acquisition unit 102 Learning data database 103 Prediction model generation unit 103a Pretreatment unit 103b Non-stationary data change unit 103c Model generation unit 103d Model evaluation unit 103e Model change unit 104 Latest data database 105 Discharge water quality prediction Part 201 Training data 204 Prediction model 205 Recent data

Claims (5)

A sensor information acquisition unit that acquires time-series data of detected values from multiple sensors installed in the sewage treatment plant,
A learning data database that stores the time-series data acquired by the sensor information acquisition unit as learning data, and
A latest data database that stores the time-series data acquired by the sensor information acquisition unit as the latest data, and
A pre-processing unit that uses the training data for the first predetermined time in the training data as a training data set and filters the training data set, and
A model generation unit that generates a prediction model for predicting the value of the discharged water quality data after the second predetermined time from the training data set after being filtered by the pretreatment unit.
A effluent water quality prediction device including a effluent water quality prediction unit that predicts the value of the effluent water quality data by applying the prediction model to the latest data. The model generator
The training data set is used as training data.
The measured value of the discharged water quality data after the second predetermined time is used as the verification data.
The effluent water quality prediction device according to claim 1, wherein a prediction model is generated by training a recurrent neural network using the training data and the verification data. The model generation unit generates a prediction model while changing the hyperparameters of the recurrent neural network.
Further, the discharged water quality prediction device uses at least a part of the training data of the training data set as evaluation data, and an error between the prediction value obtained by inputting the evaluation data into the prediction model and the verification data is predetermined. The discharge water quality prediction device according to claim 2, further comprising a model evaluation unit that sets a prediction model having a value or less in the discharge water quality prediction unit. The preprocessing unit generates a set of a plurality of parameters used in the filtering process, filters the training data set using the set of the plurality of parameters, and creates a plurality of training data sets.
The model generation unit generates a plurality of prediction models using the plurality of training data sets, and generates a plurality of prediction models.
The model evaluation unit calculates the maximum error between the predicted value obtained by inputting the evaluation data into the plurality of prediction models and the verification data for each of the plurality of prediction models.
According to claim 3, the preprocessing unit ranks the plurality of parameters according to the maximum error calculated for each of the plurality of prediction models, and performs optimization to extract the optimum set of parameters for the filtering process. The described effluent quality prediction device. Steps to acquire time-series data of detected values from multiple sensors installed in sewage treatment plants,
A step of storing the acquired time-series data as learning data,
The step of storing the acquired time-series data as the latest data, and
A step of using the training data for the first predetermined time in the training data as a training data set and filtering the training data set, and
A step of generating a prediction model for predicting the value of the discharged water quality data after the second predetermined time from the filtered training data set, and a step of generating a prediction model.
A method for predicting the quality of discharged water, which comprises a step of predicting the value of the discharged water quality data by applying the prediction model to the latest data.

Images