Water quality control indicator output device and water quality control indicator output method

Abstract

To provide a method that enables quantitative evaluation of water quality influencing factors and derivation of operating indicators for facilities to optimize water quality, which are necessary for optimal water quality management.SOLUTION: A water quality control indicator output device 10 includes a data storage part 13 that stores water quality-related data, a machine learning model estimation part 17 that outputs a trained model from the water quality-related data, a trained model storage part 18 that stores the trained model, a water quality effect quantity inference value output part 20 that infers and outputs the water quality effect quantity inference value from the water quality-related data and the trained model, a water quality effect quantity inference value storage part 21 that stores the water quality effect quantity inference value, a water quality optimization operation indicator output part 24 that infers and outputs a water quality optimization operation indicator from the water quality-related data and the trained model, and a water quality optimization operation indicator storage part 26 that stores the water quality optimization operation indicator.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to a water quality management index output device and a water quality management index output method. [Background technology]

[0002] Sewage treatment has become increasingly common using the standard activated sludge process, which uses activated sludge, a sludge made up of flocculated microorganisms, to mainly remove organic matter through an aerobic biological reaction. In this process, a process called aeration is carried out, in which oxygen is added to the mixture of activated sludge and sewage to promote the functioning of the activated sludge. Water quality management in this type of standard activated sludge process mainly involves changing the amount of activated sludge and the aeration process in response to increases or decreases in the amount of organic matter in the inflowing sewage.

[0003] However, as sewerage systems became more widespread, it became necessary to address environmental issues such as eutrophication, and treatment methods for removing nitrogen and phosphorus, in other words advanced treatment, became more widespread. The standard activated sludge process consisted only of an aerobic tank, which provided oxygen to the activated sludge through aeration, but advanced treatment introduced tanks called anaerobic and anoxic tanks, which purposely did not provide oxygen to the activated sludge, allowing for the excessive accumulation of phosphorus by organisms and special biological reactions known as denitrification, making it possible to remove nitrogen and phosphorus from sewage.

[0004] Therefore, in the current water quality management of sewage, in addition to the conventional organic matter treatment, it is also required to properly maintain the advanced treatment function. Since it is necessary to properly maintain not only the aerobic tank but also the anaerobic tank and anoxic tank at the same time, the operation for water quality management has become complicated. Therefore, in order to perform optimal water quality management that takes into account the various factors that affect water quality, a lot of experience and know-how has come to be required.

[0005] For example, when removing nitrogen through advanced treatment, the aerobic tank must not only remove organic matter, but also promote a reaction called nitrification, which oxidizes ammonia nitrogen in the sewage to nitrate nitrogen. Both can be promoted by increasing the amount of aeration, but nitrogen removal requires biological treatment in an anoxic tank. This is done by reducing the nitrate nitrogen produced in the aerobic tank to N2 and other substances through a biological reaction called denitrification. However, if oxygen remains in the anoxic tank, the aerobic biological reaction takes precedence and the denitrification reaction does not proceed. Therefore, if the amount of aeration is excessively increased in the aerobic tank, denitrification will not proceed in the anoxic tank, and if the amount of aeration is insufficient, nitrification will be insufficient, and both will hinder nitrogen removal. Furthermore, the amount and quality of inflowing sewage water vary greatly depending on the season, weather, and human activities, and optimal water quality management cannot be achieved unless the optimal operation is changed according to each situation.

[0006] Thus, with the spread of advanced treatment, water quality management has become more difficult and important, but at present, the optimal method of water quality management is largely dependent on experience, know-how, and personal intuition. However, due to the recent decline in engineers and the mass retirement of experienced personnel, it is expected that water quality management techniques will become more difficult to maintain. Therefore, there is a need for a method to obtain water quality management indicators based on data, without relying on experience, and to carry out optimal water quality management based on these indicators.

[0007] Prior art also presents a method of utilizing data through machine learning in sewage treatment and similar wastewater treatment. In Non-Patent Document 1, while considering the potential of various pieces of information held by sewers, a prediction of the quality of treated sewage water is performed through machine learning using operation management data and influent water quality data acquired daily. Calculations are performed using a deep learning model and random forest using neural networks, but sufficient accuracy cannot be obtained in verification against unknown data, and there is room for improvement.

[0008] In Non-Patent Document 2, machine learning is carried out to predict nitrate nitrogen, which is important in swine wastewater treatment. Operational management items that can be easily collected by sensors (sewage pH, sewage EC, aeration tank water temperature, aeration tank pH, and activated sludge suspended solids in the aeration tank) are used as learning data, a decision tree model is estimated, and the node branches of this decision tree are analyzed to evaluate the impact of the operational management items and create operational management guidelines. [Prior art documents] [Non-patent literature]

[0009] [Non-Patent Document 1] Nishimura, F. et al. “Sewerage as an information base for urban and river basin areas: Attempts to utilize operation and management data.” AI and Data Science Papers 1.J1 (2020):278-285. [Non-Patent Document 2] Yasuo Tanaka. "Towards improving the performance of swine wastewater treatment facilities using machine learning - Attempts to understand the nitrate nitrogen removal characteristics of individual facilities using decision tree analysis." Journal of the Japanese Society of Swine Science 58.2 (2021): 65-79. Summary of the Invention [Problem to be solved by the invention]

[0010] In Non-Patent Document 1, a machine learning model is estimated to predict the water quality of treated water based on the operation management data and water quality data of the same sewage treatment plant as the present invention, but sufficient accuracy is not achieved, and it does not reach the point of quantitatively evaluating water quality impact factors such as water quality impact amounts.

[0011] In Non-Patent Document 2, a machine learning model is estimated to predict the water quality of treated water and an impact assessment is performed based on operation and management items. However, in decision tree analysis, it is necessary to write out a decision tree for each analysis and interpret the nodes, and there is a problem that it is difficult to handle a large number of water quality impact factors.

[0012] The problem to be solved by the present invention is to provide an apparatus and method for quantitatively evaluating factors affecting water quality and deriving facility operation indicators for optimizing water quality, which are necessary for optimal water quality management. [Means for solving the problem]

[0013] In order to solve the above problems, the present invention employs the following means: That is, a water quality management index output device is provided that infers and outputs a water quality impact amount inference value and a water quality optimization operation index used for water quality management, the water quality management index output device including a data accumulation unit that accumulates water quality-related data, a machine learning model estimation unit that outputs a trained model from the water quality-related data, a trained model storage unit that stores the trained model, a water quality impact amount inference value output unit that infers and outputs a water quality impact amount inference value from the water quality-related data and the trained model, a water quality impact amount inference value storage unit that stores the water quality impact amount inference value, and a water quality optimization operation index output unit that infers and outputs a water quality optimization operation index from the water quality impact amount inference value. Effect of the Invention

[0014] It is possible to quantitatively evaluate the factors affecting water quality, which is necessary for optimal water quality management, and to derive operational indicators for facilities to optimize water quality. [Brief description of the drawings]

[0015] [Figure 1] 1 is a block diagram of a sewage treatment facility according to an embodiment of the present invention. [Diagram 2] FIG. 2 is a block diagram of a water quality management index output device. [Diagram 3] 13 is a flowchart of an optimum learning method output section of the water quality management index output device. [Figure 4] 13 is a flowchart of a machine learning model estimation unit of the water quality management index output device. [Diagram 5] 13 is a flowchart of a water quality influence amount inference value output unit of the water quality management index output device. [Figure 6]13 is a flowchart of a water quality optimization operation index output unit of the water quality management index output device. [Figure 7] 13 is an example of explanatory variable data of optimal training data created from integrated data. [Figure 8] 13 is an example of objective variable data of optimal training data created from integrated data. [Figure 9] This is a scatter plot showing the correlation between the objective variable data of the optimal teacher data (actual data in the figure) and the learned model predicted values ​​(predicted data in the figure) for both test data and training data. [Figure 10] This is an example of the estimated value of water quality impact amount for optimal training data. [Figure 11] This is an example of visualization of estimated values ​​of water quality impact amounts. [Figure 12] This is an example of visualization of the change in the inferred value of water quality impact amount. [Figure 13] This is an example of a visualization of the distribution of inferred values ​​of water quality impact amounts for each explanatory variable. [Figure 14] This is an example of a visualization of the relationship between the values ​​of the explanatory variables of interest and their average water quality impact inferred values. [Figure 15] This is an example of visualizing the relationship between the value of the explanatory variable of interest, its inferred water quality impact amount, and the value of the explanatory variable for which dependency is to be examined. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0016] Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings.

[0017] FIG. 1 is a block diagram of a sewage treatment facility in this embodiment. The sewage treatment facility 1 includes a grit tank 2, a lifting pump equipment 3, a primary sedimentation tank 4, a biological reaction tank 5, a final sedimentation tank 6, a disinfection equipment 7, a blower equipment 8, a sludge treatment equipment 9, and a water quality management index output device 10. As an example of a biological reaction tank, the structure of a step inflow type two-stage nitrification denitrification method, which is an advanced treatment for nitrogen removal, is shown. Flow meters and the like are installed in each process, and a DO (dissolved oxygen concentration) meter, a MLSS (activated sludge concentration) meter, an ORP (oxidation-reduction potential) meter, an NH4-N (ammonia nitrogen concentration) meter, and the like are installed in the biological reaction tank, and trend data of these instrumentation equipment is collected by a central monitoring system. In addition, samples of the main processes for measuring water quality and the like include inflow sewage 2w, raw water 3w, sedimentation water 4w, treated water 5w, discharge water 6w, reaction tank mixed liquid 2s, return sludge 3s, and excess sludge 4s. Samples are taken from each of these processes on a regular basis, and water quality tests are conducted for statutory inspections and water quality management.

[0018] Sewage that flows into the sewage treatment facility 1 through the sewer system, inflow sewage 2w, flows into the grit basin 2 together with on-site return water 1w generated from the sludge treatment facility 9 of the sewage treatment facility 1. In the grit basin 2, the sewage that has flowed through the sewer pipes is introduced into the grit basin 2, which has a larger cross-sectional area, greatly reducing the flow rate and removing the sand contained in the sewage. This prevents sandblasting and other phenomena at the lift pump installed in the lift pump facility 3 at the subsequent stage. In addition, a screen-based screen-based sediment removal device is also installed in conjunction with the grit basin. This device is also intended to prevent sediment from getting tangled in the lift pump.

[0019] Sewers utilize gravity flow and often collect sewage at a sewage treatment facility 1, and the installation level of the sewer pipes is often deep underground at the stage where the sewage flows into the sewage treatment facility 1. Also, because the sewage treatment facility 1 requires a large area, it is installed at ground level. This makes it necessary to use a lifting pump to lift the sewage to ground level. The lifting pump facility 3 is provided for this purpose.

[0020] The pumped sewage is introduced into the primary sedimentation tank 4, which has an even larger cross-sectional area than the gritification tank 2. Because the flow rate is even slower than in the gritification tank 2, particulate organic matter dispersed in the sewage is allowed to settle and be separated from the sewage. The sludge after settling and separation is called raw sludge 1s, and is sent to the sludge treatment facility 9. The sample before entering the primary sedimentation tank 4 is called raw water 3w, and the sewage treated in the primary sedimentation tank 4 is called settling water 4w.

[0021] The settling water 4w is introduced into the biological reactor 5 together with the return sludge 2s. In the biological reactor 5, fine pollutant particles contained in the settling water 4w and pollutants dissolved in the sewage are removed by biological reactions of activated sludge microorganisms. In the step-inflow two-stage nitrification and denitrification method shown as an example, the settling water 4w is divided into a fixed ratio and fed into each anoxic tank. A mixed state is maintained in the anoxic tank using a mechanical stirrer or other device. Aeration is performed in the aerobic tank by supplying air from the blower equipment 8 through an aeration device installed at the bottom of the reaction tank or elsewhere. The amount of aeration is controlled by a DO control that refers to a DO meter installed at the end of the biological reactor 5 and controls the amount of aeration to maintain a constant DO, or by an aeration ratio control that controls the amount of aeration with a volumetric ratio of air determined based on the amount of inflow sewage.

[0022] In the aerobic tank of the step-inflow two-stage nitrification-denitrification process, organic matter contained in the settling water 4w is treated and ammonia nitrogen is nitrified, while in the anoxic tank, nitrate nitrogen contained in the returned sludge 2s from the previous aerobic tank or the final settling tank 6 is denitrified. The anoxic state is maintained by the consumption of dissolved oxygen in the process in which the organic matter contained in the settling water 4w is utilized by living organisms, so the settling water 4w is added in portions. An ORP meter is installed to monitor this anoxic state.

[0023] The sewage treated in the biological reactor 5 is introduced into the final settling tank 6. In the final settling tank 6, the activated sludge is allowed to settle and separated, yielding treated water 5w after biological treatment has been completed. The settled and separated activated sludge is returned to the biological reactor 5 as return sludge 2s and is used repeatedly for sewage treatment. Furthermore, as microorganisms multiply during this repeated sewage treatment process, the amount of activated sludge increases. This increased amount of activated sludge is withdrawn as excess sludge 3s and sent to sludge treatment equipment 9. Furthermore, the MLSS (MLSS concentration of activated sludge) in the biological reactor 5 can be manipulated by controlling the amount of excess sludge 3s.

[0024] After the biological treatment, the treated water 5w is sent to a disinfection facility 7. Here, the treated water 5w is disinfected with sodium hypochlorite, ozone, etc., and is discharged into rivers or oceans as effluent water 6w. The raw sludge 1s and excess sludge 3s sent to a sludge treatment facility 9 are incinerated after being thickened, dehydrated, and otherwise processed.

[0025] 2 is a block diagram of the water quality management index output device 10 in this embodiment. The water quality management index output device 10 in this embodiment is targeted at the sewage treatment facility 1 as described above, and infers and outputs a water quality impact amount inference value 22 and a water quality optimization operation index 26 used for water quality management of the sewage treatment facility 1.

[0026] The water quality management index output device 10 is a device that includes a data accumulation unit 13, an optimal learning method output unit 14, an optimal learning method memory unit 15, a data acquisition unit for machine learning model estimation 16, a machine learning model estimation unit 17, a learned model memory unit 18, a data acquisition unit for water quality impact amount inference 19, a water quality impact amount inference value output unit 20, a water quality impact amount inference value memory unit 21, a data acquisition unit for water quality optimization operation index inference 23, a water quality optimization operation index output unit 24 and a water quality optimization operation index memory unit 25, and receives facility operation data 11 and water quality data 12 as input, and outputs a water quality impact amount inference value 22 and a water quality optimization operation index 26.

[0027] Among these components of the water quality management index output device 10, the optimal learning method output unit 14, the machine learning model estimation data acquisition unit 16, the machine learning model estimation unit 17, the water quality impact amount inference data acquisition unit 19, the water quality impact amount inference value output unit 20, the water quality optimization operation index inference data acquisition unit 23, and the water quality optimization operation index output unit 24 may be software or programs executed by, for example, a CPU in the information processing device. Also, the data accumulation unit 13, the optimal learning method storage unit 15, the learned model storage unit 18, the water quality impact amount inference value storage unit 21, and the water quality optimization operation index storage unit 25 may be realized by storage devices such as semiconductor memories and magnetic disks provided inside or outside the information processing device.

[0028] The water quality management index output device 10 has two input systems. One is facility operation data 11 consisting of trend data such as flow rate, ORP, DO, MLSS, and air supply volume obtained from the instrumentation equipment of the sewage treatment facility 1. The other is water quality data 12 consisting of water quality measurement results of the inflow sewage 2w, raw water 3w, settled water 4w, treated water 5w, discharged water 6w, reaction tank mixed liquid 2s, returned sludge 3s, and excess sludge 4s of the sewage treatment facility 1. Because the water quality data 12 was measured mainly as daily average water quality through periodic water quality tests, the facility operation data 11 also uses daily integrated values ​​of trend data.

[0029] As described above, the output of the water quality management index output device 10 is the water quality impact amount inference value 22 and the water quality optimization operation index 26. The water quality impact amount inference value 22 is a value obtained by inferring the water quality impact amount for each water quality impact factor in a certain treatment state for a water quality item subject to water quality management. The water quality optimization operation index 26 is an index related to the water quality impact factor that should be targeted in order to perform operation to optimize water quality, inferred from multiple water quality impact amount inference values ​​in multiple treatment states for a water quality item as a water quality management target.

[0030] In order to perform these outputs effectively, the water quality management index output device 10 has a trained model estimated by a machine learning model estimation unit 17 using data acquired by a machine learning model estimation data acquisition unit 16 as teacher data, and the trained model storage unit 18 has the trained model estimated by a machine learning model estimation unit 17. In addition, the machine learning model estimation data acquisition unit 16 and the machine learning model estimation unit 17 use the optimal teacher data, optimal learning algorithm, and optimal hyperparameters generated in the optimal learning method output unit 14 and stored in the optimal learning method storage unit 15 as the optimal learning method to acquire teacher data and estimate model parameters.

[0031] The water quality management index output device 10 uses the trained model provided in the machine learning model storage unit 18 and data acquired in the water quality impact amount inference data acquisition unit 19 by the above method, and outputs a water quality impact amount inference value 22 by a water quality impact amount inference device provided in a water quality impact amount inference value output unit 20. Also, using the trained model provided in the machine learning model storage unit 18 and data acquired in the water quality optimization operation index inference data acquisition unit 23 by the above method, a water quality impact amount inference value is output by a water quality impact amount inference device provided in a water quality optimization operation index output unit 24, and a water quality optimization operation index 26 is inferred and output based on this.

[0032] In the water quality impact amount inference device, when the trained model infers the objective variable, i.e., the water quality item subject to water quality management in this embodiment, the basis on which the inference was made is output as a water quality impact amount inference value.

[0033] The behavior of each component in the trained model estimation, the water quality impact amount inference value output, and the water quality optimization operation index output will be described below. Note that the function of the water quality impact amount inference device will be described after the explanation of these behaviors.

[0034] The facility operation data 11 is composed of trend data such as the operating state of the sewage treatment facility 1 and the measured values ​​of water quality measuring instruments installed in the sewage treatment facility 1.

[0035] The water quality data 12 includes the results of water quality measurements of samples from each process of sewage treatment conducted for water quality management, and the results of water quality measurements of effluent 6w from a sewage treatment plant conducted for statutory inspection.

[0036] The data accumulation unit 13 accumulates the integrated data obtained by integrating the facility operation data 11 and the water quality data 12. Because the water quality data 12 is mainly collected as daily average water quality, the facility operation data 11 also uses daily integrated values, and the facility operation data 11 and the water quality data 12 are combined using the date as a key and accumulated as integrated data.

[0037] Next, prior to the machine learning model estimation unit 17, the optimal learning method output unit 14 outputs an optimal learning method consisting of optimal teacher data, an optimal learning algorithm, and optimal hyperparameters. FIG. 3 is a flowchart of the optimal learning method output unit 14.

[0038] In the optimum learning method output unit 14, first, the objective variable determination s1 is executed. Since the present embodiment is aimed at water quality management, the objective variable is selected from water quality items that are important when performing water quality management. Examples include the phosphorus content and nitrogen content of the treated water 5w in the case of advanced treatment, and BOD (biochemical oxygen demand), SS (suspended solids), and ammonia nitrogen in the case of the standard activated sludge method. In addition, since there is a theoretical upper limit to the nitrogen removal rate in advanced treatment aimed at nitrogen removal, the nitrogen removal rate may be used as the objective variable instead of the nitrogen content. In addition to such water quality items, SVI (sludge volume index) focusing on the maintenance of properties of activated sludge, etc. may be used as the objective variable.

[0039] In the tentative explanatory variable extraction s2, explanatory variables are extracted from all variables in the integrated data in the data accumulation unit 13. In extracting explanatory variables, explanatory variables that are clearly not causally related to the objective variable selected in the objective variable determination s1 are excluded, including explanatory variables that are clearly not causally related to the objective variable selected in the objective variable determination s1, explanatory variables that can be calculated from other variables, explanatory variables that change due to the objective variable, explanatory variables with a small amount of data, explanatory variables that are one of the explanatory variables with a high correlation coefficient, and explanatory variables with no fluctuation in value (with extremely small variance). Between variables that can be calculated, more basic variables are extracted to increase the information available as explanatory variables. In addition, between explanatory variables with a high correlation coefficient, appropriate variables are extracted based on the distance in the process from the water quality set as the objective variable, the frequency and accuracy of measurement, and the causal relationship between the explanatory variables.

[0040] In the provisional teacher data acquisition s3, data on the variables determined in the objective variable determination s1 and the provisional explanatory variable extraction s2 is acquired as provisional teacher data from the integrated data in the data accumulation unit 13.

[0041] In outlier processing s4, abnormal values ​​and outliers in the provisional teacher data are dealt with. Methods for detecting outliers may include the 3σ method and IQR (interquartile deviation method). Sample data of detected outliers is removed from the provisional teacher data. However, if an outlier is not an abnormal value but is determined to be data indicating a situation that should be expected as a sewage treatment situation, even if it is extreme, it may not be removed from the provisional teacher data.

[0042] In the missing value processing s5, missing values ​​in the provisional teacher data after the outlier processing s4 are dealt with. Since the facility operation data 11 and the water quality data 12 have different measurement periods, the integrated data obtained by combining the two will have many missing values ​​due to their structure. To deal with such missing values, a listwise method for deleting samples containing missing values ​​or a single imputation method for filling missing values ​​with other values ​​by linear interpolation or the like may be used. In this embodiment of the present invention, the integrated data has a time series structure, and many missing values ​​are caused by differences in measurement periods, so the amount of teacher data may be secured by dealing with missing values ​​by linear interpolation.

[0043] In the provisional learning algorithm determination (S6), a provisional machine learning algorithm is determined to be used in the provisional trained model estimation (S8) using the provisional teacher data after the missing value processing (S5). When selecting a provisional machine learning algorithm, it is advisable to pay attention to the characteristics of the water quality items set as the objective variables. If the importance of the history effect is not high, it is advisable to use something similar to a gradient boosting decision tree. Conversely, if the history effect is strongly affected, it is also possible to use something similar to a regression neural network, although it is inferior to the gradient boosting decision tree in terms of the computational cost and interpretability of calculating the water quality impact amount inference value.

[0044] In the hyperparameter tuning s7, the hyperparameters are adjusted in the learning using the temporary teacher data and the temporary machine learning algorithm described above. For the hyperparameter tuning, a grid search, a random search, a Bayesian optimization method, or the like is used. In addition, an automatic hyperparameter optimization library may be used.

[0045] In tentative trained model estimation s8, a tentative trained model is estimated using the tentative teacher data, tentative machine learning algorithm, and hyperparameters described above. In the estimation, the tentative teacher data is divided into training data and test data by a hold-out method for accuracy evaluation s10, and the tentative trained model is estimated from the training data.

[0046] In the provisional trained model predicted value calculation s9, predicted data of the dependent variable is inferred using the provisional trained model and provisional teaching data described above.

[0047] In the accuracy evaluation step s10, the MAE (Mean Absolute Error), RMSE (Root Mean Square Error), and R are calculated from the objective variable data and the predicted data of the objective variable for each of the training data and test data inferred in the provisional trained model predicted value calculation step s9. 2Accuracy and generalization performance are evaluated using the score and CV score. The quantification lower limit of the water quality analysis and instrumentation equipment, and the allowable error of the objective variable of the water quality management target are used as the judgment criteria, and if the judgment criteria are not met, the process returns to Temporary Learning Algorithm Decision s6 and the learning algorithm is reviewed. If sufficient accuracy and generalization performance are obtained, the process proceeds to Optimal Learning Algorithm Output s11, and the optimal learning algorithm is output. Figure 9 is an example of a scatter plot showing the correlation between the objective variable data (actual measured data in the figure) and the predicted data of the objective variable, divided into test data and training data.

[0048] The water quality influence quantity inference unit s12 infers and outputs a provisional water quality influence quantity inference value from the provisional trained model and provisional teaching data in the optimal learning algorithm.

[0049] In the impact amount evaluation execution s13, the explanatory variables are evaluated based on the distribution of the provisional water quality impact amount inference values ​​output by the water quality impact amount inference device s12. Figure 13 is an example of visualizing the distribution of the provisional water quality impact amount inference values ​​(SHAP value in the figure) for each explanatory variable. If there is an explanatory variable for which the distribution width of the provisional water quality impact amount inference values ​​is sufficiently small, using the quantitative lower limit of the water quality analysis and instrumentation equipment, and the allowable error of the objective variable of the water quality management target as the judgment criteria, the process returns to the provisional explanatory variable extraction s2, and the explanatory variables to be extracted are reviewed so as to remove the explanatory variable. If there are no explanatory variables to be removed, the process proceeds to the optimal hyperparameter output s14 and the optimal teacher data output s15, where the optimal hyperparameters and optimal teacher data are output.

[0050] The optimal teacher data, optimal learning algorithm, and optimal hyperparameters optimized by the optimal learning method output unit 14 are output as an optimal learning method and stored in the optimal learning method storage unit 15. The optimal teacher data is stored as a method including outlier processing s4 and missing value processing s5, which are necessary for creating optimal teacher data from the data of the objective variable and explanatory variable acquired from the data accumulation unit 13.

[0051] The machine learning model estimation data acquisition unit 16 acquires machine learning model estimation data from the data accumulation unit 13 based on the optimal learning method stored in the optimal learning method storage unit 15. Fig. 7 shows an example of explanatory variable data of the machine learning model estimation data. Fig. 8 shows an example of objective variable data of the machine learning model estimation data.

[0052] 4 is a flowchart of the machine learning model estimation unit 17. In optimum learning method acquisition s16, an optimum learning method stored in optimum learning method storage unit 15 is acquired. In teacher data acquisition s17, data for machine learning model estimation is acquired by machine learning model estimation data acquisition unit 16. In trained model estimation s18, a trained model is estimated from the optimum learning method and the data for machine learning model estimation, and is output in trained model output s19.

[0053] The trained model memory unit 18 stores the trained model output from the machine learning model estimation unit 17.

[0054] The water quality influence quantity inference data acquisition unit 19 acquires water quality influence quantity inference data from the data accumulation unit 13 based on the optimum learning method stored in the optimum learning method storage unit 15 .

[0055] 5 is a flow diagram of the water quality influence amount inference value output unit 20. In water quality influence amount inference target data acquisition s20, water quality influence inference target data is acquired from the water quality influence amount inference data acquired in the water quality influence amount inference data acquisition unit 19. In addition, in trained model acquisition s21, a trained model stored in the trained model storage unit 18 is acquired.

[0056] In model predicted value inference s22, model predicted value data is inferred from the water quality influence amount inference target data and the trained model, and is output in model predicted value data output s23. In addition, in water quality influence amount inference unit s24, water quality influence amount inference value data is inferred from the water quality influence amount inference target data and the trained model using the water quality influence amount inference unit, and is output in water quality influence amount inference value data output s25.

[0057] The outputs of the model predicted value data output s23 and the water quality influence amount inferred value data output s24 are stored in the water quality influence amount inferred value storage unit 21 and become the output of the water quality management index output device 10 as the water quality influence amount inferred value 22. Fig. 10 shows an example of the output water quality influence amount inferred value.

[0058] The water quality impact amount inference value 22 output from the water quality management index output device 10 is a value obtained by inferring the water quality impact amount for each explanatory variable, i.e., water quality influence factor, for a certain sample, i.e., a certain treatment state, for the water quality item set as the objective variable as a water quality management target.

[0059] FIG. 11 is an example of visualizing the water quality impact amount inference value 22 as a waterfall plot, with the nitrogen removal rate, i.e., the nitrogen removal rate, being the objective variable, being the nitrogen removal rate removed in the process from the settling water to the treated water. In this example, it is inferred that the organic nitrogen concentration of the settling water 4w was 12 mg / L, which had the effect of increasing the nitrogen removal rate of the treated water 5w by 2.21% pt, and that the temperature of the biological reactor 5 (temperature (°C) reaction tank mixed liquid A in the figure) was 18.6°C, which had the effect of decreasing the nitrogen removal rate of the treated water 5w by 0.64% pt. Based on such inference information, it is possible to understand what water quality impact factors have influenced the water quality items subject to water quality management, and how much water quality impact they have received, resulting in the current water quality.

[0060] FIG. 12 is an example of visualizing the difference between the water quality influence amount inferred value 22 and the explanatory variable value of a sample to be compared in addition to the sample for which the water quality influence factor is to be understood, that is, the amount of change. The vertical axis shows each water quality influence factor and the amount of change of the explanatory variable in parentheses. The horizontal axis shows the amount of change of the water quality influence amount inferred value 22 for each water quality influence factor. By focusing on the amount of change in this way, it is possible to understand the relationship between the increase or decrease in the value of the water quality influence factor and the increase or decrease in the water quality influence amount inferred value 22. In this example, it is inferred that the nitrogen removal rate increased by about 0.5%pt due to the decrease in DO of the biological reaction tank by 0.88 mg / L. Naturally, the water quality influence amount inferred value 22 may be used as a material for determining the cause and measures when the water quality item to be managed by the water quality management deteriorates. In addition, from FIG. 13, which is exemplified in the influence amount evaluation of the optimal learning method output unit 13, it is also possible to read the qualitative relationship between the change in the value of the water quality influence factor and the change in the water quality influence amount inferred value 22.

[0061] As described above, by using the water quality impact amount inference value 22, water quality management based on accurate understanding of water quality impact factors, which previously required accumulated experience and know-how, can now be performed based on data without relying on them. Furthermore, it is now possible to quantitatively evaluate the water quality impact amount of water quality impact factors, making it possible to perform more precise water quality management.

[0062] The water quality optimization operation index inference data acquisition unit 23 acquires water quality optimization operation index inference data from the data accumulation unit 13 based on the optimum learning method stored in the optimum learning method storage unit 15 .

[0063] 6 is a flow diagram of the water quality optimization operation index output unit 24. In water quality optimization operation index inference target data acquisition s26, water quality optimization operation index inference target data is acquired from water quality optimization operation index inference data acquired by the water quality optimization operation index inference data acquisition unit 23. In addition, in trained model acquisition s27, a trained model stored in the trained model storage unit 18 is acquired. Note that, as the water quality optimization operation index inference target data, data from one year to several years is used so that seasonal fluctuations in the treatment state can be appropriately evaluated.

[0064] In the water quality influence amount inference unit s28, water quality influence amount inference value data is inferred from the water quality optimization operation index inference target data and the trained model using the water quality influence amount inference unit. In the water quality optimization operation index inference s29, a water quality optimization operation index is further inferred from the water quality influence amount inference value data inferred in the water quality influence amount inference unit s28, and is output in the water quality optimization operation index inference output s30, which becomes the output of the water quality management index output device 10 as the water quality optimization operation index 26.

[0065] In the water quality optimization operation index inference s29, the value of the water quality influence factor and the average value of the water quality influence amount inference value at that value are calculated for each water quality influence factor for all samples of water quality influence inference value data inferred by the water quality influence amount inferor 28, and a desirable value for each water quality influence factor that optimizes the water quality is inferred from the relationship between the two. A specific example of the inference method is shown below.

[0066] Figure 14 shows an example of visualizing the relationship between the MLSS value and the average value of the estimated water quality impact quantity corresponding to that value, when the nitrogen removal rate is the objective variable and MLSS is the water quality impact factor. An appropriate class width is set for calculating the average value. In the figure, the horizontal axis is the MLSS value, and the vertical axis is the average value of the estimated water quality impact quantity. In this example, when the MLSS is 1,400 mg / L or less, the average value of the estimated water quality impact quantity is negative, indicating that the state has a negative impact on the nitrogen removal rate. If the MLSS is further increased, the impact remains flat up to 1,600 mg / L, but once it exceeds 1,600 mg / L, it starts to rise sharply, indicating that the nitrogen removal rate increases significantly and transitions to a state of positive impact. Further increasing the MLSS after that is expected to improve the nitrogen removal rate, but the effect is not as great as the change around 1,600 mg / L. Therefore, taking into account the fact that increasing the MLSS results in increased power consumption, the water quality optimization operation index for the MLSS in this system is estimated to be 1,600 to 1,700 mg / L.

[0067] Figure 15 shows a scatter plot of the water quality impact factor values ​​and the estimated water quality impact quantity values ​​when the nitrogen removal rate is the objective variable and MLSS is the water quality impact factor. In addition, the temperature of the biological reactor is expressed by the shade of each plot. By using such plots, it is possible to infer the water quality optimization operation index by season. From Figure 15, it can be seen that the nitrogen removal rate increases significantly at 1,600 mg / L over a wide temperature range. In addition, when the MLSS is 1,600 mg / L or less, the lower the temperature, the greater the decrease in the nitrogen removal rate, and it can be seen that lowering the MLSS during the low water temperature period increases the risk of a decrease in the nitrogen removal rate. From this tendency, it can be inferred that in the low water temperature period, the MLSS should be maintained at a higher value even within the range of the water quality optimization operation index in order to avoid the risk of a decrease in the nitrogen removal rate due to a decrease in the MLSS.

[0068] As described above, the present invention makes it possible to estimate water quality prediction models for sewage treatment with high accuracy, to perform water quality management based on a quantitative evaluation of factors affecting water quality in the treatment state, to derive optimal values ​​for important factors affecting water quality and to perform water quality management aiming at these values, to take appropriate measures based on a quantitative evaluation of factors affecting water quality due to deterioration of treatment, and to achieve all of this by utilizing accumulated data without relying on experience.

[0069] Next, the behavior of the water quality impact amount inference device will be explained. In this embodiment, SHAP (SHapley Additive exPlanations) is used to infer the water quality impact amount inference value. SHAP is a theory based on the Shapley value, which is one method of distributing gains in cooperative games in game theory. First, the contents of the Shapley value will be explained.

[0070] Let's set up the following cooperative game. The distinctive feature is that it is assumed that the prize money won will change depending on the combination of people. With this setting, if three people win a prize of 600,000 yen, we will think about how to divide it fairly. Mr. A can win a prize of 200,000 yen as an individual. Mr. B can win a prize of 100,000 yen individually. Mr. C can win a prize of 50,000 yen as an individual. If person A and person B cooperate, they can win a prize of 360,000 yen. If A and C work together, they can win a prize of 220,000 yen. If Mr. B and Mr. C work together, they can win a prize of 260,000 yen. If three people work together, they can win a prize of 600,000 yen.

[0071] Rather than simply distributing the prize money based on the individual prize money, we consider the situation in which three people participate simultaneously as a superposition of events (6 ways) in which three people participate in turn, in order to take into account that the prize money changes depending on the cooperation situation. In addition, we focus on how much the prize money increases when a new player joins, that is, the marginal contribution.

[0072] If we calculate the order in which the three people participate and their marginal contributions for all cases, we get the following. TIFF2025009595000002.tif48158

[0073] For reference, we will explain what happens if A, B, and C join in that order. Person A's contribution increases from 0 people, 0 yen, to 200,000 yen per person, so Person A's marginal contribution is 200,000 yen. When Person B joins, the contribution increases to 360,000 yen with the cooperation of Persons A and B, so Person B's marginal contribution becomes 36-20=160,000 yen. When Person C joins last, the contribution increases to 600,000 yen, so Person C's marginal contribution becomes 600-36=240,000 yen.

[0074] The event in which three players win 600,000 yen can be considered to be the superposition and average of all these cases, and the prize money won by each player at that time is called the Shapley value. A's Shapley value = (20 + 20 + 26 + 34 + 17 + 34) / 6 = 25.166 B's Shapley value = (16 + 38 + 10 + 10 + 38 + 21) / 6 = 22.166 C's Shapley value = (24 + 2 + 24 + 16 + 5 + 5) / 6 = 12.666

[0075] In this way, the Shapley score can evaluate the degree of contribution taking into account the combination of players. Therefore, SHAP is an application of machine learning with players as explanatory variables and winnings as objective variables. In this embodiment, the explanatory variables are further water quality influence factors, and the objective variables are water quality items to be managed.

[0076] SHAP attempts to quantify the importance, or contribution, of each explanatory variable of the data to be inferred based on the degree of change in the predicted result of the objective variable when each explanatory variable is changed for the data to be inferred. This quantified contribution corresponds to the Shapley value, and in this embodiment corresponds to the water quality impact amount inference value.

[0077] More specifically, if the data to be inferred is a and the trained model for predicting the objective variable is f, the explanatory variables to be inferred are x=a=(a1, a2, ..., a n ), a simplified model g expressed by the following equation 1 is generated as an approximate model g that approximates f.

number

[0078] In the above formula, multiple coefficients w = (w0, w1, ..., w n ) is an approximate model parameter representing the approximate model g. Here, z(a) is expressed by the following equation 2.

number

[0079] δ in the above formula ai is the so-called Kronecker delta. In other words, in the above formula, the explanatory variables x = (x1, x2, ..., x n ) the i-th element x i is the value of the i-th element of the data to be inferred, a i If it is equal to , the value is set to 1, and if it is not equal to , the value is set to 0.

[0080] In SHAP, the calculation is constrained by three properties: local accuracy, missingness, and consistency. Local accuracy is a constraint that the explanatory variable x of the inference model f to be explained and the approximate model g(f) must be equal when the explanatory variable x is equal to the inference target data a and x=a. In other words, when the inference target x=a=(a1, a2, ..., a n ), the following equation 3 holds:

number

[0081] The missingness is the absence of explanatory factor z i Contribution of w i is zero. That is, in the above equation 1, z i For i such that (a)=0, w i =0 holds true.

[0082] Consistency is a constraint as explained below. First, between the inference model f and the approximation model g(f), f(a) and g(f)(z(a)) are assumed to be equal. In this case, the following formula 4 holds for the two inference models f1 and f2.

number

[0083] Contribution w that satisfies each of the above constraints i(f) can be uniquely determined as shown in the following equation 5.

number

[0084] In formula 5, |S| is the number of elements in the set S. In other words, the contribution of the i-th explanatory variable is calculated by averaging the sum of the differences between when the i-th factor is added and when it is not added for all combinations. This contribution, i.e., the multiple coefficients w = (w0, w1, ..., w n ) are the approximate model parameters representing the approximate model g.

[0085] In this way, the water quality impact amount inference device generates an approximation model that approximates the trained model for the inference target data based on the inference target data and a plurality of related data having values ​​related to the inference target data, and generates approximate model parameters. Approximation model parameters w=(w0, w1, ..., w n ), the i-th count w i If the value of is a large positive value, the i-th variable x i The value of a i This can be interpreted as strongly supporting that the result of the model f is f(x). Therefore, the approximate model parameters w = (w0, w1, ..., w n ) can be applied as the amount of impact that each explanatory variable has on water quality, i.e., the estimated value of the water quality impact.

[0086] Using the above method, the water quality impact amount inference device receives the data to be inferred and the trained model as inputs, and outputs the approximate model parameter w as the water quality impact amount inference value.

[0087] Although the embodiment and the modified examples of the present invention have been described above, the present invention is not limited to these. The present invention includes appropriate combinations of the embodiment and the modified examples, and appropriate modifications thereto. [Industrial Applicability]

[0088] The present invention can be used for water quality management in sewage treatment and for operation control of sewage treatment plants. [Explanation of symbols]

[0089] 1: Sewage treatment facility, 2: Grit chamber, 3: Water lift pump equipment, 4: Primary sedimentation tank, 5: Biological reaction tank, 6: Final sedimentation tank, 7: Disinfection equipment, 8: Blower equipment, 9: Sludge treatment equipment, 10: Water quality management index output device, 11: Facility operation data, 12: Water quality data, 13: Data storage unit, 14: Optimal learning method output unit, 15: Optimal learning method memory unit 16: Data acquisition unit for machine learning model estimation, 17: Machine learning model estimation unit, 18: Learned model memory unit, 19: Data acquisition unit for water quality impact amount inference, 20: Water quality impact amount inference value output unit, 21: Water quality impact amount inference value memory unit, 22: Water quality impact amount inference value, 23: Data acquisition unit for water quality optimization operation index inference, 24: Water quality optimization operation index output unit, 25: Water quality optimization operation index memory unit, 26: Water quality optimization operation index 1w: In-plant return water, 2w: Influent sewage, 3w: Raw water, 4w: Settled water, 5w: Treated water, 6w: Discharge water, 1s: Raw sludge, 2s: Reactor tank mixed liquid, 3s: Return sludge, 4s: Excess sludge s1: Determine the objective variable, s2: Extract the tentative explanatory variables, s3: Obtain the tentative teacher data, s4: Handling outliers, s5: Handling missing values, s6: Determine the tentative learning algorithm, s7: Adjust the hyperparameters, s8: Estimate the tentative trained model, s9: Calculate the tentative trained model prediction value, s10: Perform the accuracy evaluation, s11: Output the optimal learning algorithm, s12: Water quality impact amount inference device, s13: Perform the impact amount evaluation, s14: Output the optimal hyperparameters, s15: Output the optimal teacher data, s16: Obtain the optimal learning method, s 17: Obtain teacher data, s18: Estimate trained model, s19: Output trained model, s20: Obtain data to be used for inferring water quality impact amount, s21: Obtain trained model, s22: Infer model predicted value, s23: Output model predicted value data, s24: Water quality impact amount inferor, s25: Output water quality impact amount inferred value data, s26: Obtain data to be used for inferring water quality optimization operation index, s27: Obtain trained model, s28: Water quality impact amount inferor, s29: Infer water quality optimization operation index, s30: Output water quality optimization operation index

Claims

1. A water quality management index output device for inferring and outputting a water quality impact amount inference value and a water quality optimization operation index used for water quality management, a data storage unit that stores water quality related data; A machine learning model estimation unit that outputs a trained model from the water quality-related data; A trained model storage unit that stores the trained model; a water-quality impact amount inference value output unit that infers and outputs the water-quality impact amount inference value from the water-quality related data and the trained model; a water quality influence quantity inference value storage unit for storing the water quality influence quantity inference value; A water quality optimization operation index output unit that infers and outputs the water quality optimization operation index from the water quality related data and the trained model; A water quality management index output device comprising:

2. The water quality-related data is integrated data that integrates facility operation data and water quality data. The water quality management index output device according to claim 1.

3. The machine learning model estimation unit performs machine learning based on an optimal learning method on the integrated data and outputs a trained model. The water quality management index output device according to claim 1.

4. The optimal learning method is determined by optimal teacher data, an optimal learning algorithm, and an optimal hyperparameter. The water quality management index output device according to claim 3.

5. The integrated data is subjected to setting of a target variable, extraction of explanatory variables, outlier processing and missing value processing to create provisional teacher data; A provisional trained model is estimated using the provisional teacher data, the provisional machine learning algorithm, and the provisional hyperparameters; Optimizing the temporary teacher data, the temporary learning algorithm, and the temporary hyperparameters by evaluating the accuracy and the influence amount of the temporary trained model; An optimal learning method output unit that outputs the optimal learning method, The water quality management index output device according to claim 4.

6. In the accuracy evaluation, a provisional trained model predicted value is inferred using the provisional trained model and the provisional teacher data; optimizing the provisional learning algorithm and the provisional hyperparameters based on an evaluation of the prediction accuracy, generalization performance, and correlation between the objective variable and the provisional learning model prediction value of the provisional learning model; The water quality management index output device according to claim 5.

7. In the impact amount evaluation, a provisional water quality impact amount inference value is inferred for the provisional trained model and the provisional teacher data, and the provisional teacher data is optimized. The water quality management index output device according to claim 5.

8. The provisional water quality impact amount inference value is a value inferred from the provisional trained model and the provisional training data as a magnitude of contribution of the explanatory variables of the provisional training data to the provisional trained model predicted value. The water quality management index output device according to claim 7.

9. The water quality impact amount inference value output unit infers a trained model predicted value using the trained model and the water quality impact amount inference data, Inferring and outputting the water quality impact amount inference value for the trained model and the water quality impact amount inference data; The water quality management index output device according to claim 1.

10. The water quality impact amount inference value is a value inferred from the trained model and the water quality impact amount inference data as a magnitude of contribution of an explanatory variable of the water quality impact amount inference data to the trained model predicted value. The water quality management index output device according to claim 9.

11. The water quality impact amount inference data is inference target data for the water quality impact amount inference value, extracted from the integrated data by a method based on the optimal learning method. The water quality management index output device according to claim 10.

12. The water quality optimization operation index output unit infers and outputs the water quality optimization operation index using the trained model and the water quality optimization operation index inference data. The water quality management index output device according to claim 1 or 9.

13. The water quality optimization operation index inference data is inference target data of the water quality optimization operation index extracted from the integrated data by a method based on the optimal learning method. The water quality management index output device according to claim 12.

14. A method for outputting a water quality management index used in water quality management by a water quality management index output device, comprising: A trained model is estimated from water quality-related data, Inferring a water quality impact amount inference value from the trained model and the data for inferring water quality impact amount; Inferring a water quality optimization operation index from the trained model and data for inferring a water quality optimization operation index; outputting the water quality impact amount inference value and the water quality optimization operation index as a water quality management index; method.

Images