Water quality prediction device and water quality prediction method

Abstract

To provide a technique for building a model capable of predicting treated water quality with a consistent formulation without performing arbitrary data interpolation even if data acquisition intervals are undefined.SOLUTION: Provided is a water quality prediction device comprising: a training data acquisition unit that acquires data used for training to generate training data; a prediction-target data acquisition unit that acquires explanatory variable data used for prediction; a prediction model parameter updating unit that updates parameters of a prediction model using the training data to generate updated prediction model parameters; a prediction model parameter storage unit that stores the updated prediction model parameters from the prediction model parameter updating unit; and a treated-water quality prediction unit that derives the quality of treated water at a future time by time-evolving a neural ordinary differential equation by using prediction-target explanatory variable data obtained from the prediction-target data acquisition unit and the updated prediction model parameters stored in the prediction model parameter storage unit.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to a water quality prediction device and a water quality prediction method that are applied to water treatment plants and the like. [Background technology]

[0002] In recent years, advances in IoT (Internet of Things) technology have led to an increasing number of cases in which real-time measurement data from measuring instruments installed in various locations within a wastewater treatment facility, for example, is used for operational management.

[0003] Patent Document 1, an example of prior art, discloses a technology in which statistical information obtained by measuring the water quality that needs to be treated in a controlled plant and operating value information are used as input to calculate estimated values ​​of treated water quality using multiple estimation models, each with a different algorithm, and the multiple calculated estimated values ​​of treated water quality are aggregated using one aggregation calculation model selected from multiple aggregation calculation models based on the input information, thereby appropriately and automatically switching or combining multiple estimation models depending on the given situation to obtain highly accurate estimation results. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2020-113151 A Summary of the Invention [Problem to be solved by the invention]

[0005] However, the above-mentioned conventional techniques have a problem in that, although the water treatment environment that is the target of prediction for water quality control changes from moment to moment, the fact that there is a connection between before and after time is not taken into consideration. Furthermore, in the above-mentioned conventional technology, the weighting coefficients of the selection or combination from multiple estimation models are updated based on the input information, but there are problems in that the acquisition interval is indefinite and the fact that water quality data acquired by manual analysis is dominant in operational decisions is not taken into consideration.

[0006] The present invention has been made in consideration of the above, and aims to provide a technology for constructing a model that can predict treated water quality with a consistent formulation, even if the data acquisition interval is not fixed, without performing arbitrary data interpolation. [Means for solving the problem]

[0007] One aspect of the present invention that solves the above-mentioned problems and achieves the object is a water quality prediction device that includes a data accumulation unit that accumulates water quality data acquired by manual analysis at least from a prediction target, a learning data acquisition unit that acquires data used for learning from the data accumulated in the data accumulation unit and generates learning data, a prediction target data acquisition unit that acquires explanatory variable data to be used for prediction from the data accumulated in the data accumulation unit, a prediction model parameter update unit that uses the learning data acquired from the learning data acquisition unit to update parameters of a prediction model and generate updated prediction model parameters, a prediction model parameter memory unit that stores the updated prediction model parameters from the prediction model parameter update unit, and a treated water quality prediction unit that derives the treated water quality at a future time by evolving over time using a neural ordinary differential equation using the explanatory variable data of the prediction target acquired from the prediction target data acquisition unit and the updated prediction model parameters stored in the prediction model parameter memory unit.

[0008] In the above-mentioned configuration, the system further includes a data processing unit that processes real-time measurement data acquired in real time by a measuring instrument installed in the prediction target into a representative value for a predetermined period of time, and generates merged data by merging the representative value of the real-time measurement data with manual analysis water quality data of the prediction target, and the data storage unit stores the merged data as the water quality data.

[0009] Alternatively, one aspect of the present invention for solving the above-mentioned problems and achieving the object includes a data accumulation unit that accumulates water quality data acquired by manual analysis from at least a prediction target, a treated water quality prediction model training data acquisition unit that acquires data used for learning from the data accumulated in the data accumulation unit and generates treated water quality prediction model training data, a prediction target data acquisition unit that acquires explanatory variable data used for water quality fluctuation prediction and treated water quality prediction from the data accumulated in the data accumulation unit, a treated water quality prediction model parameter update unit that uses the treated water quality prediction model training data acquired from the treated water quality prediction model training data acquisition unit to update parameters of the treated water quality prediction model and generate updated treated water quality prediction model parameters, a treated water quality prediction model parameter storage unit that stores the updated treated water quality prediction model parameters from the treated water quality prediction model parameter update unit, and a water quality fluctuation prediction model training data acquisition unit that acquires data used for learning from the data accumulated in the data accumulation unit and generates water quality fluctuation prediction model training data. a water quality fluctuation prediction model parameter update unit that uses the water quality fluctuation prediction model learning data acquired from the water quality fluctuation prediction model learning data acquisition unit to update the parameters of the water quality fluctuation prediction model to generate updated water quality fluctuation prediction model parameters; a water quality fluctuation prediction model parameter memory unit that stores the updated water quality fluctuation prediction model parameters from the water quality fluctuation prediction model parameter update unit; a water quality fluctuation prediction unit that derives water quality fluctuation using explanatory variable data of the prediction target acquired from the prediction target data acquisition unit and the updated water quality fluctuation prediction model parameters stored in the water quality fluctuation prediction model parameter memory unit; and a treated water quality prediction unit that derives the treated water quality at a future time by performing time evolution with a neural ordinary differential equation using the explanatory variable data of the prediction target acquired from the prediction target data acquisition unit, the treated water quality prediction model parameters stored in the treated water quality prediction model parameter memory unit, and the water quality fluctuation data acquired from the water quality fluctuation prediction unit.

[0010] Alternatively, one aspect of the present invention that solves the above-mentioned problems and achieves the object is a water quality prediction method that includes accumulating water quality data obtained by manual analysis at least from a prediction target, obtaining data from the accumulated data to be used for learning and generating learning data, obtaining explanatory variable data from the accumulated data to be used for prediction, updating parameters of a prediction model using the learning data to generate updated prediction model parameters, storing the updated prediction model parameters, and deriving treated water quality at a future time by performing time evolution using a neural ordinary differential equation using the explanatory variable data of the prediction target obtained and the updated prediction model parameters. Effect of the Invention

[0011] According to the present invention, even if the data acquisition intervals are not fixed, it is possible to construct a model for predicting treated water quality with a consistent formulation without performing arbitrary data interpolation. [Brief description of the drawings]

[0012] [Figure 1] FIG. 1 is a diagram showing the configuration of a water treatment facility to which the water quality prediction devices according to the first and second embodiments can be applied. [Diagram 2] FIG. 2 is a functional block diagram showing the configuration of the water quality prediction device according to the first embodiment. [Diagram 3] FIG. 3 shows an example of the simplest ordinary differential solver. [Figure 4] FIG. 4 is a diagram showing an image of the forward process. [Diagram 5] FIG. 5 is a flowchart showing the learning process and the prediction process in the first embodiment. [Figure 6] FIG. 6 is a diagram for explaining an outline of the prediction process in the second embodiment. [Figure 7] FIG. 7 is a functional block diagram showing the configuration of a water quality prediction device according to the second embodiment. [Figure 8] FIG. 8 is a flowchart showing the learning process and the prediction process in the second embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0013] Hereinafter, an embodiment of the present invention will be described with reference to the accompanying drawings. However, the present invention should not be construed as being limited to the following embodiments.

[0014] <Embodiment 1> FIG. 1 is a diagram showing the configuration of a water treatment facility 200 to which a water quality prediction device 100 according to this embodiment can be applied. The water treatment facility 200 shown in Figure 1 includes a primary sedimentation tank 1, a reaction tank 2, a final sedimentation tank 3, a chlorine mixing tank 4, a blower 5, an air volume control valve 6, an aeration device 7, a first pump 8, a second pump 9, a gravity thickening tank 10, a mechanical thickening tank 11, a digestion tank 12, a dewatering tank 13, and pipes 21, 22, 23, 24, and 25.

[0015] The primary sedimentation tank 1 is a sedimentation tank into which raw water is introduced. This raw water is wastewater containing organic matter. In the primary sedimentation tank 1, solid-liquid separation of the raw water is carried out, and the effluent from the primary sedimentation tank 1 is sent to the reaction tank 2 through a pipe 21. The raw sludge that settles in the primary settling tank 1 is sent to the gravity thickening tank 10 through a pipe 25.

[0016] The reaction tank 2 contains microorganisms and purifies the effluent water from the primary sedimentation tank 1 using the microorganisms. In the reaction tank 2, the microorganisms grow by assimilating the organic matter contained in the effluent water from the primary sedimentation tank 1, and activated sludge is formed by biological treatment using the microorganisms. The effluent from the reaction tank 2 is sent to the final sedimentation tank 3 through piping 22.

[0017] The final settling tank 3 is a settling tank in which activated sludge contained in the effluent from the reaction tank 2 is allowed to settle. The supernatant of the final sedimentation tank 3 is sent to the chlorine mixing tank 4 through pipe 24. A portion of the sludge that has settled in the final settling tank 3 is returned to the reaction tank 2 through a pipe 23 by a first pump 8 and is reused. The remaining sludge that has settled in the final settling tank 3 is sent as excess sludge by a second pump 9 to a mechanical thickening tank 11. The chlorine mixing tank 4 is a mixing tank that disinfects the effluent from the final sedimentation tank 3 by chlorination. The treated water disinfected with chlorine in the chlorine mixing tank 4 is discharged outside the water treatment facility 200.

[0018] The blower 5 supplies air to the air diffuser 7 . The air volume adjustment valve 6 is provided on a pipe passing through the air diffuser 7, and adjusts the volume of air blown by opening and closing it. The air diffuser 7 is provided at the bottom of the reaction tank 2 and is connected to a pipe passing through the air volume adjustment valve 6, and supplies air whose volume is adjusted by the air volume adjustment valve 6 into the reaction tank 2. When the amount of air sent to the reaction tank 2 is adjusted in this manner, the DO (Dissolved Oxygen) value, which is the amount of dissolved oxygen in the reaction tank 2, is adjusted, and the progress of the biological treatment is adjusted.

[0019] The first pump 8 is a return sludge pump that sends a portion of the sludge that has settled in the final settling tank 3 to the reaction tank 2 through piping 23. The second pump 9 is an excess sludge extraction pump that sends the remaining sludge that has settled in the final settling tank 3 to the mechanical thickening tank 11 as excess sludge.

[0020] Measurement points MP1, MP2, and MP3 shown in FIG. 1 are measurement points where water is sampled for manual analysis of each parameter indicating water quality. The measuring point MP1 is placed at the inlet of the reaction vessel 2. Measuring point MP2 is located at the outlet of reaction vessel 2. The measuring point MP3 is placed at the outlet of the chlorine mixing tank 4. Here, examples of parameters indicating water quality include a DO value, which is the amount of dissolved oxygen, and a MLSS (Mixed Liquor Suspended Solids) value, which is the concentration of suspended matter. An operator of the water treatment facility 200 determines the manipulated variable of the controlled object by referring to each parameter indicating the water quality obtained by manual analysis of the water sampled at the measurement points MP1, MP2, and MP3.

[0021] The gravity thickening tank 10 is a tank for thickening the raw sludge that has settled in the primary sedimentation tank 1. The mechanical thickening tank 11 is a tank in which the excess sludge extracted from the final settling tank 3 by the second pump 9 is thickened. The sludge thickened in the gravity thickener 10 and the mechanical thickener 11 is sent to the digester 12.

[0022] The digestion tank 12 is a tank for carrying out digestion treatment of the thickened sludge. Here, the digestion process may be carried out, for example, by an anaerobic digestion process. In the anaerobic digestion process, organic sludge is decomposed by anaerobic microorganisms. The sludge decomposed by the digestion process is sent to the dewatering tank 13.

[0023] The dewatering tank 13 is a tank for dewatering the sludge decomposed by the digestion process, thereby reducing the water content of the sludge and reducing its volume.

[0024] Pipe 21 is disposed between primary sedimentation tank 1 and reaction tank 2, and is a pipe that conveys effluent from primary sedimentation tank 1 to reaction tank 2. The pipe 22 is disposed between the reaction tank 2 and the final sedimentation tank 3, and is a pipe that conveys the effluent from the reaction tank 2 to the final sedimentation tank 3. The pipe 23 is disposed between the first pump 8 and the reaction tank 2, and is a pipe for sending a portion of the sludge in the final settling tank 3 to the reaction tank 2. The pipe 24 is disposed between the final sedimentation tank 3 and the chlorine mixing tank 4, and is a pipe that conveys the effluent from the final sedimentation tank 3 to the chlorine mixing tank 4. The pipe 25 is disposed between the primary sedimentation tank 1 and the gravity thickening tank 10, and is a pipe for transporting raw sludge from the primary sedimentation tank 1 to the gravity thickening tank 10.

[0025] In the water treatment facility 200 shown in Figure 1, the main operational settings are the inflow volume, which is the amount of water from the primary sedimentation tank 1 to the reaction tank 2, the air supply volume, which is the amount of air supplied into the reaction tank 2, the returned sludge volume, which is the amount of sludge returned to the reaction tank 2 by the first pump 8, and the excess sludge withdrawal volume, which is the amount of sludge withdrawn from the final sedimentation tank 3 by the second pump 9. Each of these operational items has different settings depending on the treatment plant.

[0026] The reaction tank 2 can be controlled, for example, by constant air flow rate control, constant ratio control, and constant DO control. Here, the constant airflow control is a control performed so as to maintain a constant airflow rate that is set as a target airflow rate value. The constant ratio control is a control performed so that the air blowing rate corresponds to the inflow rate from the primary sedimentation tank 1 to the reaction tank 2, that is, so that the ratio between the inflow rate and the air blowing rate is constant. Moreover, the constant DO control is a control performed so that the DO value of the reaction tank 2 becomes a set target DO value.

[0027] Here, examples of operational set values, which are the operational quantities to be controlled, include the rotation speed of the first pump 8, which adjusts the amount of sludge returned from the final sedimentation tank 3 of the water treatment facility 200, the amount of sludge withdrawn per unit time or the excess sludge withdrawal time of the second pump 9, which adjusts the amount of excess sludge withdrawn from the final sedimentation tank 3 of the water treatment facility 200, and the injection rate of polymer coagulant into the dewatering tank 13 of the water treatment facility 200. These operation settings are determined by the operator's overall judgment based on the manual analysis of water quality data collected at measuring points MP1, MP2, and MP3 several times a week. In addition to the above-mentioned manual analysis water quality data, the operator may also refer to real-time measurement data in the reaction tank 2 such as inflow and outflow water quality, total phosphorus, total nitrogen, etc., DO value (dissolved oxygen), MLSS value (activated sludge suspended solids), etc., measured in real time at the water treatment facility 200, and treated water quality, such as total phosphorus and total nitrogen.

[0028] In recent years, with the advancement of IoT (Internet of Things) technology, there are an increasing number of cases where real-time measurement data from measuring instruments installed at various locations within a facility, such as a wastewater treatment facility like the water treatment facility 200, is used for operational management. However, at present, such measuring instruments are expensive and therefore can only be installed in limited locations. Even if they are installed, the accuracy varies greatly depending on the installation location, the timing of calibration, and the like. Therefore, when making operational decisions, there is a high degree of reliance on data obtained through human management, such as water quality data obtained by manual analysis, which is highly reliable. The water quality data obtained by such manual analysis is essential for operating wastewater treatment plants. However, water quality data obtained by manual analysis is usually not obtained in real time at regular intervals like real-time measurement data obtained by a measuring instrument, but is obtained at irregular intervals, such as several times a week, as described above. Meanwhile, in the field of forecasting technology, there have been many reported cases in which highly accurate predictions have been achieved by applying deep learning to time-series data. In particular, when dealing with time-series data such as the quality of treated water from wastewater treatment plants, it is common to apply recurrent neural networks as a deep learning method that handles state transitions in order to take into account time context. By using a recurrent neural network, it is possible to make highly accurate predictions of water quality data obtained by manual analysis, which is crucial for making operational decisions at wastewater treatment plants. However, recurrent neural networks are usually based on the assumption that the time interval between each state is constant. When dealing with data collected at irregular intervals, such as water quality data obtained by manual analysis, data interpolation is required to make the intervals between each state constant. In this case, the accuracy of the prediction will vary greatly depending on how the missing data is interpolated, and the decision as to which interpolation method is appropriate is left to the designer of the prediction model. Therefore, although deep learning itself is data-driven, there is a large amount of arbitrary elements that creep into predictive models, and the accuracy of the predictions depends on the skill of the person who designs the predictive models.

[0029] The water quality prediction device 100 according to this embodiment predicts future manual analysis water quality data when determining these operation items, and applies a neural ordinary differential equation method to the prediction. Then, the operator determines the operation setpoint, which is the manipulated variable of the controlled object, based on the prediction of future manual water quality data. In this way, it becomes possible for people who do not have intuition, experience, or know-how to become operators.

[0030] FIG. 2 is a functional block diagram showing the configuration of the water quality prediction device 100 according to this embodiment. The water quality prediction device 100 shown in Figure 2 includes a data processing unit 101, a data accumulation unit 102, a learning data acquisition unit 103, a prediction model parameter update unit 104, a prediction model parameter memory unit 105, a prediction target data acquisition unit 106, and a treated water quality prediction unit 107.

[0031] The data processing unit 101 processes real-time measurement data acquired in real time by a measuring instrument installed in the prediction target into a representative value for a specified time period, and generates merged data by merging the representative value of the real-time measurement data with the manual analysis water quality data of the prediction target. It should be noted that the data processing unit 101 does not need to be provided in the water quality prediction device 100, for example, when real-time measurement data from a measuring instrument is not used and only manual analysis water quality data of the prediction target is used. Examples of real-time measurement data include inflow water volume, DO set value, TP (discharge water), and TN (discharge water). TP (effluent) is the total concentration of phosphorus compounds, such as phosphorus compounds and organic phosphorus, contained in the effluent. TN (effluent) is the total concentration of nitrogen compounds, such as ammonia nitrogen and nitrate nitrogen, contained in the effluent. Examples of manual analysis water quality data include BOD (treated water), NH4-N (treated water), and PO4-P (treated water). BOD (treated water) is a value that indicates the degree of water pollution caused by organic matter in the treated water, and is a value expressed as a concentration of the amount of oxygen consumed when organic matter is oxidized or decomposed by microorganisms. NH4-N (treated water) is the value of the ammonium ions contained in the treated water expressed as the amount of nitrogen, and is an index showing the progression of nitrogen-based pollution. PO4-P (treated water) is the value that expresses the amount of phosphorus in the phosphate ions contained in the treated water, and is an index that shows the progression of phosphorus pollution. Phosphorus compounds containing phosphate are a major factor in eutrophication, so the amount of phosphate ion expressed as the amount of phosphorus is one indicator of pollution.

[0032] The data storage unit 102 stores water quality data obtained by manual analysis at least from the prediction target. When the merged data is generated by the data processing unit 101, the merged data is stored as water quality data.

[0033] The learning data acquisition unit 103 acquires data to be used for learning from the data stored in the data storage unit 102, and generates learning data. The prediction model parameter update unit 104 uses the learning data acquired from the learning data acquisition unit 103 to update the parameters of the prediction model and generate updated prediction model parameters. The prediction model parameter storage unit 105 stores the updated prediction model parameters from the prediction model parameter update unit 104. The prediction target data acquisition unit 106 acquires explanatory variable data to be used for prediction from the data stored in the data storage unit 102 . The treated water quality prediction unit 107 uses the explanatory variable data of the prediction target acquired from the prediction target data acquisition unit 106 and the updated prediction model parameters stored in the prediction model parameter memory unit 105 to perform time evolution using a neural ordinary differential equation to derive the treated water quality at a future time.

[0034] First, the difference between a normal recurrent neural network and the neural ordinary differential equation method will be explained.

[0035] In conventional time series prediction using recurrent neural networks, information at each time point is stored as a state, and the relationship between states is learned by the neural network. In contrast, in the neural ordinary differential equation method, a function that represents the time change of a state is constructed using a neural network. Therefore, when a recurrent neural network is applied to data in which the time intervals between states are not uniform, the same learning parameters will be generated for both relationships between states with long time intervals and relationships between states with short time intervals, resulting in the construction of a predictive model that is inconsistent with respect to changes in states over time. On the other hand, when using neural ordinary differential equations, a function that represents the time change of a state, rather than the relationship between states itself, is constructed in a neural network, so that learning parameters can be generated without depending on the time interval between states. Furthermore, once a function that represents the time change of a state is generated, it is possible to derive a state at any future time by evolving the state over time according to the generated function. In other words, if we express the above in a formula, in neural ordinary differential equations, when the state is represented by x, f that satisfies the following differential equation (1) is found using a neural network, and when making a prediction, this f is followed and time is evolved using integral calculations as shown in the following equation (2), making it possible to derive the state x(t0) at time t1 from the initial state x(t0) at time t0.

[0036]

number

[0037]

number

[0038] Next, the learning process in the neural ordinary differential equation, that is, the prediction parameter update process, will be described. In general, neural network training is performed by repeatedly executing forward processing and backward processing while adding training data. Here, in the forward process, a predicted value is derived. In addition, in the backward processing, the weighting coefficients of the neural network are updated so as to minimize the error between the predicted value and the actual measured value.

[0039] Similarly, in the learning process for neural ordinary differential equations, weighting coefficients are determined so as to minimize errors while repeating forward and backward processes. However, in the learning process for neural ordinary differential equations, an ordinary differential solver is used in both forward processing and backward processing. An ordinary differential solver is a function that derives the result of time evolution according to a given function from the initial state.

[0040] FIG. 3 shows an example of the simplest ordinary differential solver. FIG. 3 shows an ordinary differential solver s(x0, t0, t1, f) that evolves an initial state x0 at time t0 to time t1 according to a function f. As ordinary differential solvers different from the example shown in FIG. 3, various methods such as the Runge-Kutta method have been proposed. In the present invention, however, an appropriate method may be selected as appropriate, and the ordinary differential solver of the present invention is not limited to the example shown in FIG. 3.

[0041] FIG. 4 is a diagram showing an image of the forward process. In forward processing, an ordinary differential solver is used to evolve the input (initial state) over time for the destination time according to the neural network f[w] generated at that time. where w is a parameter of the neural network. In this way, once learning is completed and a prediction model is constructed, it is possible to derive prediction results with the time advanced by any amount.

[0042] Next, the backward processing will be described below. Generally, in backward processing in a neural network, the parameters of the neural network are updated by applying the gradient descent method to the error between the prediction result by the forward processing of the neural network and the actual measurement. For example, if the actual measured value of the objective variable for the input x is y, the predicted value of the neural network is Φ[w](x), and the error is measured as the squared error, then the error function L(w) is expressed by the following equation (3).

[0043]

number

[0044] In this case, the parameter update equation using the gradient descent method is expressed by the following equation (4).

[0045]

number

[0046] Here, t represents the number of updates to the learning parameters, and w t And w t+1 The learning is terminated when the error falls below a certain value. The learning rate η is a parameter that adjusts how much the parameters are updated in one update. Thus, once the derivative of the error function L with respect to the neural network parameters w has been calculated, the backward process can be performed. However, in the case of neural ordinary differential equations, forward processing is performed by time evolution using an ordinary differential solver, so it is not possible to obtain the differential coefficient expressed by the following equation (5) using the parameter w of the error function L in the form of a differential function as in a normal neural network.

[0047]

number

[0048] Therefore, in the case of neural ordinary differential equations, an ordinary differential solver is used to develop time backwards to calculate the differential coefficient of the above equation (5), which is the error between the predicted result based on the parameter w and the actual measurement. Let z(t) be the state at time t, and the neural network at that time be expressed as f(·, w) using parameter w, that is, let us say that the following equation (6) holds.

[0049]

number

[0050] At this time, in order to calculate the differential coefficient expressed by the above equation (5), the Pontryagin adjoint vector a(t) shown in the following equation (7) is introduced.

[0051]

number

[0052] In this case, the following equations (8) and (9) hold for the adjoint vector a(t).

[0053]

number

[0054]

number

[0055] Here, once the second equation, which is the differential coefficient of the error function L with respect to the parameter w, is found, backward processing can be performed using the gradient descent method. This can be expressed by b(t) shown in equation (10) below.

[0056]

number

[0057] Here, if z(t0), a(t0), and b(t0) at time t0 are written side by side based on the above equations (6), (8), and (10), they are expressed by the following equation (11).

[0058]

number

[0059] Here, attention is paid to the right side of the above equation (11). a(t0) can be found by substituting z(τ) into the integrand function expressed by the following equation (12). The integrand function of b(t0) can be found by substituting z(τ) into a(τ) and the equation expressed by the following equation (13).

[0060]

number

[0061]

number

[0062] Therefore, by going back in time in small increments using an ordinary differential solver for the three integrand functions shown in equation (14) below, and finding z(τ), a(τ), and b(τ) in that order, and evolving the time from time t1 to time t0, we can finally obtain the three sets at time t0 shown in equation (15) below.

[0063]

number

[0064]

number

[0065] In the above equation (15), b(t0) is the derivative of the error function L with respect to the parameter w, and is expressed by the following equation (16). Therefore, the parameters of the neural network can be updated using the gradient descent method, and backward processing is performed.

[0066]

number

[0067] Next, the water quality prediction method according to this embodiment will be described. FIG. 5 is a flowchart showing the learning process and the prediction process in this embodiment. In the learning process, first, the data processing unit 101 processes the real-time measurement data acquired by the measuring instrument and the manual analysis water quality data acquired by manual analysis into a form that can be used for learning the prediction model (data processing S11). Since the target of prediction is the treated water quality based on manual analysis, the real-time measurement data continuously acquired by the measuring instrument is processed to match the frequency of acquisition of the manual analysis water quality data, for example by taking the average over a certain period of time, and then matched with the manual analysis water quality data. If necessary, normalization processing etc. is performed on the associated real-time measurement data and manual analysis water quality data. Depending on the data, a processing method may be determined according to the distribution of the data, such as setting a threshold and binarizing the data according to size. The data storage unit 102 stores the data from the data processing unit 101 so that it can be used in the learning process and the prediction process (data storage S12). The learning data acquisition unit 103 acquires data required for constructing a prediction model from the data stored in the data storage unit 102 (learning data acquisition S13). The prediction model parameter update unit 104 updates the prediction model of the neural ordinary differential equation using the acquired data (prediction model parameter update S14). Here, the details of the update logic are as described above. The prediction model parameter storage unit 105 stores and accumulates the updated prediction model parameters so that they can be acquired at the time of prediction (prediction model parameter accumulation S15).

[0068] In the prediction process, first, the prediction target data acquisition unit 106 acquires data of the prediction target from the data stored in the data storage unit 102 (obtain prediction target data S21). The treated water quality prediction unit 107 acquires the learned prediction model parameters (prediction parameter model acquisition S22), and uses the acquired prediction model parameters to input the acquired prediction target data and derive the treated water quality at the future time to be predicted by evolving it over time using a neural ordinary differential equation, thereby predicting the treated water quality (treated water quality prediction S23).

[0069] According to this embodiment, even if the data acquisition intervals are not fixed, it is possible to construct a model for predicting treated water quality with a consistent formulation without performing arbitrary data interpolation.

[0070] <Embodiment 2> In this embodiment, a technology is provided that enables predictions that take into account sudden changes in water quality, which are important for operational decisions, by predicting the presence or absence of sudden changes in water quality as a classification problem before predicting the treated water quality of the target. Since water quality data contains a lot of data from normal times, prediction results tend to be those from normal times with little fluctuation. In this embodiment, water quality fluctuations are predicted as a separate model and input when predicting treated water quality, making it possible to make predictions that take into account sudden changes in water quality, which are important in operational decisions. In the description of this embodiment, only the differences from the first embodiment will be described, and the description of the first embodiment will be used for the points that are not specifically described.

[0071] FIG. 6 is a diagram for explaining an outline of the prediction process in this embodiment. When predicting the treated water quality, in addition to the real-time measurement data and the manual analysis water quality data, the state of the predicted water quality fluctuations, for example, a sudden rise, no change, or a sudden drop of a parameter, is input. However, in the present invention, the state of the water quality change prediction result is not limited to a sudden rise, no change, or a sudden fall, and the state of the water quality change prediction result is not limited to three.

[0072] FIG. 7 is a functional block diagram showing the configuration of a water quality prediction device 100a according to this embodiment. The water quality prediction device 100a shown in Figure 7 includes a data processing unit 101, a data storage unit 102, a treated water quality prediction model learning data acquisition unit 103a, a treated water quality prediction model parameter update unit 104a, a treated water quality prediction model parameter memory unit 105a, a water quality fluctuation prediction model learning data acquisition unit 103b, a water quality fluctuation prediction model parameter update unit 104b, a water quality fluctuation prediction model parameter memory unit 105b, a prediction target data acquisition unit 106, a treated water quality prediction unit 107, and a water quality fluctuation prediction unit 108. That is, the water quality prediction device 100a shown in Figure 7 differs from the water quality prediction device 100 shown in Figure 2 in that it has a treated water quality prediction model learning data acquisition unit 103a and a water quality fluctuation prediction model learning data acquisition unit 103b instead of the learning data acquisition unit 103, a treated water quality prediction model parameter update unit 104a and a water quality fluctuation prediction model parameter update unit 104b instead of the prediction model parameter update unit 104, and a treated water quality prediction model parameter memory unit 105a and a water quality fluctuation prediction model parameter memory unit 105b instead of the prediction model parameter memory unit 105.

[0073] Next, the water quality prediction method according to this embodiment will be described. FIG. 8 is a flowchart showing the learning process and the prediction process in this embodiment. In the learning process, first, the data processing unit 101 processes the real-time measurement data acquired by a measuring instrument and the water quality data acquired by manual analysis into a form that can be used for learning by the water quality change prediction model and the treated water quality prediction model (data processing S31). For the water quality change prediction model, the objective variable is processed into three categories as learning data: sudden increase, no change, and sudden decrease. Here, the treated water quality to be predicted is processed into three categories: if the difference from the previous manual analysis water quality data is above a threshold, it is considered to have risen sharply; if it is below the threshold, it is considered to have fallen sharply; otherwise, it is considered to have remained unchanged. Since the target of prediction is the treated water quality based on manual analysis, the real-time measurement data continuously acquired by the measuring instrument is processed to match the frequency of acquisition of the manual analysis water quality data, for example by taking the average over a certain period of time, and then matched with the manual analysis water quality data. If necessary, normalization processing etc. is performed on the associated real-time measurement data and manual analysis water quality data. Depending on the data, a processing method may be determined according to the distribution of the data, such as setting a threshold and binarizing the data according to size. The data storage unit 102 stores the data from the data processing unit 101 so that it can be used in the learning process and the prediction process (data storage S32). The treated water quality prediction model training data acquisition unit 103a acquires data required for constructing a treated water quality prediction model from the data accumulated in the data accumulation unit 102, and the water quality change prediction model training data acquisition unit 103b acquires data required for constructing a treated water quality prediction model from the data accumulated in the data accumulation unit 102 (learning target data acquisition S33). The treated water quality prediction model parameter update unit 104a updates the treated water quality prediction model based on neural ordinary differential equations using the acquired data, and the water quality change prediction model parameter update unit 104b updates the water quality change prediction model based on neural ordinary differential equations using the acquired data (prediction model parameter update S34). Here, the details of the update logic are as explained in the first embodiment. The treated water quality prediction model parameter memory unit 105a stores and accumulates the updated treated water quality prediction model parameters so that they can be acquired at the time of prediction, and the water quality fluctuation prediction model parameter memory unit 105b stores and accumulates the updated treated water quality prediction model parameters so that they can be acquired at the time of prediction (prediction model parameter accumulation S35).

[0074] In this embodiment, a water quality fluctuation prediction process is performed before the treated water quality prediction. In the water quality fluctuation prediction process, first, the prediction target data acquisition unit 106 acquires data of the water quality fluctuation prediction target and data of the treated water quality prediction target from the data accumulated in the data accumulation unit 102 (acquisition of prediction data S41). The water quality fluctuation prediction unit 108 acquires the learned water quality fluctuation prediction model parameters (fluctuation prediction parameter model acquisition S42), and uses the acquired water quality fluctuation prediction model parameters to derive water quality fluctuations at future times by developing the water quality fluctuations over time using a neural ordinary differential equation as input using the water quality fluctuation prediction target data acquired by the prediction target data acquisition unit 106 (treated water quality fluctuation prediction S43). The treated water quality prediction unit 107 acquires the learned treated water quality prediction model parameters (treated water quality prediction parameter model acquisition S44), and uses the acquired treated water quality prediction model parameters to input the data for the treated water quality prediction target acquired by the prediction target data acquisition unit 106, and derives the treated water quality at a future time by evolving it over time using a neural ordinary differential equation (treated water quality prediction S45).

[0075] According to this embodiment, in addition to the effect of embodiment 1 that a model can be constructed to predict treated water quality with a consistent formulation without arbitrary data interpolation, even if the data acquisition interval is not fixed, it is also possible to make predictions that take into account sudden fluctuations in water quality, which are important in operational decisions.

[0076] It should be noted that the present invention is not limited to the above-described embodiment, but includes various modified examples in which components are added, deleted, or converted from the above-described configuration. [Explanation of symbols]

[0077] 1 Primary sedimentation tank 2. Reactor 3 Final Sedimentation Tank 4. Chlorine mixing tank 5. Blower 6 Adjustment valve 7 Air diffuser 8. First Pump 9. Second Pump 10 Gravity thickener 11 Mechanical thickener 12 Digestion tank 13 Dehydration tank 21, 22, 23, 24, 25 Piping 100,100a Water quality prediction device 101 Data Processing Department 102 Data Storage Unit 103 Learning data acquisition unit 103a Data acquisition unit for learning treated water quality prediction model 103b Data acquisition unit for learning water quality change prediction model 104 Prediction model parameter update unit 104a Treated water quality prediction model parameter update section 104b Water quality change prediction model parameter update section 105 Prediction model parameter storage unit 105a Treated water quality prediction model parameter memory unit 105b Water quality change prediction model parameter memory unit 106 Prediction target data acquisition unit 107 Treated Water Quality Prediction Department 108 Water Quality Change Prediction Division 200 Water treatment facilities

Claims

1. a data storage unit that stores water quality data obtained by manual analysis at least from the prediction target; a learning data acquisition unit that acquires data to be used for learning from the data stored in the data storage unit and generates learning data; a prediction target data acquisition unit that acquires explanatory variable data to be used for prediction from the data stored in the data storage unit; a prediction model parameter update unit that updates parameters of a prediction model by using the learning data acquired from the learning data acquisition unit to generate updated prediction model parameters; a prediction model parameter storage unit that stores the updated prediction model parameters from the prediction model parameter update unit; A water quality prediction device comprising: a treated water quality prediction unit that derives the treated water quality at a future time by evolving the time using a neural ordinary differential equation using the explanatory variable data of the prediction target acquired from the prediction target data acquisition unit and the updated prediction model parameters stored in the prediction model parameter memory unit.

2. The present invention further includes a data processing unit that processes real-time measurement data acquired in real time by a measuring instrument installed in the prediction target into a representative value for a predetermined time period, and generates merged data by merging the representative value of the real-time measurement data with manual analysis water quality data of the prediction target, The water quality prediction device according to claim 1 , wherein the data storage unit stores the merged data as the water quality data.

3. a data storage unit that stores water quality data obtained by manual analysis at least from the prediction target; a treated water quality prediction model learning data acquisition unit that acquires data to be used for learning from the data stored in the data storage unit and generates treated water quality prediction model learning data; a prediction target data acquisition unit that acquires explanatory variable data used for predicting water quality fluctuations and predicting treated water quality from the data stored in the data storage unit; a treated water quality prediction model parameter update unit that uses the treated water quality prediction model learning data acquired from the treated water quality prediction model learning data acquisition unit to update parameters of the treated water quality prediction model to generate updated treated water quality prediction model parameters; a treated water quality prediction model parameter storage unit for storing the updated treated water quality prediction model parameters from the treated water quality prediction model parameter update unit; a water-quality-change-prediction-model learning data acquisition unit that acquires data to be used for learning from the data stored in the data storage unit and generates water-quality-change-prediction-model learning data; a water-quality-variation-prediction-model-parameter update unit that uses the water-quality-variation-prediction-model learning data acquired from the water-quality-variation-prediction-model-learning-data acquisition unit to update parameters of the water-quality-variation-prediction-model and generate updated water-quality-variation-prediction-model parameters; a water-quality-fluctuation-prediction-model-parameter storage unit that stores the updated water-quality-fluctuation-prediction-model parameters from the water-quality-fluctuation-prediction-model-parameter update unit; a water quality fluctuation prediction unit that derives water quality fluctuations using explanatory variable data of the prediction target acquired from the prediction target data acquisition unit and updated water quality fluctuation prediction model parameters stored in a water quality fluctuation prediction model parameter storage unit; A water quality prediction device comprising a treated water quality prediction unit that derives the treated water quality at a future time by evolving the time using a neural ordinary differential equation using the explanatory variable data of the prediction target acquired from the prediction target data acquisition unit, the treated water quality prediction model parameters stored in the treated water quality prediction model parameter memory unit, and the water quality fluctuation data acquired from the water quality fluctuation prediction unit.

4. A computer comprising: Accumulating water quality data obtained by manual analysis at least from the prediction target; acquiring data to be used for learning from the accumulated data to generate learning data; acquiring explanatory variable data to be used for prediction from the accumulated data; updating parameters of a predictive model using the training data to generate updated predictive model parameters; storing the updated predictive model parameters; A water quality prediction method which uses the acquired explanatory variable data of the prediction target and the updated prediction model parameters to derive the treated water quality at a future time by performing time evolution using a neural ordinary differential equation.

Images