Water treatment plant control system, water treatment plan control method, and water treatment plant control program

Abstract

To propose a water treatment plant control system, a water treatment plant control method, and a water treatment plant control program that are capable of preparing for a possible occurrence of a similar abnormal weather phenomenon at another site for an abnormal weather phenomenon that occurred at one site.SOLUTION: A plant behavior prediction model is divided into a natural phenomenon prediction model and an artifact behavior prediction model, and distributions of learning data and prediction data are taken for each of the models during initial learning and prediction. Next, a distribution difference in prediction data between an own site and another site is calculated. When the distribution difference is large, a natural phenomenon prediction model that has been already learned at the other site the artifact behavior prediction model for the own site are coupled together and transfer learning is performed.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a water treatment plant control system, a water treatment plant control method, and a water treatment plant control program related to a plant behavior prediction technology among plant equipment control technologies, and is suitable for application to a water treatment plant control system that utilizes prediction technology for plant equipment in fields such as water supply, sewerage, flood control, and water intake, where input to the plant is affected by natural phenomena. [Background technology]

[0002] Conventionally, in the operation of plant facilities, it is common to appropriately control the operating state of each device that constitutes the plant. In plants where the input to the plant can be controlled artificially, such as production plants, it is common to control the amount of input (raw materials in the case of production) input to the plant and automatically control the plant's constituent devices according to the input amount. Furthermore, in conventional plants, not only are the behavioral performance values ​​of each device in the plant fed back for control, but also the behavior within the plant is predicted using a predictive model such as machine learning, thereby controlling the plant with less delay. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent Publication No. 2022-28338 Summary of the Invention [Problem to be solved by the invention]

[0004] On the other hand, in plants where the input to the plant cannot be controlled artificially, such as water purification plants, sewage treatment plants, and storm water pumping stations, and where the input fluctuates greatly (hereinafter, abbreviated as "water treatment plants"), although automatic control of each device is performed, it is common for operators to manually intervene in the results of the automatic control according to individual circumstances. When using a known machine learning-based plant behavior prediction technology for such water treatment plants, there is a problem that learning data that is not suitable for update learning is mixed in because manual intervention according to individual circumstances is inserted into the plant control loop, which is the source of the update learning data for the prediction model.

[0005] In addition, in the case of water treatment plants, many similar sites are installed in each area, such as within a local government, but because the equipment configuration and processing capacity of each site differ, training data cannot be reused between sites, and training data must be collected separately for each site. This issue is particularly serious in water treatment plants that are affected by abnormal weather phenomena such as typhoons and heavy rains, and because the limited abnormal weather data cannot be used between multiple sites, there is a problem that the plant control algorithm cannot respond to abnormal weather that has not yet been learned.

[0006] The present invention has been made in consideration of the above points, and aims to propose a water treatment plant control system, a water treatment plant control method, and a water treatment plant control program that can prepare in advance for the possibility that a similar abnormal weather phenomenon that occurs at one site may occur at another site. [Means for solving the problem]

[0007] In order to solve such problems, the present invention provides an initial learning processing unit that uses learning data from natural phenomenon sensor data that measures natural phenomena to learn a natural phenomenon prediction model that predicts future natural phenomena, and that uses learning data from artifact sensor data that measures the behavior of artifacts to learn an artifact behavior prediction model that predicts future artifact behavior, and calculates a distribution of the learning data, and a prediction processing unit that combines the learned natural phenomenon prediction model and the learned artifact behavior prediction model to generate a prediction model, predicts artifact behavior using prediction data from the natural phenomenon sensor data and the artifact sensor data using the prediction model, and calculates a distribution of prediction data for each of the natural phenomenon sensor data and the artifact sensor data. an intra-site additional learning processing unit that compares a distribution of the learning data with a distribution of the prediction data, and performs additional learning using the prediction data belonging to an area where the distribution of the learning data and the distribution of the prediction data do not match as additional learning data for the site; and an intra-site transfer learning processing unit that compares a distribution of the learning data for a first site with a distribution of the learning data for a second site, and for the prediction data belonging to an area where the distribution of the learning data for the first site and the distribution of the learning data for the second site do not match, selects the trained natural phenomenon prediction model for the second site, and connects the trained natural phenomenon prediction model for the second site and the artificial object behavior prediction model for the first site to perform transfer learning.

[0008] Also, in the present invention, there is provided an initial learning processing step in which an initial learning processing unit uses learning data from among natural phenomenon sensor data that measures a natural phenomenon to learn a natural phenomenon prediction model that predicts future natural phenomena, while using learning data from among artifact sensor data that measures the behavior of an artifact to learn an artifact behavior prediction model that predicts future artifact behavior, and calculates a distribution of the learning data; a prediction processing unit combines the learned natural phenomenon prediction model with the learned artifact behavior prediction model to generate a prediction model, predicts artifact behavior using prediction data from among the natural phenomenon sensor data and the artifact sensor data using the prediction model, and calculates a distribution of prediction data for each of the natural phenomenon sensor data and the artifact sensor data; The method includes an intra-site additional learning processing step in which a learning processing unit compares a distribution of the learning data with a distribution of the prediction data, and performs additional learning using prediction data belonging to an area where the distribution of the learning data and the distribution of the prediction data do not match as additional learning data for the site, and an inter-site and other site transfer learning processing step in which a transfer learning processing unit compares a distribution of the learning data for a first site with a distribution of the learning data for a second site, and for prediction data belonging to an area where the distribution of the learning data for the first site and the distribution of the learning data for the second site do not match, a trained natural phenomenon prediction model for the second site is selected, and the trained natural phenomenon prediction model for the second site and the artificial object behavior prediction model for the first site are connected to perform transfer learning.

[0009] Furthermore, in the present invention, there is provided an initial learning processing step of having an initial learning processing unit learn a natural phenomenon prediction model that predicts future natural phenomena using learning data from natural phenomenon sensor data that measures natural phenomena, and also learn an artifact behavior prediction model that predicts future artifact behavior using learning data from artifact sensor data that measures artifact behavior, and calculate a distribution of the learning data; a prediction processing step of having a prediction processing unit generate a prediction model by combining the learned natural phenomenon prediction model and the learned artifact behavior prediction model, predict artifact behavior using prediction data from the natural phenomenon sensor data and the artifact sensor data using the prediction model, and calculate a distribution of prediction data for each of the natural phenomenon sensor data and the artifact sensor data; an intra-site additional learning processing step of comparing a distribution of the learning data with a distribution of the prediction data, and using the prediction data belonging to an area where the distribution of the learning data and the distribution of the prediction data do not match as additional learning data for the site for additional learning; and an inter-site transfer learning processing step of causing an inter-site transfer learning processing unit to compare a distribution of the learning data for a first site with a distribution of the learning data for a second site, and for the prediction data belonging to an area where the distribution of the learning data for the first site and the distribution of the learning data for the second site do not match, causing the inter-site transfer learning processing unit to select a trained natural phenomenon prediction model for the second site, and then connecting the trained natural phenomenon prediction model for the second site and the artificial object behavior prediction model for the first site to perform transfer learning. Effect of the Invention

[0010] According to the present invention, it is possible to realize a water treatment plant control system, a water treatment plant control method, and a water treatment plant control program that can prepare in advance for the possibility that a similar abnormal weather phenomenon that occurs at one site may occur at another site. [Brief description of the drawings]

[0011] [Figure 1] 1 is a block diagram showing a schematic configuration of a water treatment plant control system according to a first embodiment. [Diagram 2] FIG. 1 is a diagram showing an example of application of this embodiment to inflow prediction of a storm water pumping station. [Diagram 3] FIG. 1 is a diagram showing an example of an overall overview when this embodiment is applied to a storm water pumping station. [Figure 4] FIG. 2 is a diagram illustrating an example of a data structure of rainfall data. [Diagram 5] FIG. 13 is a diagram illustrating an example of a data structure of water level data. [Figure 6] FIG. 13 is a diagram illustrating an example of the structure of distribution data. [Figure 7] 13 is a flowchart illustrating an example of a procedure for an initial learning process. [Figure 8] 13 is a flowchart illustrating an example of a procedure for a prediction process. [Figure 9] 13 is a flowchart showing an example of a procedure for an additional learning process within one's own site. [Figure 10] FIG. 13 is a diagram showing an example of an additional learning screen within one's own site. [Figure 11] 13 is a flowchart illustrating an example of a procedure for a self-other-site transfer learning process. [Figure 12] FIG. 13 is a diagram illustrating an example of a self-site / other-site transfer learning screen. [Figure 13] FIG. 11 is a diagram showing an example in which the water treatment plant control system according to the second embodiment is applied to water quality prediction at a water purification plant. [Figure 14] FIG. 11 is a block diagram showing an example of a schematic system configuration when a water treatment plant control system according to a second embodiment is applied to water quality prediction at a water purification plant. [Figure 15] FIG. 11 is a diagram showing the data structure of water quality data in the second embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0012] An embodiment of the present invention will now be described in detail with reference to the drawings.

[0013] (1) First embodiment (1-1) Basic system configuration FIG. 1 is a block diagram showing a schematic configuration of a water treatment plant control system 100 according to the first embodiment.

[0014] The water treatment plant control system 100 according to the first embodiment includes sensor data D100 including natural phenomenon sensor data D110, artifact sensor data D120, and distribution data D130, a prediction model M100 including a natural phenomenon prediction model M110 and an artifact behavior prediction model M120, a control unit 300, and a storage device (not shown) that stores the data and the model. The water treatment plant control system 100 is a computer system including a display device that displays a screen described below, and an input device that accepts input from a user.

[0015] The control unit 300 includes an initial learning processing unit 100S, a prediction processing unit 200S, an in-house site additional learning processing unit 300S, and an in-house / other site transfer learning processing unit 400S. The initial learning processing unit 100S, the prediction processing unit 200S, the in-house site additional learning processing unit 300S, and the in-house / other site transfer learning processing unit 400S correspond to each function of the water treatment plant control program that operates on the above-mentioned computer system.

[0016] In this embodiment, learning and prediction of various behaviors by machine learning are assumed, and the model refers to statistical model data. For example, when multiple regression analysis is used, the statistical model data corresponds to weight coefficient data of explanatory variables of a multiple regression equation, and for example, when a neural network is used, the statistical model data corresponds to weight coefficient data of each layer and each node inside the network. The statistical model data is stored in the storage device, similar to the sensor data D100. In the following description, the statistical model data is also simply referred to as a "model".

[0017] In this embodiment, it is assumed that plants having the same purpose, such as water purification plants, are scattered at multiple locations. Therefore, in the following, each individual installation location (facility) of a group of these plants will be referred to as a "site."

[0018] The water treatment plant control system 100 in this embodiment includes two types of data, sensor data D100 and a prediction model M100, an initial learning processing unit 100S, a prediction processing unit 200S, an in-site additional learning processing unit 300S, and an inter-site transfer learning processing unit 400S.

[0019] The sensor data D100 includes two types of sensor data, natural phenomenon sensor data D110 and artificial object sensor data D120, and distribution data D130 that examines the statistical distribution of each sensor data. These sensor data are stored individually for each site. The details of the data structure of these sensor data will be described later.

[0020] The sensor data D100 is data indicating values ​​measured by a plurality of sensors, for example, of natural phenomena such as weather or the behavior of man-made objects such as piping in a plant.

[0021] The prediction model M100 is a machine learning model for predicting the behavior of a plant. The machine learning method applicable to this embodiment is not limited to a specific method, and any regressor that can be used for future prediction may be used. For example, a multiple regression equation or a regressor based on deep learning may be used. In the following, in this embodiment, a deep learning model will be described as an example.

[0022] The feature quantities (explanatory variables) and true values ​​(objective variables) of the prediction model depend on the behavior of the plant to be predicted. In the following description, the feature quantities and true values ​​(and predicted values) for the machine learning model may be set in any way. For example, when there is one type of sensor, sensors S1, S2, S3, . . . , S10, in the sensor data D100, the values ​​of sensors S1 to S9 may be set as feature quantities to predict the value of sensor S10 (i.e., the value of sensor S10 may be assigned as the objective variable), or the values ​​of sensors S2 to S10 may be set as feature quantities to predict the value of sensor S1.

[0023] The initial learning processing unit 100S performs initial learning of the prediction model M100. At this time, the prediction model is learned using trainable data, which is data for which the true values ​​corresponding to the feature quantities to be taught to the machine learning model are known, from among the natural phenomenon sensor data D110 and the artificial object sensor data D120, and at the same time, the statistical distribution of this learning data is taken and stored as distribution data.

[0024] The initial learning processing unit 100S uses learning data from the natural phenomenon sensor data D110 that measures natural phenomena to learn a natural phenomenon prediction model that predicts future natural phenomena, while using learning data from the artificial object sensor data D120 that measures the behavior of artificial objects to learn an artificial object behavior prediction model that predicts future artificial object behavior, and calculates the distribution of the learning data.

[0025] As a learning data distribution calculation process, the initial learning processing unit 100S calculates the distribution of each of the learning features and learning true values ​​for the learning data used to learn the prediction model, from the natural phenomenon sensor data D110 and the artificial object sensor data D120, and stores the distribution data of the learning data.

[0026] The prediction processor 200S predicts the plant behavior using the trained prediction model M100. The prediction data used at this time is the latest sensor data at the time of prediction, and the true value at the time of prediction is not yet known. Therefore, the prediction processor 200S obtains the statistical distribution of only the feature data and stores it as distribution data. At this stage, it is possible to obtain a prediction using the initial learning model and the distribution of the learning and prediction data.

[0027] The prediction processing unit 200S generates a prediction model by combining a learned natural phenomenon prediction model and a learned artificial object behavior prediction model, and uses this prediction model to predict artificial object behavior using prediction data from the natural phenomenon sensor data and the artificial object sensor data, and calculates the distribution of prediction data for each of the natural phenomenon sensor data and the artificial object sensor data.

[0028] As the prediction data distribution calculation process, the prediction processing unit 200S calculates the distribution of prediction features for prediction data to be used for prediction out of the natural phenomenon sensor data and the artificial object sensor data, and stores the distribution data of the prediction data.

[0029] The in-site additional learning processing unit 300S checks the distribution difference between the distribution of the learning data and the distribution of the prediction data for the learning and prediction data within the in-site. If the distribution difference is large, that is, if an unlearned distribution is found within the in-site, the in-site additional learning processing unit 300S recommends additional learning of the unlearned data for the in-site prediction model to the user of the water treatment plant control system 100, and performs the additional learning.

[0030] The additional learning processing unit 300S within its own site compares the distribution of the learning data with the distribution of the prediction data, and performs additional learning using the prediction data that belongs to an area where the distribution of the learning data and the distribution of the prediction data do not match as additional learning data for the site.

[0031] The in-site additional learning processing unit 300S compares the distribution of features included in the learning data for the first site with the distribution of features included in the prediction data for the first site, and if the difference in distribution is large, displays a message encouraging additional learning of the prediction model for the first site using the prediction data, and if an instruction to perform additional learning is received, sets the actual measurement value obtained after the elapse of the prediction destination time that is the target of prediction of the prediction model as a true value for learning, and performs additional learning of the prediction model for the first site using the features included in the prediction data and the true value for learning.

[0032] The own-and-other-site transfer learning processing unit 400S examines the distribution difference between the learning data of the own site and the other site. If the distribution difference is large, that is, if a distribution that has been learned at the other site but not yet learned at the own site is found, the own-and-other-site transfer learning processing unit 400S recommends transfer learning of the own-site prediction model that divertes at least a part of the other site's prediction model to the user of the water treatment plant control system 100, and executes transfer learning.

[0033] The self-other site transfer learning processing unit 400S compares the distribution of the learning data for the first site with the distribution of the learning data for the second site, and for prediction data that belongs to an area where the distribution of the learning data for the first site and the distribution of the learning data for the second site do not match, selects a trained natural phenomenon prediction model for the second site, connects the trained natural phenomenon prediction model for the second site with the artificial object behavior prediction model for the first site, and performs transfer learning.

[0034] The self-other-site transfer learning processing unit 400S compares the distribution of features and true values ​​included in the learning data for the first site with the distribution of features and true values ​​included in the learning data for the second site, and if the difference in distribution is large, displays a message encouraging transfer learning to the prediction model for the first site using the natural phenomenon prediction model for the second site, and if an instruction to perform transfer learning is received, combines the natural phenomenon prediction model for the second site with the artificial object behavior prediction model for the first site, and performs transfer learning using the learning data for the first site.

[0035] The water treatment plant control system 100 according to this embodiment executes a water treatment plant control method, and this water treatment plant control method includes an initial learning processing step in which an initial learning processing unit 100S uses learning data from natural phenomenon sensor data D110 that measures a natural phenomenon to learn a natural phenomenon prediction model that predicts future natural phenomena, while using learning data from artificial object sensor data D120 that measures the behavior of an artificial object to learn an artificial object behavior prediction model that predicts future artificial object behavior, and calculates a distribution of the learning data; and a prediction processing unit 200S combines the learned natural phenomenon prediction model and the learned artificial object behavior prediction model to generate a prediction model, predicts artificial object behavior using prediction data from the natural phenomenon sensor data D110 and the artificial object sensor data D120 using the prediction model, and calculates a distribution of the learning data. D120, an additional learning processing step within the own site in which an additional learning processing unit 300S compares the distribution of the learning data with the distribution of the prediction data, and performs additional learning using the prediction data belonging to an area where the distribution of the learning data and the distribution of the prediction data do not match as additional learning data for the site, and an internal / other site transfer learning processing step in which an internal / other site transfer learning processing unit 400S compares the distribution of the learning data for the first site with the distribution of the learning data for the second site, and for the prediction data belonging to an area where the distribution of the learning data for the first site and the distribution of the learning data for the second site do not match, a trained natural phenomenon prediction model for the second site is selected, and the trained natural phenomenon prediction model for the second site and the artificial object behavior prediction model for the first site are connected to perform transfer learning.

[0036] In addition, the water treatment plant control program of the water treatment plant control system 100 according to this embodiment causes the initial learning processing unit 100S to execute the above-mentioned initial learning processing step, the prediction processing unit 200S to execute the above-mentioned prediction processing step, the in-house site additional learning processing unit 300S to execute the above-mentioned in-house site additional learning processing step, and the between-own and other sites transfer learning processing unit 400S to execute the between-own and other sites transfer learning processing step in a computer.

[0037] (1-2) Application to inflow prediction for storm water pumping stations Fig. 2 shows an example of application of this embodiment to inflow prediction for a stormwater pumping station. A stormwater pumping station is a plant that uses a pump to forcibly drain stormwater collected through a pipe network into a river or the like in order to prevent inland flooding. Inland flooding here refers to flooding that can occur when natural drainage through the land cannot cope with the flooding, such as when heavy rain occurs in an area with a high paved area such as an urban area and a stormwater sewer pipe network.

[0038] When this embodiment is applied to predicting the inflow of a storm water pumping station, the natural phenomenon prediction model M110 corresponds to a rainfall model, and the artificial object behavior prediction model M120 corresponds to a pipe network water level model.

[0039] In this embodiment, the first site and the second site are storm water pumping stations, and the prediction processing unit 200S inputs rainfall data as natural phenomenon sensor data into a rainfall model serving as a natural phenomenon prediction model M110 as input for the prediction model, and inputs water level data as artificial object sensor data into a pipe network water level model serving as an artificial object behavior prediction model M120, and combines the output of the rainfall model and the output of the pipe network water level model as output for the prediction model to predict the amount of rainwater inflow into the storm water pumping station.

[0040] In this embodiment, rainfall enters the pipe network 210, flows into the pump well 220 of the pump station 200, and is forcibly discharged into the river 240 using the stormwater pump 230. For this reason, in this embodiment, the feature values ​​input to the rainfall model are rainfall sensor mesh data, and the feature values ​​input to the pipe network water level model are pipe network water level sensor data and the output of the rainfall model. The true value taught to the entire prediction model combining the rainfall model and the pipe network water level model is the inflow sensor data of the pump well.

[0041] Fig. 3 shows an example of the overall overview when this embodiment is applied to a storm water pumping station. This embodiment is applicable to all water treatment plants that are subject to plant input disturbances due to natural phenomena. Therefore, even when this embodiment is applied to a storm water pumping station, the overall data configuration and processing configuration are the same as those of the first embodiment.

[0042] When this embodiment is applied to a storm water pumping station, the natural phenomenon sensor data D110 corresponds to rainfall data, and the artificial object sensor data D120 corresponds to water level data. Similarly, for the prediction model, the natural phenomenon prediction model M110 corresponds to a rainfall model, and the artificial object behavior prediction model corresponds to a pipe network water level model.

[0043] (1-3) Data structure of rainfall data FIG. 4 shows an example of the data structure of rainfall data. The rainfall data D110 corresponds to actual sensor information obtained by observing rainfall conditions using a rainfall radar or the like, or rainfall forecast information calculated based on the sensor information. This rainfall data D110 is two-dimensional mesh information that corresponds one-to-one with the map. Therefore, mesh information 410 consisting of w×h cells composed of the X-axis and Y-axis is three-dimensional information collected for multiple times at time t=t0 to t0-1, t0-2, . . . , t0-n on the time axis. In other words, the interval of the time axis corresponds to the sampling rate of the rainfall mesh. Each cell that composes the mesh indicates the amount of rainfall within a predetermined vertical and horizontal size. The vertical and horizontal size is, for example, 100M×100M.

[0044] (1-4) Data structure of water level data 5 shows an example of the data structure of the water level data D210. The water level data D210 has a two-dimensional table structure in which the vertical axis is the time axis and the horizontal axis is the pipe network water level sensor. The interval of the vertical time axis corresponds to the sampling rate of the pipe network water level sensor.

[0045] On the other hand, the horizontal axis shows the values ​​of water level sensors installed at multiple locations in the pipeline. The sensor group also includes the pump well inflow, which is the true value to be predicted. Therefore, we will build a prediction model in which the water level sensor values ​​from this sensor group are used as features and the pump well inflow is used as the true value.

[0046] (1-5) Data structure of distribution data 6 shows an example of the structure of the distribution data D130. The distribution data D130 is calculated by the initial learning processing unit 100S and the prediction processing unit 200S. The distribution data D130 is information on the distribution of the rainfall data D110 and the water level data D120. The distribution data D130 includes distribution data 600 of the rainfall data D110 and distribution data 610 of the water level data D120.

[0047] Both the rainfall data and the water level data have multiple samples in the time axis direction, and therefore show distribution information on the sample group included in the entire interval of those samples (or the sample group for the period specified on the intra-site additional learning screen and intra-site transfer learning screen described later). In this embodiment, basic statistics (maximum value, third quartile value, median value, average value, first quartile value, minimum value, variance, standard deviation) are shown as examples of distribution information, but are not limited to these. For example, by clustering the sample group, the number of samples in each cluster, the position of the cluster center of gravity, intra-cluster variance, inter-cluster variance, the distance between the cluster centers of gravity, etc. can also be used as distribution data.

[0048] (1-6) Initial learning process 7 is a flowchart showing an example of a procedure for the initial learning process. The initial learning process is executed by the initial learning processing unit 100S.

[0049] The initial learning processing unit 100S first reads the rainfall data and the water level data, including the pipeline water level sensor data, as feature data (step S110). Next, the initial learning processing unit 100S reads the pump well inflow data as the true value (step S120). Note that the specific sensor assignments corresponding to the feature values ​​and true values ​​are defined in advance according to the application target of this embodiment, and the sensor assignment pattern is not limited to this.

[0050] Next, the initial learning processing unit 100S calculates the above-mentioned basic statistics for the entire period actually used for initial learning for each of the rainfall data and water level data that have been read (step S130), and records them in the storage device as distribution data.

[0051] The initial learning processing unit 100S performs a learning process (step S140) and records the data in the storage device as learned prediction model data. The data structure of the prediction model data depends on the machine learning method adopted in this embodiment. For example, when a multiple regression equation is used, the partial regression coefficients for each explanatory variable (rainfall sensor and water level sensor) become the prediction model data, and in the case of a neural network, the weight information connecting the nodes included in each layer of the network becomes the prediction model data.

[0052] (1-7) Prediction processing 8 is a flowchart showing an example of a procedure for the prediction process. The prediction process is executed by the prediction processing unit 200S.

[0053] The prediction processing unit 200S first reads the rainfall data and the water level data, which are the pipe network water level sensor data, as feature data (step S210). The pipe network water level sensor data read here has the same data structure as the pipe network water level sensor data executed by the initial learning processing unit 100S, but the pipe network water level sensor data read in the prediction process is a feature used for future prediction, so the read target time is the true value, that is, the sensor data at the time when there is no inflow at the prediction target future time.

[0054] In step S220, the prediction processing unit S100 calculates the basic statistics described above for the feature quantities of the prediction target for each of the read rainfall data and water level data, and records them in the storage device as distribution data.

[0055] The prediction processor 200S performs prediction processing (step S240) and outputs a predicted value. The predicted value output here is used, for example, in another plant equipment control system.

[0056] (1-8) Additional learning process within your own site 9 is a flowchart showing an example of a procedure for the in-own-site additional learning process. The in-own-site additional learning process is executed by the in-own-site additional learning processing unit 300S.

[0057] The in-site additional learning processing unit 300S first compares the distribution data of the learning data with the distribution data of the prediction data (step S310). At this time, since the learning data has true values ​​but the prediction data does not have true values, the distribution difference between the feature amounts, which are comparable data, is compared.

[0058] In this embodiment, basic features are exemplified as distribution information, and more specifically, the distribution difference can be compared by checking whether the range of maximum to minimum values ​​of the training data overlaps with the range of maximum to minimum values ​​of the predicted data, or whether the distribution of the training data and the distribution of the predicted data are close to each other based on information on the mean (or median) and variance.

[0059] In this embodiment, the method of expressing the distribution is not limited. For example, if cluster information is used instead of basic statistics, the distance between cluster centroids or the ratio of intra-cluster variance to inter-cluster variance can be used as an index of distribution difference.

[0060] Next, the in-site additional learning processing unit 300S determines whether the distribution difference is large (step S320). If the in-site additional learning processing unit 300S determines that the distribution difference is not large, it ends the in-site additional learning processing. On the other hand, if it determines that the distribution difference is large, it displays a screen recommending to the user of the water treatment plant control system to perform additional learning of the prediction data with respect to the prediction model of the site, although the details will be described later (steps S320, S330). If there is an input to perform additional learning, the in-site additional learning processing unit 300S executes the additional learning of the prediction data (steps S340 < S350) and records it in the storage device as the learned prediction model data.

[0061] Here, at the time when the prediction data is generated, since the true value data paired with the feature amount does not exist yet, learning is started after waiting for the occurrence of the true value data that is determined after the passage of the prediction destination time. For example, when predicting the inflow volume in the future 5 minutes using rainfall and water level data for the past 30 minutes at the time t = t0 on the time axis, the rainfall and water level sensor data that are the feature amounts are already measured at the time t = t0, but the true inflow volume paired with the feature amount data is not determined until 5 minutes later. Therefore, the operation is to start additional learning 5 minutes after the true value is determined.

[0062] (1-9) In-site additional learning screen FIG. 10 shows an example of the in-site additional learning screen 1000. The in-site additional learning screen 1000 is displayed by the in-site additional learning processing unit 300S and has sample range specifying means 1005, 1010, a sample distribution display area 1015, a guidance display area 1045, additional learning condition specifying means 1050, 1055, and an additional learning execution means 1060. Note that the sample range specifying means may include, for example, numerical input means for sample times such as year, month, day, hour, and minute, and range specifying means using a slide bar or the like.

[0063] In the local site additional learning screen 1000, first, as a default display, the entire learning data is selected as the sample range, and the distribution 1030 of the learning data is displayed in the display area 1015. In this case, in this embodiment, an example is shown in which it is displayed as a two-dimensional distribution map, and the vertical axis and horizontal axis are respectively selected one by one from each sensor in the rainfall data and water level data (pull-down menus 1020, 1025). Note that this embodiment is not limited to this, and other forms are also possible.

[0064] For example, if a separate principal component analysis is applied in the calculation of the distribution, the horizontal axis can be set to the first principal component and the vertical axis to the second principal component, so that even if there are a large number of sensors, it is possible to display the distribution as a two-dimensional distribution. In another example, the distribution of a large number of sensors can be regarded as a multidimensional vector, and a self-organizing map can be applied to the vector to map it into two dimensions, so that it can be displayed in the display area 1015. Furthermore, although the present embodiment shows an example of a two-dimensional distribution display in the display area 1015, it may also be displayed as a three-dimensional cube map.

[0065] By specifying a sample range of the predicted data, distribution areas 1035, 1040 of the predicted data in the specified range are displayed superimposed on the display area 1015 of the sample distribution.

[0066] Next, as a result of executing steps S310 and S320 in the additional learning processing unit 300S within the own site, a distribution area 1040 that is determined to have a large distribution difference is highlighted 1040a, and guidance recommending additional learning using the predicted data for that distribution area 1040 is displayed in the guidance display area 1045.

[0067] For example, when a user performs additional learning according to specified guidance, the user can select the radio button for specifying additional learning conditions 1050 to execute the learning immediately (strictly speaking, immediately after the true value paired with the predicted data feature is determined), or can select the radio button for specifying additional learning conditions 1055 and input a date and time to select whether to schedule execution at the specified date and time, and then press the additional learning execution means 1060, which is an additional learning button, to execute additional learning using the predicted data.

[0068] In this embodiment, the timing of additional learning is specified based on a time such as immediately or a specified date and time as described above, but this embodiment is not limited to this. For example, the timing may be specified based on the following event criteria. For example, as a method of specifying the event criteria, a method of performing additional learning when a peak portion of an inflow increase, peak value, and decrease based on a peak value of the inflow is confirmed may be adopted. More specifically, based on the peak value of the inflow to be the target of additional learning, a section in which the inflow decreases to 5% of the peak value may be set as a confirmed event for the target period of additional learning.

[0069] 11 is a flowchart illustrating an example of a procedure for the self-and-other-site transfer learning process. The self-and-other-site transfer learning process is executed by the self-and-other-site transfer learning processing unit 400S.

[0070] First, the own-site / other-site transfer learning processing unit 400S compares the distribution data of the own-site learning data with the distribution data of the other-site learning data (step S410). Unlike the intra-site additional learning processing by the intra-site additional learning processing unit 300S described above, in the own-site / other-site transfer learning processing, the own-site / other-site transfer learning processing unit 400S compares not only the feature quantities but also the distributions of the true values, since the distribution comparison targets between the own and other sites are all learning data and include true values.

[0071] Next, the transfer learning processing unit between the own and other sites 400S judges whether the distribution difference is large (step S420). If the transfer learning processing unit between the own and other sites judges that the distribution difference is not large, the transfer learning processing between the own and other sites is terminated, whereas if the transfer learning processing unit between the own and other sites judges that the distribution difference is large (step S420), a warning is displayed to the user of the system that data with a large distribution difference has occurred at the other site (for example, a torrential rain that has not been learned at the own site has occurred at a nearby storm water pump station) (step S430). In addition, the transfer learning processing unit between the own and other sites 400S displays a screen that recommends transfer learning to the prediction model of the own site based on the learned prediction model of the other site (step S440). If there is an input to perform transfer learning, the transfer learning is performed by the processing of steps S460 to S480 thereafter.

[0072] Here, the in-site additional learning processor 300S additionally learns data generated within the in-site. Since the configuration of the pipeline is the same before and after the additional learning, the in-site additional learning processor 300S does not need to change the prediction model structure itself, and only needs to add data.

[0073] On the other hand, in the inter-site transfer learning processing unit 400S, since the pipe network configuration and the pipe network water level sensor configuration are different between the inter-site and other sites, the entire prediction model cannot be reused. Therefore, for the rainfall model that deals with natural phenomena (in other words, the same phenomenon can occur at neighboring sites), a model that has already been trained on data with large distribution differences at the other site is reused, and for the pipe network model, a model for the own site that matches the pipe network configuration of the own site is used.

[0074] Specifically, first, the self-other-site transfer learning processing unit 400S loads the precipitation model α as a model of the other site that has already been learned using data with large distribution differences at the other site (step S460), combines the loaded precipitation model α with the self-site pipe network model (step S470), and transfer-learns the self-site forecast model using the precipitation model α as the initial learning value (step S480).

[0075] 12 shows an example of an own-and-other-site transfer learning screen 1200. The own-and-other-site transfer learning screen 1200 is displayed by the own-and-other-site transfer learning processing unit 400S, and includes an own-site sample range designation means 1205, a other-site selection means 1210 to be compared with the own site, a sample distribution display area 1215, a guidance display area 1245, transfer learning execution means 1250, 1255, and a transfer learning execution means 1260. The sample range designation means may include a numerical input means for the sample time such as date, hour, and minute, and a range designation means such as a slide bar.

[0076] In displaying the own-site / other-site transfer learning screen 1200, first, as a default display, the entire own-site learning data is selected as the sample range, and the distribution 1230 of the own-site learning data is displayed in the sample distribution display area 1215. In this case, in this embodiment, an example is shown in which it is displayed as a two-dimensional distribution map, and the vertical axis and horizontal axis are selected one by one from each sensor in the rainfall data and water level data, respectively (pull-down menus 1220, 1225).

[0077] Next, by selecting another site to be compared using the other site selection means 1210 and specifying the sample range of the learning data of the other site, the distribution (1235, 1240) of the other site learning data of the specified range is displayed superimposed on the display area 1015 of the sample distribution.

[0078] Next, as a result of the execution of the above-mentioned steps S410 and S420 by the self-other-site transfer learning processing unit 400S, a separate highlighting 1240a is displayed for the area 1240 that is determined to have a large distribution difference, and guidance recommending that transfer learning be performed using the other site rainfall model for that area 1240 is displayed in the guidance display area 1245.

[0079] For example, when a user performs transfer learning according to a prescribed guidance, the user presses a radio button for transfer learning execution means 1250 to perform immediate execution, or presses a radio button for transfer learning execution means 1255 and inputs a date and time to select whether to perform scheduled execution at the specified date and time, and then presses a transfer learning button, that is, transfer learning execution means 1260, which executes transfer learning using a rainfall model from another site.

[0080] In this embodiment, the timing of transfer learning is specified based on time such as immediately or a specified date and time, but this is not a restriction of the present invention. For example, if site B specified by the other site selection means 1210 is in learning at the specified execution time of transfer learning of the own site (the time if immediate, or the scheduled time if scheduled), the rainfall model of site B, which is the source of transfer learning, will be in an undetermined state, so the transfer learning process to the own site may be started with the learning end event of site B as a trigger.

[0081] As described above, the water treatment plant control system 100 according to this embodiment learns a natural phenomenon prediction model that predicts future natural phenomena using learning data from the natural phenomenon sensor data D110 that measures natural phenomena, while learning an artifact behavior prediction model that predicts future artifact behavior using learning data from the artifact sensor data D120 that measures the behavior of artifacts, and generates a prediction model by combining the trained natural phenomenon prediction model and the trained artifact behavior prediction model with the initial learning processing unit 100S that calculates the distribution of the learning data, and predicts artifact behavior using prediction data from the natural phenomenon sensor data D110 and the artifact sensor data D120 using this prediction model, while also predicting the artifact behavior using the respective natural phenomenon sensor data D110 and the artifact sensor data D120. the distribution of the learning data for the first site and the distribution of the learning data for the second site, and for prediction data belonging to an area where the distribution of the learning data for the first site and the distribution of the learning data for the second site do not match, a trained natural phenomenon prediction model for the second site is selected for prediction data belonging to an area where the distribution of the learning data for the first site and the distribution of the learning data for the second site do not match, and then the trained natural phenomenon prediction model for the second site and the trained artificial object behavior prediction model for the first site are connected to perform transfer learning.

[0082] Furthermore, the water treatment plant control method of the water treatment plant control system 100 according to this embodiment includes an initial learning processing step in which an initial learning processing unit 100S uses learning data from the natural phenomenon sensor data D110 that measures a natural phenomenon to learn a natural phenomenon prediction model that predicts future natural phenomena, while using learning data from the artificial object sensor data D120 that measures the behavior of an artificial object to learn an artificial object behavior prediction model that predicts future artificial object behavior, and calculates a distribution of the learning data; and a prediction processing unit 200S generates a prediction model by combining the learned natural phenomenon prediction model and the learned artificial object behavior prediction model, predicts artificial object behavior using prediction data from the natural phenomenon sensor data D110 and the artificial object sensor data D120 using the prediction model, and calculates a distribution of the natural phenomenon sensor data D110 and the artificial object sensor data D120. the distribution of the learning data for the first site and the distribution of the learning data for the second site, and for the prediction data belonging to the region where the distribution of the learning data for the first site and the distribution of the learning data for the second site do not match, a transfer learning processing step between self and other sites in which a transfer learning processing unit between self and other sites compares the distribution of the learning data for the first site and the distribution of the learning data for the second site, and for the prediction data belonging to the region where the distribution of the learning data for the first site and the distribution of the learning data for the second site do not match, a trained natural phenomenon prediction model for the second site is selected, and the trained natural phenomenon prediction model for the second site and the artificial object behavior prediction model for the first site are connected to perform transfer learning.

[0083] Furthermore, the water treatment plant control program of the water treatment plant control system 100 according to this embodiment includes an initial learning processing step of causing the initial learning processing unit 100S to learn a natural phenomenon prediction model that predicts future natural phenomena using learning data from the natural phenomenon sensor data D110 that measures natural phenomena, while causing the initial learning processing unit 100S to learn an artifact behavior prediction model that predicts future artifact behavior using learning data from the artifact sensor data D120 that measures the behavior of artifacts, and to calculate a distribution of the learning data; and causing the prediction processing unit 200S to generate a prediction model by combining the learned natural phenomenon prediction model and the learned artifact behavior prediction model, and to predict artifact behavior using prediction data from the natural phenomenon sensor data D110 and the artifact sensor data D120 using the prediction model, and to calculate predictions for each of the natural phenomenon sensor data D110 and the artifact sensor data D120. the computer executes a prediction processing step of calculating a distribution of the measurement data, an intra-site additional learning processing step of causing an intra-site additional learning processing unit to compare the distribution of the learning data with the distribution of the prediction data, and perform additional learning using the prediction data belonging to an area where the distribution of the learning data and the distribution of the prediction data do not match as additional learning data for the site, and an intra-site / other-site transfer learning processing step of causing an intra-site / other-site transfer learning processing unit to compare the distribution of the learning data for the first site with the distribution of the learning data for the second site, and for the prediction data belonging to an area where the distribution of the learning data for the first site and the distribution of the learning data for the second site do not match, select a trained natural phenomenon prediction model for the second site, and connect the trained natural phenomenon prediction model for the second site and the artificial object behavior prediction model for the first site to perform transfer learning.

[0084] In this way, by dividing the prediction model into two, a natural phenomenon prediction model and an artificial object behavior prediction model, a natural phenomenon prediction model that has learned about abnormal weather phenomena at a specific site can be reused at another site. This makes it possible to learn in advance about abnormal weather phenomena that occur at a specific site in preparation for cases where similar abnormal weather phenomena occur across multiple sites.

[0085] In addition, by applying this embodiment to the inflow prediction of a storm water pumping station, when a torrential rain or the like occurs, data on the torrential rain can be additionally learned at the site, and the additionally learned natural phenomenon prediction model can be diverted to another site for transfer learning. As a result, before experiencing a torrential rain similar to the torrential rain at another site, it becomes possible to operate appropriately even in an unexperienced torrential rain by using a prediction model that has already been trained on the torrential rain.

[0086] In this embodiment, the first site and the second site are storm water pumping stations, and the prediction processing unit 200S inputs rainfall data as natural phenomenon sensor data to a rainfall model as a natural phenomenon prediction model, and inputs water level data as artificial object sensor data to a pipe network water level model as an artificial object behavior prediction model, as an input of the prediction model, and predicts the amount of rainwater inflow to the storm water pumping station by combining the output of the rainfall model and the output of the pipe network water level model as an output of the prediction model. In this way, it is possible to prepare in advance for a case where a similar abnormal weather phenomenon may occur in another storm water pumping station as a second site, in response to an abnormal weather phenomenon that occurs in a storm water pumping station as a first site.

[0087] In this embodiment, the initial learning processing unit 100S, as a learning data distribution calculation process, calculates the distribution of each of the learning features and learning true values ​​for the learning data used to learn the prediction model from among the natural phenomenon sensor data and the artificial object sensor data, and stores it as distribution data of the learning data, and the prediction processing unit 200S, as a prediction data distribution calculation process, calculates the distribution of the prediction features for the prediction data used for prediction from the natural phenomenon sensor data and the artificial object sensor data, and stores it as distribution data of the prediction data.In this way, the learning data used to learn the prediction model and the prediction data used for prediction can be used reliably.

[0088] In this embodiment, the in-site additional learning processing unit 300S compares the distribution of features included in the learning data for the first site with the distribution of features included in the prediction data for the first site, and if the difference in distribution is large, displays a message encouraging additional learning of the prediction model for the first site using the prediction data, and if an instruction to perform additional learning is received, sets the actual measurement value obtained after the elapse of the prediction destination time to be predicted by the prediction model as the true value for learning, and performs additional learning on the prediction model for the first site using the features included in the prediction data and the true value for learning. In this way, a more suitable prediction model can be created.

[0089] In this embodiment, the self-other-site transfer learning processing unit 400S compares the distribution of feature quantities and true values ​​included in the learning data for the first site with the distribution of feature quantities and true values ​​included in the learning data for the second site, and when the difference in distribution is large, displays a message encouraging transfer learning to the prediction model for the first site using the natural phenomenon prediction model for the second site, and when an instruction to perform transfer learning is received, combines the natural phenomenon prediction model for the second site with the artifact behavior prediction model for the first site, and performs transfer learning using the learning data for the first site. In this way, it is possible to prepare in advance for a case in which a similar abnormal weather phenomenon may occur at another second site with respect to an abnormal weather phenomenon that has occurred at a certain first site.

[0090] In this embodiment, the in-site additional learning processing unit 300S displays an in-site additional learning screen shown in Fig. 10. This in-site additional learning screen has a sample range designation means 1005 as an example of an input part for designating a range of prediction model learning data for the first site, a sample range designation means 1010 as an example of an input part for designating a range of prediction data for the first site, a sample distribution display area 1015 as an example of a part for displaying the distribution of learning data and prediction data, a guidance display area 1045 as an example of an area for displaying a sentence recommending additional learning when the in-site additional learning processing unit 300S determines that the distribution difference between the learning data and the prediction data is large, and an additional learning execution means 1060 as an example of an input part for instructing the execution of additional learning. In this way, the sentence recommending additional learning displayed in the guidance display area 1045 makes it easier to visually recognize that additional learning should be performed.

[0091] In this embodiment, the self-other-site transfer learning processing unit 400S displays a self-other-site transfer learning screen. This self-other-site transfer learning screen has a self-site sample range designation means 1205 as an example of an input part for designating the range of the prediction model learning data of the first site, a other site selection means 1210 as an example of an input part for designating the second site, an input part for designating the range of the prediction model learning data of the second site, a sample distribution display area 1215 as an example of a part for displaying the distribution of the learning data of the first site and the learning data of the second site, a guidance display area 1245 as an example of an area for displaying a sentence recommending transfer learning when the self-other-site transfer learning processing unit 400S determines that the distribution difference between the first site learning data and the second site learning data is large, and a transfer learning execution means 1260 as an example of an input part for instructing the execution of transfer learning. In this way, the sentence recommending transfer learning displayed in the guidance display area 1245 makes it easier to visually recognize that transfer learning should be performed.

[0092] (2) Second embodiment The water treatment plant control system 100a according to the second embodiment has a configuration and operation that is almost the same as that of the water treatment plant control system 100 according to the first embodiment, so a description of the similar configuration and operation will be omitted and the following description will focus on the differences.

[0093] In the water treatment plant control system according to the second embodiment, the first site and the second site are water purification plants, and the prediction processing unit 200S inputs rainfall data as natural phenomenon sensor data D110 into a rainfall model as a natural phenomenon prediction model as input to the prediction model, and inputs water quality data as artificial object sensor data D120 into a water quality model as an artificial object behavior prediction model, and combines the output of the rainfall model and the output of the water quality model as output of the prediction model to predict the water quality of the water purification plant.

[0094] That is, the second embodiment shows an example in which the water treatment plant control system 100, the water treatment plant control method, and the water treatment plant control program according to the first embodiment are applied to water quality prediction at a water purification plant. Note that this embodiment is not limited to such a water purification plant, but is applicable to all water treatment plants in which the plant input may be disturbed by natural phenomena. The second embodiment differs from the first embodiment only in the correspondence relationship between the natural phenomenon sensor data D110 and the artificial object sensor data D120 described above, and the configuration and processing are almost the same, so the following description will focus on the differences.

[0095] 13 is a diagram showing an example in which the water treatment plant control system according to the second embodiment is applied to water quality prediction at a water purification plant. When the water treatment plant control system according to this embodiment is applied to water quality prediction at a water purification plant, the natural phenomenon prediction model M110 corresponds to a raw water quality model, and the artificial object behavior prediction model M120 corresponds to an in-plant water quality model.

[0096] Rainfall flows down river 1310 and into the water intake equipment 1320 of the water purification plant 1300, and then passes through sedimentation equipment 1330, filtration equipment 1340, disinfection equipment 1350, and distribution equipment 1360 before being distributed through the water pipe network. Therefore, in this embodiment, the features input to the raw water quality model are rainfall sensor mesh data and raw water turbidity sensor data, and the features input to the on-site water quality model are sensor data from each process in the water purification plant, as well as the plant control data and the output of the raw water quality model described below. Furthermore, the true value taught to the entire predictive model combining the raw water quality model and the on-site water quality model is, for example, the purified water turbidity sensor data from the distribution equipment.

[0097] Fig. 14 is a block diagram showing an example of a schematic system configuration when the water treatment plant control system according to the second embodiment is applied to water quality prediction at a water purification plant. The system configuration shown in Fig. 14 is almost the same as the system configuration shown in Fig. 3 described above, so the following description will focus on the differences.

[0098] When the water treatment plant control system according to this embodiment is applied to water quality prediction at a water purification plant, the data structure, configuration, and processing are almost the same as when the water treatment plant control system according to the first embodiment is applied to the amount of stormwater inflow at a stormwater pumping station. When the water treatment plant control system according to this embodiment is applied to water quality prediction at a water purification plant, the natural phenomenon sensor data D110 corresponds to rainfall data, and the artificial object sensor data D120 corresponds to water quality data. Similarly, with regard to the prediction model, the natural phenomenon prediction model M110 corresponds to a rainfall model, and the artificial object behavior prediction model corresponds to a water quality model.

[0099] Fig. 15 is a diagram showing the data structure of water quality data in the second embodiment. The water quality data D120 in the second embodiment has a two-dimensional table structure in which the vertical axis is the time axis and the horizontal axis is the value of the sensor in the water purification plant. The intervals of the time axis on the vertical axis correspond to the sampling rate of the sensor in the water purification plant. Meanwhile, the sensor group on the horizontal axis shows the values ​​of various sensors installed at multiple locations in the purified water (for example, the results of measuring the water itself, such as turbidity and pH, and the control values ​​of water treatment plant equipment, such as the amount of coagulant added).

[0100] The sensor group also includes the true value of the prediction target. In this embodiment, the purified water turbidity is exemplified as the true value, but this embodiment is not limited to this, and other sensor values ​​may be used as long as they are values ​​measured in the water purification plant and in the water distribution route. For example, instead of the purified water turbidity, the residual chlorine concentration in the water distribution facility may be used.

[0101] Therefore, in this embodiment, a prediction model is constructed in which the values ​​of any of the sensors measured at the water intake equipment to disinfection equipment, such as raw water turbidity, coagulant injection amount, pH value, etc., from among the values ​​of these sensor groups, are used as feature values, and the purified water turbidity is used as the true value.

[0102] In the second embodiment, the specific data structure of the natural phenomenon sensor data D110 can be the same as the data structure of the rainfall data in the first embodiment. In addition, the natural phenomenon prediction model M110 of the prediction model M100 can also use the rainfall model in the same manner as in the first embodiment.

[0103] In the water treatment plant control system according to the second embodiment, the first site and the second site are water purification plants, and the prediction processing unit 200S inputs rainfall data as natural phenomenon sensor data into a rainfall model as a natural phenomenon prediction model as input to the prediction model, and inputs water quality data as artificial object sensor data into a water quality model as an artificial object behavior prediction model, and combines the output of the rainfall model and the output of the water quality model as output of the prediction model to predict the water quality of the water purification plant.

[0104] In this way, by applying this embodiment to water quality prediction at a water purification plant, when a flood, drought, etc. occurs, data on the abnormal event can be additionally learned at the site, and the additionally learned natural phenomenon prediction model can be diverted to another site for transfer learning. As a result, before an abnormal event similar to the abnormal event occurs at another site, more appropriate water quality can be predicted using a prediction model that has already learned the abnormal event.

[0105] (3) Other embodiments In the above-described embodiment, the water treatment plant control system uses rainfall data to predict the inflow volume of a storm water pumping station and predict the water quality of a water purification plant. However, this embodiment is not limited to the use of rainfall data and can also be applied to a form that does not use rainfall data, such as predicting the water intake volume of a plant that takes in seawater and desalinates it. [Industrial Applicability]

[0106] The present invention can be applied to water treatment plant control systems that utilize prediction technology for plant equipment in fields such as water supply, sewerage, flood control, and water intake, where the input to the plant is affected by natural phenomena. [Explanation of symbols]

[0107] 100...water treatment plant control system, 100S...initial learning processing unit, 200S...prediction processing unit, 300S...additional learning processing unit within own site, 400S...transfer learning processing unit between own and other sites, D100...learning data, D110...natural phenomenon sensor data, D120...artificial object sensor data, D130...distribution data, M100...prediction model, M110...natural phenomenon prediction model, M120...artificial object behavior prediction model.

Claims

1. an initial learning processing unit that uses learning data from natural phenomenon sensor data that measures natural phenomena to learn a natural phenomenon prediction model that predicts future natural phenomena, and also uses learning data from artifact sensor data that measures behavior of artifacts to learn an artifact behavior prediction model that predicts future artifact behavior, and calculates a distribution of the learning data; a prediction processing unit that generates a prediction model by combining the trained natural phenomenon prediction model and the trained artificial object behavior prediction model, predicts artificial object behavior using prediction data from the natural phenomenon sensor data and the artificial object sensor data using the prediction model, and calculates a distribution of prediction data for each of the natural phenomenon sensor data and the artificial object sensor data; an additional learning processing unit within the own site that compares a distribution of the learning data with a distribution of the prediction data, and performs additional learning using the prediction data that belongs to an area in which the distribution of the learning data and the distribution of the prediction data do not match as additional learning data for the site; an inter-site transfer learning processing unit that compares a distribution of the learning data for a first site with a distribution of the learning data for a second site, selects the trained natural phenomenon prediction model for the second site for prediction data belonging to an area where the distribution of the learning data for the first site and the distribution of the learning data for the second site do not match, and then connects the trained natural phenomenon prediction model for the second site and the trained artificial object behavior prediction model for the first site to perform transfer learning; A water treatment plant control system comprising:

2. the first site and the second site are storm water pumping stations; The prediction processing unit inputs rainfall data as the natural phenomenon sensor data into a rainfall model as the natural phenomenon prediction model as an input of the prediction model, and inputs water level data as the artificial object sensor data into a pipe network water level model as the artificial object behavior prediction model, and predicts the amount of rainwater inflow into the storm water pumping station by combining the output of the rainfall model and the output of the pipe network water level model as an output of the prediction model.

2. The water treatment plant control system according to claim 1.

3. the first site and the second site are water purification plants; The prediction processing unit inputs, as an input of the prediction model, rainfall data as the natural phenomenon sensor data into a rainfall model as the natural phenomenon prediction model, and inputs water quality data as the artificial object sensor data into a water quality model as the artificial object behavior prediction model, and combines, as an output of the prediction model, an output of the rainfall model and an output of the water quality model to predict a predicted water quality in the water purification plant.

2. The water treatment plant control system according to claim 1.

4. the initial learning processing unit, as a learning data distribution calculation process, calculates a distribution of each of learning feature values ​​and learning true values ​​for learning data used in learning of the prediction model from among the natural phenomenon sensor data and the artificial object sensor data, and stores the calculated distribution data as distribution data of learning data; the prediction processing unit, as the prediction data distribution calculation process, calculates a distribution of prediction features for prediction data to be used for prediction from among the natural phenomenon sensor data and the artificial object sensor data, and stores the calculated distribution as distribution data of prediction data; 2. The water treatment plant control system according to claim 1.

5. The in-site additional learning processing unit, comparing a distribution of features included in the learning data for the first site with a distribution of features included in the prediction data for the first site; When the difference in distribution is large, a display is displayed to prompt additional learning of the prediction model for the first site using the prediction data; when an instruction to perform the additional learning is received, an actual measurement value obtained after the elapse of a prediction destination time to be predicted by the prediction model is set as a true value for learning, and additional learning is performed on the prediction model for the first site using the feature amount included in the prediction data and the true value for learning.

2. The water treatment plant control system according to claim 1.

6. The self-site / other-site transfer learning processing unit, comparing a distribution of features and true values ​​included in the learning data for the first site with a distribution of features and true values ​​included in the learning data for the second site; When the difference in distribution is large, a display is displayed to encourage transfer learning of the prediction model for the first site using the natural phenomenon prediction model for the second site; when an instruction to perform the transfer learning is received, combining the natural phenomenon prediction model for the second site and the artifact behavior prediction model for the first site, and performing transfer learning using learning data for the first site; 2. The water treatment plant control system according to claim 1.

7. the own-site additional learning processing unit displays an own-site additional learning screen, The additional learning screen on your own site is an input section for specifying a range of prediction model training data for the first site; and an input section for specifying a range of prediction data for the first site; a section for displaying a distribution of the learning data and the prediction data; an area for displaying a message recommending additional learning when the in-site additional learning processing unit determines that the distribution difference between the learning data and the prediction data is large; and an input unit for instructing execution of the additional learning; 2. The water treatment plant control system according to claim 1, further comprising:

8. The self-site / other-site transfer learning unit displays a self-site / other-site transfer learning screen, The self-other site transfer learning screen is An input section for specifying a range of prediction model training data of the first site; an input section for designating the second site; An input section for specifying a range of prediction model training data of the second site; a section for displaying a distribution of the learning data of the first site and the learning data of the second site; an area for displaying a sentence recommending transfer learning when the self-and-other-site transfer learning processing unit determines that the distribution difference between the first-site learning data and the second-site learning data is large; and An input unit for instructing the execution of the transfer learning; 2. The water treatment plant control system according to claim 1, further comprising:

9. an initial learning processing step in which an initial learning processing unit uses learning data from among natural phenomenon sensor data that measures natural phenomena to learn a natural phenomenon prediction model that predicts future natural phenomena, and also uses learning data from among artifact sensor data that measures behavior of artifacts to learn an artifact behavior prediction model that predicts future artifact behavior, and calculates a distribution of the learning data; a prediction processing step in which a prediction processing unit generates a prediction model by combining the trained natural phenomenon prediction model and the trained artificial object behavior prediction model, predicts artificial object behavior using prediction data from the natural phenomenon sensor data and the artificial object sensor data using the prediction model, and calculates a distribution of prediction data for each of the natural phenomenon sensor data and the artificial object sensor data; an additional learning processing step in which an additional learning processing unit in the own site compares a distribution of the learning data with a distribution of the prediction data, and performs additional learning using the prediction data belonging to an area in which the distribution of the learning data and the distribution of the prediction data do not match as additional learning data for the site; an own-other-site transfer learning processing step in which an own-other-site transfer learning processing unit compares a distribution of the learning data for a first site with a distribution of the learning data for a second site, selects a trained natural phenomenon prediction model for the second site for prediction data belonging to an area where the distribution of the learning data for the first site does not match the distribution of the learning data for the second site, and connects the trained natural phenomenon prediction model for the second site and the artificial object behavior prediction model for the first site to perform transfer learning; A water treatment plant control method for a water treatment plant control system comprising:

10. the first site and the second site are storm water pumping stations; In the prediction processing step, the prediction processing unit inputs rainfall data as the natural phenomenon sensor data into a rainfall model as the natural phenomenon prediction model, and inputs water level data as the artificial object sensor data into a pipe network water level model as the artificial object behavior prediction model, as an input of the prediction model, and predicts the amount of rainwater inflow into the storm water pumping station by combining the output of the rainfall model and the output of the pipe network water level model as an output of the prediction model. The water treatment plant control method according to claim 9, wherein the water treatment plant control system is a water treatment plant control method.

11. the first site and the second site are water purification plants; In the prediction processing step, the prediction processing unit inputs rainfall data as the natural phenomenon sensor data into a rainfall model as the natural phenomenon prediction model as an input of the prediction model, and inputs water quality data as the artificial object sensor data into a water quality model as the artificial object behavior prediction model, and combines the output of the rainfall model and the output of the water quality model as an output of the prediction model to predict the predicted water quality in the water purification plant. The water treatment plant control method according to claim 9, wherein the water treatment plant control system is a water treatment plant control method.

12. an initial learning processing step of causing an initial learning processing unit to learn a natural phenomenon prediction model that predicts future natural phenomena using learning data from natural phenomenon sensor data that measures natural phenomena, and to learn an artifact behavior prediction model that predicts future artifact behavior using learning data from artifact sensor data that measures artifact behavior, and to calculate a distribution of the learning data; a prediction processing step of causing a prediction processing unit to generate a prediction model by combining the trained natural phenomenon prediction model and the trained artificial object behavior prediction model, and causing the prediction processing unit to predict artificial object behavior using prediction data from the natural phenomenon sensor data and the artificial object sensor data using the prediction model, and calculating a distribution of prediction data for each of the natural phenomenon sensor data and the artificial object sensor data; an additional learning processing step within the own site for causing an additional learning processing unit within the own site to compare a distribution of the learning data with a distribution of the prediction data, and to perform additional learning using the prediction data belonging to an area in which the distribution of the learning data and the distribution of the prediction data do not match as additional learning data for the own site; an own-other-site transfer learning processing step of causing an own-other-site transfer learning processing unit to compare a distribution of the learning data for a first site with a distribution of the learning data for a second site, select a trained natural phenomenon prediction model for the second site for prediction data belonging to an area where the distribution of the learning data for the first site and the distribution of the learning data for the second site do not match, and then connect the trained natural phenomenon prediction model for the second site and the artificial object behavior prediction model for the first site to perform transfer learning; A water treatment plant control program for causing a computer to execute the above.

13. the first site and the second site are storm water pumping stations; In the prediction processing step, the prediction processing unit inputs rainfall data as the natural phenomenon sensor data into a rainfall model as the natural phenomenon prediction model as an input of the prediction model, and inputs water level data as the artificial object sensor data into a pipe network water level model as the artificial object behavior prediction model, and predicts the amount of rainwater inflow into the rainwater pumping station by combining the output of the rainfall model and the output of the pipe network water level model as an output of the prediction model.

12. The water treatment plant control program according to claim 11.

14. the first site and the second site are water purification plants; In the prediction processing step, the prediction processing unit inputs rainfall data as the natural phenomenon sensor data into a rainfall model as the natural phenomenon prediction model as an input of the prediction model, and inputs water quality data as the artificial object sensor data into a water quality model as the artificial object behavior prediction model, and predicts the predicted water quality in the water purification plant by combining the output of the rainfall model and the output of the water quality model as an output of the prediction model.

12. The water treatment plant control program according to claim 11.

Images