Control device, method for calculating manipulated variable, and program

Abstract

This invention provides a device that is effective in suppressing control delays in the control of water treatment plants using machine learning. [Solution] The control device 100 is a device for controlling a water treatment plant 1 and includes: an accumulation unit 112 that accumulates multiple performance patterns, which are sets of multiple performance records each containing time, an operation amount for the water treatment plant 1, and the state of the water treatment plant 1; a model generation unit 113 that generates a model representing the relationship between the state and time of the water treatment plant 1, multiple explanatory variables related to the operation amount, and a target variable indicating the operation amount, by machine learning based on the multiple performance patterns accumulated in the accumulation unit 112; and a calculation unit 115 that calculates the operation amount based on the current state of the water treatment plant 1, the current time, and the model.

Description

Technical Field

[0005]

[0001] The present disclosure relates to a control device, an operation amount calculation method, and a program.

Background Art

[0002] Patent Document 1 discloses an information processing device for optimizing the injection rate of a flocculant. This information processing device includes a learned model, a parameter generation unit, an information acquisition unit, and a calculation unit. The learned model is configured to receive a predetermined explanatory variable and output a predetermined objective variable. The predetermined explanatory variable includes the injection rate of the flocculant (EV1) and at least one water quality data and / or operation condition data (EV2) before the injection of the flocculant. The predetermined objective variable includes at least one water quality data (OV) after the injection and flocculation of the flocculant. The information acquisition unit is configured to acquire EV2 of the water treatment plant. The parameter generation unit is configured to generate a plurality of EV1. The learned model is configured to output OV from the acquired EV2 and the generated EV1. The calculation unit determines the injection rate of the flocculant based on the generated EV1 and the output OV. The determination of the injection rate of the flocculant is determined based on the absolute value of the change rate of OV with respect to EV1 being less than or equal to a predetermined reference value or less than a predetermined reference value.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] The present disclosure provides an effective device for suppressing control delay in the control of a water treatment plant by machine learning.

Means for Solving the Problems

[0005] A control device relating to one aspect of this disclosure is a device for controlling a water treatment plant and comprises: an accumulation unit that accumulates multiple performance patterns, which are sets of multiple performance records each containing a time, an operator variable for the water treatment plant, and the state of the water treatment plant; a model generation unit that generates a model representing the relationship between multiple explanatory variables related to the state and time of the water treatment plant and a target variable indicating the operator variable, by machine learning based on the multiple performance patterns accumulated in the accumulation unit; and a calculation unit that calculates the operator variable based on the current state of the water treatment plant, the current time, and the model.

[0006] A method for calculating manipulated quantities relating to another aspect of this disclosure includes: accumulating multiple performance patterns, which are sets of multiple performance records each containing a time, the state of the water treatment plant, and the manipulated quantities output to the water treatment plant; generating a model representing the relationship between the state of the water treatment plant, the time, multiple explanatory variables related to the state of the water treatment plant, and a target variable indicating the manipulated quantities, using machine learning based on the accumulated multiple performance patterns; and calculating the manipulated quantities based on the current state of the water treatment plant, the current time, and the model.

[0007] A program relating to yet another aspect of this disclosure involves causing the device to perform the following actions: accumulate multiple performance patterns, which are collections of multiple performance records each containing a time, the state of the water treatment plant, and the manipulated quantity output to the water treatment plant; generate a model representing the relationship between the state of the water treatment plant, the time, multiple explanatory variables related to the water treatment plant, and a target variable indicating the manipulated quantity, using machine learning based on the accumulated multiple performance patterns; and calculate the manipulated quantity based on the current state of the water treatment plant, the current time, and the model. [Effects of the Invention]

[0008] According to this disclosure, it is possible to provide a device that is effective in suppressing control delays in the control of a water treatment plant using machine learning. [Brief explanation of the drawing]

[0009] [Figure 1] This is a block diagram illustrating the configuration of a water treatment plant. [Figure 2] This is a block diagram illustrating the functional configuration of a control device. [Figure 3] This is a table that illustrates actual performance records. [Figure 4] This is a schematic diagram illustrating the model. [Figure 5] This graph illustrates the difference between performance patterns and baseline patterns. [Figure 6] This is a schematic diagram illustrating typical patterns, models, and baseline patterns for each weather condition. [Figure 7] This is a schematic diagram illustrating example performance patterns, models, and standard patterns for each type of drug. [Figure 8] This is a block diagram illustrating the hardware configuration of the control device. [Figure 9] This is a flowchart illustrating the control procedure. [Figure 10] This flowchart illustrates the procedure for updating the reference pattern and model. [Modes for carrying out the invention]

[0010] The embodiments will be described in detail below with reference to the drawings. In the description, the same elements or elements having the same function will be denoted by the same reference numeral, and redundant descriptions will be omitted.

[0011] [Control device] The control device 100 shown in Figure 1 is a device that controls the water treatment plant 1. The water treatment plant 1 is a plant that performs at least a part of water treatment. Water treatment includes water purification treatment, which involves taking raw water from a water source in the natural environment, treating it to a water quality suitable for its intended use, and supplying it to the customer, and sewage treatment, which involves taking wastewater from the customer, purifying it, and returning it to the natural environment. Figure 1 illustrates a case where the water treatment plant 1 is a plant that performs water purification treatment, but the water treatment plant 1 may also be a plant that performs sewage treatment. A water treatment plant 1 that performs water purification treatment may include, as an example, a rapid mixing tank 20, a flocculation tank 30, a sedimentation tank 40, a filtration tank 50, a purified water tank 60, a pump 70, and a distribution tank 80.

[0012] The rapid mixing tank 20 adds chemicals (carbon dioxide, sodium hypochlorite, polyaluminum chloride, aluminum sulfate, etc.) to the water taken from the water source 8 and mixes it rapidly. This process makes it easier for fine particles to aggregate. The water mixed with the chemicals is sent to the flocculation tank 30. In the flocculation tank 30, the chemicals added in the rapid mixing tank 20 cause flocs (clumps) of small particles to form. The sedimentation tank 40 allows the flocs formed in the flocculation tank 30 to settle by gravity. This easily removes solid matter from the water. The settled flocs are sent to a sludge tank (not shown), etc. The water is sent to the filtration tank 50. The filtration tank 50 filters the water sent from the sedimentation tank 40 using a filter containing sand, charcoal, etc. This further removes fine particles, etc. The water purification tank 60 stores the water filtered in the filtration tank 50. Pump 70 pumps water from the water treatment reservoir 60 to the distribution reservoir 80. The water pumped to the distribution reservoir 80 is sent to one or more customer destinations 9.

[0013] The water treatment plant 1 further includes one or more chemical injection devices. For example, the water treatment plant 1 includes a chemical injection device 21 that injects chemicals (such as carbon dioxide, sodium hypochlorite, polyaluminum chloride, aluminum sulfate, etc.) mixed with water into the water in the rapid mixing tank 20. The water treatment plant 1 may further include a chemical injection device 41 that injects chemicals (such as sodium hypochlorite, etc.) into the water in the sedimentation tank 40, and may further include a chemical injection device 51 that injects chemicals (such as sodium hypochlorite, sodium hydroxide, etc.) into the water sent from the filtration tank 50 to the clean water tank 60.

[0014] The control device 100 controls at least a part of the water treatment plant 1. For example, the control device 100 controls the water treatment plant 1 based on the operation of the operator. For example, the control device 100 controls the water treatment plant 1 based on the operation amount input by the operator (for example, input to the user interface 195 described later). The operation amount is, for example, the amount directly operated by the control device 100 in order to make the control amount approach the target. For example, when the control amount is water quality, the chemical injection amounts by the chemical injection devices 21, 41, 51 (the chemical injection amount for the water to be treated) are an example of the operation amount. As an example, the control device 100 causes the chemical injection amounts by at least any one of the chemical injection devices 21, 41, 51 to follow the chemical injection amount input by the operator.

[0015] Thus, when performing control based on the input of the operator, it is necessary for the operator to determine the operation amount based on the state of the water treatment plant. Determining the optimal operation amount from the state of the water treatment plant depends largely on the experience and intuition of the operator, and there are problems in the concentration of work on skilled operators and the inheritance of techniques that depend on each person. Therefore, the control device 100 may be configured to generate a model representing the relationship between the state of the water treatment plant and the operation amount by machine learning, and calculate the operation amount based on the generated model and the state of the water treatment plant. Thereby, it becomes possible to determine the operation amount without relying on the experience and intuition of the operator.

[0016] However, when based on a model representing the relationship between the state of the water treatment plant and the operation amount, after the state of the water treatment plant actually changes, the operation amount for coping with the change will be calculated. As a result, since the change in the operation amount corresponding to the change in the state of the water treatment plant is delayed, there is a possibility that a delay in control for coping with the change in the state of the water treatment plant will occur.

[0017] In contrast, the control device 100 accumulates a plurality of performance patterns which are aggregates of a plurality of performance records each including time, the operation amount for the water treatment plant 1, and the state of the water treatment plant 1, generates a model representing the relationship between the explanatory variables related to the state and time of the water treatment plant 1 and the objective variable indicating the operation amount by machine learning based on the plurality of accumulated performance patterns, and is configured to calculate the operation amount based on the current state of the water treatment plant 1, the current time, and the model.

[0018] A plurality of explanatory variables of the model are related to both the state of the water treatment plant and time. Therefore, before the state of the water treatment plant changes, based on time, the change of the operation amount can be started and the delay of control can be suppressed.

[0019] FIG. 2 is a block diagram illustrating a functional configuration of the control device 100. As shown in FIG. 2, the control device 100 includes, as functional components (hereinafter referred to as “functional blocks”), a data collection unit 111, a storage unit 112, a model generation unit 113, a model storage unit 114, a calculation unit 115, a display unit 116, and a control unit 117. The data collection unit 111 collects data representing the state of the water treatment plant 1. For example, the data collection unit 111 collects data representing the state of the water treatment plant 1 from one or more sensors (for example, a water volume sensor, a water quality sensor, etc.) provided in the water treatment plant 1. For example, the data collection unit 111 repeatedly collects data such as water volume, water quality (turbidity, PH, residual chlorine, odor, alkalinity, water temperature, etc.) from the sensor 22 provided in the rapid mixing tank 20, the sensor 42 provided in the sedimentation tank 40, the sensor 52 provided in the filtration tank 50, and the like.

[0020] The storage unit 112 stores multiple performance patterns, which are collections of multiple performance records based on the data acquired by the data collection unit 111. Each performance record includes the time, the amount of operation performed on the water treatment plant 1, and the state of the water treatment plant 1. Figure 3 is a table illustrating performance records. Each row in the table represents one performance record. Each of the multiple performance records includes the time, the amount of operation performed on the water treatment plant 1, the state of the water treatment plant 1, and identification information of the worker who determined the amount of operation.

[0021] The multiple performance records stored in the storage unit 112 have a certain degree of periodicity. Utilizing this, the storage unit 112 can classify the multiple performance records into multiple performance patterns, each representing a time change between the manipulated volume and the state of the water treatment plant 1 at predetermined intervals. Therefore, storing multiple performance records in the storage unit 112 is an example of storing multiple performance patterns in the storage unit 112. Each of the multiple performance patterns is a collection of multiple performance records within a predetermined period. For example, if the multiple performance records have periodicity corresponding to daily demand patterns, the multiple performance records can be classified into multiple performance patterns, each representing a time change between the manipulated volume and the state of the water treatment plant 1 in a single day. If the workers change daily, each of the multiple performance records will be associated with a single worker.

[0022] Returning to Figure 2, the model generation unit 113 generates a model representing the relationship between explanatory variables related to the state and time of the water treatment plant 1 and an objective variable representing the manipulated quantity, using machine learning based on multiple accumulated performance patterns, and stores the generated model in the model storage unit 114. The objective variable representing the manipulated quantity may be an objective variable that directly represents the manipulated quantity, or it may be an objective variable that indirectly represents the manipulated quantity. For example, the objective variable may represent the difference between the manipulated quantity and a predetermined reference pattern. With an objective variable that represents the difference, the manipulated quantity is represented by the combination with the reference pattern. With a model in which the objective variable is the difference, the deviation of the manipulated quantity from the reference pattern can be kept smaller compared to a model in which the objective variable is the entire manipulated quantity, thereby improving the reliability of the control of the water treatment plant 1.

[0023] A reference pattern represents the relationship between time and a reference value of the manipulated variable. For example, a reference pattern represents the trend of the reference value of the manipulated variable per day. A reference pattern is generated in advance, for example, based on multiple actual patterns. For example, the control device 100 may further have a pattern storage unit 121 that stores reference patterns. The model generation unit 113 may generate a model that represents the relationship between a target variable showing the difference between the manipulated variable and the reference pattern for the water treatment plant 1 and multiple explanatory variables, by machine learning based on multiple actual patterns stored in the storage unit 112 and reference patterns stored in the pattern storage unit 121.

[0024] The control device 100 may further include a reference generation unit 123. The reference generation unit 123 generates a reference pattern based on the relationship between time and manipulated variable in a plurality of actual patterns stored in the storage unit 112, and stores it in the pattern storage unit 121. For example, the reference generation unit 123 generates a reference pattern by calculating a reference value at each time based on a plurality of manipulated variables each of the plurality of actual patterns contained at each time within the predetermined period. For example, the reference generation unit 123 calculates the average value of the plurality of manipulated variables at each time as the reference value at each time.

[0025] Since the model and the baseline pattern are generated based on multiple historical data sets that are the same, the correlation between the multiple explanatory variables and the dependent variable showing the difference increases. This improves the reliability of the model's output, allowing for the calculation of manipulated variables with greater confidence.

[0026] Figure 4 is a graph illustrating multiple performance patterns and a reference pattern. The horizontal axis represents time, and the vertical axis represents the manipulated variable. Figure 4 shows two sets of performance patterns P1 and P2, and a reference pattern RP. Due to the periodicity described above, there is a certain degree of similarity between performance patterns P1 and P2, but there is variability between them due to factors such as the operator's habits and the state of the water treatment plant 1 on that day. The reference generation unit 123 generates the reference pattern RP based on performance patterns P1 and P2. In this way, by combining multiple performance patterns with variability into one reference pattern, a model with the difference as the target variable can be easily generated.

[0027] The model generation unit 113 generates several explanatory variables for the model, which may include a variable that shows the difference between the manipulated variable at the previous time (prior to the current time) and the reference pattern, as an explanatory variable related to time. When the control of the water treatment plant 1 is repeatedly executed, the previous time is, for example, the time in the previous cycle. An explanatory variable related to time means a variable that is uniquely determined according to time. A variable that is uniquely determined according to time includes a variable that is uniquely determined according to time and other variables. The difference between the manipulated variable and the reference pattern at the previous time is the difference between the manipulated variable at the previous time and the reference value corresponding to the previous time in the reference pattern. The reference value corresponding to the previous time is a value that is uniquely determined by the reference pattern and the previous time. The variable that shows the difference between the manipulated variable and the reference pattern at the previous time is a variable related to time because it is uniquely determined according to the manipulated variable at the previous time and the previous time.

[0028] According to a variable that represents the relationship with time based on the difference between the manipulated quantity and the reference pattern at the previous time, it is possible to suppress excesses or deficiencies in the manipulated quantity based on the magnitude of the difference, even before a change in the state of water treatment plant 1 occurs.

[0029] The model generation unit 113 generates several explanatory variables for the model, which may include, as explanatory variables related to time, a variable indicating the manipulated quantity at the previous time prior to the current time, and a variable indicating the reference value corresponding to the previous time in the reference pattern. By using variables that represent the relationship with time based on the manipulated quantity at the previous time and the reference value at the previous time, it is possible to suppress excesses or deficiencies in the manipulated quantity by using both the manipulated quantity at the previous time and the reference value at the previous time, even before a change appears in the state of the water treatment plant 1.

[0030] The explanatory variables of the model generated by the model generation unit 113 may include variables that directly indicate time, as explanatory variables related to time. By using explanatory variables that directly indicate time, it is possible to determine the manipulated variables with even greater accuracy.

[0031] The model generation unit 113 may generate a model using the random forest method. This allows for the easy generation of highly reliable models. Figure 5 schematically illustrates a model generated by the random forest method. The model 320 shown in Figure 5 includes multiple decision trees 321 and an output selection unit 322. The decision trees 321 represent the relationship between inputs and outputs using conditional branching in multi-stage branches 323. The output selection unit 322 selects representative values ​​from the outputs of the multiple decision trees 321. Here, the representative values ​​may be the mean, median, or predicted values ​​of a specified decision tree. Furthermore, the mean may be the average of the predicted values ​​of a specified set of decision trees. In the random forest method, the model generation unit 113 generates multiple decision trees 321 by bagging.

[0032] As shown in Figure 5, each of the multiple decision trees 321 receives input of explanatory variables related to the state of the water treatment plant 1 and the time. For example, each of the multiple decision trees 321 receives input of multiple explanatory variables, including a variable representing the state of the water treatment plant 1 and a variable showing the difference between the manipulated variable and the reference pattern at the previous time, and outputs an objective variable showing the difference between the manipulated variable and the reference pattern at the current time. The output selection unit 322 selects a representative value from the outputs of the multiple decision trees 321 and outputs the selected representative value as the objective variable.

[0033] The method for generating the model is not limited to the random forest method. For example, the model generation unit 113 may generate multiple decision trees 321 by gradient boosting instead of bagging. The model generation unit 113 may also generate a neural network as the model using deep learning.

[0034] Returning to Figure 2, the calculation unit 115 repeatedly executes a cycle to calculate the manipulated variable at the current time based on the current state of the water treatment plant 1, the current time, and the model stored in the model storage unit 114. For example, the calculation unit 115 repeats the above cycle at a predetermined processing interval. "Current" means the cycle being executed, and "current time" means a specific point in time within the cycle being executed. For example, the calculation unit 115 calculates the difference between the manipulated variable and the reference pattern at the current time based on the current state of the water treatment plant 1, the current time, and the model, and then calculates the manipulated variable based on the calculated difference, the reference pattern, and the current time. The calculation unit 115 may also identify the previous time based on the current time and input multiple explanatory variables into the model, including a variable representing the current state of the water treatment plant 1 and a variable indicating the difference between the manipulated variable and the reference pattern at the previous time. As a result, the model outputs a target variable representing the difference between the manipulated variable and the reference pattern at the current time. The calculation unit 115 calculates the manipulated variable by adding the calculated difference to the reference value corresponding to the current time in the reference pattern. Addition here includes adding negative values ​​(subtracting positive values).

[0035] As described above, the model stored in the model storage unit 114 represents the relationship between the state and time of the water treatment plant 1, explanatory variables related to that state, and a target variable indicating the manipulated quantity. Even before the amount of water to be treated starts to rise from zero, the model can output a non-zero manipulated quantity (for example, a manipulated quantity greater than zero) if the difference between the manipulated quantity at the previous time and the reference pattern is a non-zero value. For this reason, the calculation unit 115 may calculate a non-zero manipulated quantity before the amount of water to be treated starts to rise from zero, based on the current time and the model. For example, a chemical injection amount greater than zero may be calculated before the amount of water to which the chemical solution is injected starts to rise from zero.

[0036] The display unit 116 displays the manipulated quantity calculated by the calculation unit 115. For example, the display unit 116 displays the manipulated quantity on a display device of the user interface 195, which will be described later. By making the operator aware of the manipulated quantity calculated by the calculation unit 115, the reliability of the control of the water treatment plant 1 can be further improved.

[0037] The control unit 117 controls the water treatment plant 1 based on the manipulated variable calculated by the calculation unit 115. For example, the control unit 117 makes the water treatment plant 1 follow the manipulated variable calculated by the calculation unit 115. The control unit 117 may also make the water treatment plant 1 follow the manipulated variable determined by the operator after recognizing the manipulated variable displayed by the display unit 116. For example, the control unit 117 may make the water treatment plant 1 follow the manipulated variable entered by the operator into the input device of the user interface 195. When the operator determines the manipulated variable after recognizing the manipulated variable displayed by the display unit 116, the manipulated variable determined by the operator is based on the manipulated variable calculated by the calculation unit 115. Therefore, making the water treatment plant 1 follow the manipulated variable determined by the operator after recognizing the manipulated variable displayed by the display unit 116 is also included in controlling the water treatment plant 1 based on the manipulated variable calculated by the calculation unit 115.

[0038] For example, the control unit 117 controls at least one of the chemical injection devices 21, 41, and 51 based on the chemical injection amount calculated by the calculation unit 115. For instance, the control unit 117 makes the chemical injection amount from at least one of the chemical injection devices 21, 41, and 51 follow the chemical injection amount calculated by the calculation unit 115. The control unit 117 may also make the chemical injection amount from at least one of the chemical injection devices 21, 41, and 51 follow the operating amount determined by the operator after recognizing the chemical injection amount displayed by the display unit 116.

[0039] The storage unit 112 may continue to store actual records even after the model and reference pattern have been generated. If multiple actual patterns are updated by the storage unit 112, the reference generation unit 123 may regenerate the reference pattern based on the relationship between time and manipulated variable in the updated actual patterns. The model generation unit 113 may regenerate the model based on the updated actual patterns and the updated reference pattern. By updating both the reference pattern and the model in accordance with actual results, the reliability of the control of the water treatment plant 1 can be maintained and improved.

[0040] An example of a case where multiple performance patterns are updated is when a predetermined number of performance patterns are newly stored in the storage unit 112. The predetermined number may be 1 or 2 or more.

[0041] Each of the multiple performance records may further include weather information at the location of the water treatment plant 1. The model generation unit 113 may generate a model for each weather category. By not including weather that has a large impact on the manipulated variables as explanatory variables and by separating the model for each weather category, the reliability of the model output can be improved.

[0042] The reference generation unit 123 may generate reference patterns for each weather category. The model generation unit 113 may generate a model for each weather category that shows the relationship between a target variable representing the difference between the manipulated variable and the reference pattern corresponding to the weather category, and multiple explanatory variables. By separating the reference patterns by weather category, the correlation between the multiple explanatory variables and the target variable representing the difference is increased. This further improves the reliability of the model output.

[0043] Figure 6 is a schematic diagram illustrating actual patterns, models, and baseline patterns for each weather category. The storage unit 112 groups multiple actual patterns into a predetermined number of weather categories. For example, based on the weather information contained in each of the multiple actual records, the storage unit 112 labels each of the multiple actual patterns with one of the multiple weather categories. This groups multiple actual patterns into multiple weather categories.

[0044] In the diagram, Group G1 is the group labeled with the weather category "sunny". Group G2 is the group labeled with the weather category "cloudy". Group G3 is the group labeled with the weather category "rainy".

[0045] The reference generation unit 123 generates a reference pattern for each of the multiple weather categories, based on the multiple actual patterns within each group, where multiple actual patterns belong to each group. This generates multiple reference patterns corresponding to each of the multiple groups. In the diagram, reference pattern RP1 corresponds to group G1, reference pattern RP2 corresponds to group G2, and reference pattern RP3 corresponds to group G3.

[0046] The model generation unit 113 generates a model for each of the multiple groups that represents the relationship between the manipulated variable, the target variable showing the difference between it and the corresponding reference pattern, and multiple explanatory variables. This generates multiple models corresponding to each of the multiple groups. For example, the model generation unit 113 generates model M1, which represents the relationship between the manipulated variable, the target variable showing the difference between it and the reference pattern RP1, and multiple explanatory variables, based on the multiple actual patterns of group G1 and the reference pattern RP1. The model generation unit 113 generates model M2, which represents the relationship between the manipulated variable, the target variable showing the difference between it and the reference pattern RP2, and multiple explanatory variables, based on the multiple actual patterns of group G2 and the reference pattern RP2. The model generation unit 113 generates model M3, which represents the relationship between the manipulated variable, the target variable showing the difference between it and the reference pattern RP3, and multiple explanatory variables, based on the multiple actual patterns of group G3 and the reference pattern RP3.

[0047] As described above, the manipulated quantity may include the injection amounts of multiple types of chemicals into the water to be treated. The reference generation unit 123 may generate a reference pattern for each type of chemical. The model generation unit 113 may generate a model for each type of chemical that represents the relationship between a target variable showing the difference between the injection amount and the reference pattern corresponding to the type of chemical, and multiple explanatory variables. The calculation unit 115 may calculate the injection amount for each type of chemical based on the corresponding reference pattern and the corresponding model. The injection amount can be accurately determined for each of the multiple types of chemicals.

[0048] Figure 7 is a schematic diagram illustrating example performance patterns, models, and standard patterns for each type of drug. The storage unit 112 stores multiple performance patterns for each type of drug.

[0049] In the diagram, group G11 represents multiple performance patterns accumulated for drug A, group G12 represents multiple performance patterns accumulated for drug B, and group G13 represents multiple performance patterns accumulated for drug C.

[0050] The reference generation unit 123 generates a reference pattern for each of the multiple groups of drugs, based on the multiple performance patterns within that group, given that each group of multiple types of drugs has multiple performance patterns. As a result, multiple reference patterns corresponding to multiple groups are generated. In the figure, reference pattern RP11 corresponds to group G11, reference pattern RP12 corresponds to group G12, and reference pattern RP13 corresponds to group G13.

[0051] The model generation unit 113 generates a model for each of the multiple groups that represents the relationship between the manipulated variable, the target variable showing the difference between it and the corresponding reference pattern, and multiple explanatory variables. This generates multiple models corresponding to each of the multiple groups. For example, the model generation unit 113 generates a model that represents the relationship between the manipulated variable, the target variable showing the difference between it and the reference pattern RP11, and multiple explanatory variables, based on the multiple actual patterns of group G11 and the reference pattern RP11. The model generation unit 113 generates a model that represents the relationship between the manipulated variable, the target variable showing the difference between it and the reference pattern RP12, and multiple explanatory variables, based on the multiple actual patterns of group G12 and the reference pattern RP12. The model generation unit 113 generates a model that represents the relationship between the manipulated variable, the target variable showing the difference between it and the reference pattern RP13, and multiple explanatory variables, based on the multiple actual patterns of group G13 and the reference pattern RP13.

[0052] The calculation unit 115 calculates the injection amount of drug A based on multiple explanatory variables obtained for drug A, the reference pattern RP11, and the model M11. The calculation unit 115 calculates the injection amount of drug B based on multiple explanatory variables obtained for drug B, the reference pattern RP12, and the model M12. The calculation unit 115 calculates the injection amount of drug C based on multiple explanatory variables obtained for drug C, the reference pattern RP13, and the model M13.

[0053] Figure 8 is a block diagram illustrating the hardware configuration of the control device 100. The control device 100 has a circuit 190. The circuit 190 has a processor 191, a memory 192, a storage 193, an input / output port 194, and a user interface 195. The storage 193 includes, for example, one or more non-volatile storage media. The non-volatile storage media includes one or more storage devices. Examples of one or more storage devices include hard disk drives, solid-state drives, flash memory, etc. The non-volatile storage media may also include portable storage media such as optical discs. Storage 193 stores a program that instructs the control device 100 to perform the following: accumulate multiple performance patterns, which are collections of multiple performance records each containing time, an operation quantity for water treatment plant 1, and the state of water treatment plant 1; generate a model representing the relationship between explanatory variables related to the state and time of water treatment plant 1 and a target variable indicating the operation quantity, using machine learning based on the multiple accumulated performance patterns; and calculate the operation quantity based on the current state of water treatment plant 1, the current time, and the model. For example, storage 193 stores a program that instructs the control device 100 to configure the above-mentioned multiple functional blocks.

[0054] Memory 192 includes one or more volatile storage media. The volatile storage media includes one or more memory devices. An example of one or more memory devices is random access memory. Memory 192 temporarily stores programs loaded from storage 193. Processor 191 includes one or more arithmetic devices. An example of an arithmetic device is a CPU (Centreal Processing Unit) or a GPU (Graphics Processing Unit). Processor 191 executes the programs loaded into memory 192, thereby causing the control device 100 to configure the above-mentioned functional blocks. Processor 191 may temporarily store the calculation results in memory 192.

[0055] The input / output port 194 performs electrical signal input and output with components of the water treatment plant 1 (e.g., chemical injection devices 21, 41, 51, and sensors 22, 42, 52, etc.) in response to requests from the processor 191. The user interface 195 performs information input and output to the operator. For example, the user interface 195 includes an output device that outputs information to the operator and an input device that acquires input from the operator. The output device outputs information to the operator in the form of sound or images, etc. Examples of output devices include liquid crystal monitors or organic EL (Electro Luminescence) monitors. Examples of input devices include keyboards and mice. The output device and the input device may be integrated, such as in a so-called touch panel.

[0056] [Control Procedure] Next, as an example of a control method, a control procedure executed by the control device 100 is illustrated. This procedure includes a manipulated variable calculation procedure executed by the control device 100 as an example of a manipulated variable calculation method. The manipulated variable calculation procedure includes accumulating multiple actual patterns, which are sets of multiple actual records each containing time, the state of the water treatment plant 1, and the manipulated variable output to the water treatment plant 1; generating a model representing the relationship between the state of the water treatment plant 1, time, multiple explanatory variables related to the state and time of the water treatment plant 1, and a target variable indicating the manipulated variable, using machine learning based on the accumulated multiple actual patterns; and calculating the manipulated variable based on the current state of the water treatment plant 1, the current time, and the model.

[0057] Figure 9 is a flowchart illustrating the control procedure. The procedure shown in Figure 9 is executed with pre-generated reference patterns for each weather category stored in the pattern storage unit 121 and pre-generated models for each weather category stored in the model storage unit 114. Furthermore, the control procedure shown in Figure 9 is repeatedly executed at a predetermined processing cycle.

[0058] As shown in Figure 9, the control device 100 executes steps S01 and S02. In step S01, the data acquisition unit 111 acquires data representing the current state of the water treatment plant 1. In step S02, the calculation unit 115 determines which of several weather categories the current weather falls into, selects a reference pattern corresponding to the determination result from multiple reference patterns stored in the pattern storage unit 121, and selects a model corresponding to the determination result from multiple models stored in the model storage unit 114.

[0059] Next, the control device 100 executes step S03. In step S03, the calculation unit 115 calculates the difference between the manipulated amount and the reference pattern at the previous time (for example, the time one cycle ago) prior to the current time (hereinafter referred to as the "previous difference"). For example, the calculation unit 115 identifies the previous time based on the current time, identifies the manipulated amount at the previous time from multiple actual records stored in the storage unit 112, and identifies the reference value corresponding to the previous time in the reference pattern. The calculation unit 115 calculates the difference between the identified manipulated amount and the identified reference value as the previous difference.

[0060] Next, the control device 100 executes step S04. In step S04, the calculation unit 115 calculates the difference between the manipulated variable and the reference pattern at the current time (hereinafter referred to as the "current difference"). For example, the calculation unit 115 inputs several explanatory variables into the model, including a variable representing the current state of the water treatment plant 1 and a variable indicating the previous difference. As a result, the target variable representing the current difference is output from the model.

[0061] Next, the control device 100 executes step S05. In step S05, the calculation unit 115 calculates the manipulated variable based on the current difference and the reference pattern. For example, the calculation unit 115 identifies a reference value corresponding to the current time in the reference pattern and calculates the manipulated variable based on the current difference and the identified reference value.

[0062] Next, the control device 100 executes steps S06, S07, S08, and S09. In step S06, the display unit 116 displays the manipulated quantity. In step S07, the control unit 117 controls the water treatment plant 1 based on the manipulated quantity entered by the operator. In step S08, the storage unit 112 stores a performance record that includes the current time, the manipulated quantity in step S07, and data representing the state of the water treatment plant 1 acquired in step S01. In step S09, the calculation unit 115 waits for the elapsed time from the start of step S01 to reach the processing cycle.

[0063] Subsequently, the control device 100 returns the process to step S01. The control device 100 then repeatedly executes the above process.

[0064] The control procedure may include procedures for updating the reference pattern and model. Figure 10 is a flowchart illustrating an example of a procedure for updating the reference pattern and model. This update procedure is performed while the control procedure described above is being executed.

[0065] As shown in Figure 10, the control device 100 executes steps S11, S12, and S13. In step S11, the storage unit 112 waits for new historical patterns to be accumulated. In step S12, the storage unit 112 labels the newly added historical patterns with one of several weather categories. In step S13, the storage unit 112 checks whether a number of historical patterns suitable for updating the reference pattern and model have been newly accumulated. For example, the storage unit 112 checks whether multiple historical patterns have been newly accumulated for each of the several weather categories.

[0066] In step S13, if the control device 100 determines that a sufficient number of actual patterns suitable for updating the reference pattern and model have not been accumulated, the control device 100 returns to step S11. In step S13, if the control device 100 determines that a sufficient number of actual patterns suitable for updating the reference pattern and model have been accumulated, the control device 100 executes steps S14 and S15. In step S14, the reference generation unit 123 generates (updates) a reference pattern for each weather category based on the newly accumulated actual patterns. In step S15, the model generation unit 113 generates (updates) a model for each weather category based on the newly accumulated actual patterns and the newly generated reference pattern. After that, the control device 100 returns to step S11. The control device 100 repeatedly executes the above process. Note that the control device 100 may update the reference pattern and model at predetermined intervals, regardless of the number of accumulated actual patterns.

[0067] 〔summary〕 The above disclosure includes the following components: (1) A control device 100 for controlling a water treatment plant 1, comprising: an accumulation unit 112 that accumulates multiple performance patterns which are sets of multiple performance records each containing time, an operation amount for the water treatment plant 1, and the state of the water treatment plant 1; a model generation unit 113 that generates a model representing the relationship between multiple explanatory variables related to the state and time of the water treatment plant 1 and an objective variable indicating the operation amount, by machine learning based on the multiple performance patterns accumulated in the accumulation unit 112; and a calculation unit 115 that calculates the operation amount based on the current state of the water treatment plant 1, the current time, and the model. The manipulated variables required for control change according to changes in the state of the water treatment plant 1. Therefore, it is conceivable to generate a model using machine learning that represents the relationship between one or more explanatory variables related to the state of the water treatment plant 1 and the target variable that represents the manipulated variables. Based on a model generated by machine learning, highly reliable manipulated variables can be easily calculated. However, a time lag is unavoidable between the recognition of a change in the state of the water treatment plant 1 and the modification of the manipulated variables in response to that change, resulting in a control delay. In contrast, with this control device 100, multiple explanatory variables in the model are related to both the state of the water treatment plant 1 and time. Therefore, it is possible to start changing the manipulated variables based on time before a change in the state of the water treatment plant 1 is recognized, thereby suppressing the control delay.

[0068] (2) The control device 100 according to (1), further comprising a storage unit 121 that stores a reference pattern representing the relationship between time and a reference value of the manipulated quantity, the model generation unit 113 generates a model representing the relationship between a target variable showing the difference between the manipulated quantity and the reference pattern for the water treatment plant 1 and a plurality of explanatory variables by machine learning based on a plurality of actual patterns stored in the storage unit 112 and the reference pattern stored in the storage unit 121, the calculation unit 115 calculates the difference based on the current state of the water treatment plant 1, the current time and the model, and calculates the manipulated quantity based on the calculated difference, the reference pattern and the current time. By limiting the model's output to the difference relative to a reference pattern, rather than the entire manipulated variable, the reliability of the control can be further improved.

[0069] (3) The control device 100 as described in (2), wherein the multiple explanatory variables include, as explanatory variables related to time, a variable that shows the difference between the manipulated variable at the previous time prior to the current time and the reference pattern. According to a variable that represents the relationship with time based on the difference between the manipulated quantity and the reference pattern at the previous time, it is possible to suppress excesses or deficiencies in the manipulated quantity based on the magnitude of the difference, even before a change in the state of the water treatment plant 1 occurs.

[0070] (4) The control device 100 according to (2) or (3), further comprising a reference generation unit 123 that generates a reference pattern based on the relationship between time and operation amount in multiple performance patterns and stores it in the storage unit 121. Since the model and the baseline pattern are generated based on multiple historical data sets that are the same, the correlation between the multiple explanatory variables and the dependent variable showing the difference increases. This improves the reliability of the model's output, allowing for the calculation of manipulated variables with greater confidence.

[0071] (5) When multiple performance patterns are updated by the storage unit 112, the reference generation unit 123 regenerates a reference pattern based on the relationship between time and manipulated variable in the updated multiple performance patterns, and the model generation unit 113 regenerates a model based on the updated multiple performance patterns and the updated reference pattern, the control device 100 as described in (4). By updating both the standard patterns and models in accordance with actual performance, the reliability of the control system for water treatment plant 1 can be maintained and improved.

[0072] (6) Each of the multiple performance records further includes weather information at the location of the water treatment plant 1, and the model generation unit 113 generates a model for each weather category, the control device 100 according to any one of (2) to (5). By excluding weather, which has a significant impact on the manipulated parameters, from the explanatory variables and instead dividing the model by weather category, the reliability of the model's output can be improved.

[0073] (7) Each of the multiple performance records further includes weather information at the location of the water treatment plant 1, and the control device 100 according to (4) or (5), wherein the reference generation unit 123 generates a reference pattern for each weather category. By dividing the reference patterns into weather categories, the correlation between multiple explanatory variables and the dependent variable showing the difference increases. This further improves the reliability of the model's output.

[0074] (8) The control device 100 described in (7) generates a model generation unit 113 which generates a model for each weather category that shows the relationship between a target variable that shows the difference between the manipulated variable and a reference pattern corresponding to the weather category, and a plurality of explanatory variables. By separating both the baseline pattern and the model by weather category, the correlation between multiple explanatory variables and the dependent variable showing the difference is further increased. This can further improve the reliability of the model's output.

[0075] (9) The control device 100 according to (2), wherein the multiple explanatory variables include, as explanatory variables relating to time, a variable indicating the manipulated quantity at the previous time prior to the current time, and a variable indicating the reference value corresponding to the previous time in the reference pattern. According to a variable that represents the relationship with time using the manipulated quantity at the previous time and the reference value at the previous time, even before a change in the state of the water treatment plant 1 occurs, it is possible to suppress an excess or deficiency of the manipulated quantity by using both the manipulated quantity at the previous time and the reference value at the previous time.

[0076] (10) The control device 100 according to (2), wherein the multiple explanatory variables include a variable that directly indicates time as an explanatory variable related to time. By using explanatory variables that directly indicate time, the manipulated variables can be determined more accurately.

[0077] (11) The control device 100 according to any one of (2) to (10), wherein the manipulated amount includes the amount of chemical injected into the water to be treated. The amount of medication to be injected can be determined with high accuracy.

[0078] (12) The control device 100 described in (11), wherein the calculation unit 115 calculates a non-zero manipulated amount based on the current time and the model before the amount of water to be treated starts to rise from zero. This allows for more precise determination of the amount of medication to be injected.

[0079] (13) The control device 100 according to any one of (4) to (12), wherein the manipulated quantity includes the injection amounts of multiple types of chemicals into the water to be treated, the reference generation unit 123 generates a reference pattern for each type of chemical, the model generation unit 113 generates a model for each type of chemical that represents the relationship between an objective variable showing the difference between the injection amount and the corresponding reference pattern and multiple explanatory variables, and the calculation unit 115 calculates the injection amount for each type of chemical based on the corresponding reference pattern and the corresponding model. The injection amount for each of multiple types of medications can be determined with high accuracy.

[0080] (14) The model generation unit 113 is a control device 100 according to any one of (1) to (13) that generates a model by the random forest method. Highly reliable models can be easily generated.

[0081] (15) A control device 100 according to any one of (1) to (14), further comprising a display unit 116 that displays the manipulated amount calculated by the calculation unit 115. By making the operator aware of the calculated operation amount, the reliability of the control of the water treatment plant 1 can be further improved.

[0082] (16) A method for calculating manipulated quantities, comprising: accumulating multiple performance patterns, which are sets of multiple performance records each containing a time, the state of water treatment plant 1, and a manipulated quantity output to water treatment plant 1; generating a model representing the relationship between the state and time of water treatment plant 1, multiple explanatory variables related to the state and time of water treatment plant 1, and a target variable indicating the manipulated quantity, using machine learning based on the multiple accumulated performance patterns; and calculating the manipulated quantity based on the current state of water treatment plant 1, the current time, and the model.

[0083] (17) A program that causes the device to perform the following: accumulate multiple performance patterns, which are sets of multiple performance records each containing the time, the state of water treatment plant 1, and the manipulated quantity output to water treatment plant 1; generate a model representing the relationship between the state and time of water treatment plant 1, multiple explanatory variables related to the state and time of water treatment plant 1, and a target variable indicating the manipulated quantity, using machine learning based on the accumulated multiple performance patterns; and calculate the manipulated quantity based on the current state of water treatment plant 1, the current time, and the model.

[0084] Although embodiments have been described above, this disclosure is not necessarily limited to the embodiments described above, and various modifications are possible without departing from its essence. [Explanation of Symbols]

[0085] 1...Water treatment plant, 100...Control device, 112...Storage unit, 121...Pattern storage unit, 123...Reference generation unit, 113...Model generation unit, 115...Calculation unit, 116...Display unit.

Claims

1. A device for controlling a water treatment plant, A storage unit that stores multiple performance patterns, which are collections of multiple performance records each containing the time, the amount of operation performed on the water treatment plant, and the state of the water treatment plant, A model generation unit generates a model representing the relationship between multiple explanatory variables related to the state and time of the water treatment plant and a target variable indicating the manipulated quantity, using machine learning based on multiple actual patterns stored in the storage unit. A calculation unit that calculates the manipulated quantity based on the current state of the water treatment plant, the current time, and the model, A control device equipped with the following features.

2. The system further includes a storage unit that stores a reference pattern representing the relationship between the time and the reference value of the manipulated variable, The model generation unit generates a model representing the relationship between the target variable, which shows the difference between the manipulated amount and the reference pattern for the water treatment plant, and the plurality of explanatory variables, by machine learning based on a plurality of actual patterns stored in the storage unit and the reference pattern stored in the memory unit. The calculation unit calculates the difference based on the current state of the water treatment plant, the current time, and the model, and calculates the manipulated quantity based on the calculated difference, the reference pattern, and the current time. The control device according to claim 1.

3. The aforementioned plurality of explanatory variables include, as explanatory variables related to time, a variable showing the difference between the manipulated quantity at the previous time prior to the current time and the reference pattern, The control device according to claim 2.

4. The system further includes a reference generation unit that generates a reference pattern based on the relationship between the time and the manipulated amount in the plurality of performance patterns and stores it in the storage unit. The control device according to claim 2.

5. When the multiple performance patterns are updated by the storage unit, The reference generation unit regenerates the reference pattern based on the relationship between the time and the manipulated amount in the updated plurality of performance patterns. The model generation unit regenerates the model based on the updated multiple performance patterns and the updated reference pattern. The control device according to claim 4.

6. Each of the aforementioned multiple performance records further includes weather information at the location of the water treatment plant, The model generation unit generates the model for each weather category. The control device according to any one of claims 2 to 5.

7. Each of the aforementioned multiple performance records further includes weather information at the location of the water treatment plant, The aforementioned reference generation unit generates the reference pattern for each weather category. The control device according to claim 4 or 5.

8. The model generation unit generates a model for each weather category that represents the relationship between the target variable, which shows the difference between the manipulated variable and the reference pattern corresponding to the weather category, and the plurality of explanatory variables. The control device according to claim 7.

9. The aforementioned plurality of explanatory variables include, as explanatory variables related to time, a variable indicating the manipulated quantity at the previous time prior to the current time, and a variable indicating the reference value corresponding to the previous time in the reference pattern. The control device according to claim 2.

10. The aforementioned multiple explanatory variables include, as explanatory variables related to time, a variable that directly indicates time. The control device according to claim 2.

11. The aforementioned operating amount includes the amount of chemical injected into the water to be treated. The control device according to any one of claims 2 to 5.

12. The calculation unit calculates a non-zero manipulated quantity based on the current time and the model, before the amount of water to be treated starts to rise from zero. The control device according to claim 11.

13. The aforementioned operating amount includes the amount of multiple types of chemicals injected into the water to be treated. The aforementioned reference generation unit generates the reference pattern for each type of chemical, The model generation unit generates a model for each type of drug that represents the relationship between the injection amount, the target variable showing the difference between that amount and the corresponding reference pattern, and the plurality of explanatory variables. The calculation unit calculates the injection amount for each type of drug based on the corresponding reference pattern and the corresponding model. The control device according to claim 4 or 5.

14. The model generation unit generates the model using the random forest method. The control device according to any one of claims 1 to 5.

15. The system further includes a display unit that displays the manipulated quantity calculated by the calculation unit. The control device according to any one of claims 1 to 5.

16. This involves accumulating multiple performance patterns, which are collections of performance records each containing the time, the status of the water treatment plant, and the manipulated quantities output to the water treatment plant. A model representing the relationship between multiple explanatory variables related to the state and time of the water treatment plant and the target variable indicating the manipulated quantity is generated by machine learning based on multiple accumulated historical patterns. The operation quantity is calculated based on the current state of the water treatment plant, the current time, and the model. A method for calculating manipulated variables, including the method itself.

17. This involves accumulating multiple performance patterns, which are collections of performance records each containing the time, the status of the water treatment plant, and the manipulated quantities output to the water treatment plant. A model representing the relationship between multiple explanatory variables related to the state and time of the water treatment plant and the target variable indicating the manipulated quantity is generated by machine learning based on multiple accumulated historical patterns. The operation quantity is calculated based on the current state of the water treatment plant, the current time, and the model. A program that causes a device to execute a command.

Images