Prediction device, prediction method, and program

Abstract

To easily and accurately predict the inflow volume at a desired particle size. 【Solution means】A prediction device according to one aspect of the present disclosure has a first prediction unit that predicts the temporal change in the inflow volume of water into a dam by repeatedly inputting rainfall and the inflow volume into a first neural network based on a first neural network that predicts the minute temporal change in the inflow volume of water into the dam and a continuous time simulation.

Description

Technical Field

[0001] The present disclosure relates to a prediction device, a prediction method, and a program.

Background Art

[0002] At the site of dam management, the operation is carried out while predicting the inflow of water into the dam so as not to cause emergency discharges or overflows. As technologies for predicting the inflow and water level of dams, or technologies related to or applicable to them, for example, Patent Documents 1 to 6 and Non-Patent Documents 1 to 8 are known.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Patent Document 2

Patent Document 3

Patent Document 4

Patent Document 5

Patent Document 6

Non-Patent Documents

[0004]

Non-Patent Document 1

Non-Patent Document 2

Non-Patent Document 3

Non-Patent Document 4

Non-Patent Document 5

Non-Patent Document 6

Non-Patent Document 7

Non-Patent Document 8

Summary of the Invention

Problems to be Solved by the Invention

[0005] However, in the prior art, an operator could not easily and accurately predict the inflow into the dam at any granularity.

[0006] In view of the above, the present disclosure is made, and an object thereof is to easily and accurately predict the inflow at a desired granularity.

Means for Solving the Problems

[0007] A prediction device according to an aspect of the present disclosure has a first prediction unit that predicts a temporal change in the inflow by repeatedly inputting rainfall and the inflow to a first neural network that predicts a minute temporal change in the water inflow into a dam, based on a continuous time simulation.

Effects of the Invention

[0008] The inflow can be easily and accurately predicted at a desired granularity.

Brief Description of the Drawings

[0009]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

[0010] Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings.

[0011] [Background, Prior Art and Its Problems] ·Background and Prior Art At the dam management site, the dam is operated by an operator while predicting the inflow of water into the dam so as not to cause emergency discharges or overflows. To support such operations, various prediction techniques for predicting the inflow have been proposed so far, and prediction systems for realizing those prediction techniques have also been developed.

[0012] As an example, a technique of constructing a prediction model based on physical phenomena and then predicting by that prediction model is known. For example, Patent Document 1 discloses a prediction technique using a tank model. Also, for example, Patent Document 2 discloses a prediction technique that utilizes a distributed outflow model, divides the prediction target area into a plurality of meshes, and constructs a detailed physical model using information such as the characteristics and gradients of the soil in each mesh.

[0013] In addition, there is also a technique of constructing a statistical model as a prediction model from given data using a neural network. For example, Patent Document 3 and Non-Patent Document 1 disclose a technique for predicting future water levels by utilizing a recurrent neural network (RNN). Further, for example, Patent Document 4 discloses a prediction technique that utilizes a fully connected neural network. Furthermore, for example, Non-Patent Document 2 reports a case where the prediction by a statistical model using a neural network is more accurate than the prediction by a physical model.

[0014] On the other hand, as a prediction technique using a statistical model with a neural network, Neural ODE (Neural Ordinary Differential Equations) has been proposed in Non-Patent Document 3. Non-Patent Document 3 discloses a technique for highly accurately predicting in combination with a continuous-time simulation after constructing a neural network that approximates minute changes. Further, Non-Patent Document 3 discloses a case where the prediction accuracy is higher for an object in which the time change is determined by a differential change (e.g., an object that changes according to physical principles) than when utilizing a recurrent neural network. Furthermore, Non-Patent Document 4 and Non-Patent Document 5 disclose cases where higher-accuracy prediction is possible by using Neural ODE than LSTM (Long Short Term Memory), which is one of the methods of using a neural network.

[0015] In addition, for example, Patent Document 5 and Patent Document 6 disclose cases where Neural ODE is utilized as a prediction model. In particular, Patent Document 5 discloses a technique for predicting the future state from the initial state of a plant by utilizing Neural ODE for a chemical plant, and also mentions its application to a dam.

[0016] Note that in Neural ODE, even for data measured discretely, predictions can be made at any time granularity from such data. Therefore, for example, using the prediction of rainfall at a one-hour granularity, predictions can be made at the granularity at which an operator changes the set value (e.g., a ten-minute granularity, etc.). On the other hand, in Neural ODE, it is difficult to predict a target whose state changes with future inputs, and as a technique for predicting the state with future inputs, it is necessary to utilize a technique called Neural CDE (Neural Controlled Differential Equations) (Non-Patent Documents 7-8).

[0017] ·Problems of the Prior Art In predictions using physical models (e.g., Patent Documents 1-2, etc.), for example, it is necessary to set parameters such as the penetration condition of the soil and the gradient direction in accordance with the actual situation, and it cannot be easily used. Also, for example, it is difficult to construct a physical model that can appropriately represent complex physical phenomena until rainwater flows into a dam.

[0018] RNN and LSTM, which are statistical models, can easily construct a highly accurate prediction model from data, but they cannot predict the inflow volume at any granularity according to the operator's operation.

[0019] On the other hand, Neural ODE can accurately predict the inflow volume at any granularity for the initial state, but there are no specific application cases for predicting the inflow volume into a dam. Also, when predicting the inflow volume into a dam, the forecast value of future rainfall is used as an input, but in Neural ODE, it was not possible to predict from future inputs.

[0020] In particular, in a situation where information can only be obtained discretely (e.g., at a one-hour granularity) like the forecast value of rainfall, it is more difficult to utilize Neural ODE that requires continuous input values. Note that there are also no specific application cases for Neural CDE, which adds future input elements to that Neural ODE, for predicting the inflow volume into a dam.

[0021] Therefore, hereinafter, a prediction device 10 that can easily and accurately predict the inflow volume into a dam at any granularity by utilizing a Neural ODE will be described. In addition, a learning device 20 that learns a neural network used when the prediction device 10 predicts the inflow volume into the dam will also be described. Note that the prediction device 10 and the learning device 20 may be configured by a single device or may be a system configured by a plurality of devices.

[0022] [Prediction Device 10] Hereinafter, a prediction device 10 that can easily and accurately predict the inflow volume into a dam at any granularity by utilizing a Neural ODE will be described. Note that hereinafter, the neural network used in the Neural ODE is denoted as f, and it is assumed that the neural network f has been trained.

[0023] Note that the neural network f is not limited to a neural network with a specific configuration, and it is possible to use a neural network with any configuration. As an example, it is possible to use a fully connected neural network as the neural network f.

[0024] <Hardware Configuration Example of Prediction Device 10> FIG. 1 is a diagram showing an example of the hardware configuration of a prediction device 10 according to an embodiment. As shown in FIG. 1, the prediction device 10 according to an embodiment includes an input device 101, a display device 102, an external I / F 103, a communication I / F 104, a RAM (Random Access Memory) 105, a ROM (Read Only Memory) 106, an auxiliary storage device 107, and a processor 108. These pieces of hardware are communicably connected to each other via a bus 109.

[0025] The input device 101 is, for example, a keyboard, a mouse, a touch panel, physical buttons, etc. The display device 102 is, for example, a display, a display panel, etc. Note that the prediction device 10 may not have at least one of the input device 101 and the display device 102, for example.

[0026] The external I / F 103 is an interface with an external device such as the recording medium 103a. Examples of the recording medium 103a include a CD (Compact Disc), a DVD (Digital Versatile Disk), an SD memory card (Secure Digital memory card), a USB (Universal Serial Bus) memory card, etc.

[0027] The communication I / F 104 is an interface for connecting to a communication network. The RAM 105 is a volatile semiconductor memory (storage device) that temporarily holds programs and data. The ROM 106 is a non-volatile semiconductor memory (storage device) that can hold programs and data even when the power is turned off. The auxiliary storage device 107 is a non-volatile storage device (storage device) such as an HDD (Hard Disk Drive), an SSD (Solid State Drive), a flash memory, etc. The processor 108 is an arithmetic device such as a CPU (Central Processing Unit), for example.

[0028] Note that the hardware configuration shown in FIG. 1 is an example, and the hardware configuration of the prediction device 10 is not limited to this. For example, the prediction device 10 may have a plurality of auxiliary storage devices 107 and a plurality of processors 108, may not have a part of the illustrated hardware, or may have various hardware other than the illustrated hardware.

[0029] <Example of the functional configuration of the prediction device 10> FIG. 2 is a diagram showing an example of the functional configuration of the prediction device 10 according to an embodiment. As shown in FIG. 2, the prediction device 10 according to an embodiment includes a rainfall forecast acquisition unit 201, an actual result acquisition unit 202, and an inflow prediction unit 203. Each of these units is realized, for example, by a process in which one or more programs installed in the prediction device 10 cause a processor 108 or the like to execute.

[0030] The rainfall forecast acquisition unit 201 acquires a rainfall forecast (hereinafter also referred to as a "rainfall forecast value") of a target area, which is an area that affects the inflow of a dam (hereinafter also referred to as the "target dam") for predicting the inflow. Specific examples of the target area include an area to which a river existing upstream of the target dam belongs, etc. Note that there may be a plurality of target areas. Also, the rainfall forecast value may be referred to as, for example, a "rainfall prediction value" or the like.

[0031] Hereinafter, the rainfall forecast value is denoted as {u a (t)|t = t1 a , ···, t s a}}. Also, for i = 1, ···, s - 1, let Δ a = t i+1 a - t i a . Here, Δ a is the prediction granularity of the rainfall forecast value. As Δ a , for example, a discrete granularity such as one hour is assumed. Note that the rainfall forecast acquisition unit 201 may acquire the rainfall forecast value from an external system that provides the rainfall forecast value, for example. Note that t ∈ {t1 a , ···, t s a} is a future time.

[0032] The actual result acquisition unit 202 acquires the actual rainfall value (hereinafter also referred to as the "rainfall actual result value") in the target area and the actual inflow value (hereinafter also referred to as the "inflow actual result value") of the target dam.

[0033] Hereinafter, the rainfall actual result value is denoted as {u b (t)|t = t1 b , ···, tr b Let it be so. Also, for i = 1, ···, r - 1, Δ b = t i+1 b - t i b Let it be so. Here, Δ b is the measurement granularity of the rainfall actual value. Δ b As, for example, Δ b = Δ p etc., discrete granularities are assumed. Note that t ∈ {t1 b , ···, t r b} are past times.

[0034] Similarly, hereinafter, the inflow actual values are {y c (t)|t = t1 c , ···, t k c} Let it be so. Also, for i = 1, ···, k - 1, Δ c = t i+1 c - t i c Let it be so. Here, Δ c is the measurement granularity of the inflow actual value. Δ c As, for example, Δ c = Δ b etc., discrete granularities are assumed. Note that t ∈ {t1 c , ···, t k c} are past times.

[0035] The inflow prediction unit 203 has the rainfall forecast values {u a (t)|t = t1 a , ···, t s a}, the rainfall actual values {u b (t)|t = t1 b , ···, t r b}, and the inflow actual values {y c (t)|t = t1 c , ···, t k cBased on this, the future inflow of the target dam is predicted. Here, the inflow prediction unit 203 includes an interpolation unit 211, a continuous-time simulation unit 212, and a neural network f. Note that the continuous-time simulation unit 212 and the neural network f constitute a Neural ODE.

[0036] The interpolation unit 211 uses a predetermined interpolation method (e.g., linear interpolation, interpolation by rounding, etc.) to interpolate the rainfall forecast values {u a (t)|t=t1 a ,···,t s a} to create continuous-granularity rainfall forecast values {u a (t)|t1 a ≦t≦t s a}. Similarly, the interpolation unit 211 uses a predetermined interpolation method to interpolate the rainfall actual values {u b (t)|t=t1 b ,···,t r b} to create continuous-granularity rainfall actual values {u b (t)|t1 b ≦t≦t r b}. Similarly, the interpolation unit 211 uses a predetermined interpolation method to interpolate the inflow actual values {y c (t)|t=t1 c ,···,t k c} to create continuous-granularity inflow actual values {y c (t)|t1 c ≦t≦t k c}. Note that continuous granularity means that the time t takes continuous values.

[0037] The continuous-time simulation unit 212 includes the neural network f, the rainfall forecast values {u a (t)|t1 a ≦t≦t s a}, and the rainfall actual values {u b (t)|t1 b ≦t≦t rb} and the actual inflow value {y c (t)|t1 c ≦ t ≦ t k c} are used to predict the future inflow of the target dam by continuous-time simulation. That is, the continuous-time simulation unit 212 inputs the rainfall amount and the inflow amount into the neural network f to calculate the minute-time change of the inflow amount, and repeats predicting the inflow amount at the next simulation time from the minute-time change of the inflow amount, thereby calculating the predicted value of the future inflow amount (hereinafter, also referred to as the "inflow amount predicted value").

[0038] Hereinafter, the inflow amount predicted value is {y p (t)|t1 ≦ t ≦ t n} is set. Also, for i = 1, ···, n - 1, h = t i+1 -t i is set. Here, h is the minute-time representing the simulation time width of the continuous-time simulation. Note that t ∈ [t1, t n is the future time.

[0039] <Operation example of prediction device 10> FIG. 3 is a flowchart showing an example of the operation of the prediction device 10 according to an embodiment.

[0040] The rainfall forecast acquisition unit 201 acquires the rainfall forecast value {u a (t)|t = t1 a , ···, t s a} of the target area (step S101).

[0041] The actual result acquisition unit 202 acquires the actual rainfall value {u b (t)|t = t1 b , ···, t r b} of the target area and the actual inflow value {y c (t)|t = t1 c , ···, t k c} of the target dam (step S102).

[0042] The interpolation unit 211 of the inflow prediction unit 203 uses a predetermined interpolation method to interpolate the rainfall forecast values {u a (t)|t=t1 a ,···,t s a}, the rainfall actual values {u b (t)|t=t1 b ,···,t r b}, and the inflow actual values {y c (t)|t=t1 c ,···,t k c} respectively (step S103). Thereby, continuous-granularity rainfall forecast values {u a (t)|t1 a ≦t≦t s a}, continuous-granularity rainfall actual values {u b (t)|t1 b ≦t≦t r b}, and continuous-granularity inflow actual values {y c (t)|t1 c ≦t≦t k c} are obtained.

[0043] The continuous-time simulation unit 212 of the inflow prediction unit 203 uses a neural network f, the rainfall forecast values {u a (t)|t1 a ≦t≦t s a}, the rainfall actual values {u b (t)|t1 b ≦t≦t r b}, and the inflow actual values {y c (t)|t1 c ≦t≦t k c} to predict the inflow prediction values {y p (t)|t1≦t≦t n} by continuous-time simulation (step S104). The continuous-time simulation unit 212 can perform continuous-time simulation using, for example, the Runge-Kutta method or the like.

[0044] When using the Runge - Kutta method, the continuous - time simulation unit 212 calculates the predicted inflow value \(y^{(i)}\) (\(t^{(i + 1)}\)) (\(1\leq i\leq n - 1\)) for the next time \(t^{(i + 1)}\) of the current simulation time \(t\) as follows. Here, assume that \(\frac{dy}{dt}=f(y,u)\). Also, let \(t_1\in[t_1,t]\), \(y(t_1)=y(t_1)\), and \(t\leq t\leq t + h\). i at the next time \(t\) i+1 (1\leq i\leq n - 1) p (t i+1 ). c ,t k c ,y p (t_1)=y c (t_1),t n \leq t s a .

[0045] y p (t i+1 )=y p (t i )+(h / 6)(L_1 + 2L_2+2L_3 + L_4) L_1=f(y p (t i ),u(t i )) L_2=f(y p (t i )+(h / 2)L_1,u(t i +(h / 2))) L_3=f(y p (t i )+(h / 2)L_2,u(t i +(h / 2))) L_4=f(y p (t i )+hL_3,u(t i +h)) Here, when \(t\in[t_1,t]\), \(u(t)=u(t)\); when \(t\in[t_1,t + h]\), \(u(t)=u(t)\). i ) is, when \(t i \in[t_1 a ,t s a ,u(t i )=u a (t i ); when \(t i \in[t_1 b ,t r b ,u(t i )=u b (ti ) is as follows.

[0046] As described above, the predicted inflow value {y p (t)|t1 ≤ t ≤ t n} can be obtained. However, the Runge - Kutta method is just an example of a method for realizing continuous - time simulation and is not limited to this. For example, continuous - time simulation may be realized by the Euler method or the like.

[0047] Note that the neural network f is defined as dy / dt = f(y, u), but this is just an example. For example, other information may be input to the neural network f. For example, when the flow rate z of a river with continuous or discrete granularity is obtained, dy / dt = f(y, z, u) may be used. In this case, when the flow rate z of the river has a discrete granularity, the interpolation unit 211 may create a flow rate z of the river with continuous granularity by interpolation.

[0048] Also, the inflow y and rainfall u input to the neural network f may be, for example, integral elements (e.g., moving average), differential elements (e.g., change amount in a short time), etc.

[0049] [Learning device 20] Hereinafter, a learning device 20 for learning the neural network f used in the prediction device 10 will be described. Note that the learning device 20 may be the same device as the prediction device 10 or may be a different device from the prediction device 10.

[0050] <Hardware configuration example of learning device 20> The hardware configuration example of the learning device 20 may be the same as that of the prediction device 10, so the description thereof is omitted.

[0051] <Functional configuration example of learning device 20> FIG. 4 is a diagram showing an example of the functional configuration of the learning device 20 according to an embodiment. As shown in FIG. 4, the learning device 20 according to an embodiment includes a rainfall forecast acquisition unit 201, an actual result acquisition unit 202, an inflow prediction unit 203, and a learning unit 204. Each of these units is realized, for example, by processing in which one or more programs installed in the learning device 20 cause a processor 108 or the like to execute. Note that since the rainfall forecast acquisition unit 201, the actual result acquisition unit 202, and the inflow prediction unit 203 of the learning device 20 are the same as those of the rainfall forecast acquisition unit 201, the actual result acquisition unit 202, and the inflow prediction unit 203 of the prediction device 10, respectively, their descriptions are omitted.

[0052] The learning unit 204 uses the predicted inflow values {y p (t)|t1≦t≦t n} and the actual inflow values {y c (t)|t=t1 c ,···,t k c} to train the neural network f (that is, update the parameters of the neural network f).

[0053] <Operation Example of Learning Device 20> FIG. 5 is a flowchart showing an example of the operation of the learning device 20 according to an embodiment. Note that since steps S201 to S204 in FIG. 5 are the same as steps S101 to S104 in FIG. 4, respectively, their descriptions are omitted.

[0054] Subsequent to step S204, the learning unit 204 uses the predicted inflow values {y p (t)|t1≦t≦t n} and the actual inflow values {y c (t)|t=t1 c ,···,t k c} to train the neural network f (step S205). Hereinafter, for simplicity, [t1 c ,t k c ⊂[t1,t nIt is assumed to be so. At this time, the learning unit 204 may update the parameters of the neural network f so as to minimize the loss function loss shown in the following equation (1).

[0055] loss=(1 / |T|)Σ t∈T |y c (t)-y p (t)| (1) Here, T = {t1 c ,···,t k c} and |T| = k.

[0056] Note that, for example, when a plurality of inflow amount actual values and a plurality of inflow amount predicted values are obtained, the loss function loss may be calculated using these. Specifically, when N discrete-granularity (rainfall forecast value, rainfall actual value, inflow amount actual value) are obtained, and N continuous-granularity (inflow amount actual value, inflow amount predicted value) are obtained from these N (rainfall forecast value, rainfall actual value, inflow amount actual value), the loss function loss shown in the following equation (2) may be calculated.

[0057] loss=(1 / N)(loss1+···+loss N ) (2) Here, loss i is the loss calculated by the above equation (1) using the i-th inflow amount actual value and inflow amount predicted value.

[0058] In addition, for example, instead of the above equation (1), the loss function loss shown in the following equation (3) may be calculated.

[0059] loss=(1 / |T|)Σ t∈T max(|y c (t)-y p (t)|) (3) In addition, for example, instead of the above equation (1), the loss function loss shown in the following equation (4) may be calculated.

[0060] loss=(1 / |T|)|Σ t∈T (y c(t)) - Σ t∈T (y p (t))| (4) Note that the loss functions loss shown in the above formulas (1) to (4) are all examples and are not limited thereto. For example, the error in the change in the water level of the target dam over a certain period may be minimized, or a predetermined index value (e.g., Nash - Sutcliffe coefficient, peak error (Non - Patent Document 6), etc.) may be minimized.

[0061] [Modification Example] Hereinafter, modification examples of the above - described embodiment will be described. Note that, as long as the following modification examples do not conflict with each other, a plurality of modification examples can be combined as appropriate.

[0062] · Modification Example 1 When the dam existing on the downstream side of the water system called the connected water system is the target dam, the inflow volume into the target dam is affected by the discharge volume of the dam on the upstream side. Therefore, when the target dam is the dam on the downstream side of the connected water system, it is preferable to predict the inflow volume using the planned value and the actual value of the discharge volume of the dam on the upstream side. Here, it is common that the planned value and the actual value of the discharge volume of the dam on the upstream side are obtained at discrete granularities.

[0063] Therefore, after making the planned value and the actual value of the discharge volume of the dam on the upstream side continuous by interpolation by the interpolation unit 211, the predicted value of the inflow volume is calculated by the neural network f using these planned value and actual value. Thereby, when the target dam is the dam on the downstream side of the connected water system, it becomes possible to predict the inflow volume with higher accuracy.

[0064] Note that the discharge volume of the dam on the upstream side may be the discharge volume of emergency gate discharge, the discharge volume used for power generation, or both of them.

[0065] · Modification Example 2 Similar to the above Modification Example 1, when the target dam is a dam on the downstream side of a linked water system, as shown in FIG. 6, the prediction device 10 includes an inflow prediction unit 203-1 that predicts an inflow prediction value from rainfall (forecast value, actual value) and inflow (actual value), an inflow prediction unit 203-2 that predicts an inflow prediction value from the discharge (planned value, actual value) of the upstream dam, and a total calculation unit 205 that calculates the sum of these inflow prediction values at the same time. In this case, the inflow prediction unit 203-1 is the same as the inflow prediction unit 203 in FIG. 2 (that is, the interpolation unit 211-1, the continuous time simulation unit 212-1, and the neural network f included in the inflow prediction unit 203-1 are the same as the interpolation unit 211, the continuous time simulation unit 212, and the neural network f in FIG. 2, respectively). On the other hand, the inflow prediction unit 203-2 includes an interpolation unit 211-2 that interpolates the discrete planned discharge value and the discrete measured actual discharge value into continuous granularity, a neural network g that predicts the minute time change of the inflow with the continuous granularity discharge (planned value, actual value) as input, and a continuous time simulation unit 212-2 that performs continuous time simulation. Note that the interpolation unit 211-2, the continuous time simulation unit 212-2, and the neural network g only have different inputs and outputs from the interpolation unit 211, the continuous time simulation unit 212, and the neural network f in FIG. 2, and the processes they execute are the same.

[0066] · Modification Example 3 The prediction device 10 may have a water level conversion unit that converts the inflow prediction value into a predicted value of the water level of the target dam (hereinafter, also referred to as "water level prediction value"). The water level conversion unit may convert the inflow prediction value {y p (t)|t1≦t≦t n} into the water level prediction value w(t) at time t, for example, as follows.

[0067] w(t)=w(t1)+Δw(t1)+···+Δw(t) Δw(t)=a×y(t) Here, a is a constant.

[0068] In addition, when the target dam discharges water, let the discharge at time t be x(t), and Δw(t)=a×(y(t)-x(t)).

[0069] However, although the conversion formula for the water level prediction value w(t) is a linear conversion, this is just an example. For example, depending on the shape of the target dam, etc., piecewise linear or non-linear conversions may be performed.

[0070] ·Modification Example 4 In the above embodiment, the neural network f is a model for predicting the minute-time change in the inflow rate. However, when it is desired to predict the water level prediction value w(t), the neural network f may be a model for predicting the minute-time change in the water level.

[0071] ·Modification Example 5 When learning the neural network f according to the above formula (2), if there is data on sunny days among the N discrete-granularity (rainfall forecast value, rainfall actual value, inflow rate actual value), that data may be deleted. This makes it possible to reduce data that does not contribute to the learning of the neural network f, and it becomes possible to speed up the learning process.

[0072] ·Modification Example 6 In the above embodiment, Neural ODE was used, but Neural CDE may be used instead of Neural ODE.

[0073] [Example] As an example of the above prediction device 10, the inflow rate to the target dam of the water system shown in FIG. 7 was predicted. That is, River A and River B exist upstream of the target dam, and the inflow rate to the target dam was predicted with the point immediately before the target dam as the prediction point of the inflow rate. Also, as the target area, an area within a predetermined range centered on latitude a1 and longitude b1, and an area within a predetermined range centered on latitude a2 and longitude b2 were used.

[0074] At this time, as the rainfall forecast value, the forecast values for six hours every hour announced by the Japan Meteorological Agency were used. Also, with h being 10 minutes, the predicted values of the inflow for up to 24 hours ahead were predicted using the Runge-Kutta method as a continuous-time simulation. When these predicted values of the inflow and the actual measured values of the actual inflow were compared, they were within the allowable error. Therefore, according to the prediction device 10 according to one embodiment, it can be said that the inflow could be predicted easily and accurately at a desired granularity (h = 10 minutes).

[0075] [Summary] As described above, in the prediction device 10 according to one embodiment, the predicted value of the inflow to the target dam is calculated by Neural ODE using the rainfall forecast value (that is, the predicted value of future rainfall), the actual measured value of the rainfall, and the actual measured value of the inflow to the target dam. At this time, in the prediction device 10 according to one embodiment, the predicted value of the inflow can be calculated at an arbitrary time granularity. Therefore, the operator of the target dam can easily obtain a predicted value of the inflow with good accuracy at the granularity he desires.

[0076] Note that although the above prediction device 10 is assumed to present the predicted value of the inflow to the operator, this is just an example. For example, the predicted value of the inflow may be output to the control device that controls the target dam. As a result, it becomes possible to automatically control the target dam based on the predicted value of the inflow, and the automation of the operation of the target dam can be realized.

[0077] The present invention is not limited to the above specifically disclosed embodiments, and various modifications, changes, combinations with known technologies, etc. are possible as long as they do not deviate from the gist described in the claims.

Explanation of Reference Numerals

[0078] 10 Prediction device 20 Learning device 101 Input device 102 Display device 103 External I / F 103a Recording medium 104 Communication I / F 105 RAM 106 ROM 107 Auxiliary storage device 108 Processor 109 Bus 201 Rainfall forecast acquisition unit 202 Performance acquisition unit 203 Inflow prediction unit 204 Learning unit 211 Interpolation unit 212 Continuous time simulation unit

Claims

According to claim 1, based on the predicted value of discretely predicted rainfall, the actual value of the inflow of water into the dam measured discretely, and a predetermined interpolation method, an interpolation unit that creates a continuous predicted value of the rainfall and a continuous actual value of the inflow volume. A first neural network that predicts the minute time change of the inflow of water into the dam, and based on continuous time simulation, repeatedly inputs the continuous predicted value of the rainfall and the continuous actual value of the inflow volume into the first neural network, thereby predicting the temporal change of the inflow volume at the granularity specified by the operator of the dam. A first prediction unit. It has The first prediction unit A prediction device that predicts the temporal change of the inflow volume by performing the first neural network and the continuous time simulation using the method of Neural ODE or Neural CDE. According to claim 2, a first neural network that predicts the minute time change of the inflow of water into the dam, and based on continuous time simulation, repeatedly inputs rainfall and the inflow volume into the first neural network, thereby predicting the temporal change of the inflow volume at the granularity specified by the operator of the dam, and predicting a first inflow volume representing the inflow of water into the dam. A first prediction unit. A second neural network that predicts the minute time change of the inflow of water into the dam, and based on continuous time simulation, repeatedly inputs the planned value and the actual value of the discharge of the upstream dam existing upstream of the dam into the second neural network, thereby predicting the temporal change of the inflow volume at the granularity and predicting a second inflow volume representing the inflow of water into the dam. A second prediction unit. A total calculation unit that calculates the total of the first inflow volume predicted by the first prediction unit and the second inflow volume predicted by the second prediction unit. It has The first prediction unit Predicts the temporal change of the inflow volume by performing the first neural network and the continuous time simulation using the method of Neural ODE or Neural CDE. The second prediction unit A prediction device that predicts the temporal change of the inflow volume by performing the second neural network and the continuous time simulation using the method of Neural ODE or Neural CDE.

3. A prediction device according to claim 1, comprising: an acquisition unit that acquires a predicted value of the rainfall predicted discretely and an actual value of the inflow measured discretely.

4. The acquisition unit acquires a predicted value of the rainfall in one or more regions that affect the inflow into the dam, the prediction device according to claim 3.

5. The first prediction unit also repeatedly inputs continuous planned values and actual values of the discharge of the dam to the first neural network, the prediction device according to claim 1.

6. A prediction device according to claim 1, comprising: a conversion unit that converts a temporal change in the inflow into a temporal change in the water level of the dam.

7. Based on a predicted value of rainfall predicted discretely, an actual value of the inflow of water into the dam measured discretely, and a predetermined interpolation method, an interpolation procedure for creating a continuous predicted value of the rainfall and a continuous actual value of the inflow; a first neural network that predicts a minute temporal change in the inflow of water into the dam, and a first prediction procedure for predicting a temporal change in the inflow at a granularity specified by the operator of the dam by repeatedly inputting the continuous predicted value of the rainfall and the continuous actual value of the inflow to the first neural network based on a continuous time simulation; which is executed by a computer, The first prediction procedure is a prediction method for predicting a temporal change in the inflow by performing the first neural network and the continuous time simulation using a method of Neural ODE or Neural CDE.

8. A first neural network that predicts a minute temporal change in the inflow of water into the dam, and a first prediction procedure for predicting a temporal change in the inflow at a granularity specified by the operator of the dam by repeatedly inputting rainfall and the inflow to the first neural network based on a continuous time simulation, and predicting a first inflow representing the inflow of water into the dam; Based on a second neural network that predicts minute temporal changes in the water inflow rate into the dam and a continuous-time simulation, by repeatedly inputting the planned and actual discharge values of an upstream dam existing upstream of the dam into the second neural network, a second prediction procedure for predicting the temporal change in the inflow rate at the said granularity and predicting a second inflow rate representing the water inflow rate into the dam. A total calculation procedure for calculating the total of the first inflow rate predicted by the first prediction procedure and the second inflow rate predicted by the second prediction procedure. which is executed by a computer. The first prediction procedure is By using the method of Neural ODE or Neural CDE to perform the first neural network and the continuous-time simulation, the temporal change in the inflow rate is predicted. The second prediction procedure is A prediction method for predicting the temporal change in the inflow rate by using the method of Neural ODE or Neural CDE to perform the second neural network and the continuous-time simulation. According to claim 9, an interpolation procedure for creating a continuous predicted rainfall value and a continuous actual inflow rate value based on a predicted value of discretely predicted rainfall, an actual value of the water inflow rate into the dam measured discretely, and a predetermined interpolation method. Based on a first neural network that predicts minute temporal changes in the water inflow rate into the dam and a continuous-time simulation, by repeatedly inputting the continuous predicted rainfall value and the continuous actual inflow rate value into the first neural network, a first prediction procedure for predicting the temporal change in the inflow rate at the granularity specified by the operator of the dam. to be executed by a computer. The first prediction procedure is A program for predicting the temporal change in the inflow rate by using the method of Neural ODE or Neural CDE to perform the first neural network and the continuous-time simulation.

10. A first prediction procedure for predicting a first inflow rate representing the water inflow rate into the dam by repeatedly inputting rainfall and the inflow rate into a first neural network based on a first neural network that predicts minute-time changes in the water inflow rate into the dam and a continuous-time simulation, thereby predicting the temporal change in the inflow rate at a granularity specified by the operator of the dam. A second prediction procedure for predicting a second inflow rate representing the water inflow rate into the dam by repeatedly inputting planned and actual discharge values of an upstream dam existing upstream of the dam into a second neural network based on a second neural network that predicts minute-time changes in the water inflow rate into the dam and a continuous-time simulation, thereby predicting the temporal change in the inflow rate at the granularity. A total calculation procedure for calculating the total of the first inflow rate predicted by the first prediction procedure and the second inflow rate predicted by the second prediction procedure. To be executed by a computer. The first prediction procedure is as follows. By using the method of Neural ODE or Neural CDE, performing the first neural network and the continuous-time simulation to predict the temporal change in the inflow rate. The second prediction procedure is as follows. A program for predicting the temporal change in the inflow rate by using the method of Neural ODE or Neural CDE and performing the second neural network and the continuous-time simulation.

Images