Learning device, prediction device, and prediction system

Abstract

The present invention provides a learning device, a prediction device, and a prediction system that are capable of adjusting the prediction accuracy of a wind speed range of interest. [Solution] A learning device according to this embodiment includes a model learning unit and a loss function generation unit. The model learning unit uses explanatory variables as input and learns a prediction model so as to reduce the loss value of a loss function that includes an error between a first objective variable, which is a teacher signal, and a second objective variable, which is the output of the prediction model. The loss function generation unit changes the contribution of the loss function to the loss value based on at least one of the error between the first objective variable and the second objective variable and the value of the first objective variable.

Description

[Technical Field]

[0001] FIELD Embodiments of the present invention relate to a learning device, a prediction device, and a prediction system. [Background technology]

[0002] In the operation management of transportation infrastructure such as roads, railways, and airports, actual and predicted wind speeds along routes are monitored and used as a basis for determining restrictions due to strong winds. Wind speed predictions are sometimes made using a prediction model that learns from actual data using machine learning. In these cases, it is common to predict future wind speeds in the time direction from wind speed values ​​from the past to the present, or to predict future wind speeds at a destination in the space direction from wind speed forecasts, which are future wind speeds at a given location. In wind speed predictions for such transportation infrastructure operation management, if the predicted value for the strong wind side is output on the lower wind side than the actual wind speed, this could pose a safety issue. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Laid-Open No. 2017-187371 Summary of the Invention [Problem to be solved by the invention]

[0004] However, in general prediction model training, the loss function used is the mean squared error or mean absolute error, so the prediction accuracy (the predicted value in the prediction model) is uniform regardless of wind speed, making it difficult to adjust the prediction accuracy for the wind speed range of interest.

[0005] The problem to be solved by the present invention is to provide a learning device, a prediction device, and a prediction system that can improve the prediction accuracy of a wind speed range of interest. [Means for solving the problem]

[0006] The learning device according to this embodiment includes a model learning unit and a loss function generation unit. The model learning unit uses explanatory variables as input and learns a prediction model so as to reduce the loss value of a loss function that includes an error between a first objective variable, which is a teacher signal, and a second objective variable, which is the output of the prediction model. The loss function generation unit changes the contribution of the loss function to the loss value based on at least one of the error between the first objective variable and the second objective variable and the value of the first objective variable. [Effects of the Invention]

[0007] This makes it possible to improve the accuracy of predictions for the wind speed range of interest. [Brief explanation of the drawings]

[0008] [Figure 1] FIG. 1 is a block diagram showing the configuration of a prediction system according to an embodiment of the present invention. [Figure 2] FIG. 10 is a diagram showing an example map illustrating an example of wind condition measurements. [Figure 3] FIG. 4 is a block diagram showing a detailed configuration example of a learning unit. [Figure 4] FIG. 4 is a diagram schematically showing the flow of data in a learning unit. [Figure 5] FIG. 10 is a diagram showing an example of a frequency distribution when a loss function is used. [Figure 6] FIG. 10 is a diagram showing an example of frequency distribution when a loss function that improves the accuracy of the predicted value on the strong wind side is used. [Figure 7] FIG. 10 is a diagram showing an example of a prediction result of a prediction model trained using a loss function. [Figure 8] FIG. 8 is a diagram showing the relationship between the predicted values ​​and the actually measured values ​​of FIG. 7. [Figure 9] FIG. 8 is a diagram showing the frequency of errors between the predicted values ​​and the measured values ​​of FIG. 7. [Figure 10] FIG. 10 is a diagram showing an example of a prediction result of a prediction model trained using a loss function. [Figure 11] FIG. 11 is a graph showing the relationship between the predicted values ​​and the measured values ​​shown in FIG. 10 . [Figure 12] FIG. 11 is a graph showing the frequency of errors between the predicted values ​​and the measured values ​​of FIG. 10 . [Figure 13] 10 is a flowchart showing an example of calculation processing of the learning device. [Figure 14] 10 is a flowchart showing an example of calculation processing of a prediction device. [Figure 15] FIG. 11 is a diagram showing an example of a map illustrating an example of measurement of wind conditions according to the second embodiment. [Figure 16] FIG. 11 is a block diagram of a learning device according to a third embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0009] Hereinafter, a learning device, a prediction device, and a prediction system according to embodiments of the present invention will be described in detail with reference to the drawings. Note that the embodiments described below are examples of embodiments of the present invention, and the present invention should not be interpreted as being limited to these embodiments. Furthermore, in the drawings referred to in this embodiment, identical parts or parts having similar functions are denoted by the same or similar reference numerals, and repeated explanations thereof may be omitted. Furthermore, for convenience of explanation, the dimensional ratios of the drawings may differ from the actual ratios, and some components may be omitted from the drawings.

[0010] (First embodiment) Fig. 1 is a block diagram showing the configuration of a prediction system 1 according to this embodiment. As shown in Fig. 1, the prediction system 1 according to this embodiment is a system that can prioritize improving the prediction accuracy of an output value range of interest when learning a prediction model, and includes a learning device 10 and a prediction device 20.

[0011] The learning device 10 is a device that learns a prediction model based on wind speed and wind direction at a predetermined point. The learning device 10 has a first acquisition unit 100, a learning unit 102, a first storage unit 104, and an operation unit 106. The learning device 10 is configured to include, for example, a CPU (Central Processing Unit).

[0012] The prediction device 20 is a device that predicts wind speed using a prediction model learned by the learning unit 102. The prediction device 20 includes a second acquisition unit 200, a prediction unit 202, a second storage unit 204, a display control unit 206, and a display unit 208. The prediction device 20 is configured to include, for example, a CPU. Details of the prediction device 20 will be described later.

[0013] FIG. 2 is a diagram showing an example of a map illustrating an example of wind condition measurements. The mesh of horizontal lines 30 and vertical lines 32 corresponds to the mesh of the Local Forecast Model (LFM) of the Japan Meteorological Agency, for example. In this way, this mesh divides a two-dimensional plane into grid-like sections. The mesh interval is, for example, 2 kilometers. Note that in this embodiment, a mesh corresponding to the Local Numerical Prediction Model is used, but this is not limiting. For example, a mesoscale model (MSM) or a weather model (WRF), which are numerical prediction models calculated with a grid interval (5 km), may also be used.

[0014] An anemometer or the like is installed at point 42a of transportation infrastructure 40 where strong winds blow. The mesh intersection of the Japan Meteorological Agency's local numerical forecast model nearest to point 42a is designated as 42b. The anemometer at point 42a measures time history data of wind conditions, including wind speed and wind direction. These anemometers measure wind speed and wind direction, for example, every 30 minutes. At mesh intersection 42b, meteorological data, including wind speed and wind direction, based on the local numerical forecast model is forecast in 30-minute increments for up to 10 or 18 hours in advance.

[0015] The first acquisition unit 100 communicates with the anemometer at point 42a via the network NW to acquire observation data including wind speed and wind direction, and forecast data. The observation data and forecast data are stored in the first storage unit 104, which will be described later, in association with the observation point, mesh intersection, and time.

[0016] The network NW includes, for example, a wide area network (WAN), a local area network (LAN), the Internet, a dedicated line, a wireless base station, a provider, etc. The observation data received by the first acquisition unit 100 is stored in the first storage unit 104. The first acquisition unit 100 also acquires wind direction and wind speed, which are forecast data (analysis data) for mesh intersections in 30-minute increments, by communicating with a server device of the Japan Meteorological Agency via the network NW, and stores the data in the first storage unit 104.

[0017] The learning unit 102 learns a prediction model for predicting the wind speed at a target point. Details of the learning unit 102 will be described later.

[0018] The first storage unit 104 is configured, for example, by a hard disk drive (HDD) or a solid state drive (SSD). The first storage unit 104 stores observation data and data related to local numerical weather forecast models in association with locations and times. The first storage unit 104 also stores learning programs. As a result, the learning unit 102 configures each unit by, for example, executing the programs stored in the first storage unit 104.

[0019] The operation unit 106 is configured with, for example, a mouse and a keyboard. The operation unit 106 can set parameters in equations (1) to (5) described below. For example, the operator can use the operation unit 106 to set α, β, th, etc. in equations (1) to (5).

[0020] Here, an example configuration of the learning unit 102 will be described with reference to Fig. 3 and Fig. 4. Fig. 3 is a block diagram showing a detailed example configuration of the learning unit 102. The learning unit 102 includes a learning data generation unit 106, a loss function generation unit 108, and a model learning unit 110. Fig. 4 is a diagram schematically showing the flow of data in the learning unit 102.

[0021] In the case of prediction in the time direction, the learning data generation unit 106 generates learning data for training a prediction model using the observation data stored in the first storage unit 104. In this embodiment, a learning set is generated using the average wind speed and wind direction every 30 minutes from the present time onward at point 42a, which is the target location, as the input signal, and the maximum instantaneous wind speed at point 42a from the present time onward as the teacher signal.

[0022] Furthermore, in the case of spatial prediction, the input signal uses the forecast value obtained by the local numerical prediction model at the mesh intersection 42b. That is, in this embodiment, there is a high correlation between the forecast value obtained by the local numerical prediction model at the mesh intersection 42b and the actual measured value at the point 42a. Furthermore, the time interval between the average values ​​of the actual measured values ​​at the point 42a is made to correspond to the time interval between the average values ​​predicted by the local numerical prediction model at the mesh intersection 42b. This makes it possible to predict the maximum instantaneous wind speed at the point 42a within 30-minute intervals based on the forecast value obtained by the local numerical prediction model at the mesh intersection 42b 10 hours or 18 hours in advance, as will be described later.

[0023] The loss function generating unit 108 generates a loss function, for example, one of the equations (1) to (5) described below, depending on the purpose. For example, equations (1) to (3) are loss functions that calculate a loss value (loss) used to reduce underestimation of predicted values. For example, equations (4) and (5) are loss functions used to improve prediction accuracy in strong winds more than in weak winds. Equation (6) is a loss function of a comparative example. These will be described in detail later.

[0024] The loss function generation unit 108 generates values ​​such as the coefficient α and constant β of the loss function of one of the equations (1) to (5), for example, based on at least one of the prediction error, which is the difference between the output value yi of the prediction model and the actual measurement value yi, and the magnitude of the teacher signal (wind speed). In this way, the loss function generation unit 108 generates a loss function using the coefficient α, constant β, etc. generated by the loss function generation unit 108. For example, the model training unit 110 uses explanatory variables as input and trains a prediction model so as to reduce the loss value of the loss function, which includes the error between the first objective variable, which is the teacher signal, and the second objective variable, which is the output of the prediction model.

[0025] For example, the loss function generation unit 108 changes the contribution of the loss function to the loss value based on at least one of the error between the first objective variable and the second objective variable and the value of the first objective variable. For example, the explanatory variables may be weather forecast information including wind speed and wind direction at a time after the present at a mesh point of the weather analysis, and the first objective variable may be wind speed at a time after the present at the prediction target point. Alternatively, the explanatory variables may be actual weather information including wind speed and wind direction at a time before the present at the prediction target point, and the first objective variable may be wind speed at a time after the present at the prediction target point. In this way, the loss function generation unit 108 changes the contribution of the loss function to the loss value (loss) by changing values ​​of the coefficient a, constant β, etc. based on at least one of the error and the magnitude of the teacher signal (wind speed).

[0026] As shown in FIG. 4, the prediction model according to this embodiment is a model in which, for example, a 30-minute average wind speed and a 30-minute average wind direction are input In110, and a maximum instantaneous wind speed is output Out110. The prediction model is, for example, a regression model using a neural network. The model learning unit 110 learns the prediction model by sequentially updating the inter-node coupling parameters of the neural network so that the loss value (loss) is minimized. For example, the model learning unit 110 learns the inter-node coupling parameters of the neural network by backpropagation or the like so that the loss value (loss) of the loss function is minimized.

[0027] A general learning algorithm can be used as the learning algorithm in the model learning unit 110. Note that a linear regression model, a decision tree-based regression model, or the like can also be used as the prediction model. In this case, the model learning unit 110 learns the linear regression model, the decision tree-based regression model, or the like so that the loss value (loss) of the loss function becomes the minimum value.

[0028] Here, equations (1) to (5) will be explained with reference to FIGS. 5 and 6. FIG. 5 is a diagram showing an example of a frequency distribution when a loss function used to reduce underestimation of a predicted value is used. The horizontal axis represents the error of the predicted value of the prediction model after learning, and the vertical axis represents the frequency of occurrence of the error. The error here is the difference between the output value yi hat of the prediction model in equations (1) to (3) and (6) and the actual measured value yi. The frequency distribution line E10 is an example of a frequency distribution when a loss value "loss" used to reduce underestimation of a predicted value using equations (1) to (3), for example, is used. The frequency distribution line N10 is a frequency distribution when a loss value "loss" of a comparative example is used using equation (6). When the predicted value yi hat is larger than the actual measured value yi, it is considered an overestimation, and when the predicted value yi hat is smaller than the actual measured value yi, it is considered an underestimation.

[0029]

number

[0030] n is the number of training data. As described above, in the model training unit 110, the prediction model is trained so that the loss value (loss) of the loss function is minimized. As can be seen from this, the prediction model is trained preferentially to minimize the difference between the predicted value yi hat, which is assigned a coefficient α greater than 1 and has a large contribution, and the actual measured value yi. In other words, the prediction model is trained preferentially to minimize the difference between the predicted value yi hat to which the coefficient α is assigned and the actual measured value yi, over the difference between the predicted value yi hat and the actual measured value yi to which the coefficient α is not assigned. Therefore, the underestimated difference value, in which the predicted value yi hat is smaller than the actual measured value yi, is trained preferentially so that the difference approaches zero. As a result, the frequency distribution line E10 of the predicted values ​​obtained by the trained prediction model shifts the error of the underestimated predicted value toward zero relative to the frequency distribution line N10, and the error of the overestimated predicted value toward a larger error. This makes it possible to further suppress the predicted value on the underestimated side of the maximum instantaneous wind speed.

[0031]

number

[0032] As can be seen from this, the prediction model is trained preferentially so that the difference between the predicted value yi, which is assigned a constant β greater than 0, and the actual measured value yi becomes smaller. Therefore, training is preferentially performed so that the underestimated difference value, where the predicted value yi is smaller than the actual measured value yi, approaches 0. As a result, in the predicted values ​​obtained by the trained model, the error in the underestimated predicted value shifts toward 0, and the error in the overestimated predicted value shifts toward a larger error. This makes it possible to further suppress the predicted value on the underestimated side of the maximum instantaneous wind speed.

[0033]

number

[0034] Here, a case where a predicted value on the excessive side of the maximum instantaneous wind speed is suppressed will be described. In equations (1) to (3), when the predicted value yi is underestimated (i.e., when the predicted value yi is smaller than the measured value yi), a coefficient α greater than 1 and a constant β greater than 0 are assigned. Conversely, when the predicted value yi is overestimated, a coefficient α greater than 1 and a constant β greater than 0 may be assigned. More specifically, in equations (1) to (3), when the predicted value yi is overestimated (i.e., when the predicted value yi is larger than the measured value yi), a coefficient α greater than 1 or a constant β greater than 0 is assigned. This shifts the error in the overestimation case toward 0, and the error in the underestimation case shifts toward a larger error. This makes it possible to further suppress predicted values ​​on the excessive side of the maximum instantaneous wind speed.

[0035] Figure 6 shows an example of a frequency distribution when using a loss value (loss) used to further improve the accuracy of the strong wind side prediction value. The horizontal axis represents the error in the prediction value of the prediction model after learning, and the vertical axis represents the frequency of occurrence of the error. The error here is the difference between the output value yi (hat) of the prediction model in equations (4), (5), and (6) and the actual measured value yi. The frequency distribution line H10 is an example of a frequency distribution of the strong wind side error due to the loss value (loss) used to further improve the accuracy of the strong wind side prediction value, for example, using equations (4) and (5). The frequency distribution line L10 is an example of a frequency distribution of the weak wind side error due to the loss value (loss) used to further improve the accuracy of the strong wind side prediction value, for example, using equations (4) and (5).

[0036]

number

[0037]

number

[0038] As can be seen from this, the prediction model is trained preferentially so that the difference between the predicted value yihat, which is assigned a coefficient α greater than 1, and the actual measured value yi becomes smaller. Therefore, the actual measured value yi is trained preferentially so that the error approaches 0 on the strong wind side, where the wind speed is greater than the threshold th. As a result, the distribution width centered on error 0 on the error frequency distribution line H10 on the strong wind side, where the wind speed is greater than the threshold th, becomes narrower than the distribution width centered on error 0 on the error frequency distribution line L10 on the weak wind side, where the wind speed is slower than the threshold th. In this way, by assigning the coefficient α, it is possible to adjust the frequency distribution of errors produced by the prediction model.

[0039]

number

[0040] 1, the prediction device 20 will be described in detail. The prediction device 20 includes a second acquisition unit 200, a prediction unit 202, a second storage unit 204, a display control unit 206, and a display unit 208.

[0041] Like the first acquisition unit 100, the second acquisition unit 200 communicates with the anemometer at the mesh intersection 42b via the network NW and can acquire observation data 132, including wind speed and wind direction. This observation data is associated with the observation point and time and stored in the second storage unit 204 (described later). When using data from the Japan Meteorological Agency's local numerical weather forecast model, it is possible to acquire forecast data for the mesh intersection, such as wind direction and wind speed, in 30-minute increments for up to 10 or 18 hours in advance. The network NW includes, for example, a WAN, a LAN, the Internet, a dedicated line, a wireless base station, and a provider. The second acquisition unit 200 also acquires a forecast model for point 42a from the learning device 10 and stores it in the second storage unit 204.

[0042] The prediction unit 202 predicts the instantaneous wind speed at point 42a using the prediction model for point 42a. This prediction unit 202 uses as input the 30-minute average wind speed and wind direction actually measured by the anemometer at mesh intersection 42b. In this case, it is possible to predict the maximum instantaneous wind speed up to 30 minutes into the future.

[0043] On the other hand, the prediction unit 202 can also predict the maximum instantaneous wind speed at point 42a using predicted values ​​up to 10 or 18 hours ahead based on a local numerical forecast model at mesh intersection 42b. In this case, it becomes possible to predict the maximum instantaneous wind speed at point 42a up to 10 or 18 hours ahead in 30-minute increments.

[0044] The second storage unit 204 is configured with, for example, an HDD (hard disk drive) or an SSD (solid state drive), similar to the first storage unit 104. The second storage unit 204 stores a prediction program. As a result, the prediction device 20 configures each unit by executing, for example, the program stored in the second storage unit 204.

[0045] The display control unit 206 causes the display unit 208 to display result data based on the prediction result of the prediction unit 202. For example, the display control unit 206 can create charts such as those shown in Figs. 7 to 12 (described later) and display them as image information on the display unit 208. The display unit 208 is, for example, a monitor, and displays the image information generated by the display control unit 206.

[0046] Here, an example of a prediction result of a prediction model for point 42a trained using the loss value loss shown in equation (1) will be described with reference to Figures 7 to 9. Also, as a comparative example, an example of a prediction result of a prediction model for point 42a trained using the loss value loss shown in equation (6) will be described with reference to Figures 10 to 12.

[0047] FIG. 7 shows an example of a prediction result of a prediction model trained using the loss value (loss) shown in equation (1). The horizontal axis represents time, and the vertical axis represents the predicted value. The prediction interval is 30 minutes. Here, the maximum instantaneous wind speed within the next 30 minutes of the input value is predicted.

[0048] Figure 8 is a diagram showing the relationship between the predicted values ​​and the actual measured values ​​in Figure 7. The horizontal axis represents the actual measured values, and the vertical axis represents the predicted values. Dashed line L8 is the line where the actual measured values ​​and the predicted values ​​are the same. Values ​​below dashed line L8 are underestimated values, and values ​​above dashed line L8 are overestimated values.

[0049] Fig. 9 is a diagram showing the frequency of errors between the predicted values ​​and the measured values ​​in Fig. 7. The horizontal axis indicates the error, and the vertical axis indicates the frequency value.

[0050] FIG. 10 shows an example of a prediction result of a prediction model trained using the loss value (loss) shown in Equation (6). The horizontal axis represents time, and the vertical axis represents the predicted value. The prediction interval is 30 minutes. Here, the maximum instantaneous wind speed within the next 30 minutes of the input value is predicted.

[0051] Fig. 11 is a diagram showing the relationship between the predicted values ​​and the actual measured values ​​in Fig. 10. The horizontal axis represents the actual measured values, and the vertical axis represents the predicted values. The dashed line L8 is the line where the actual measured values ​​and the predicted values ​​are the same. Values ​​below dashed line L8 are underestimated values, and values ​​above dashed line L8 are overestimated values.

[0052] Fig. 12 is a diagram showing the frequency of errors between the predicted values ​​and the measured values ​​in Fig. 10. The horizontal axis indicates the error, and the vertical axis indicates the frequency value.

[0053] Comparing the predicted values ​​in Fig. 7 with those in Fig. 10, the prediction results of the prediction model trained using the loss value shown in equation (1) tend to show larger values ​​overall. This becomes clearer when comparing Fig. 8 with Fig. 11. In Fig. 8, the number of distributions on the underestimation side is smaller than the number of distributions on the underestimation side shown in Fig. 11. In other words, by using the loss value shown in equation (1), the predicted values ​​on the underestimation side of the maximum instantaneous wind speed are further suppressed.

[0054] Furthermore, when comparing the error distribution values ​​in Fig. 8 with those in Fig. 10, the error distribution values ​​in Fig. 8 are shifted toward values ​​greater than 0. Comparing these figures also shows that by using the loss value (loss) shown in equation (1), predicted values ​​on the undersized side of the maximum instantaneous wind speed can be further suppressed. In this way, as shown in equation (1), by applying the coefficient α, it is possible to further suppress predicted values ​​on the undersized side of the maximum instantaneous wind speed.

[0055] The above is a description of the configuration of the prediction system 1 according to this embodiment. Below, we will explain an example of the operation of the prediction system 1. Figure 13 is a flowchart showing an example of the calculation process of the learning device 10 according to this embodiment. Here, we will explain an example of the process after generating learning data.

[0056] 13, the model learning unit 110 generates a prediction model using a neural network with two input cells, one output cell, and two hidden layers (step S100). The coupling coefficient (parameter) of each cell is set based on a random number, and i is set to 1.

[0057] Next, the i-th input signal of the training data is input to the prediction model, a predicted value is calculated, and 1 is added to i (step S102). Subsequently, the loss function generation unit 108 calculates the difference between the actual measurement value, which is the teacher signal, and the predicted value. Subsequently, the loss function generation unit 108 determines whether the predicted value is smaller than the actual measurement value, and generates a coefficient α if the predicted value is smaller than the actual measurement value, and generates a coefficient 1 if the predicted value is equal to or greater than the actual measurement value (step S104).

[0058] Next, the loss function generation unit 108 multiplies the difference between the actual measurement value and the predicted value by the coefficient generated by the loss function generation unit 108 to generate a loss value (step S106). Next, the model learning unit 110 changes the coupling coefficient (parameter) of the prediction model so as to reduce the loss value (step S108).

[0059] Next, the model learning unit 110 determines whether i is equal to or greater than n (step S110), and if i is less than n (N in step S110), repeats the processing from step S102. On the other hand, if i is equal to or greater than n (Y in step S110), the model learning unit 110 determines whether the number of learning iterations has exceeded a predetermined number of times (step S112), and if it has not exceeded that number (N in step S112), repeats the processing from step S102. On the other hand, if it has exceeded that number (Y in step S112), learning is terminated. In this way, the prediction model is learned while generating a loss value "loss" to which the coefficient α is assigned.

[0060] 14 is a flowchart showing an example of calculation processing of the prediction device 20 according to this embodiment. Here, an example will be described in which predicted values ​​up to 18 hours ahead based on a local numerical prediction model are used.

[0061] As shown in FIG. 14, the second acquisition unit 200 acquires the prediction model of the point 42a from the learning device 10, stores it in the second storage unit 204, and sets it in the prediction unit 202 (step S200).

[0062] Next, the second acquisition unit 200 acquires predicted values ​​for the mesh intersection 42b up to 18 hours ahead based on a local numerical weather forecast model via the network (step S202). Subsequently, the prediction unit 202 sequentially inputs predicted values ​​for the mesh intersection 42b up to 18 hours ahead in 30-minute increments using the prediction model for point 42a, predicts the maximum instantaneous wind speed at point 42a in 30-minute increments (step S204), and ends the prediction process. In this way, when predicted values ​​based on a local numerical weather forecast model are used, predictions up to 18 hours ahead are possible.

[0063] As described above, according to this embodiment, when generating the loss value (loss) of the loss function, the learning unit 102 assigns the coefficient α and the constant β according to the prediction error of the prediction model or the value of the actual measurement value, which is the teacher signal. This makes it possible to adjust the prediction accuracy for the wind speed range of interest.

[0064] (Second embodiment) The prediction system 1 according to the second embodiment differs from the prediction system 1 according to the first embodiment in that the prediction system 1 according to the second embodiment uses predicted values ​​obtained by a local numerical prediction model of a plurality of mesh intersections 42b to 42e for the point 42a. The differences from the prediction system 1 according to the first embodiment will be described below.

[0065] The learning data generation unit 106 generates learning data that associates predicted values ​​(analysis values) for multiple mesh intersections 42c-g for up to 10 hours or 18 hours ahead based on a local numerical weather forecast model at point 42a with the maximum instantaneous wind speed at point 42a. The learning data generation unit 106 generates learning data that predicts the maximum instantaneous wind speed at point 42a using predicted values ​​for multiple mesh intersections 42c-g for up to 10 hours or 18 hours ahead based on a local numerical weather forecast model at point 42a. The model learning unit 110 uses this learning data to train a prediction model.

[0066] Similarly, the prediction unit 202 predicts the maximum instantaneous wind speed at point 42a using predicted values ​​at multiple mesh intersections 42c to 42g for up to 10 hours or 18 hours ahead based on the local numerical weather forecast model at point 42a. Here, for simplicity of explanation, only point 42a will be described.

[0067] FIG. 15 is a diagram showing an example of a map illustrating an example of wind condition measurements according to the second embodiment. A mesh of horizontal lines 30 and vertical lines 32 corresponds to, for example, a mesh of the Local Forecast Model (LFM) of the Japan Meteorological Agency. The mesh interval is, for example, 2 kilometers. Point 42a is surrounded by mesh intersections 42b-e. In the model learning unit 110, the prediction unit 202 weights and adds the 30-minute average wind speed and 30-minute average wind direction of the multiple mesh intersections 42b-e surrounding point 42a.

[0068] For example, the prediction unit 202 performs a weighted addition of the 30-minute average wind speed and 30-minute average wind direction of the multiple mesh intersections 42b-e using the inverse of the distance from point 42a to mesh intersections 42b-e. The sum of the weighted values ​​is normalized to 1.0. For example, if the 30-minute average wind speed and 30-minute average wind direction obtained by adding the weighted values ​​of the multiple mesh intersections 42b-e are input to the prediction model, the prediction stability can be improved in complex terrain such as mountainous areas.

[0069] Furthermore, the training data generation unit 106 changes the positions of the multiple mesh intersections 42b-e so that the magnitude of the loss function at the end of training of the training model in the model training unit 110 is reduced. In this case, the prediction unit 202 may not limit the mesh intersections 42b-e surrounding point 42a to the four nearest points, but may also include the 30-minute average wind speed and 30-minute average wind direction of surrounding mesh intersections. For example, in cases where wind is blowing into point 42a from multiple valleys, the prediction accuracy may improve as the range of mesh intersections is expanded.

[0070] As described above, according to this embodiment, the predicted values ​​(analysis values) of the plurality of mesh intersections 42b to 42g are used for prediction by the prediction model. This allows for more stable prediction even when point 42a is located in complex terrain.

[0071] (Third embodiment) The prediction system 1 according to the third embodiment differs from the prediction system 1 according to the first embodiment in that it further includes an analysis unit that analytically sets α, β, th, etc. in equations (1) to (5). The differences from the prediction system 1 according to the first embodiment will be described below.

[0072] FIG. 16 is a block diagram of the learning device 10 of the prediction system 1 according to the third embodiment. As shown in FIG. 16, the learning device 10 further includes an analysis unit 114 that analytically sets α, β, th, and the like in equations (1) to (5). After the model learning unit 110 has completed learning, the analysis unit 114 generates a distribution diagram showing the frequency of errors between predicted values ​​and actual measured values ​​for the learning data (see FIG. 9). The analysis unit 114 also calculates the skewness of the generated distribution diagram and determines whether it matches the desired distribution. If the distribution is not the desired distribution, the analysis unit 114 can have the model learning unit 110 repeat learning by changing the values ​​of the parameters α, β, th, and the like, to generate parameters α, β, and th that approach the desired skewness.

[0073] As described above, according to this embodiment, the analysis unit 114 evaluates the error distribution (see FIG. 9) using the training data after the model training unit 110 has trained the parameters α, β, th, etc., and generates parameters α, β, th that approach the desired distribution. In this way, the analysis unit 114 evaluates the error distribution (see FIG. 9) using the training data at the end of model training, and changes the method of contribution of the loss function to the loss value (loss). This makes it possible to bring the distribution diagram showing the frequency of errors between predicted values ​​and actual values ​​closer to the desired shape, even for untrained data.

[0074] At least a portion of the prediction system 1 described in the above embodiment may be configured with hardware or software. If configured with software, a program that realizes at least a portion of the functions of the learning device 10 and the prediction device 20 may be stored on a recording medium such as a flexible disk or CD-ROM and read and executed by a computer. The recording medium is not limited to removable recording media such as magnetic disks or optical disks, but may also be fixed recording media such as a hard disk unit or memory.

[0075] In addition, a program that realizes at least some of the functions of the prediction system 1 may be distributed via a communication line (including wireless communication) such as the Internet. Furthermore, the program may be encrypted, modulated, or compressed and distributed via a wired line or wireless line such as the Internet, or stored on a recording medium.

[0076] Although several embodiments have been described above, these embodiments are presented only as examples and are not intended to limit the scope of the invention. The novel devices, methods, and programs described in this specification can be embodied in various other forms. Furthermore, various omissions, substitutions, and modifications can be made to the forms of the devices, methods, and programs described in this specification without departing from the spirit of the invention. [Explanation of symbols]

[0077] 1: prediction system, 10: learning device, 20: prediction device, 100: first acquisition unit, 102: learning unit, 104: first storage unit, 106: operation unit, 114: analysis unit, 200: second acquisition unit, 202: prediction unit, 204: second storage unit, 206: display control unit, 208: display unit

Claims

1. a model learning unit that uses explanatory variables as input and learns the prediction model so as to reduce a loss value of a loss function that includes an error between a first objective variable that is a teacher signal and a second objective variable that is an output of the prediction model; a loss function generation unit that changes a contribution of the loss function to a loss value based on at least one of an error between the first objective variable and the second objective variable and a value of the first objective variable; A learning device comprising:

2. The learning device according to claim 1, wherein the explanatory variables are weather forecast information including wind speed and wind direction at a time later than the present at a mesh point of weather analysis, and the first objective variable is wind speed at a time later than the present at a prediction target point.

3. 2. The learning device according to claim 1, wherein the explanatory variables are actual weather information including wind speed and wind direction at a time prior to the present at the prediction target location, and the first dependent variable is wind speed at a time after the present at the prediction target location.

4. The learning device according to claim 2 , wherein the loss function generation unit increases the contribution of the loss function to the loss value when the error is negative compared to when the error is positive.

5. The learning device according to claim 2 , wherein the loss function generation unit reduces a contribution of the loss function to a loss value when the error is negative compared to when the error is positive.

6. The learning device according to claim 2 , wherein the loss function generation unit changes a contribution of the loss function to a loss value in accordance with the magnitude of the wind speed.

7. The learning device according to claim 1 , wherein the loss function generation unit changes a magnitude of a coefficient by which a square or an absolute value of the error is multiplied when changing the contribution of the loss function to the loss value.

8. The learning device according to claim 7 , wherein the loss function generation unit increases the value of the coefficient by a positive value when increasing the contribution of the loss function to the loss value.

9. The learning device according to claim 8 , wherein the loss function generation unit changes the magnitude of a constant to be added to the square of the error or an absolute value when changing the contribution of the loss function to the loss value.

10. The learning device according to claim 9 , wherein the loss function generation unit increases the constant by a positive value when increasing a contribution to the loss function.

11. The learning device according to claim 10 , wherein the magnitude of the coefficient or the constant is changed based on an evaluation of an error distribution of the prediction model.

12. a prediction unit that uses the prediction model created by the learning device according to claim 2; an acquisition unit that acquires weather forecast information including wind speed and wind direction at a time after the present at a mesh point of weather analysis; a prediction unit that predicts a wind speed at the prediction target point at a time after the present time using the weather forecast information acquired by the acquisition unit; A prediction device comprising:

13. a prediction unit that uses the prediction model created by the learning device according to claim 3; an acquisition unit that acquires weather history information including wind speed and wind direction at a time prior to the present at a prediction target location; a prediction unit that predicts a wind speed at the prediction target point at a time after the present time using the weather record information acquired by the acquisition unit; A prediction device comprising:

14. The prediction device according to claim 12 or 13, further comprising a display control unit that causes a display unit to display result data based on the prediction result of the prediction unit.

15. The learning device according to claim 2; The prediction device according to claim 12; A prediction system comprising:

16. The learning device according to claim 3; The prediction device according to claim 13; A prediction system comprising:

Images