Meteorological observation device, meteorological observation method and program

Abstract

To improve measurement accuracy on a cloud mass to be performed using a radar.SOLUTION: A meteorological observation device 1 has: a model reliability identification unit 136 that identifies model reliability serving as model reliability of a learned model on the basis of state data acquired by inputting explanation variable data into the learned model; a data reliability identification unit 137 that identifies data reliability indicative of reliability of the state data corresponding to each of a plurality of spatial locations on the basis of at least the model reliability; and an output unit 135 that outputs estimation meteorological data indicative of a state of a cloud mass on the basis of the state data. A reflection wave data acquisition unit 131 is configured to acquire a plurality of pieces of reflection wave data corresponding to low reliability locations over the number of times exceeding the number of times of acquiring reflection wave data from a prescribed range including locations other than the low reliability location, from a prescribed range including one or more low reliability locations relatively low in data reliability of a plurality of locations.SELECTED DRAWING: Figure 4

Description

[Technical Field]

[0001] The present invention relates to a weather observation device, a weather observation method, and a program. [Background technology]

[0002] Conventionally, there is known a technology for observing rain clouds using radar waves. Patent Document 1 discloses a system that uses a trained model created by machine learning based on past observation data to estimate the location of rainfall based on information acquired from radar waves. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Publication No. 2022-139033 Summary of the Invention [Problem to be solved by the invention]

[0004] When observing rain clouds using a radar device, the amount of cloud water or water vapor contained in the cloud body is estimated based on the measurement results of the reflected waves generated when radar radio waves are reflected by the cloud body. Depending on the state of the cloud body or water vapor, there are problems such as a decrease in the accuracy of estimation of the amount of cloud water or water vapor based on the reflected waves due to large attenuation of the reflected waves, fluctuations in the received power of the reflected waves due to changes in air currents or external disturbances, and superposition of noise.

[0005] The present invention has been made in consideration of these points, and aims to improve the accuracy of measurements relating to cloud bodies made using radar. [Means for solving the problem]

[0006] A meteorological observation device according to a first aspect of the present invention includes a reflected wave data acquisition unit that causes a radar device to emit radar waves and acquires reflected wave data based on the reflected waves of the radar waves reflected by cloud bodies in the air, a meteorological data acquisition unit that acquires meteorological data indicating the atmospheric pressure and temperature at the positions of the cloud bodies from an external device, a reflection factor identification unit that identifies the radar reflection factor observed for each position of the cloud bodies based on the reflected wave data, and a state data acquisition unit that acquires the state data by inputting explanatory variable data including the meteorological data and the radar reflection factor corresponding to the time when the reflected wave data was acquired into a trained model that indicates the state of the cloud bodies and outputs a probability distribution of state data that is objective variable data. the reflected wave data acquisition unit acquires the reflected wave data corresponding to the low-reliability positions from within a predetermined range including one or more low-reliability positions where the data reliability is relatively low among the plurality of positions, a number of times greater than or equal to the number of times the reflected wave data is acquired from within a predetermined range including positions other than the low-reliability positions.

[0007] The model reliability specifying unit may specify the model reliability based on a spread of a probability distribution of the state data.

[0008] The data reliability determination unit may determine the data reliability based on the model reliability and a data error function that takes as arguments at least one of a received signal strength indicator (RSSI) and a signal-to-noise ratio (SNR) of the reflected wave data.

[0009] The system may further include a memory unit that stores the trained model, which is a Gaussian process regression model that uses multiple training data sets consisting of explanatory variable data including training weather data and training radar reflectivity factors, and status data indicating the status of training cloud bodies to learn model parameters that determine a probability density function corresponding to the probability distribution of the status data.

[0010] The state data acquisition unit may input, into the trained model as explanatory variables, the weather data indicating the air pressure and temperature at the position of the cloud body corresponding to the time point when the reflected wave data acquisition unit acquired the reflected wave data, among the multiple weather data corresponding to the multiple time points acquired by the weather data acquisition unit, together with the radar reflectivity factor.

[0011] The reflected wave data acquisition unit may further emit radar radio waves with a power greater than the power of the radar radio waves emitted by the radar device before the data reliability identification unit identifies the data reliability, and acquire the reflected wave data from within a predetermined range that includes the one or more low reliability positions.

[0012] When the data reliability is less than a threshold value, the reflected wave data acquisition unit may acquire the multiple reflected wave data from positions within the specified range that includes a position corresponding to the data reliability and from which the reflected wave data has not been acquired.

[0013] The output unit may output an average value of the state data determined from a probability distribution of the state data as the estimated weather data indicating the state of the cloud body.

[0014] The output unit may output the average of the individual average values ​​determined from the probability distribution of the individual state data corresponding to the individual reflected wave data corresponding to a plurality of positions within a specified range as the estimated weather data corresponding to the positions within the specified range.

[0015] The output unit may output the estimated weather data indicating the state of the cloud body based on an average value obtained by weighting the average value of the individual state data by at least one of the reception strength and the signal-to-noise ratio of the individual reflected wave data.

[0016] A meteorological observation method according to a second aspect of the present invention is a computer-executed method, comprising the steps of: causing a radar device to emit radar radio waves; and acquiring reflected wave data based on the reflected waves of the radar radio waves reflected by cloud bodies in the air; acquiring meteorological data indicating the atmospheric pressure and temperature at the positions of the cloud bodies from an external device; identifying radar reflectivity factors observed for each position of the cloud bodies based on the reflected wave data; and inputting explanatory variable data including the meteorological data and the radar reflectivity factors corresponding to the time when the reflected wave data was acquired into a trained model that indicates the state of the cloud bodies and outputs a probability distribution of state data, which is objective variable data, thereby outputting the state data. a step of acquiring data representing the state of the cloud body; a step of specifying, based on the state data, a model reliability which is the reliability of the trained model when the explanatory variable data is input, based on the state data; a step of specifying, based on at least the model reliability, a data reliability which indicates the reliability of the state data corresponding to each of a plurality of spatial positions; a step of outputting estimated weather data which indicates the state of the cloud body based on the state data; and a step of acquiring, from within a predetermined range which includes one or more low-reliability positions among a plurality of positions where the data reliability is relatively low, a plurality of pieces of reflected wave data corresponding to the low-reliability positions a number of times greater than or equal to the number of times the reflected wave data is acquired from within a predetermined range which includes positions other than the low-reliability positions.

[0017] A third aspect of the present invention provides a program for causing a computer to perform the following steps: cause a radar device to emit radar waves, and acquire reflected wave data based on the reflected waves of the radar waves reflected by a cloud body in the air; acquire meteorological data indicating the atmospheric pressure and temperature at the position of the cloud body from an external device; identify a radar reflectivity factor observed for each position of the cloud body based on the reflected wave data; and acquire state data by inputting explanatory variable data including the meteorological data and the radar reflectivity factor corresponding to the time when the reflected wave data was acquired into a trained model that indicates the state of the cloud body and outputs a probability distribution of state data that is objective variable data. a step of specifying, based on the state data, a model reliability which is the reliability of the trained model when the explanatory variable data is input, a step of specifying, based on at least the model reliability, a data reliability which indicates the reliability of the state data corresponding to each of a plurality of spatial positions, based on at least the model reliability, a step of outputting estimated weather data which indicates the state of the cloud body based on the state data, and a step of acquiring, from within a predetermined range which includes one or more low-reliability positions among a plurality of positions where the data reliability is relatively low, a plurality of reflected wave data corresponding to the low-reliability positions a number of times greater than or equal to the number of times the reflected wave data is acquired from within a predetermined range which includes positions other than the low-reliability positions. [Effects of the Invention]

[0018] According to the present invention, it is possible to improve the accuracy of measurements relating to cloud bodies performed using radar. [Brief explanation of the drawings]

[0019] [Figure 1] FIG. 1 is a diagram illustrating an overview of a weather observation system. [Figure 2] FIG. 10 is a diagram illustrating state data. [Figure 3] FIG. 10 is a diagram illustrating state data. [Figure 4]FIG. 1 is a diagram illustrating a configuration of a meteorological observation device. [Figure 5] 10 is a flowchart illustrating an example of a learning process executed by the learning device. [Figure 6] 10 is a flowchart illustrating an example of an estimation process executed by a meteorological observation device. DETAILED DESCRIPTION OF THE INVENTION

[0020] [Outline of the Weather Observation System S] Figure 1 is a diagram for explaining the overview of a meteorological observation system S. The meteorological observation system S is a system for observing the state of clouds, water vapor, etc. (hereinafter referred to as "cloud bodies") using radar radio waves. A meteorological radar device for observing clouds measures the presence and distribution of clouds by emitting radio waves of a certain frequency from the radar device and observing the radar reflectivity factor of the waves reflected by the clouds received by the radar device. However, a precipitation radar device using X-band or the like cannot capture the waves of radar waves reflected by clouds or water vapor that are composed of water droplets with a diameter of less than 0.1 mm, and therefore cannot directly observe the state of cloud bodies.

[0021] To overcome these drawbacks of conventional radar devices, one method is to passively observe the state of water vapor by observing the radiation intensity of water vapor in the microwave frequency range using, for example, a microwave radiometer. In this case, by receiving radiation intensities at multiple frequencies between 20 and 30 GHz or 50 and 60 GHz at time intervals of several seconds to several minutes, the state of water vapor in the atmosphere can be observed based on the radiation intensities observed at these frequencies. However, when using a microwave radiometer, one measurement takes several minutes, and if rain falls on the sensor, measurements cannot be made until the sensor dries.

[0022] In light of these challenges, attempts have been made to emit radar waves at multiple frequencies, including the terahertz frequency band, from a radar device and observe the state of water vapor based on the difference in the attenuation of the radar waves at each of the multiple frequencies due to water vapor. Radar devices using radar waves at multiple frequencies like this can observe cloud bodies in three dimensions, covering a 360-degree area above and horizontally from the radar device, by rotating the transmitting and receiving antenna to change the horizontal direction of the transmitted and received waves and by changing the elevation angle. The radar observation control parameters, such as the output power of the radar device's transmitted waves, the azimuth angle increments of the received waves, and the elevation angle increments, are always kept constant and do not change during observation.

[0023] However, depending on the state of cloud bodies, radar waves may be attenuated significantly as they pass through clouds or the atmosphere, or the strength of radar waves may fluctuate over time due to air currents, making it difficult to properly receive the reflected waves generated by radar waves reflected by cloud bodies. As a result, there is a problem with insufficient accuracy in estimating the state of cloud bodies. Therefore, the weather observation system S utilizes a trained model created through machine learning to estimate the state of cloud bodies, and if there are locations where the reliability of the estimation results is insufficient, the estimation accuracy is improved by increasing the number of observations.

[0024] The meteorological observation system S includes a meteorological observation device 1, a radar device 2, and a learning device 3. Some or all of the components of the learning device 3 may be included in the meteorological observation device 1. Below, with reference to FIG. 1, an overview of the process flow for creating a trained model and the process flow for observing the state of cloud bodies using the trained model will be described.

[0025] [Outline of estimation process by meteorological observation device 1] The meteorological observation device 1 is a computer that executes an estimation process to estimate the state of cloud bodies based on the reflected waves of radar radio waves and outputs estimated meteorological data based on the estimation results. The meteorological observation device 1 has an estimation unit 11 and a radar control unit 12.

[0026] The estimation unit 11 estimates the state of the cloud body based on the objective variable data y0 indicating the state of the cloud body acquired from the trained model by inputting explanatory variable data x0 based on observation data created by the radar device 2 based on reflected wave data derived from reflected waves and meteorological data near the cloud body at the time the reflected wave data was acquired into the trained model. The objective variable data y0 is used as state data indicating at least one of the cloud water content Cv and the water vapor content Vv. Details of the explanatory variable data x0 and the objective variable data y0 will be described later.

[0027] The radar control unit 12 controls the radar device 2 to cause the radar device 2 to transmit radar radio waves. The radar control unit 12 controls the radar device 2 based on the data reliability corresponding to the reliability of the received reflected wave data (i.e., the magnitude of the data error) and the reliability of the trained model under the observation conditions corresponding to the observation data. If the data reliability is below a threshold, the radar control unit 12 causes the radar device 2 to emit radar radio waves again so that the objective variable data y0 for the observation conditions under which the data reliability is below the threshold can be acquired again.

[0028] When it is estimated that the data reliability is low, the radar control unit 12 controls the radar device 2 to increase the intensity of the radar radio waves, refine the azimuth angle increments for acquiring the observation data, refine the elevation angle increments, etc., in order to further acquire highly reliable observation data in areas where the data reliability is low.

[0029] Note that the information estimated for the objective variable data y0 has a probability distribution, and the reliability of the trained model corresponds to, for example, the magnitude of the spread (e.g., variance) of the probability distribution. In other words, the greater the spread of the probability distribution of the objective variable data y0 output by the trained model, the lower the reliability of the trained model under the observation conditions corresponding to the objective variable data y0.

[0030] [Overview of Radar Device 2] The radar device 2 has a radar transmitter 21 and a radar receiver 22. The radar transmitter 21 emits radar radio waves under the control of the radar control unit 12. The radar receiver 22 receives the radar radio waves reflected by a cloud body, and outputs observation data based on the reflected wave data.

[0031] The radar device 2 is designed to be able to detect cloud droplets that are much smaller than raindrops, and emits radar radio waves in the Ka band (e.g., 35 GHz, wavelength of approximately 8.5 mm), W band (e.g., 94 GHz, wavelength of approximately 3.2 mm), or D band (e.g., 148 GHz, wavelength of approximately 2.0 mm). As described above, two frequency bands from the Ka band, W band, and D band may be used to observe water vapor based on the difference in attenuation depending on the frequency.

[0032] The radar device 2 is, for example, a frequency-modulated continuous-wave (FMCW) radar device having a pulse compression means using linear frequency modulation (LFM) or chirp modulation. The radar device 2 outputs, as observation data, the linear distance (slant range) r to the cloud body, time t, radar reflectivity factor Z(r,t,θ,φ), RSSI(r,t,θ,φ) indicating the received strength of the reflected wave, SNR(r,t,θ,φ) indicating the signal-to-noise ratio of the reflected wave, and the azimuth angle θ and elevation angle φ that identify the radar reflectivity factor.

[0033] Here, RSSI(r,t,θ,φ) is the reception level directly observed by the radar device 2, and SNR(r,t,θ,φ) indicates the reception likelihood including the process gain due to pulse compression in post-demodulation processing. The radar device 2 emits radar waves based on modulated signals having multiple pulse widths (time required to sweep a predetermined frequency width). The radar device 2 can change the intensity of the emitted radar waves by switching the process gain.

[0034] [Overview of learning process by learning device 3] The learning device 3 executes a learning process for creating a trained model. When explanatory variable data x0 is input from the estimation unit 11, the learning device 3 outputs response variable data y0.

[0035] The learning device 3 includes a learning data generation unit 31, a learning data storage unit 32, a model creation unit 33, and a trained model storage unit 34. The learning data generation unit 31 generates learning data including explanatory variable data and target variable data obtained based on a meteorological model, a radio wave attenuation model, and cloud body parameters, etc. The explanatory variables are, for example, the distance r from the radar device 2 to the cloud body C, the azimuth angle θ of the cloud body relative to the radar device 2, the elevation angle φ, the altitude of the cloud body, the air pressure P at the position of the cloud body, the temperature T at the position of the cloud body, and the radar reflectivity factor Z of the radar reflected wave.

[0036] The learning data generation unit 31 identifies the atmospheric pressure P and the temperature T based on data acquired from another device that can measure atmospheric pressure, temperature, etc. in synchronization with the radar device 2, for example. The learning data generation unit 31 identifies the distance from the radar device 2 to the cloud body based on measurement results from a radiosonde, an aircraft, a microwave radiometer, etc. The learning data generation unit 31 may calculate the atmospheric pressure P and the temperature T at the position of the cloud body based on the statistical function of ITU-R P.835, using the measurement results of the temperature, atmospheric pressure, and humidity on the ground.

[0037] The dependent variable data is, for example, at least one of cloud water content Cv and water vapor content Vv. The training data generation unit 31 generates a training dataset including these explanatory variable data and dependent variable data. When generating the training dataset, the training data generation unit 31 may use explanatory variable data and dependent variable data based on actual measurement results, or may use explanatory variable data and dependent variable data calculated based on a theoretical formula.

[0038] The learning data generation unit 31 generates a learning data set X trainIn order to generate the above-mentioned data, a meteorological model defined by ITU-R called Mean Annual Global Reference Atmosphere (MAGRA) can be used, for example, to determine the temperature, pressure, water vapor density, etc. in the sky from the temperature, pressure, and water vapor density at the ground level. According to this meteorological model, the temperature, pressure, and water vapor density in the sky above the radar device 2 can be obtained by selecting the latitude model of the location where the radar device 2 is installed and its surroundings from high latitude, mid-latitude, low latitude, or standard region, setting the water vapor density at the surrounding ground level, and selecting the season from summer or winter. Note that the meteorological model is not limited to MAGRA, and other meteorological models can also be used.

[0039] Furthermore, the training data generation unit 31 can calculate a radar reflectivity factor by applying the central altitude of the cloud body, the thickness of the cloud body, and the peak water vapor density of the cloud body to a known theoretical formula. The training data generation unit 31 can also calculate the attenuation of radar radio waves as they pass through the cloud body and the atmosphere using the known theoretical formula, and correct the calculated radar reflectivity factor based on the attenuation, thereby calculating a training radar reflectivity factor corresponding to the observed reflected waves. The training data generation unit 31 can generate a large amount of training data sets each consisting of a set of distance r, altitude h, air pressure P, temperature T, water vapor content Vv, cloud water content Cv, and radar reflectivity factor Z by varying various meteorological conditions and cloud body conditions.

[0040] The learning data generation unit 31 stores the generated learning data in the learning data storage unit 32. The learning data storage unit 32 is a storage medium that stores the learning data set generated by the learning data generation unit 31.

[0041] The model creation unit 33 performs machine learning on the machine learning model by inputting the training datasets stored in the training data storage unit 32 into the machine learning model. The model creation unit 33 creates a trained model by performing machine learning using all training datasets, and stores the created trained model in the trained model storage unit 34. The model creation unit 33 may perform machine learning and update the trained model every time a new training dataset is stored in the training data storage unit 32.

[0042] The model creation unit 33 generates learning data X, which is a set of explanatory variable data x0 and objective variable data y0. train (x train (i),y train (i) / i=1,2,...,N train )(N train is the number of training datasets), we use the probability density function p(y0|Θ * ,x0) to create a trained model.

[0043] The target of machine learning performed by the model creation unit 33 is the learning model parameters (parameters) Θ that determine the probability density function, and the optimal learned model parameters are calculated by learning. For example, if the learned model is a Gaussian process regression model, the probability density function p(y0|Θ) corresponding to the objective variable data y0 output by the learned model is * , x0) is expressed by a normal distribution (Gaussian distribution). The objective variable data y0 is expressed by, for example, a probability density function p(y0|Θ * , x0), and the parameters of the kernel function for expressing the variance-covariance matrix. train may also include: X train ={x train (i),y train (i)} x train (i)=(r(i),θ(i),φ(i),P(i),T(i),Z(i)) ytrain (i)=(Cv(i),Vv(i)) where i=1,2,...,N train is.

[0044] In the Gaussian process regression model, the probability density function p(y0|Θ * , x0) is estimated by the probability density function of the normal distribution. The probability distribution of the objective variable data y0 is X train The model parameter Θ depends on * and the explanatory variable data x0 corresponding to the observed data. The model parameters Θ * is the training data X train ={x train (i),y train (i) / i=1,2,...,N train} explanatory variable data x train (mean value μ*) of (i), variance-covariance matrix K * (N train Row N train column symmetric matrix), kernel function parameter λ * , and X train That is, Θ * ={μ * ,K * ,λ * ,X train}.

[0045] Here, the variance-covariance matrix K * The (i,j) component is X train Explanatory variable data x train (i) (i=1,2,...,N train ) is calculated using the kernel function k(x train (i),x train (j)), that is, K * (i,j)=k(x train (i),x train (j)). The kernel function used is, for example, the Gaussian kernel k(x train (i),x train (j))=exp(-c×||x train (i)-x train (j)|| 2), but other functions are also possible. Note that c is the kernel function parameter λ * The kernel function parameter λ is equivalent to * In the case of a Gaussian kernel, is calculated by the maximum likelihood estimation method, the Markov chain Monte Carlo method, or the like.

[0046] The trained model storage unit 34 is a non-volatile storage medium such as an SSD (Solid State Drive). When explanatory variable data including weather data such as atmospheric pressure P and temperature T corresponding to the time when the reflected wave of the radar radio wave was acquired and radar reflectivity factors are input, the trained model stored in the trained model storage unit 34 outputs, as state data, objective variable data y0 that indicates the state of the cloud body and is objective variable data that has a probability distribution. As described above, the objective variable data y0 has a probability distribution, and the trained model may output data indicating the mean value and variance of the probability distribution as state data.

[0047] [Details of estimation process by meteorological observation device 1] The following describes details of the estimation process performed by the meteorological observation device 1. The estimation unit 11 calculates the distribution of the objective variable data y0, which is the cloud water content and water vapor density, from the explanatory variable data x0, which includes newly observed radar reflectivity factors, etc., using a probability density function p(y0|Θ * , x0). The estimation unit 11 obtains state data y0 corresponding to the multiple positions from the trained model by sequentially inputting multiple pieces of explanatory variable data x0 based on observation data corresponding to multiple positions in three-dimensional space and weather data at the positions of cloud bodies corresponding to the observation data into the trained model. The estimation unit 11 outputs estimated weather data based on the state data y0 corresponding to the multiple positions. The estimation unit 11 outputs estimated weather data indicating, for example, the average value of the state data y0 output by the trained model.

[0048] 2 and 3 are diagrams for explaining the state data y0. For easy understanding, in Fig. 2 and Fig. 3, the explanatory variable data x0 and the response variable data y0 are assumed to be one-dimensional normal distributions. In Fig. 2, the position of x is xr1 When x is at x, r2 The two corresponding probability density functions (curves) P1 and P2 are shown. The mean of the probability distribution corresponding to the probability density function P1 is μ_x r1 and the mean value of the probability distribution corresponding to the probability density function P2 is μ_x r2 is.

[0049] In Figure 3, curve M indicates the average value of the probability density function corresponding to each explanatory variable data x0, and curve D indicates the standard deviation added to curve M. L ,D, which is the curve M minus the standard deviation S As an example, the standard deviation corresponding to probability density function P1 is smaller than the standard deviation corresponding to probability density function P2, so r2 The variance of the response variable data y0 corresponding to the position x r1 is larger than the variance of the response variable data y0 corresponding to the position x r2 The reliability of the target variable data y0 corresponding to the position x r1 The estimation unit 11 estimates the reliability of the response variable data y0 by using, for example, the probability density function p(y0|Θ * ,x0) is used as the standard deviation (square root of the variance) of the covariance matrix.

[0050] Here, the estimation unit 11 defines a learning model reliability function for calculating the reliability C(x) of the learned model for each x as C(x)=c Θ Here, the multidimensional normal distribution p(x) of a random variable with mean μ and variance-covariance matrix K is expressed as p(x)=N(x|μ,K). In this case, the estimation unit 11 calculates C(x)=c Θ* (x)=1 / k x The reliability C(x) is calculated by k sim =(k(x train (1),x),k(x train (2),x),···, k(x train (N train ),x) T (Size N train × 1), the estimation unit 11 calculates kx =k(x,x)-k sim T (K * ) -1 k sim By k x Calculate y train =(y train (1),y train (2),···,y train (N train )) T (Size N train ×1), the estimation unit 11 calculates μ x =k sim T ((K * ) -1 y train By μ x Calculate.

[0051] Furthermore, depending on the condition of the cloud body, there may be cases where radar radio waves are significantly attenuated while passing through clouds or the atmosphere, or where reflected waves cannot be received properly due to time fluctuations of the reflected waves caused by air currents. Therefore, the estimation unit 11 uses the data error function E(x,q)=ε(x,q) that calculates the error E(x,q) of the received data using quality information q indicating the quality of the observation data = (r, θ, φ, P, T, Z) as the reliability of the observation data. In other words, the data error function ε(x,q) is a function that takes the quality information q as an argument. The quality information q is, for example, the values ​​of RSSI and SNR.

[0052] The estimation unit 11 calculates data reliability based on the reliability of the trained model (i.e., the reliability of the objective variable data y0) and the reliability of the observed data. The estimation unit 11 adaptively and preferentially acquires additional observed data in the direction (θ, φ) where the data reliability is low, thereby improving the quality of the estimation result.

[0053] [Configuration of weather observation device 1] Next, the configuration of the meteorological observation device 1 will be described in detail. Fig. 4 is a diagram showing the configuration of the meteorological observation device 1. The meteorological observation device 1 has a communication unit 110, a storage unit 120, and a control unit 130. The control unit 130 has a reflected wave data acquisition unit 131, a reflection factor identification unit 132, a meteorological data acquisition unit 133, a state data acquisition unit 134, an output unit 135, a model reliability identification unit 136, and a data reliability identification unit 137.

[0054] The communication unit 110 includes a communication interface for transmitting and receiving data to and from the radar device 2, the learning device 3, or another device. The communication unit 110 transmits, for example, data for controlling the radar device 2, which is input from the reflected wave data acquisition unit 131, to the radar device 2. The communication unit 110 inputs the observation data received from the radar device 2 to the reflected wave data acquisition unit 131. The communication unit 110 also receives meteorological data indicating atmospheric pressure and temperature from an external device, and inputs the received meteorological data to the meteorological data acquisition unit 133.

[0055] Furthermore, the communication unit 110 transmits explanatory variable data x0 based on the observation data input from the state data acquisition unit 134 to the learning device 3, and inputs the objective variable data y0 received from the learning device 3 to the state data acquisition unit 134. The communication unit 110 transmits the estimated weather data output by the output unit 135 to an external device (e.g., a computer).

[0056] The storage unit 120 has storage media such as a ROM (Read Only Memory), a RAM (Random Access Memory), and an SSD (Solid State Drive). The storage unit 120 stores programs executed by the control unit 130. The storage unit 120 stores various data used by the control unit 130 to create estimated weather data.

[0057] The storage unit 120 may store a trained model created by the learning device 3. That is, the storage unit 120 may store a trained model that is a Gaussian process regression model obtained by machine learning model parameters that determine a probability density function corresponding to the probability distribution of state data using a plurality of training data sets that are configured of explanatory variable data including training weather data and training radar reflectivity factors, and state data that is objective variable data indicating the state of training cloud bodies.

[0058] The control unit 130 has, for example, a CPU (Central Processing Unit). The control unit 130 executes the programs stored in the storage unit 120 to function as a reflected wave data acquisition unit 131, a reflection factor identification unit 132, a meteorological data acquisition unit 133, a state data acquisition unit 134, an output unit 135, a model reliability identification unit 136, and a data reliability identification unit 137.

[0059] The reflected wave data acquisition unit 131 also functions as the radar control unit 12 shown in Fig. 1, causes the radar device 2 to emit radar radio waves, and acquires reflected wave data based on the radar radio waves reflected by cloud bodies in the air. The reflected wave data acquisition unit 131 inputs the acquired reflected wave data to the reflection factor identification unit 132. The reflected wave data acquisition unit 131 acquires multiple pieces of reflected wave data corresponding to low-reliability positions from within a predetermined range that includes one or more low-reliability positions with relatively low data reliability among the multiple positions, a number of times greater than or equal to the number of times that reflected wave data is acquired from within a predetermined range that includes positions other than the low-reliability positions.

[0060] That is, the reflected wave data acquisition unit 131 acquires more reflected wave data around a location where the data reliability is relatively low than around a location where the data reliability is relatively high. This improves the accuracy of estimated weather data at a location where the data reliability is determined to be low in the first observation. The predetermined range is a range within which there is no substantial difference in the state of cloud bodies, for example, a range within 50 m. The predetermined range may be set by the administrator of the weather observation device 1.

[0061] After the data reliability determination unit 137 determines the data reliability, the reflected wave data acquisition unit 131 may further emit radar radio waves (for example, radar radio waves with a large pulse width) with a power greater than the power of the radar radio waves emitted by the radar device 2 before the data reliability determination unit 137 determined the data reliability, and acquire reflected wave data from within a predetermined range that includes one or more low reliability positions.

[0062] When the data reliability is less than the threshold, the reflected wave data acquisition unit 131 may acquire multiple pieces of reflected wave data from positions within a predetermined range that includes the position corresponding to that data reliability and from which reflected wave data has not been acquired. This predetermined range is a range that is smaller than the interval between positions from which reflected wave data was previously acquired. In other words, when the data reliability is less than the threshold, the reflected wave data acquisition unit 131 acquires multiple pieces of reflected wave data at a smaller granularity (position interval) than when the reflected wave data was previously acquired. This allows the status data acquisition unit 134 to acquire a large amount of status data around positions with low data reliability, thereby improving the accuracy of the estimated weather data.

[0063] The reflection factor identifying unit 132 identifies the radar reflection factor observed for each cloud body position based on the reflected wave data. For example, the reflection factor identifying unit 132 identifies the radar reflection factor Z(r, t, θ, φ) corresponding to each position based on the reflected wave data corresponding to multiple positions with different combinations of the linear distance r to the cloud body, the azimuth angle θ, and the elevation angle φ. The reflection factor identifying unit 132 inputs the identified radar reflection factor to the state data acquiring unit 134.

[0064] The weather data acquisition unit 133 acquires weather data indicating the atmospheric pressure and temperature at the location of a cloud body from an external device. For example, the weather data acquisition unit 133 acquires weather data associated with time and location from an external server that stores weather data. The weather data acquisition unit 133 notifies the status data acquisition unit 134 of the acquired weather data.

[0065] When the state data acquisition unit 134 receives explanatory variable data x0 including meteorological data corresponding to the time point at which the reflected wave data was acquired and a radar reflectivity factor, the state data acquisition unit 134 acquires the dependent variable data y0 as state data by inputting the explanatory variable data x0 to a trained model that indicates the state of the cloud body and outputs a probability distribution of the state data which is dependent variable data y0. The state data acquisition unit 134 inputs, among the multiple meteorological data corresponding to the multiple time points acquired by the meteorological data acquisition unit 133, meteorological data indicating the air pressure and temperature at the position of the cloud body corresponding to the time point at which the reflected wave data acquisition unit 131 acquired the reflected wave data, together with the radar reflectivity factor, to the trained model as explanatory variable data x0.

[0066] The state data acquisition unit 134 inputs explanatory variable data x0, including the distance r from the radar device 2 to the cloud body C, the azimuth angle θ of the cloud body relative to the radar device 2, the elevation angle φ, the altitude of the cloud body, the air pressure P at the position of the cloud body, the temperature T at the position of the cloud body, and the radar reflection factor Z of the radar reflected wave, to the trained model storage unit 34 shown in FIG. 1 via, for example, the communication unit 110. The state data acquisition unit 134 acquires the objective variable data y0 output from the trained model storage unit 34 as state data. The state data acquisition unit 134 sequentially inputs multiple pieces of explanatory variable data x0 corresponding to multiple pieces of observation data corresponding to multiple positions to the state data acquisition unit 134, and acquires state data corresponding to the multiple positions.

[0067] The output unit 135 outputs estimated weather data indicating the state of the cloud body based on the state data. The output unit 135 outputs, for example, the average value of the state data determined from the probability distribution of the objective variable data y0 as the estimated weather data indicating the state of the cloud body. The output unit 135 may also output, as the estimated weather data corresponding to the positions within the predetermined range, the average value of the individual average values ​​determined from the probability distribution of the objective variable data y0, which is individual state data corresponding to individual reflected wave data corresponding to multiple positions within the predetermined range. The predetermined range is, for example, a range determined as a range in which there is no substantial difference in the state of the cloud body.

[0068] The individual mean values ​​are the mean values ​​μ of the probability distribution, e.g., at position x in Figure 2.r1 The mean value μ_x corresponding to r1 and position x r2 The mean value μ_x corresponding to r2 The average of the individual averages is the average μ_x r1 and the average value μ_x r2 This is the average value of

[0069] In this case, the output unit 135 may output estimated weather data indicating the state of the cloud body based on the average value calculated by weighting the average value of each state data by at least one of the RSSI and the SN ratio of each reflected wave data. For example, at the position x r1 The RSSI of the reflected wave data at position x r2 If the RSSI of the reflected wave data is better than the RSSI of the reflected wave data at r1 The weight of the average value μ_x r2 The weight of the calculated value is set to be larger than the weight of the calculated value, and the weighted average is output as the estimated weather data.

[0070] The model reliability identification unit 136 identifies model reliability C(x), which is the reliability of the trained model when explanatory variable data x0 is input, based on the state data. The model reliability identification unit 136 identifies model reliability C(x), for example, based on the magnitude of variance of the state data. The model reliability identification unit 136 lowers the model reliability C(x) as the variance increases. The model reliability identification unit 136 notifies the data reliability identification unit 137 of the identified model reliability C(x).

[0071] The data reliability determination unit 137 determines data reliability indicating the reliability of state data corresponding to each of a plurality of spatial positions based on at least the model reliability C(x). Specifically, the data reliability determination unit 137 determines data reliability G(x) based on the model reliability C(x) and a data error function E(x) that uses at least one of the received signal strength (RSSI) and the signal-to-noise ratio (SNR) of the reflected wave data as an argument.

[0072] The data reliability determination unit 137 may, for example, calculate E(x)=γ(xr)×((RSSI(r,t,θ,φ)+o RSSI ) / r RSSI +(SNR(r,t,θ,φ)+o SNR ) / r SNR ) to calculate the data error function E(x). Here, γ(xr) is a proportionality constant determined by the reflected wave data xr, and r RSSI and r SNR is the range width for normalization used to average the influence of the RSSI value and the SNR value (for example, 20 (dB) for the SNR value), RSSI and SNR is also an offset value of the range of RSSI and SNR values ​​for normalization.

[0073] Next, the data reliability specification unit 137 calculates the data reliability G(x) based on the model reliability C(x) and the data error function E(x). The data reliability specification unit 137 calculates the data reliability G(x) based on the N quantized values ​​of {azimuth angle θ, elevation angle φ} using a predetermined grid width. d A set of observation directions D:={d(1),d(2),···,d(N d )}, the data reliability G(x) corresponding to each observation direction is calculated.

[0074] The data reliability specification unit 137 determines the x r The data reliability specification unit 137 calculates the data reliability for (d(i)). For example, the data reliability specification unit 137 calculates the data reliability for the function E(x) by using the range R E Using G(x)=C(x)(1-E(x) / R E ) or G(x)=C(x) / E(x) with x=x r By substituting (d(i)), the data reliability G(x) for the observation data in the d(i) direction is calculated.

[0075] Next, the data reliability specifying unit 137 calculates data reliability map information M(D). Specifically, when data reliability map information for the j-th observation direction d(j)=(θ(j),φ(j)) is defined as M(D)=M(d(j)), the data reliability specifying unit 137 calculates M(d(j))=Σx in Xd(j) G(x) / N d(j) The data reliability map information M(D) is calculated by the following equation. d(j) is the set X consisting of all x in the d(j) direction. d(j) The data reliability specifying unit 137 notifies the reflected wave data acquiring unit 131 of the calculated data reliability map information M(D).

[0076] [Radar Device 2 Control] As described above, the reflected wave data acquisition unit 131, for example, causes the radar device 2 to re-emit stronger radar radio waves in order to improve the accuracy of the observation data for each observation direction indicated by the data reliability map information M(D) where the data reliability corresponding to that observation direction is equal to or less than a threshold value.

[0077] The reflected wave data acquisition unit 131 determines the observation control parameter Φ by Φ = h(M) using the input reliability map information M(D). The reflected wave data acquisition unit 131 determines the following parameter as the observation control parameter Φ for each d(j) in the observation direction set D.

[0078] Observation flag: F obj (j)=1(M(d(j)) <Th obj ),F obj (j)=0(M(d(j))≧Th obj ) The reflected wave data acquisition unit 131 is F obj When (j)=1, radar device 2 emits radar waves and performs re-observation. obj is the threshold for whether or not to perform re-observation.

[0079] Observation time: T obj (j)=α T / M(d(j)) α T is a constant. The lower the data reliability, the longer the reflected wave data acquisition unit 131 makes the radar device 2 emit radar waves and performs re-observation.

[0080] Observation azimuth angle increment: Δθ obj(j)=α θ M(d(j)) α θ is a constant. The reflected wave data acquisition unit 131 increases the azimuth angle interval Δθ as the data reliability decreases. obj Decrease (j) and observe again.

[0081] Observation elevation angle increment: Δφ obj (j)=α φ M(d(j)) α φ is a constant. The lower the data reliability, the larger the elevation angle increment Δφ obj Decrease (j) and observe again.

[0082] Observed pulse width (signal width): B obj (j)=α B / M(d(j)) α B is a constant. The lower the data reliability, the longer the pulse width at which the reflected wave data acquisition unit 131 makes the radar device 2 emit radar radio waves. In other words, the lower the data reliability, the higher the power of the radar radio waves at which the reflected wave data acquisition unit 131 makes re-observation.

[0083] The reflected wave data acquisition unit 131 calculates the observation control parameter Φ j The observation control parameter Φ is transmitted to the radar device 2 as follows: Φ j =h(M(d(j))) =(F obj (j),T obj (j),Δθ obj (j),Δφ obj (j),B obj (j))

[0084] The reflected wave data acquisition unit 131 may transmit the data reliability map information M(D) to the radar device 2, and the radar device 2 may modulate and demodulate the pulse width based on the data reliability map information M(D) and emit radar radio waves.

[0085] In this way, when the reflected wave data acquisition unit 131 acquires reflected wave data corresponding to the same direction (θ, φ) multiple times, the condition data acquisition unit 134 acquires multiple condition data based on the multiple reflected wave data. In this case, the output unit 135 outputs the average value of the average values ​​of the multiple condition data corresponding to each direction as the estimated weather data corresponding to that direction.

[0086] [Learning process flowchart] Fig. 5 is a flowchart showing an example of the learning process executed by the learning device 3. The flowchart shown in Fig. 5 shows the flow of the process when the learning device 3 cooperates with the radar device 2 to acquire part of the learning data.

[0087] The radar device 2 emits radar waves while changing its direction (S11), and the learning data generation unit 31 acquires radar data indicating the radar reflection factor, the emission direction of the radar waves, etc. based on the reflected wave data acquired by the radar receiver 22 (S12). The learning data generation unit 31 also acquires weather data indicating the air pressure and temperature at a position corresponding to the radar reflection factor from an external device that manages weather data (S13). The learning data generation unit 31 generates learning data based on the radar data and weather data (S14). Note that the learning data generation unit 31 may use the radar data and weather data calculated based on a theoretical formula as at least a part of the learning data.

[0088] The model creation unit 33 creates a trained model by performing machine learning using the training data generated in this way as training data (S15). The model creation unit 33 stores the created trained model in the trained model storage unit 34 (S16).

[0089] [Flowchart of estimation process] 6 is a flowchart showing an example of estimation processing executed by the meteorological observation device 1. The reflected wave data acquisition unit 131 acquires reflected wave data by causing the radar device 2 to emit radar radio waves (S21). The reflection factor identification unit 132 identifies a radar reflection factor based on the reflected wave data acquired by the reflected wave data acquisition unit 131 (S22). Furthermore, the meteorological data acquisition unit 133 acquires meteorological data such as atmospheric pressure and temperature (S23).

[0090] The state data acquisition unit 134 acquires state data by inputting explanatory variable data based on radar reflectivity factors, weather data, etc., into a trained model stored in the learning device 3 (S24). Furthermore, the model reliability determination unit 136 determines model reliability based on the variance of the acquired state data (S25). The data reliability determination unit 137 calculates a data error function based on the reception state of the reflected wave data, etc. (S26), and determines data reliability based on the model reliability and the data error function (S27).

[0091] If the data reliability in all directions is equal to or greater than the threshold (YES in S28), the output unit 135 outputs estimated weather data based on the status data acquired in S24 (S29). If there is a direction for which the data reliability is less than the threshold (NO in S28), the reflected wave data acquisition unit 131 changes the transmission conditions of the radar wave, causes the radar device 2 to emit radar wave again, acquires reflected wave data (S21), and repeats the processes from S21 to S27. By following the above procedure, the meteorological observation device 1 can improve the accuracy of the estimated weather data.

[0092] The present invention has been described above using embodiments, but the technical scope of the present invention is not limited to the scope described in the above embodiments, and various modifications and alterations are possible within the scope of the gist thereof. For example, all or part of the device can be configured by functionally or physically distributing or integrating in any unit. Furthermore, new embodiments resulting from any combination of multiple embodiments are also included in the embodiments of the present invention. The effects of the new embodiments resulting from the combination also have the effects of the original embodiments. For example, the machine learning model used may be another model that performs estimation using probability distribution. [Explanation of symbols]

[0093] 1. Weather observation equipment 2. Radar equipment 3 Learning Device 11 Estimation part 12 Radar control unit 21 Radar transmitter 22 Radar receiver 31 Learning data generation unit 32 Learning data storage unit 33 Model Creation Department 34 Trained model memory 110 Communications Department 120 Storage section 130 Control Unit 131 Reflected wave data acquisition unit 132 Reflection factor identification part 133 Weather Data Acquisition Unit 134 Status data acquisition unit 135 Output section 136 Model Reliability Identification Unit 137 Data Reliability Identification Unit

Claims

1. a reflected wave data acquisition unit that causes the radar device to emit radar waves and acquires reflected wave data based on the waves reflected by cloud bodies in the air; a meteorological data acquisition unit that acquires meteorological data indicating atmospheric pressure and temperature at the position of the cloud body from an external device; a reflection factor identifying unit that identifies a radar reflection factor observed for each position of the cloud body based on the reflected wave data; a state data acquisition unit that, when explanatory variable data including the weather data corresponding to the time point when the reflected wave data was acquired and the radar reflectivity factor is input, acquires the state data by inputting the explanatory variable data into a trained model that indicates the state of the cloud body and outputs a probability distribution of state data that is objective variable data; a model reliability determination unit that determines, based on the state data, model reliability, which is the reliability of the trained model when the explanatory variable data is input; a data reliability determination unit that determines a data reliability indicating a reliability of the state data corresponding to each of a plurality of spatial positions based on at least the model reliability; an output unit that outputs estimated weather data indicating the state of the cloud body based on the state data; and the reflected wave data acquisition unit acquires, from within a predetermined range including one or more low-reliability positions among the plurality of positions where the data reliability is relatively low, the plurality of reflected wave data corresponding to the low-reliability positions a number of times equal to or greater than the number of times the reflected wave data is acquired from within the predetermined range including positions other than the low-reliability positions; Weather observation equipment.

2. the model reliability specifying unit specifies the model reliability based on a spread of a probability distribution of the state data. The meteorological observation device according to claim 1.

3. the data reliability determination unit determines the data reliability based on the model reliability and a data error function that uses at least one of a received signal strength indicator (RSSI) and an signal-to-noise ratio (SNR) of the reflected wave data as an argument. The meteorological observation device according to claim 1.

4. The method further comprises a storage unit for storing the trained model, which is a Gaussian process regression model obtained by machine learning model parameters that determine a probability density function corresponding to a probability distribution of state data using a plurality of training data sets that are composed of explanatory variable data including training weather data and training radar reflectivity factors, and state data that indicate the state of training cloud bodies. The meteorological observation device according to claim 1.

5. the state data acquisition unit inputs, into the trained model as explanatory variables, the weather data indicating the atmospheric pressure and temperature at the position of the cloud body corresponding to the time point when the reflected wave data acquisition unit acquired the reflected wave data, among the plurality of weather data corresponding to the plurality of time points acquired by the weather data acquisition unit, together with the radar reflectivity factor; The meteorological observation device according to claim 1.

6. the reflected wave data acquisition unit further emits radar radio waves having a power greater than the power of the radar radio waves emitted by the radar device before the data reliability determination unit identified the data reliability, and acquires the reflected wave data from within a predetermined range including the one or more low-reliability positions. The weather observation device according to any one of claims 1 to 5.

7. When the data reliability is less than a threshold, the reflected wave data acquisition unit acquires the plurality of reflected wave data from positions within the predetermined range including a position corresponding to the data reliability and from which the reflected wave data has not been acquired. The weather observation device according to any one of claims 1 to 5.

8. The output unit outputs an average value of the state data determined from a probability distribution of the state data as the estimated weather data indicating the state of the cloud body. The weather observation device according to any one of claims 1 to 5.

9. the output unit outputs, as the estimated weather data corresponding to the position within the predetermined range, an average value of the individual average values ​​determined from a probability distribution of the individual state data corresponding to the individual reflected wave data corresponding to a plurality of positions within the predetermined range. The meteorological observation device according to claim 8.

10. The output unit outputs the estimated weather data indicating the state of the cloud body based on an average value obtained by weighting the average value of the individual state data by at least one of the reception intensity and the S / N ratio of the individual reflected wave data. The meteorological observation device according to claim 9.

11. The computer executes a step of emitting radar waves from a radar device and acquiring reflected wave data based on the waves reflected by cloud bodies in the air; acquiring meteorological data indicating atmospheric pressure and temperature at the location of the cloud body from an external device; Identifying a radar reflectivity factor observed for each position of the cloud body based on the reflected wave data; a step of acquiring the state data by inputting explanatory variable data including the meteorological data corresponding to the time point when the reflected wave data was acquired and the radar reflectivity factor into a trained model that indicates the state of the cloud body and outputs a probability distribution of state data that is objective variable data; A step of identifying a model reliability, which is a reliability of the trained model when the explanatory variable data is input, based on the state data; determining a data reliability indicating a reliability of the state data corresponding to each of a plurality of spatial locations based at least on the model reliability; outputting estimated weather data indicating the state of the cloud body based on the state data; acquiring, from within a predetermined range including one or more low-reliability positions among the plurality of positions where the data reliability is relatively low, a plurality of pieces of reflected wave data corresponding to the low-reliability positions a number of times greater than or equal to the number of times the reflected wave data is acquired from within a predetermined range including positions other than the low-reliability positions; A meteorological observation method comprising:

12. On the computer, a step of emitting radar waves from a radar device and acquiring reflected wave data based on the waves reflected by cloud bodies in the air; acquiring meteorological data indicating atmospheric pressure and temperature at the location of the cloud body from an external device; Identifying a radar reflectivity factor observed for each position of the cloud body based on the reflected wave data; a step of acquiring the state data by inputting explanatory variable data including the meteorological data corresponding to the time point when the reflected wave data was acquired and the radar reflectivity factor into a trained model that indicates the state of the cloud body and outputs a probability distribution of state data that is objective variable data; A step of identifying a model reliability, which is a reliability of the trained model when the explanatory variable data is input, based on the state data; determining a data reliability indicating a reliability of the state data corresponding to each of a plurality of spatial locations based at least on the model reliability; outputting estimated weather data indicating the state of the cloud body based on the state data; acquiring, from within a predetermined range including one or more low-reliability positions among the plurality of positions where the data reliability is relatively low, a plurality of pieces of reflected wave data corresponding to the low-reliability positions a number of times greater than or equal to the number of times the reflected wave data is acquired from within a predetermined range including positions other than the low-reliability positions; A program to execute.

Images