Supervised neural network to predict unlabeled rain rates

Patent metadata

Field	Value
Publication number	US-10754063-B2
Application number	US-201615182138-A
Country	US
Kind code	B2
Filing date	Jun 14, 2016
Priority date	Jun 14, 2016
Publication date	Aug 25, 2020
Grant date	Aug 25, 2020

Abstract

In an embodiment, radar observation data for time points are received at an input layer of a rain rate prediction neural network. The radar observations are forward propagated via hidden layers of the network to determine rain rates for the time points. The rain rates are integrated over a time period, determined based on the time points, to determine a predicted rainfall amount. The predicted rainfall amount is compared with an actual rainfall amount, determined based on received rainfall measurements, to determine an error. If the error does not satisfy certain criteria, then the error is apportioned to each of the time points, the apportioned errors are back propagated via the hidden layers, and weights associated with nodes in the hidden layers are updated. The radar observation data is again forward propagated via the layers, multiplied by the updated weights, and used to determine new rain rates.

First claim

What is claimed is:

1. A computer-implemented data processing method comprising:
using instructions that are programmed in a computer system comprising one or more processors and computer memory for;
receiving, at an input layer of a rain rate prediction neural network, radar observation data for a plurality of time points;
forward propagating the radar observation data via the rain rate prediction neural network to predict a plurality of rain rates for the plurality of time points;
determining, based on the plurality of rain rates, a predicted rainfall amount for a time period determined based on the plurality of time points;
comparing the predicted rainfall amount with an actual rainfall amount determined based on received rainfall measurements to determine an error difference value;
in response to determining that the error difference value does not satisfy one or more criteria:
apportioning the error difference value to each time point of the plurality of time points; wherein apportioning an error for a kth node in the rain rate prediction neural network having a kth weight includes computing an apportioned error difference value for the kth node by dividing the error difference value by a count of nodes and multiplying the computed apportioned error difference value by the kth weight to obtain a result;
backward propagating the result via the rain rate prediction neural network;
receiving new radar observation data; and
forward propagating the new radar observation data to predict a new plurality of rain rates for the plurality of time points.

2. The method of claim 1 , further comprising:
computing and outputting one or more predicted rain rate values with the rain rate prediction neural network, in response to determining that the error difference value satisfies the one or more criteria.

3. The method of claim 1 , further comprising, using programmed instructions:
setting a weight value for each node of the rain rate prediction neural network to a specified constant value;
predicting the plurality of rain rates for the plurality of time points by multiplying the radar observation data at each node in one or more hidden layers of the rain rate prediction neural network by a corresponding weight value.

4. The method of claim 3 , wherein multiplying the radar observation data at each node in the one or more hidden layers of the rain rate prediction neural network by a corresponding weight value includes transforming the radar observation data into predicted rain rates.

5. The method of claim 1 , further comprising determining the predicted rainfall amount using programmed instructions by integrating the plurality of rain rates for the plurality of time points over the time period; wherein integrating the plurality of rain rates for the plurality of time points over the time period to determine the predicted rainfall amount comprises applying a trapezoidal rule of numerical integration to the plurality of rain rates.

6. The method of claim 1 , wherein the error difference value represents an error between an hourly prediction of rain for the time period and rain gauge readings for the time period.

7. The method of claim 1 , wherein apportioning the error difference value to each time point of the plurality of time points includes dividing the error difference value by the count of the nodes in the rain rate prediction neural network.

8. The method of claim 1 , wherein the radar observation data comprises reflectivity measures that have been determined based on echo signals that were received by a radar system in response to, and based upon, emitting radar signals toward clouds; wherein the echo signals are the emitted radar signals reflected from the clouds.

9. The method of claim 1 , wherein the received rainfall measurements comprise actual rainfall measurement values from a plurality of rain gauges.

10. The method of claim 1 , wherein the one or more criteria comprise a threshold value, and further comprising comparing the error difference value to the threshold value to determine whether the error difference value exceeds the threshold value.

11. A data processing system comprising:
a memory;
one or more processors coupled to the memory and programmed, using instructions that are part of a rain rate prediction neural network, to:
receive, at an input layer of the rain rate prediction neural network, radar observation data for a plurality of time points;
forward propagate the radar observation data via the rain rate prediction neural network to predict a plurality of rain rates for the plurality of time points;
determine, based on the plurality of rain rates, a predicted rainfall amount for a time period determined based on the plurality of time points;
compare the predicted rainfall amount with an actual rainfall amount determined based on received rainfall measurements to determine an error difference value;
in response to determining that the error difference value does not satisfy one or more criteria:
apportion the error difference value to each time point of the plurality of time points;
wherein apportioning an error for a kth node in the rain rate prediction neural network having a kth weight includes computing an apportioned error difference value for the kth node by dividing the error difference value by a count of nodes and multiplying the apportioned error difference value by the kth weight to obtain a result;
backward propagate the result via the rain rate prediction neural network;
receive new radar observation data, and
forward propagate the new radar observation data to predict a new plurality of rain rates for the plurality of time points.

12. The data processing system of claim 11 , wherein the one or more processors are programmed to compute and output one or more predicted rain rate values with the rain rate prediction neural network, in response to determining that the error difference value satisfies the one or more criteria.

13. The data processing system of claim 11 , wherein the one or more processors are programmed to:
set a weight value for each node of the rain rate prediction neural network to a specified constant value;
predict the plurality of rain rates for the plurality of time points by multiplying the radar observation data at each node in one or more hidden layers of the rain rate prediction neural network by a corresponding weight value.

14. The data processing system of claim 13 , wherein multiplying the radar observation data at each node in one or more hidden layers of the rain rate prediction neural network by a corresponding weight value includes transforming the radar observation data into predicted rain rates.

15. The data processing system of claim 11 ,
wherein the predicted rainfall amount is determined by integrating the plurality of rain rates for the plurality of time points over the time period;
wherein integrating the plurality of rain rates for the plurality of time points over the time period to determine the predicted rainfall amount comprises applying a trapezoidal rule of numerical integration to the plurality of rain rates.

16. The data processing system of claim 11 , wherein the error difference value represents an error between an hourly prediction of rain for the time period and rain gauge readings for the time period.

17. The data processing system of claim 11 , wherein apportioning the error difference value to each time point of the plurality of time points includes dividing the error difference value by the count of nodes in the rain rate prediction neural network.

Assignees

• Climate Corp

Inventors

• Lakshmanan Valliappa

Classifications

• G06F16/535

  Filtering based on additional data, e.g. user or group profiles · CPC title

• G01W1/10 Primary

  Devices for predicting weather conditions (computers per se G06; display devices G09) · CPC title

• G06Q10/04

  Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem" (market predictions or forecasting for commercial activities G06Q30/0202) · CPC title

• G01S13/951

  ground based · CPC title

• G01S7/417

  involving the use of neural networks · CPC title