Precipitation Estimation Using Segmented Occurrence and Intensity Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current modeling techniques for estimating rainfall at precise locations face challenges such as neglecting spatial correlation, computational inefficiency, and inability to propagate probabilistic estimates, leading to inaccurate large-scale precipitation forecasts.
Innovation Solution
A computer-implemented two-stage modeling approach that separately estimates precipitation occurrence and intensity using expectation propagation and fully independent training conditional approximation, creating probabilistic estimates for large-scale fields by accounting for neighboring locations and reducing computational complexity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If current modeling techniques are used to estimate rainfall at given locations, then rainfall estimates can be obtained, but spatial correlation of rainfall is not taken into account causing probability of rainfall to sharply increase if neighboring gauges observed rainfall
Solution Approach 1:
The patent segments the rainfall estimation process into two separate models: a precipitation occurrence model that estimates the probability of rainfall occurrence, and a precipitation intensity model that estimates the intensity given occurrence. This segmentation allows the occurrence model to properly account for spatial correlation through Gaussian process regression, while the intensity model handles the conditional intensity estimation, resolving the contradiction between measurement precision and spatial correlation reliability.
Solution Approach 2:
The patent introduces an intermediary variable - the precipitation occurrence probability - that mediates between the observed gauge data and the final rainfall intensity estimate. This intermediary allows the system to first determine whether rainfall occurs at a location (accounting for spatial correlation), and then estimate the intensity conditional on that occurrence, thereby resolving the sharp increase problem in probability estimation.
2Measurement precision
If current models estimate rainfall at given locations, then point estimates are obtained, but creating large scale models based on wide array of observations becomes computationally prohibitive
Solution Approach 1:
The patent divides the computational task into two separate modeling stages: occurrence estimation using Gaussian process regression, and intensity estimation using a conditional model. This segmentation reduces the overall computational burden compared to estimating the full joint distribution, as each stage operates on simplified assumptions and requires fewer computational resources while maintaining accuracy.
Solution Approach 2:
The patent changes the parameter representation by separating the rainfall field into occurrence probability parameters and intensity parameters. This parameter transformation allows the use of computationally efficient Gaussian process regression for occurrence, followed by a simpler conditional intensity model, thereby improving productivity while maintaining measurement precision.
3Measurement precision
If many modeling techniques estimate rainfall at a given location, then point estimates are obtained, but they are unable to propagate probabilistic estimates in creating large scale fields
Solution Approach 1:
The patent implements feedback by using the estimated occurrence probability field to inform the intensity estimation process across the entire domain. The occurrence model's output feeds into the intensity model, allowing probabilistic estimates to be propagated consistently across large scale fields. This feedback mechanism ensures that uncertainty and probability information are maintained throughout the estimation process rather than being lost.
Solution Approach 2:
The patent applies local quality by allowing the precipitation occurrence and intensity to have different spatial characteristics and probability distributions at different locations. The Gaussian process regression captures local spatial correlation in occurrence probabilities, while the conditional intensity model captures local intensity characteristics, together enabling proper propagation of probabilistic estimates across the entire field.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method for estimating precipitation values and associated uncertainties is provided. In an embodiment, precipitation records that indicate the occurrence and intensity of precipitation at specific locations are received by a weather computing system. The weather computing system uses the gauge information to separately create multiple realizations of precipitation occurrence fields and precipitation intensity fields. The weather computing system may model the occurrence of precipitation by proposing a value for each point independently and using the proposed value to update all prior proposals. The weather computing system may model the intensity of precipitation by modeling the spatial correlation of precipitation intensity and sampling from distributions at each location to determine the intensity of precipitation at each location. The weather computing system may then combine the precipitation intensity and occurrence fields into one or more final estimate fields.