Data-Driven Prediction Using Spatial Information Constraints
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for data-driven prediction, particularly in fields like earth science, face challenges due to insufficient consideration of spatial correlation between data points, limiting the applicability of intelligent learning methods.
Innovation Solution
A method and system for data-driven prediction based on spatial information constraints, utilizing collocated Co-Kriging for interpolation, sequential Gaussian simulation for sample weight calculation, and deep fully connected neural networks with a spatially constrained loss function to address the scarcity of representative learning samples.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning methods are applied to professional research fields with expensive data acquisition, then prediction accuracy can be improved, but the scarcity and lack of representativeness of learning sample data prevents effective application
Solution Approach 1:
The patent uses Co-Kriging interpolation to create synthetic training samples that copy the spatial correlation characteristics of expensive-to-acquire geophysical observation data. By interpolating from a small set of real measurement points to generate a large set of synthetic points that preserve spatial relationships, the method replicates the valuable information content without requiring actual field measurements at every location.
Solution Approach 2:
The patent introduces Co-Kriging interpolation as an intermediary step between the scarce real measurement data and the deep learning model training process. This intermediary generates intermediate synthetic samples that bridge the gap between limited real data and the large data quantity required by deep learning, allowing the model to learn from data that reflects true spatial correlations.
2Reliability
If convolution neural networks are used to consider spatial distribution locally, then spatial correlation can be captured, but the limited input sample size prevents global spatial correlation consideration
Solution Approach 1:
The patent transforms the problem from a two-dimensional limitation (local spatial correlation only) to a three-dimensional solution by adding the temporal/dimension of synthetic data generation. By creating synthetic samples through Co-Kriging interpolation that preserve global spatial correlations, the method enables deep learning models to consider spatial relationships across the entire domain rather than just local neighborhoods.
3Device complexity
If deep fully connected neural networks treat each sampling point as isolated, then computational simplicity is maintained, but spatial correlation information is completely lost
Solution Approach 1:
The patent performs preliminary action by applying Co-Kriging interpolation to the training data before it enters the deep learning model. This preliminary processing embeds spatial correlation information into the synthetic samples themselves, so that when the data-driven prediction model processes these pre-processed samples, the spatial relationships are already encoded in the data structure, allowing the model to learn spatial patterns without complex architectural modifications.
Data Source
AI summary
Provided herein is a method and a system for data-driven prediction based on spatial information constraints, belonging to the technical field of intelligent information processing. The method comprises: prediction target interpolation based on collocated Co-Kriging; sample weight calculation based on sequential Gaussian simulation and loss function construction based on spatial information constraints; optimization of loss function and data-driven prediction based on deep fully connected neural network. The system comprises: data acquisition module, data preprocessing module, prediction target interpolation module, sample weight calculation module, loss function construction module, loss function optimization module, data-driven prediction module. It realizes the expansion of learning samples under the restriction of spatial information, and uses the spatial information to optimize the loss function, thus improving the utilization rate of data information, facilitating guiding the learning process to converge to reasonable assumptions, thereby improving the performance of the prediction method based on data-driven.