Probabilistic Prediction Model Quantile Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current learning models are inadequate in quantifying uncertainties and providing prediction intervals that meet required confidence levels, particularly in industrial applications like oil and gas production wells and electric batteries, where risk management requires more comprehensive prediction tools.
Innovation Solution
A method for training a probabilistic prediction model that incorporates a predefined distribution quantile to optimize model parameters, allowing for accurate prediction of quantiles and uncertainty incorporation in prediction intervals, using techniques such as cross-validation and distance minimization between covariance matrices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional bootstrap or maximum likelihood estimation is used for prediction, then the method is simple and widely applicable, but the prediction intervals do not meet required confidence levels and fail to accurately quantify uncertainties
Solution Approach 1:
The patent changes the optimization parameter from traditional point prediction error to quantile-based criteria. By optimizing model parameters to match predicted quantiles with empirical quantiles from bootstrap samples, the method achieves accurate prediction intervals that meet required confidence levels while maintaining the simplicity of bootstrap resampling.
Solution Approach 2:
The patent implements a feedback mechanism where predicted quantiles are compared with empirical quantiles from bootstrap samples, and model parameters are iteratively adjusted to minimize the difference. This feedback loop ensures that prediction intervals accurately reflect the true uncertainty distribution.
2Loss of information
If point prediction is used to produce a single value, then the prediction is simple and direct, but information about prediction uncertainty and distance from true values is lost
Solution Approach 1:
The patent transitions from one-dimensional point prediction to two-dimensional prediction by simultaneously estimating both the central tendency (mean/median) and the uncertainty distribution (quantiles). This dimensional expansion preserves uncertainty information while maintaining prediction efficiency through integrated model training.
3Reliability
If maximum likelihood estimation is used to fit parameters, then the solution fits observations in average, but coverage and width of prediction intervals are not optimized
Solution Approach 1:
The patent changes the optimization objective from minimizing point prediction error to minimizing quantile prediction error. By focusing on matching empirical quantiles with predicted quantiles, the method optimizes prediction interval coverage and reliability without sacrificing point prediction accuracy, as both are estimated within the same probabilistic framework.
Data Source
AI summary
This method for predicting value(s) of a quantity relative to a target device is implemented by an electronic prediction system and comprises the following steps:training a prediction probabilistic model, said training including:receiving measured values of the quantity for N devices, N>1,calculating predicted values of the quantity for said N devices with the prediction probabilistic model,computing a criteria based on the measured values and on the predicted values,modifying model parameter(s) of the prediction probabilistic model according to the computed criteria,updating the prediction probabilistic model with the modified model parameter(s),predicting value(s) of the quantity relative to the target device with the trained prediction probabilistic model;wherein the criteria depends on a predefined distribution quantile.