Two Efficient Twin ELM Methods With Prediction Interval.

Area of Science:

• Computational Intelligence
• Machine Learning
• Industrial Process Modeling

Background:

• Industrial processes like iron/steel and microelectronics require accurate prediction of operational indices and parameters.
• Uncertainties in input/output variables necessitate the construction of reliable prediction intervals (PIs).
• Existing prediction models often struggle with inherent data uncertainties and measurement errors.

Purpose of the Study:

• To propose two novel twin extreme learning machine (TELM) models for constructing accurate prediction intervals (PIs).
• To address uncertainties in industrial process predictions by developing robust regression models.
• To enhance the reliability of predictions in operational optimization and scheduling.

Main Methods:

• Developed a regularized asymmetric least squares extreme learning machine (RALS-ELM) with differential error weighting and Tikhonov regularization.
• Proposed an asymmetric Bayesian extreme learning machine (AB-ELM) utilizing an asymmetric Gaussian distribution and type II maximum likelihood.
• Employed a pair of reciprocal weights to generate lower and upper bound regressors for PI calculation.

Main Results:

• The proposed RALS-ELM and AB-ELM models demonstrated effectiveness in constructing prediction intervals.
• Numerical comparisons on synthetic, benchmark, and industrial datasets confirmed the models' performance.
• The methods successfully handled uncertainties, leading to more reliable predictions.

Conclusions:

• The developed TELM models offer a robust solution for prediction interval construction in industrial settings.
• The asymmetric weighting and regularization techniques improve prediction accuracy and reduce overfitting.
• Future work includes adaptive weight adjustment and application to nonlinear quantile regression.