Learning, fast and slow: a two-fold algorithm for data-based model adaptation

TL;DR

This paper introduces a two-fold adaptive modeling approach combining slow ensemble learning for out-of-domain uncertainties with fast Gaussian process updates for in-domain variability, enhancing model accuracy over time.

Contribution

The paper presents a novel two-fold architecture that integrates slow ensemble learning and fast Gaussian process updates for dynamic model adaptation.

Findings

Improved model accuracy over standard methods.

Effective detection of new operating conditions.

Enhanced adaptation to both out-of-domain and in-domain uncertainties.

Abstract

This article addresses the challenge of adapting data-based models over time. We propose a novel two-fold modelling architecture designed to correct plant-model mismatch caused by two types of uncertainty. Out-of-domain uncertainty arises when the system operates under conditions not represented in the initial training dataset, while in-domain uncertainty results from real-world variability and flaws in the model structure or training process. To handle out-of-domain uncertainty, a slow learning component, inspired by the human brain's slow thinking process, learns system dynamics under unexplored operating conditions, and it is activated only when a monitoring strategy deems it necessary. This component consists of an ensemble of models, featuring (i) a combination rule that weights individual models based on the statistical proximity between their training data and the current operating condition, and (ii) a monitoring algorithm based on statistical control charts that supervises the ensemble's reliability and triggers the offline training and integration of a new model when a new operating condition is detected. To address in-domain uncertainty, a fast learning component, inspired by the human brain's fast thinking process, continuously compensates in real time for the mismatch of the slow learning model. This component is implemented as a Gaussian process (GP) model, trained online at each iteration using recent data while discarding older samples. The proposed methodology is tested on a benchmark energy system referenced in the literature, demonstrating that the combined use of slow and fast learning components improves model accuracy compared to standard adaptation approaches.