Federico Carollo
Adaptive and Causally-Consistent Conformal Calibration for Non-Stationary Multi-Step Time Series Forecasting.
Rel. Michela Meo, Greta Vallero, Daniela Renga. Politecnico di Torino, Corso di laurea magistrale in Ingegneria Informatica (Computer Engineering), 2026
|
Preview |
PDF (Tesi_di_laurea)
- Tesi
Licenza: Creative Commons Attribution Non-commercial No Derivatives. Download (6MB) | Preview |
|
|
Archive (ZIP) (Documenti_allegati)
- Altro
Licenza: Creative Commons Attribution Non-commercial No Derivatives. Download (648kB) |
Abstract
Accurate, reliable uncertainty quantification is vital for time series forecasting, especially in energy systems where prediction errors directly affect grid stability, operating decisions, and costs. Deep neural networks like TiDE, iTransformer, and TSMixer handle quantile regression well, but they offer no formal statistical guarantees and often fall short under non-stationary conditions. Conformal Prediction (CP) solves part of this problem by guaranteeing valid marginal coverage without assuming an underlying distribution. Yet standard methods like Conformalized Quantile Regression (CQR) depend on temporal exchangeability, while current dynamic variants fail to respect the observation lag present in multi-step-ahead forecasting. This thesis introduces a causally consistent extension of the Reservoir Conformal Prediction (ResCP) framework, designed for non-stationary, multi-step time series forecasting.
First, to capture evolving context, we use Echo State Networks (ESNs) with orthogonal, simple cycle, and random sparse initializations to map past dynamics into latent memory states
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Corso di laurea
Classe di laurea
Aziende collaboratrici
URI
![]() |
Modifica (riservato agli operatori) |
