The prediction of physical assets future usage behavior is critical for asset-owning companies and leasing providers, as it directly affects operational continuity, maintenance planning, and financial strategies. Telemetry data, such as energy consumption or usage time, provides the basis for this analysis and forecasting. Telemetry time series often exhibit non-stationary behavior, with seasonalities and trends that can abruptly or gradually change over time. Traditional statistical forecasting models struggle to capture such regime shifts, while modern deep learning approaches typically require large datasets that are difficult to obtain in practice and lack interpretability. In this thesis, we present a novel forecasting algorithm specifically designed for asset telemetry in non-stationary environments. The method decomposes the time series into trend and seasonal components, employs Bayesian techniques to detect abrupt changepoints, and integrates this information into a Prophet-based forecasting framework. To improve reliability, the algorithm introduces mechanisms for robustness against noise and outliers, as well as a sliding window approach to adapt to gradual evolving patterns. We validate the proposed algorithm through extensive experiments on both synthetic and real-world datasets. Results show that our solution not only outperforms traditional baselines such as Auto-ARIMA and ETS (Exponential Time Smoothing), but also improves upon Prophet in terms of accuracy. The result is a model that is accurate, interpretable, robust and adaptable across different asset categories and operational regimes.
La previsione del comportamento futuro di utilizzo degli asset fisici è fondamentale per le aziende e le società di leasing, poiché influisce direttamente sulla continuità operativa, sulla pianificazione della manutenzione e sulle strategie finanziarie. I dati telemetrici, come il consumo energetico o il tempo di utilizzo, costituiscono la base per tale analisi e previsione. Le serie temporali telemetriche presentano spesso un comportamento non stazionario, con stagionalità e trend che possono variare in modo improvviso o graduale nel tempo. I modelli statistici tradizionali di previsione faticano a catturare questi cambiamenti di regime, mentre gli approcci moderni basati sul deep learning richiedono in genere grandi quantità di dati, difficili da reperire nella pratica, e risultano poco interpretabili. In questa tesi presentiamo un nuovo algoritmo di previsione, progettato specificamente per le serie telemetriche degli asset in ambienti non stazionari. Il metodo scompone la serie temporale nei suoi componenti di trend e stagionalità, utilizza tecniche bayesiane per rilevare improvvisi punti di cambiamento e integra queste informazioni in un framework di previsione basato su Prophet. Per migliorarne l’affidabilità, l’algoritmo introduce meccanismi di robustezza al rumore e agli outlier, oltre a un approccio a finestra mobile per adattarsi a pattern che evolvono gradualmente. Abbiamo validato l’algoritmo proposto attraverso esperimenti estensivi sia su dataset sintetici che reali. I risultati mostrano che la nostra soluzione non solo supera modelli tradizionali di riferimento come Auto-ARIMA ed ETS (Exponential Time Smoothing), ma migliora anche le prestazioni di Prophet in termini di accuratezza. Il risultato è un modello accurato, interpretabile, robusto e adattabile a diverse categorie di asset e regimi operativi.
A forecasting algorithm for asset telemetries in non-stationary environments
FARACE, GABRIELE
2024/2025
Abstract
The prediction of physical assets future usage behavior is critical for asset-owning companies and leasing providers, as it directly affects operational continuity, maintenance planning, and financial strategies. Telemetry data, such as energy consumption or usage time, provides the basis for this analysis and forecasting. Telemetry time series often exhibit non-stationary behavior, with seasonalities and trends that can abruptly or gradually change over time. Traditional statistical forecasting models struggle to capture such regime shifts, while modern deep learning approaches typically require large datasets that are difficult to obtain in practice and lack interpretability. In this thesis, we present a novel forecasting algorithm specifically designed for asset telemetry in non-stationary environments. The method decomposes the time series into trend and seasonal components, employs Bayesian techniques to detect abrupt changepoints, and integrates this information into a Prophet-based forecasting framework. To improve reliability, the algorithm introduces mechanisms for robustness against noise and outliers, as well as a sliding window approach to adapt to gradual evolving patterns. We validate the proposed algorithm through extensive experiments on both synthetic and real-world datasets. Results show that our solution not only outperforms traditional baselines such as Auto-ARIMA and ETS (Exponential Time Smoothing), but also improves upon Prophet in terms of accuracy. The result is a model that is accurate, interpretable, robust and adaptable across different asset categories and operational regimes.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_10_Farace_Executive_Summary_02.pdf
non accessibile
Descrizione: testo tesi
Dimensione
760.37 kB
Formato
Adobe PDF
|
760.37 kB | Adobe PDF | Visualizza/Apri |
|
2025_10_Farace_Tesi_02.pdf
non accessibile
Descrizione: executive summary
Dimensione
5.01 MB
Formato
Adobe PDF
|
5.01 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/243569