Dissertation/ Thesis
Implementation of the temporal fusion transformer for demand forecasting
| Τίτλος: | Implementation of the temporal fusion transformer for demand forecasting |
|---|---|
| Συγγραφείς: | Llaquet Vélez, Santiago |
| Συνεισφορές: | Zamora Fernandez, Sergi, Universitat de Barcelona, Accenture |
| Πηγή: | UPCommons. Portal del coneixement obert de la UPC Universitat Politècnica de Catalunya (UPC) |
| Στοιχεία εκδότη: | Universitat Politècnica de Catalunya, 2025. |
| Έτος έκδοσης: | 2025 |
| Θεματικοί όροι: | Classificació AMS::91 Game theory, economics, social and behavioral sciences::91B Mathematical economics, Àrees temàtiques de la UPC::Matemàtiques i estadística, Time Series Forecasting, Demand Forecasting, Multi-Horizon Forecasting, social and behavioral sciences::91B Mathematical economics, Classificació AMS::91 Game theory, economics, Transformer-based models, Deep Learning, Predicció (Estadística), Classificació AMS::68 Computer science::68T Artificial intelligence, Machine learning, Explainable AI, Supply Chain, Aprenentatge automàtic, Classificació AMS::62 Statistics::62M Inference from stochastic processes, Interpretability, Forecasting, Aprenentatge profund |
| Περιγραφή: | Time Series Forecasting plays a pivotal role across many domains, with Demand Forecasting being particularly essential for effective business planning and decision-making. While classical time series methods and Machine Learning models have achieved considerable success, they face limitations in capturing complex temporal dynamics, scaling to high-dimensional data, and offering robust interpretability. Deep Learning offers a promising alternative capable of addressing these limitations. The Temporal Fusion Transformer (TFT), a state-of-the-art Deep Learning architecture, advances multi-horizon forecasting by combining superior performance with interpretability. TFT addresses key challenges in traditional approaches through sequence encoders for local temporal processing, self-attention mechanisms for long-term dependencies, and gating mechanisms to filter irrelevant inputs, ensuring adaptability and scalability across diverse scenarios. This thesis presents an implementation of TFT for Demand Forecasting, focusing on architectural enhancements, improved performance over established methods, enhanced explainability, and practical applicability. It also evaluates TFT’s ability to address the common challenges of Transformer-based models in Time Series Forecasting. The findings assert TFT’s potential as a transformative tool for advancing Demand Forecasting methodologies and bridging the gap between state-of-the-art research and real-world business needs. |
| Τύπος εγγράφου: | Master thesis |
| Περιγραφή αρχείου: | application/pdf |
| Γλώσσα: | English |
| Σύνδεσμος πρόσβασης: | https://hdl.handle.net/2117/424614 |
| Rights: | CC BY NC ND |
| Αριθμός Καταχώρησης: | edsair.dedup.wf.002..cb559f90b6fbd77f8b5812cc985d90ff |
| Βάση Δεδομένων: | OpenAIRE |
καταχωρήστε σχόλιο πρώτοι!