A Comparative Forecasting Performance Analysis of Symbolic Representation Techniques for Time Series Using Transformer Models and Multi-Head Attention Mechanisms
الملخص
The prediction of future values from historical time series data is widely applied infields such as industrial monitoring, finance, and decision support systems. Whilemost forecasting approaches operate directly on numerical time series, symbolicrepresentations have emerged as an alternative that can reduce dimensionality, fil-ter noise, and improve interpretability. However, their effectiveness for forecastingtasks particularly when combined with modern deep learning architectures—remainsinsufficiently explored. This thesis conducts a structured comparison between nu-merical and symbolic approaches to time series forecasting, employing deep learn-ing techniques with particular emphasis on Long Short-Term Memory (LSTM) mod-els and Transformer-based architectures. Numerical time series are transformedinto symbolic sequences using Symbolic Aggregate approXimation (SAX) and Com-pressed adaptive Brownian bridge-based symbolic aggregation (ABBA)-inspired rep-resentation, and both representations are evaluated under identical experimentalconditions. Experiments are conducted on datasets of increasing complexity, includ-ing simple synthetic signals, medium complexity averaged time series, and complexreal-world datasets. The results show that symbolic representations can outperformnumerical representations on simple datasets and achieve comparable performanceon medium complexity datasets, while no clear advantage is observed for highlycomplex and volatile time series. Symbolic inputs demonstrate robustness to noiseand strong pattern-learning capabilities, particularly for smooth or quasi-periodicsignals commonly found in industrial and IoT monitoring applications. However,for complex datasets, symbolic abstraction may lead to temporal misalignment andloss of fine-grained information. Additionally, the study highlights limitations ofthe Mean Absolute Percentage Error (MAPE) metric for complex and non-stationarydata, emphasizing the need for complementary evaluation measures. Overall, thiswork provides a structured evaluation of symbolic time series forecasting in the con-text of modern deep learning, clarifying when symbolic representations are benefi-cial and identifying their limitations as data complexity increases.
الكلمات المفتاحية
غير متوفر
التخصصات
رسائل وأطاريح مشابهة
- Multi-Scale Deep Learning Frameworks for Trajectory Prediction, Scene Understanding, and Perception in Autonomous Driving Xingnan Zhou · Concordia University · 2026
- Interaction-Aware Maneuver and Lane Change Prediction Using Deep Learning Architectures for Autonomous Driving Petrit Rama · Kaiserslautern University of Technology · 2026
- On the limits of privacy in federated learning : personalization and adversarial analysis Sargam Gupta · KTH Royal Institute of Technology · 2026
- Essays on Time Series Forecasting and Machine Learning Frank, Johannes · 2025
- Applications of artificial intelligence for explainability and uncertainty quantification for performance-based design and damage classification in civil engineering Shabbir, Khurram · Aalto University · 2025
- Hybrid Machine Learning and Time-Series Modelling to Forecast Product Demand: A CRISP-DM Approach Niklas Virtanen · University of Vaasa · 2026
