FORECASTING METHODS OF TIME SERIES WITH EXPRESSED SEASONALITY BASED ON TRANSFORMER

Abstract

Forecasting time series data with seasonal patterns remains a significant challenge in various domains. The complex and non-stationary nature of such data requires advanced modeling techniques capable of capturing intricate temporal dependencies and periodic fluctuations. Traditional statistical methods like ARIMA models often struggle to adequately represent the underlying dynamics, motivating the exploration of more flexible and powerful approaches. In recent years, deep learning architectures, particularly transformer-based models, have demonstrated remarkable success in handling sequential data and capturing long-range dependencies. Their self-attention mechanisms and parallel processing capabilities make them well-suited for time series forecasting tasks involving seasonality. This study provides a comprehensive evaluation of three prominent transformer-based models: Temporal Fusion Transformer (TFT), PatchTST, and DLinear. Each model brings unique architectural innovations and training strategies to tackle the intricacies of seasonal time series forecasting effectively. The experimental results, evaluated on a real-world product sales dataset, reveal the potential of these models to outperform traditional methods and achieve superior forecasting accuracy. However, it is crucial to consider the trade-off between model complexity, computational resources, and the specific requirements of the forecasting task at hand. As the field of time series forecasting continues to evolve, transformer-based models offer a promising direction for handling complex seasonal patterns, paving the way for more accurate and reliable predictions across various industries. Nevertheless, further research is necessary to address the challenges of model interpretability, efficient parameter optimization, and the incorporation of domain-specific knowledge into these architectures.