Improving Forecasting Performance for Abnormal Time Series Data with the TFT-TPE Integrated Model and Google Trends

Abstract

Forecasting is essential in manufacturing and business, but is hindered by abnormal events like COVID-19. This paper proposes a model that integrates Temporal Fusion Transformer (TFT) with Tree-Structured Parzen Estimator (TPE), in which TFT is a deep neural network specifically designed for processing time series data to capture trends and model complex data variations and, at the same time, TPE is an optimization technique that uses a tree-like data structure to determine the best set of hyperparameters for TFT. The TFT-TPE integrated model, therefore, provides an effective solution to the forecasting problem, especially for abnormal data. The study proposes a combination of forecasting historical data, considering the COVID-19 period, and utilizing Google Trends to enhance forecasting accuracy. The experimental results show that the TFT-TPE integrated model achieves forecasting results better than traditional forecasting models, especially the ability to overcome the anomalies in time series data.

Keywords

• transformer
• tft-tpe integration
• abnormal time series data
• google trends
• tourism demand forecasting