Multi-Horizon Direction of Arrival Forecasting With Temporal Fusion Transformers: A Time-Series Learning Approach to Target Tracking

This paper explores the application of Temporal Fusion Transformers (TFTs) to target recognition and multi-horizon tracking, with a particular emphasis on enhancing Direction of Arrival (DoA) estimation in complex signal environments. By leveraging the advanced temporal modeling capabilities and dynamic feature selection mechanisms inherent to the TFT architecture, we adapt and optimize the model for improved performance in antenna signal processing, leading to a more accurate and robust estimation framework. Our analysis underscores the architectural strengths of TFTs, especially their capacity to model long-range temporal dependencies and to dynamically prioritize sensor-derived features—capabilities that are essential for precise target tracking and recognition. We train and evaluate the model under diverse conditions, including multi-horizon forecasting and varying prediction latencies across models. The proposed method is benchmarked against state-of-the-art algorithms, including recurrent neural networks and standard transformer-based models. Extensive experimental results demonstrate that the TFT-based approach consistently outperforms existing techniques in forecasting accuracy, multi-target recognition, and computational complexity, highlighting its potential to advance the field of DoA estimation.