Multi-Scale Transformer for Long-Term Time Series Forecasting

In order to make well-informed decisions, long-term time series forecasting is crucial for a number of applications in finance, energy, and environmental science. While traditional transformer models have demonstrated a strong ability to capture temporal dependencies, they frequently struggle to handle the lengthy sequences and multi-scale temporal patterns present in time series data. In this paper, we propose a novel architecture to improve forecasting performance over long horizons: the Multi-Scale Transformer. Our model effectively captures both short-term fluctuations and long-term trends by extending the transformer by integrating multi-scale processing. In addition, we suggest a feature interaction module that allows the model to learn intricate interdependencies in the data by explicitly modeling the relationships between individual features. We compare our models’ performance to the most advanced techniques using a variety of benchmark datasets. According to experimental results, our method continuously outperforms current models and achieves notable gains in forecasting accuracy. A promising path forward for long-term time series forecasting is the transformer framework's incorporation of multi-scale analysis.