Abstract
Time series forecasting plays a vital role across scientific, industrial, andenvironmental domains, especially when dealing with high-dimensional andnonlinear systems. While Transformer-based models have recently achievedstate-of-the-art performance in long-range forecasting, they often suffer frominterpretability issues and instability in the presence of noise or dynamicaluncertainty. In this work, we propose DeepKoopFormer, a principled forecastingframework that combines the representational power of Transformers with thetheoretical rigor of Koopman operator theory. Our model features a modularencoder-propagator-decoder structure, where temporal dynamics are learned via aspectrally constrained, linear Koopman operator in a latent space. We imposestructural guarantees-such as bounded spectral radius, Lyapunov based energyregularization, and orthogonal parameterization to ensure stability andinterpretability. Comprehensive evaluations are conducted on both syntheticdynamical systems, real-world climate dataset (wind speed and surfacepressure), financial time series (cryptocurrency), and electricity generationdataset using the Python package that is prepared for this purpose. Across allexperiments, DeepKoopFormer consistently outperforms standard LSTM and baselineTransformer models in terms of accuracy, robustness to noise, and long-termforecasting stability. These results establish DeepKoopFormer as a flexible,interpretable, and robust framework for forecasting in high dimensional anddynamical settings.