KODA: A Data-Driven Recursive Model for Time Series Forecasting and Data Assimilation using Koopman Operators

Ashutosh Singh,Ashish Singh,Tales Imbiriba,Deniz Erdogmus,Ricardo Borsoi

2024-09-29

Abstract:Approaches based on Koopman operators have shown great promise in forecasting time series data generated by complex nonlinear dynamical systems (NLDS). Although such approaches are able to capture the latent state representation of a NLDS, they still face difficulty in long term forecasting when applied to real world data. Specifically many real-world NLDS exhibit time-varying behavior, leading to nonstationarity that is hard to capture with such models. Furthermore they lack a systematic data-driven approach to perform data assimilation, that is, exploiting noisy measurements on the fly in the forecasting task. To alleviate the above issues, we propose a Koopman operator-based approach (named KODA - Koopman Operator with Data Assimilation) that integrates forecasting and data assimilation in NLDS. In particular we use a Fourier domain filter to disentangle the data into a physical component whose dynamics can be accurately represented by a Koopman operator, and residual dynamics that represents the local or time varying behavior that are captured by a flexible and learnable recursive model. We carefully design an architecture and training criterion that ensures this decomposition lead to stable and long-term forecasts. Moreover, we introduce a course correction strategy to perform data assimilation with new measurements at inference time. The proposed approach is completely data-driven and can be learned end-to-end. Through extensive experimental comparisons we show that KODA outperforms existing state of the art methods on multiple time series benchmarks such as electricity, temperature, weather, lorenz 63 and duffing oscillator demonstrating its superior performance and efficacy along the three tasks a) forecasting, b) data assimilation and c) state prediction.

Machine Learning,Artificial Intelligence

What problem does this paper attempt to address?

The paper attempts to address the difficulties faced by existing Koopman operator-based methods in long-term prediction and data assimilation when dealing with time series data generated by complex nonlinear dynamical systems (NLDS). Specifically:

1. **Error accumulation in long-term prediction**: Although Koopman operator-based methods can capture the latent state representation of NLDS, these methods still face the problem of error accumulation in long-term prediction when applied to real data, leading to significant drift in state estimation.

2. **Non-stationarity issue**: Many real-world NLDS exhibit time-varying behavior, making non-stationarity difficult to capture with existing models.

3. **Lack of systematic data assimilation methods**: Existing methods lack a systematic data-driven approach to real-time data assimilation for prediction tasks using noisy measurement data.

To alleviate the above issues, the authors propose a Koopman operator-based method (named KODA - Koopman Operator with Data Assimilation) that integrates prediction and data assimilation into NLDS. Specifically, KODA uses Fourier domain filters to decompose data into physical components and residual components, and captures local or time-varying behavior through a flexible and learnable recurrent model. Additionally, KODA introduces an online data assimilation strategy that utilizes new measurement data for correction during inference. Through extensive experimental comparisons, the authors demonstrate the superior performance of KODA on multiple time series benchmark datasets, including electricity, temperature, weather, Lorenz 63, and Duffing oscillator.