utokyo-hackathon's Introduction

Hybrid Forecasting of Chaotic Processes

This repository contains implementations and presentation materials on our group research conducted as a part of the UTokyo IST Research Hackathon organized from September 20th – 24th, 2021.

A huge thanks to everyone involved in this project:

Akshat Verma
Pongsakorn Chairatanakul
Alexander Thomas Magro (Mentor)
Wentao Sun (Mentor)

Overview

We replicate the experiment to forecast a Lorenz system by using a hybrid model presented in this paper. Specifically, we compare the performance of a traditional ESN model and a hybrid model (i.e. a combination between ESN and an imperfect knowledge-based model) on the generated benchmark data:
- Knowledge-based model: an imperfect model of Lorenz system. Imperfections can be represented by a slight error in parameter b (as described in the paper) or a difference in timestep between the benchmark data and that of a knowledge-based model when we observe the system (our original work).
- Echo State Network (ESN): a simple model in reservoir computing. A simple explanation of the model is described here.
- Hybrid-Model: a combination of knowledge-based model and ESN, proposed in the paper.
We further investigate the performance of these models under a different type of imperfection, a difference in timestep. The experiment suggested that the hybrid model also outperform the original one.
We study the relationship between the allocation ratio (time used knowledge-based model divided by ESN model) and valid time (time until predictions become diverged).

Recommend Projects