Minimizing Chaos in Echo State Networks: A Hybrid Approach Using the Lorenz System

Abstract

In recent years, machine learning, especially deep neural networks, has made substantial progress, consistently surpassing conventional time-series forecasting methods across various domains.This paper introduces a novel hybrid approach that combines the Lorenz system and the echo state network (ESN) to tackle and reduce the "butterfly effect" in chaos forecasting.The core contribution lies in harnessing the Lorenz system's unique properties, where initially converging trajectories gradually diverge, to train the ESN-a neural network celebrated for its non-linear computational capabilities, echo state property, and input forgetting capability.The primary aim is to establish a more robust and precise framework for predicting chaotic systems, given their sensitivity to initial conditions.This research endeavors to provide a versatile tool with wide-ranging applications, particularly in areas like stock price prediction, where accurately forecasting chaotic behavior holds paramount importance.The Lorenz system initiates with nearly identical initial states, differing by a mere 10 -3 in the x-coordinate at t=0.Initially, these trajectories seem to overlap, but after t=1000, they significantly diverge.In this proposed approach, data from t=0 to t=1000 serves as the training input for the ESN.Once the training phase concludes, the ESN's formidable non-linear computational capabilities, echo state property, and input forgetting capability render it exceptionally well-suited for stepwise predictions and tasks sensitive to initial conditions.The simulation results demonstrate that over the subsequent 360 prediction steps conducted by the ESN, the "butterfly effect" stemming from the slightly varying initial states provided to the Lorenz System is effectively minimized.Notably, the simulation results underscore the superior performance of our hybrid approach, revealing a minimal root mean square error (RMSE) of less than 1.0.In contrast, a prior study introduced the MrESN (Multiple Reservoir Echo State Network) approach, which is a specific type of Echo State Network (ESN) used for forecasting multivariate chaotic time series.It employs multiple internal reservoirs within the network architecture to handle the complex dynamics of chaotic data but achieved lower accuracy with a larger RMSE of 43.70.Another preceding algorithm, BFA-DRESN, aimed at enhancing forecasting accuracy but yielded an RMSE value of 18.83.This research advances ESNbased predictability, offering a promising solution for addressing the challenges posed by chaos.