Better Multi-step Time Series Prediction Using Sparse and Deep Echo State Network

Abstract

Multi-step time series prediction is essential in real-world applications but challenging to obtain accurately due to a fallacy accumulation. Incrementing the required future steps typically results in performance degradation. Data-driven machine learning techniques have the potential to tackle this task but demand significant or special computing powers such as memory and graphics processing units (GPUs). This work demonstrates that Deep Echo State Network (DeepESN) with a sparse configuration can capture multi-step prediction in a comparable or even better performance while demanding lower resources and processing times. Most experimental results documented in the literature examine only one or a few multi-step ahead. Here we report the prediction of up to 250 future steps with better correlation-of-coefficient contrasting to the baseline models. Sparsing the projection of the input signal to each reservoir of the DeepESN can reduce the circumstances of overfitting in time series learning. This finding could lead to utilizing deep learning models with affordable resources and processing times.