Forecasting chaotic weather variables with echo state networks and a novel swing training approach

Abstract

Physical systems like weather variables, the behavior of oceanic bodies, economic variables, and similar systems possess a considerable amount of chaos, and deducing predictive patterns from them is almost not feasible. Forecasting weather parameters has always been a major objective due to several reasons, but observably not all weather variables are equally influential. Parameters like temperature, humidity, and radiation, though easier to predict due to obvious patterns, are found to be less impactful in practical circumstances. But variables that are more turbulent and difficult to model, like wind flow, cloud cover, sea level pressure, etc. are very influential in the context of managing disasters like cyclones, estimating trades, estimating power generation and many other causes. In this study, the aim is to accurately forecast the horizontal wind velocity field which individually considers the longitudinal and latitudinal components of velocity and sea level pressure over a certain region. Initial attempts in this domain used to be purely mathematical with deterministic results, but predicting real phenomena by overpowering the chaos required enormous computational overhead. The results obtained were not satisfactory. With the evolution of statistical modeling, machine learning and deep learning, these systems were majorly formulated as spatiotemporal forecasting problems. Modified deep learning approaches like ConvLSTM, transformer-based models, CNN-GRU models, and others were applied on the spatiotemporal prediction problem producing more accurate results as compared to the primary methods. But the introduction of reservoir computing framework once again invoked the possibilities of improvements in this domain. An instance of reservoir computing, known as the echo state Network (ESN) was found to simulate and learn the complex dynamics of physical systems with significantly less effort. This boiled down to be a major motivation behind this paper. In this study, we have selected two regions over the Indian peninsula and the Mexican gulf and collected hourly data of horizontal wind velocity and sea level pressure with a spatial granularity of 0.25 degrees on these regions. This data has been used to train our proposed model which is an echo state network and convolution-based encoder–decoder architecture. The model forecasts the same parameters over the same spatial zone for some forecast horizons. The results have been rigorously compared with potential baselines where the ESN was replaced by conventional recurrent networks, and it was observed that our proposed approach outperforms all the other baselines.