Adapted symbolic dynamic networks for multi-step forecasting of chaotic wind speed time series

Accurate short-term wind speed forecasting is crucial for efficient wind power generation and distribution. However, many existing prediction models often falter due to the inherently nonlinear and chaotic nature of wind dynamics, apart from being computationally intensive. This paper presents a multi-step wind speed forecasting method based on a complex network constructed from wind speed time series. The proposed framework builds upon a recently developed complex network-based approach with two key modifications specifically suitable for wind dynamics. We introduce an adaptive time series discretisation using k-centroid clustering, replacing the original uniform binning strategy, and a pattern matching scheme via Levenshtein string similarity instead of the Euclidean distance-based matching approach. These refinements adapt to the heavy-tailed distribution of wind speeds and offer a computationally efficient and robust approach to pattern matching that enhances prediction accuracy. Applied to wind speed data from multiple locations, our model outperforms conventional nonlinear and machine learning methods in terms of several error metrics. The model also exhibits strong spatial generalisability, delivering reasonably accurate predictions across 40 nearby locations, chosen for their attractor similarity, using a model built on data from a single reference site. This work presents one of the first successful demonstrations of a complex network-based multi-step forecasting method applied to real-world wind speed prediction.