Recursive Aggregation of Liquid Neural Networks for Wind Power Prediction in Poland

Abstract

This study presents an application of Liquid Neural Networks (LNNs) for medium-term wind power forecasting in Poland, integrating meteorological variables derived from Numerical Weather Prediction (NWP) models. This paper proposes a neural network consisting of two LNN architectures. Each LNN models a different dynamic range for wind generation prediction in Poland. One- and two-layer LNN architectures were analyzed for their predictive accuracy in wind generation forecasting. The input dataset included wind speed, direction, gust intensity, atmospheric pressure, temperature, and directional change, collected from the Polish Transmission System Operator (TSO) and NWP data. The study further introduced an adaptive prediction fusion framework based on the Recursive Least Squares (RLS) algorithm, which dynamically weights model outputs to minimize error. The proposed method for incorporating weights into LNN outputs is characterized by lower prediction error, resulting in greater accuracy in wind generation predictions in Poland. A comparative analysis of deep learning techniques with exogenous inputs confirmed that the proposed RLS-LNN-based fusion model achieved the highest accuracy on wind data for Poland developed by the authors over the last 5 years.