WIND SPEED FORECASTING IN KAZAKHSTAN USING DEEP LEARNING

Abstract

Accurate wind speed forecasting is crucial for grid stability and for renewable energy sources optimization. In this study we will be evaluating performance of 5 models: 1D Convolutional Neural Network (1D CNN), Long-Short Term Memory (LSTM), WaveNet, Transformer, and Swin Transformer for Time Series (Swin4TS); which was used/adapted for wind speed forecasting across different regions of the Kazakhstan. We utilized 3 different datasets: NASA, KazHydroMet, and one we collected ourselves using the Yurt facility at the territory of Nazarbayev University. Models' evaluation is set across different forecast horizons: ultra-short-term (ULTF), short-term (STF), medium-term (MTF), and long-term (LTF). Our results showed that there is no model that universally superior for every dataset and every forecast horizon. LSTM showed excellent performance for USTF oncomplex and cyclic data. WaveNet showed an ability to adapt for long-term forecasting patterns in the challenging NASA dataset. In the KazHydroMet dataset's domain the Transformer model was proven to be the top performer. 1D CNN showed a baseline consisted mid-tier performance across all datasets. The novel architecture Swin4TS in general showed poor results, but managed to narrow the gap in LFT and was fairly comparable with other models.