A comparative study of classical and deep learning models for wind speed forecasting

This study examined the performance of several time series models (both classical models like ARMA, ARIMA and deep learning models such as LSTM, CNN) are compared in terms of both classical and deep learning models for predicting wind speed based on two meteorological datasets. The literature review presented strengths and weaknesses for each approach, and the influence of data scale, resolution, and quality on the forecasting results. The quantitative result demonstrated that for the high-frequency data LSTM reduces the MSE ~99.3% comparing with ARIMA, when ARIMA is good for short-term prediction and small datasets. Then, the models were used for a small Turkish dataset, and the proposed curves were tested against the large group of Jena database in Germany. Results indicated that while classical models are better suited for forecasting over the short term with limited data, deep learning methods require substantial, high-quality data to surpass traditional forecasting methods. These findings provide valuable recommendations for renewable energy planners on model selection with respect to the characteristics of datasets and their operational requirements.