Comparative Analysis of LSTM Architectures for Wind Speed Forecasting: A Case Study in Muş, Turkey

Abstract

This study assesses the effectiveness of five distinct Long Short-Term Memory (LSTM) architectures for forecasting wind speed in Muş, Turkey. The models include Vanilla LSTM, Stacked LSTM, Bidirectional LSTM, Attention LSTM, and Residual LSTM. The data, obtained from the Muş Meteorological Office, underwent preprocessing to handle missing values by averaging the same day and month values between 1969 and 2023. The dataset, containing 20,088 daily wind speed measurements, was split into training and test sets, with 80% allocated for training and 20% for testing. Each model was trained over 100 epochs with a batch size of 32, and performance was assessed using Mean Squared Error (MSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). The Vanilla LSTM model showed the lowest MSE and MAE values, indicating superior overall performance, while the Attention LSTM model achieved the lowest MAPE, demonstrating better percentage accuracy. These findings indicate that the Vanilla and Attention LSTM models are the most effective for wind speed forecasting, with the choice between them depending on the prioritization of total error versus percentage error.