Contributions to Meteorological Forecasting: A Comparative Study of Machine Learning-Driven Temperature Predictions at Menara Station in Morocco

Abstract

The present research study conducts an evaluation of some machine learning and deep learning models to forecast temperature at the Menara station in Morocco. Four different machine learning models; Prophet, XGBoost, LSTM, and BiL-STM; were assessed to identify the most effective approach for temperature forecasting. For each model, we proceed by fine-tuning to optimize its performance. The models efficacy was assessed using five different partitions of training and testing datasets for the test sizes of 40%, 30%, 20%, 10%, and 5%. The comparative analysis was based on selected metrics commonly used in machine learning, including MAE, MSE, RMSE, and R2, across both training and testing phases. The investigation reveals a significant performance disparity, with XGBoost, LSTM, and BiLSTM outperform Prophet in terms of accuracy. XGBoost shows better accuracy in learning from data training with the lowest MAE values of 2.277, 2.324, and 2.267 for the 40%, 30%, and 10% test sizes respectively. On the other hand, BiLSTM excels in generalizing the forecasting capabilities across different data partitions, providing the best MAE outcomes of 2.299, 2.265, and 2.263 for 30%, 20%, and 10% test sizes respectively. LSTM was relaitvely close to BiLSTM in terms of generalizing the forecasting capabilities. The study recomands BiLSTM with its effective utilization of bidirectional temporal dynamics for enhanced prediction accuracy as the optimal model for reliable and generalizable temperature predictions at Menara station.