Evaluating hybrid and standard deep learning models for maximum temperature forecasting in a semi-arid region

Abstract

Temperature forecasting is important for industries affected by climate, especially in semi-arid regions where the weather can change quickly and is hard to predict over time. Many studies have examined various deep learning models, including long short-term memory (LSTM), gated recurrent unit (GRU), convolutional neural networks (CNNs), and transformer-based hybrids. However, their performance in data-limited semi-arid environments is often unclear and inconsistent. This study compares six deep learning methods for predicting daily maximum temperatures in Settat, Morocco. It uses 11 years of ground-observed meteorological data. The models examined include a baseline artificial neural network (ANN) and five hybrid structures: ANN-LSTM, ANN-GRU, ANN-CNN, ANN–random forest (RF), and ANN-transformer. The results indicate that the ANN performs the best overall, with MAE = 0.0432, root mean square error (RMSE) = 0.0543, and R² = 0.8820. It surpasses all hybrid models. When using a relative improvement metric, the ANN shows accuracy gains of 32% to 42% compared to the recurrent, convolutional, and attention-based hybrids. These results suggest that in semi-arid climates, where maximum temperature mainly depends on the same-day atmospheric conditions, simpler feedforward models work better than more complex temporal models. The study underscores the need to match model complexity with climatic factors and dataset size, offering a useful benchmark for temperature forecasting in regions with limited data.