Accurate forecasting of temperature extremes is critical for climate-sensitive sectors such as energy, agriculture, and public health. This study proposes a hybrid Deep Learning (DL) framework based on Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (BiLSTM) networks for predicting daily maximum and minimum temperature deviations ( \(\:{DT}_{max}\) and \(\:{DT}_{min}\) ), across three stations in southern Morocco. We rigorously evaluated the model’s performance against a suite of conventional Machine Learning (ML) and DL benchmarks, including Linear Regression (LR), Random Forest (RF), Extreme Gradient Boosting (XGBoost), Support Vector Regression (SVR), Multi-Layer Perceptron (MLP), CNN, and LSTM, using multiple evaluation metrics. The proposed CNN–BiLSTM model demonstrated superior performance across all study sites. At Errachidia, it achieved the lowest Root Mean Square Error (RMSE) for both \(\:{DT}_{max}\) and \(\:{DT}_{min}\) , reducing the DTmin RMSE by \(\:41\%\) compared to the standalone CNN model. At Midelt, a region with high climatic variability, the hybrid model maintained the lowest errors and robustly achieved Nash–Sutcliffe Efficiency (NSE) values above \(\:0.89\) . Similar trends were observed at Ouarzazate, where the model delivered the lowest RMSE and the highest Pearson correlation coefficient ( \(\:R=0.97\) ), while improving the \(\:{DT}_{min}\) RMSE by \(\:54\%\) over the CNN baseline. These findings highlight the effectiveness of combining convolutional feature extraction with bidirectional temporal modeling to significantly enhance predictive accuracy. The proposed framework outperforms conventional methods and offers a reliable tool for improving temperature forecasting and supporting risk management strategies in vulnerable regions.
