Predicting temperature accurately is crucial for agriculture, disaster management, and climate monitoring. This paper proposes XAI-TempSpikeDetector, a framework for temperature prediction and severe event detection that combines deep learning (DL), machine learning (ML), and time series models. Linear regression, Support Vector Machine (SVM), XGBoost, Random Forest, Gated Recurrent Unit (GRU), Long-Short Term Memory (LSTM), Bidirectional LSTM (BiLSTM), Auto Regressive Moving Average (ARIMA) and Neural Prophet are used for temperature prediction, and the results are compared. The GRU model outperformed the other approaches with the lowest mean squared error (MSE) of 0.0003, the root mean squared error (RMSE) of 0.0176, the mean absolute error (MAE) of 0.0116, and the highest R-squared error (R2) score of 0.9853, achieving the highest prediction accuracy among them. Using statistical models, ML and DL techniques, temperature spikes-extreme cold events crucial for early warning systems were detected, helping mitigate the risks associated with severe weather conditions. The type of precipitation, humidity, and wind speed are the main variables affecting temperature variations, according to a SHapley Additive Explanations (SHAP) study. The results highlight the importance of incorporating explainability strategies into predictive models, enhancing the detection of severe events while ensuring interpretability and confidence.