Explainable AI-Driven Dew Point Forecasting With Attention-Based Temporal Convolutional Networks

Abstract

Climate change poses global challenges, including rising temperatures and extreme weather, severely affecting regions such as Bangladesh. Predictions of dew point temperature, a key indicator for climate change, are vital for climate adaptation, agriculture, energy planning, weather forecasting, and public health. This research analyzed dew point temperature data from 35 weather stations across Bangladesh to find patterns and trends over four decades since 1980. This study developed an attention-based Temporal Convolutional Networks (TCN) model for predicting dew point temperature. The model was enhanced with Explainable AI (XAI) techniques to facilitate feature selection and provide interpretable insights into the factors influencing the predictions. In order to assess the performance of the proposed model this work utilized different performance measures and achieved test loss of 0.002, validation loss of 0.00212, Mean Absolute Error (MAE) of 0.032, and Root Mean Squared Error (RMSE) of 0.044, R2 and adjusted R2 values of 0.938, demonstrating higher prediction accuracy compared to other models such as Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN), Bi-directional Long Short-Term Memory (BiLSTM), LSTM_CNN hybrid, and Gated Recurrent Unit (GRU). Consistent performance across datasets from other weather stations confirmed the robustness of the suggested model. The findings provide valuable insights into dew point trends and contribute a reliable predictive model for practical applications in climate-sensitive domains.