One-Hour-Ahead Mean Radiant Temperature Forecasting in Jabodetabek Using CNN-LSTM and Temporal Convolutional Networks

Abstract

Mean Radiant Temperature (MRT) represents the combined shortwave and longwave radiant load experienced by a human body, but continuous observations are rarely available across large metropolitan areas. This study developed and compared a Convolutional Neural Network–Long Short-Term Memory (CNN-LSTM) model and a Temporal Convolutional Network (TCN) for one-hour-ahead MRT forecasting in Jabodetabek. The dataset comprised 210,528 hourly records from 2023–2024 at 12 ERA5 grid points. Five radiation variables and two near-surface thermal variables were used as predictors, while ERA5-HEAT MRT was the target. Each sample contained a 24-hour historical window. Three chronological train-validation-test splits were evaluated: 80:10:10, 70:10:20, and 70:15:15. Both architectures achieved R² values above 0.93 in all scenarios. Under the 80:10:10 split, TCN produced the lowest RMSE value of 2.8104 °C, the highest R² value of 0.9518, the lowest validation loss value of 5.9592, and a residual bias of −0.4863 °C. CNN-LSTM achieved the lowest MAE value of 1.8874 °C and MAPE value of 5.76% and was more stable when the training proportion decreased. Overall, TCN 80:10:10 was selected as the best configuration, although field validation is required before operational deployment.