Uncertainty and sensitivity analysis of deep learning models for diurnal temperature range (DTR) forecasting over five Indian cities.

In this article, the maximum and minimum daily temperature data for Indian cities were tested, together with the predicted diurnal temperature range (DTR) for monthly time horizons. RClimDex, a user interface for extreme computing indices, was used to advance the estimation because it allowed for statistical analysis and comparison of climatological elements such time series, means, extremes, and trends. During these 69 years, a more erratic DTR trend was seen in the research area. This study investigates the suitability of three deep neural networks for one-step-ahead DTR time series (DTRTS) forecasting, including recurrent neural network (RNN), long short-term memory (LSTM), gated recurrent unit (GRU), and auto-regressive integrated moving average exogenous (ARIMAX). To evaluate the effectiveness of models in the testing set, six statistical error indicators, including root mean square error (RMSE), mean absolute error (MAE), coefficient of correlation (R), percent bias (PBIAS), modified index of agreement (md), and relative index of agreement (rd), were chosen. The Wilson score approach was used to do a quantitative uncertainty analysis on the prediction error to forecast the outcome DTR. The findings show that the LSTM outperforms the other models in terms of its capacity to forget, remember, and update information. It is more accurate on datasets with longer sequences and displays noticeably more volatility throughout its gradient descent. The results of a sensitivity analysis on the LSTM model, which used RMSE values as an output and took into account different look-back periods, showed that the amount of history used to fit a time series forecast model had a direct impact on the model's performance. As a result, this model can be applied as a fresh, trustworthy deep learning method for DTRTS forecasting.