Prediction of Long-Term Near-Surface Temperature Based on NA-CORDEX Output

Temperature is one of the most important parameters in climate modeling, as it has significant impacts on various geophysical processes such as evaporation and precipitation. Applying multiple climate models for prediction generally outperforms the use of individual climate models, and neural networks perform well at capturing nonlinear relationships, which can provide more reliable temperature projections. In this study, three neural network algorithms, including Multi-layer Perceptron (MLP), Time-lagged Feed-forward Neural Networks (TLFN) and Nonlinear Auto-Regressive Networks with exogenous inputs (NARX), were used to develop data-driven models for predicting daily mean near-surface temperature based on North American Coordinated Regional Downscaling Experiment (NA-CORDEX) output. A case study of Big Trout Lake in Ontario, Canada was carried out to demonstrate the applications and to evaluate the performance of the proposed neural network based models. The results showed that MLP, TLFN, and NARX performed well in generating accurate daily near-surface temperature predictions with the coefficient of determination (R2) values above 0.84. The three neural network based models had similar performance with no significant difference in terms of root mean square error and R2. Neural network based climate prediction models outperformed each of the individual regional climate models and generated smoother predictions with less fluctuation. This study provides a technical basis for generating reliable predictions of daily temperature using neural networks based model.