A novel hyper-tuned hybrid deep learning architecture for extended forecasting horizons: Application for soil moisture

This paper proposes a novel hybrid deep learning architecture in which multi-horizon forecasting is combined with hyperparameter optimization, ensuring a model structure that is both adaptive and straightforward. In the proposed framework, Gated Recurrent Units (GRU) are coupled with Temporal Convolutional Networks (TCN), while hyperparameters are tuned automatically using the Newton–Raphson-Based Optimizer (NRBO). This optimization improves generalization, reduces human bias, and speeds up convergence by efficiently finding optimal configurations. The proposed model was evaluated for 1, 7, 15, and 30-day ahead soil moisture forecasting and compared with several models. Results show that the proposed NRBO–TCN–GRU achieved the best overall performance, with RMSE values of 0.831, 1.852, 2.211, and 3.191 and corresponding R² values of 0.989, 0.921, 0.864, and 0.757, respectively. Compared with the PSO-based TCN–GRU, the proposed approach reduced RMSE by 6.1% to 15.8% while improving R² by up to 18.1%, and was nearly three times faster in running time.