Optimizing soil temperature in the root zone is essential for crop health and productivity, particularly in India, where a huge population relies on agriculture for their livelihood. This study proposes a deep learning ensemble model combining Temporal Convolutional Networks (TCNs) and Artificial Neural Networks (ANNs) to forecast soil heat flux, which is a key determinant of Root Zone Temperature (RZT). The proposed model is trained on an extensive historical soil heat flux dataset collected by UAF Pakistan over one year. TCNs effectively capture temporal dependencies in soil data, while ANNs model complex feature interactions. The predictions are fed into a Bayesian optimization framework that targets a near-ideal RZT of 25 \(^\circ \) C. The TCN-ANN ensemble achieved an MAE of 0.1559 and an R-squared error of 0.945 at 50 epochs, outperforming individual models such as TCN, ANN and LSTM. Post-optimization, the RZT stabilized within the range of 18–25 \(^\circ \) C, achieving accuracy 99.44%, within a threshold ±3.45 \(^\circ \) C.