The l1-penalizing method introduced in "Learning Multiple Quantiles With Neural Networks" provides non-crossing quantiles estimates of response variable for given explanatory data.
@article{moon2021learning, title={Learning multiple quantiles with neural networks}, author={Moon, Sang Jun and Jeon, Jong-June and Lee, Jason Sang Hun and Kim, Yongdai}, journal={Journal of Computational and Graphical Statistics}, volume={30}, number={4}, pages={1238--1248}, year={2021}, publisher={Taylor \& Francis} }