The papers are listed from newest to oldest. Each paper is tagged with the directions it belongs to: A physical dynamics, B neural solvers, and C Model2Action. A paper can belong to more than one.
2026
A graph memory stores local LiDAR geometry and provides waypoint-specific context at each step of recurrent diffusion denoising. The generated trajectory is evaluated against the geometry it moves through.
2025
A probabilistic electricity-cost forecast is trained through the stochastic water-supply schedule it induces. The SPTO+ variant uses a differentiable convex upper bound on the optimality gap as the loss.
2024
Lag-penalty attention and a masked weighted Markov chain reduce the delayed response of a learned demand model near turning points.
The Coder models the pressure response to measured variables. The Critic models the residual variation caused by unmeasured disturbances.
A DCLS-Net predicts the energy of each rolling pass. Series concatenation links the passes so that the error stops accumulating along the sequence.
Nonlinear model predictive control initializes the controller weights. A control barrier function with a nonquadratic loss enforces the state and input constraints during learning.
2023
CEEMDAN reduces sensor noise, pump-state signals add operating context, and gradient correction extends the horizon of the recurrent pressure model.
Eligibility traces at several time scales give the critic credit signals accumulated over different horizons.
2021
A daily demand forecast is decomposed into reusable fine-grained patterns, which reconstruct an approximately 300-step minute-level profile.
A weighted-voting ensemble of deep classifiers locates leakage events in a simulated water-distribution network.