WA10 - Stochastic and Robust Optimization 3

We propose an event-driven re-optimization framework for multi-commodity flow with stochastic demand. A dual-inspired trigger controls update frequency, trading off optimality and computation. Experiments reveal near-optimal performance at small thresholds, monotone regret behavior, and structured inter-event times that admit simple statistical characterization.

Neural network surrogates have emerged as a promising alternative for accelerating electromagnetic transients simulations of the power electronic converters used to interface renewable sources of energy, such as solar and wind, to the electric power grid. However, their sensitivity to input perturbations, arising from measurement noise, control errors, and variability in operating conditions, poses a major barrier to their use in safety-critical contexts. In this work, we propose a robustness certification framework for neural network-based models of power electronic converters, focusing on physics-informed temporal convolutional networks (TCNs) modelling boost DC–DC converter dynamics. The proposed architecture leverages dilated causal convolutions with ReLU activations, enabling compatibility with linear relaxation-based verification techniques. In particular, we employ the optimization-based CROWN bound propagation via the auto_LiRPA framework to compute certified output bounds under bounded input perturbations using linear relaxations of the activation functions. The feedforward structure of TCNs, combined with smoothing and Lipschitz reglarization strategies, yields tight and numerically stable bounds while preserving high predictive accuracy. The verification is performed on sequential data using a chunked formulation, allowing scalable certification. Experimental results on a large dataset of converter operating points demonstrate that the certified bounds consistently enclose the model predictions and remain sufficiently tight to be practically meaningful. For representative perturbation levels, the average bound width remains small relative to the signal magnitude, even during transient regimes, indicating that the surrogate model maintains robust behaviour under uncertainty. These results highlight the effectiveness of combining architectural design and verification-aware training to improve certifiability.

This study introduces a variant of the multi-period vehicle routing problem in which collections may occur between release and due dates. The release date is defined as the time that the products become available for collection on farms. The release date for each product is stochastic, and farm products are collected over a one-week planning horizon. The due date for each collection is time-dependent, specified by the farmer, and late collections are not allowed. A fixed number of single-compartment homogeneous vehicles is booked throughout the week. Although the internal fleet size remains fixed at the weekly level, daily routes and farm assignments can be adjusted as availability information is revealed. If realized service requirements exceed the capacity of the committed internal fleet, additional transportation can be procured from a third-party logistics provider (3PL) at a higher marginal cost. The objective is to minimize the total collection cost, which includes transportation and expected 3PL costs. We propose a decomposition-based solution approach to solve the resulting two-stage stochastic vehicle routing problem with outsourcing.