March 2022
1st of March - Hybrid Format
Speaker : Olga Iufevera (PhD student with Aneel Tanwani)
Title : Suboptimal Filters with Poisson-Sampled Measurements
Abstract : We consider continuous-discrete filters, i.e. a class of systems with an unobserved continuous-time stochastic process and discrete-time observations. The observations are assumed to arrive randomly in time under a Poisson distribution. We design a suboptimal estimator with a reset map which updates whenever a new measurement is received. We provide error bounds on the covariance of the estimation error which depend on the sampling rate.
Title : Suboptimal Filters with Poisson-Sampled Measurements
Abstract : We consider continuous-discrete filters, i.e. a class of systems with an unobserved continuous-time stochastic process and discrete-time observations. The observations are assumed to arrive randomly in time under a Poisson distribution. We design a suboptimal estimator with a reset map which updates whenever a new measurement is received. We provide error bounds on the covariance of the estimation error which depend on the sampling rate.
Speaker : Adrien Le Franc (Post Doc)
Title : Parametric stochastic optimization for day-ahead and intraday co-management of a power unit
Abstract : We consider renewable power units equipped with a battery and engaged in day-ahead load scheduling. In this context, the unit manager must submit a day-ahead power production profile prior to every operating day, and is engaged to deliver power accordingly. During the operating day, the unit manager is charged penalties if the delivered power differs from the submitted profile. First, we model the problem of computing the optimal production profile as a parametric multistage stochastic optimization problem. The production profile is modeled as a parameter which affects the value of the intra-day management of the power unit, where the photovoltaic production induces stochasticity. Second, we introduce parametric value functions for solving the problem. Under convexity and differentiability assumptions, we are able to compute the gradients of these value functions with respect to the parameter. When the differentiability assumption breaks, we propose two approximation methods. One is based on a smooth approximation with the Moreau envelope, the other one is based on a polyhedral approximation with the SDDP algorithm. We showcase applications in the context of the French non-interconnected power grid and benchmark our method against a Model Predictive Control approach.
Title : Parametric stochastic optimization for day-ahead and intraday co-management of a power unit
Abstract : We consider renewable power units equipped with a battery and engaged in day-ahead load scheduling. In this context, the unit manager must submit a day-ahead power production profile prior to every operating day, and is engaged to deliver power accordingly. During the operating day, the unit manager is charged penalties if the delivered power differs from the submitted profile. First, we model the problem of computing the optimal production profile as a parametric multistage stochastic optimization problem. The production profile is modeled as a parameter which affects the value of the intra-day management of the power unit, where the photovoltaic production induces stochasticity. Second, we introduce parametric value functions for solving the problem. Under convexity and differentiability assumptions, we are able to compute the gradients of these value functions with respect to the parameter. When the differentiability assumption breaks, we propose two approximation methods. One is based on a smooth approximation with the Moreau envelope, the other one is based on a polyhedral approximation with the SDDP algorithm. We showcase applications in the context of the French non-interconnected power grid and benchmark our method against a Model Predictive Control approach.