No abstract

In the Paris subway system, stations represent about one third of the overall energy consumption. Within stations, ventilation is among the top consuming devices; it is operated at maximum airflow all day long, for air quality reasons. In this paper, we present a concept of energy system that displays comparable air quality while consuming much less energy. The system comprises a battery that makes it possible to recover the trains braking energy, arriving under the form of erratic and strong peaks. We propose an energy management system (EMS) that, at short time scale, controls energy flows and ventilation airflow. By using proper optimization algorithms, we manage to match supply with demand, while minimizing energy daily costs. For this purpose, we have designed algorithms that take into account the braking variability. They are based on the so-called Stochastic Dynamic Programming (SDP) mathematical framework. We fairly compare SDP based algorithms with the widespread Model Predictive Control (MPC) ones. First, both SDP and MPC yield energy/money operating savings of the order of one third, compared to the current management without battery (our figure does not include the cost of the battery). Second, depending on the specific design, we observe that SDP outperforms MPC by a few percent, with an easier online numerical implementation.

No abstract

Multistage stochastic optimization problems are, by essence, complex because their solutions are indexed both by stages (time) and by uncertainties (scenarios). Their large scale nature makes decomposition methods appealing. The most common approaches are time decomposition -and statebased resolution methods, like stochastic dynamic programming, in stochastic optimal control -and scenario decomposition -like progressive hedging in stochastic programming. We present a method to decompose multistage stochastic optimization problems by time blocks, which covers both stochastic programming and stochastic dynamic programming. Once established a dynamic programming equation with value functions defined on the history space (a history is a sequence of uncertainties and controls), we provide conditions to reduce the history using a compressed "state" variable. This reduction is done by time blocks, that is, at stages that are not necessarily all the original unit stages, and we prove a reduced dynamic programming equation. Then, we apply the reduction method by time blocks to two time-scales stochastic optimization problems and to a novel class of so-called decision-hazard-decision problems, arising in many practical situations, like in stock management. The time blocks decomposition scheme is as follows: we use dynamic programming at slow time scale where the slow time scale noises are supposed to be stagewise independent, and we produce slow time scale Bellman functions; then, we use stochastic programming at short time scale, within two consecutive slow time steps, with the final short time scale cost given by the slow time scale Bellman functions, and without assuming stagewise independence for the short time scale noises.

No abstract