Optimización del almacenamiento de energía en redes eléctricas
Storage technologies are a fundamental component of electric power systems, ranging from slow, high capacity systems (such as hydro reservoirs) to fast, high ramping, low capacity systems (such as utility-scale batteries). In this thesis we deal with optimizing the operation of energy storage system...
Saved in:
| Main Author: | |
|---|---|
| Format: | masterThesis |
| Language: | Spanish |
| Published: |
2021
|
| Subjects: | |
| Online Access: | https://hdl.handle.net/20.500.12008/30024 |
| Tags: |
No Tags, Be the first to tag this record!
|
| Summary: | Storage technologies are a fundamental component of electric power systems, ranging from slow, high capacity systems (such as hydro reservoirs) to fast, high ramping, low capacity systems (such as utility-scale batteries). In this thesis we deal with optimizing the operation of energy storage systems from a broad perspective, focusing on both the large- and low-scale systems. Operating a grid with storage is a difficult task, in the sense that i) there is inherent uncertainty from the stochastic variables involved (such as the demand and renewable energy available), and ii) storage dynamics couple decisions across time, implying that actions must be taken with respect to some global goal. Accordingly, we formulate the optimal dispatch problem as a multi-stage dynamic programming problem, subject to various control and state constraints. We study both these cases and consider their applications on the Uruguayan grid. In the case of hydro-reservoirs, we model the cost-to-go functions as convexquadratic in the reservoirs. This leads to an approximate dynamic programming algorithm which at each stage samples state-cost pairs and fits convex-quadratic functions in a recursive manner. We implement this efficiently via modern optimization solvers, and our results show that the control policy learned in this fashion exceeds the performance of a naïve myopic policy. We also consider the operation of a bulk battery storage system in a single-bus model of the Uruguayan grid. In this regard, we consider learning the controller via Q-learning, the quintessential algorithm in the field of Reinforcement Learning. With no prior information on the transition model and on the stochastic variables involved, we obtain an agent that makes hourly decisions based on the state of the system, namely the state of charge of the battery, the time of day and the forecasted wind and demand. We train the controller with real data of three winters, and obtain a policy that operates the system with good performance, charging the battery —even at expenses of fuel generation— when generation is cheap and renewable energies abundant, and turning that surplus back to the grid when demand peaks. |
|---|