Técnicas de Inteligência Artificial Aplicadas ao Controlo Preditivo de Baterias Estacionárias

• Autor(a) principal: Ricardo Emanuel Gomes Fernandes da Silva
• Data de Publicação: 2018
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: https://hdl.handle.net/10216/114216
• topic: Engenharia electrotécnica, electrónica e informática; Electrical engineering, Electronic engineering, Information engineering

The expected trend of decreasing feed-in tariffs in the upcoming years meets the current necessity to secure a more sustainable and autonomous electric power grid, capable of integrating more renewable energy resources. This trend turns self-consumption particularly relevant for prosumers (consumers that own small distributed generation units), namely at the household and small industry levels. With the growing advances in stationary storage technologies, reflected utmost at their economic viability, stationary batteries along with electric vehicles are viewed as one of the best solutions to maximize such self-consumption levels of prosumers. Today's storage controllers, used on the management of charging and discharging these batteries present a reactive and immediate response. It can although be more interesting, for a prosumer, that such controllers could present a more predictive action, i.e., capable of understanding how consumption and production profiles will evolve, in order to maximize the self-consumption. If we also consider the prosumer to be involved in a market dynamic pricing scheme, the controller should also behave opportunistically, taking into account the market prices so that the energy requirements made to grid would be deviated to time windows were prices were cheaper. This problem can be mathematically framed on the definitions of multi-objective and multi-temporal. Associating the elevated number of state variables and the error possibilities inherent to the data's forecasting nature makes this problem extremely complex, narrowing its resolution to techniques based on artificial intelligence. In the present work, the capability of artificial intelligence techniques in predictively controlling stationary storage when coupled with photovoltaic generation units, is evaluated. Namely it is used the Proximal Policy Gradient method, made available by OpenAI and inserted in the category of Deep Reinforcement Learning which combine neural networks with the training of artificial intelligence agents through Reinforcement Learning. The comparison with genetic algorithms is made in order to infer the viability of this methodology in the resolutions of the problem at hand.