Tesis (Ingeniería de Sistemas y Automática)

URI permanente para esta colecciónhttps://hdl.handle.net/11441/11345

Examinar

Examinando Tesis (Ingeniería de Sistemas y Automática) por Autor "Aguilar Guisado, Juan"

Seleccione resultados tecleando las primeras letras
Mostrando 1 - 1 de 1
  • Miniatura
    Acceso AbiertoTesis Doctoral
    Stochastic Model Predictive Control and Machine Learning for the Participation of Virtual Power Plants in Simultaneous Energy Markets
    (2022-07-07) Aguilar Guisado, Juan; Arce Rubio, Alicia; Bordons Alba, Carlos; Universidad de Sevilla. Departamento de Ingeniería de Sistemas y Automática
    The emergence of distributed energy resources in the electricity system involves new scenarios in which domestic consumers (end-users) can be aggregated to participate in energy markets, acting as prosumers. Every prosumer is considered to work as an individual energy node, which has its own renewable generation source, its controllable and non-controllable energy loads, or even its own individual tariffs to trade. The nodes can build aggregations which are managed by a system operator. The participation in energy markets is not trivial for individual prosumers due to different aspects such as the technical requirements which must be satisfied, or the need to trade with a minimum volume of energy. These requirements can be solved by the definition of aggregated participations. In this context, the aggregators handle the difficult task of coordinating and stabilizing the prosumers' operations, not only at an individual level, but also at a system level, so that the set of energy nodes behaves as a single entity with respect to the market. The system operators can act as a trading-distributing company, or only as a trading one. For this reason, the optimization model must consider not only aggregated tariffs, but also individual tariffs to allow individual billing for each energy node. The energy node must have the required technical and legal competences, as well as the necessary equipment to manage their participation in energy markets or to delegate it to the system operator. This aggregation, according to business rules and not only to physical locations, is known as virtual power plant. The optimization of the aggregated participation in the different energy markets requires the introduction of the concept of dynamic storage virtualization. Therefore, every energy node in the system under study will have a battery installed to store excess energy. This dynamic virtualization defines logical partitions in the storage system to allow its use for different purposes. As an example, two different partitions can be defined: one for the aggregated participation in the day-ahead market, and the other one for the demand-response program. There are several criteria which must be considered when defining the participation strategy. A risky strategy will report more benefits in terms of trading; however, this strategy will also be more likely to get penalties for not meeting the contract due to uncertainties or operation errors. On the other hand, a conservative strategy would result worse economically in terms of trading, but it will reduce these potential penalties. The inclusion of dynamic intent profiles allows to set risky bids when there exist a potential low error of forecast in terms of generation, load or failures; and conservative bids otherwise. The system operator is the agent who decides how much energy will be reserved to trade, how much to energy node self consumption, how much to demand-response program participation etc. The large number of variables and states makes this problem too complex to be solved by classical methods, especially considering the fact that slight differences in wrong decisions would imply important economic issues in the short term. The concept of dynamic storage virtualization has been studied and implemented to allow the simultaneous participation in multiple energy markets. The simultaneous participations can be optimized considering the objective of potential profits, potential risks or even a combination of both considering more advanced criteria related to the system operator's know-how. Day-ahead bidding algorithms, demand-response program participation optimization and a penalty-reduction operation control algorithm have been developed. A stochastic layer has been defined and implemented to improve the robustness inherent to forecast-dependent systems. This layer has been developed with chance-constraints, which includes the possibility of combining an intelligent agent based on a encoder-decoder arquitecture built with neural networks composed of gated recurrent units. The formulation and the implementation allow a total decouplement among all the algorithms without any dependency among them. Nevertheless, they are completely engaged because the individual execution of each one considers both the current scenario and the selected strategy. This makes possible a wider and better context definition and a more real and accurate situation awareness. In addition to the relevant simulation runs, the platform has also been tested on a real system composed of 40 energy nodes during one year in the German island of Borkum. This experience allowed the extraction of very satisfactory conclusions about the deployment of the platform in real environments.