La charla será dictada por Thomas Capelle, Machine Learning engineer de la empresa estadounidense Weights & Biases. Él es responsable de mantener activo y actualizado el repositorio WandB/Examples. Su experiencia es en planificación urbana, optimización combinatoria, economía del transporte y matemáticas aplicadas.
El evento incluirá una competencia de clasificación en Kaggle con premios para las mejores propuestas. Los y las participantes deben llevar sus laptops y deben tener acceso a WandB, Colab y Kaggle.Inscripción gratuita: https://forms.gle/UVEBVJy3NpQHdYB29
Due to climate change concerns, many governments have pushed for higher penetration of intermittent renewable energy sources. Among these energy sources, photovoltaic (PV) generation is one of the most sought-off, particularly by domiciliary users and small industries. However, the main drawback of this energy source is its variability and intermittency, not being available for the whole day. One way of diminishing this drawback is to use energy storage systems like batteries. In Chile, as in several other countries, the new regulation allows selling excess household generation, albeit at a price significantly lower than the consumer price. With this new setting, the residential sector user, with solar panels installed in their home, can not only use the energy from that source and connect to the electrical distribution network when required. In addition, she can also sell the excess energy generated to the distribution network, getting an economic benefit from this sale. The decision becomes complex when storage capacity, like batteries, is added to the user. In this new case, the decision process must consider when and how much to store or sell to the grid; and whether energy should be used, sent to the network, or stored in the system.
This work presents a novel approach to scheduling these storage units in a PV generation system based on stochastic optimization. A common approach to using historical data for stochastic optimization has been to use machine learning techniques to compute relevant scenarios. Instead of this “predict then optimize” strategy, we show that using a combined “predict and optimize” approach results in better recommendations. The resulting scenarios capture the relevant effects on the decision process and not just data features. We show experimental results applied to a real-life control system with limited computation capacity and further validate our results by testing the resulting schedules in an actual prototype.
This is joint work with Helena García, Tito Homem-de-Mello, Gonzalo Ruz, Francisco Jara, and Carlos Silva. This work was partially funded by FONDEF grant ID19I10158.