Week 16.5: Challenge 2 Final Solution

The final solution - M[AI]nager - consists in a project that uses both machine learning and optimization algorithms to manage the energy of a community and help the energy aggregators make better energetic and financial decisions not only for themselves but also for all the members of the community.

The teams' final thoughts

Bruno Ribeiro - Machine learning and optimisation algorithms are very efficient and powerful tools. Together they can make robust systems, capable of solving very unique problems in efficient and modern manner.

Bruno Veiga - The community energy prediction along side with the use of the the particle swarm optimization constitute a unique way of managing the community energy very early and with very positive results saving money not only to the energy aggregator but also to all members of the community.

Carlos Coelho - For the integration of renewable energies in the electric grid, it is crucial that the grid is able to deal with its variance and intermittency. The first step is being able to predict the energy generation and consumption, so that the community manager can plan ahead. Our solution, M[AI]nager, is able to both forecast the hourly energy balance of a community as well as establishing whole-day action plan, that minimizes the costs of buying energy from the wholesale market. Although it is still mostly a proof of concept, we are very confident that at the very least, its solar generation forecasting capabilities can be applied in the real world.

Carlos Veiga - For the use of machine learning in real situations, knowledge of machine learning techniques and an adequate volume of data for learning is necessary. With these elements, a useful and efficient system can be created for the customer.

Miguel Silva - It was an enlightening challenge as we combined the use of machine learning techniques with optimization algorithms. With the goal to achieve a community optimization, in terms of energy expenses and consumptions, we successfully handled an untreated dataset, to make predictions for a community consumption and solar energy generation, reaching an optimization solution to propose to a community energy manager.

The architecture

This time, every module besides the Web Interface used the same technology, allowing for a more compact setup:

• Web Interface: User Interface that allows the user to easily utilize and interpret the developed forecast and planning systems.

• Forecast model creation script: Jupyter Notebook implementation that was used to study and develop the forecast models.

• Main Server: Serves as a mediator that facilitates the access to the forecast and planning modules.

Forecast Module: Module that uses the forecast Models to generate predictions for the consumption or generation of energy in a given period.

Planning Module: Module that uses forecasted energy consumption and generation and creates the best course of action, in order to minimize the total monitary cost to the community.

The M[AI]nager

Most of the Pl[AI]tform remains similar to the one presented in Week 8.5. Nonetheless, here's a video showing a quick rundown of the new funcionalities:

It is worth noting that in the planning, one of the most important parameters is the financial balance. As such, the system tends to sell all the energy at the end of the prediction period, in order to maximize it - we'll call that phenomenon sale rush. As such, the system always plans 12h above requested, and only presents the results for the requested period. This means that its results are unaffected by the sale rush period.