On the use of data-driven techniques to solve the microgrid sizing problem

Different optimization methods can be used for the optimal sizing of microgrids but all face a trade-off in terms of level of detail and computation effort. This paper aims at demonstrating that using supervised learning algorithms, it is possible to correctly approximate almost instantly the optimal sizing while keeping a sufficient degree of detail.
To this end, a methodology for optimal microgrid sizing is applied to a wide range of case studies in different locations around the globe. The generated results, grouped in a database, are then used to train a supervised learning method that can approximate the outcome of the sizing optimization problem in almost no time, allowing for quasi-instantaneous pre-designs, with an average relative error of 14%. The resulting approximation can then be used directly or to warm-start the solving of the sizing problem.