| Summary: | Micro Grids constitute a growing sector of the energetic industry, representing a paradigm shift from the central power generation plans to a more distributed generation. The capacity to work isolated from the main electric grid make the MG resilient system, capable of conducting flexible operations while providing services that make the network more competitive. Additionally, Micro Grids supply clean and efficient low-cost energy, enhance the flexible assets coordination and improve the operation and stability of the of the local electric grid, through the capability of providing a dynamic response to the energetic resources. For that, it is required an intelligent coordination which balances all the available technologies. With this, rises the need to integrate accurate and robust load and production forecasting models into the MG management platform, thus allowing a more precise coordination of the flexible resource according to the emerging demand needs. For these reasons, the HALOFMI methodology was developed, which focus on the creation of a precise 24-hour load forecast model. This methodology includes firstly, a hybrid multi-level approach for the creation and selection of features. Then, these inputs are fed to a Neural Network (Multi-Layer Perceptron) with hyper-parameters tuning. In a second phase, two ways of data operation are compared and assessed, which results in the viability of the network operating with a reduced number of training days without compromising the model's performance. Such process is attained through a sliding window application. Furthermore, the developed methodology is applied in two case studies, both with 15-minute timesteps: the first one is composed by aggregated load profiles of Standard Low Voltage clients, including production and self-consumption units. This case study presents regular and very smooth load profile curves. The second case study concerns a touristic island and represents an irregular load curve with high granularity with abrupt variations. From the attained results, it is evaluated the impact of integrating a recursive intelligent feature selection routine, followed by an assessment on the sliding window application and at last, a comparison on the errors coming from different estimators for the model, through several well-defined performance metrics.
|
|---|