Illuminance prediction through extreme learning machines

Planning, managing, and operating power grids using mixed traditional and renewable energy sources requires a reliable forecasting of the contribution of the renewable sources, due to their variable nature. Besides, the short-term prediction of the climatic conditions finds application in other fields (e.g., Climate Sensitive Buildings). In particular, this work is related to the solar radiation forecasting, that affects the photovoltaic production. The variability of the weather phenomena and climate features make the prediction a difficult task. In fact, the amount of solar radiation that reaches a particular geographical location depends not only by its latitude, but also by the geographical characteristics of the region that can create local climate conditions. In order to capture such variability, the data collected in the past can be used. Several sources can provide the data needed for the prediction (satellite and ground images, numerical weather predictions, ground measurement stations) with different resolution in time and space. In this paper, a new learning paradigm, the Extreme Learning Machine, is used to train a neural network model for the prediction of the solar illuminance. The neural networks are challenged on a two-year ground solar illuminance dataset measured in Milan, and the results are compared with those of simple predictors and results in literature.