Multi-horizon Irradiation Forecasting for Mediterranean Locations Using Time Series Models

Abstract

As fossil fuels combustion poses a real public health problem, PV and wind energy sources seem good alternatives. The main advantage is the renewable and inexhaustible aspects and the main disadvantages are related to their intermittencies. This paper deals with a solution to solve this problem: the forecasting of the renewable energy sources and more precisely the forecasting of solar irradiation. Several methods have been developed by experts and can be divided in two main groups: (i) methods using mathematical formalism of Times Series (TS) and (ii) Numerical Weather Prediction (NWP) models. Depending on the horizon of prediction or by the spatial resolution to be considered some of these methods are more effective compared to others. In this work we focus on the grid manager's point of view interested by four horizons: d+1; h+24, h+1 and m+5. Thus we tested different time series forecasting models for Mediterranean locations in order to prioritize different predictors. For the d+1 horizon, we conclude to use an approach based on neural network being careful to make stationary the time series, and to use exogenous variables. For the h+1 horizon, a hybrid methodology combining the robustness of the autoregressive models and the non-linearity of the connectionist models provides satisfactory results. For the h+24 case, neural networks with multiple outputs give very good results. For m+5 horizon, even if neural networks are the most effective, the simplicity and the relatively good results shown by the persistence-based approach, lead us to recommend it. All the proposed methodologies and results are complementary to the prediction studies available in the literature. We can also conclude that the methodologies developed could be included as prediction tools in the global command control systems of energy sources.