Modelos híbridos estocástico - matemático para previsão de velocidade do vento

Abstract

In the last 10 years, it is possible to identify greater insertions of renewable energy resources for electricity generation in the national energy matrix, such as wind generation. This can be justified in several aspects, for example, in relation to the wind power source; it has played an important role in the Brazilian matrix by providing a financially viable alternative to the main electricity generator in the country that is the hydraulic source. Another point in favor of the wind generation is the environmental issue, the use of this resource must be considered in this aspect, in addition in the Northeast of Brazil there is the so-called complementarity, this when the river flows to generate energy are low, second half of the year, in much of it according to climatology the winds are more intense. The amount of work on wind generation in the country grows each day bringing benefits to the particular sector providing guarantees of exploitation of the local winds. In this sense, the objective of this work is to present two innovative hybrid models, which can help in the wind sector, by being able to predict wind velocities with good accuracy based on validation measures of the models. The models were elaborated from the mathematical combinations of two classic time-series models (Integrated Automatic Regressive of Moving Averages and Exogenous inputs (ARIMAX) and Exponen Tial Exponential Smoothing with Holt-Winters - HW) with an artificial intelligence model (Artificial Neural Network (RNA)). In the comparisons between the time series (observed and adjusted in terms of monthly and hourly averages) it is possible to identify, for example, Nash-Sutcliffe (NS) efficiency value of approximately 98%, and also percentage error value around 4.5%, which according to the literature (as will be shown) confirm the good accuracy of the models. A large differential of the proposed hybrid models, when compared to other traditional ones is that they can incorporate both linear and nonlinear characteristics, which are often found in time series, being this important condition to provide greater precision of the velocities of the predicted wind, thus providing greater reductions in statistical error measures (for example, in some cases of the order of 50%) when compared with the classical models that compose them. The proposed hybrid models may represent important tools for decision-makers in the wind generation sector in terms of wind exploitation; however, other areas of interest can be analyzed for their feasibility.