SOLAR RADIATION FORECAST USING ARTIFICAL NEURAL NETWORKS IN SOUTH BRAZIL Ricardo A. Guarnieri *, Enio B. Pereira, Sin Chan Chou Instituto Nacional de Pesquisas Espaciais (INPE), São José dos Campos, São Paulo, Brazil ABSTRACT Forecasts from Eta/CPTEC model, expressing the future atmospheric conditions, are used as inputs in Artificial Neural Networks (ANNs), in order to achieve more reliable short-term forecasts for the incident solar radiation. Global solar radiation measurements performed by two stations of the SONDA project located in south Brazil (Florianópolis and São Martinho da Serra) are used as the targets during ANNs training and for forecasts evaluation. Solar radiation forecasts from ANNs present higher correlation coefficients and lower errors than the Eta model output for shortwave radiation on ground. The well-know bias observed in solar radiation forecasts by the Eta model was removed by the use of ANNs. The improvement in RMSE obtained with ANNs over the Eta model was higher than 30%, estimated with a skill- score. This improvement is a response to a constant demand from the energy sector for more accurate ways of forecasting the solar energy power, so as to support the management of the national generation and distribution systems of electricity. 1. INTRODUCTION * The study of the incident solar radiation has several implications for agriculture, illumination and heating of buildings and residences, and, of course, for meteorological research. In addition, owing to the fast increase in importance of the solar energy resource as viable and promising source of renewable energy, its demand for solar radiation studies has expanded accordingly. Economical and environmental reasons have motivated the increasing use of alternative and renewable sources of energy in Brazil and in the rest of the world: environmental damages caused by fossil fuels consumption; concerns about the elevation of atmospheric carbon levels and consequent temperature increasing and climate changes; the commitment for reduction of carbon dioxides (and other greenhouse gases) emissions by the countries that ratified Kyoto Protocol; the perspectives of oil depletion in next decades (Bentley, 2002; Geller, 2003); the increasing demand for energy to support the new expanding economies such as China, India and Brazil (Goldenberg and Villanueva, 2003); the demand from energy matrixes for complementary resources to overcome instabilities such as that observed in hydroelectric generation during dry seasons; and causes such as the international crises that impact the oil price. Solar energy is one of the most promising options of renewable energy resources in Brazil. Since most of the Brazilian territory is located in the inter-tropical region, a *Corresponding author adress: Ricardo A. Guarnieri, Instituto Nacional de Pesquisas Espaciais, Centro de Previsão de Tempo e Estudos Climáticos (CPTEC), Divisão de Clima e Meio Ambiente, São José dos Campos, SP, Brazil. E-mail: ricardog@cptec.inpe.br high potential of solar energy is accessible along whole year (Colle e Pereira, 1998). The current disadvantages of this energy resource are the high costs, the inconstant and unknown availability, and the dependence on the weather and climate conditions. Solar energy costs are expected to fall in next decades, due to technologic advances and market demands. On the other hand, while technologic advances are foreseen, studies are required for a more reliable assessment of the regional availability, the temporal variability and the predictability. There is a worldwide demand from the electricity energy sector for accurate forecasts of solar energy resources so as to manage co- generation systems. In addition, accurate short- term forecasts of solar radiation is an important information for the management of energy dispatch in transmission lines, since the solar radiation influences the heat dissipation by the cables. Forecasting solar irradiation, even one day in advance, is a complicated task. Part of the difficulties arises from the solar radiation dependence on clouds and meteorological conditions, which intrinsically involves non-linear physical processes. Other difficulties are linked with the inaccuracy of weather forecasts by numerical models, due to the complexity of the non-linear processes involved, and also due to the difficulties of forecasting the optical properties for the future state of the atmosphere. Mesoscale weather forecast models usually have radiation parameterization codes, since solar radiation is the main energy source for atmospheric processes. The Eta model that runs operationally in the Brazilian Center of Weather Forecast and Climate Studies (CPTEC/INPE) has outputs for many meteorological variables, including solar radiation incidence on ground. However, these radiation forecasts are greatly overestimated. 1777