부산대학교 도서관

One key obstacle to high penetration of solar energy into electric grids is the intermittency of the energy source. Solar power forecasting can be the key to overcome intermittency by allowing energy providers to plan for fluctuations in advance and avoid disruptions to the grid. Over the years, various forecasting models have been presented in the literature that makes use of environmental variables such as temperature, humidity, soiling, and other parameters to forecast the output of solar panels in a given environment. In regions with stable environmental conditions, many parameters may not be necessary for accurate prediction. This paper presents a minimalistic prediction model based on feedforward artificial neural networks that predicts the power output of a soiled panel using only time of the day and the number of days the solar panel has not been cleaned for as inputs. This model can be used to determine a cleaning schedule to prevent losses due to soiling. The model was trained using six months' worth of hourly data measured from a 100-Watt monocrystalline solar panel. The model was able to predict maximum power with an RMSE of 0.069. Two models are presented in this paper. One used random discrete grid search for the hyperparameters selection and other was a hand-crafted model; both performed similarly.