Analysis of time series forecasting in application to solar energy harvest

Abstract

The promised future applications in solar energy harvest have been remarkably recognized. However, the hourly forecasting of normal solar irradiance (NSI) outputs is considered a problem due to the dynamic nature of meteorological information not only in a day but also across days. The thesis proposed three neural network models including a dense layer without a hidden layer (DNN_h0), a dense neural network with two hidden layers (DNN_h2), a dense neural network with two hidden layers associated with one intermediate metrological feature (air temperature: T) (DNN_h2T), and dense neural network with two hidden layers associated with 7 intermediate metrological features (DNN_h2F). These models would be used to forecast an hourly prediction of normal solar irradiance (NSI) across an entire day. As well as, we proposed two configurations to represent our datasets: FTC (sine-cosine) and 1H (one-hot) encodings. In addition, we used metrological features such as air temperature T and others to determine the effectiveness of a model’s performance in terms of mean absolute error (MAE). We conducted two groups of experiments: single-step and multi-step prediction models by using one real-world dataset (NREL). As a result, the comparison is revealed that the (NSI) has an acceptable model performance in both FTC and 1H encodings for the multi-step models by using an intermediate metrological feature: air temperature T in the (DNN_h2T) model. Whereas the single-step model (DNN_h0) has shown slightly acceptance to find a well performance to predict the (NSI), while the (DNN_h2) model shows a significant (MAE) values in both encodings.