Abstract:
One of the primary challenges in this context is to predict with high accuracy Global
Horizontal Irradiance (GHI) for an efficient integration of solar energy into power grids,
especially within areas that possess different climates. In this work, we tackle the
solar irradiance forecasting problem using a substantially large and complicated data
instead of typical ones used in existing literature. The study attempts to improve
the generalizability and accurate prediction of GHI by using historical time series for
Pakistan. This approach employs several machine and and deep learning models (e.g.,
Multilayer perceptron, Random Forest, AdaBoost, SVR, RNN, GRU and LSTM). The
performance was evaluated on each dataset separately, as well as over their combination.
All together to sum up, in deep learning models the LSTM made a perfect job —
R2=0.99; RMSE = 2.0134; MAE = 1.2354. While in machine learning models the
Multilayer Perceptron model again had a high R2 value of 0.9998, RMSE of 4.3135 and
MAE was found to be only at 2.5969. As a result, the usage of bigger datasets for GHI
forecasting provided better and dependable predictions. In addition to providing new
fundamental understanding, the work also stands to inform practical applications related
to operational optimization of photovoltaic (PV) power plants and acumen regarding
energy trading strategies that could be leveraged for better integration of solar energy
onto a grid. Moreover, the developed methodologies are transferable to other areas as
well and thus support globally sustainable energy facilitation.
