Comparative optimization of global solar radiation forecasting using machine learning and time series models

Area of Science:

• Renewable Energy
• Data Science
• Meteorology

Background:

• Photovoltaic (PV) power output forecasting is increasingly important due to the rise of solar energy.
• Global Solar Radiation (GSR) forecasting is complex, heavily influenced by unpredictable weather patterns.
• Accurate GSR forecasting is essential for reliable PV energy generation.

Purpose of the Study:

• To propose an algorithm for selecting optimal machine learning and time series models for GSR forecasting.
• To minimize forecasting errors for GSR data specific to Tetouan, Morocco.
• To compare the performance of selected models against a persistence model.

Main Methods:

• The study evaluated Autoregressive Integrated Moving Average (ARIMA), Feed Forward Neural Network with Back Propagation (FFNN-BP), k-Nearest Neighbour (k-NN), and Support Vector Machine (SVM).
• These models were compared against a persistence model as a baseline.
• Statistical metrics were employed to rigorously evaluate and compare the forecasting performance of each method.

Main Results:

• Machine learning and time series models proved straightforward to implement for GSR forecasting.
• The Feedforward Neural Network (FFNN) and ARIMA models demonstrated superior performance.
• Both FFNN and ARIMA provided accurate approximations for the GSR output.

Conclusions:

• The proposed algorithmic approach effectively identifies optimal models for GSR forecasting.
• FFNN and ARIMA are highly effective for predicting GSR, offering reliable approximations.
• This research contributes to improving the accuracy of solar energy resource assessment.