Performance assessment of SARIMA, MLP and LSTM models for short-term solar irradiance prediction under different climates in Morocco

Abstract

Photovoltaic (PV) production is highly dependent on global solar irradiance (GHI) and often experiences irregular fluctuations. With the continuous increase in PV penetration rates, GHI forecasting methods are becoming important to ensure optimal management of the energy produced. In this study, three forecasting models – Multilayer Perceptron (MLP), Long Short-Term Memory (LSTM), and Seasonal Autoregressive Integrated Moving Average (SARIMA) were evaluated to predict GHI one day ahead. Ground measurement data from four sites in Morocco and weather forecasts from the Global Forecast System (GFS) were used to perform the forecast. Results show that forecasts based on MLP and LSTM are more accurate than SARIMA and persistence even if under complicated weather conditions. For clear days with low variability, the RMSE for LSTM, MLP M1, and MLP M2 are 18.58, 12.97, and 45.20 W/m². For cloudy days with high variability, the RMSE are 103, 60.80 W/m², and 89.17 W/m², respectively.

KEYWORDS:

• Forecasting
• solar irradiance
• machine learning
• PV production

Acknowledgements

The authors would like to express their gratefulness to Green Energy Park research platform, IRESEN (Research Institute in Solar Energy and New Energies) for providing ground measurements. The authors would like to thank Dr. Natalie Hanrieder from DLR for providing EnerMENA ground measurements.

Disclosure statement

No potential conflict of interest was reported by the author(s).