EVALUATION OF SOLAR RADIATION FORECAST MODELS: LIGHTGBM, LSTM AND GRU

Abstract

Accurate solar radiation forecasting is a crucial technical factor in optimizing the performance of solar power systems. This paper evaluates the performance of
three advanced machine learning models: Light Gradient Boosting Machine (LightGBM), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) in solar
radiation forecasting. The models were trained and tested on a real-world dataset, including meteorological parameters and measured solar radiation. Evaluation
criteria such as Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) were used to compare the models' performance. Results show that each model has
its strengths and weaknesses. The LightGBM model demonstrated superior performance during training with faster training times and higher accuracy (0,5s;
RMSE = 54.8 W/m2
và MAE = 27.6 W/m2
) so với LSTM (456,5s; RMSE = 59.2W/m2
; MAE = 34.7W/m2
) và GRU (397,2s; RMSE= 59.3W/m2
; MAE = 34.7W/m2
). The
three models showed comparable accuracy in forecasting scenarios, but LightGBM had significantly lower prediction times than LSTM and GRU. LSTM and GRU,
although more complex and requiring longer training times, also demonstrated good forecasting capabilities with complex time series data characteristics. The paper
provides a comprehensive view of these models' performance and recommends selecting appropriate models in solar radiation forecasting applications.