Hybrid Photovoltaic Power Forecasting Using Onsite and ERA5 Data Based on LSTM Algorithm in an Islanded System: A Case Study of Selayar Island

Abstract

Reliable photovoltaic (PV) power forecasting is important for islanded power systems, where power balance must be maintained locally and operational flexibility is limited. This study develops a short-term PV power forecasting framework for the Selayar Island hybrid PV system by combining onsite measurements and ERA5 reanalysis data. Three input scenarios are evaluated, namely onsite, ERA5, and hybrid data. Long Short-Term Memory (LSTM) is used as the main forecasting model, while Gated Recurrent Unit (GRU), Convolutional Neural Network (CNN), and Feedforward Neural Network (FFN) are used as comparison models. The dataset consists of hourly PV power, onsite irradiance, and ERA5 meteorological variables, including irradiance, cloud cover, temperature, wind speed, and precipitation. A 48-hour sliding window is applied to predict PVpower for one-, two-, and three-hour-ahead forecasts. The results show that irradiance is the most dominant variable affecting PVpower output. Among the evaluated models, the hybrid LSTM model achieves the best performance for the one-hour-ahead forecast, with an RMSE of 74.00 kW, MAE of 41.28 kW, and NRMSE of 7.17%. Compared with the onsite-only LSTM model, the improvement is moderate, with an absolute NRMSE reduction of 0.06 percentage points. Nevertheless, the hybrid approach provides additional weather context that can help the model better represent changing atmospheric conditions. The multi-horizon results also show that prediction errors increase from one to three hours ahead, indicating higher uncertainty at longer forecasting horizons in islanded PV systems.