Hybrid Neural Network–ARIMA for Time-Series Bias Correction of GFS Wind Speed Data to Support Renewable Energy Assessment in Java, Indonesia

Bias correction of Global Forecast System (GFS) wind speed data is essential for accurate wind resource assessment in Indonesia, particularly in regions where observations from Automatic Weather Stations (AWS) are sparse and wind variability is high. This study develops a Model Output Statistics (MOS)-based post-processing framework that combines nonlinear deep learning models, including a Convolutional Neural Network–Long Short-Term Memory (CNN–LSTM) model, linear statistical models such as Autoregressive Integrated Moving Average (ARIMA) and Seasonal Autoregressive Integrated Moving Average with eXogenous variables (SARIMAX), and hybrid configurations to correct time-series bias in GFS wind speed at five locations in Java. One year of GFS hindcast data and AWS observations was used to train and validate six predictive model schemes under single- and multi-predictor settings using an expanding- window cross-validation strategy. Model performance was evaluated using multiple error metrics and a composite index to identify the best-performing configuration at each site. The results show consistent improvements relative to the GFS baseline, with performance differences associated with local wind variability and the interaction between linear and nonlinear components in the time series. Overall, the proposed framework provides a robust and adaptable approach for improving GFS-based wind information, with practical relevance for wind energy assessment and operational forecasting in Indonesia.