Short-Term Wind Power Forecasting in East Java Using Gated Recurrent Unit

Wind power is one of Indonesia's renewable energy resources. One of the challenges in using wind power is wind speed forecasting. An accurate wind speed forecasting can help determine the output of wind turbine power, which is helpful for generator scheduling to ensure stable electricity supply. Recurrent Neural Network (RNN) is one of the common forecasting methods. Its advantages lie in its feedback loop structure that allows RNN to learn the order and dependence between time lags. This study proposed the use of Gated Recurrent Unit (GRU) as a wind turbine's power forecasting method by forecasting the wind speed. When compared to Long-short term memory (LSTM), GRU as an RNN variant also offers simpler formulation and improving the training efficiency. The simulation result with M2T1NXSLV dataset showed that the GRU model trained using 24-hour time lag, 50 epochs, and external variables (i.e., pressure, humidity, and temperature) exhibited the most accurate forecasting result with MAE and MAPE of 0.107m/s and 8.06% respectively. These metrics were better than LSTM trained with similar parameters, except the number of epochs (i.e., 40), with MAE and MAPE of 0.109 m/s and 8.22%, respectively.