Continuous-Time Recurrent Neural Networks for Wind Power Prediction

Nowadays, wind power energy is a efficient renewable resource due to its low hazardous environmental emissions. Therefore, it is widely utilized to help meet electricity demand by providing stable and sustainable power generation in everyday grid operations and plays a crucial role for wind power industries in the power supply market. To ensure the continuous power supply from this sustainable power resource, accurate wind power forecasting within a specific time period is essential. Also, it necessary for effective wind power management, wind turbine maintenance, scheduling, and energy production forecasting. In recent time, advanced forecasting architectures like RNN, LSTM, and GRU are working in a efficient way for wind power forecasting. As these types of architectures mainly rely on extracting features from time series data, significantly improving forecasting accuracy. However, these architectures take more computational time when dealing with large or complex datasets, making them less practical for the wind power management industry. To address this key limitation, this study proposes a Continuous-Time Recurrent Neural Network (CTRNN) architecture that requires less time for training and validation while maintaining almost same accuracy compared to other architectures. This effectiveness of the proposed architecture has been explained by applying CTRNN to wind power data. And also evaluate its performance and compare with above mentioned architectures using various metrics such as MAE, MSE, RMSE, and R2 score. The results of short-term wind power forecasting by CTRNN are showing that the proposed model achieves comparable values at 1500 epochs for these metrics such as MAE: 0.102, MSE: 0.0305, RMSE: 0.1746, and R2 score: 0.9726. While its computation time (997.51s) is lower than that of other architectures such as RNN, LSTM, and GRU.