Improving voltage harmonics forecasting at a wind farm using deep learning techniques

Abstract

The South African renewable energy mix includes wind, solar, hydro and ocean. This energy mix contributes to the nation energy requirements while reducing dependency on fossil fuel and in the process mitigating the emission of green-house gases. Wind power generation is always associated with the generation of voltage harmonics. Precise predictions of the presence of voltage harmonics is of vital importance in order to ensure clean voltage is coupled to the national grid. A total of 8103 voltage harmonics, measured at Jeffreys Bay Wind Farm in the Eastern Cape Province have been used in our experiments. The proposed model would take two steps to extract important features present in the voltage harmonics signals. The mean voltage amplitude is extracted using moving window segmentation. Long short-term memory (LSTM), a deep learning method, is used in the prediction of voltage harmonics generation based on the voltage features extracted. LSTM is a special kind of recurrent neural network (RNN) capable of learning long-term dependencies. For simplicity the model uses one LSTM layer with 128 hidden neurons. 8103 calculated mean values were used as the expected data to train the model in Matlab. The LSTM model could predict the next 3800 sample mean values with low root mean square error (RMSE).