Abstract:
In recent times, the world has inclined towards using renewable energy sources since they have relatively lower greenhouse gas emissions, occur freely in nature, and unlike fossil fuels, cannot be depleted. Wind power is one such renewable energy source that has attracted a lot of research and interest in the power industry. With the growing quantities of wind power generation incorporated into power systems, grid reliability is at risk since wind power is highly intermittent. Wind power forecasts help incorporate wind in a grid’s power mix more efficiently and reduce the quantity of power reserves allocated to cater to the intermittency of wind. This makes adopting more wind power resources into the grid more economical. This work developed an approach to wind power forecasting using Bidirectional Long Short-Term Memory (BiLSTM) Neural Networks hybridized with data decomposition techniques and a wind power curve layer. First, a wind power curve was modelled from the historical wind speed and wind power datasets using the Avrami equation and the best line of fit determined. Next, the wind time series data was decomposed into several Intrinsic Mode Functions (IMFs) and a Residual Function using Empirical Mode Decomposition (EMD). Finally, the BiLSTM model enhanced with the Avrami Power curve was used to forecast future wind power values. The developed model was tested on an online-based dataset and compared with the traditional LSTM, BiLSTM and hybrid (Bi)LSTM - data decomposition models. Using the developed BiLSTM + EMD enhanced with an Avrami Power Curve layer, wind power prediction RMSE improved by at least 50% for the 24-hour forecast compared to hybrid BiLSTM-data decomposition models.
