Temporal Fusion Transformer for Day-Ahead Wind Power Forecasting in the South African Context

The rising integration of wind power generation sources in electricity grids threatens grid stability due to the stochastic and variable nature of this energy source. Wind power generation forecasting is a possible solution to mitigate risks associated with unbalanced electrical supply and demand in a power grid with high levels of wind power generation sources. This paper proposes a wind power forecasting technique used to mitigate the uncertainties associated with this variable input source. The forecasting method proposed is based on the Temporal Fusion Transformer technique, which is a state-of-the-art time series forecasting model. The wind power forecasting model is predominantly aimed towards achieving accurate day-ahead wind power forecasts but also considers shorter forecasting horizons. The paper further explores adapting the wind power forecasting model to the prevailing South African electricity market for wind power as governed by the Standard Power Purchase Agreements applicable to the Renewable Energy Independent Power Producers Procurement program. This study represents an application of the temporal fusion transformer directly to raw wind power output data. The study finds that the proposed forecasting model is comparable in performance to commonly used recurrent neural network based wind power forecasting techniques. The model degrades appreciably 10 to 12 hours after the last model input is accessible. The study further indicates possible drawbacks with the proposed wind power forecasting model and recommendations applicable to the Renewable Energy Independent Power Producers Procurement program power forecasting constraints.