LOWESS smoothing and Random Forest based GRU model: A short-term photovoltaic power generation forecasting method

Accurate prediction of photovoltaic power generation is vital to guarantee smooth operation of power stations and ensure users’ electricity consumption. As a good forecasting tool, Gated Recurrent Unit method has been widely used in different forecasting areas. However, the existing studies ignore the impact of data fluctuations on prediction accuracy, to fill the gaps and enhance prediction accuracy, several different data smoothing techniques are introduced and compared to reduce fluctuations, Random Forest method is used for feature selection, and RepeatVector layer extended by attribute dimensions and TimeDistributed layer with full connectivity are utilized to optimize the Gated Recurrent Unit model. A real-world case from the photovoltaic power plant in Xuhui District, Shanghai, China, is adopted to evaluate the performance of proposed method. The comparing results with Recurrent Neural Networks and Long Short-Term Memory, and the actual data as well, show that the proposed prediction method can effectively improve the prediction accuracy of photovoltaic power generation. We also use the daily and monthly data of The Desert Knowledge Australia Solar Centre in Australia to investigate whether the proposed method is suitable for short-term or medium and long-term prediction. The results indicate that our method is more appropriate for short-term prediction.