Improving ultra-short-term photovoltaic power forecasting using advanced deep-learning approach

Forecasting photovoltaic (PV) power output is of high importance to optimal operation, scheduling and planning of power systems with much PV due to the weather uncertainty. Here, an advanced deep-learning method employing Gated Recurrent Unit (GRU) network with dual attention (DA) mechanism and Encoder-Decoder framework is applied to improve ultra-short-term PV output forecasting up to a 2-hour time range. The novel DA-GRU model showed better forecasting accuracy over other models such as neural networks and machine learning under single and multi-step forecasting. The mean absolute percentage error of 1-step (15 min), 4-step(1 h) and 8-step (2 h) forecasting with DA-GRU was 12.4 %, 20 % and 22 %, respectively. Increasing the number of forecast steps and time range improved the accuracy of DA-GRU over the other forecasting models. The better feature extraction ability based on feature and time attention has greater impact on the forecasting results at different time points yielding a better prediction accuracy.