A short-term photovoltaic power prediction model using cyclical encoding and STL decomposition based on LSTM

Photovoltaic (PV) power generation is greatly influenced by time series, and the accurate short-term prediction of PV power is helpful to the planning and scheduling of power grid. Traditional forecasting methods have a poor ability to capture periodic and seasonal information and temporal dependency. To address this problem, a short-term PV power prediction model based on cyclical encoding (CE), seasonal and trend decomposition using loess (STL), and long short-term memory (LSTM) is proposed. Firstly, the input sequence is decomposed into three components, including seasonal, trend, and remainder components, using STL decomposition. Moreover, the timestamps are encoded using a cyclical encoding method to represent the positional feature of input sequence. The positional features are combined with the three components, respectively, to form three new feature sets. Subsequently, the three feature sets are separately inputted into three independent LSTM models to obtain three corresponding predicted components, which are summed up as the final PV power prediction results. Lastly, two real datasets from the PV power stations in Ningxia and Xinjiang Provinces, China, are used to verify the proposed model. The results indicate it has better performance than existing models and achieves the smallest prediction error on the two datasets, i.e., mean absolute errors of 1.32 MW and 0.99 MW, respectively.