Combining transfer learning and constrained long short-term memory for power generation forecasting of newly-constructed photovoltaic plants
Xing Luo, Dongxiao Zhang, Xu Zhu
Abstract
Photovoltaic power generation (PVPG) forecasting has attracted increasing research and industry attention due to its significance for energy management, infrastructure planning, and budgeting. Emerging deep learning (DL) models based on historical data have provided effective solutions for PVPG forecasting with great success. However, newly-constructed photovoltaic (NCPV) plants often lack collections of historical data, and thus it is difficult to forecast their future generation accurately. In this work, combining transfer learning (TL) and DL models, we initially propose two parameter-transferring strategies and a constrained long short-term memory (C-LSTM) model, to address the hourly day-ahead PVPG forecasting problem of NCPV plants. The K-nearest neighbors (KNN) algorithm is utilized to extract prior knowledge as physical constraints, which can guide the training process of C-LSTM. The performances of different TL methods combined with C-LSTM are evaluated specifically, and appropriate ones are determined accordingly. The proposed models are evaluated based on real-life datasets collected from actual PV plants in Australia. The results demonstrate that the proposed C-LSTM model outperforms the standard LSTM model with higher forecasting accuracy. In addition, the results also indicate that significant improvements in forecasting accuracy and stability can be obtained by the proposed TL strategies combined with C-LSTM, regardless of different sky conditions (i.e., clear sky, partly cloudy sky, and overcast sky), compared to the conventional machine learning and statistical models in the literature. The forecasting skill of the combined model has improved up to 68.4% compared with the reference persistence model.