Photovoltaic power forecasting: A hybrid deep learning model incorporating transfer learning strategy

Abstract

Accurate forecasting of photovoltaic power is essential in the integration, operation, and scheduling of hybrid grid systems. In particular, modeling for newly built photovoltaic sites is restricted by insufficient data and training burden. In this study, a novel hybrid photovoltaic power forecasting model assisted with a transfer learning strategy is proposed. The hybrid model, named the attention-dilate convolution neural network-bidirectional long short-term memory network, consists of three steps. Step 1 - Input reconstruction: the historical power and meteorological factors are reconstructed as new inputs based on their relevance to the forecast by introducing a long short-term memory-based attention mechanism; Step 2 - Feature extraction: a hybrid structure is applied to extract spatial and temporal features from new inputs in parallel; Step 3 - Feature mapping: the extracted features are mapped into the forecasted photovoltaic output. Furthermore, to address the modeling for new sites, a transfer learning strategy that fine-tunes the pre-trained model is proposed in this work. The structure by step-wise division allows fine-tuning to be applied to the necessary parts rather than the entire model. Subsequently, the data from the actual photovoltaic system was acquired to validate the proposed model and transfer learning strategy. The proposed model showed significantly superior performance than the other models in the tests, and the parameter transferring not only makes up for the data shortage but also effectively accelerates the model training. With the transfer learning strategy, the maximum improvement in accuracy and training efficiency reached 69.51% and 71.42%, respectively.