Abstract：Most photovoltaic (PV) power prediction models adopt batch offline training, which poses a challenge on dealing with limited training data for newly established PV power plants. In order to address this issue, a PV power prediction model based on a combination of convolutional neural network (CNN) and long short-term memory (LSTM) network using incremental learning is proposed. Firstly, the CNN is used to extract the features of the meteorological data, and the power prediction is carried out through the LSTM network. The CNN-LSTM hybrid model is used for background learning, to train a baseline model that can be used for incremental learning. Secondly, incremental learning training is carried out according to different time spans to realize the online update of the model. In order to solve the problem of catastrophic forgetting in incremental learning, this paper uses the elastic weight consolidation (EWC) algorithm and the online elastic consolidation (Online_EWC) algorithm. Experimental results show that, compared with unconstrained incremental learning, incremental learning using EWC and Online_EWC methods can significantly alleviate the problem of catastrophic forgetting and reduce the prediction mean absolute error (MAE) and root mean square error (RMSE), up to 21.7% and 18.3%, respectively. At the same time, the time cost of incremental learning is significantly lower than that of traditional batch learning.
2024, Vol. 25 
