Ultra Short Term Photovoltaic Power Prediction Based on Reinforcement Learning and Combined Deep Learning Model

Ultra short term photovoltaic power prediction is of great significance for the safe operation of photovoltaic grid connected system. In order to solve the problem that the traditional single prediction model is affected by the random fluctuation of photovoltaic power, which leads to the poor prediction accuracy, a combined deep learning prediction model is proposed. Firstly, wavelet packet decomposition is used to decompose the original PV power sequence, which initially reduces the instability of the original PV power. Secondly, based on the first mock exam, three single models of long and short time memory network, gated loop unit and recurrent neural network are used to predict PV power, and three prediction results are weighted together. Finally, the Q-learning algorithm of reinforcement learning is used to optimize the combination weight, so as to maximize the prediction performance of the combination model. The experimental results based on the measured data of a photovoltaic power station show that the proposed combined forecasting model is superior to other forecasting models, and the effectiveness of the proposed model is verified. © 2021, Power System Technology Press. All right reserved.