Abstract: In order to further improve the accuracy of ultra-short-term photovoltaic power prediction, a multi-model fusion photovoltaic power prediction method based on reinforcement learning is proposed. Firstly, the local outlier factor（LOF） algorithm is used to detect and remove outliers, and a multilayer perceptron regression algorithm is employed to correct the data anomalies. Then, the data is divided into training, validation, and testing sets. In the training set, models such as support vector Regression（SVR）, multiple linear regression（MLR）, Bayesian ridge regression（BRR）, convolutional neural network-long short term memory（CNN-LSTM） and particle swarm optimization-gated recurrent unit（PSO-GRU） are trained. These trained models are validated on the validation set to select the best-performing models as sub-models. Finally, in the testing set, the five sub-models are used for forecasting, and their predictions are fused using a reinforcement learning method. The fusion value is taken as the final prediction result. Experimental results show that the proposed method significantly reduces the mean absolute error, mean squared error, root mean squared error, and relative error compared to single-model methods and other traditional fusion methods, verifying the effectiveness of this approach.