Abstract: Aiming at the problem of strong randomness of photovoltaic power data and difficulty in accurate prediction, a combined prediction method based on temporal convolutional network（TCN）, bidirectional long short-term memory network（BiLSTM） and echo state network（ESN） is proposed. Firstly, the complete ensemble empirical mode decomposition with adaptive noise analysis（CEEMDAN） is used to decompose the power data into a series of relatively stable sub power subsequences. Then, the decomposed and reconstructed power sequence and other feature sequences are input into the TCN-BiLSTM Attention-ESN. TCN-BiLSTM Attention-ESN is applied to extract features and then spatiotemporal feature vectors are constructed. Finally, the extracted spatiotemporal feature vectors are input into ESN to obtain the prediction results. The proposed method is validated using photovoltaic power data from photovoltaic power stations in Xinjiang, China. The results showe that compared with current advanced prediction methods, the proposed method has higher prediction accuracy, which helps to increase the proportion of photovoltaic power generation and ensure the balance and operation safety of the power system.