Due to the complex environment of the fishery in the south Pacific, it is difficult for traditional methods to extract effective features from the fishery impact factor data. A fusion deep learning model CNN-GRU-Attention is proposed to realize the prediction of CPUE (catch per unit effort) in the Thunnus alalunga fishery in the south Pacific. First CNN (convolution neural network) is utilized to extract characteristic data of impact factor within the scope of a single fishery, so as to obtain one dimensional characteristic data matching the CPUE value of the fishery, then GRU (gated recurrent unit) is used to predict CPUE value of the fishery, and finally Attention mechanism is used to accelerate the convergence of the model and optimize the model by weight allocation, and the same data is used to conduct comparative experiments on different models. The results show that: compared with the traditional prediction model multi-layer back-propagation (BP) networks, the absolute error decreases by 0.047 and the root-mean-square error decreases by 0.352. Compared with GRU, the absolute error is 0.012 lower and the root-mean-square error is 0.055 lower. The experimental results prove the effectiveness of CNN-GRU-Attention model and provide a new idea for the method of prediction of fishery.