Based on deep learning, a new prediction model was proposed to preliminarily explore the feature interpretability. The core idea of the proposed method is that the local functional connectivity density(lFCD) and four dimensional(spatial, temporal) consistency of local neural activities(FOCA) are fed to the autoencoder(AE) separately to extract the temporal and spatial features automatically. And then the features are used to predict the individual fluid intelligence score with deep neural network and ensemble learning. Finally, Pearson correlation coefficient and mean absolute error are used to assess the performance of the proposed method. The results of proposed method show that the mean absolute error between the predicted value and the real value was 4.1±3.2, and the Pearson correlation coefficient was 0.55(P=1.9×10-18). Compared with CPM(Connectome Predictive Model)and dimension reduction-predictive methods, the proposed method achieved the best performance. In addition, the highly selective spatial pattern mainly include the parietal lobe, occipital lobe area, basal ganglia network, and its surrounding deep nucleus area of the brain. This study improves the model of local functional connectivity to predict fluid intelligence, and the prediction error is lower than that of CPM(connectome predictive model). At the same time, the visualization of features reflects the spatial pattern of brain function activity related to fluid intelligence, which indicates that the proposed method can help us understand the age-related brain function change pattern, and has good application value.