The size effect has an important influence on the mechanical properties and structural design of concrete materials. Currently, the test is still the main way to study the size effect of concrete. Limited by the sample preparation period, complex boundary, and loading conditions, the comprehensive cost is high, and the result is discrete. With the help of deep learning and the Bayesian optimization algorithm, based on a large number of test data, this paper established a deep neural network model (BO-DNN) of concrete compressive strength size effect without any simplified calculation assumptions. The model was compared with the existing size effect model, and the influence of each parameter on the size effect of compressive strength was investigated by changing the value of selected characteristic parameters. The results show that the size effect of the water-binder ratio on compressive strength is significant, and the smaller the water-binder ratio is, the more obvious the size effect is. The size effect of compressive strength increases with the increase of the aggregate size, but the increase slows down with the increase of the aggregate size. The size effect of compressive strength of specimens with an aspect ratio less than 2 increases with the increase of the aspect ratio, and the size effect tends to be stable after exceeding 2. The shape of the specimen has little influence on the size effect of compressive strength. The longer the curing age is, the more significant the size effect is, but the size effect tends to be stable after the curing age exceeds 90 d. The prediction model proposed in this paper has strong generalization ability, higher accuracy, and stability, and can better mine the complex nonlinear relationship between the characteristic parameters, providing a theoretical basis and reference for the engineering design of concrete materials and structures.