Exploring phase transition behaviors and constructing phase diagrams are of importance for theoretically and experimentally studying ferroelectric physics and materials. Because of the rapid development of computers and artificial intelligence, especially machine learning methods combined with other computational methods such as first principle calculation, it is possible to predict and choose appropriate materials that meet the target requirements from a large number of material data, which greatly saves the cost of experiments. In this work, we use neural network method and phenomenological theoretical calculations to accurately predict the phase structures that may appear in the phase diagrams of different orientated Pb(Zr<sub>0.52</sub>Ti<sub>0.48</sub>)O<sub>3</sub> ferroelectric films, and establish the temperature-strain phase diagrams of (001), (110) and (111) oriented thin film, and calculate the polarization and dielectric properties of different oriented films at room temperature. By analyzing the changes of prediction accuracy and loss with the number of iterations, it is found that the deep neural network method has the advantages of high accuracy and speed in the construction of the film temperature-strain phase diagram and the prediction of the types of phases. Through the analysis of the room temperature polarization and dielectric properties, it is found that the (111)-oriented PbZr<sub>0.52</sub>Ti<sub>0.48</sub>O<sub>3</sub> film has the largest out-of-plane polarization and the smallest out-of-plane dielectric coefficient, and they are insensitive to misfit strain. This work provides guidelines for designing micro-nano devices that require the stable dielectric coefficient and polarization performance in the special working environment and operation.