<p>This paper investigates data-driven localized wave solutions and parameter discovery of the Boussinesq equation via an improved Physics-Informed Neural Networks(PINNs) method–Multi-Output Physics-Informed Neural Networks with differential order reduction(TSDOR-MPINNs). This method converts the original equation into a pair of conservation law equations with two outputs by introducing a potential function transformation, and also reduces the order of both the time and space derivatives of the Boussinesq equation. By using the conservation law equations and their corresponding data, the prediction accuracy is significantly improved. Research results show that this method achieves high accuracy in solving all kinds of localized wave solutions for the Boussinesq equation, achieving an improvement of two to three orders of magnitude compared with standard PINNs. Finally, this paper also applies this method to solve the inverse problem of the Boussinesq equation. The results indicate that this method can still effectively identify the unknown coefficient parameters in the equation even when the data contains strong noise. This achievement not only expands the application of PINNs in high-order nonlinear partial differential equations but also provides a more accurate and efficient approach for the numerical simulation of complex wave phenomena.</p>