Physics-informed neural network (PINN) offer a novel numerical approach for solving partial differential equations (PDEs). To investigate their potential in multi-physics simulations of electrical equipment, this study examines the accuracy of PINN for solving electrostatic fields with complex boundary conditions using a needle-plane electrode geometry, representing insulation defects. A PINN model with five hidden layers was constructed. The loss function was formulated by integrating the electrostatic PDE, Dirichlet boundaries (high voltage and grounded electrodes), Neumann boundaries (zero charge), and floating potential constraints (equipotential conditions and charge conservation). Electrostatic field solutions under both simple and complex boundary conditions, including a small circular area with a floating potential boundary representing conductive impurities, were compared with finite element method (FEM) results. Under simple conditions, PINN converged after 6,000 iterations, yielding results closely matching FEM solutions with average absolute error of 0.0015 V and 0.010 V/m for potential and electric field intensity, respectively. However, introducing complex floating potentials caused convergence difficulties even after 20,000 iterations, significantly impacting accuracy and indicating sensitivity to loss function weighting.