Abstract: The Physics-Informed Neural Network (PINN) can be used to simultaneously predict the unsteady-state temperature field and unknown heat sources in two/three dimensions. Even when using a single measurement point, the solution strategy proposed still demonstrates excellent anti-noise performance. This method changes the law that traditional neural networks require a large number of sample response data. It only needs to randomly sample the calculation domain and time interval to generate a sample set and, has the ability to flexibly add constraint conditions. It is a new type of intelligent algorithm that can be used to solve forward and inverse problems simultaneously. To reduce the degrees of freedom of the optimization parameters when predicting the full-field temperature and the unknown heat source, the unknown heat source is expanded by using the complete polynomial, and an improved adaptive weighting method is proposed to enhance prediction accuracy. Compared with the traditional iterative inversion method based on the finite element method, the method proposed has greater advantages in the number of measurement points, in calculation cost and in noise resistance performance. Results from several typical numerical examples demonstrate that this method completely avoids the tedious preprocessing element division and exhibits higher prediction accuracy than reference solutions.