To address the high computational cost and low efficiency in high-resolution heat-conduction topology optimization, this paper employs the Problem independent machine learning (PIML) method to achieve fast and high-accuracy topology optimization for large-scale heat conduction structures. The core idea is to condense the interior-node degrees of freedom of substructures onto the boundary nodes using numerical shape functions, thereby effectively reducing the number of degrees of freedom in the solution process. Meanwhile, an artificial neural network is employed to learn the numerical shape functions, avoiding the situation where the online computation of the condensed stiffness matrix would otherwise negate the efficiency gains and even lead to lower overall efficiency. For the first time, the body force term (i.e., internal nodal thermal loading) in the Problem independent machine learning method is discussed, and the temperatures at the interior nodes of substructures are corrected accordingly. The results show that, with acceptable error, the problem-independent machine learning method can significantly improve solution efficiency compared with the conventional finite element method.