In order to improve reconstruction accuracy to resolve temperature distribution reconstruction problems, a method for optimal sensor placement based on random forest algorithm is proposed. Denoting different measurement sites as different sample features, a series of different sensor placements and the reconstruction errors which are calculated by these placements constitute a sample dataset. A random forest model is setting up by the sample dataset, feature importance is also evaluated, then the optimal sensor placement is determined by feature importance. Simulation test and combustion test are set up to verify the feasibility and practicability of the proposed method. Testing data shows that comparing the original method, the proposed method can improve the reconstruction accuracy by at least 20%. Research results indicate that the proposed method has a good practical value, it also provides a new probe of using random forest algorithm to solve industrial problems.