The principle and characteristics of BP neural network and genetic algorithm are introduced and Pidgeon process is briefed. The standard BP neural network has disadvantages of slow-rate convergence and getting easily into local minima value. The prediction model based on BP neural network optimized by genetic algorithm with input of calcined dolomite activity, silicon ratio, pelletizing pressure, reduction time, reduction temperature, and vacuum degree is established to study the relationship between process parameters and magnesium reduction degree. The model is rehearsed and tested by the screening production data. The results show that the prediction model can precisely predict the magnesium reduction degree, the hit rate of the model with ΔηMg≤ ±1. 0% is about 96%, the maximum error is less than 1. 3%. To some extent the selection of process parameters in Pidgeon process can be extracted by the model.