According to the invention, a probability integral method parameter prediction model for optimizing the BP neural network based on a combined algorithm (GP) of a genetic algorithm and a particle swarmoptimization algorithm, and an input layer of the BP neural network is optimized by adopting an average influence value algorithm (MIV), so that the complexity of the network is reduced, and the purpose of improving the prediction precision is achieved. An MIV-GP-BP model is established by taking actual measurement data of 50 working surfaces as a training set and a test set of the BP neural network, and the precision and reliability of a model prediction result are analyzed; the results show that: in five parameters, the root mean square error ranges from 0.0058 to 1.1575; the maximum relative error of q, tan beta, b and theta is not more than 5.42%, the average relative median error is less than 2.81%, the s / H relative error is not more than 9.66%, the average relative median error is less than 4.31% (the parameter itself is small), and the optimized neural network model has higher prediction precision and stability.