This paper proposes a novel BP network model based on nonlinear iterative partial least-squares algorithm which can fit nonlinear data. The novel BP network model can reduce iterative step number and advance learning effieieney. This paper pretreats data by nonlinear iterative partial least-squares algorithms. The weights initialization of input floor and output floor are set by applying the loading weights of dependent variable and cause variable, the member of hidden nodes are set by applying factor numbers of nonlinear iterative partial leastsquares algorithm, the connection co- efficient is set by applying the connection matrix B. Performances of the BP, PLS, and PLS-BP are analyzed and compared. The results show that the PLS-BP has better fitting and forecasting than BP and PLS.