Abstract: To solve the deficiency of time-lag effect and time cumulative effect in traditional artificial neural networks during dealing with input and output,the unit impulse response function of system theory is brought into neural network model,and a new type of neuron model named impulse response neuron is established.Then,feed-forward impulse response neural network(IRNN) model with one hidden layer is constructed based on the neuron.Furthermore,BP(backpropagation) learning algorithm is deduced for training IRNN,which lays a foundation for IRNN from theory to application.Finally,good results are obtained by IRNN via practical application in rainfall-runoff system simulation,which shows that IRNN is of greater adaptability and superiority in highly non-linear complex mapping system.