Under the influence of multiple factors, runoff time series shows nonlinear and chaotic characteristics. Back propagation neural network (BPNN) model alone can be used to the medium and long term prediction of the runoff, but it has the disadvantage of insufficient quantification of runoff influencing factors. Single chaos theory model can quantify the influencing factors of runoff, but it is only applicable to the short term prediction. In view of this situation, a medium and long term runoff prediction model coupled with chaos theory and BPNN was established. Time delay τ, embedding dimension m and maximum Lyapunov exponent λmax of the runoff series of Yingluoxia Hydrological Station in the upper reaches of Heihe River from 1944 to 2017 were calculated by chaos theory, and the phase space reconstruction of the runoff series was carried out to determine the number and values of neurons in the input layer of BPNN, and the time span of the prediction. Then BPNN was used to train the runoff data from January 1944 to December 2012, based on which two prediction models of chaos-BPNN and chaos-BPNN equal dimension replenishment were established. The runoff data from January 2013 to December 2017 were used to verify the models. The result showed that the prediction accuracy of the chaos-BPNN equal dimension replenishment model reached 91.84%, with a satisfactory prediction effect. The runoff prediction models coupled with chaos theory and BPNN combine the advantages of the two methods, especially the chaos-BPNN equal dimension replenishment model. It can not only supplement new information but also eliminate the old data which has lost its purpose due to the development of the system, thus reducing the time span of BPNN training and improving the prediction accuracy. This model is a new effective approach to the prediction of medium and long term runoff.