Detection and location of high impedance faults in power distribution systems by means of online condition monitoring and protection devices is a challenge, especially in systems with distributed generation. This paper describes a wavelet transform feature extraction method combined with a pair of online adaptive neural networks to detect and locate high impedance faults in time-varying distributed generation systems. Empirically validated IEEE models were used to generate data streams containing faulty and normal occurrences. Comparative results considering feed-forward, radial-basis, and recurrent neural networks as well as the proposed hybrid wavelet-adaptive neural network approach are shown. Interesting results in the sense of accuracy for different scenarios were achieved. Robustness to the effect of distributed generation and to transient events is attained through the ability of the neural network to adapt parameters, number of hidden neurons, and connection weights on the fly. New conditions could be captured by changing the structure of the neural model.