This paper concerns the application of a neuro-fuzzy learning method based on data streams for high impedance fault (HIF) detection in medium-voltage power lines. A wavelet-packet-transform-based feature extraction method combined with a variation of evolving neuro-fuzzy network with fluctuating thresholds is considered for recognition of spatial–temporal patterns in the data. Wavelet families such as Haar, Symlet, Daubechie, Coiflet and Biorthogonal were investigated as a way to provide the most discriminative features for fault detection. The proposed evolving neuro-fuzzy classification model has shown to be particularly suitable for the problem because the HIF environment is subject to concept changes. Different from other statistical and intelligent approaches to the problem, the developed neuro-fuzzy model for HIF classification is not only parametrically, but also structurally adaptive to cope with nonstationarities and novelties. New neurons and connections are incrementally added to the neuro-fuzzy network when necessary for the identification of new patterns, such as faults and usual transients including sag, swell and spikes due to the switching of 3-phase capacitors and energization of transformers. Experimental evaluations compare the proposed classifier with other well-established computational intelligence methods, viz. multilayer perceptron neural network, learning vector quantization neural network and a support vector machine model. Results have shown that the evolving neuro-fuzzy system is effective and robust to changes. The system is able to maintain its detection and classification accuracy even in situations in which other classifiers exhibit a significant drop in accuracy due to gradual and abrupt changes of the fault patterns. Fuzzy rules are useful for interpretability purposes and help to enhance model credibility for decision making.