•A new method for power quality disturbance detection and classification based on deep learning at the edge.•Deep learning is used for automatic feature extraction and selection.•The proposed methodology is embedded in a low-cost smart meter.•Performance evaluation using accuracy, precision, recall and F1-Score. The large amounts of data collected by smart meters (SM), such as electric energy, water gas consumption and power quality (PQ) metrics, can create a massive flow of data transmitted between consumers and utilities. In this context, an edge-fog-cloud architecture based on a low-cost SM is proposed. The employed SM acquires voltage and current signals to obtain their frequency and amplitude, allowing PQ to be monitored through methods of detection and classification of disturbances in order to send only information about the detected disturbances to the utility, thus reducing network traffic associated with PQ disturbances in Smart Grids. The proposed methodology was embedded at a low-cost SM to enable data exchange with the utility, offering an enormous potential for real scenarios.