In recent years, optimized power line communication techniques (PLC) for smart grid metering systems are significant to existing electrical infrastructure for transmitting data effectively. Traditional approaches for optimizing power line communications in smart grids had faced several challenges which include Attenuation, Signal Degradation and noise interferences. Therefore, this research proposes Temporal Convolutional Network–Multi-Head Attention (TCN-MHA) for optimizing PLC in smart grid metering systems. Initially, the data is collected from publicly available Fühler‐im‐Netz (FiN) dataset and this dataset consists of real-world data of PLC. Then, FiN dataset is pre-processed by using Hilbert-Huang Transform (HHT) and z-score normalization which converted frequency domain signals from time domain signals and normalized signals into similar ranges of values. After that, features are extracted by using Short-Time Fourier Transform (STFT) that provides effective characteristic extraction and time–frequency analysis. Finally, proposed TCN-MHA is utilized for optimizing PLC in smart grids. The proposed TCN-MHA attained better results in terms accuracy (99.95%), precision (99.26%), recall (99.94%), and F1 score (99.16%) when compared to existing Convolutional Neural Network with Gated Recurrent Unit (CNN-GRU) method.