This paper proposes an unsupervised adversarial autoencoder (AAE) for detecting false data injection attacks (FDIAs) in unbalanced power distribution grids. The AAE model uses long short-term memory (LSTM) to capture temporal dependencies in time-series data and leverages generative adversarial networks (GANs) for better reconstruction of input data. The model does not rely on abstract models or mathematical representations, making it effective for detecting anomalies in systems with nonlinear characteristics and unbalanced configurations. The AAE is tested on IEEE 13-bus and 123-bus systems with historical meteorological and load data under three types of data falsification functions. The results show that the AAE outperforms other unsupervised learning methods in detecting FDIAs. The model uses two discriminators to distinguish between real and generated data, with CNNs employed for each to identify hidden patterns. The AAE's performance is evaluated using metrics such as accuracy, precision, recall, and F-1 score, demonstrating its effectiveness in detecting different types of FDIAs. The model is compared with other data-driven methods, including autoencoders, CNNs, and clustering algorithms, showing superior performance in terms of detection accuracy and F-1 scores. The AAE model is capable of detecting anomalies in both balanced and unbalanced power distribution systems, making it a promising solution for securing smart grids against cyber threats.This paper proposes an unsupervised adversarial autoencoder (AAE) for detecting false data injection attacks (FDIAs) in unbalanced power distribution grids. The AAE model uses long short-term memory (LSTM) to capture temporal dependencies in time-series data and leverages generative adversarial networks (GANs) for better reconstruction of input data. The model does not rely on abstract models or mathematical representations, making it effective for detecting anomalies in systems with nonlinear characteristics and unbalanced configurations. The AAE is tested on IEEE 13-bus and 123-bus systems with historical meteorological and load data under three types of data falsification functions. The results show that the AAE outperforms other unsupervised learning methods in detecting FDIAs. The model uses two discriminators to distinguish between real and generated data, with CNNs employed for each to identify hidden patterns. The AAE's performance is evaluated using metrics such as accuracy, precision, recall, and F-1 score, demonstrating its effectiveness in detecting different types of FDIAs. The model is compared with other data-driven methods, including autoencoders, CNNs, and clustering algorithms, showing superior performance in terms of detection accuracy and F-1 scores. The AAE model is capable of detecting anomalies in both balanced and unbalanced power distribution systems, making it a promising solution for securing smart grids against cyber threats.