Smart grids have recently encountered several challenges, particularly the growing threat of fraud facilitated by technological advancements. In this article, we explore novel methodologies for fraud detection using artificial intelligence. We started our analysis by scrutinizing statistical techniques such as the ARIMA model, which has proven effective in anomaly detection due to its time series forecasting capabilities. Unfortunately, this method suffers from moderate accuracy and a high rate of false positives. We also used deep learning models such as CNN-1D, CNN-2D, and CNN-LSTM, which demonstrated much better performance than ARIMA. The model is able to capture complex dependencies in the data used as a case study. CNN-2D performed best, with a reasonable number of false positives; its accuracy was 93.18%. We show the weaknesses and strengths of each approach chosen. As a summary, we can say that the CNN model can achieve excellent performance in fraud detection, but ARIMA is not completely ruled out for basic anomaly detection with nonlinear cases and complex patterns.