Anomaly detection plays a significant role in the area of Smart Grids: many algorithms were devised and applied, from intrusion detection to power consumption anomalies identification. In this paper, we focus on detecting anomalies from smart meters power consumption data traces. The goal of this paper is to replicate to a much larger dataset a previously proposed approach by Chou and Telaga (2014) based on ARIMA models. In particular, we investigate different model training approaches and the distribution of anomalies, putting forward several lessons learned. We found the method applicable also to the larger dataset. Fine-tuning the parameters showed that adopting an accumulating window strategy did not bring benefits in terms of RMSE. While a 2s rule seemed too strict for anomaly identification for the dataset.