The paradigm of electrical networks is changing rapidly due to the large penetration of renewable energy sources and other distributed generation and storage assets. The variability in the generation profile hinders the operation of electrical systems and forces to improve control systems as well as to have a better knowledge about how distribution networks perform under intermittent conditions in order to ensure its reliability. However, the network capacity and reliability itself are compromised if the distribution system operators don't fully control nontechnical and non-expected losses in the grid. Therefore, fraud and other grid anomalies' detection and assessment, as well as a detailed evaluation of consumers' electricity loads, becomes a priority for paving the way to smarter low-voltage electricity grids. In this framework, this paper provides a methodology to improve the reliability of lowvoltage networks by developing a dynamic method based on data-driven models which is able to detect and evaluate the origin of non-technical losses.