Load Forecasting and Fault Prediction Framework for Distribution Grids: A Bus- and Topology-Agnostic Solution

The rapid electrification of energy systems and aging power grid infrastructure have created a pressing need for effective predictive maintenance and fault forecasting. This paper introduces a critical event forecasting framework for medium and low-voltage distribution grids, leveraging smart meter data to predict faults and provide actionable insights that help grid operators manage systems closer to their operational limits. Designed to handle the complex task of load forecasting for small, low-voltage consumer clusters, the framework incorporates three core modules: a forecasting module using Long Short-Term Memory (LSTM) layers for precise load predictions, a network module that aggregates data from smart meters, and an event prediction module that identifies potentially fault-prone operations. By adopting a bus-semi-agnostic and topology-agnostic clustering approach, the framework not only boosts computational efficiency but also enables short-term scenario analysis and power flow optimization. Tested on a real Danish distribution system, the framework outperforms traditional methods, such as the Standardized Load Profile (SLP), in both load forecasting and fault prediction, especially for predicting rare, high-impact events like overloads and under-voltages. Our evaluation indicates improved prediction precision for voltage and load extremes compared to the SLP baseline. Our evaluation indicates improved prediction precision for voltage and load extremes compared to the SLP baseline. These results underscore the framework’s potential as a non-wires alternative for grid reinforcement, allowing operators to extend asset life and delay major investments without sacrificing system reliability.