Temporal Fusion Transformer for Alarm Forecasting in a Danish DSO: Embedding Pearson Correlation

Effective situational awareness in power distribution systems hinges on accurate forecasting of critical grid parameters. This paper leverages a Temporal Fusion Transformer-based framework for short-term load forecasting tailored explicitly for alarm forecasting in a Danish distribution system operator. Unlike most conventional load forecasting approaches, our model prioritizes the prediction of busbar voltage magnitudes and line currents, parameters directly related to alarm triggers. These forecasts provide operators with early warnings of potential voltage violations and line overloads, enabling timely corrective actions to maintain grid stability and reliability. The model uses three years of operational data, wavelet-transformed features, and a Pearson correlation matrix to embed spatial interdependencies and physical grid topology into the forecasting process. By focusing on alarm-related parameters, the framework enhances the operator's ability to manage critical events proactively. This multi-horizon approach ensures precise forecasting of voltage and currents, enabling reliable alarm prediction, as confirmed by comparative analysis.