Causal-Augmented Forecasting for Energy-System Decision Support

Local energy systems–microgrids coupling renewable generation with electrical and thermal storage–must be sized, operated, and offered into ancillary-services markets under uncertainty about demand, generation, and prices. Optimal configurations cannot be derived analytically, and the data-driven tools used in practice are largely correlational: they degrade under regime changes and offer operators little transparent reasoning. We present \emph{causal-augmented forecasting}, which pairs neural and statistical time-series models with structural causal models so that decision support reflects the causal drivers of demand, generation, and price, supporting intervention-aware, explainable recommendations with improved robustness under structural change. The method is embedded in a \emph{deterministic-first} architecture: classical, auditable computation produces every figure, language models are confined to explanation, and a validation guard rejects any numerical claim not traceable to a computed value. On 3.5 years of real European public energy data (2022--2025, hourly), the strongest model attains 4.92% day-ahead MAPE with calibrated 90% prediction intervals (empirical coverage 0.89-0.92); across the 2022 European energy-price shock the causal-augmented model degrades less than a purely correlational one x 1.16$ vs. x 1.31$–a modest, honestly reported robustness gain. Point accuracy is delivered by the neural model, while the causal layer's value is explainability, intervention-aware decision support, and robustness. We show how the forecasts drive a decision layer–sizing optimisation, predictive control, and ancillary-service provision–over an energy-system digital twin, and release the evaluation harness openly for reproducibility.