Uncertainty-aware probabilistic energy forecasting for low-carbon industrial energy management systems

Industrial energy systems integrating on-site renewable generation are subject to increased operational variability and uncertainty, particularly under high renewable penetration and carbon-constrained operation. Conventional deterministic forecasting approaches are insufficient for capturing uncertainty in renewable generation, grid import behaviour, and associated carbon intensity signals. Reliable probabilistic forecasting is therefore required to support next-generation industrial energy management systems. This paper presents a probabilistic multi-horizon forecasting framework based on a multi-output quantile Long Short-Term Memory (LSTM) architecture. The model generates probabilistic forecasts of photovoltaic (PV) generation, grid electricity import, and grid carbon intensity. Prediction interval reliability is improved using conformal prediction, enabling statistically consistent uncertainty calibration without distributional assumptions. The framework is evaluated using high-resolution industrial energy datasets under winter and summer operating conditions. Results demonstrate robust forecasting performance across multiple horizons, achieving median forecast errors below approximately 0.7 in winter and approximately 1.1 in summer. Calibrated prediction intervals achieve empirical coverage levels of approximately 0.80-0.83 in winter and 0.71-0.74 in summer. The results demonstrate the suitability of probabilistic deep learning methods for uncertainty-aware forecasting in industrial energy planning and management.