Probabilistic Grid Load Forecasting: Surpassing TSO Baselines via Temporal Fusion Transformers

Introduction
Accurate electricity load forecasting is essential for smart grids and energy trading. While imbalance price forecasting is bounded by extreme volatility, grid load exhibits robust auto-regressive patterns. Market participants typically rely on deterministic day-ahead forecasts from Transmission System Operators (TSOs). However, relying solely on external providers limits Profit and Loss (P&L) risk management. This research introduces a probabilistic deep learning framework to predict load, systematically benchmarking against and improving upon official forecasts published by the Italian TSO (Terna).

Methods
The forecasting engine adapts a Temporal Fusion Transformer (TFT) architecture. The model ingests historical grid loads, spatial-temporal static covariates, and exogenous variables. Crucially, the TFT decoder incorporates Terna’s day-ahead load forecasts via real-time API as future-known deterministic inputs. Employing Quantile Loss for calibrated uncertainty bands and Variable Selection Networks (VSNs) to dynamically weigh features, the architecture is trained to identify and correct residual error patterns inherent in the TSO's baseline predictions.

Results
Backtesting on a representative validation window (July 1, 2025) compared TFT median predictions against the official ENTSO-E/Terna day-ahead baseline. The TFT achieved a Mean Absolute Error (MAE) of 172.61 MW, Root Mean Squared Error (RMSE) of 218.37 MW, and Weighted Mean Absolute Percentage Error (WMAPE) of 0.5%. The TSO baseline generated an MAE of 420.69 MW, RMSE of 500.87 MW, and WMAPE of 1.1%. This yields relative reductions of 59.0% (MAE), 56.4% (RMSE), and 54.5% (WMAPE). The 10th and 90th quantiles provided calibrated bounds successfully enveloping the actual load curve.

Conclusions
These findings demonstrate TSO load forecasts can be systematically refined using probabilistic architectures that capture residual error and temporal dynamics. Transitioning to an independent, calibrated probabilistic prediction allows energy traders to mitigate volume risk. This framework enables robust bid optimization in day-ahead and intraday markets, enhancing P&L risk management and reducing imbalance cost exposure.