Temporal Fusion Transformers for Commercial Building Energy Forecasting: A UK Half-Hourly Settlement Case Study

Accurate half-hourly energy forecasting is a commercial and regulatory imperative for UK businesses following the introduction of the Mandatory Half-Hourly Settlement (MHHS) programme, which required all Measurement Class C and D electricity customers to settle in half-hourly intervals from 2025, now in effect. Despite the urgency of this requirement, existing forecasting approaches have predominantly relied on traditional statistical methods or shallow machine learning models inadequately suited to the multi-variate, multi-horizon nature of half-hourly settlement forecasting. This paper presents the first application of the Temporal Fusion Transformer (TFT) architecture to UK commercial building energy forecasting under MHHS conditions, evaluated against a dataset of 1,112 commercial buildings with 36.4 million half-hourly electricity records and 5.2 million gas records drawn from the Building Data Genome Project 2 (BDG2). The TFT electricity model achieves a Mean Absolute Percentage Error (MAPE) of 10.31% — below the 15% commercial benchmark — with RMSE of 63.85 kWh and MAE of 19.83 kWh on a 48-step (24-hour) forecast horizon. Benchmark comparisons against Linear Regression (45.80%), Random Forest (40.70%), XGBoost (40.70%), Standard LSTM (53.80%), Robust LSTM (55.00%), and Optimised LSTM (57.90%) demonstrate the TFT's 74.7% error reduction relative to the best baseline. The TFT's interpretable attention mechanism reveals that 24-hour lag (weight 0.31), hour-of-day (0.24), and 7-day lag (0.18) are the dominant predictive features, supporting the paper's dual-utility framing. These findings establish a performance benchmark for the emerging MHHS compliance software market and confirm that transformer architectures specifically designed for multi-horizon temporal forecasting substantially outperform general-purpose methods on this task.