Day-ahead net load forecasting in ZEBs: A comparative ML study with novel feature engineering and simulation-based evaluation

The growing deployment of zero-energy buildings supports the transition toward a more sustainable building stock. In practice, many of these buildings operate under behind-the-meter configurations to reduce infrastructure costs and simplify data management. However, this monitoring approach constrains the effective exploitation of their flexibility potential, as forecasting shifts from load to net load and must account for the combined uncertainty of demand and photovoltaic generation. To address this challenge, we systematically benchmark a broad range of classical machine-learning and state-of-the-art deep-learning models within a forecasting pipeline that directly predicts the net load profile of a zero-energy building. We further propose a joint hyperparameter and feature selection optimisation framework that integrates model-based feature importance with greedy forward feature addition inside a Bayesian optimisation procedure. Model performance is evaluated using conventional forecasting metrics as well as newly introduced domain-informed measures designed to assess peak and valley prediction accuracy. To reflect the ultimate objective of flexibility provision, we also embed the forecasts in a simulation-based battery dispatch framework that enhances photovoltaic self-consumption and provides peak-shaving support to a distribution transformer. Results show that although several models achieve strong statistical accuracy, XGBoost delivers superior operational performance in the simulation-based evaluation, an advantage not fully reflected by statistical error metrics. The proposed domain-informed metrics, however, successfully identify LightGBM as a consistently strong alternative in capturing operationally relevant dynamics. A sensitivity analysis across different battery capacities further reveals that the selection of forecasting model has a limited impact at low storage levels, where outcomes remain comparable, but becomes increasingly critical as storage capacity grows, significantly influencing planning and flexibility provision.