Data-Efficient Electricity Consumption Forecasting with a Tabular Foundation Model

We address building-level electricity consumption forecasting under limited data. We introduce a simple pipeline that couples feature engineering with an electricity consumption predictor comprising a tree-based model and a tabular foundation model. A lightweight meta-learner aggregates their outputs to exploit complementary strengths of these models. We assess the effectiveness of the proposed approach on the KEA-2025 dataset. Experimental results on varying training-history lengths demonstrate that the foundation model excels in low-data regimes, the tree-based model improves as data grow, and the ensemble consistently achieves the best performance.