This study proposes a new technique for real-time building energy modelling and event detection using kernel regression. We show that this technique can exceed the performance of conventional neural network algorithms, and do so by a large margin when the available training dataset is small. Furthermore, unlike the synapse weights in a neural network, the parameters of our kernel regression models are amenable to human interpretation and can give useful information about the building being studied. We extensively test our proposed algorithms using a new dataset consisting of 1.5 years of power and environmental measurements for four buildings, in addition to benchmarking against the ASHRAE Predictor Shootout dataset. On the new dataset, our kernel regression algorithm gave the best prediction performance in three of four cases and significantly outperformed neural networks (the nearest competitor) with training sets of 1/2 a year or less.