Högskolan Dalarnas logga och länk till högskolans webbplats

Accurate building energy consumption prediction is crucial for building energy management. However, a substantial number of buildings lack sufficient data that hinders the application of data-driven models for energy prediction. Transfer learning emerges as a powerful strategy to address the challenge posed by limited data availability. This research proposes a novel model ensemble method based on Multi-Layer Perception (MLP) structure, which can realise selfadaptive weight, to exploit the advantage of existing transfer learning strategies for building energy sequence-to-sequence (Seq2seq) prediction. The overall and stepwise model performance comparisons between traditional transfer models and traditional ensemble methods in 4 different transfer scenarios are conducted to prove the superior performance of proposed method under all the investigated transfer conditions. The impact of prediction step length on the model performance is also investigated. The results show that the proposed method outperforms traditional transfer models and ensemble methods at different prediction steps in all the investigated transfer conditions. Compared to the best performing transfer model, the proposed method can reduce prediction error by 6.83 %-25.08 %. Compared to the best performing ensemble method, the proposed method can reduce prediction error by 6.32 %-36.54 %. The analysis of the selfadaptive weight reveals that the proposed method is capable of dynamically allocating weights to the two transfer models to enhance the prediction accuracy.