Few-sample model training assistant: A meta-learning technique for building heating load forecasting based on simulation data

Accurately predicting building heating loads with limited data is a challenge. Utilizing data from other buildings to create pre-trained models can assist in model training for the target building, such as transfer learning. However, obtaining high-quality source data for different target buildings is difficult in practice, impacting the implementation of this method. This paper introduces a Few-Sample Model Training Assistant (FSMTA) for predicting building heating loads. Firstly, prototype building simulation models of target building type are constructed, followed by Monte Carlo method for a large amount of simulation data of buildings with different parameters. Subsequently, a meta-model is developed by MAML meta-learning method to extract common features of these buildings through the simulation data, which is used as the FSMTA. The FSMTA was validated in high-rise residential buildings. 600 simulated buildings were used as source data to obtain the FSMTA. Another 40 simulated buildings and 1 actual building were used for testing under 7 different few-sample conditions. The results show that the FSMTA can reduce the average error by 4.11 %–36.75 % compared to direct learning and by 4.98 %–10.28 % compared to the multi-source transfer learning method.