A parameter-based multi-source transfer learning method for building load forecasting with sparse data scenarios

Accurate forecasting of building loads is essential for optimizing energy management and reducing energy consumption. However, forecasting accuracy is often compromised in new buildings or sparse data scenarios. Most existing research relies heavily on large amounts of historical data, with limited studies addressing the challenges posed by data sparsity. To address this issue, this paper proposes a forecasting method based on multi-source transfer learning and a Convolutional Neural Network-Gated Recurrent Unit (CNN-GRU) model. The proposed approach first trains models separately on multiple data-rich source domains and then transfers deeper layers of these models to smaller target domain datasets. This enables the target domain model to incorporate both knowledge from multiple source domains and its own learned patterns. Additionally, a Maximum Mean Discrepancy (MMD) strategy is employed for source domain selection, mitigating distributional differences between source and target domains. To validate the effectiveness of the proposed method, open-source data from U.S. commercial buildings was used for forecasting experiments. The results demonstrate that the proposed method improves forecasting accuracy, reducing the Root Mean Square Error (RMSE) by 44.15 %, the Mean Absolute Error (MAE) by 46.71 %, and increasing the R-squared value by 2.38 % (0.988) compared to the non-transferred CNN-GRU model. These findings indicate that the proposed method is a feasible solution for load forecasting in sparse data scenarios. • The model reduced RMSE by 44.15 % and MAE by 46.71 %, while improving R² by 2.38 % over the non-transferred model on the test set. • The smaller the MMD value of similar buildings, the better the forecasting effect of migration learning. • Multi-source migration learning models have overall better predictive performance compared to single-source migration learning models. As the number of source domains increases, the accuracy improves. • For different combinations of the same number of source domains, the lower the total MMD value the better the forecasting. • The MMD value should be kept below 0.021 for a significant lifting effect.