Forecasting masked-load with invisible distributed energy resources based on transfer learning and Bayesian tuning

Abstract

Load forecasting with distributed energy resources (DERs) behind-the-meter is more challenging owing to transformed data patterns. Traditional forecasting method which is only based on unmasked-load could not suit the present limited masked-load. To bridge the divergence between unmasked-load and masked-load, this article proposes a masked-load forecasting (MLF) method based on transfer learning technique and Bayesian optimization, which is Maximum Mean Discrepancy-Neural Network with Bayesian optimization (MMD-NN^b). At first, common feature vectors between unmasked-load and masked-load are extracted and an outcome predictor could be established based on feature vectors from historical unmasked-load. The feature vectors from masked-load could therefore accommodate to the outcome predictor, and the masked-load could be forecast. Owing to the excessive hyperparameters involved in training, Bayesian optimization is adopted for hyperparameters fine-tuning. MMD-NN^b was tested and compared with four related models. The improvements from MMD-NN^b were observed in all comparison scenarios. Also, MMD-NN^b was proved to have high resilience to the different DERs and not requiring additional DERs-data.