Dynamic Load Forecasting with Adversarial Domain Adaptation Based LSTM Neural Networks

Electric load forecasting technologies determine the constraints of power supply to balance the demand in optimizing power system operation. The load demand becomes uncertain as the statistical characteristics/distributions of power consumption frequently change due to higher turnovers of consumers, such as temporary tenants flowing into or out of apartments. Hence, this study focuses on the dynamic load whose historical consumption readings are limited following a unique characteristic and whose power demand is thus hard to predict. Motivated by the implicit continuity each time the characteristic changes (i.e., distribution shifts), we introduce an adversarial domain adaptation based on the deep recurrent neural architecture for the dynamic load in a specific period. The domain adaptation learns the statistical shift from labeled records of prolonged historical loads and the target load in a data-driven manner so that the architecture can finally only utilize limited records to accurately predict the demand for the target load. Finally, we generated synthetic profiles with the developed model and other leading methods, e.g., support vector regression (SVR), gradient boosting decision tree (GBDT), random forest (RF), fully-connected feedforward neural network (FFNN), long short-term memory (LSTM) recurrent neural network (RNN), gated recurrent unit (GRU) RNN, sequence-to-sequence (Seq2Seq) RNN, and Transformer neural network, and compared them with true values. Experimental results confirmed the effectiveness of using domain adaptation and transfer learning.