Improving short-term power load forecasting with a domain adversarial transfer network

Data-driven approaches have been extensively utilized in power load forecasting to enhance predictive performance. Nevertheless, when electricity retailers integrate new consumers, the absence of historical consumption data often restricts the effectiveness of traditional data-driven techniques. To tackle this challenge, this study introduces a short-term load forecasting approach based on a domain adversarial transfer network. The model adopts a Transformer structure as its feature extraction module to capture dynamic patterns and temporal relationships within load sequences. Following this, the load predictor leverages these features to generate precise forecasts of future electricity demand. By implementing adversarial training between the domain discriminator and the feature extractor, the model promotes the learning of domain-invariant representations. Moreover, techniques such as multi-kernel maximum mean discrepancy and correlation alignment are integrated to minimize distribution shifts between source and target domains. Validation was conducted using power consumption data from industrial users in a southern Chinese province. Experimental outcomes indicate that the proposed method attains satisfactory accuracy and exhibits robust adaptability in small-sample scenarios.