Probabilistic Individual Short-Term Load Forecasting Using Conditional Variational Autoencoder

A novel probabilistic load forecasting method based on a Conditional Variational Autoencoder (CVAE) architecture is proposed in this paper. The proposed framework has been tested at the meter level, feeder level, and the Operating Center (“OC”, regional aggregate) level with real data. At the meter level, the proposed approach has been tested with data from 5000 different smart meters. Analysis at the feeder level included comparing the individual forecasts against the aggregate forecasts for 30 different feeders with a total over 16000 individual meters. At the OC level, the model was tested on 50 different OCs. According to the results, the proposed approach is capable of accurately and efficiently forecasting hourly load over a 4-day period at various levels of aggregation. To demonstrate the robustness of the proposed method, the proposed method has been tested on unseen customers from 10 different feeders with total number of 20880 customers. Results show that the proposed method is robust to unseen customers. Additionally, the proposed model has been compared with a range of benchmarks, including the wellestablished long short-term memory (LSTM) model and multiple linear regression (MLR). The CVAE algorithm is more accurate than the other algorithms in forecasting short-term loads.