Short-Term Load Probability Prediction Based on Conditional Generative Adversarial Network Curve Generation

To ensure the stable and secure operation of the power system, we propose a method for short-term load probability prediction based on Conditional Generative Adversarial Network (CGAN) curve generation. Initially, an adaptive integrated prediction model was developed for daily load key values. This model is based on Multi-Attention Bidirectional Long Short-Term Memory (MABi-LSTM) neural networks, utilizing features including date, temperature, and historical load as inputs. Secondly, the maximum information coefficient method is used to weigh the load characteristics, and a similar curve dataset is constructed based on the weighted K-nearest neighbor algorithm and weighted resampling. Then, using load fundamental values and similar curve datasets as conditions and training sets, respectively, a load curve generation model based on CGAN is constructed, and a loss function for correcting numerical deviation and curve shape deviation is proposed. Finally, considering the uncertainty of the model and noise, a mapping relationship from noise to the probability distribution of the model output is constructed for short-term load probability prediction. Taking the load data of a power grid in a particular area of North China, the proposed method has been verified to have higher prediction accuracy than traditional methods.