Efficient probabilistic forecasting for power system time series via conditional variational autoencoders

Abstract

Forecasting the power system time series (PSTS) distribution provides valuable priors for dispatch operators, enhancing both economic efficiency and operational security. This is especially pertinent given the abundance of data now available from advanced metering infrastructure. However, current quantile-based and artificial neural network (ANN)-based probabilistic forecasting (PF) methods suffer from the quantile crossing problem. This leads to the non-monotonic nature of the cumulative distribution function and reduces power system dispatchers’ confidence in the PF results. To address this problem, this paper proposes an efficient PF method called conditional variational autoencoder (CVAE) based PF (CPF) for PSTS with monotonicity guarantees. The CPF model (CPFM) integrates the principles of CVAE with the characteristics of PF, thereby enabling effective learning and inference of PSTS probabilistic distributions. The adaptation of CVAE for PF involves several critical steps. First, the encoder–decoder structure of the CVAE model is modified for compatibility with both the PSTS input–output formats and the existing ANN architectures. Second, multivariate Gaussian distributions are employed as the latent variables to fit PSTS’ distributions of any continuous shape through the probabilistic processes learned by the CVAE. Third, the CVAE training process is derived and utilized to transform traditional point PSTS into PF by maximizing the evidence lower bound. Furthermore, the inference process within the CVAE is modified to generate the complete distribution of PSTS and to ensure its monotonicity. Empirical results on the Switzerland load dataset, the Denmark solar PV power dataset, and the Palo Alto Electric Vehicle charging dataset demonstrate that the CPFM achieves nearly double the performance of the baseline models in providing accurate and reliable PSTS distributions.