Bayesian Matrix Factorization for Electricity Load Imputation

The missing values in electricity load are a critical issue in grid applications. The electricity demand time series data are usually large-scaled with complicated patterns and cause difficulties for imputation. This paper presents a Bayesian matrix factorization (BMF)-based imputation method for large-scale electricity load missing value imputation. Through factorizing the original electricity load matrix into two latent matrices, the intrinsic information of the electricity load matrix can be discovered. Two Bayesian inference algorithms, Gibbs sampling and iterated conditional models are applied to solve the BMF model. The affect of the matrix rank to the electricity load imputation task is empirically studied. Experimental results on three real-world electricity load datasets are presented to show the superiority of the proposed method against five benchmark algorithms.