Design of Experiments for Pix2Pix Hyperparameter Analysis in Energy Consumption Data Imputation

The occurrence of missing values in time series is a recurring problem caused by sensor failures or communication losses, compromising the quality of data-driven analyses. This work investigates the influence of Pix2Pix Generative Adversarial Network (GAN) hyperparameters on the imputation of missing values under different missing data mechanisms. Using the Electricity Load Diagrams 2011-2014 dataset, 40% of the observable values were artificially removed according to the Missing at Random (MAR), Missing Completely at Random (MCAR), and hybrid (MAR+MCAR) mechanisms. The Design of Experiments (DOE) approach was employed to systematically evaluate the effect of the Pix2Pix GAN hyperparameters on the Adjusted Symmetric Mean Absolute Percentage Error (ASMAPE). In addition to identifying the combination of learning rate and number of steps as the most influential hyperparameters in imputation performance, the results showed that the model maintains similar performance under different missing data mechanisms, highlighting the robustness and reliability of the Pix2Pix GAN in imputation tasks. The study contributes a statistically grounded methodology for hyperparameter analysis in energy consumption time series.