FILLING MISSING VALUES FOR AI-BASED (LOAD) FORECASTS WITHIN THE INTERFLEX MICRO GRIDDEMO IN SIMRIS, SWEDEN

Missing data impairs the performance of most neural networks with a particularly strong effect on time series prediction networks. Imputation addresses this issue and by replacing missing values with substitute values. The choice of a suitable imputation method requires fundamental knowledge of the dataset.Autoencoders (AE) have been widely applied in representation learning and feature extraction. In this paper we usea stacked denoising overcomplete autoencoder for imputation in multi-variate time series. We assess the model’s feature reproduction capability and compare its effect to simple mean imputation on a open source dataset. Moreover, we assess the imputation’s influence on a recurrent neural network’s short-term load forecasting results and show that our proposed autoencoder model yields better results in feature imputation and significantly improves the forecasting accuracy for low and high fractions of missing data