Forecast Adaptation to Charge Transfers

This paper presents a study in neural networks for short-term predictions, in the scope of electric load forecasting in the presence of charge transfer scenarios. Charge transfers are humanprovoked transfers of electrical charge from one subnetwork to another, due to a shift in both subnetwork’s needs. This work studied the problem of how to dynamically adapt the model when a situation of charge transfer arises, presenting a simple method to increase robustness of predictive models through situations of charge transfer, decreasing damage to the subsequent days’ predictions. The study gathered data preprocessing, auto-correlation methods to choose model’s input attributes, decreasing of backpropagation’s parameters and an extensive evaluation of configurations. This study also introduces the concept of simulating random charge transfers on the train set, which proved to achieve better performance in charge transfer situations, and in the subsequent days. Key-words: Load forecast, charge transfers, neural networks, time series correlation.