Emotions do not emerge randomly out of the sudden over time. In fact, there is a serial regularity in the way emotions appear in textual sequences. Consequently, by targeting the ordered interdependencies of the occurrences of the emotional attitudes, we can model the corresponding latent distributional emotional information. This is done through a distributional emotion embedding methodology that we have previously introduced. At the same time, energy load fluctuations constitute a phenomenon that is subdetermined by the current context and behaviors of human communities. These, however, are expressed linguistically. Therefore, by modeling them appropriately, we can potentially improve our energy load forecasting schemes. The present work incorporates emotional features extracted from a distributional embedding methodology applied to world news data, and by exploiting the corresponding latent information of the serial dependencies between emotions, significantly improves a series of machine learning regression schemes in the task of forecasting the Greek energy system load. On this basis, we demonstrate the universal prevalence of the proposed methodology, highlighting a series of extensions.