Introducing explainability in sequence-to-sequence learning for short-term load forecasting

Methods to forecast electric loads and generation timeseries are widely applied in power system operation balancing. For this purpose, most sophisticated forecasting methods complex, since the net electricity consumption is not dependent on explainable by a single cause. The increasing complexity furthers trend towards machine-learning that achieve more accurate load forecasts with exogenous features. However, well-known downside of machine learning users will face non-transparent, inexplicable models difficult relate to. In paper, we propose graphical representation salient features as functions time for day-ahead residual forecasting. presented work extends sequence-to-sequence recurrent neural network model visually accommodate explanation produced forecast. resulting saliency maps reduce high input–output relationships two-dimensional plane over steps.