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