Explainable Needn't Be (Much) Less Accurate:Evaluating an Explainable AI Dashboard for Energy Forecasting

摘要

This paper presents the evaluation results of an improved version of an interactive tool for energy demand and supply forecasting, based on the combination of explainable machine learning with visual analytics. The prototype applies a kNN algorithm to forecast energy demand and supply from historical data (consumption, production, weather) and presents the results in an interactive visual dashboard. The dashboard allows the user to understand how the forecast relates to the input parameters and to analyse different forecast alternatives. It provides small utilities not familiar with AI with an easily understandable, while sufficiently accurate tool for energy forecasting in prosumer scenarios. The evaluation of the forecast accuracy has shown our method to be only 0.26%-1 .73% less accurate than more sophisticated, but less explainable machine learning methods. Moreover, the achieved accuracy (MAPE 5.06%) is sufficient for practical needs of the application scenario. The evaluation with potential end-users also provided positive results regarding the usability, understandability and usefulness for the intended application context.