Comparison of Inputs Correlation and Explainable Artificial Intelligence Recommendations for Neural Networks Forecasting Electricity Consumption

The energy sector explores various paths to improve the energy management of buildings. Nowadays a frequent path is to schedule load forecasting activities due to the accessibility of reliable forecasting algorithms. Data scientists usually take advantage of a large historic of consumption with weekly patterns and sensors data presenting a higher correlation with the consumption variable. However, specialists in the explainable artificial intelligence area focus on studying the positive or negative impact of each variable to the prediction accuracy. In this paper, a correlation analysis evaluates in the first stage the most reliable sensors to be used during training and forecasting tasks. In the second stage, the Local Interpretable Model-Agnostic Explanations (LIME) explainable artificial intelligence method is applied to determine which features have a stronger positive or negative influence on the prediction accuracy. The training and forecasting tasks are supported in this paper by the forecasting algorithm Artificial Neural Networks. In the case study, a historic of two years and six months is used to estimate the consumption values of a targeted week considering periods of five minutes. The results section calculates the confidence of each sensor to the prediction accuracy provided by LIME method and compares the obtained insights with the correlation analysis. The results and conclusions sections state that the two sensors more correlated with the consumption variable either contribute negatively to the prediction performance or do not contribute at all on most test targets.