Analysis of input parameters for deep learning-based load prediction for office buildings in different climate zones using eXplainable Artificial Intelligence

Abstract

Simplification of input variables can increase the applicability of Artificial Intelligence (AI) in building load prediction. The most essential inputs for AI therefore need to be identified via a significance level test. In this study, the significance of the input parameters was evaluated using the standardized regression coefficient (SRC) and Explainable AI methods, i.e., Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP). To consider various types of office buildings operating under different climates, the constraints of the U.S. Department of Energy reference buildings with the Latin hypercube sampling were used to generate two thousand building models for each of five climate zones classified by the Köppen and Geiger system. The number and types of essential inputs for deep learning-based building load prediction were identified based on SRC, LIME, and SHAP. By comparison, the SHAP method gave the smallest number of essential inputs for accurate load prediction. In addition, the types of essential inputs varied with different climates. This study presented the potential elimination of weather sensors with time variables depending on the climate conditions. It would help building practitioners and non-experts to determine the essential inputs to build up a simple but accurate deep learning-based building system.