A Comparative Study of Explainable AI Techniques for Short-Term Energy Forecasting: Evaluating WindowSHAP and TimeSHAP across Time-series Prediction Models

This paper analyses the compatibility of the different explanation models for time-series load fore- casting by evaluating them based on the quantitative method. The aim of the analysis is inspired by the digital twin development project for efficient energy distribution of Ecofactorij, in which the energy load is expected to be accurately forecasted using AI tools. However, the current pre- diction models are often described as black-box, which means the model cannot project insights into its own decision-making process for the outputs. In order to solve this, several explanation methods have been developed and deployed to address this lack of transparency. Therefore, this analysis aims to reveal the advantages, tendencies, and limitations of each combination of an explanation model and a prediction model for time-series forecasting. In particular, this paper argues for WindowSHAP and TimeSHAP, the two different variants of Shapley Additive Explana- tion(SHAP) by quantitatively scoring them based on their fidelity, stability, and sparsity to exam- ine the characteristics of each coupling in combination with a time-series prediction model with a black-box problem, namely Recurrent Neural Network(RNN), Long Short-term Memory(LSTM), and Gated Recurrent Unit (GRU). A human-subject Likert scale survey was used to qualitatively evaluate each explanation method’s performance, which was also evaluated using fidelity, sparsity, and stability metrics. While both approaches show limitations in stability and user interpretability, the results show that TimeSHAP performs slightly better than WindowSHAP across all metrics, especially fidelity. The results support the continuous efforts towards transparent AI systems for sustainable energy management and provide useful insights into the application of XAI in time- series forecasting.