An explainable artificial intelligence benchmarking with deep reinforcement learning for telecom battery backup prediction

Abstract

Accurate prediction of battery backup time is essential for ensuring service reliability and efficient energy management in modern telecommunication networks, particularly during power outages. This study proposes a comprehensive benchmarking framework to evaluate the performance of eleven advanced time-series forecasting models, including deep learning architectures (multi-layer perceptron (MLP), convolutional neural network 1-dimension (CNN1D), long short-term memory (LSTM), bidirectional long short-term memory (BiLSTM), gated recurrent unit (GRU), and temporal convolutional network (TCN)), Transformer-based models (Transformer, Informer-Lite, and patch time-series Transformer-lite (PatchTST-Lite)), and linear baseline approaches (decomposition linear (DLinear) and normalized linear (NLinear)). The models are trained and validated using real-world telecom power system data under varying load and environmental conditions. To improve model transparency and reliability, an explainable artificial intelligence (XAI) framework is incorporated to interpret feature contributions and identify key factors influencing battery performance. Experimental results indicate that Transformer-based models outperform conventional deep learning and linear models in terms of prediction accuracy and robustness across different operating scenarios. Furthermore, XAI analysis highlights that load current, ambient temperature, and battery health indicators are the most influential variables affecting backup time prediction. The proposed framework provides both high accuracy forecasting and interpretability, enabling more effective decision-making and enhancing energy resilience in telecommunication infrastructure.