XDL-energy: Explainable hybrid deep learning framework for smart campus energy forecasting and anomaly justification using the DPU-ALDOSKI dataset

Mohammed A. M.Sadeeq,
XDL-energy: Explainable hybrid deep learning framework for smart campus energy forecasting and anomaly justification using the DPU-ALDOSKI dataset,
Energy and Buildings,
Volume 347, Part B,
2025,
116305,
ISSN 0378-7788,
https://doi.org/10.1016/j.enbuild.2025.116305.
(https://www.sciencedirect.com/science/article/pii/S0378778825010357)
Abstract: Managing energy efficiently and in a way that is understandable is essential for the sustainable and intelligent operation of modern smart campuses. In this paper, we introduce XDL-Energy, a novel framework that combines a hybrid Convolutional Neural Network–Long Short-Term Memory (CNN-LSTM) model with SHapley Additive exPlanations (SHAP) to deliver both high-accuracy energy consumption forecasting and interpretable anomaly detection. The model is trained on the real-world DPU-ALDOSKI dataset, which contains over 3 million multi-sensor IoT records from distributed campus buildings. By fusing CNN’s spatial pattern recognition with LSTM’s temporal sequence modeling, the framework captures complex energy behaviors across time and space. SHAP analysis reveals key drivers of predictions, such as temperature, motion, and humidity, offering both global and local interpretability. Experimental results demonstrate that XDL-Energy achieves robust performance (MAE: 0.045, RMSE: 0.062, MAPE: 4.8%) and reliably flag anomalies using dynamic rolling thresholds. Additional evaluations including ablation studies, SHAP dependence plots, and multiple anomaly case analyses confirm the system’s practical relevance. XDL-Energy bridges the gap between deep learning and transparency, making it a deployable and trustworthy solution for smart campus energy management.
Keywords: Smart campus; Energy forecasting; IoT; CNN-LSTM; Explainable AI; SHAP; Anomaly detection; Time series prediction; Energy management; Hybrid models