کنفرانس بین المللی محاسبات نرم

صفحه اصلی / ششمین کنفرانس بین المللی محاسبات نرم

A Robust Deep Learning Framework for Urban Resource Forecasting under Non-Stationary Dynamics

نویسندگان :

Mazaher Saeedi Siraee¹

1- دانشگاه جامع علمی ـ کاربردی

کلمات کلیدی :

Intelligent Systems،City Management،Artificial Intelligence،Transformer

چکیده :

Efficient management of urban resources, such as energy, water, and transportation, requires accurate forecasting models that can handle the nonlinear, dynamic, and non-stationary nature of urban data. In this study, a hybrid machine learning framework is proposed that integrates discrete cosine transform (DCT)-based feature extraction, frequency-domain feature fusion, and an adaptive optimization strategy using the Covariance Matrix Adaptation Evolution Strategy (CMA-ES). The proposed model effectively captures both temporal and frequency characteristics of urban signals while automatically tuning its hyperparameters for enhanced adaptability. Experimental evaluations conducted on a real-world urban energy dataset demonstrate that the proposed approach outperforms conventional Recurrent Neural Network (RNN) and tree-based models. Specifically, it achieves 62.7% lower MAE, 59.1% lower RMSE, and 69.4% lower MAPE, while reducing computational time by 66.7%. These results confirm the model’s robustness, efficiency, and suitability for intelligent, data-driven urban resource management systems.