LOAD ANALYSIS AND FORECASTING IN INTERNET OF THINGS NETWORKS

Abstract

This paper addresses the challenge of accurate load forecasting in distributed Internet of Things networks characterized by highly variable traffic patterns, heterogeneous device behavior, and strict energy limitations, particularly for nodes powered by hybrid solar-based sources. The rapid expansion of IoT infrastructures and the increasing complexity of network interactions highlight the need for forecasting models capable of jointly considering both traffic dynamics and the energy state of sensor nodes. The aim of the study is to develop a hybrid forecasting model that integrates deep learning techniques with an energy-aware adaptation mechanism, enabling the generation of network activity predictions that remain consistent with the available energy resources. The proposed architecture incorporates two complementary predictors: a GRU-based module responsible for capturing temporal traffic dependencies, and a 1D-CNN encoder designed to evaluate the node’s accessible energy. A key component of the model is the adaptive scaling coefficient, which adjusts the predicted load according to the current energy profile. To validate the proposed approach, a full software pipeline was implemented, providing data preprocessing, model training, performance evaluation, and visualization capabilities. Experiments conducted on a real-world IoT traffic dataset demonstrate that the hybrid model yields lower prediction errors compared to a baseline LSTM approach and the naïve persistence method, while also exhibiting high robustness under conditions of weak feature correlation and significant noise. Moreover, the model successfully performs energy-aware adjustment of the forecast, preventing network overload in low-energy scenarios and enhancing the operational reliability of constrained IoT nodes. The results confirm the effectiveness and practical relevance of energyoriented forecasting methods for smart city infrastructures, agricultural IoT deployments, industrial sensor networks, and other resource-sensitive application domains.