Intelligent Workload Prediction and Mobile Task Offloading in Cloud Environments Using Hybrid Deep Learning and Reinforcement Learning

Abstract

Cloud computing provides users with access to data, data services, and a variety of physical computing resources on demand. Such capabilities can be provided based on explicit requests from individuals or organizations to augment their local organizational compute infrastructure by provisioning requested hardware and software from cloud service provider data centers. They can also be based on implicit requests generated by users’ mobile devices. These devices can then benefit from treating all of these resources as one integrated system, here referred to as the digital continuum. There are many issues that should be considered to guarantee seamless and efficient operation of such an ecosystem, but we limit the scope of this work to only two major ones. As such, in this thesis, I focus on predicting workloads, which is critical for efficient data center operations due to the dynamic nature of workloads. I also consider the issue of offloading mobile device tasks to improve user experience and better address processing needs at the edge.

In data centers of cloud service providers (CSPs), inaccurate workload forecasting can lead to poor resource utilization and an unsatisfactory user experience. It can result in suboptimal resource provisioning that wastes energy, breaks service‑level agreements (SLAs), and affects performance. Therefore, it is important to achieve high forecast accuracy for dynamic resource management to optimize reliable operation in data centers despite changing demand. At the same time, mobile devices have become relatively powerful small computers that can run more applications that are sensitive to delay. However, these devices will always be limited in energy and processing resources, and some of their tasks may still need to be carefully balanced between local processing, additional capabilities at the edge, and offloading tasks to the data center when it makes sense, thus optimizing mobile task offloading in Mobile Edge Cloud Collaboration (MECC) environments.

Thus, this dissertation first proposes a novel hybrid predictive model to improve workload‑forecasting accuracy. The methodology begins with two stages of signal decomposition via a two-step strategy: First, I apply Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) to extract intrinsic structures and remove noise artifacts from high-dimensional workload traces. Then I apply Variational Mode Decomposition (VMD) to filter out high‑frequency signals. The decomposed signals are then passed to a deep‑learning model comprising parallel 1D Convolutional Neural Networks (CNN1D) and Bidirectional Long Short‑Term Memory (Bi‑LSTM) networks, which will allow for learning short‑, medium‑, and long‑term temporal patterns. This yields adaptive and accurate prediction models for intelligent resource provisioning within data centers.

Addressing the second challenge, the dissertation explores dynamic multi‑objective task offloading within the MECC framework. It does that by formulating the problem as a Markov Decision Process (MDP) designed to jointly minimize delay, reduce energy consumption, and control computational cost. This formulation takes into consideration changing network conditions, different task requirements, and real‑time availability of resources across mobile devices, edge nodes, and the cloud. A deep reinforcement learning (DRL) algorithm is also proposed, which combines the sequential modeling of Gated Recurrent Units (GRUs) with the stable optimization technique of Proximal Policy Optimization (PPO). The outcome is an offloading policy that prioritizes executing tasks locally and, on the edge, and only sends them to the cloud if it is strictly necessary. To maintain secure task execution, we integrate a lightweight Hash‑based Message Authentication Code (HMAC) that allows data integrity to be verified with minimal overhead.

The proposed CVCBM system was shown to improve the accuracy of workload prediction by 73.8% in related work. Moreover, the proposed HPPO-GRU system was shown to enable safe and efficient task offloading, while reducing total computing cost by approximately 7% to related works. Collectively, these advances improve the reliability of distributed computing across mobile, edge, and cloud layers, enabling next‑generation intelligent infrastructure to respond dynamically to modern digital demands.