The rapid expansion of edge-cloud infrastructures and latency-sensitive Internet of Things (IoT) applications has intensified the challenge of intelligent task offloading in dynamic and resource-constrained environments. This paper presents an Adaptive and Intelligent Customized Deep Q-Network (AICDQN), a novel reinforcement learning-based framework for real-time, priority-aware task scheduling in mobile edge computing systems. The proposed model formulates task offloading as a Markov Decision Process (MDP) and integrates a hybrid Gated Recurrent Unit-Long Short-Term Memory (GRU-LSTM) load prediction module to forecast workload fluctuations and task urgency trends. This foresight enables a Dynamic Dueling Double Deep Q-Network \((\text {D}^{4}\text {QN})\) agent to make informed offloading decisions across local, edge, and cloud tiers. The system models compute nodes using priority-aware M/M/1, M/M/c and M/M/ \(\infty\) queuing systems, enabling delay-sensitive and queue-aware decision-making. A dynamic priority scoring function integrates task urgency, deadline proximity, and node-level queue saturation, ensuring real-time tasks are prioritized effectively. Furthermore, an energy-aware scheduling policy proactively transitions underutilized servers into low-power states without compromising performance. Extensive simulations demonstrate that AICDQN achieves up to 33.39% reduction in delay, 57.74% improvement in energy efficiency, and 81.25% reduction in task drop rate compared with existing offloading algorithms, including Deep Deterministic Policy Gradient (DDPG), Distributed Dynamic Task Offloading (DDTO-DRL), Potential Game based Offloading Algorithm (PGOA), and the User-Level Online Offloading Framework (ULOOF). These results validate AICDQN as a scalable and adaptive solution for next-generation edge-cloud systems requiring efficient, intelligent, and energy-constrained task offloading.
