Energy-efficient wireless network control via spatio-temporal deep learning and multi-agent reinforcement learning Authors: Khalil M Abdelnaby, Mohammed A F Al-Husainy, Mohammad O Alhawarat, Mohamed A Rohaim, Khairy M Assar Published: 2026 ABSTRACT AbstractThe rapid expansion of wireless data traffic is placing increasing strain on the energy consumption of current communication networks, intensifying the tension between performance and sustainability objectives. In interference-intensive multiple access scenarios such as power-domain non-orthogonal multiple access (NOMA), energy-efficient optimization is particularly challenging due to the strong coupling between power control and resource allocation decisions. In order to solve this issue, this paper introduces an AI-Enhanced Energy Optimization Framework (AEEOF), which uses deep spatio-temporal learning and reinforcement learning to provide adaptive and energy-aware network control. The proposed framework incorporates a Spatio-Temporal Graph Convolutional Network (ST-GCN) to learn spatial interference relationships and a Gated Recurrent Unit (GRU) to capture temporal traffic dynamics, embedding the resulting representations into a Multi-Agent Deep Deterministic Policy Gradient (MADDPG) controller to support sequential decision-making. Such a design provides the framework with the ability to dynamically distribute power and timing policies based on changes in network conditions. Extensive simulations in a realistic 5G-oriented environment of high interference levels prove the significant performance improvement. The suggested solution can save up to 15% of energy and make the energy use more efficient by about 40%, which is the number of bits delivered per joule. The overall system throughput goes up by 6.25%, the cell-edge user data rate goes up by up to 60%, the fairness goes up by 20%, and the chance of an outage goes down by 70%. A systematic ablation study with three architectural variants validates the individual contribution of each core component — the ST-GCN spatial module, the GRU temporal module, and the MADDPG reinforcement learning controller. Comparative evaluation against conventional orthogonal and non-AI-assisted baselines further supports the effectiveness of the proposed framework within the studied setting. These findings indicate that intelligent spatio-temporal learning is a promising approach for improving energy efficiency and network performance in interference-intensive wireless environments, as demonstrated within the studied 5G-oriented simulation setting.Keywords: Energy-efficient wireless networks, Artificial intelligence, Spatio-temporal learning, Reinforcement learning, Interference management, Resource allocation, Multi-agent systems, Non-orthogonal multiple access, Sustainable communications, 5G and beyondSubject terms: Engineering, Mathematics and computing Introduction The rapid proliferation of Internet of Things (IoT) devices, smart city infrastructure, autonomous systems, and data-intensive applications is imposing increasingly complex demands on modern wireless communication networks. In addition to the support of the high data rates and very low latency, current networks must be able to work in an energy-efficient and environmentally sustainable way, especially as wireless networks are now a leading source of global energy use1,2. This challenge is further compounded by the emerging requirements of 5G-Advanced and early 6G networks, which target massive connectivity, intelligent service delivery, and pervasive automation across heterogeneous deployment environments3–5. Energy efficiency is becoming a central design goal of the next-generation wireless systems, both due to a reduction in cost of operation and due to a more general sense of sustainability and climate concerns6–8. The systematic investigations and systemwide measurements have repeatedly demonstrated that radio access networks, dense deployment of base stations, and dynamic changes of traffic are the leading factors contributing to excessive energy usage in cellular and IoT-enabled systems2,9. As a result, the optimization of the energy consumption and the quality-of-service (QoS) guarantees has become a primary research question in both the physical-layer design, resource allocation, network control, and cross-layer optimization3,10. Classical optimisation methods struggle to manage the growing complexity, non-convexity, and uncertainty of contemporary wireless systems. Time-varying traffic loads, stochastic channel conditions, user mobility, and heterogeneous service demands collectively limit the applicability of static or model-driven approaches in realistic deployment scenarios11,12. Such constraints have encouraged a shift in paradigm to data-driven and learning-based methods that can be driven by dynamic environments with minimum human intervention13,14. The artificial intelligence (AI) and machine learning (ML) have also become important facilitators to intelligent and energy- conscious wireless networks. Recent research shows the usefulness of AI in the optimization of spectrum allocation, power control, user association, and management of interference in complex and uncertain conditions13,14. More specifically, the deep learning and reinforcement learning methods have demonstrated great promise to be able to capture the spatio-temporal dependencies and learn near-optimal control policies directly on network data, without making recourse to overly simplistic analytical models15,16. The capabilities are inherently suitable to the vision of 6G networks as self-optimizing, context-conscious, and sustainability-oriented communication systems4,17. In addition to communications, AI-based energy optimization has been intensively explored in other related fields like smart grids, smart buildings, industrial automation, and renewable energy systems, and this is evidence of its potential to vastly decrease energy usage without impacting system performance18–20. Such interdisciplinary developments also enhance the appropriateness of AI as a unifying construct in dealing with energy efficiency issues across cyber-physical systems, such as wireless communication systems21–23. In the wireless world, resource management of energy efficiency is closely related to radio resource management and multiple access schemes. Non-Orthogonal Multiple Access (NOMA) has been widely investigated as a possible method of enhancing spectral efficiency and supporting massive connectivity by sharing spectral resources, including time-frequency, by multiplexing in the power domain24–27. The systems based on NOMA present special issues of interference coupling, power, and successive interference cancellation that make energy-efficient operation quite difficult, particularly in dense and nonhomogeneous networks16,28,29. Although NOMA cannot be implemented as a compulsory feature of the existing 5G standards, this framework is still an excellent research tool to address more complex interference-aware optimization techniques and stress-testing techniques based on intelligent control actions in highly coupled resource constraints. The recent developments in AI-assisted wireless communication also demonstrate that the intelligent optimization of future networks is highly important. Deep learning methods have also been applied successfully on the physical layer to channel estimation, signal detection, and resource allocation, and have been shown to have better robustness and adaptability than traditional techniques14. Besides, reinforcement learning-related methods have demonstrated the potential to balance energy and quality of service needs in ultra-dense and heterogeneous network deployments15,30. Such advances follow the wider trends in 6G research, which imagine close collaboration between AI and communication system design4,17. Although these have been made, whether in the next generation wireless networks, energy-efficient operation can be achieved is one of the open challenges in wireless networks. Current solutions have a tendency to treat individual components in isolation, and to take a static environment, or ignore the spatio-temporal correlations of real network traffic and channel dynamics2,9. Moreover, most of the learning-based methods have a problem of scalability, lack of generalization, or systematic validation in real operating environments10,31. Such challenges can only be dealt with the aid of combined structures that collectively take into account the temporal evolution, spatial interference patterns, and long-term energy optimization objectives. Based on these observations, this paper will suggest a smart AI-based framework of dynamic energy-efficient resource optimization in higher-level wireless networks. The proposed approach will aim at capturing the dynamics of the complex network, adapting to time-varying conditions, and improving the energy efficiency in a sustained manner without affecting the service performance by utilizing the principles of deep spatio-temporal learning and reinforcement learning. The real-world network conditions are considered, and the structure is tested with the traditional optimization strategies to prove the efficiency, strength, and scalability. The paper deals with the problem of the attainment of the energy-efficient operation of advanced wireless networks with dense deployments of the wireless networks, heterogeneous services, and dynamic traffic conditions. In contrast to current research that deals with the individual elements of optimization, or a simplified setting, the paper presents a compound and intelligence-led structure, which collectively addresses spatio-temporal dynamics, interference coupling, and long-term energy efficiency goals. The key findings of this paper are as follows: Integrated AI-Driven Energy Optimization Framework: An elaborate AI-driven model is put forward in the case of dynamic optimization of resource energy efficiently within advanced wireless networks. The framework combines both the deep spatio-temporal learning and the reinforcement learning so as to make adaptive decisions to the time-varying channel conditions, traffic variation, and interference. As opposed to the traditional model-based procedures, the presented solution learns policies that are optimal in controlling the network with the use of network observations and, thus, can operate in a complex and uncertain environment. Integrated AI-Driven Energy Optimization Framework: An elaborate AI-driven model is put forward in the case of dynamic optimization of resource energy efficiently within advanced wireless networks. The framework combines both the deep spatio-temporal learning and the reinforcement learning so as to make adaptive decisions to the time-varying channel conditions, traffic variation, and interference. As opposed to the traditional model-based procedures, the presented solution learns policies that are optimal in controlling the network with the use of network observations and, thus, can operate in a complex and uncertain environment. Spatio-Temporal Feature Modelling for Wireless Energy Dynamics: A deep spatio-temporal learning architecture is used to capture the inherent time dynamics of the traffic demand and the spatial interaction of inter-user and inter-cell interference. With this design, the model will learn long-term time dependencies and space correlation patterns in a joint manner, which are the most important in energy-aware optimization but most commonly overlooked in current learning-based methods. Spatio-Temporal Feature Modelling for Wireless Energy Dynamics: A deep spatio-temporal learning architecture is used to capture the inherent time dynamics of the traffic demand and the spatial interaction of inter-user and inter-cell interference. With this design, the model will learn long-term time dependencies and space correlation patterns in a joint manner, which are the most important in energy-aware optimization but most commonly overlooked in current learning-based methods. Reinforcement Learning-Based Adaptive Resource Control: The suggested model creates minimization of energy efficiency as a sequential decision-making and uses reinforcement learning to adapt resource allocation and power control strategies. The learning agent obtains a balance between performance and sustainability over longer operational time scales by directly incorporating energy consumption and quality-of-service constraints into the rewarding entity. Reinforcement Learning-Based Adaptive Resource Control: The suggested model creates minimization of energy efficiency as a sequential decision-making and uses reinforcement learning to adapt resource allocation and power control strategies. The learning agent obtains a balance between performance and sustainability over longer operational time scales by directly incorporating energy consumption and quality-of-service constraints into the rewarding entity. Interference-Aware Optimization Under Challenging Multiple Access Conditions: The structure is tested in a representative multiple access scenario with strong interference coupling, which is a review of the energy-efficient optimization. It is important to note that the choice points out that the proposed approach is capable of dealing with complex interference relationships and shows its strength in challenging network conditions. Interference-Aware Optimization Under Challenging Multiple Access Conditions: The structure is tested in a representative multiple access scenario with strong interference coupling, which is a review of the energy-efficient optimization. It is important to note that the choice points out that the proposed approach is capable of dealing with complex interference relationships and shows its strength in challenging network conditions. Comprehensive Performance Evaluation and Reproducibility: There are a number of simulations that are carried out to test the proposed framework against realistic network conditions. The analysis involves rigorous validation steps, convergence tests, and comparison of the test with the traditional optimization baselines. Full experimental configurations, hyperparameter specifications, and dynamics of the learning are given so that one can reproduce them and have a clear understanding of their performances. Comprehensive Performance Evaluation and Reproducibility: There are a number of simulations that are carried out to test the proposed framework against realistic network conditions. The analysis involves rigorous validation steps, convergence tests, and comparison of the test with the traditional optimization baselines. Full experimental configurations, hyperparameter specifications, and dynamics of the learning are given so that one can reproduce them and have a clear understanding of their performances. The rest of this paper will be structured in the following way. Section II is an overview of related literature in energy-efficient wireless communication, AI-based resource optimization, and interference-aware multiple access strategies. Part III presents the conceptual framework of the intended AI-Enhanced Energy Optimization Framework (AEEOF). Section IV includes the system model and the formulated energy efficiency optimization problem. Section V elaborates on the proposed energy optimization structure, which involves the spatio-temporal learning and the multi-agent reinforcement learning frameworks. Section VI assesses the effectiveness of the proposed strategy as a result of large-scale simulations in real network conditions. Lastly, Section VII summarizes the paper and forms the research directions. Literature review The optimization of energy efficiency of wireless communication networks has been broadly studied from various angles, such as the design of the network architecture, radio resource management, and cross-layer optimization. The initial research largely revolved around analytical and heuristic methods of lowering power usage in cellular systems, and primarily on energy-conscious scheduling, adaptive modulation, and sleep modes of base stations9,11. Although these techniques are important in giving a theoretical understanding, their applicability is usually constrained by oversimplifying assumptions that may not reflect the complexity of the modern heterogeneous and ultra-dense networks. As the IoT and smart environments emerged, the focus of research shifted to the associated field of energy efficiency in networking in various areas of application. Research papers on wireless IoT devices and machine-type communications underlined the trade-offs of reliability of communications in terms of latency and energy consumption12,31. Correspondingly, network topologies and wireless access technology performance studies demonstrated that fixed settings cannot be used to address the dynamic needs of modern services32,33, which drives the adaptive and intelligent control measures. Simultaneously with the development of wireless networking, AI-controlled energy optimization has made a big breakthrough in smart grids, in renewable energy systems, and in industrial robotics. In energy consumption prediction, fault diagnosis, and operational optimization of power systems and smart infrastructures, machine learning models have been utilized21,34–36. The use of deep learning and hybrid AI models has proven to be more accurate and resistant to forecasting nonlinear non-stationary data in the energy domain37. These achievements are excellent indicators that data-driven intelligence will be relevant and able to respond to challenging issues of energy management, which motivates similar approaches in wireless communication networks. The AI methods have been gaining popularity in the context of wireless systems to optimize the allocation of radio resources and the control of networks. It has been suggested that reinforcement learning-based models can dynamically adjust the level of power, scheduling decisions, and network configurations depending on time-varying traffic and channel characteristics15,30. These methods allow independent decision-making and lifelong learning and are therefore appropriate in ultra-dense and heterogeneous deployments. Yet, most of the available literature is dedicated to a particular network layer or a small number of scenarios, which limit their generalizability. The concept of NOMA has attracted a significant amount of research interest due to its ability to increase spectral efficiency and its capability to enable high connectivity. Principles of power-domain multiplexing and successive interference cancellation in cooperative and relay-assisted NOMA systems were developed based on the foundational studies24–27. Then, works delved into resource allocation, relay selection, and the management of interference techniques that would enhance the throughput and energy efficiency of NOMA-based networks16,28,29. In spite of these contributions, a majority of the current approaches of NOMA optimization are based on assumptions that are static or use the convex approximation, thus being less versatile to dynamic network conditions. Current research has recently started to combine AI methods and NOMA systems to overcome these limitations. Resource allocation schemes that are based on deep learning have been proposed to manage interference patterns and imperfect channel state information16. However, such strategies are mostly short-term based in terms of performance optimization and fail to explicitly consider the long-term energy efficiency or long-term sustainability goals. In addition, most of the current research on AI-based assistance of NOMA does not include extensive validation in real traffic dynamics and heterogeneous networks. In addition to NOMA, new research has also examined novel wireless technology to enhance energy efficiency, such as reconfigurable intelligent surfaces, wireless power transfer, and cognitive radio networks38,39. Although these technologies have potential prospects of saving energy, their application to practice network structures presents new complexity and control issues. The proper utilization of such technologies would need clever coordination systems that can combine to harm a myriad of system parameters in the long run. At a bigger picture, recent surveys and vision papers on 6G wireless networks underscore the core of AI in supporting sustainable, intelligent, and adaptive communication systems4,5,17. As these works point out, it is important to go beyond the fragmented methods of optimization to a comprehensive approach that combines learning, prediction, and control across the different layers of the network. Nonetheless, numerous of the suggested architectures are just conceptual, with very little concern for the actual algorithmic design as well as empirical validation. Altogether, current literature shows that significant advances in the field of AI-enhanced energy optimization and advanced multiple access methods have been achieved. Nevertheless, there are still serious gaps concerning the integration of spatio-temporal modeling, energy-conscious long-term decisions, and support for realistic wireless conditions. The existing methods focus on one or two parts of the issue, or they are based on limiting assumptions that restrict their application. These constraints drive the creation of combined AI-based models that can collaboratively address the time dynamics, space interference interaction, and energy-saving goals, which is the subject of the current study. Several advanced resource management frameworks have recently emerged that are relevant to the comparative context of this study. Multi-agent proximal policy optimisation (MAPPO) has been applied to cooperative spectrum and power allocation in multi-cell wireless networks, demonstrating decentralised execution with convergence behaviour comparable to centralised genie-aided algorithms40. Attention-based and transformer-enhanced architectures have been proposed for dynamic SIC ordering and power allocation in NOMA systems, leveraging self-attention mechanisms to capture inter-user dependencies and achieve near-optimal fairness without exhaustive search41. Additionally, model-based reinforcement learning frameworks that integrate explicit world models for channel occupancy and traffic dynamics have shown meaningful reductions in sample complexity for wireless channel access problems42. While these approaches represent important recent advances, they differ from the proposed AEEOF framework in one or more of the following respects: they do not jointly model spatial interference topology and temporal traffic dynamics within a unified spatio-temporal representation; they are evaluated in single-agent or simplified multi-agent settings without explicit outage and fairness constraints in the reward design; or they focus on channel access rather than end-to-end energy efficiency optimisation. A comprehensive comparative evaluation of AEEOF against these more recent frameworks constitutes an important direction for future work, as discussed in Section VII. Conceptual architecture of the AI-enhanced energy optimization framework (AEEOF) This paragraph introduces the conceptual design of the proposed AI-Enhanced Energy Optimization Framework (AEEOF), which will help to overcome the energy efficiency challenge that emerges in the dense, interference-limited wireless networks that utilize NOMA. Rather than relying on fixed optimisation rules, AEEOF adopts a learning-based paradigm capable of jointly addressing the spatial interference structure and temporal traffic dynamics that govern energy consumption in contemporary wireless systems32,43,44. The framework creates a perception-decision-adaptation cycle that monitors the network condition and optimizes policies of resource allocation, as well as communicates with the environment to enhance long-term energy performance and Quality of Service (QoS) assurances. As depicted in Fig. 1, AEEOF is structured into four closely coupled functional layers, which together process raw network observations into control actions that are energy efficient. The conceptual description of these layers follows; the mathematical formulation and algorithmic implementation of the layers appear in Sections IV and V. Conceptual architecture of the proposed AI-enhanced energy optimization framework (AEEOF). Network state perception and problem contextualization The AEEOF layer 1 causes the building of a detailed and orderly portrayal of the network state. Dynamic interference coupling, heterogeneous user demands, and time-varying wireless channels provide the main reasons of energy inefficiency in NOMA-based systems. In order to encompass these factors, the framework monitors a multidimensional state vector that comprises channel gains, inter-user interference, user traffic demands, buffer states, and transmission power levels. The framework does not consider users individually, and this is clearly seen in the fact that it explicitly recognizes that users with the same time-frequency resources in NOMA are coupled in terms of interference. This inspires the graphical representation of the network, where nodes are the users or base stations and the edges represent the relationship of interference and channel correlation. This explicit model allows the learning model to reason on collective behavior and space relations that are essential in power allocation, that is, energy conscious. Formal descriptions of the system model, state variables, constraints, and optimization objectives are given in Section IV. Spatio-temporal feature extraction via deep representation learning AEEOF layer two is concerned with a high-level extraction of spatio-temporal features of the raw network state. Consumption patterns of the energy in wireless networks are spatially correlated because of interference effects as well as temporally dependent because of traffic variation, user movement, and channel attenuation. These dependencies should be captured for predictive and proactive energy optimization. The framework makes use of the ST-GCN to have a model of spatial relationships. The ST-GCN works on the built interference graph, which allows the model to incorporate the information shared with the neighboring nodes and learn the effects of local decisions on the overall energy efficiency. This enables the framework to detect dominant patterns of interference and spatial bottlenecks of energy that cannot be detected using traditional learning (based on vectors). The framework combines a GRU network, which is used to operate chains of past network states45,46. GRU captures the short-term variations as well as long-term variations in the demand and channel conditions to the traffic, thus the model forecasts the future operations in the network as opposed to its reaction being myopic. ST-GCN + GRU becomes a single whole spatio-temporal representation that is informed by the changing operational conditions of the network. Multi-agent decision making and energy-efficient control The third tier of AEEOF converts the acquired spatio-temporal representations into action control decisions. Due to the distributed character of wireless networks as well as the interaction of users in NOMA, the multi-agent sequential decision-making process is formulated as an energy-optimizing problem. Every base station or user agent seeks to maximize its local transmission strategy whilst taking into consideration the actions of others. In order to overcome this difficulty, AEEOF uses a Multi-Agent Deep Deterministic Policy Gradient (MADDPG) framework. In centralized training, agents use the information of the global state to acquire cooperative policies that reduce the long-term energy consumption, without violating the QoS constraints, including minimum data rates, outage probability constraints, and fairness requirements. Agents act in a decentralized way during execution, and they take real-time choices by relying on local readings and received policies. Continuous control actions, such as transmission power allocation and scheduling decisions, are the output of this layer, and they have a direct impact on the instantaneous and cumulative network energy efficiency. Section V expounds on the reinforcement learning formulation, reward design, and policy update mechanisms. Environment interaction and performance evaluation loop The last level of AEEOF is the one that defines the feedback loop between the learning actors and the wireless setting. The chosen actions are input into a 5G-NOMA simulator implemented in high-fidelity in MATLAB, which runs simulations of realistic channel propagation, interference dynamics, and traffic dynamics. The environment reacts by changing the network condition and supplying feedback as rewards and performance measures. This is a closed-loop relation that allows the agents to improve their policies on the basis of trial-and-error learning. Measures used in performance assessments are energy efficiency (bits/Joule), overall energy usage, system throughput, user fairness, and outage probability. The framework is evaluated by means of benchmarking against traditional OMA schemes, heuristic schemes of allocating the power of the NOMA, and non-AI baselines. In addition to a systematic ablation study detailed in Section VI-G, where three variants of the framework are assessed by selectively removing the ST-GCN, GRU, and MADDPG modules, respectively, to isolate and quantify the individual contributions of each architectural component to overall performance. Section VI gives detailed results and settings of the experiment. To conclude, the AEEOF offers a unified and scalable framework combining spatio-temporal deep learning with multi-agent reinforcement learning to address the energy efficiency optimization in wireless networks with interference issues. The framework can explicitly model the structures of spatial interference and temporal evolution of traffic, thereby surpassing major limitations of conventional optimization strategies. This conceptual architecture has been used as a blueprint to guide the formal system modeling and algorithm development that is done in the following sections. System model and problem formulation This section shows the system model and the energy optimization problem that is taken into consideration in this work. The aim is to define the wireless environment, interference processes, and energy consumption patterns of the proposed AI-Enhanced Energy Optimization Framework (AEEOF). This mathematical formulation, as stated in this part, forms the basis of the learning-based solution that was formulated in Section V. Network model We refer to a downlink 5G wireless network with a single cell based on NOMA, in which a base station (BS) serves a group of K User Equipments (UEs) sharing time-frequency resource blocks. Users are clustered into NOMA groups, and superposition coding is employed at the transmitter in order to multiplex several users in the same resource block. Let K= {1, 2,…,K} denote the set of users. The wireless channel between the BS and user k is modeled as: where represents the small-scale fading coefficient, is the distance between the BS and user k, and denotes the path loss exponent. The channel state is assumed to vary over time, capturing the effects of user mobility and environmental dynamics. NOMA transmission and interference model In the NOMA framework, the BS transmits a superimposed signal composed of the signals intended for multiple users. Let denote the transmit power allocated to user k, subject to a total power budget . The transmitted signal can be expressed as: Here, denotes transmit power and is the amplitude scaling coefficient that is used to normalize the symbol that is meant for user k. At the receiver side, successive interference cancellation (SIC) is employed. Users are ordered based on their channel gains, and users with stronger channels decode and subtract the signals of weaker users before decoding their own signal. The signal-to-interference-plus-noise ratio (SINR) for user k is given by: where denotes the set of users whose signals are not canceled at user k, and represents the noise power. This interference-coupled structure introduces strong spatial dependencies among users, making energy-efficient power allocation a challenging task and motivating the use of graph-based learning techniques. Achievable rate model Based on the SINR expression, the achievable data rate of user k is computed using the Shannon capacity formula: where B denotes the system bandwidth. The achievable rates depend jointly on channel conditions, power allocation, and interference levels, all of which vary dynamically over time. Energy consumption model The total energy consumption of the network consists of both transmission power and static circuit power. It is modeled as where represents the circuit power consumption associated with signal processing, baseband operations, and hardware components at the BS. This model captures the trade-off between increasing transmit power to enhance throughput and minimizing energy usage to improve sustainability. Energy efficiency metric and QoS constraints The network energy efficiency (EE) is defined as the ratio of the aggregate system throughput to the total energy consumption: To ensure acceptable service quality, the optimization problem is subject to the following constraints47,48: Minimum Rate Constraint Minimum Rate Constraint where is the minimum required data rate for user k. Power Budget Constraint Power Budget Constraint Non-Negativity Constraint Non-Negativity Constraint Fairness Constraint Fairness Constraint A fairness constraint based on Jain’s fairness index may also be imposed to prevent excessive performance disparity among users. Energy Efficiency Optimization Problem Energy Efficiency Optimization Problem The energy efficiency optimization problem can now be formulated as49–51: s.t. The problem is a non-convex problem by its nature, as it contains a fractional objective function and a coupled term of interference, as well as dynamic channel variations. In addition, the stochastic character of wireless environments and the fact that decisions should be made in real-time make conventional optimization techniques computationally infeasible 52-54. The components of this optimisation problem — specifically the energy efficiency objective in Eq. (10), the minimum rate constraint in Eq. (11), and the fairness requirement — are directly encoded into the reward function of the reinforcement learning controller, as formalised in Eq. (15) of Section V-D. This alignment ensures that the learning agents optimise a reward signal that is mathematically consistent with the stated optimisation problem. Motivation for an AI-driven solution The formulated problem is quite complex and non-stationary, which means that a learning-based solution is needed in order to acquire scalable and adaptive energy optimization. This inspires the suggested AI-Enhanced Energy Optimization Framework (AEEOF), which uses the approaches of spatio-temporal deep learning and multi-agent reinforcement learning to acquire near-optimal policies of power allocation by virtue of interacting with the wireless environment. Section V has the detailed architecture and learning algorithm. Proposed AI-driven energy optimization framework This section provides the suggested AI-Enhanced Energy Optimization Framework (AEEOF) as a framework that combines spatio-temporal deep learning and multi-agent reinforcement learning to facilitate adaptive and power-efficient interference-intensive wireless network control. The structure would meet the non-stationary and highly coupled nature of power allocation decisions and scheduling decisions in the NOMA-based systems. AEEO allows improvement in energy efficiency through proactive optimization of spatial interference structures and temporal traffic dynamics, with Quality of Service (QoS) guarantees, by learning the relationship between these two factors jointly. Overview and design rationale The essence of the problem on energy efficient optimization of wireless networks is the fact that users sharing radio resources are highly connected, i.e., in the power domain, non-orthogonal access. Conventional optimization methods can only poorly address the non-convexity that arises, as well as dynamic uncertainty, whereas the solutions that are static or rules-based are not adaptable. AEEOF attempts to address these drawbacks by posing the energy optimization as a series of sequential decision-making and resolving the problem with learning-based methods. The framework is based on three major design principles. The first is that space awareness is required to model the relationship of interference between users sharing the same spectrum. Second, there should be a sense of time so as to record the progression of traffic and the variability of channels with time. Third, adaptive decisions are required to keep on optimizing power and scheduling policy based on the realized results. In order to meet such demands, AEEOF incorporates a ST-GCN and a GRU into a multi-agent deep deterministic policy gradient (MADDPG) architecture. The general architecture is the closed-loop interaction as shown in Fig. 2, whereby the learned policies interact continuously with the wireless environment. Architecture of the proposed AI-driven energy optimization framework. Network state representation and data generation The wireless network is defined at every decision interval by a high-dimensional state vector that represents physical-layer as well as traffic-level information. The state representation covers gains of channel, level of inter-user interference, signal-to-interference-plus-noise ratios (SINRs), demand of traffic per user, occupancy of buffer, and the current power allocation rates. In order to explicitly model interference coupling, they model the network state as a graph, with users or base stations as nodes and the strength of interference and channel correlations as edges52–54[55-57]. The high-fidelity MATLAB-based 5 G-oriented NOMA simulator is used to generate training data. The simulator gives a real-life channel propagation, fading dynamics, user movement, and bursty traffic models. The combination of network states, actions, and the resultant performance measures is gathered per simulation episode. The produced data is brought to normalcy and arranged into a time series to allow repeated learning. This simulated model helps ensure that there is controlled. experimentation and, at the same time, offering enough diversity to facilitate sound learning of policies. Spatio-temporal feature learning (ST-GCN + GRU) In order to derive informative representations of the structured network state, AEEOF tries to use a two-stage spatio-temporal feature learning module. A spatio-temporal graph convolutional network (ST-GCN) is used to implement the spatial feature extractor. The ST-GCN uses the information of the neighboring nodes directly on the interference graph; it learns the value of the local power allocation decisions to the global interference and energy consumption. This allows the framework to find spatial energy bottlenecks and dominant interference paths which are not modeled through traditional vector-based models. The interference graph is constructed dynamically at each decision interval. The node set corresponds to the K active user equipment’s in the cell. A directed edge is established between users i and j if the interference power caused by user i at the receiver of user j exceeds a predefined threshold ε, i.e., , where is the cross-channel coefficient. The edge weight is set proportional to the instantaneous interference-to-noise ratio (INR) between the two users: . The graph is updated at each time step to reflect real-time changes in channel conditions, user positions, and power allocation states, ensuring that the ST-GCN operates on a current and accurate representation of the interference topology. The implementation of the temporal feature extractor is based on the GRU network. GRU processes time series of states of historical network and learns both the short-term dynamics and long-term traffic patterns. It is such a time modelling capacity that enables the framework to predict future network situations and proactively modify control measures instead of responding to them45,58,59,60. The outputs of the ST-GCN and GRU modules are combined to create a single spatio-temporal latent representation, and that is the input of the decision module of the reinforcement learning. Two-stage learning pipeline The overall learning procedure of AEEOF follows a two-stage design. In the first stage, the ST-GCN–GRU module is pre-trained in a supervised fashion using labelled samples generated by the high-fidelity MATLAB-based 5G-NOMA simulator. Each sample consists of a network state snapshot — capturing channel gains, inter-user interference levels, SINRs, per-user traffic demands, buffer occupancies, and current power allocation levels — paired with a ground-truth label derived from the simulator’s known optimal control response under that state. The supervised objective is to train the ST-GCN–GRU encoder to produce latent representations that reliably identify and distinguish energy-critical network states, including congestion-prone, interference-dominated, and power-inefficient configurations, that require adaptive control intervention. The classification metrics reported in Section VI-C (accuracy, precision, recall, F1-score) quantify the reliability and consistency of this representation learning stage. In the second stage, the pre-trained ST-GCN–GRU module serves as a fixed state encoder, and its output latent representations are passed as state inputs to the MADDPG reinforcement learning agents. The Reinforcement Learning (RL) stage then learns sequential power allocation and scheduling policies entirely through trial-and-error interaction with the wireless environment, using the encoded state representations as the basis for policy gradient computation. This two-stage design ensures that the RL agents operate on structured, discriminative state abstractions, which reduces noise in policy updates and promotes stable convergence of the multi-agent learning process. Multi-agent reinforcement learning formulation (MADDPG) Considering the distributed and interference-coupled nature of the NOMA systems, the energy optimization problem is defined as a multi-agent reinforcement learning (MARL) problem. Each agent is associated with a manageable network entity, like a base station or a group of users, and makes decisions on continuous control, such as the transmission power levels and the scheduling decisions. Each learning agent in the MADDPG framework corresponds to the base station managing a distinct NOMA user cluster. The local observation space of agent i at time step t is defined as: where is the channel gain, is the signal-to-interference-plus-noise ratio, is the aggregate interference level, is the traffic demand, and is the buffer occupancy of user k belonging to cluster . The continuous action of agent i is a vector of per-user power allocation coefficients , subject to . During centralised training, each critic network additionally receives the joint observation and action vectors of all agents to enable cooperative policy learning. The Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm is adopted by AEEOF. In training, the agents use centralized critics, who have the information about the global state, with which they can coordinate their learning and achieve converged learning. In the implementation, the agents act in a decentralized fashion based on observations that are locally available, and as such, the framework can be utilized in real-time implementation. The space of actions is continuous, as the control of power is fine-grained, whereas the policy networks transform the representations of the state space (spatio-temporally) to those of an optimal control. This model allows the agents to train collaborative techniques, reducing the long-run energy use, and meets the QoS limitations. Reward function and policy optimization The rewarding feature is intended to directly relate the goal of the energy efficiency and punish the breach of service limitations. The reward function is designed to directly operationalise the energy efficiency objective while penalising violations of QoS constraints. At each discrete time step t, the scalar reward signal is defined as: where is the instantaneous network energy efficiency in bits/Joule; is the instantaneous network outage probability; is the mean user data rate; is the minimum per-user rate requirement; F(t) is Jain’s fairness index computed over the current user rate distribution; and are weighting coefficients that govern the relative importance of each objective component. Actions that maximize delivered data per unit energy are rewarded, while actions that cause outages, rate violations, or unfair resource distribution incur proportional penalties. The weighting coefficients were tuned empirically on the validation set and their final values are reported in Table 2. Table 2AEEOF training hyperparameters and network architecture.ParameterValue Spatio-Temporal Feature Learning Module ST-GCN layers3 graph convolutional layers (64, 128, 64 units)GRU hidden units128 units, 2 stacked layersDropout rate0.3 (applied to ST-GCN and GRU layers)L2 regularisation coefficient1 × 10⁻⁴Pre-training dataset size150,000 samples (70/15/15 split)Early stopping patience15 epochs (validation loss) MADDPG Reinforcement Learning Controller Actor network3 FC layers (256, 128, 64 units), ReLU activationCritic network3 FC layers (256, 128, 64 units), ReLU activationActor learning rate1 × 10⁻⁴Critic learning rate3 × 10⁻⁴Discount factor (γ)0.95Batch size256 transitionsReplay buffer size10⁶ transitionsTarget network update rate (τ)0.01 (soft update)Exploration noiseOrnstein–Uhlenbeck: σ = 0.2, θ = 0.15, linear decay to σ = 0.01Training episodes2,000Time steps per episode200 Reward Function Parameters Energy efficiency weight (η)1.0Outage penalty weight (λ₁)0.5Rate violation penalty weight (λ₂)0.8Fairness penalty weight (λ₃)0.3 Computational Configuration Training hardwareNVIDIA A100 GPUSimulation platformMATLAB R2023b/SimulinkApproximate training duration~ 6 h (~ 1.5 h ST-GCN–GRU; ~4.5 h MADDPG) AEEOF training hyperparameters and network architecture. Algorithmic flow and computational complexity The general algorithmic procedure of AEEOF is presented in Fig. 3. At every episode, the environment is launched, and the agents are in the initial network state. The spatio-temporal feature learning module commits the state, the MADDPG agents choose control actions, and the environment updates states and rewards. The agents update their policies, then, according to the experience gained. Integrated ST-GCN–GRU–MADDPG Framework for Energy-Efficient NOMA Control. Computationally, the overwhelming complexity is due to the ST-GCN and GRU modules and the policy network updates. Although the training phase is more expensive to compute, inference at execution is only forwarding of trained networks, and thus, real-time deployment is possible. The decentralized implementation with the centralized training paradigm also contributes to the scalability as the size of the network grows. The entire operation flow of the proposed AI-Enhanced Energy Optimization Framework (AEEOF) that combines spatio-temporal deep learning and multi-agent reinforcement learning to optimize energy efficiencies in interference-intensive NOMA-based wireless networks is outlined in Algorithm 1. Based on the principle of centralized training and decentralized implementation, the algorithm allows cooperative learning and, at the same time, provides scalability and relevance in real-time. The algorithm starts with the phase of initialization, during which each learning agent initializes its actor and critic networks. The actor network produces unceasing control activities, i.e., power levels of transmissions, timing choices, and the critic network determines the long-term worth of the activities. Also, target networks are initialized to stabilize training. A common replay buffer is developed to save previous experiences so that off-policy learning is achieved, along with better sample usage. Also, the original interference graph is built according to channel conditions and user proximity, explicitly modeling the relationship between spatial interferences. The algorithm runs in the course of training, and during each episode, it is evaluated as an independent realization of network dynamics. The wireless environment is perturbed to the initial state at the start of every episode, and the initial network state is monitored. The network state is fed into the spatio-temporal feature learning module as the algorithm operates on time steps. In particular, the ST-GCN propagates spatial features by incorporating a summary of information on adjacent nodes in the interference graph, which captures the propagation of interference among users using the same spectrum. Meanwhile, the GRU takes the sequences of the states over time to learn the temporal relation that would be based on traffic and channel changes. The spatial and temporal features are then combined into a single latent representation, which is the input to the multi-agent decision-making module. In this representation, agents choose a continuous action based on their actor network, and exploration noise is introduced to the training to motivate policy exploration. Joint action vector is applied to the wireless setting, leading to the re-estimation of network states and a scalar reward signal that jointly measures network energy efficiency and Quality of Service (QoS) compliance. Training strategy summary The complete training procedure of AEEOF proceeds in two sequential stages. In Stage 1, the ST-GCN–GRU module is pre-trained in a supervised fashion using 105,000 labelled network state samples. Training employs mini-batch gradient descent with the Adam optimiser, dropout and L2 regularisation for overfitting prevention, and validation-based early stopping to select the best-generalising model checkpoint. In Stage 2, the pre-trained ST-GCN–GRU encoder is frozen and its output representations are used as state inputs to the MADDPG controller. The MADDPG agents are trained under the centralised training with decentralised execution (CTDE) paradigm: during training, centralised critics access the joint observation and action vectors of all agents to support cooperative policy learning, while actors rely only on local observations. Policy updates are performed using sampled mini-batches drawn uniformly from a shared experience replay buffer, with soft target network updates applied after each gradient step to stabilise training. Exploration is governed by Ornstein–Uhlenbeck noise added to the actor outputs, with linear decay across training episodes to transition from exploration-dominant to exploitation-dominant behaviour. All hyperparameters governing both training stages are consolidated in Table 2 for reproducibility. Simulation results and performance evaluation This part includes a critical analysis of the suggested AI-Enhanced Energy Optimization Framework (AEEOF) with detailed simulations carried out in a 5G-oriented wireless network environment with high fidelity. The evaluation addresses four complementary objectives: (i) verifying the convergence and stability of the ST-GCN–GRU supervised pre-training stage through training and validation loss analysis; (ii) verifying the convergence of the MADDPG reinforcement learning controller through episode vs. cumulative reward analysis; (iii) quantifying the network-level performance gains of the complete AEEOF framework across key metrics including energy consumption, throughput, fairness, and outage probability; and (iv) comparing the proposed framework against conventional orthogonal and non-AI-assisted baselines under identical operating conditions, supported by rigorous statistical significance testing. All the findings are achieved by using large-scale Monte Carlo simulations of user distributions and channel realizations with randomized user dissemination and channel realizations to guarantee statistical rigor and reproducibility, and by subsequent tests of formal statistical significance. The subsections that follow describe the experimental framework, dynamics of learning, analysis of comparative performance, scalability behavior, and robustness of the proposed framework, and hence present a complete evaluation of the effectiveness of the proposed framework in wireless networks that are limited by interference. Experimental testbed and simulation configuration In order to critically assess the performance, resilience, and scalability of the proposed AI-Enhanced Energy Optimization Framework (AEEOF), a high-fidelity MATLAB-based digital twin of a 5G-based wireless network with a heavy load of interference was simulated. A simulation environment has been created with MATLAB and Simulink with the support of Deep Learning Toolbox and Reinforcement Learning Toolbox, which allows the easy combination of the components of the spatio-temporal deep learning and multi-agent reinforcement learning. Since the physical layer and propagation characteristics are realistically presented in terms of Table 1, they are in accordance with the 3GPP Technical Report TR 38.901 specifications of urban macro-cell (UMa) models and are well adjusted to the modern cellular standards. One isolated hexagonal cell at a radius of 800 m is taken into account at a carrier frequency of 3.5 GHz and an aggregate system bandwidth of 100 MHz and subcarrier separation of 30 kHz. These parameters represent a realistic 5G deployment in the middle band, and enable the direct evaluation of the energy use tradeoffs in realistic spectral and interference environments31,58,59. Table 1Core simulation parameters for the AEEOF testbed.ParameterValueDeployment ScenarioUrban Macro-Cell (UMa), isolated hexagonal cellCell Radius800 mCarrier Frequency3.5 GHzSystem Bandwidth100 MHzSubcarrier Spacing30 kHzUser DistributionClustered (Thomas Process)Mobility ModelCorrelated Random WalkMultiple Access SchemeDownlink Power-Domain NOMASIC ModelImperfect SIC with residual interferenceChannel Model3GPP TR 38.901 UMa + CDL-CBase Station Power ModelNonlinear PA with sleep modeTraining HardwareNVIDIA A100 GPUSimulation PlatformMATLAB/Simulink Core simulation parameters for the AEEOF testbed. The locations of user equipment (UE) are created as a clustered spatial point process to create hotspot-driven traffic patterns that are commonly found in urban high-density areas. A correlated random walk is used to model the temporal user mobility, both short-term and long-term trends of movement. Power-domain NOMA is used, where the time-frequency resources are shared among multiple users and differentiated by assigning them different power levels. In order to make it more realistic, imperfect SIC is modelled explicitly, considering the residual interference and decoding errors. The wireless channel takes into consideration path loss based on distance, log-normal shadowing, and small-scale fading based on the clustered delay line (CDL-C) model as developed in 3GPP TR 38.901. The model of base station power consumption is based on a nonlinear power amplifier profile with sleep-mode capabilities, which allows the framework to utilize the transmission-level and hardware-level energy-saving. Monte Carlo simulation using over 100,000 independent realizations of the channel used was done in each of the experimental configurations to ensure statistical reliability. Any reported results are averaged across at least 1,000 independent simulation runs with randomized user drops, traffic realizations, and channel conditions. AEEOF is implemented on the computational level to take advantage of distributed processing on multi-core CPUs and NVIDIA A100 GPUs to effectively train the spatio-temporal ST-GCN-GRU module and the multi-agent deep deterministic policy gradient (MADDPG) controller. This configuration will have the characteristics of training times, convergence behavior, and scalability that will be reflective of real deployment scenarios. To support full reproducibility of the reported results, Table 2 provides a comprehensive summary of all training hyperparameters, network architecture specifications, and computational configuration parameters used in the implementation of AEEOF. The training follows a two-stage procedure: supervised pre-training of the ST-GCN–GRU feature learning module, followed by MADDPG reinforcement learning controller training under the centralised training with decentralised execution paradigm described in Section V. All hyperparameters were finalised prior to the test set evaluation reported in Sections VI-B through VI-G, ensuring that no test set information influenced the configuration of the framework. Training dynamics and convergence analysis In order to define the reliability, stability, and overall generalization of the proposed AI-Enhanced Energy Optimization Framework (AEEOF), the training dynamics and convergence behaviour are analyzed in detail. With the complexity and hybridism of the ST-GCN-GRU representation unit and implementing it with a multi-agent reinforcement learning architecture, the stability of learning must be explicitly checked out first before any of the gains of the system-level performance can be interpreted. Dataset partitioning and evaluation protocol The 5G-NOMA digital twin in Sections III and IV was used to generate the dataset that was used to train and evaluate the spatio-temporal learning aspects of AEEOF. The samples of network states that were labelled were 150,000, all of which were snapshots of channel conditions, interference topology, traffic demand, and optimal control responses. To achieve adherence to modern IEEE requirements of machine learning-based communication systems, the dataset was randomly shuffled and divided into three subsets that do not overlap: 70% (105,000 samples) for training, 70% (105,000 samples) for training, 15% (22,500 samples) for validation, 15% (22,500 samples) for validation, 15% (22,500 samples) for independent testing. 15% (22,500 samples) for independent testing. The training set was only used to update model parameters; the validation set assisted in the decision-making process concerning hyperparameter tuning and early stopping, and the test set was absolutely isolated and not accessed during the training process. This protocol allows each measure of learning to be a true generalization and not a rote learning of data. Learning dynamics and convergence behaviour The training of the ST-GCN-GRU module was observed through the loss during the training as well as during the validation by epoch, and this is shown in Fig. 4. The loss function indicates the difference between the predicted and target spatio-temporal representations that dictate further choices in reinforcement learning. Training and validation loss curves across epochs. In the first training stage, loss curves are decreasing rapidly and monotonically, which predicts the effective extraction of the fundamental spatial interference structure and temporal traffic pattern. On the curves, the convergence rate decreases gradually as the training proceeds, and the curves converge towards an identical minimum. Notably, the validation loss follows the training loss significantly during the whole training process, and there is no apparent deviation or late-stage escalation. This property gives excellent empirical support that the model is converging to a stable solution, and it is not over-fitted as a result of its architectural complexity. The lack of fluctuations or instability also proves that the optimization procedure is numerically strong and the choice of the learning rate and batch size is appropriate. Figure 4. Training and Validation Loss Curves Across Epochs is direct and unambiguous evidence that the presented performance improvements of AEEOF are obtained with a well-converged, well-regularized, and reproducible learning process. The absence of overfitting and the smooth convergence justify the use of the ST-GCN -GRU architecture as a strong spatio-temporal feature extractor of energy-conscious decision-making in the interference-prone wireless networks. More importantly, the analysis will form a firm basis for the network-level evaluations in the subsequent subsections, as any future progress in energy efficiency, throughput, fairness, and outage probability can be explained by meaningful learning in place of training instability or data leakage artifacts. Early stopping and model selection An early stopping mechanism based on the validation was used to avoid over-training and an optimal bias-variance trade-off. Training was stopped automatically because the validation loss was not decreasing after 15 epochs, and the model parameters associated with the minimum validation loss were deployed. The approach will ensure the final model employed in the AEEOF reflects a solution that is most generalizable instead of a training endpoint that was picked arbitrarily. It is also consistent and reproducible, even in independent training runs. MADDPG policy convergence analysis While the ST-GCN–GRU loss curves in Fig. 4 demonstrate the convergence of the supervised pre-training stage, a complete evaluation of the AEEOF training pipeline requires separate assessment of the MADDPG reinforcement learning controller’s learning convergence. Figure 5 presents the episode vs. cumulative reward curve of the MADDPG agents recorded over the full 2,000 training episodes, with a 50-episode moving average overlaid on the raw per-episode values. MADDPG learning convergence curve showing episode vs. cumulative reward averaged across all agents over 2,000 training episodes. The bold line represents a 50-episode moving average smoothed over the raw per-episode values (shown in light shading). The vertical dashed line marks the approximate onset of policy convergence. The convergence trajectory in Fig. 5 exhibits three distinct phases consistent with well-behaved multi-agent deep reinforcement learning dynamics. In the initial exploration phase (approximately episodes 1–400), the cumulative reward is low and exhibits high variance, reflecting the agents’ reliance on Ornstein–Uhlenbeck exploration noise and the absence of a developed policy. During this phase, the agents gather diverse experience across the state-action space, populating the replay buffer with samples that cover a broad range of network conditions. In the policy improvement phase (approximately episodes 400–1,400), the cumulative reward increases rapidly and monotonically as the agents exploit the structured spatio-temporal state representations provided by the pre-trained ST-GCN–GRU encoder to identify increasingly effective power allocation and scheduling strategies. The rate of improvement during this phase reflects the quality of the state representations — the discriminative features learned during supervised pre-training enable the agents to associate meaningful state abstractions with high-reward actions more efficiently than would be possible with raw state inputs. In the convergence phase (approximately episodes 1,400–2,000), the moving average stabilises within a narrow band, indicating that the agents have reached a stable policy equilibrium. The persistence of moderate episode-level variance in this phase is attributable to the stochastic nature of the wireless environment — including random channel realisations, user mobility, and bursty traffic — rather than instability in the learned policies. This distinction is confirmed by the narrow confidence interval of the mean energy efficiency distribution reported in Section VI-E, which demonstrates low cross-run variance in the final performance of the converged policies. Taken together, Figs. 4 and 5 provide a complete two-part convergence analysis covering both the supervised pre-training stage and the reinforcement learning control stage of the AEEOF training pipeline, confirming that the reported performance improvements reflect fully converged, stable, and reproducible learned policies. Regularization and overfitting mitigation Since the spatio-temporal learning module has a large representational power, various regularization methods were systematically combined into the training pipeline. To avoid co-adaptation of neurons, dropout regularization was applied to both the temporal GRU layers and the spatial graph convolution layers. Moreover, L2 weight regularization was applied to all trainable parameters to deter large weight values and smooth decision boundaries. The combined effect of dropout regularisation, L2 weight regularisation, and validation-based early stopping is clearly illustrated in Fig. 4, where the training and validation loss curves remain closely aligned throughout the entire optimisation process. This proves that the learnt representations are strong and can be transferred across unknown network conditions. Decision reliability and learning stability analysis The regularisation strategy applied during training (comprising dropout on both the GRU and ST-GCN layers and L2 weight regularisation on all trainable parameters) is described in Section VI-B-iv and contributes directly to the stability and generalisability of the learned representations evaluated here. Evaluation objective and metric interpretation In contrast to standard supervised classification problems, the learning goal of AEEOF is not to classify continuous network states (high-dimensional spatio-temporal) into discrete ones, but to convert high-dimensional spatio-temporal network states into useful continuous control actions. The spatio-temporal feature extractor (ST-GCN-GRU) is, however, evaluated against auxiliary decision-consistency measures during training and validation to determine the accuracy and consistency of the learned representations by the feature extractor. In particular, the accuracy, precision, recall, and F1-score, which are reported, are used to measure the prediction capabilities of the model to identify and distinguish energy-critical network states, including congestion-prone states, interference-dominated states, and power-inefficient allocations, that need adaptive control responses. These measures are hence indicators of the reliability of decisions, as opposed to traditional classification accuracy. Quantitative decision reliability results The proposed AEEOF achieves: 94% decision accuracy, indicating a high level of consistency between predicted and target energy-aware control responses; 94% decision accuracy, indicating a high level of consistency between predicted and target energy-aware control responses; 95% precision, demonstrating that the model rarely triggers unnecessary or suboptimal power reconfiguration actions; 95% precision, demonstrating that the model rarely triggers unnecessary or suboptimal power reconfiguration actions; 95% recall, confirming that the vast majority of energy-saving opportunities are correctly identified; 95% recall, confirming that the vast majority of energy-saving opportunities are correctly identified; 96% F1-score, reflecting an optimal balance between conservative and aggressive control decisions. 96% F1-score, reflecting an optimal balance between conservative and aggressive control decisions. All the metrics reported are only computed on the held-out test set, averaged over several independent training runs to reduce stochastic influences in weight initialization and mini-batch sampling. The low variance of the results in the different runs is another indicator that the learned representations are stable. Implications for reinforcement learning control The high decision reliability of the ST-GCN–GRU module provides the MADDPG reinforcement learning controller with a stable and discriminative foundation for policy learning. Specifically, the quality of the learned state representations ensures that: (i) policy gradients are computed over meaningful state abstractions, reducing noise in actor network updates; (ii) critic networks converge more smoothly, since state-action value estimates remain consistent across training steps; and (iii) exploration is better directed, avoiding unnecessary oscillations in power allocation and scheduling decisions. Consequently, the reinforcement learning agents are able to focus on long-term energy optimisation rather than compensating for representational errors introduced by an unreliable state encoder. The strong quantitative reliability metrics reported above (94% accuracy, 95% precision, 95% recall, and 96% F1-score) provide direct empirical support for this conclusion. Network performance gains: before and after optimization With the reliability and stability of the acquired spatio-temporal representations established in the previous section, VI-C, we now consider the practical network-level influence of the implementation of the proposed AI-Enhanced Energy Optimization Framework (AEEOF). This sub-section is devoted to before-and-after performance analysis, which illustrates how reliable decision-making is converted into real energy, throughput, coverage, fairness, and outage probability benefits in a realistic 5G-based NOMA setting. To make the entire evaluation clear and complete, there are several complementary figures and summary tables that demonstrate each of the different operational aspects of the network. Aggregate network performance comparison Figure 6 shows the overall results of the implementation of the proposed AI-Enhanced Energy Optimization Framework (AEEOF) regarding the comparison of important network performance indicators before and after optimization. The figure brings out three basic aspects of network operation: energy consumption, signal quality, and coverage capability. Network performance comparison before and after deployment of the AI-Enhanced Energy Optimization Framework (AEEOF). Figure 6 depicts that the implementation of AEEOF will lead to a decrease in the total energy consumption by 15%, where the total energy consumption will drop to 42 J. This enhancement is not obtained by simply being naive about power reduction, but by being smart and context-sensitive power management because it considers spatial interference relationships as well as temporal traffic dynamics. The framework allows removing the wasteful transmissions by predicting traffic changes and alleviating inter-user interference without compromising the reliability of services. At the same time, the mean received signal strength is raised by around 28.6%, with the figure reaching 70% and 90%. This improvement is attributed to the performance of the spatio-temporal learning architecture in detecting positive transmission opportunities and redistributing power to users and time slots with which it can produce the greatest marginal benefit. This improvement indicates that signal quality does not suffer in the process of increasing energy efficiency; rather, the two targets are enhanced simultaneously. However, notably, the effective coverage rate is enlarged by 30%, as the proportion grows to 78%. What it means is that AEEOF provides a high level of reliability within a significantly extended geographical coverage without the need to deploy extra infrastructure. The coverage enhancement is especially significant since coverage expansion normally involves increased transmit power that is incompatible with the objectives of energy conservation. The suggested architecture mitigates this drawback through the utilization of interference-aware scheduling and predictive resource allocation, which enables the network to provide more efficient service to distant and disadvantaged users. Combined with Figs. 4 and 6 gives solid empirical evidence that AEEOF has been able to overcome the traditional trade-off between energy efficiency and network performance. The framework is able to provide coordinated gains on energy use, signal quality, and coverage rather than maximizing each of these areas independently. These findings support the central hypothesis of this study that intelligent spatio-temporal and reinforcement learning-based control can meaningfully expand the performance operating region of an interference-limited wireless network within the studied simulation conditions. Spatial energy efficiency analysis The spatial distribution of energy consumption of the cell coverage area among users’ distance to the base station is shown in Fig. 7. This discussion offers a critical view of the manner in which the proposed AI-Enhanced Energy Optimization Framework (AEEOF) handles energy efficiency on aggregate and spatial levels, especially to users that are located at the cell edges, which are the traditional consumers of the most power. Energy consumption as a function of user distance from the base station before and after applying the proposed AI-Enhanced Energy Optimization Framework (AEEOF). As anticipated, energy is used more as distance grows in both cases because of path-loss attenuation and the need to use more transmit power to ensure the links are reliable. Nevertheless, it has been discovered that a steady and significant decrease in energy expenditure is noted at all ranges following the implementation of AEEOF. The consistent separation between the two curves across all user distances confirms that AEEOF delivers energy efficiency gains throughout the entire cell coverage area, rather than concentrating improvements only among users near the base station. The energy saving as a percentage is stronger at large distances. At 800 m away from the base station, the energy used is lower than around 39.5 J to 33.4 J, which is a considerable consumption, in the most energy-consuming area of the cell. This is an indication that AEEOF is especially efficient in reducing the unnecessary power consumption that is normally attributed to cell-edge communication. The main feature of this improvement is the ST-GCN element’s ability to model spatial interference, which lets the model accurately get the inter-user relations of interference and stop power from going up unnecessarily. The framework doesn’t just raise the power level to deal with channel inefficiencies; it also trains allocation schemes that are aware of interference so that communication can happen reliably at a lower energy cost. Also, the GRU’s ability to model time lets the system predict changes in traffic and stop transmissions that are reactive and use a lot of energy. Three interconnected factors explain why cell-edge users benefit more substantially from AEEOF’s energy efficiency gains than cell-centre users. First, cell-edge users have worse path-loss conditions and more interference from other users who are using the same time-frequency resources in the NOMA framework. Without intelligent control, the usual way to deal with these situations is to turn up the transmit power, which wastes a lot of energy. By explicitly modeling the interference graph topology, the ST-GCN module lets the framework find the specific interference relationships that affect each cell-edge user and give them power in a structured and targeted way, which means that power amplification is not needed. This power allocation that takes interference into account saves more energy for cell-edge users because their baseline energy use without optimization is much higher than average. Second, cell-center users already benefit from good channel conditions, short propagation distances, and low levels of interference. This means that even without smart optimization, they use relatively little energy, which is close to the theoretical minimum. So, cell-center users have less room for improvement, while cell-edge users have a lot more room for improvement that the framework can take advantage of. Third, the GRU temporal modeling part helps cell-edge users by predicting times of high traffic demand and changing resource allocation before congestion happens. Cell-edge users are more susceptible to traffic-induced outages and reactive power increases. Proactive prediction and management of these conditions therefore yields a proportionally greater energy benefit for cell-edge users, compared to cell-centre users whose favourable channel conditions allow reactive management at relatively low energy cost. These three mechanisms explain why AEEOF’s energy efficiency gains aren’t the same everywhere and are always stronger the farther away from the base station. The slope of the optimized curve is further flatter, which further demonstrates the higher scalability of energy with distance. The property of this feature is a more balanced policy of distributing spatial resources that increases fairness by avoiding the disproportionate use of energy by the remote users. Thus, the spatial discrimination is avoided to bring about the energy saving in Fig. 7, which strengthens the fairness improvements mentioned in the rest of the section. Generally, Fig. 7 suggests very strong results that the suggested AEEOF offers spatially uniform energy efficiency benefits, which tackles one of the most complex issues of the wireless networks with interference limitation: energy-efficient service delivery to cell-edge users. Throughput, fairness, and outage improvements Figure 8 presents a holistic comparison of the proposed AI-Enhanced Energy Optimization Framework (AEEOF) against three representative baseline schemes using a radar chart that captures multiple key performance indicators (KPIs) simultaneously. This visualization provides an intuitive assessment of the trade-offs inherent in each approach and highlights the overall operational balance achieved by AEEOF. Normalized key performance indicators for OMA, NOMA with fixed power allocation, NOMA-DQN, and the proposed AEEOF framework. The radar chart reveals that the AEEOF polygon dominates all competing schemes across every evaluated dimension, indicating Pareto superiority. In terms of energy efficiency, AEEOF achieves the highest normalized score, reflecting its ability to minimize energy expenditure through interference-aware power allocation and proactive traffic prediction. The significant margin over both NOMA-FPA and NOMA-DQN underscores the importance of joint spatio-temporal learning rather than static or purely reactive control. For system throughput, AEEOF maintains near-optimal performance, exceeding OMA and NOMA-FPA while closely approaching the theoretical upper bound. This result confirms that energy savings are not achieved at the cost of capacity. Instead, intelligent resource coordination allows the network to sustain high data rates with lower power consumption. Fairness, measured using Jain’s fairness index, highlights a critical distinction between the schemes. While OMA exhibits high fairness due to orthogonal resource allocation, it suffers from limited efficiency and throughput. NOMA-FPA, by contrast, sacrifices fairness to improve spectral utilization. AEEOF successfully reconciles these competing objectives, achieving both high fairness and high efficiency by learning. balanced power and scheduling policies that explicitly account for disadvantaged users. The outage robustness metric further demonstrates the resilience of AEEOF under challenging channel conditions. By minimizing the likelihood of service disruption through anticipatory decision-making and interference mitigation, AEEOF achieves the highest robustness score among all schemes. This improvement is particularly important for reliability-sensitive applications and cell-edge users. Overall, Fig. 8 confirms that the proposed AEEOF does not merely optimize individual KPIs in isolation. Instead, it delivers a well-balanced, system-level optimization that simultaneously enhances energy efficiency, throughput, fairness, and reliability. This holistic improvement is a direct consequence of integrating spatio-temporal deep learning with multi-agent reinforcement learning, enabling the framework to transcend the limitations of conventional optimization strategies. Energy–spectral efficiency trade-off Figure 9 illustrates the fundamental trade-off between spectral efficiency and energy efficiency for the evaluated network schemes. This trade-off represents one of the central challenges in green wireless communication systems, as achieving higher data rates typically requires increased transmit power, leading to reduced energy efficiency. Within the evaluated simulation environment, the proposed AEEOF achieves consistently higher energy efficiency across all compared baseline schemes and across all evaluated KPIs. These results indicate that the integration of spatio-temporal feature learning with multi-agent reinforcement learning offers measurable advantages over the studied conventional and single-component baselines under the specific conditions examined. Energy efficiency versus spectral efficiency trade-off for different multiple access and optimization schemes, illustrating the Pareto frontier achieved by the proposed AEEOF framework. At a representative operating point of 10 bps/Hz, AEEOF achieves an energy efficiency of approximately 3.2 Mbits/Joule, corresponding to a 40% improvement over the NOMA-DQN baseline. This gain is particularly significant because NOMA-DQN already represents an AI-driven optimization approach, highlighting the added value of explicitly modeling spatial interference and temporal traffic dynamics. The shape of the AEEOF curve provides insight into the underlying optimization mechanism. By leveraging spatio-temporal feature learning, the framework reduces the marginal power cost of achieving higher spectral efficiency. The ST-GCN component mitigates inter-user interference by learning structured spatial relationships, while the GRU anticipates traffic fluctuations, enabling proactive resource provisioning. Together, these capabilities allow AEEOF to shift the entire energy–spectral efficiency trade-off curve upward. In contrast, the baseline schemes exhibit inferior frontiers due to their limited adaptability. OMA suffers from inefficient spectrum utilization, while NOMA-FPA relies on static power allocation that cannot respond to dynamic interference conditions. Although NOMA-DQN improves adaptability through reinforcement learning, its lack of explicit spatio-temporal modeling constrains its ability to achieve global Pareto optimality. Overall, Fig. 9 demonstrates that, within the studied single-cell simulation setting, the proposed AEEOF consistently achieves a more favourable energy–spectral efficiency operating region compared to the evaluated baseline schemes. These results suggest that joint spatio-temporal learning and multi-agent reinforcement learning can meaningfully improve the energy–capacity operating point relative to conventional and single-component optimization approaches, though generalisation to multi-cell or heterogeneous deployments remains a subject for future investigation. Scalability under increasing user density Figure 10 investigates the scalability of the suggested AI-Enhanced Energy Optimization Framework (AEEOF) through the assessment of energy efficiency as the number of users that use the service grows. The analysis will be critical in determining whether the framework can be deployed in dense and ultra-dense network conditions that are envisaged in 5G-Advanced and above. With higher user density, energy efficiency diminishes in all the schemes considered because of the growing inter-user interference, the increased contention of common resources, and the increased signaling overhead. The degradation rate, however, is much different among schemes. The suggested AEEOF can sustain maximum efficiency of the energy consumption, regardless of the number of users, and exhibits a high scalability over the benchmark methods. Energy efficiency as a function of the number of active users, illustrating the scalability of the proposed AEEOF framework under increasing network density. NOMA-FPA, NOMA-DQN, and AEEOF) have the same mean energy efficiency. With low to medium user densities, AEEOF attains significant benefits because it effectively uses both space reuse and predictive scheduling. As the network approaches the highly congested regime, the structure remains superior to the competing approaches and maintains a large performance gap even when 100 users are active. This action suggests that the policies acquired generalize well to large-scale state spaces and do not suffer from a combinatorial explosion that often impedes solutions of reinforcement learning in dense environments. The smooth nature of the performance degradation of AEEOF could be ascribed to the hybrid spatio-temporal architecture. The spatial encoder of the graph form involves a natural scaling with network size that the model of the interference through the localized neighborhood aggregation, whereas the GRU includes aggregate traffic dynamics that do not decrease even when the number of users increases. Conversely, the OMA and NOMA-FPA schemes have no flexibility and hence decay quickly, whereas the NOMA-DQN baseline has difficulties in keeping the best policies because it does not take into account the state in a structured way but rather has a limited state representation and lacks coherent spatial awareness. On the whole, Fig. 10 proves that AEEOF can not only perform well in a nominal setting but also withstands and scales to a high user density, which would be suitable in future wireless deployments that will be characterized by massive connectivity as well as heterogeneous service demands. Statistical significance and robustness validation A detailed statistical significance and robustness analysis was done to ensure that the performance improvements attained by the proposed AI-Enhanced Energy Optimization Framework (AEEOF) are not a product of stochastic variability or good simulation conditions. The consistency, reliability, and reproducibility of the observed improvements of the system are strictly validated in this subsection through further investigation of the reliability of the results across various independent simulation runs. Statistical evaluation methodology All statistical analyses have been conducted on the energy efficiency measure (Mbits/Joule), which is a composite measure of both energy consumption and presented throughput and hence the most reflective measure of system-level performance. To determine the energy efficiency of each of the considered schemes (OMA, NOMA-FPA, NOMA-DQN, and AEEOF), the energy efficiency turned out to be recorded in 1,000 distinct Monte Carlo simulation runs, where the user locations, channel realizations, and traffic patterns were randomized. Such a large sample size provides statistical strength as well as avoids bias because of particular network instantiations. Since the present comparison was multi-group with somewhat normal distribution of the values of energy efficiency, Analysis of Variance (ANOVA) was used to test the null hypothesis of making the schemes of the same scheme having the same mean performance. To further determine the pairwise differences of schemes, the Tukey Honestly Significant Difference (HSD) post hoc test was applied at a significance level of α = 0.01. One-way ANOVA results As depicted in Table 3, a one-way ANOVA (Analysis of Variance) was conducted to determine whether the observed differences in energy efficiency among the evaluated schemes are statistically significant. The null hypothesis states that all schemes (OMA, Table 3One-way ANOVA results for energy efficiency.Source of VariationSum of SquaresDegrees of FreedomMean SquareF-Statisticp-value Between Schemes 18.7236.24520< 0.001 Within Schemes 49.7839960.012—— Total 68.503999——— One-way ANOVA results for energy efficiency. The ANOVA is run with 1000 independent Monte Carlo simulation runs of schemes with randomized user positions, channel conditions, and traffic demands. The following table shows how much greater (i.e., + 520) the computed F-statistic is than the sum of all the differences within each evaluation scheme and their respective mean energy efficiency distributions. Based on the computed F-statistic, the average energy efficiencies for all evaluated schemes are estimated to have significantly greater variability than those for other evaluated schemes. (p < 0.001) For this reason, there is sufficient evidence to reject the null hypothesis that the mean energy efficiencies for each of the evaluated schemes have no statistically significant difference between them. In summary, they demonstrate significant performance gain differences between the proposed AEEOF method and performance gains attributable to random fluctuations in underlying network conditions and the randomness of the stochastic simulation used in this experiment. Post-hoc pairwise comparisons were conducted using a Tukey HSD test to determine which of the evaluated schemes demonstrate significant differences compared to each other. The Tukey HSD test controls for multiple comparisons while providing estimates of the magnitude of pairwise differences in average mean energy efficiencies. As shown in Table 4, the proposed AEEOF execution has demonstrated superior performance over all benchmark methods, including the AI-based NOMA-DQN approach. Although there was little difference in performance between the baseline methods, the differences were statistically significant, indicating incremental gains from adaptive power control as well as reinforcement-learning; therefore, the AEEOF provides these substantial advantages due to its unique combination of spatio-temporal feature learning with multi-agent reinforcement-learning techniques for consistent improvements that cannot be achieved using traditional or single-path optimization techniques. Table 4Tukey’s HSD pairwise comparison results.ComparisonMean Difference (Mbits/J)p-valueSignificance AEEOF vs. OMA +1.45< 0.001Significant AEEOF vs. NOMA-FPA +0.85< 0.001Significant AEEOF vs. NOMA-DQN +0.60< 0.001Significant NOMA-DQN vs. NOMA-FPA +0.25< 0.01Significant Tukey’s HSD pairwise comparison results. Distributional analysis of energy efficiency It is also important when evaluating robustness and variance to look at distributional characteristics beyond simply comparing means. The mean energy efficiency (AEEOF) for the proposed scheme is 3.05 Mbits/J (see Fig. 11), which confirms what has been previously discussed in prior evaluations of the performance characteristics of the scheme. The very low standard deviation of 0.15 Mbits/J — the lowest among all evaluated schemes — implies that AEEOF exhibits a high level of robustness and consistency across different network realisations. Statistical distribution of energy efficiency for OMA, NOMA-FPA, NOMA-DQN, and AEEOF over 1,000 Monte Carlo simulation runs, highlighting both mean performance and variance. Since there is very little dispersion present in the data set on energy efficiencies, it can be inferred that the control policies learned are likely to have similar performance when deployed in networks with adverse traffic and channel conditions; therefore, these control policies should have very little performance degradation under extreme or dynamic conditions. In contrast, the baseline schemes demonstrate significantly higher variability compared to the proposed scheme, particularly with respect to the NOMA-FPA (i.e., the performance of the NOMA-FPA scheme is very sensitive to abrupt changes in either traffic demand or channel conditions). Robustness and stability discussion The statistical distributions reveal that, in addition to achieving the scheme, AEEOF has both high mean energy efficiency (AEEOF) and the lowest variance of all schemes tested, indicating a high level of stability and resilience across varying network conditions. The reduced variance supports the effectiveness of spatiotemporal learning to address uncertainty created by varying traffic demands and channel characteristics. The narrow confidence intervals of AEEOF also show that the learned control policies from AEEOF produced equivalent generalizations across independent experiments, rather than exploiting one or more isolated favourable situations. This robustness is critical in the implementation of control policies for actual operational environments due to the inherently non-stationary and unpredictable nature of operational environments. Discussion and key insights The results from Sections VI-A through VI-E are integrated to extract the core insights resulting from the abstracted AEEOF (AI-Enhanced Energy Optimization Framework). Instead of repeating numerical outcomes of each of the sections previously referenced, a discussion of how the abstracted AEEOF works is provided, along with validation of the design principles used and the way the traditional assumptions regarding interference-limited wireless networks have been affected. From learning reliability to system-level gains The experimental evaluation’s central outcome shows a strong link between the learning of a system and how well it performs at the network level. The convergence behavior and stability analyses of the model have indicated that AEEOF’s spatio-temporal model learns valid and repeatable representations of the network’s dynamics. The reliability associated with the learning does not result solely from machine learning reasons but shows up as calculable operational benefits. The consistent reductions in energy use and, at the same time, increases in throughput and coverage indicate that these policies learned were not overfit to any individual channel realizations or traffic patterns. They show that there is substantial structural knowledge concerning the relationships between interference, traffic evolution, and power allocation through space and time within the network, confirming that there must be robust learning done before one can be assured of the reliability of autonomous control of a network. Furthermore, the MADDPG policy convergence curve in Fig. 5 confirms that the reported system-level performance improvements reflect fully converged reinforcement learning policies, providing the evidential link between learning process stability and operational performance gains that completes the convergence analysis of the proposed framework. Breaking the energy–performance trade-off Typical network-stabilizing optimizations are based upon a continuous tradeoff constrained by the need for one of the following three factors to be improved: (a) energy efficiency, (b) throughput, and (c) coverage/fairness. The simulation results indicate that, within the evaluated setting, AEEOF achieves simultaneous improvements in energy consumption, throughput, and coverage, suggesting that the tension between these objectives can be partially mitigated through intelligent spatio-temporal coordination. These findings are encouraging, though whether such simultaneous gains generalise beyond the specific simulation conditions studied here (including single-cell topology, NOMA-based access, and the defined traffic model) requires further empirical validation. The role of spatio-temporal intelligence in interference management The importance of modeling the spatial dimension and the temporal dimension together in interference-heavy environments has been indicated through analyzing the results of both types of analysis. The observed energy savings across the entire cell area, as well as for users located far away, show that the spatial learning component of AEEOF is effective at managing interference relationships by representing the network as a graph and learning the structure of interactions within the network; as such, it does not require excessive amounts of power to be used as a result of poor channel characteristics when using other means to compensate for interference (through increased transmit power). At the same time, by using modelled data to anticipate changes in traffic over time, the AEEOF framework can take preemptive action on the basis of anticipated congestion rather than waiting to respond after congestion has already occurred. By being able to anticipate changes in traffic, the AEEOF framework reduces marginal costs due to transience and eliminates oscillatory behaviour in control. Together, the integration of both forms of intelligence provides a framework for developing an optimal coordinated approach to optimizing the performance of the entire network compared to using either a spatial approach or a temporal approach independently of each other. Fairness and cell-edge performance as first-class objectives But also, practical deployment feasibility about its ability to leverage information from both the past (temporal aggregation) and the present (spatial encoding). Because of this combination of structured representations, the optimal actions can be effectively identified based on previously successful actions, rather than through exhaustive searching in multiple dimensions that could potentially fail. Similarly, AEEOF’s ability to utilize these two types of information leads to a general framework that is amenable to many new applications of reinforcement learning, as well as new areas of research into reinforcement learning as a whole due to the flexible nature of the architecture. Statistical robustness and trustworthiness All rated gains are supported by Statistical significance testing. The rejection of the null hypothesis with a high level of confidence and the continued AEEOF’s performance in post-test based on quality superiority supports the theory of systematic rather than incidental improvements. The decreased variation across energy efficiency distributions indicates very stable operations across multiple operating conditions. The statistical stability of the results is very important for real-world applications since network operators need reliable and consistent results when operating their networks. The results suggest that AEEOF may be an acceptable autonomous control methodology that can provide performance guarantees when operating at unknowns. Implications for next-generation wireless networks The collective findings suggest that spatio-temporal learning offers a promising direction for energy-efficient optimization in wireless systems, at least within the interference-limited, single-cell NOMA conditions studied here. Broader conclusions regarding the optimization landscape of next-generation wireless networks will require validation across more diverse deployment scenarios. Wireless systems will benefit from holistic controls based on AI, allowing them to operate adaptively, similarly aligning types of efficiency, performance, fairness, and reliability rather than asking for optimal solutions via static heuristics or objectives. These types of learning dynamics are already impacting future wireless solutions, experiencing denser infrastructure deployments, increasing variety in use cases, and larger environmental sustainability requirements. This study indicates that AI-enabled learning systems capable of understanding and utilizing structures will play a significant role in achieving these principles. Ablation study To assess the individual performance of different architectural components of the AI-Enhanced Energy Optimization Framework (AEEOF), an extensive ablation study was performed. A total of three ablation variants were created by disabling one of the core modules, while the other modules and their corresponding simulation settings remained unchanged. The definitions for the three variants are as follows: AEEOF w/o ST-GCN: The spatiotemporal graph convoluted module has been removed where only the temporal features are extracted using the GRU and aggregated with the MADDPG controller. This version helps to estimate the separate contribution of modeling the spatial interference. AEEOF w/o ST-GCN: The spatiotemporal graph convoluted module has been removed where only the temporal features are extracted using the GRU and aggregated with the MADDPG controller. This version helps to estimate the separate contribution of modeling the spatial interference. AEEOF w/o GRU: The AEEOF scheme without GRU retains only the spatial features of the ST-GCN along with the MADDPG controller. The AEEOF w/o GRU scheme is proposed to analyze and validate the effect of traffic modeling in the temporal domain separately. AEEOF w/o GRU: The AEEOF scheme without GRU retains only the spatial features of the ST-GCN along with the MADDPG controller. The AEEOF w/o GRU scheme is proposed to analyze and validate the effect of traffic modeling in the temporal domain separately. AEEOF w/o MADDPG: The multi-agent deep deterministic policy gradient reinforcement learning controller is replaced with a greedy heuristic control policy, and the ST-GCN and GRU feature extractors are retained. This setup accounts for the independent impact of adaptive reinforcement learning-based control. AEEOF w/o MADDPG: The multi-agent deep deterministic policy gradient reinforcement learning controller is replaced with a greedy heuristic control policy, and the ST-GCN and GRU feature extractors are retained. This setup accounts for the independent impact of adaptive reinforcement learning-based control. Simulations for the four set-ups, including AEEOF, were performed under similar criteria, as previously stated in Section VI-A, for over 1000 independent Monte Carlo simulations with randomized user locations, channel implementations, and traffic. The results are tabulated in Table 5. The ablation results reveal several important and consistent insights regarding the role of each architectural component. Table 5Ablation study results across key performance metrics.MetricAEEOF w/o ST-GCNAEEOF w/o GRUAEEOF w/o MADDPGAEEOF (Full) Energy Efficiency (Mbits/J) 2.152.422.783.05 Total Energy Consumption (J) 58.452.146.842 System Throughput (Mbps) 125.8134.5142.2155.4 Jain’s Fairness Index 0.740.790.850.92 Outage Probability (%) 12.59.86.43.2 Ablation study results across key performance metrics. Contribution of ST-GCN (Spatial Modeling): Removing the ST-GCN module produces the most significant performance degradation across all evaluated metrics. Energy Efficiency drops to 3.05 from 2.15 Mbits/J with a drop of 29.5%. Total Energy Consumption also increases from 42.0 J to 58.4 J, with a total increase of 39.0%. Throughput drops from 155.4 to 125.8 Mbps, Jain’s Fairness Index drops from 0.92 to 0.74, and Outage Probability increases from 3.2% to 12.5%. This indicates that the most significant component of the proposed approach to identifying the dominant interference paths and spatial-energy bottlenecks is the spatial modeling in ST-GCN. Without ST-GCN, the proposed approach is not able to explicitly identify the interference patterns between users that share the same spectrum. Instead of relying on the weights for the different user paths, it resorts to simply increasing the transmit power to guarantee the required transmission performance. This solution results in high-energy consumption without significantly improving fairness across users. Impact of the Within-GRU (Temporal Modeling): With full removal of the temporal module, performance degradation averaged over all metrics is moderate but consistent. Energy efficiency decreased from 3.05 to 2.42 Mbits/J (−20.7%); total energy consumption increased from 42.0 to 52.1 J. System throughput shows a decrease from 155.4 to 134.5 Mbps and the fairness index from 0.92 to 0.79, while outage probability doubled from 3.2 to 9.8%. These observations confirm the importance of temporal traffic modeling in proactive and anticipatory resource allocation and management. Without GRU, the framework loses its predictability and foresight capacity in estimating future changes for traffic demands and channel state information, which forces the system into an immediate-reactive control, which triggers energy-demanding transmissions and increases the occurrences of service outages. Impact of Removing MADDPG (Reinforcement Learning Control): Removing the MADDPG RL controller and replacing it with a greedy heuristic policy incurs the lowest and yet significant performance loss among the three ablation variants. Energy efficiency drops from 3.05 to 2.78 Mbits/J, an 8.9% efficiency loss, with total energy consumption collating to 42.0 J and 46.8 J. System throughput drops from 155.4 to 142.2 Mbps, with the fairness index dropping from 0.92 to 0.85 and the outage probability doubling from 3.2% to 6.4%. These results indicate that the spatio-temporal representation needs the MADDPG RL sequential adaption in decision making to optimally control the system. Without MADDPG RL, the accurate representation of the system state would not yield energy-efficient power allocation and scheduling decisions over the long operational time scale. From the ablation study, we can observe that a clear order can be established based on the importance of each component: ST-GCN > GRU > MADDPG. However, all three components had a significant improvement in AEEOF’s performance, indicating that no one specific component can solely provide a solution to produce AEEOF’s full performance. AEEOF outperforms other methods in all five metrics considered when all three modules are used together. This indicates that neither module could outperform using ST-GCN with either GRU or MADDPG, meaning that spatial modelling, temporal modelling, and reinforcement learning-based control are all concurrently required and sufficient to enable the performance improvements in energy efficiency, throughput, fairness, and outage probability. These results provide support to the motivation behind AEEOF’s architecture, as discussed in Section V-A, as well as the architectural design choices seen in the proposed framework. Limitations and scope of the present study While the simulation results reported in Sections VI-A through VI-G are encouraging and demonstrate consistent performance improvements under the studied conditions, it is important to clearly delineate the scope and boundaries of the present study to avoid overgeneralization of the findings. Single-cell topology All experiments are conducted within an isolated single-cell hexagonal deployment. Inter-cell interference, cooperative multi-cell coordination, handover management, and the effects of adjacent cell loading are therefore absent from the evaluation. Real-world deployments, particularly in ultra-dense 5G and beyond networks, involve complex inter-cell interactions that may affect the behavior and generalizability of the proposed framework. Extension of AEEOF to multi-cell and heterogeneous network topologies is an important direction for future work, as noted in the conclusion. Simulation-based evaluation The framework is evaluated exclusively within a MATLAB/Simulink-based digital twin environment. While the simulation incorporates 3GPP TR 38.901-compliant channel models, imperfect SIC, and realistic traffic and mobility patterns, it does not constitute a hardware testbed or real channel measurement campaign. Performance under real propagation conditions, hardware impairments, and implementation constraints may differ from the reported simulation results. Validation on software-defined radio platforms or standardized system-level simulators is identified as a priority for future experimental work. Specific NOMA formulation The study focuses on downlink power-domain NOMA with a fixed user clustering strategy. Uplink NOMA, cooperative and relay-assisted NOMA, rate-splitting multiple access (RSMA), and grant-free access schemes are outside the scope of the present evaluation. Furthermore, while imperfect SIC is modelled, the SIC ordering is assumed to be known and fixed; dynamic SIC ordering under mobility is not considered. Centralized training assumption The MADDPG framework relies on centralized training with global state information, which may not be directly available in fully distributed real-network deployments. Although decentralized execution is supported, the training phase requires access to joint observations and actions of all agents, which implies coordination overhead. Fully decentralized or federated training variants of the framework are left for future investigation. Baseline comparisons The comparative evaluation includes OMA, NOMA with fixed power allocation (NOMA-FPA), and a single-agent DQN-based NOMA baseline (NOMA-DQN). While these represent a principled set of conventional and AI-assisted reference points that enable systematic isolation of each architectural contribution, the evaluation does not extend to more recent approaches such as MAPPO-based cooperative resource management, transformer-enhanced power allocation frameworks, or model-based reinforcement learning methods with integrated channel prediction. These more recent approaches represent relevant and increasingly competitive reference points, and a broader comparative study incorporating them would further strengthen the evaluation. This is acknowledged as a limitation of the present work and is identified as a concrete direction for future investigation. These limitations do not diminish the validity of the reported findings within their stated scope, but they do define the boundaries within which the conclusions of this study should be interpreted. Conclusion and future work This study introduces an AI-Enhanced Energy Optimization Framework (AEEOF), which has been developed to solve the main challenge of providing energy efficiency when working with wireless networks that share a lot of interference while maintaining performance, fairness, and scalability. The motivation for the development of AEEOF comes from looking at how the exponential growth in traffic, i.e., data, and the sustainability needs of today’s cellular networks are conflicting. AEEOF integrates deep spatio-temporal learning with multi-agent reinforcement learning to enable intelligent and adaptive network control. The key innovation behind AEEOF is that it deals with all aspects of network dynamics holistically. It does this by modelling the spatial interference relationships contained within the Spatio-Temporal Graph Convolutional Network, along with the capturing of the temporal traffic evolution using Recurrent Learning to create a very detailed and structured representation of what is happening in the network state. Once this representation is created, it can be embedded into the Multi-Agent Reinforcement Learning in order to provide for Decentralized but Co-ordinated decision-making to evolve the power allocation and scheduling policies as a result of two separate criteria: the current and the future availability of traffic data (instantaneous vs. time). This addresses the principal constraints of traditional optimisation approaches, which typically rely on static assumptions and optimise individual objectives in isolation rather than in a coordinated manner. AEEOF’s huge statistical performance gains across every metric are evident in simulated 5 g oriented NOMA environments. AEEOF reduces total energy usage by up to 15% while enhancing energy efficiency by approximately 40%, with simultaneous improvements in system throughput, fairness, effective coverage, and outage probability. The statistically significant analysis also indicates that the improvement in performance across every metric is a consequence of systematic processes, as well as the lack of stochastic variability. These results are obtained within a controlled single-cell, MATLAB-based simulation environment under the specific system model described in Section IV, and should be interpreted accordingly. Convergence of both the ST-GCN–GRU supervised pre-training stage and the MADDPG reinforcement learning controller is demonstrated through dedicated convergence analyses in Section VI-B, providing complete empirical evidence that the reported gains reflect stable and fully trained policies. The findings also provide a much more in-depth understanding of the role that intelligent, spatio-temporally learning plays in wireless network optimization. Learning the structure of interference and the flow of traffic through the network, it allows for breaking through traditional, long-standing tradeoffs that can exist between energy efficiency, spectral efficiency, and user fairness. Therefore, AEEOF signifies a pivot away from the current state of being reactive and heuristic in nature toward a more proactive and data-driven manner in terms of communication, which aligns with efforts to create genuine AI-native communication systems. The limitations of the present study, including its single-cell simulation scope, the MATLAB/Simulink-based evaluation platform, and the specific NOMA formulation adopted, are discussed explicitly in Section VI-H. The directions for future work outlined below are motivated directly by these boundaries. Even though these results are very encouraging, there are many ways in which future research could be explored. Several concrete directions for future work are identified. First, the framework should be extended to multi-cell and heterogeneous network topologies, incorporating inter-cell interference coordination and handover management to broaden its applicability beyond the single-cell setting studied here. Second, a comprehensive comparative evaluation against more recent state-of-the-art baselines — including MAPPO-based cooperative resource management, transformer-enhanced resource allocation frameworks, and model-based reinforcement learning approaches with integrated channel prediction — would further contextualise the contribution of AEEOF and establish its performance relative to the current frontier of the field. Third, online learning and continuous adaptation mechanisms should be explored to enable real-time deployment without dependence on pre-collected training datasets. Fourth, validation on hardware testbeds or software-defined radio platforms would provide important evidence of real-world feasibility beyond the simulation environment studied here. In summary, the findings of this research show that Artificial Intelligence-based Spatial and Temporal Intelligence can be a significant factor in enabling the creation of sustainable, efficient wireless networks. This paper provides an integrated optimization framework for combining deep learning with reinforcement learning. Overall, the basis for future research into intelligent resource management has been established through the use of AEEOF and will help facilitate energy-efficient and autonomous control in 5G and future generations of Communication Systems. Acknowledgements This research paper is not supported by any company or organization; it relies solely on the efforts of the authors. Abbreviations AI-Enhanced Energy Optimization Framework Artificial Intelligence Analysis of Variance Base Station Clustered Delay Line-C Centralised Training with Decentralised Execution Deep Q-Network Deep Reinforcement Learning Energy Efficiency Fixed Power Allocation Gated Recurrent Unit Honestly Significant Difference Interference-to-Noise Ratio Internet of Things Key Performance Indicator Multi-Agent Reinforcement Learning Multi-Agent Deep Deterministic Policy Gradient Multi-Agent Proximal Policy Optimisation Machine Learning Non-Orthogonal Multiple Access Orthogonal Multiple Access Power Amplifier Quality of Service Reinforcement Learning Rate-Splitting Multiple Access Successive Interference Cancellation Signal-to-Interference-plus-Noise Ratio Spatio-Temporal Graph Convolutional Network User Equipment Urban Macro-cell Author contributions All authors contributed to the study’s conception and design. The specific contributions of each author are outlined below:- **Khalil M. Abdelnaby: ** Conceptualization, methodology development, and primary design of the spatio-temporal deep learning architecture. Conducted the formal analysis and wrote the original draft.- **Mohammed A. F. Al-Husainy: ** Development and implementation of the Multi-Agent Reinforcement Learning (MARL) algorithms. Responsible for software simulation environment setup and data validation.- **Mohammad O. Alhawarat: ** Data curation and preprocessing of wireless network traffic patterns. Contributed to the mathematical modeling of energy efficiency metrics and visualization of experimental results.- **Mohamed A. Rohaim: ** Supervision and project administration. Provided critical revisions for intellectual content and assisted in the optimization of the neural network hyperparameters.- **Khairy M. Assar: ** Resources, funding acquisition, and final validation of the research findings. Performed the final review and editing of the manuscript to ensure alignment with industry standards. Data availability All data generated or analysed during this study are included in this published article. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement All data generated or analysed during this study are included in this published article.