QS-MADS: quorum sensing perturbation-driven adaptive multi-agent scheduling for distributed power grids Authors: Kun Zeng, Yixiao Cao, Hui Sun Published: 2026 ABSTRACT AbstractThe rapid proliferation of distributed energy resources (DERs) in modern power grids introduces unprecedented complexity into real-time dispatch scheduling, demanding solutions that can operate under high uncertainty, partial observability, and volatile renewable generation. Existing multi-agent reinforcement learning (MARL) approaches for grid dispatch often suffer from weak inter-agent coordination, sample inefficiency, and fragility under measurement noise, largely because they rely on either independent exploration or expensive hidden-state exchange. We propose QS-MADS, a quorum-sensing-driven adaptive multi-agent scheduling framework that combines a topology-aware graph attention encoder with a lightweight scalar QS signal. During centralised training and optional online adaptation, each agent emits one 32-bit QS criticality signal; the trainer aggregates these scalars into a population consensus that governs a differentiable adaptive perturbation module used only for policy-gradient exploration. At deployment, the stochastic perturbation is removed and each deterministic actor executes from its local k-hop graph embedding, avoiding pairwise hidden-state communication. On IEEE 118-bus, RTS-96, and the CIGRE medium-voltage microgrid benchmarks, QS-MADS achieves up to 25% higher cumulative dispatch reward, reduces voltage violation events by 31%, and converges in up to 40% fewer training episodes relative to the strongest baseline, while maintaining millisecond-level inference latency and robust performance under measurement noise.Keywords: Multi-agent reinforcement learning, Distributed energy resources, Quorum sensing, Deep learning, Power grid dispatch, Graph attention networkSubject terms: Energy science and technology, Engineering, Mathematics and computing Introduction Modern power grids are undergoing a structural transition from centralised generation dominated by large thermal plants toward heterogeneous networks of distributed energy resources (DERs) encompassing solar photovoltaics, wind turbines, battery storage, and flexible demand-response units1,2. This transition introduces profound operational challenges: the grid operator must continuously balance generation and load while satisfying voltage and frequency constraints across a network whose generation capacity fluctuates on sub-minute timescales3,4. Traditional model-based optimal power flow (OPF) solvers are ill-suited to this setting because they rely on accurate, globally known system models and incur prohibitive computational latency when re-solved at high frequency5,6. The scale and stochasticity of DER-rich grids therefore motivate data-driven, model-free scheduling approaches7,8. Recent optimisation studies have also improved day-ahead scheduling and security-constrained unit commitment under renewable integration through mixed-integer quadratically constrained programming (MIQCP), highlighting the continuing need for methods that combine operational feasibility with fast adaptive control9. Deep reinforcement learning (DRL) has shown considerable promise for power system control10,11. Among DRL paradigms, multi-agent reinforcement learning (MARL) is particularly attractive for distributed grids because each controllable asset can be modelled as an independent agent that optimises local objectives while remaining coordinated with the wider system12,13. Seminal frameworks such as MADDPG13 and MAPPO14 adopt a centralised training with decentralised execution (CTDE) paradigm that has been transferred to grid dispatch with encouraging results8,15. More recent works incorporate graph neural networks (GNNs) to capture grid topology, enabling agents to exchange learned representations along physical power-flow edges4,16. Despite these advances, three fundamental weaknesses persist. First, explicit message-passing schemes scale poorly with the number of agents and introduce communication overhead that may be infeasible on hardware-constrained grid edge controllers5. Second, standard policy gradient methods for cooperative MARL are highly sensitive to the exploration strategy: naive ε-greedy or Gaussian noise injection treats all agents identically, ignoring the heterogeneous state conditions across the grid and often resulting in either excessive conservatism or destructive oscillation7,14. Third, the robustness of learned dispatch policies under realistic measurement noise and sensor fault conditions remains underexplored, limiting real-world deployment confidence6,10. The fundamental challenge in distributed grid dispatch is achieving collective intelligence with bounded, low-bandwidth coordination: each agent must make locally executable decisions that are globally coherent, even though its observation is incomplete and noisy, and full exchange of policy, value, or high-dimensional hidden representations is undesirable. This challenge is further compounded by the non-stationarity introduced when multiple agents update their policies concurrently, which can cause training instability and reward collapse17,18. Inspired by how bacteria achieve colony-level behavioural coordination through quorum sensing (QS)—a chemical signalling mechanism whereby individual cells broadcast molecular signals and collectively respond once an aggregate population density threshold is exceeded—we propose QS-MADS, as illustrated in Figure 1. In QS-MADS, each dispatch agent emits a scalar QS signal derived from a compact graph-attention encoding of its local grid state; during CTDE training these signals are aggregated into a population consensus vector that triggers a differentiable adaptive perturbation module for structured exploration noise. The key insight is that the QS consensus signal plays the role of a shared training-time coordination context that modulates exploration in a population-aware manner, while deployed actors remain deterministic and locally executable without full policy, value, or hidden-state exchange. Overview of the QS-MADS framework. During CTDE training, grid nodes emit lightweight scalar QS criticality signals that are aggregated into a population-level consensus vector. The consensus vector drives the adaptive perturbation module, which shapes training exploration; at deployment, the stochastic perturbation is removed and the deterministic local actors execute projected dispatch actions on the physical grid. QS-MADS comprises five tightly integrated components: (i) a graph attention encoder (GAT) that maps local observations and physical topology into agent-level embeddings; (ii) a QS signal generator that projects each embedding onto a scalar signal reflecting local state criticality; (iii) a differentiable QS aggregator, used in the centralised training learner, that forms a population consensus without pairwise exchange of hidden representations; (iv) an adaptive perturbation module that uses the consensus vector to parametrise structured noise distributions injected into actor gradient updates during training; and (v) a QS-shaped reward function that attributes individual rewards relative to the population consensus, stabilising cooperative learning. Together, these components enable QS-MADS to exhibit emergent coordination, robust exploration, and stable convergence in large-scale distributed grid settings, while deployed control remains local and deterministic. We summarise our main contributions as follows: Novel QS-driven perturbation mechanism. We introduce the first principled transplantation of biological quorum sensing into cooperative MARL for power grids, replacing ad hoc exploration noise with a population-aware, differentiable perturbation module that adapts exploration intensity to the collective grid state. Novel QS-driven perturbation mechanism. We introduce the first principled transplantation of biological quorum sensing into cooperative MARL for power grids, replacing ad hoc exploration noise with a population-aware, differentiable perturbation module that adapts exploration intensity to the collective grid state. Low-bandwidth emergent coordination. QS-MADS achieves inter-agent coordination during training through a compact scalar signal pathway rather than pairwise policy, value, or hidden-state message passing, reducing the training-time coordination payload by over 90% compared with graph-message-passing baselines while improving coordination quality. Low-bandwidth emergent coordination. QS-MADS achieves inter-agent coordination during training through a compact scalar signal pathway rather than pairwise policy, value, or hidden-state message passing, reducing the training-time coordination payload by over 90% compared with graph-message-passing baselines while improving coordination quality. Unified QS-shaped reward. We design a novel reward shaping mechanism that anchors individual agent rewards to the population QS consensus, mitigating the non-stationarity problem in cooperative MARL without introducing value decomposition complexity. Unified QS-shaped reward. We design a novel reward shaping mechanism that anchors individual agent rewards to the population QS consensus, mitigating the non-stationarity problem in cooperative MARL without introducing value decomposition complexity. State-of-the-art performance and robustness. QS-MADS achieves up to 25% higher dispatch reward and 31% fewer voltage violations than strong baselines across three public benchmark grids, maintains stable performance under measurement noise levels up to σ = 0.30, and exhibits millisecond-level inference latency. State-of-the-art performance and robustness. QS-MADS achieves up to 25% higher dispatch reward and 31% fewer voltage violations than strong baselines across three public benchmark grids, maintains stable performance under measurement noise levels up to σ = 0.30, and exhibits millisecond-level inference latency. Related work Deep reinforcement learning for power grid dispatch Deep reinforcement learning has emerged as a leading paradigm for model-free power system optimisation, with applications spanning voltage regulation, economic dispatch, and demand response2,3,10. Duan et al.10 demonstrated that actor-critic architectures can solve security-constrained OPF problems faster than interior-point methods. Subsequent works introduced recurrent network architectures to capture the temporal dynamics of renewable generation forecasts6,8. Li et al.2 showed that incorporating physical network constraints as differentiable penalty terms in the reward function markedly improves constraint satisfaction rates. However, single-agent DRL treats the entire grid as a monolithic control problem and does not scale to systems with hundreds of independently controllable DERs, motivating the shift to multi-agent formulations4,5. Inverter-based Volt-VAR control is particularly relevant to our setting: Liu and Wu introduced a two-stage DRL scheme for robust inverter control under model mismatch19, and later proposed an online multi-agent constrained soft actor-critic framework for decentralised Volt-VAR control20. These studies demonstrate the value of learning-based distributed control, but they do not directly address population-aware exploration and scalar consensus for large-scale cooperative scheduling. Multi-agent reinforcement learning and coordination Cooperative MARL has been extensively studied in the context of robotics and network resource management13,14,17,18. The CTDE paradigm, as instantiated by MADDPG13 and QMIX18, allows agents to share global state information during training while executing decentralised policies at inference time. More recent methods such as MAPPO14 and FACMAC21 extend this paradigm to continuous action spaces relevant for power dispatch. For grid-specific settings, several works embed GNNs within the CTDE framework to exploit topology-aware representations12,16. Recent attention- enhanced MARL has further improved distributed Volt-VAR control under observation perturbations, showing that attention mechanisms can increase robustness to imperfect measurements22. Despite strong results, these methods typically rely on explicit hidden-state exchange, centralised value factorisation, or attention communication among agents, which can become costly in large systems. Our work departs from this paradigm by using a biologically inspired scalar QS signal pathway as a low-bandwidth training-time coordination substrate rather than exchanging full embeddings or value messages. Biological mechanisms in machine learning The transfer of biological principles into machine learning has a rich history, from evolutionary algorithms to immune-system-inspired anomaly detection23. More recently, stigmergy and swarm intelligence concepts have been adapted for distributed optimisation24. In the reinforcement learning literature, population-based training methods use multiple concurrent agents to stabilise exploration and hyperparameter search25. Quorum sensing specifically has been applied in robotics for distributed task allocation and fault detection26, yet its integration into MARL policy gradient methods, particularly for grid dispatch, remains entirely unexplored. Our work establishes this connection, demonstrating that QS-derived consensus signals can serve as a principled, lightweight coordination substrate for cooperative deep RL agents. The QS formulation is theoretically preferable to standard density-based clustering or basic statistical consensus in this setting for three reasons. First, density-based clustering requires an explicit metric and produces discrete group assignments, whereas grid stress is continuous and can propagate across weak electrical couplings. Second, a simple average consensus treats all agents symmetrically and can dilute rare but security-critical voltage events. Third, the QS analogy motivates a differentiable threshold-like response: agents emit bounded criticality signals, critical agents receive larger consensus weights, and the resulting population signal modulates exploration intensity. This gives QS-MADS a continuous, selective, and trainable mechanism for amplifying sparse high-risk states, rather than a post hoc clustering rule. Microgrid stability and practical operation control Beyond learning-based dispatch, recent microgrid studies emphasise that operational scheduling must be evaluated against practical stability regions and preventive control constraints. Amato et al. proposed a geometrical method for fast practical static stability-region evaluation of smart microgrids27, and later extended this line of work toward a preventive control strategy for microgrids28. These works are complementary to QS-MADS: they provide stability-aware feasibility assessment, whereas QS-MADS learns a low-latency distributed scheduling policy. In practical deployments, the QS-MADS action projection layer and stability-region filters can be combined so that learned setpoints are both coordinated and security screened before execution. Method System overview QS-MADS models a distributed power grid as a graph where each node represents a controllable DER or grid bus, and each edge encodes physical power-flow admittance. At each decision timestep t, agent i observes a local state comprising bus voltage magnitude, voltage angle, active and reactive injections, renewable forecast error, storage state of charge when available, and neighbouring setpoints within a fixed k-hop electrical neighbourhood. The global dispatch problem is modelled as a partially observable cooperative Markov game: agents jointly maximise a team reward that penalises voltage violations, generation cost, and line congestion, while each agent executes a locally feasible continuous action (active and reactive setpoint adjustment). The method proceeds through five sequential steps as illustrated in Figure 2: Step 1 encodes local grid observations into structured embeddings via a graph attention network; Step 2 derives per-agent QS signals from these embeddings and aggregates them into a population-level consensus vector; Step 3 feeds the training-time consensus vector into a differentiable perturbation module that generates structured exploration noise; Step 4 updates actor and critic networks using the perturbed policy gradient augmented by QS-shaped reward; Step 5 deploys deterministic projected actions during evaluation and observes the resulting state transition. The five-step pipeline of QS-MADS. Each step corresponds to a dedicated neural module; arrows denote information flow during centralised training. The perturbation module (Step 3) is driven by the QS consensus vector from Step 2 and is disabled at deterministic deployment time. Training–deployment semantics and communication budget To remove ambiguity about decentralisation, we explicitly distinguish the training and deployment modes of QS-MADS. During CTDE training, the learner can access all simulated rollouts and therefore computes the consensus vector in the training process, not in a required cloud controller for real-time grid execution. The only quantity treated as a QS communication payload is the scalar criticality value . With 32-bit floating-point encoding, this corresponds to 4 bytes per agent per decision step; even for 64 agents, the aggregate scalar payload is only 256 bytes, before standard packet headers. In contrast, exchanging a 256-dimensional embedding would require 1024 bytes per sender before topology-dependent replication, and pairwise GNN message passing scales with the number of edges or agent pairs. At deployment, the stochastic perturbation and QS-shaped reward are disabled. Each agent executes after the hard feasibility projection, using only its local k-hop measurement window and the parameters learned during CTDE training. Thus, QS-MADS is not claimed to be communication-free; it is a low-bandwidth training-time coordination method with deterministic decentralised execution. If an operator enables online adaptation, the same scalar QS telemetry can be sent to a substation controller or regional coordinator, but the paper’s reported test-time actions do not require collection of global hidden vectors. Graph attention encoding Each agent i receives a zero-padded raw observation vector consisting of its own bus features and the features of nodes in a two-hop electrical neighbourhood. The per-node feature vector contains voltage magnitude, voltage angle, active and reactive injections, load demand, renewable forecast error, and storage state of charge when applicable. Table 1 gives the exact padded dimensions and action types used in the three systems. Table 1State dimensions and action types used by each dispatch agent. Dimensions are after zero-padding/truncation to the maximum two-hop neighbourhood size in each benchmark.SystemneighboursDimension Action type/dimension IEEE 118-Bus1268Continuous , RTS-961058Continuous , CIGRE MV Microgrid638Continuous , State dimensions and action types used by each dispatch agent. Dimensions are after zero-padding/truncation to the maximum two-hop neighbourhood size in each benchmark. A multi-head graph attention network (GAT)29 aggregates neighbourhood information to produce an agent embedding: where is a shared linear projection, σ denotes a LeakyReLU activation, and the attention coefficient is computed as: with being a learnable attention vector. The resulting embedding captures both local bus state and topology-weighted neighbourhood information, forming the input to all subsequent QS modules. Using multi-head attention with K=4 heads and concatenating their outputs allows the encoder to simultaneously attend to multiple aspects of the grid state, such as voltage sensitivity and power balance. We use k=2 rather than only immediate 1-hop neighbours because voltage and reactive-power effects are not strictly confined to adjacent buses in meshed or feeder-connected distribution networks. A validation sweep on IEEE 118-Bus showed that 1-hop observations reduce CDR to 0.819 and VSI to 91.1%, while 2-hop observations reach 0.861 CDR and 93.7% VSI; moving to 3 hops gives only marginal improvement (0.864 CDR and 93.9% VSI) while increasing inference latency from 2.83 ms to 4.61 ms. Therefore, k=2 provides the best accuracy–latency trade-off for the tested benchmarks. Quorum sensing signal aggregation The QS signal for agent i at time t is defined as a scalar measure of local state criticality: where and are learnable parameters, and maps the signal to (0, 1). Intuitively, when agent i detects a high-criticality condition (e.g., voltage approaching limit), analogous to high local bacterial density in biological QS. During centralised training, the population-level consensus vector is formed by a differentiable aggregation of all agent signals: where the aggregation weights are: with for numerical stability. This soft-weighting ensures that agents in critical states contribute disproportionately to the training consensus vector, analogous to the threshold-triggered collective response in biological QS. The high-dimensional embeddings in Eq. (4) are available inside the CTDE learner; they are not collected by a mandatory runtime cloud controller. The runtime communication payload is therefore not a full hidden vector exchange, but at most the optional scalar QS telemetry used for online monitoring or adaptation. Adaptive perturbation module The core novelty of QS-MADS is the adaptive perturbation module (APM), which uses the consensus vector to generate structured, heterogeneous exploration noise for each agent. Rather than injecting fixed isotropic Gaussian noise into the policy, APM parametrises a diagonal Gaussian perturbation distribution whose mean and variance are conditioned on the consensus: During training, the perturbed exploratory action is: where λ is a learnable scalar temperature that controls the overall perturbation magnitude. During validation and deployment, the stochastic term is removed and the controller uses the deterministic projected action , where enforces generator, inverter, and storage limits. Thus, random noise is never injected into physical pumps, inverters, or generators in real operation. Since is sampled from a differentiable distribution, gradients flow through the reparametrisation trick30, enabling end-to-end training of via the actor-critic policy gradient. When the consensus signal indicates that many agents are simultaneously in critical states (high ), the APM increases to encourage wider exploration, preventing premature convergence to local optima. Conversely, when consensus is low (all agents near comfortable operating points), the APM shrinks , concentrating exploitation of the learned policy. This population-aware modulation of exploration is precisely the adaptive quality that fixed noise schemes lack and that makes QS-MADS robust under non-stationary grid conditions. At test time, coordination is retained through the learned actor parameters and local GAT encoder; the APM has shaped the policy during training but does not add operational randomness. QS-shaped reward and policy update Naive cooperative MARL rewards treat all agents identically, which fails to credit agents in critical positions whose locally suboptimal actions may be globally beneficial. We introduce a QS-shaped individual reward that re-weights the team reward by each agent’s relative criticality: where is an exponential moving baseline of the team reward with η =0.01, and is a shaping coefficient. Agents with high QS signals receive amplified credit when the team reward exceeds its recent baseline, and amplified penalty when it falls below that baseline. This encourages critical agents to internalise the team-level consequences of their actions, reducing the lazy agent problem without requiring counterfactual value decomposition. The full algorithm is presented in Algorithm 1. Each agent’s actor and critic are updated via proximal policy optimisation (PPO) with the QS-shaped reward signal. The APM parameters θ are updated jointly with the actor via the reparametrised policy gradient. The QS signal parameters are updated by backpropagation through the consensus vector into the actor loss. QS-MADS Training Procedure (Training-Time Perturbation Only). Dispatch execution and feedback Following policy update, the dispatch actions are translated into physical generation setpoints via a projection layer that enforces hard box constraints using a differentiable clipping function. The projected setpoints are dispatched to the grid simulation (or physical hardware), and the resulting bus voltages and power flows are measured, forming the next-step observations . The feedback loop is closed by recomputing the team reward from measured power quality metrics. A key design choice is that the QS signal is computed from post-execution measurements rather than planned actions, ensuring the consensus vector reflects true grid state rather than anticipated state. This choice makes the QS aggregation robust to execution discrepancies caused by forecast error or hardware saturation. Evaluation Experimental setup All experiments are conducted on a server with four NVIDIA A100 80 GB GPUs and dual 64-core AMD EPYC 7763 CPUs under PyTorch 2.1. The GAT encoder uses hidden units, K = 4 attention heads, and two message-passing layers, giving ∼1.2M learnable parameters per agent. The APM is a two-layer MLP with hidden size 128 and ELU activations. Each actor is a three-layer MLP (256-256-) and each critic a three-layer MLP (256-256-1). Training uses Adam with learning rate , gradient clipping at norm 10, PPO clip , QS shaping coefficient , perturbation temperature λ initialised to 0.1 and learned jointly. Rollout length is T = 256 steps, batch size B = 1024, and each experiment runs for 2000 training iterations with five independent random seeds. Hyperparameters are selected via grid search on a held-out validation episode set. To ensure a fair comparison, all baselines were tuned with the same validation budget as QS-MADS. For each method, we swept learning rate , hidden width , batch size , and method-specific exploration or entropy coefficients. The reported baseline results use the best validation configuration averaged over the same five random seeds, rather than unmodified default literature values. Early stopping was not used; all methods were trained for the full 2000-iteration budget. Datasets We evaluate on three established power-system benchmarks. The IEEE 118-Bus system31 provides 118 buses, 54 generators, and 186 transmission lines, and is a standard testbed for large-scale grid control. The RTS-96 reliability test system32 models a 73-bus, 33-generator grid representative of deregulated market environments with stochastic renewable injection. The CIGRE MV Microgrid33 is a 14-bus medium-voltage benchmark with heterogeneous DER types including photovoltaic, wind, and battery storage, stressing the coordination of temporally correlated renewable resources. For all three datasets, we simulate renewable generation variability using realistic historical traces and inject synthetic measurement noise (σ = 0.05) during evaluation. Evaluation metrics We report four metrics that collectively capture both economic and power quality objectives: Cumulative Dispatch Reward (CDR): Mean episodic team reward, reflecting economic dispatch cost minimisation and constraint satisfaction jointly. Cumulative Dispatch Reward (CDR): Mean episodic team reward, reflecting economic dispatch cost minimisation and constraint satisfaction jointly. Voltage Stability Index (VSI): Fraction of timesteps in which all bus voltages remain within the permissible band [0.95, 1.05] p.u., higher is better. Voltage Stability Index (VSI): Fraction of timesteps in which all bus voltages remain within the permissible band [0.95, 1.05] p.u., higher is better. Voltage Violation Count (VVC): Total count of bus-timestep pairs violating voltage limits per episode, lower is better. Voltage Violation Count (VVC): Total count of bus-timestep pairs violating voltage limits per episode, lower is better. Convergence Episode (CE): Number of training episodes required to reach 95% of final policy performance, measuring sample efficiency. Convergence Episode (CE): Number of training episodes required to reach 95% of final policy performance, measuring sample efficiency. Baseline methods We compare against five representative baselines spanning rule-based and learning-based dispatch methods: MADDPG13: Multi-agent deep deterministic policy gradient with centralised critic and decentralised actors; a canonical continuous-action MARL algorithm widely benchmarked in grid control. MADDPG13: Multi-agent deep deterministic policy gradient with centralised critic and decentralised actors; a canonical continuous-action MARL algorithm widely benchmarked in grid control. MAPPO14: Multi-agent proximal policy optimisation with shared global state critic; represents the state of the art in cooperative MARL with on-policy training. MAPPO14: Multi-agent proximal policy optimisation with shared global state critic; represents the state of the art in cooperative MARL with on-policy training. SAC-MA: Multi-agent soft actor-critic with independent entropy regularisation per agent; adapted for distributed grid dispatch following11. SAC-MA: Multi-agent soft actor-critic with independent entropy regularisation per agent; adapted for distributed grid dispatch following11. CTDE-GNN4: A CTDE framework that replaces the centralised critic MLP with a graph neural network, exploiting grid topology to improve coordination. Unlike QS-MADS, CTDE-GNN uses graph representations inside the centralised critic/message-passing pathway; QS-MADS uses GAT only to build local agent embeddings and coordinates training through a scalar QS consensus plus perturbation module. CTDE-GNN4: A CTDE framework that replaces the centralised critic MLP with a graph neural network, exploiting grid topology to improve coordination. Unlike QS-MADS, CTDE-GNN uses graph representations inside the centralised critic/message-passing pathway; QS-MADS uses GAT only to build local agent embeddings and coordinates training through a scalar QS consensus plus perturbation module. QMIX18: Value decomposition method using a monotonic mixing network; included as a representative of factorised cooperative value methods adapted to continuous action spaces via action discretisation. QMIX18: Value decomposition method using a monotonic mixing network; included as a representative of factorised cooperative value methods adapted to continuous action spaces via action discretisation. Overall performance Tables 2 and 3 present the quantitative results on the IEEE 118-Bus and RTS-96 benchmarks respectively; results on the CIGRE MV Microgrid follow in Table 4. Figures 3a and 3b visualise normalised metric profiles across baselines. Table 2Overall Performance on IEEE 118-Bus System. CDR is reported as mean±std over five seeds. Best results are bolded; second-best are .MethodCDR ↑VSI (%) ↑VVC ↓CE ↓MADDPG0.621± 0.02781.4142.31841MAPPO0.593± 0.03178.9158.71623SAC-MA0.648± 0.02483.1131.81712CTDE-GNNQMIX0.634± 0.02682.7137.61582QS-MADS (Ours)93.771.9836Improvement+21.0%+7.3pp−31.0%−40.0% Overall Performance on IEEE 118-Bus System. CDR is reported as mean±std over five seeds. Best results are bolded; second-best are . Table 3Overall Performance on RTS-96 System. CDR is reported as mean±std over five seeds.MethodCDR ↑VSI (%) ↑VVC ↓CE ↓MADDPG0.598± 0.02979.3157.41889MAPPO0.572± 0.03476.1173.21701SAC-MA0.619± 0.02680.8144.91795CTDE-GNNQMIX0.608± 0.02880.2149.71614QS-MADS (Ours)91.881.7893Improvement+24.4%+8.1pp−31.0%−38.5% Overall Performance on RTS-96 System. CDR is reported as mean±std over five seeds. Table 4Overall Performance on CIGRE MV Microgrid. CDR is reported as mean±std over five seeds.MethodCDR↑VSI (%)↑VVC↓CE↓MADDPG0.573± 0.03277.2168.11932MAPPO0.551± 0.03674.5184.31748SAC-MA0.592± 0.02878.4155.61824CTDE-GNNQMIX0.581± 0.03077.9160.81657QS-MADS (Ours)90.189.3912Improvement+25.0%+8.9pp−29.9%−38.7% Overall Performance on CIGRE MV Microgrid. CDR is reported as mean±std over five seeds. Normalised utility profiles across baselines and datasets. For lower-is-better metrics (VVC and CE), values are inverse-normalised before plotting, so taller bars always indicate better performance. QS-MADS consistently achieves the highest profile across all four metrics. Analysis of Results. On the IEEE 118-Bus benchmark (Table 2), QS-MADS achieves a CDR of 0.861, outperforming the next-best method, CTDE-GNN (0.712), by 21.0%. The voltage stability index improves from 87.3% (CTDE-GNN) to 93.7%, reflecting that the QS consensus mechanism actively redirects exploration toward near-limit voltage conditions. Voltage violation counts drop from 104.2 (CTDE-GNN) to 71.9 per episode, a reduction of 31.0%, demonstrating that the QS-shaped reward successfully encourages agents in critical states to prioritise constraint satisfaction over myopic reward maximisation. Notably, QS-MADS achieves 95% of its final performance in only 836 training episodes, compared to 1394 for CTDE-GNN and 1841 for MADDPG, reflecting that population-aware perturbation dramatically accelerates exploration coverage of the state space. On RTS-96 (Table 3), QS-MADS maintains its superiority with a CDR of 0.847, an improvement of 24.4% over CTDE-GNN (0.681). The larger relative improvement on RTS-96 compared to 118-Bus can be attributed to the higher renewable penetration in RTS-96, which creates more frequent high-criticality grid events that the QS signal is particularly well-suited to detect and coordinate around. VVC drops from 118.3 to 81.7, and convergence accelerates from 1451 to 893 episodes. MADDPG and MAPPO both struggle on RTS-96, suggesting that naive policy gradient methods are poorly equipped for the non-stationarity introduced by volatile renewable generation, whereas the QS perturbation module provides a principled mechanism for agents to collectively adapt their exploration under these conditions. On the CIGRE MV Microgrid (Table 4), QS-MADS achieves the largest relative improvement across all three datasets: 25.0% CDR gain over CTDE-GNN. The microgrid benchmark features the smallest number of buses but the highest proportion of heterogeneous DERs, making intra-agent coordination harder and making the QS-driven low-bandwidth coordination advantage most pronounced. The convergence benefit is particularly striking: 912 episodes versus 1488 for CTDE-GNN, because in smaller systems the QS consensus signal has a higher signal-to-noise ratio and can more precisely guide exploration toward globally critical regions. Ablation Study Figure 4 and Table 5 present component-level ablation results on the IEEE 118-Bus benchmark. Ablation study comparing QS-MADS component variants on the IEEE 118-Bus system. Removing any single component degrades both dispatch reward and voltage stability. Table 5Ablation Study on IEEE 118-Bus System.VariantCDR↑VSI (%)↑VVC↓w/o QS Signal0.69484.2119.7w/o Perturbation0.73186.1108.4w/o Graph Attention0.76287.999.3w/o QS Reward0.74887.2103.8Full QS-MADS0.86193.771.9 Ablation Study on IEEE 118-Bus System. Removing the QS signal entirely (“w/o QS Signal”) causes the largest performance drop, with CDR falling from 0.861 to 0.694 (−19.4%) and VVC increasing from 71.9 to 119.7 (+66.5%). This confirms that the QS signal is the primary coordination substrate: without it, agents revert to independent policy gradient with isotropic noise, losing population-aware exploration. Removing only the perturbation module while retaining the QS signal (“w/o Perturbation”) yields CDR = 0.731, demonstrating that the QS signal alone—used only for reward shaping without the adaptive noise component—still provides meaningful coordination benefit but cannot match the full system. Removing the graph attention encoder and replacing it with a flat MLP (“w/o Graph Attention”) reduces CDR to 0.762, confirming that topology-aware embedding is essential for the QS signal to accurately reflect grid-structural criticality. Removing QS-shaped reward while retaining all other components (“w/o QS Reward”) reduces CDR to 0.748, indicating that the reward shaping complements but does not substitute for the perturbation mechanism. Together, these ablations confirm that all four components contribute unique and additive value, and that the QS signal generation is the most critical design choice. Importantly, the ablation results also explain the later robustness findings: the GAT filters local measurement noise, the QS signal compresses criticality into a bounded scalar, the APM learns where additional exploration is useful, and QS reward shaping aligns critical agents with team-level constraint satisfaction. Robustness Study Figure 5 shows dispatch reward as a function of measurement noise level for all methods on the IEEE 118-Bus benchmark. Dispatch reward under increasing measurement noise standard deviation on the IEEE 118-Bus system. QS-MADS maintains substantially higher reward across all noise levels, demonstrating robustness to sensor uncertainty. At zero noise, QS-MADS leads with CDR = 0.861 versus CTDE-GNN’s 0.712. As noise increases to σ = 0.30, QS-MADS degrades only mildly to 0.719 (a 16.5% drop), while MADDPG collapses to 0.247 (a 60.2% drop) and CTDE-GNN degrades to 0.389 (a 45.4% drop). This remarkable noise robustness of QS-MADS arises from two complementary properties. First, the QS signal is a sigmoid-compressed scalar derived from a multi-hop neighbourhood embedding, making it naturally resistant to individual sensor outliers—a single corrupted observation perturbs modestly because the GAT aggregates over multiple neighbours. Second, the APM’s perturbation variance adapts to the consensus signal: when noise corrupts several QS signals simultaneously, the consensus vector shifts, causing the APM to widen the exploration distribution, effectively performing implicit noise-aware exploration rather than blindly committing to corrupted policy gradients. These results indicate that QS-MADS is practically deployable in real grid environments where measurement faults and communication noise are routine. Scalability Study Figure 6 evaluates dispatch reward as the number of agents is varied from 4 to 64 on a parametrically scaled version of the IEEE 118-Bus system. Scalability evaluation: dispatch reward versus number of dispatch agents. QS-MADS scales nearly flat up to 64 agents, while baselines degrade significantly. QS-MADS maintains a CDR of 0.812–0.861 across 4 to 64 agents, degrading by only 5.7% at 64 agents relative to 8. CTDE-GNN degrades from 0.751 at 4 agents to 0.638 at 64 (−15.0%), and MAPPO from 0.728 to 0.589 (−19.1%). The superior scalability of QS-MADS stems directly from the O(N) communication cost of the QS broadcast mechanism: each agent emits only a scalar signal, and the consensus aggregation is a weighted mean computable in O(N) without pairwise communication. CTDE-GNN’s degradation reflects its message-passing bottleneck; MAPPO’s degradation reflects the increased variance of the centralised critic’s value estimation with larger agent populations. These results confirm that QS-MADS is architecturally suited for deployment in large-scale DER grids with hundreds of controllable assets. Inference latency and communication cost Reviewer feedback highlighted that sub-minute scheduling requires explicit runtime evidence. We therefore measured single-decision inference latency using 10,000 forward passes with batch size 1, including local GAT encoding, actor inference, and feasibility projection, but excluding the training-only stochastic perturbation. Table 6 and Figure 7 show that QS-MADS remains within millisecond-level latency on all benchmarks. Table 6Inference latency per dispatch decision in milliseconds, reported as mean±std over 10,000 forward passes. The stochastic APM is disabled because deployment uses deterministic actors.MethodIEEE 118-BusRTS-96CIGRE MVMADDPG1.74± 0.081.69± 0.070.82± 0.04MAPPO1.91± 0.101.86± 0.080.91± 0.05SAC-MA2.13± 0.112.07± 0.091.04± 0.05CTDE-GNN6.42± 0.245.97± 0.212.31± 0.10QMIX2.28± 0.122.17± 0.111.12± 0.06QS-MADS Inference latency per dispatch decision in milliseconds, reported as mean±std over 10,000 forward passes. The stochastic APM is disabled because deployment uses deterministic actors. Inference latency comparison. QS-MADS adds a small local GAT overhead relative to MLP-only baselines but is substantially faster than CTDE-GNN and remains far below the sub-minute dispatch requirement. The IEEE 118-Bus latency of QS-MADS is 2.83± 0.13 ms, compared with 6.42± 0.24 ms for CTDE-GNN. This confirms that the local GAT plus deterministic actor is compatible with fast real-time scheduling. The scalar QS payload is also bounded: 64 agents require only 256 bytes per decision step for optional QS telemetry, whereas exchanging 256-dimensional embeddings would require at least 65,536 bytes for all directed pairwise pairs in a 64-agent fully connected message-passing design. Convergence analysis Figure 8 plots training curves for all methods on the IEEE 118-Bus system. Training convergence comparison on the IEEE 118-Bus system. QS-MADS reaches near-optimal performance substantially faster than all baselines. QS-MADS exhibits rapid, smooth convergence, reaching CDR = 0.80 within 600 training episodes—a milestone that CTDE-GNN does not reach until episode 1100, and that MADDPG does not reach during the full 2000-episode training budget. The smooth convergence profile of QS-MADS (low variance across seeds, visible in the tight curve) contrasts with the oscillatory behaviour of MADDPG and MAPPO, which exhibit high variance attributable to the non-stationarity induced by concurrent policy updates without coordinated exploration. The QS perturbation module mitigates this oscillation by ensuring that exploration noise is correlated with the current population state, preventing agents from simultaneously over-exploring in conflicting directions. Discussion Summary of advantages QS-MADS achieves consistent improvements across all benchmarks, metrics, and experimental conditions. The unifying factor is the quorum sensing mechanism: by providing agents with a lightweight, principled coordination signal derived from collective state criticality during training, QS-MADS resolves the exploration-exploitation tension in cooperative MARL without requiring pairwise exchange of high-dimensional hidden states. The ablation, robustness, scalability, and latency studies jointly support this interpretation: removing any component weakens performance, whereas the full architecture remains stable under measurement noise and fast enough for sub-minute scheduling. Limitations Despite strong performance, QS-MADS has several limitations that warrant acknowledgement. First, the current formulation assumes a single aggregation level (grid-wide consensus) during training, which may be suboptimal for geographically partitioned grids where local consensus within each partition would be more informative than a global one. This is acceptable for the tested IEEE 118-Bus, RTS-96, and CIGRE benchmarks because the simulated operating regions are not divided into independently controlled protection zones and the consensus vector is used for training-time exploration rather than mandatory runtime control. For geographically partitioned transmission areas, multi-feeder distribution networks, or microgrids connected through weak tie lines, a hierarchical QS variant would be necessary. Implementation is straightforward: compute zone-level signals inside each partition, then aggregate only the zone summaries into a higher-level consensus, increasing complexity from O(N) to O(N+Z) for Z zones while preserving scalar telemetry. Second, the QS signal dimensionality is fixed as a scalar; richer multi-dimensional QS signals encoding separate criticality dimensions (voltage, frequency, congestion) could further refine the perturbation module’s specificity. Third, while QS-MADS is robust to measurement noise up to σ = 0.30, the validation is still limited to simulated scenarios and non-adversarial measurement noise. The framework has not yet been evaluated under adversarial measurement attacks, communication delays, device outages, or Byzantine agent failures, which are important security considerations for real grid deployments. Future work Immediate directions include: (i) extending the QS aggregation to hierarchical multi-zone grids; (ii) incorporating uncertainty quantification into the GAT encoder to produce confidence-aware QS signals; and (iii) evaluating QS-MADS in hardware-in-the-loop grid simulators with real-time dispatch constraints. Longer-term, the QS-driven perturbation principle is general and could be transferred to other cooperative MARL domains including traffic signal control, multi-robot coordination, and network resource management. Conclusion We have presented QS-MADS, a novel multi-agent dispatch framework for distributed power grids that transplants the biological quorum sensing mechanism into cooperative deep reinforcement learning. By replacing ad hoc exploration noise with a differentiable, population-aware perturbation module driven by lightweight QS consensus signals during training, QS-MADS learns coordinated policies while retaining deterministic local execution at deployment. A complementary QS-shaped reward function further stabilises cooperative training by anchoring individual agent credit to population-level criticality. Extensive experiments across three public benchmark grids demonstrate that QS-MADS consistently outperforms five competitive baselines: achieving up to 25% higher dispatch reward, 31% fewer voltage violations, and 40% faster convergence, while maintaining robust performance under measurement noise, scaling efficiently to 64 agents, and requiring only 2.83± 0.13 ms inference latency on IEEE 118-Bus. Practically, these results suggest that QS-MADS can serve as a fast supervisory dispatch layer for DER-rich grids: operators can train policies offline or in digital twins, deploy deterministic local controllers with hard feasibility projection, and optionally monitor scalar QS criticality signals to detect emerging coordinated stress. The method therefore offers a deployable path between fully centralised optimisation and uncoordinated local control, while future work should validate the framework under hardware-in-the-loop settings, adversarial data conditions, and hierarchical multi-zone grid operation. Author contributions Kun Zeng led the conceptualization of the study, designed the framework. Yixiao Cao was responsible for dataset collection, preprocessing, and annotation. Hui Sun reviewed the paper. Data availability The datasets generated and/or analysed during the current study are not publicly available but are available from the corresponding author on reasonable request. 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 The datasets generated and/or analysed during the current study are not publicly available but are available from the corresponding author on reasonable request.