An explainable reinforcement learning framework for energy-aware cluster head selection in smart-city IoT wireless sensor networks

Abstract

Cluster head (CH) selection is a decisive control mechanism for prolonging the operational lifetime of Internet-of-Things (IoT)-enabled wireless sensor networks (WSNs) that underpin smart-city sensing. Although reinforcement learning (RL) has produced increasingly capable CH selectors, three weaknesses persist across the recent literature: the learned policies are opaque and therefore unsuitable for mission-critical operations; they are validated almost exclusively under stationary traffic that does not reflect urban traffic loads; and they ignore the energy-harvesting heterogeneity that characterises real deployments. This study presents eXplainable Energy-Aware Clustering Hierarchy (XEACH), an explainable, energy-aware RL framework that addresses all three limitations within a single design. We formalise CH selection as a constrained network-lifetime maximisation problem, establish the NP-hardness of its per-round subproblem, and recast the sequential decision process as a Markov decision process in which each node learns a state-conditioned value for cluster-head service and selection is driven directly by the learned policy. We prove the convergence of the learner under standard stochastic-approximation conditions and bound the per-round cluster-head count. To render the policy auditable, we attach a post hoc explainability module based on SHapley Additive exPlanations (SHAP) that attributes every CH decision to its driving features. The framework is evaluated on smart-city scenarios of 50–300 nodes, incorporating a non-stationary traffic model with diurnal congestion and stochastic event spikes, together with 30% energy-harvesting nodes. All results are averaged over multiple independent seeds and reported with standard deviations. Against LEACH, SEP, and HEED, XEACH extends first-node-death by 88% over LEACH at 100 nodes and consistently outperforms all three baselines across network sizes, with the margin over LEACH widening as the network grows. An ablation confirms that the reinforcement-learning policy, rather than any fixed heuristic, drives these gains: replacing the learned policy with a random policy collapses lifetime by an order of magnitude. The SHAP analysis of fidelity 0.87 identifies residual energy and distance to the base station as the dominant, physically coherent drivers of selection. We additionally calibrate the energy model to the measured datasheet parameters of a current LoRa transceiver to establish the real-time feasibility of deployment.