Energy sustainability and secure operation are persistent challenges in Internet-of-Things (IoT) wireless sensor networks (WSNs), where limited battery capacity, heterogeneous traffic, and security procedures jointly drive premature node depletion and service degradation. This paper proposes an uncertainty-aware bilevel co-optimization framework that unifies residual-energy prediction with robust, energy-aware scheduling for clustered IoT-WSNs. At the lower level, a lightweight temporal predictor (TCN + LSTM with stochastic sampling) learns short-horizon residual-energy evolution from multivariate, dataset-aligned windows capturing sensing/communication activity, proximity-to-cluster-head effects, and security overhead (authentication latency, key exchange, and rekeying), and produces both point forecasts and uncertainty estimates to enable risk-sensitive control. At the upper level, a constrained, horizon-based scheduler selects per-node actions (duty cycle, sensing rate, transmission power) to extend network lifetime and balance residual energy while enforcing safety thresholds and operational bounds; bilevel coupling is realized via differentiable hypergradient updates, complemented by trust-region action smoothing and adaptive primal–dual constraint handling to suppress energy-critical states under uncertainty. On a real-world WSN energy–security dataset, the proposed model attains the best lower-level learning performance with $\mathrm{M}\mathrm{A}\mathrm{E}=0.004$, $\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=0.006$, and $R^2=0.995$ for residual-energy regression, and up to $0.98$ accuracy/ $0.98$ F1 for secure-and-efficient classification. End-to-end scheduling results show that the full framework improves estimated network lifetime by up to $1.60\times$, reduces residual-energy variance to $0.60\times$, and lowers safety violations to $0.35\times$ relative to a fixed-policy baseline, demonstrating robust, secure, and sustainable IoT-enabled WSN operation.