Causal Invariant Pareto Surrogates: A Hybrid Machine Learning and Multi-Objective Optimization Framework for Resilient Smart Campus Resource Management

Smart campus resource management faces a critical challenge: optimization strategies trained on historical data often fail when environmental conditions shift, such as seasonal transitions or unexpected occupancy changes. We propose a causal invariant Pareto surrogate framework that couples causal representation learning with multi-objective evolutionary optimization to address this distribution shift problem. The architecture begins with a Variational Causal Encoder that disentangles high-dimensional telemetry into invariant causal factors and spurious context factors, where an invariance penalty minimizes mutual information between causal factors and environment labels. These invariant factors then serve as inputs to a surrogate model that replaces computationally expensive physics-based simulations during the optimization loop. A counterfactual testing mechanism evaluates candidate solutions under perturbed environmental contexts, and solutions exhibiting inconsistent predictions are penalized within the evolutionary selection process. The surrogate is integrated into a Pareto-based evolutionary algorithm, where dominance relations are determined using predicted sustainability objectives including energy consumption, carbon emissions, comfort, and cost. Moreover, the selection pressure incorporates a causal consistency score, which guides the population toward strategies effective due to underlying physical mechanisms rather than transient correlations. We evaluate the framework on heterogeneous smart campus telemetry partitioned into meteorological, occupancy, and academic calendar domains. The primary contribution is a unified pipeline that produces Pareto-optimal resource management strategies with demonstrated robustness across environmental distribution shifts. Furthermore, this approach offers a generalizable methodology for sustainable building control, where long-term adaptability is essential and where conventional correlation-based surrogates prove insufficient.