Optimización del consumo energético en centros de datos mediante modelos de regresión múltiple basados en carga de procesamiento y métricas de red

Abstract

Global digitalization has established data centers as critical infrastructure, yet their energy consumption is increasing disproportionately, posing significant sustainability challenges. The absence of physical measurement sensors in virtualized servers has driven the adoption of software-based predictive models. This study conducted a systematic review using the PRISMA protocol (2021-2026) focused on Multiple Linear Regression (MLR) models for estimating energy consumption in data centers. SNMP metrics from CPU, network, and memory were analyzed, with preprocessing steps including imputation, normalization, and mitigation of multicollinearity. Hybrid MLR models incorporating Gradient Boosting achieved a mean absolute error (MAE) of 1.46 and an R² exceeding 95%. These findings demonstrate that software models can effectively replace physical sensors, supporting dynamic virtual machine consolidation and predictive HVAC control. Future research directions include the application of Causal ML to identify the etiological origins of thermal stress, Symbolic Regression adaptable to hardware changes, and extrapolation to hybrid renewable networks in rural edge computing environments. Hybrid MLR serves as a precise substitute for physical instrumentation, reduces idle consumption to within 10% of the ideal limit, and enables proactive climate control. The integration of Causal ML and Symbolic Regression is necessary to address stochasticity in future architectures.