Energy-Efficient Optimization of Cloud-Native Deep Learning Services for Power Load Forecasting

Abstract

Power-load forecasting models are increasingly deployed as containerized cloud services. Variable and bursty demand can lead to resource overprovisioning, whereas CPU-based horizontal scaling often reacts too late. We propose PEAS, an energy-aware predictive autoscaler for online power-load forecasting. A GRU forecasts near-future request volume, and an upper prediction bound determines the target Pod count. Asymmetric cooldown and hysteresis suppress unnecessary scaling. The system deploys an LSTM forecasting service on Kubernetes and uses Prometheus and Kepler for service, resource, and energy measurements. Across steady, periodic, burst, and mapped-trace workloads, PEAS reduced mean P95 latency, failure rate, and energy per successful request by 55.1%, 84.0%, and 25.4% relative to CPU-HPA. CPU utilization increased by 15.3%. PEAS also remained robust under request noise and across model sizes. These results show that predictive lead time, uncertainty compensation, and scaling stability can jointly improve the performance-energy trade-off of cloud-native power-system inference.