AI-Driven Demand Forecasting for Capacity Planning in Multi-Tenant Systems

Abstract

Capacity planning in multi-tenant systems is a critical challenge due to dynamic workloads, resource contention, and the need for tenant isolation. Traditional forecasting methods, such as ARIMA and exponential smoothing, struggle to adapt to the heterogeneity and volatility of multi-tenant environments. This paper proposes a scalable AI-driven framework for demand forecasting and capacity planning, leveraging hybrid architectures like LSTM-Transformer ensembles and transfer learning. We validate our approach using a cloud-based Kubernetes testbed, synthetic datasets, and anonymized real-world workload traces. Experimental results demonstrate a 34% improvement in forecasting accuracy (RMSE) over statistical baselines and a 40% reduction in over-provisioning costs. The framework achieves sub-second latency for real-time decision-making while scaling to 1,000+ tenants.