A Multi-Task Learning Approach for Predicting Capacity Expansion Timing and Requirements in Colocation Datacenters

Abstract

Colocation data centers are essential infrastructure for the digital economy, supporting scalable and secure operations across industries. With the global colocation market exceeding $218 billion—fueled by data growth, technology, and economic expansion—intelligent long-term capacity planning is critical. Traditional reactive methods often miss localized demand shifts, leading to under-provisioning or costly over-investment. A key challenge lies in forecasting when and how much capacity will be needed across infrastructure segments, such as power delivery domains. We propose a multi-task learning framework to jointly predict the timing and magnitude of future capacity expansions. Our hybrid Transformer-based architecture integrates static and temporal features, such as facility telemetry, sector metadata, macroeconomic indicators, and sentiment signals, into a unified temporal embedding space with a static feature layer. It generates dual outputs: a binary classifier for expansion events and a conditional regressor for size. By modeling long-range dependencies and uncertainty, our approach enables accurate, adaptive forecasts that support proactive procurement, smarter resource allocation, and improved infrastructure agility