Intelligent Electrical Capacity Forecasting for AI Workloads in Modular Data Centre Expansion Planning

As the number of workloads for AI training and inference grows at a rapid rate, hyperscale and modular data centres have experienced unprecedented volatility and variability in electrical demand profiles, which is threatening the traditional deterministic capacity expansion planning approaches. This paper combines recent developments in electrical load forecasting, deep learning architectures, and risk-aware infrastructure planning, and proposes an integrated framework for intelligent forecasting capability of electrical capacity in the context of AI-workload based modular expansion of the data centre. The framework integrates probabilistic time-series forecasting models, such as long short-term memory (LSTM) networks and temporal fusion transformers (TFT), with a conditional value-at-risk (CVaR) constrained optimization layer that maps in uncertain demand forecasts into capacity expansion decisions, done in steps and in sections. When compared to other forecasting methods, the hybrid attention-based models show mean absolute percentage errors of less than four percent, much lower than the mean absolute percentage errors of classical statistical baseline models that are typically over nine percent when applied to similar short-term forecasting problems. The analysis further shows that in successive phases of expanded use, the combination of forecast-driven provisioning and on-site renewable generation and energy storage can reduce the marginal carbon intensity by over 50% compared to the traditional static deterministic approach to provisioning, while also improving capital efficiency and grid resilience. Comparative benchmarking is done to assess efficiency, cost, scalability, and regulatory dimensions, and shows several times greater growth in facility demand for AI workloads when compared to enterprise workloads in more traditional baselines. However, the synthesis identifies longstanding issues such as congestion in interconnection queues, forecast uncertainty propagation, and harmonisation of data-sharing processes between the operators of data centres and the operators of the transmission system. The proposed framework provides planners with a repeatable, scientifically rigored and mathematically based approach that is applicable to the ever expanding and rapidly evolving electrical use of AI computing.