Predictive Electricity Price Forecasting for Generative AI: Anticipating Future Energy Costs for Proactive Computational Scheduling

The increasing computational requirements of generative artificial intelligence (AI) systems have made electricity expenditure an increasingly important component of inference infrastructure management. Large language model (LLM) inference, proactive response preparation, model loading, speculative computation, and other GPU-intensive operations may be performed across different time periods, during which electricity prices can vary substantially. Although existing economic-aware predictive computing approaches consider dynamic electricity pricing when selecting computational execution periods, incorporating electricity prices into scheduling decisions is not equivalent to predicting future electricity prices with sufficient temporal and probabilistic accuracy. Without explicit forecasting of future energy costs, computational schedulers may rely on historical averages, currently observed prices, or externally supplied pricing information, potentially overlooking future price fluctuations and the opportunities they create for advance computational preparation. This paper introduces Predictive Electricity Price Forecasting for Generative AI (PEPF-GAI), a specialized framework that investigates the prediction of future electricity prices and their integration into proactive computational scheduling. Unlike conventional price-aware scheduling, which primarily evaluates the cost of computation using available electricity price information, PEPF-GAI explicitly models the future temporal evolution of electricity prices and translates probabilistic price forecasts into computational execution decisions. The proposed architecture integrates historical electricity price analysis, temporal pattern extraction, renewable energy availability estimation, electricity demand forecasting, market condition modeling, multi-horizon price prediction, prediction uncertainty estimation, and cost-aware computational scheduling. It considers different electricity pricing environments, including time-of-use tariffs, real-time pricing, day-ahead electricity markets, and location-dependent electricity prices. A unified mathematical formulation is developed to estimate future electricity expenditure, quantify the expected economic value of shifting computational workloads across time periods, and jointly optimize computational execution schedules under GPU availability, task deadlines, prediction uncertainty, and service-level constraints. The framework further distinguishes between forecasting electricity prices, forecasting total electricity expenditure, and predicting the economic benefits of computational rescheduling. These represent related but fundamentally different prediction problems. An experimental methodology is established to compare historical-average pricing, persistence forecasting, conventional time-series forecasting, machine-learning-based price prediction, and uncertainty-aware predictive scheduling. Evaluation metrics include price forecasting accuracy, directional prediction accuracy, economic regret, realized electricity expenditure, computational scheduling efficiency, and total system operating cost. The proposed research extends economic-aware predictive computing from the utilization of available electricity price information toward the explicit anticipation of future electricity prices. Rather than assuming that lower-price periods can always be identified reliably or that computation should automatically be shifted whenever electricity prices decrease, PEPF-GAI investigates the conditions under which electricity price forecasting can support economically beneficial and operationally feasible predictive computation. Its effectiveness remains subject to empirical validation across different electricity markets, tariff structures, forecasting horizons, and computational workloads.