Temporal Request-Frequency Prediction for Predictive Computing in Generative AI: Forecasting Future Query Volumes Across Multiple Time Horizons

The increasing adoption of generative artificial intelligence has created substantial and continuously changing computational demand for large language model (LLM) inference services. Although existing inference optimization approaches primarily respond to incoming requests or reuse previously generated computational results, predictive computing introduces the possibility of preparing computational resources before actual demand materializes. However, effective predictive computation requires more than identifying which questions may appear in the future. It also requires estimating how frequently different questions, query categories, and information topics will be requested across future time intervals. This paper introduces Temporal Request-Frequency Prediction (TRFP), a forecasting framework designed to estimate future request volumes for generative AI services across multiple temporal horizons. Unlike binary query-occurrence prediction, which primarily estimates whether a future request will appear, TRFP focuses on estimating the expected number and temporal distribution of future requests associated with particular questions or query categories. The proposed framework integrates historical request statistics, temporal behavioral patterns, emerging information trends, external event signals, query-category representations, and short-term demand fluctuations to generate multi-horizon request-frequency forecasts. Its architecture incorporates temporal demand representation, query classification and clustering, multi-horizon request-count prediction, probabilistic uncertainty estimation, prediction updating, and forecasting-aware computational planning. A mathematical formulation is developed to estimate future request counts, distinguish request occurrence probability from expected request frequency, and quantify forecasting errors across different time intervals. Furthermore, the paper establishes a theoretical connection between predicted request frequency and proactive computational preparation, demonstrating how estimated future query volumes can inform the selection, prioritization, and scheduling of advance-generated responses and reusable computational resources. An experimental methodology is proposed to evaluate forecasting accuracy, temporal consistency, prediction lead time, and the resulting computational implications under chronological testing conditions. The framework does not assume that future request frequencies are perfectly predictable or that accurate forecasting necessarily guarantees computational savings. Instead, it investigates the conditions under which multi-horizon request-frequency prediction can provide useful information for allocating computational resources in generative AI systems. The proposed research extends Predictive Computing for Generative AI from anticipating future information requirements toward quantitatively forecasting the intensity and temporal distribution of future computational demand.