Year: 2026
Type: article
Abstract: This study develops an artifact that improves monitoring efficiency and forecasts monthly Total Load (KW) for the X data hall. The manual Power Distribution Unit (PDU) monitoring process was mapped using an As-Is value stream map, yielding a cycle time of 190 minutes and a lead time of 220 minutes. A To-Be automated monitoring design reduces the expected cycle time to 90 minutes and lead time to 120 minutes, supported by sampling checks for data integrity. Forecasting couples SARIMAX with an Artificial Intelligence (AI) residual learner. Using 36 months of observations (Oct 2021 to Sep 2024), SARIMAX captures linear dynamics, annual seasonality, and exogenous drivers from ambient temperature, cooling degree days (18°C baseline), and heatwave and maintenance flags. Support vector regression then models nonlinear residual structure. Evaluation showed stable accuracy with holdout MAPE of 2.11% and rolling-origin MAPE of 3.03%. The refit hybrid model generated two-year forecasts and scenario sensitivity outputs to support capacity planning and load balancing. A Power BI dashboard integrates PDU readings, forecasts, and anomaly context for robust decision support. Limitations include monthly granularity and modest sample size; future work should use higher-frequency telemetry and probabilistic uncertainty calibration. An indicative economic evaluation shows that the proposed automation can achieve a payback below one year under conservative energy saving assumptions.
Source: Applied Operations and Analytics
Author: Samer Abaddi
Institution: Identity
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7147545317