LiteCast: A Lightweight Forecaster for Carbon Optimizations

Mathew Joseph, Tanush Savadi, Abel Souza · 2025

Read LiteCast: A Lightweight Forecaster for Carbon Optimizations by Mathew Joseph, Tanush Savadi, Abel Souza (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-06187.txt.

Introduction

Mathew Joseph, Tanush Savadi, and Abel Souza present LiteCast, a lightweight way to forecast the carbon intensity of an electricity grid. Carbon-aware computing shifts work to cleaner times, but it usually leans on heavy forecasting models fed with long histories. The authors argue that exact accuracy matters less than ranking future hours in the right order.

LiteCast models a region's energy mix from only a few days of energy and weather data. It runs fast and adapts quickly when grid conditions change suddenly. Tested in fifty regions worldwide under real workloads, it delivered twenty percent higher savings than leading forecasters and reached ninety-seven percent of the best savings possible.

Catalog record

CREATORMathew Joseph, Tanush Savadi, Abel Souza
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE51 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-06187

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

LiteCast: A Lightweight Forecaster for Carbon Optimizations Mathew Joseph, Tanush Savadi, Abel Souza Abstract . Over recent decades, electricity demand has experienced sustained growth, driven primarily by structural shifts such as the widespread electrification of transportation and the accelerated expansion of Artificial Intelligence (AI). Grids have managed the resulting surges by scaling generation capacity, incorporating additional resources such as solar and wind, and implementing demand-response mechanisms. Altogether, these policies influence a region’s carbon intensity by affecting its energy mix: in some cases, they displace higher-emission sources during peak periods and lower emissions, while in others, upward variability derives from carbon-intensive peaking units that increase emissions. To mitigate the environmental impacts of consumption, carbon-aware optimizations often rely on long-horizon, high-accuracy forecasts of the grid’s carbon intensity that typically use compute intensive models with extensive historical energy mix data. In addition to limiting scalability, accuracy improvements do not necessarily translate into proportional increases in savings. Highlighting the need for more efficient forecasting strategies, we argue that carbon forecasting solutions can achieve the majority of savings without requiring highly precise and complex predictions. Instead, it is the preservation of the ranking of forecasts relative to the ground-truth that drives realized savings. In this paper, we present LiteCast , a lightweight time series forecasting method capable of quickly modeling a region’s energy mix to estimate its carbon intensity. LiteCast requires only a few days of historical energy and weather data, delivering fast forecasts that can quickly adapt to sudden changes in the electrical grid. Our evaluation in 50 worldwide regions under various real-world workloads shows that LiteCast outperforms state-of-the-art forecasters, delivering 20% higher savings with near-optimal performance, achieving 97% of the maximum attainable average savings, while remaining lightweight, efficient to run, and adaptive to new data. 1. Introduction The last decades have witnessed global electricity consumption increase at an accelerated pace, driven by a combination of structural and technological shifts. Contributing factors include electrification of the residential ( 4 ) , industrial, and transportation sectors, alongside the rapid proliferation of Artificial Intelligence (AI) applications and the large-scale expansion of data center infrastructures ( 2 ) . This surge has underscored the growing need to monitor and mitigate the environmental impacts of energy use. Many of these initiatives have focused primarily on improving energy efficiency and implementing accounting offsets, in which certificates are swapped in exchange for the procurement of clean energy from solar and wind projects ( 19 ) . While these policies have accelerated grid decarbonization, they also present new challenges: the variable and less predictable nature of renewable generation disrupts traditional power flows and system dynamics. As renewable integration deepens, ensuring grid reliability, controlling costs, and minimizing environmental impacts requires increasingly effective and coordinated, real-time planning. In response, researchers have begun exploring alternative strategies to reduce the environmental impact and improve the carbon efficiency of electricity consumption. To achieve carbon-efficiency, both supply- and demand-side approaches have been adopted. On the supply side, consumers use power purchase agreements (PPAs) of renewable energy sources, such as wind and solar, through long-term contracts to power their operations, commonly used in large-scale industries, e.g., data centers ( 19 ) . On the demand side, researchers have developed techniques to modulate demand and its associated emissions by shifting it to times and locations with low-carbon electricity. For instance, in buildings, flexible loads such as laundry, water heating, and residential electric vehicle (EV) can be deferred based on knowledge of low-intensity grid conditions. Likewise, “spatial strategies” can be employed to cloud workloads with significant geodistributed flexibility that allow them not only to change time, but also location to better align execution with the availability of low carbon energy ( 29 ; 14 ; 32 ; 12 ; 11 ) . Generally, most carbon-aware optimizations assume near perfect knowledge of the grid’s carbon intensity ( 35 ) , which may not always be accurate, limiting the potential of their solutions ( 23 ) . These optimizations have been made possible by the emergence of third-party services such as Electricity Maps ( 21 ) and WattTime ( 34 ) that provide real-time and predictive mechanisms with carbon and energy data, enabling the coordinated optimization of demand-side consumers and distributed compute workloads ( 22 ; 35 ) . State-of-the-art models apply complex, compute-intensive deep neural networks and ensemble models to understand the intra- and inter- spatiotemporal relationships between the various energy sources to forecast up to the next four days of a grid’s carbon intensity ( 7 ; 36 ; 39 ) . For instance, CarbonCast ( 7 ) applies a two-tier neural network using historical energy, carbon, and weather data to improve carbon intensity forecasts. Although these methods generally achieve high prediction accuracy across many regions—with few exceptions in grids with high wind penetration—their effectiveness in scheduling demand loads to optimally reduce their carbon footprint remains uncertain. In addition, while these solutions are theoretically sound, many lack computational scalability, as they rely heavily on large volumes of historical data to perform well. This hampers low-latency performance, making real-time, adaptive operations more complex, time-consuming and costly in practice. This is especially important in today’s evolving power grids, which undergo frequent upgrades and not only enhance inter-grid connections but also incorporate increasingly complex local energy systems, often featuring significant energy storage such as batteries. In this paper, we introduce LiteCast , a lightweight forecasting mechanism that enables fast, adaptive spatiotemporal carbon-optimizations. In contrast to state-of-the-art solutions that prioritize accuracy through complex models, LiteCast leverages the insight that carbon-aware scheduling does not always require precise carbon intensity data to minimize emissions. Instead, we argue that for effective scheduling, models should focus less on minimizing error metrics (such as MAPE, RMSE, or MSE), and more on accurately capturing the timing and ordering of carbon intensity forecasts because they better drive carbon-aware scheduling decisions to reflect real emissions. LiteCast uses a SARIMAX method (Seasonal Auto Regressive Integrated Moving Average with Exogenous Regressor) that incorporates seasonality along with exogenous variables such as weather and demand forecasts across various worldwide regions. More importantly, LiteCast requires less than 5% of the historical data required by complex models, and is capable of being retrained in a very low-latency manner, without using large-scale resources. We experiment LiteCast with various types of continuous and non-continuous loads and across various worldwide regions with varying power grid profiles. Compared to an oracle with perfect future knowledge, our approach achieves up to 97% of the potential savings while remaining highly lightweight and efficient to run. In evaluating our hypothesis, this paper makes the following contributions: • Lightweight Forecasting. We present LiteCast , a lightweight model that uses a standard time-series method along with exogenous data to understand and forecast the grid’s carbon-intensity. • Large-scale data analysis. We analyze the performance of LiteCast across 50 regions worldwide. • Together with LiteCast , we introduce a heuristic that enhances forecasts by leveraging the temporal flexibility of jobs, dynamically shifting their schedules as new information becomes available while still meeting their SLO guarantees. • Large-scale analysis of forecasting implications on carbon-aware optimizations. We leverage state-of-the-art carbon-aware temporal workload shifting approaches, and spatial workload migration strategies, to quantify the actual carbon savings using LiteCast . Our analysis in

Read the complete text