Infrastructure Systems: Smart, Resilient, Energy Systems

REU Scholar: Rasel Ali

REU Scholar Home Institution: Lehman College

REU Mentor: Dr. Yufei Tang

Zero-Shot Forecasting for Smart Cities: Evaluating TimeGPT Against LSTM and TCN Models

Accurate time series forecasting is essential for optimizing critical smart city systems, from energy grid management to climate modeling. Among these, predicting electricity demand is particularly vital for informed decision-making and future preparedness. Traditional deep learning models, such as Long Short-Term Memory (LSTM) networks and Temporal Convolutional Networks (TCNs), have achieved strong performance but require extensive dataset-specific training, hyperparameter tuning, and lengthy training cycles. This study presents a comparative analysis of these established architectures against TimeGPT, a state-of-the-art pre-trained foundational model for time series forecasting. We evaluate performance across diverse real-world datasets, including high-frequency power consumption and multivariate climate data, examining accuracy, generalizability, and operational efficiency. Our findings reveal a paradigm shift in forecasting methodologies. While custom-trained TCN and LSTM models deliver high accuracy, their performance is often matched or surpassed by TimeGPT's zero-shot capabilities. By leveraging its broad pre-trained knowledge of temporal patterns, TimeGPT consistently produces robust forecasts with minimal setup and no additional training, effectively addressing challenges such as missing data and complex seasonality. These results highlight foundational models as transformative tools for rapid, scalable, and reliable forecasting, accelerating the development of intelligent solutions for smart cities.

Watch video presentation
REU Scholar: Rasel Ali