Comparative Analysis of Zero-Shot Capability of Time-Series Foundation Models in Short-Term Load Prediction

TL;DR

This study investigates the zero-shot forecasting capability of Time-Series Foundation Models (TSFMs) for short-term load prediction (STLP) by comparing Chronos, Moment, TimesFM, Lag-llama, and TimeGPT against Gaussian Process ($GP$) and Support Vector Regression ($SVR$). The results show that, without task-specific training, TSFMs like Chronos and TimeGPT can outperform traditional baselines in both probabilistic and point predictions, with Chronos particularly strong in probabilistic forecasting. TSFMs offer practical advantages in data-scarce or privacy-constrained settings due to minimal or no additional training requirements. A key limitation is their inability to incorporate external conditions (e.g., weather), pointing to future work that integrates exogenous factors to further boost STLP accuracy.

Abstract

Short-term load prediction (STLP) is critical for modern power distribution system operations, particularly as demand and generation uncertainties grow with the integration of low-carbon technologies, such as electric vehicles and photovoltaics. In this study, we evaluate the zero-shot prediction capabilities of five Time-Series Foundation Models (TSFMs)-a new approach for STLP where models perform predictions without task-specific training-against two classical models, Gaussian Process (GP) and Support Vector Regression (SVR), which are trained on task-specific datasets. Our findings indicate that even without training, TSFMs like Chronos, TimesFM, and TimeGPT can surpass the performance of GP and SVR. This finding highlights the potential of TSFMs in STLP.

Figures (4)

• Figure 1: Two tokenization approaches for time series. $t$ is the length of the time series, and $b$ is the length of the token. The Lag-llama's approach is to segment time series equally to create a sequence of tokens (assume $t$ is divisible by $b$). The Chronos's approach is to map the real values to a token.
• Figure 2: Two output mechanisms. $\mathbf{z_i}$ is the token which is obtained by tokenization. $\mathbf{z_i}=\{x_{i\times b+1},x_{i\times b+2},...,x_{i\times b+b}\}$ in Lag-llama, and $\mathbf{z_i}=q(x_i)$ in Chronos.
• Figure 3: The prediction process of Chronos. $\{\mathbf{z}^{1:}_{t+1} \}$ is sampled from $\{ \mathbf{z}_{1:t}\}$ based on the probability $\{p_{\mathbf{z}_i}; i=1, 2,.., t \}$, where $\sum_i^t p_{\mathbf{z}_i} =1$.
• Figure 4: Prediction examples of GP, TimeGPT, Lag-llama, and Chronos.