Bridging Smart Meter Gaps: A Benchmark of Statistical, Machine Learning and Time Series Foundation Models for Data Imputation

TL;DR

This study tackles missing data in smart-meter time series by benchmarking a spectrum of models—statistical baselines, traditional ML, and time-series foundation models—via a forward–backward–interpolation gap-filling framework. Using a public London dataset with artificially introduced gaps, it reveals that time-series foundation models, notably TimeMoE, can achieve superior imputation accuracy without dataset-specific training, albeit with substantial computational costs. Traditional methods like Holt-Winters and Random Forest remain strong, reliable baselines, while simple baselines such as Linear Interpolation perform surprisingly well in some settings. The work highlights the practical trade-offs between accuracy and compute, and suggests future work in fine-tuning open TS models and exploring longer or differently structured gaps to optimize data imputation in smart grids.

Abstract

The integrity of time series data in smart grids is often compromised by missing values due to sensor failures, transmission errors, or disruptions. Gaps in smart meter data can bias consumption analyses and hinder reliable predictions, causing technical and economic inefficiencies. As smart meter data grows in volume and complexity, conventional techniques struggle with its nonlinear and nonstationary patterns. In this context, Generative Artificial Intelligence offers promising solutions that may outperform traditional statistical methods. In this paper, we evaluate two general-purpose Large Language Models and five Time Series Foundation Models for smart meter data imputation, comparing them with conventional Machine Learning and statistical models. We introduce artificial gaps (30 minutes to one day) into an anonymized public dataset to test inference capabilities. Results show that Time Series Foundation Models, with their contextual understanding and pattern recognition, could significantly enhance imputation accuracy in certain cases. However, the trade-off between computational cost and performance gains remains a critical consideration.

Figures (5)

• Figure 1: Distribution of randomly generated gaps across the evaluated data points
• Figure 2: Prompts used for GPT-4o and Llama 3.1
• Figure 3: and of the best-performing models (green) and least-performing models (blue) per category
• Figure 4: Relationship Between Parameters and : Bubble Size Represents Parameter Count, Color Indicates (Lower is Better)
• Figure 5: Gap-filling results for the best and least-performing models on a household electricity consumption time series