SmartMeterFM: Unifying Smart Meter Data Generative Tasks Using Flow Matching Models

TL;DR

SmartMeterFM proposes a unifying flow-matching framework to address generation, imputation, and super-resolution for high-dimensional smart meter time series. A Transformer-based velocity network learns a conditional velocity field $u_\theta(x,t)$ to transform a Gaussian prior into realistic monthly 15-minute profiles, with inference-time guidance enabling task adaptation without re-training. The method achieves superior or competitive performance across tasks compared to specialized baselines, demonstrated on a large Dutch dataset with multiple customer categories and realistic constraints. This approach offers a flexible, privacy-conscious tool for grid planning and operation by producing high-fidelity, condition-consistent synthetic data across diverse scenarios.

Abstract

Smart meter data is the foundation for planning and operating the distribution network. Unfortunately, such data are not always available due to privacy regulations. Meanwhile, the collected data may be corrupted due to sensor or transmission failure, or it may not have sufficient resolution for downstream tasks. A wide range of generative tasks is formulated to address these issues, including synthetic data generation, missing data imputation, and super-resolution. Despite the success of machine learning models on these tasks, dedicated models need to be designed and trained for each task, leading to redundancy and inefficiency. In this paper, by recognizing the powerful modeling capability of flow matching models, we propose a new approach to unify diverse smart meter data generative tasks with a single model trained for conditional generation. The proposed flow matching models are trained to generate challenging, high-dimensional time series data, specifically monthly smart meter data at a 15 min resolution. By viewing different generative tasks as distinct forms of partial data observations and injecting them into the generation process, we unify tasks such as imputation and super-resolution with a single model, eliminating the need for re-training. The data generated by our model not only are consistent with the given observations but also remain realistic, showing better performance against interpolation and other machine learning based baselines dedicated to the tasks.

Figures (6)

• Figure 1: Demonstration of projection-based guidance. At every step, we shift the velocity by projecting an estimation of the destination into the feasible region.
• Figure 2: Diagram of SmartMeterFM sampling process. The teal lines represent the data generation process, with the orange indicating the normal conditional generation part, and the red showing the additional guidance. The velocity network is a neural network.
• Figure 3: Architectural design of the velocity network, as used in Fig. \ref{['fig:ours-diagram']}. Red arrows indicate input to the neural network, while green arrow indicates the output. The details of folding, initial convolution and final projection can be found in lin2025EnergyDiffUniversalTimeseriesa.
• Figure 4: MMD permutation test results of SmartMeterFM. Blue areas are the histograms of permuted MMD values, with the top bar indicating the maximum, the bottom bar indicating the minimum, and the middle bar indicating the $0.95$ quantile. The red dots present the actual MMD value. Low MMD and high p-value suggest a small difference between the generated data and the real data.
• Figure 5: Imputation results of one sample using three different approaches. The gray areas are observed while the white areas are missing, divided by the dotted red lines.