A Unified Variational Imputation Framework for Electric Vehicle Charging Data Using Retrieval-Augmented Language Model

TL;DR

This work addresses the challenge of missing EV charging data by introducing PRAIM, a unified probabilistic imputation framework that leverages retrieval-augmented language model embeddings to fuse multimodal context (time-series, calendar, geospatial features) into a single universal model. PRAIM enriches context with a retrieval-augmented memory (RAG), encodes it in a variational latent space, and uses a FiLM-modulated transformer decoder to produce a Gaussian distribution for imputed demands, offering both predictions and uncertainty estimates. Across four public datasets, PRAIM outperforms a broad set of baselines in imputation accuracy and in preserving the data’s distribution, with notable gains in downstream forecasting performance. The approach demonstrates strong robustness to high missingness, adapts via non-parametric memory for concept drift, and remains scalable for large EV charging networks, marking a significant advance in multimodal, uncertainty-aware time-series imputation. This framework sets the stage for applying LLM/RAG-based imputation to other complex domains with sparse, heterogeneous data streams.

Abstract

The reliability of data-driven applications in electric vehicle (EV) infrastructure, such as charging demand forecasting, hinges on the availability of complete, high-quality charging data. However, real-world EV datasets are often plagued by missing records, and existing imputation methods are ill-equipped for the complex, multimodal context of charging data, often relying on a restrictive one-model-per-station paradigm that ignores valuable inter-station correlations. To address these gaps, we develop a novel PRobabilistic variational imputation framework that leverages the power of large lAnguage models and retrIeval-augmented Memory (PRAIM). PRAIM employs a pre-trained language model to encode heterogeneous data, spanning time-series demand, calendar features, and geospatial context, into a unified, semantically rich representation. This is dynamically fortified by retrieval-augmented memory that retrieves relevant examples from the entire charging network, enabling a single, unified imputation model empowered by variational neural architecture to overcome data sparsity. Extensive experiments on four public datasets demonstrate that PRAIM significantly outperforms established baselines in both imputation accuracy and its ability to preserve the original data's statistical distribution, leading to substantial improvements in downstream forecasting performance.

Figures (9)

• Figure 1: Distribution of missing daily record percentages in the four datasets. The dashed and solid lines indicate the mean and median, respectively.
• Figure 2: An example of prompt construction for EV charging station data.
• Figure 3: Illustration of RAG for embedding augmentation.
• Figure 4: The workflow of PRAIM. Inputs per station $i$ and day $t$: $v_i$ (station ID), $\mathbf{m}_{t,i}$ (missing data indicators), $\tilde{\mathbf{c}}_t$ (calendar), $\mathbf{g}_i$ (coordinates), $\mathcal{POI}_{t,i}$, and $\mathbf{d}_{t,i}^\mathrm{ma, no}$ (historical demands). The LLM generates an embedding $\boldsymbol{\omega}_{t,i}$ based on the structured prompt, further refined by RAG as $\hat{\boldsymbol{\omega}}_{t,i}$. The following variational layer samples the latent variable $\mathbf{z}_{t,i}$ based on the posterior $q(\mathbf{z}_{t,i} | \hat{\boldsymbol{\omega}}_{t,i})$, which is fed into FiLM to produce parameters, $\boldsymbol{\gamma}_{t,i}$ and $\boldsymbol{\beta}_{t,i}$, to modulate the input demands into $\boldsymbol{\gamma}_{t,i}$ and $\boldsymbol{\beta}_{t,i}$. The final sequence decoder maps the modulated demands into the Gaussian distribution of reconstructed charging demands: $\mathcal{N}( \hat{\mathbf{d}}_{t,i}^\mathrm{no} | \hat{\boldsymbol{\mu}}_{t,i}, \hat{\boldsymbol{\sigma}}_{t,i}^2 )$.
• Figure 5: MAE comparison between PRAIM and benchmarks across used four datasets.

Theorems & Definitions (4)

• Remark 1: Adapativity to Concept Drift
• Remark 2: Data Integrity in RAG
• Remark 3: LLM Hallucination
• Remark 4: Model Extensibility to Finer Temporal Resolution