Institution profile

EDF R&D

Industry researcheurope · fr
Official website
Research library71linked papers
Opportunities0open roles
Selected work

Representative Papers

DeviceScope: An Interactive App to Detect and Localize Appliance Patterns in Electricity Consumption Time Series

Jun 06, 2025

This work addresses the challenge of detecting and localizing appliance-level on/off events from aggregated smart meter data—without access to device-level ground-truth labels—specifically targeting non-expert end users. Method: We propose CamAL, a weakly supervised localization framework that leverages Class Activation Mapping (CAM) integrated with temporal convolutional networks, requiring only household-level appliance presence labels for training. We further introduce an interactive visualization system (built with React and D3) enabling user-driven pattern verification and active learning feedback. Contribution/Results: Evaluated on real-world smart meter datasets, CamAL achieves an 89.2% F1-score and a mean localization error of ±12.3 seconds, substantially reducing annotation effort. The system has been deployed in pilot programs across three European utility providers, advancing the practical adoption of fine-grained electricity consumption behavior analysis.

1 citations1 influentialRead paper

Few Labels are all you need: A Weakly Supervised Framework for Appliance Localization in Smart-Meter Series

Jun 06, 2025

To address the scarcity of fine-grained appliance-level labels in non-intrusive load monitoring (NILM), this paper proposes CamAL, a weakly supervised appliance pattern localization framework that requires only household-level appliance presence labels—eliminating the need for time-stamped, instance-level annotations. CamAL integrates deep classifier ensembles, gradient-weighted class activation mapping (Grad-CAM), and weakly supervised temporal modeling to achieve interpretable, precise localization of appliance operational periods. Evaluated on four real-world datasets, CamAL significantly outperforms existing weakly supervised NILM methods. Remarkably, it achieves performance comparable to state-of-the-art fully supervised approaches using only a minimal number of coarse labels, thereby drastically reducing annotation effort and cost. This work establishes a highly efficient and practical paradigm for NILM under severe label scarcity, advancing both interpretability and scalability in real-world energy disaggregation applications.

1 citationsRead paper

Variational Parameter Calibration with Physics-Aware Latent-Space Surrogates

Aug 11, 2026

This work addresses the challenge of inverse parameter calibration in computational physics, where existing surrogate models often lack end-to-end differentiability and physical awareness, limiting their effectiveness. To overcome this, the authors propose a physics-informed latent-space framework based on an autoencoder architecture. The approach enables offline training of a differentiable surrogate model under observable supervision, mapping physical parameters to flow field predictions while embedding variational calibration directly in the latent space. By seamlessly integrating physical constraints with data-driven learning, the method achieves fully end-to-end differentiable surrogate modeling—a first in this domain. Evaluations on two computational fluid dynamics benchmarks demonstrate that, under realistic conditions including noise, low resolution, and partial observability, the proposed framework significantly reduces both calibration error and solution variability.

0 citationsRead paper

Balanced Twins: Causal Inference on Time Series with Hidden Confounding

Jun 17, 2026

This study addresses causal inference in time series settings with latent confounding and staggered interventions, where conventional methods are constrained by explicit temporal assumptions or convex combination requirements. The authors propose a deep representation learning framework that jointly learns low-dimensional individual latent representations and propensity scores, enabling estimation of individual counterfactual outcomes through a flexible nonparametric matching mechanism—without presupposing a specific temporal dynamics structure. By circumventing the convexity constraints inherent in synthetic control methods, the approach effectively mitigates bias arising from unobserved confounders. Empirical evaluations on both electricity demand response and ICU clinical datasets demonstrate that the proposed method significantly improves the accuracy of counterfactual prediction and average treatment effect estimation in nonstationary dynamic environments.

0 citationsRead paper
Recent publications

Latest Papers

Variational Parameter Calibration with Physics-Aware Latent-Space Surrogates

Aug 11, 2026

This work addresses the challenge of inverse parameter calibration in computational physics, where existing surrogate models often lack end-to-end differentiability and physical awareness, limiting their effectiveness. To overcome this, the authors propose a physics-informed latent-space framework based on an autoencoder architecture. The approach enables offline training of a differentiable surrogate model under observable supervision, mapping physical parameters to flow field predictions while embedding variational calibration directly in the latent space. By seamlessly integrating physical constraints with data-driven learning, the method achieves fully end-to-end differentiable surrogate modeling—a first in this domain. Evaluations on two computational fluid dynamics benchmarks demonstrate that, under realistic conditions including noise, low resolution, and partial observability, the proposed framework significantly reduces both calibration error and solution variability.

0 citationsRead paper

Balanced Twins: Causal Inference on Time Series with Hidden Confounding

Jun 17, 2026

This study addresses causal inference in time series settings with latent confounding and staggered interventions, where conventional methods are constrained by explicit temporal assumptions or convex combination requirements. The authors propose a deep representation learning framework that jointly learns low-dimensional individual latent representations and propensity scores, enabling estimation of individual counterfactual outcomes through a flexible nonparametric matching mechanism—without presupposing a specific temporal dynamics structure. By circumventing the convexity constraints inherent in synthetic control methods, the approach effectively mitigates bias arising from unobserved confounders. Empirical evaluations on both electricity demand response and ICU clinical datasets demonstrate that the proposed method significantly improves the accuracy of counterfactual prediction and average treatment effect estimation in nonstationary dynamic environments.

0 citationsRead paper

Zero-shot generalization of transformer neural operators to larger domains

Jun 12, 2026

This work addresses the limited zero-shot generalization of existing Transformer-based neural operators to larger spatial domains, a limitation stemming from their assumption of a fixed solution domain. The authors propose an architecture-agnostic solution that introduces a factorizable attention logit bias to enforce fine-grained, controllable spatial locality while remaining compatible with efficient attention mechanisms. Coupled with rotation-based positional encoding, this approach explicitly models translation equivariance, thereby enabling seamless extension to arbitrarily large spatial regions. Evaluated on two canonical PDE benchmarks and a realistic 3D industrial atmospheric flow task, the method demonstrates significantly improved zero-shot generalization performance when extrapolating to larger domains.

0 citationsRead paper

TS-ICL: A Flexible Time-Indexed Foundation Model for Time Series via In-Context Learning

Jun 04, 2026

This work addresses the limitation of existing time series foundation models, which are predominantly confined to forecasting and struggle to unify diverse real-world tasks such as handling irregular observations, imputing missing values, and managing degraded sampling. The authors propose TS-ICL, the first framework that integrates in-context learning with causal data priors through a probabilistic Transformer-based encoder-regressor architecture, casting all tasks into timestamp-aligned regression problems. Trained on synthetically generated causal dependency structures, TS-ICL achieves state-of-the-art performance in missing value imputation and maintains leading accuracy on both univariate and covariate-aware forecasting benchmarks, particularly excelling under partial observation conditions within the look-back window.

0 citationsRead paper