I develop statistical and machine learning methods to build reliable models from noisy, incomplete, and irregularly sampled data. My research spans robust estimation, uncertainty quantification, data fusion, and dynamical modeling, with a focus on inference and decision-making under uncertainty.
Distribution grids combine sensors with different measurement types, reporting rates, and missing data. Het-ODE uses sensor-specific encoders and attention to learn a shared continuous-time graph representation, then refines fused reconstructions with a graph-transformer residual estimator. Tested on three IEEE feeders, the approach estimates states at sparsely observed buses while retaining the distinction between measured inputs and inferred states.
How can we estimate a changing system when its sensors measure different quantities, report at different rates, and leave gaps? In this work, each sensor type has its own encoder to handle its available measurement channels. An observation mask records which values are present. The encoders produce shared latent features, a common representation that also retains the identity of each sensor type.
Sensor-typed attention
The next challenge is deciding which observations should inform one another. Sensor-typed attention learns these relationships using the measurement features, sensor identities, and time between observations. A query represents the information being sought; keys and values represent the information other observations can contribute. Combining this evidence across observation times gives each bus a probability distribution for its initial latent state.
Continuous graph ODE
Between observations, the system continues to change. We represent that evolution with a graph ordinary differential equation, which describes how a compact latent state changes over time. Shared interaction functions exchange information between neighboring buses, while sensor-specific readouts determine their latent derivatives. A typed decoder converts the resulting trajectory into a fused reconstruction of electrical quantities for the state estimator.
Residual state correction
The fused reconstruction provides an initial estimate, which the state estimator then refines. A graph transformer learns a correction for each bus using that estimate, the observation mask, and connections between buses. The final state estimate is the fused estimate plus this learned correction. This residual formulation lets the estimator refine information already recovered by data fusion. In the three test feeders, the refinement improves accuracy over fusion alone.
Citation / BibTeX
@inproceedings{algikar2026sensoraware,
title={Sensor-Aware Joint Data Fusion and State Estimation under Sparse Multi-rate Distribution Grid Sensing},
author={Algikar, Pooja and Fenu, Stefano and Jhala, Kumarsinh and Plathottam, Siby},
booktitle={NeurIPS Workshop: AI Foundations for Power Grids: From Models to Deployment at Scale},
year={2026},
eprint={2610.15156},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2610.15156}
}
Elapsed0:00Remaining1:47Total1:47
01Probe evidence and attacker belief02Planning guides the learned policy03Physical probing and measured feedback
Virtual power plants often lack the feeder model needed by network-based cyber defenses. This method tests small, secret DER perturbations through measured voltage responses, updates a Bayesian belief over attacker profiles, and uses Stackelberg guidance within reinforcement learning. The learned policy balances informative probing with perturbation cost, adapting defense using DER-level observations rather than explicit feeder topology.
A virtual power plant needs to detect interference even when it lacks a detailed feeder model. We test the integrity of its observations by applying small, secret power changes through selected distributed energy resources. These changes should create a recognizable voltage response at observing resources. A statistical detector compares the measured response with the expected signature, accounting for noise, and reports a decision with a measure of confidence.
Bayesian attacker belief
Detection evidence also helps us learn what kind of attacker may be present. A Gaussian process relates candidate attacker profiles to detection outcomes. Combining this likelihood with the previous belief produces an updated probability distribution over profiles. Hamiltonian Monte Carlo draws samples from that distribution. Their mean and covariance summarize the estimated attacker behavior and the uncertainty that the defender should consider when planning its next action.
Stackelberg-guided RL
The defender must choose a useful probe while anticipating how an attacker might respond. We model this as a Bayesian leader-follower game: the defender proposes an action, accounting for possible attacker responses under its current belief. The proposed action guides reinforcement learning through the policy input and a reward term penalizing deviation. Voltage and risk feedback help the policy adapt the final probing action.
DER probing feedback
The chosen action becomes a small adjustment to active and reactive power, subject to the resources' operating limits. Observing resources measure the voltage response, providing fresh evidence for detection and the attacker-belief update. Repeating this cycle links learning to physical feedback. The objective is useful detection information with limited probing effort and voltage disturbance. The strategy uses local voltage measurements and does not require explicit network connectivity.
Citation / BibTeX
@article{algikar2026gametheoretic,
title={Game-Theoretic Reinforcement Learning based Topology-Agnostic Defense for Virtual Power Plants},
author={Algikar, Pooja and Fenu, Stefano and Jhala, Kumarsinh and Plathottam, Siby},
journal={IEEE Transactions on Smart Grid},
year={2026},
note={Early Access},
doi={10.1109/TSG.2026.3741535},
eprint={2610.15183},
archivePrefix={arXiv},
url={https://arxiv.org/abs/2610.15183}
}
Elapsed0:00Remaining1:57Total1:57
01Sensing constraints and method families02The approved research-dependency map03Uncertainty-aware decision support
1 / 3
01Sensing constraints
02Method families
03Coupled research needs
04Decision support
IEEE Access
Real-Time Grid Visibility in Active Distribution Networks and Microgrids: A Review of Data Fusion, State Estimation, and Topology Identification
Grid visibility depends on more than an accurate state estimator: sensing, communication, data fusion, and topology inference interact. This review organizes classical, probabilistic, learning-based, and physics-informed methods around their operational requirements. It examines how uncertainty and residual evidence should pass between these functions, and identifies gaps in field validation, adaptation, and decision support for active networks and microgrids.
Reliable grid operation requires knowing the system's electrical state and how its equipment is connected. Yet measurements are often sparse, noisy, delayed, and collected at different rates. This review examines how those sensing and communication constraints affect grid visibility. It brings data fusion, state estimation, and topology identification into one discussion, so methods can be assessed against the information actually available to an operator.
Method families
Different methods address different parts of this challenge. Classical estimators use electrical models to relate measurements to system states. Robust methods reduce the influence of unreliable observations, while probabilistic methods represent uncertainty explicitly. Learning and graph methods capture relationships in data, and physics-informed methods incorporate electrical constraints. The review compares these choices through their assumptions, data requirements, computational demands, and suitability for real-time operation.
Coupled research needs
Grid visibility depends on several interacting tasks. Data fusion combines heterogeneous observations for state and topology inference. The inferred network configuration determines how a state estimator interprets those observations. Disagreement between predicted and measured values can motivate checking data quality or connectivity. A central research need is to share uncertainty and residual evidence across these tasks, so each inference can benefit from the others.
Decision support
For operational decisions, an estimate must come with evidence of when it can be trusted. The review identifies field validation, adaptation to changing topology, uncertainty-aware fusion, and scalable computation as important research priorities. It also highlights the need for fallback when sensing or inference becomes unreliable. Its contribution is a coordinated view of what these methods must deliver before they can support dependable real-time decisions.
Citation / BibTeX
@article{algikar-grid-visibility-review,
title = {Real-Time Grid Visibility in Active Distribution Networks and Microgrids: A Review of Data Fusion, State Estimation, and Topology Identification},
author = {Pooja Algikar and Kumarsinh Jhala and Stefano Fenu and Siby Plathottam},
journal = {IEEE Access},
note = {Accepted; publication details pending}
}
