Scalable Representation and Processing of Synchro-Waveform Data

Representing High-dimensional Power Data

January 2026 1

Vishwa Saragadam, Hamed Mohsenian-Rad

Energy Systems Research Workshop – Seed Grants

Vishwanath Saragadam     Assistant Professor of ECE at UCR.

    ◦ Expertise: Computational Imaging, Machine Learning, Computer Vision.

    ◦ Research Interests: Thermal Vision, Ultra-Broadband Imaging, Scalable High-Dimensional Representations

Hamed Mohsenian-Rad   Professor of ECE at UCR.

    ◦ Expertise: Power Systems, Sensing, Machine Learning, Optimization.

    ◦ Research Interests: Large-scale data-driven methods with applications in monitoring, control, and optimization of power systems.

Multi-scale Event Analysis

Spatio-Temporal

Correlation Analysis

Live Grid Health Monitoring

Fault Detection/Prediction

Time

(Input)

Systems Dynamics Tracking

Implicit Neural Representation of Waveform Measurements

Representing High-dimensional Power Data

January 2026 2

Vishwa Saragadam, Hamed Mohsenian-Rad

Energy Systems Research Workshop – Seed Grants

Background and Motivation

​

• Traditional filtered/processed representations of voltage and current miss some important details in waveforms.
• Modern sensors like Waveform Measurement Units (WMUs) enable high-resolution waveform capture.
• Direct waveform analytics is now possible—but introduces a big data challenge.

A single 3-phase WMU: ~ 4 billion samples/day (>1 GB/day)

​

Our Contribution

​

• Implicit Neural Representations (INRs): compact and continuous signal models for voltage/current waveform, using sinusoidal activations to capture the periodic nature of waveforms.

t → MLP with sinusoidal activations → v(t)

Single vs. Double Hidden Layer INRs

Representing High-dimensional Power Data

January 2026 3

Vishwa Saragadam, Hamed Mohsenian-Rad

Energy Systems Research Workshop – Seed Grants

Single-layer INR ≈ Fourier Transform

​

​

​

​

• Parameter count:

​

→ Approximates Fourier series 

(captures steady-state only)

​

​

Double-layer INR:

​

​

​

• Parameter count:

​

→ Captures transients and steady-state

Case Study – Sub-Cycle Oscillation

~3 × higher accuracy with same parameter count

Additional Real-world Case Studies

Representing High-dimensional Power Data

January 2026 4

Vishwa Saragadam, Hamed Mohsenian-Rad

Energy Systems Research Workshop – Seed Grants

INR model: double layer - (MSE: 0.77%–2.85%)

​

Dataset: three-phase SEL-735 sensor data (480V, 128 samples/cycle) diverse voltage and current signatures.

Sensitivity Analysis

Representing High-dimensional Power Data

January 2026 5

Vishwa Saragadam, Hamed Mohsenian-Rad

Energy Systems Research Workshop – Seed Grants

Sensitivity to Number of Parameters

​

Accuracy improves with more neurons (especially in second layer).

Trade-off: model size vs. accuracy

MSE for different number of parameter: a) voltage  waveforms; and b) current waveforms c) parameter count.

Sensitivity to Number of Layers

​

Two-layer INR is more stable, simpler, and almost always sufficient.

MSE for different number of layers: a) voltage  waveforms; and b) current waveforms

Modeling Correlated Waveforms

Representing High-dimensional Power Data

January 2026 6

Vishwa Saragadam, Hamed Mohsenian-Rad

Energy Systems Research Workshop – Seed Grants

Power system waveforms are correlated 

• Across 3 phases
• Across multiple location
• Between voltage and current

​

A single INR can represent correlated waveforms

​

→ Reduce model size and increase efficiency.

​

Compared 3 separate models vs. 1 combined INR for 3-phase waveform modeling:

​

Three Separate INRs:

​

​

Joint INR (3 outputs):

​

MSE for separate vs. combined INR models: a) voltage  waveforms and b) current waveforms.

Example of Application: Oscillation Analysis with INR Models

Representing High-dimensional Power Data

January 2026 7

Vishwa Saragadam, Hamed Mohsenian-Rad

Energy Systems Research Workshop – Seed Grants

• Apply Discrete Fourier Transform (DFT) to both raw waveforms and INR outputs

​

Single-mode oscillation: dominant frequency: (900 Hz)

INR matches DFT spectrum.

​

Dual-mode modulated: sidebands at (60 Hz ± f_sideband)

INR recovers sidebands precisely.

​

​

Modeling Compression:

​

Separate INRs: 8103 parameters

Combined INR (3 outputs): 5503 parameters

Raw waveform: 23,808 points

​

Three separate INR models:

One combined INR models:

→ INR achieves  ~4 × compression

Key Takeaways

Representing High-dimensional Power Data

January 2026 8

Vishwa Saragadam, Hamed Mohsenian-Rad

Energy Systems Research Workshop – Seed Grants

• INRs are an alternative to sampled waveforms: 

Modeling voltage and current as continuous neural functions preserves sub-cycle detail and enables compact, differentiable representations.

​

• Architecture matters:

1-layer INR ≈ Fourier representation → good for steady-state

2-layer INR adds nonlinear frequency mixing → captures transients and oscillations

Sinusoidal activations are essential

​

• Strong accuracy–efficiency trade-off:�Real-world case studies show <1–2% MSE with 3–4× reduction compared to raw waveform.

​

• Multi-signal modeling improves efficiency: 

Shared-parameter INR jointly for three-phase, voltage–current, or synchro-waveforms

Processing Terabytes of Power Grid Data Efficiently

Representing High-dimensional Power Data

January 2026 9

Vishwa Saragadam, Hamed Mohsenian-Rad

Energy Systems Research Workshop – Seed Grants

Preliminary Work: Demonstrated highly compact and non-linear representation of high-resolution synchro-waveform data across phases and across geographical locations

​

Target Agency: NSF: Cyber Physical Systems, California State Grants, Industry

Timeline: 2026 Summer

Estimated Request: ~$500,000

​

Research Focus: Compact and Streaming Representations, Event detection and forecasting

​

1. Multi-scale and Correlated Synchro-Waveform Representations
• Data correlated across times, and across geographical locations
• Leverage for compact representation and cross-location prediction

​

• Foundational Research in Non-Linear Representations
• Harmonic Analysis-inspired representations for periodic data with compact anomalies
• Bridging the gap between fast classical approaches and modern AI-driven approaches