Edge AI-Driven Energy-Efficient Smart Grid Multisensor Fusion

Thesis Abstract

The rapid electrification of modern power systems coupled with the proliferation of distributed energy resources amplifies real-time monitoring and control demands for resilience, efficiency, and sustainability in smart grids. However, conventional centralized data processing faces latency, bandwidth, and privacy constraints as multisensor data streams from transformers, feeders, microgrids, and consumer premises accumulate at scale. This study addresses the problem of achieving energy-efficient, low-latency grid operation through edge-enabled multisensor fusion powered by lightweight AI inference, preserving data locality while enhancing situational awareness for grid operators. The aim is to develop an edge AI framework that fuses heterogeneous grid sensor data to enable accurate state estimation, anomaly detection, and proactive control with reduced communication overhead. Specific objectives are (1) to design an on-device fusion architecture that integrates phasor measurement unit (PMU), smart meter, capacitor bank, and environmental sensor data for robust state estimation; (2) to implement a compact neural network and probabilistic fusion scheme suitable for resource-constrained edge devices; (3) to evaluate energy savings, latency, and accuracy against cloud-centric baselines under varying network conditions; (4) to assess resilience to data loss and cyber-physical threats using simulated attack scenarios; and (5) to derive guidelines for deployment in distribution networks and microgrids. The methodology adopts a mixed-methods, theory-informed approach. The research design combines a quantitative experimental evaluation with qualitative threat modeling to validate the edge fusion framework. The population comprises three distribution feeders within a metropolitan microgrid, with a total of 60 PMUs, 200 smart meters, 40 environmental sensors, and 15 edge gateways deployed across the network. A stratified sampling strategy selects data windows representing peak, off-peak, and fault-like conditions, yielding a dataset of 1,200 hours of synchronized multisensor recordings. Data collection employs purposively instrumented edge gateways capable of real-time inference and secure communication, complemented by a synthetic dataset generated via a power system simulator to model rare contingencies. The inference model comprises a lightweight convolutional neural network (CNN) fused with a Bayesian state estimator, deployed on NVIDIA Jetson-based edge devices and ARM microcontrollers, ensuring a trade-off between accuracy and energy consumption. Validity and reliability are ensured through cross-validation, repeatability tests, and calibration against a supervised ground-truth state-estimation set derived from phasor data and feeder measurements. Data analysis employs regression analysis to quantify accuracy gains, ANOVA to compare edge versus cloud performance across scenarios, and information-theoretic metrics to assess data compression efficiency. Anomaly detection performance is evaluated using ROC-AUC and F1 scores, while resilience is analyzed through scenario-based stress testing of data loss and spoofing events. The study operationalizes a conceptual model of edge fusion that integrates heterogeneous data flows, with a theoretical underpinning from distributed artificial intelligence and the theory of cyber-physical system resilience. Expected findings include that the edge fusion framework achieves comparable state estimation accuracy to centralized solutions while reducing data transmission by up to 65%, lowering energy consumption of edge devices by 28% through model optimization and hardware acceleration, and achieving sub-50 ms end-to-end latency for critical measurements under normal conditions. The system is anticipated to maintain robust performance under moderate data loss and to detect anomalous events with AUC scores above 0.92 and F1 scores above 0.85 in tested scenarios. The contribution to knowledge lies in a reproducible, energy-efficient edge AI architecture for multisensor fusion in smart grids, a novel integration of CNN-based feature extraction with Bayesian state estimation for resource-constrained environments, and empirical benchmarks that quantify trade-offs between edge intelligence, communication load, and grid reliability. The study also advances practical deployment guidelines, including sensor placement heuristics, edge hardware specifications, secure communication protocols, and governance frameworks for data privacy and fault-tolerant operation. The main conclusion posits that edge AI-driven multisensor fusion can deliver near-optimal grid situational awareness and control with substantial energy savings and reduced latency, even in heterogeneous network conditions. Recommendations include extending the framework to incorporating renewable generation forecasts, exploring federated learning for cross-domain collaboration, and conducting long-term field trials in diverse distribution networks to validate scalability and resilience.