Adaptive Smart Grid Systems Using Machine Learning Techniques
DOI:
https://doi.org/10.5281/zenodo.19152724Keywords:
smart grid, machine learning, load forecasting, demand response, fault detection, reinforcement learning, graph neural networks, temporal fusion transformerAbstract
The accelerating integration of distributed renewable energy resources, electric vehicle charging infrastructure, and demand-responsive prosumers
into modern power networks is rendering traditional rule-based grid management approaches operationally inadequate. Smart grid systems require
adaptive control mechanisms capable of real-time load forecasting, dynamic demand response coordination, fault detection, and voltage stability
management under conditions of high renewable penetration and stochastic demand variability. This study presents AdaptGrid, a unified machine
learning framework for adaptive smart grid management integrating a temporal fusion transformer (TFT) for multi-horizon load forecasting, a
multi-agent deep reinforcement learning (MADRL) controller for real-time demand response, and a graph convolutional network (GCN)-based fault
detection system. AdaptGrid was validated on the IEEE 118-bus test system and deployed on the IREN SpA medium-voltage distribution network in
Reggio Emilia, Italy (847 nodes, 23,400 prosumers). AdaptGrid achieved load forecasting MAPE of 1.84% (day-ahead) and 2.31% (week-ahead),
surpassing persistence (6.74%), ARIMA (4.12%), and LSTM (3.07%) baselines. Demand response coordination reduced peak load by 18.7% and
grid operational costs by 14.3%. Fault detection achieved F1-score of 0.974 with mean detection latency of 83 ms. These results establish AdaptGrid
as an accurate, responsive, and practically deployable ML framework for adaptive smart grid management.
