INTELLIGENT DEEP LEARNING-BASED SDN FRAMEWORK FOR REAL-TIME FAULT MANAGEMENT IN SMART WIND ENERGY SYSTEMS

Abstract

The modern smart grid system using renewable energy (wind energy) as the basic energy-producing source requires an Artificial Intelligence (AI) based fault management system for handling the structural and operational limitations of the wind energy system to have high efficiency and standability. This study proposes real-time fault detection and mitigation in wind turbine systems using an SDN-POX controller integrated with the Temporal Convolutional Method (TCM). The POX controller deployed in Mininet extracts the real-time sensor data streams (e.g., vibration, temperature, rotor speed, pitch angle, and torque) of a wind turbine via Message Queuing Telemetry Transport (MQTT). Incoming data from each sensor channel is first structured into normalized multi-channel auto-updating sliding windows to facilitate the automatic extraction of reference operating points and dynamic tolerances via real-time streaming mean and standard deviation estimations. In order to fuse sensor modalities, the TCM features residual stacks (to record both short-term and long-term dependencies without recurrence), parallel sequence modeling, and channel mixing via 1-D causal-dilated convolutions. A SoftMax-based multi-class head is used to classify fault states, and an anomaly score measures deviations from nominal circumstances. The controller compares real-time feature vectors against learned reference-tolerance pairs to trigger Finite State Machine (FSM)-based mitigation in MATLAB/Simulink, including torque limitation, pitch correction, or braking. However, the fault management (involving detection and mitigation) take place within one control loop, with a fault detection delay (TFD) under 300 ms and a fault mitigation delay (TFSM)of less than 50 ms. By conducting experiments, we discovered that our approach detects faults more accurately and faster in mitigating and normalizing faults as compared to traditional methods (Graph Neural Network-GNN, K-Nearest Neighbors-KNN, and Supervisory Control and Data Acquisition-SCADA models).