https://github.com/Pushkkaarr/Anomaly-and-Fault-Detection-in-Nuclear-Power-Plants-with-Machine-Learning

WHAT IT DOES

AI-powered autonomous control system for nuclear reactor stability using reinforcement learning, continuously monitoring reactor states and making real-time control decisions to maintain 100% nominal power.

Learns simultaneous control of control rod position and coolant flow rate, handling complex transients and emergency scenarios such as Loss of Flow Accidents (LOFA) without human intervention.

TECH STACK

Core: Python, Gymnasium (OpenAI Gym), Stable-Baselines3, PyTorch

Scientific Computing: NumPy, SciPy (RK45 ODE solver), Matplotlib

ML Architecture: Soft Actor-Critic (SAC), MLP-based policy networks, Experience Replay

Monitoring: TensorBoard (optional), checkpoint-based model recovery

KEY FEATURES

Physics-Based Reactor Simulation – Implements point kinetics neutronics with delayed neutron precursors, coupled thermal-hydraulic modeling, Doppler feedback, and pressure dynamics calibrated for a 2000 MWt PHWR reactor.

Continuous Dual-Action Control – Uses a continuous action space to simultaneously regulate control rod motion and coolant flow rate with safety-constrained scaling to prevent sudden transients.

Reward-Shaped Stability Learning – Employs a staircase reward function incentivizing progressively tighter power regulation, with strict penalties for thermal violations and instability.

Safety-Critical Design – Enforces hard termination thresholds for catastrophic states, conservative action limits, flow clamping, and robust numerical stability handling.

Performance Monitoring & Visualization – Tracks power deviation, temperatures, control signals, and cumulative rewards in real time with post-episode diagnostic plots for analysis.

IMPLEMENTATION HIGHLIGHTS

• Reinforcement Learning Control: Trained a Soft Actor-Critic agent over 200,000+ timesteps to achieve stable reactor operation under varying initial conditions and transient scenarios.
• Conservative Training Strategy: Reduced entropy exploration, extended training duration, and frequent checkpointing to prioritize safety, stability, and interpretability over aggressive optimization.

TECHNICAL CHALLENGES SOLVED

• Multi-Variable Stability Control: Balanced competing objectives such as power tracking, thermal limits, and actuator constraints where traditional PID controllers struggle.