APWinston/AI-Smart-Battery-Orchestration-For-Mini-Grids

AI-powered battery dispatch system for Ghana SREP solar mini-grids using LSTM forecasting and PPO reinforcement learning .It is trained on 6 years of real ERA5 climate data from 3 different locations in Ghana AI Smart Battery Orchestration for Ghana SREP Mini-Grids This project develops an AI system to manage battery storage in solar mini-grids across Ghana's Scaling-Up Renewable Energy Programme (SREP). The system learns when to charge and discharge a battery, hour by hour, to maximise electricity supply to rural communities while protecting battery health. It was trained on six years of real ERA5 climate data across six Ghana locations and evaluated against a conventional rule-based controller. SYSTEM SPECIFICATIONS The trained model represents the average Ghana SREP mini-grid site. Battery capacity 650 kWh LiFePO4 Solar PV 132.5 kWp (176.7 m2 panel area) Community size Approximately 1,318 people Mean load 18.96 kW Max charge rate 130 kW (0.2C) SOC operating range 20% to 90% Solar/Load ratio 1.51 TRAINING PIPELINE The system is built in five sequential phases. Phase 1 phase1_build_dataset.py Loads ERA5 climate data for Tamale, Kumasi, and Axim. Merges with a Nigeria national load profile scaled to the SREP community size. Outputs master_dataset_scaled.csv. Phase 2 phase2_train_lstm.py Trains a two-layer LSTM with 128 hidden units to forecast the next 24 hours of solar irradiance and load demand from an 8-feature, 24-hour lookback window. Outputs best_lstm_scaled.pth. Phase 3 phase3_environment.py Defines a custom Gymnasium environment wrapping the LSTM forecaster. Models battery physics with charge and discharge efficiency of 0.95, a solar performance ratio of 0.75, and a three-mechanism LiFePO4 degradation model: rainflow cycle aging (Xu et al. 2016), Arrhenius calendar aging with asymmetric SOC stress (Wang et al. 2014), and a lithium plating penalty below 30% SOC. Phase 4 phase4_ppo_training.py Trains a PPO agent using Stable-Baselines3 over three curriculum phases totalling 12 million steps. The observation space is 52 inputs covering SOC, SOH, a 24-hour solar forecast, a 24-hour load forecast, …