Deep Reinforcement Learning for Autonomous Energy Management in Isolated Microgrids Under Extreme Weather Events in the Niger Delta

This research develops a deep reinforcement learning framework for autonomous energy management in isolated microgrids operating under the extreme weather conditions of the Niger Delta, South South Nigeria, making original theoretical contributions at the intersection of reinforcement learning theory and power systems control under environmental uncertainty. Niger Delta microgrids face compound challenges including high humidity degrading solar panel output, frequent tropical storms causing intermittent wind generation, and flash flooding isolating fuel supply logistics for diesel backup generation, collectively creating a highly stochastic and non-stationary operational environment that renders model-based optimisation controllers unreliable. The theoretical contribution consists of a novel DRL agent architecture incorporating a long short-term memory prediction module that encodes multi-day weather pattern memory into the state representation, enabling the agent to distinguish short-term weather anomalies from the onset of extended adverse conditions requiring conservative energy dispatch. A novel reward function formulation is developed that balances immediate cost minimisation against the risk of future energy insecurity during extended storm periods, addressing the credit assignment challenge that prevents conventional DRL agents from developing adequate precautionary storage behaviour. The DRL agent is trained in a high-fidelity Niger Delta microgrid simulation environment built on real weather data from three meteorological stations in Bayelsa and Delta States. Transfer learning experiments demonstrate that agents trained on one microgrid achieve near-optimal performance on novel microgrids after 15 percent of full training interactions, significantly reducing deployment cost. Keywords: deep reinforcement learning, microgrid energy management, extreme weather, Niger Delta, autonomous control.