Economic and Energy Security Optimisation of a Hybrid Microgrid Using Machine Learning

Amidst South Africa's persistent energy crisis, many large power users have chosen to install embedded generation, creating 'microgrids' where local generation supplements the national grid to meet electrical loads. The presence of local generation and storage offers an opportunity to optimize both the quantity and timing of energy procured from the national grid. The study follows a two-stage approach. Firstly., machine learning is used to predict the load requirements for a microgrid for the next 24 hours. Secondly, reinforcement learning is used to train an agent to purchase the optimal amount of energy at the most opportune time. This approach demonstrates a 5-6% reduction in energy costs, a drastic 98% decrease in diesel generation, and the elimination of unmet demand due to load shedding for such a microgrid.