A prescriptive Dirichlet power allocation policy with deep reinforcement learning

Prescribing optimal operation based on the condition of system, and thereby potentially prolonging its remaining useful lifetime, has tremendous potential in terms actively managing availability, maintenance, costs complex systems. Reinforcement learning (RL) algorithms are particularly suitable for this type problem given their capabilities. A special case a prescriptive is power allocation task, which can be considered as sequential whereby action space bounded by simplex constraint. general continuous action-space solution such problems still remained an open research question RL algorithms. In space, standard Gaussian policy applied reinforcement does not support constraints, while Gaussian-softmax introduces bias during training. work, we propose Dirichlet tasks analyze variance gradients. We demonstrate that bias-free provides significantly faster convergence, better performance, robustness to hyperparameter changes compared policy. Moreover, applicability proposed algorithm evaluate performance study set multiple lithium-ion (Li-I) battery The experimental results prescribe operation, improving efficiency sustainability multi-power source • framework proposed. with method outperforms current SOTA methods shows robustness. End-to-end learning. good scalability transferability.