Hyper-parameter Optimization of Multi-attention Recurrent Neural Network for Battery State-of-Charge Forecasting

In the past years, a rapid deployment of battery energy storage systems for diverse smart grid services has been seen in electric power systems. However, a cost-effective and multi-objective application of these services necessitates a utilization of forecasting methods for a development of efficient capacity allocation and risk management strategies over the uncertainty of battery state-of-charge. The aim of this paper is to assess the tuning efficiency of multi-attention recurrent neural network for multi-step forecasting of battery state-of-charge under provision of primary frequency control. In particular, this paper describes hyper-parameter optimization of the network with a tree-structured parzen estimator and compares such optimization performance with random and manual search on a simulated battery state-of-charge dataset. The experimental results demonstrate that the tree-structured parzen estimator enables 0.6% and 1.5% score improvement for the dataset compared with the random and manual search, respectively.