Adaptive Transportable Energy Storage Expansion Planning via AI-Assisted Distributionally Robust Optimization under Renewable Energy and Flood Scenarios

Recently, higher penetration of Renewable Energy Generation (REG) into the grid, rapid increase in electricity demand, and natural disasters (such as floods in Malaysia), leads to line congestions, blackouts, and greater demands on resiliency of power systems. Lately, a new concept called Transportable Energy Storage Systems (TESS) was proposed, where a battery is installed on trucks or railways (“battery on wheels”) to enhance its mobility. The applications of TESS include peak load shaving, energy sharing among prosumers and alleviating power transmission congestion, under normal operating conditions. Furthermore, TESS can enhance grid resiliency under extreme conditions, by supplying electricity to blacked-out areas during and after natural disasters. However, research on the long-term (20 years) expansion for TESS is still infancy due to its fundamental complexity and technical challenges Therefore, this project aims to investigate the potential of large-scale TESS integration in future grids, using a novel adaptive stochastic TESS expansion planning model. The proposed model integrates the stochastic unit commitment with the vehicle routing problem, formulated as temporal-spatial optimization problem, to determine “when” and “where” the TESS should charge and discharge. Future load demand, solar REG, traffic flow, along with the flood model in Kuala Selangor, will be utilized in deciding the allocation and sizing of TESSs to be deployed. A convolution neural network (CNN) assisted optimization model, coupled with Adaptive Wasserstein Distributionally Robust Optimization and Wasserstein Distributionally Robust Chance Constraint (WDRO-WDRCC) approach, will be developed to solve the above-mentioned problem. Flexibility and resiliency indices will be formulated to validate the efficiency of the proposed approach. Moreover, CNN-assisted WDRO-WDRCC is expected to have good out-of-sample performance, producing cost-effective solutions with user-definable confidence levels (e.g. to cope with 95% uncertainty), capable of avoiding over-conservativeness solution, whilst remaining computationally tractable and efficient (up to 17 times faster than conventional model).