Microgrid Operational Planning Using a Hybrid Neural Network with Resource-aware Scenario Selection

Modern power systems continue to increase in complexity due to the intermittent behavior of distributed energy resources (DERs) and stochastic climate factors. Considering these factors simultaneously generates enumerable scenarios for decision-making within a microgrid (MG) dispatch policy. An MG dispatch policy can be formulated as a mixed integer linear program (MILP) and solved using traditional solution methods such as simplex, graphical method, etc. However, given the number of potential scenarios, these traditional solution methods incur a high computational burden to obtain near real-time solutions. Thus, dimension reduction techniques are needed to reduce the MILP's complexity. To address the need for a scalable solution method and reduced model complexity, we propose a 2-stage reconfigurable framework for near real-time MG operational planning. The proposed 2-stage reconfigurable framework is comprised of four modules: i) a resource-aware scenario selection (RSS) algorithm to identify the most probable scenarios, ii) an MG dispatch policy to obtain solutions or dispatch decisions for an MG system, iii) a neural network (NN) to map the most probable scenarios from RSS to their corresponding solutions from the MG dispatch policy and iv) a rule-based policy to monitor depreciating predictions from the NN. The RSS module is a probabilistic variate of a full factorial design (FFD) that assigns weights to test points defining the design space based on stochastic climate factors and the number of quartiles used to partition each factor's likelihood function. The most probable scenarios from RSS and their corresponding solutions from the proposed MG dispatch policy were both synthesized into a training set. NNs were then trained using this dataset to predict dispatch decisions for the IEEE 18-bus and the IEEE 33-bus. The proposed RSS reduced the state space by 77%, where the remaining states reflect 98% of the original state space. Furthermore, the NNs applied for MG operational planning showed robust predictive performance with loss function values of 0.0228 and 0.0563 for the IEEE 18-bus and the IEEE 33-bus, respectively.