A reliable hybrid autoregressive integrated moving average and deep reinforcement machine learning strategy for resiliency enhancement in microgrid

Abstract

Machine learning algorithms are crucial for energy transition and climate change system resilience. This paper presents the essential issue of improved microgrid resilience after catastrophes, with a focus on the Root Mean Square Error (RMSE) assessment metric. By using a Hybrid Autoregressive Integrated Moving Average- Deep Reinforcement Learning (HARIMA-DRL) based machine learning microgrid restoration method for post-disaster power recovery, resilience in the operation aspect is enhanced. The local power grid's resilience during a simulated disaster is examined using a quantitative resilience measure that is based on two significant quality functions. When the hybrid HARIMA-DRL model final resilience value is 3.07 %, 7.78 %, and 11.51 % higher for 5000, 8000, and 10000 training datasets, as compared to Convolutional Neural Network and Long-Short Term Memory (CNN-LSTM) and Gauss Reinforcement Memory (GRM). Simulation findings comparing the proposed HARIMA-DRL method to CNN-LSTM and GRM models trained on separate datasets show that the HARIMA-DRL enhances grid instantaneous and average resilience by 3.38 % over T1 and 4.96 % over T2. Flexible kernel settings and smooth nonlinear prediction behaviour allow the proposed HARIMA-DRL approach to provide the highest resilience enhancement and lowest prediction error.