Overload Alarm Forecasting in Power Grids Using Q-Learning and Transformer Architectures

The growing electricity demand, leading to an increased burden on power grid components in the future, has the potential to undermine the reliability of the electrical grid. To tackle this challenge, a common but costly approach is to allocate resources toward enhancing the capacity of the grid infrastructure. However, a more effective strategy involves the utilization of software-driven solutions that leverage the current infrastructure. This paper introduces a novel framework for predicting overload alarms resulting from severe loading situations (SLSs) in electrical grids operating near their operational limits. The proposed approach utilizes wavelet transform and different machine learning algorithms to forecast critical events caused by SLS. Wavelet transform is utilized to extract various signal frequency levels, which are then input into different models including Transformer and some traditional models. Subsequently, these outputs are fed into a Q-learning algorithm to facilitate optimal model selection at each time step. To obtain more practical results, the loading signal undergoes a filtering process aimed at removing short-duration loadings that exceed the threshold. The threshold value signifies the maximum network capacity within a specified time-frame, determined by an expert. The effectiveness of the suggested reinforcement learning approach is demonstrated through its ability to accurately forecast overload alarms by incorporating both industrial and residential loading.