Consequence-Aware Prescriptive Maintenance Framework With Transformer-KAN Forecasting and PPO-Controlled Grid Reconfiguration

The rapid increase in electrification has pushed electrical grids to operate near their limits, raising concerns about the emergence of alarms and the consequent decrease in reliability. Predicting these alarms allows for proactive corrective measures, enhancing grid stability. However, limited resources necessitate an effective prioritization process for each alarm. Prescriptive maintenance not only predicts impending alarms but also prescribes automatic corrective actions to mitigate them before they occur. This paper proposes a novel prescriptive maintenance framework for distribution grids employing a hybrid Transformer Kolmogorov-Arnold Network (Transfomre-KAN) and Proximal Policy Optimization (PPO). Initially, load and production values at each node are predicted using two Transfomer-KAN models. To ensure scalability and increased accuracy, the prediction algorithm incorporates features derived from wavelet transform, Tangent Pearson Embedding (TPE) and statistical analysis. Subsequently, load flow analysis is performed to determine node voltages, line currents, and consequent alarms. The Chernoff-Bound algorithm is then employed to assess the significance of each alarm, taking into account the potential consequences of failures. Corrective solutions involve adjusting the statuses of circuit breakers within the grid. To enable a fast and automatic prescription process, PPO algorithm, which is in the field of reinforcement learning, is utilized to train a neural network that inputs alarm significances and outputs optimal switching combinations. The practicality of the proposed method is demonstrated through a case study on a real Danish distribution grid. For the sake of benchmarking simulations are performed on a modified IEEE 13-Node test feeder. In addition Long Short-Term Memory (LSTM) model, Double Deep Q-Network (DDQN), Soft Actor Critic (SAC), and Asynchronous Advantage Actor Critic (A3C) are added. Comparisons show that the proposed method outperforms in terms of prediction and prescription.