Optimizing Artificial Neural Networks for Smart Grid Stability Analysis using Tabu Search-Based Hyperparameter Tuning

Smart grids are supposed to make electricity networks more reliable, yet in practice their behavior can be surprisingly unstable when the load changes too quickly. During this work, we tried to understand that instability and to build a model that predicts when a grid remains stable or begins to drift. We used an Artificial Neural Network and tuned it with a Tabu Search routine that keeps adjusting the number of neurons and the activation functions until the results stop improving. It was not a single run; several training attempts were needed before the model settled. After applying the Tabu optimization, the network reached 98.2% accuracy and about 96% recall, which is higher than what we obtained from SVM or KNN in the same conditions. This enhancement was mainly due to the fact that the search process was made to search with unusual network topologies and not specific configurations. In addition, the process was also faster since it did not require re-training every model until it was ready. Although the context of the investigation is on the predictability of stability, the same methodological framework can also be applied to other systems in need of expediency and accuracy in forecasting complex nonlinear dynamics.