The problem of Power Quality analysis is becoming crucial to ensuring the proper functioning of complex systems and big plants. In this regard, it is essential to rapidly detect anomalies in voltage and current signals to ensure a prompt response and maximize the system’s availability with the minimum maintenance cost. In this paper, anomaly detection algorithms based on machine learning, such as One Class Support Vector Machine (OCSVM), Isolation Forest (IF), and Angle-Based Outlier Detection (ABOD), are used as a first tool for rapid and effective clustering of the measured voltage and current signals directly on-line on the sensing unit. If the proposed anomaly detection algorithm detects an anomaly, further investigations using suitable classification algorithms are required. The main advantage of the proposed solution is the ability to rapidly and efficiently detect different types of anomalies with low computational complexity, allowing the implementation of the algorithm directly on the sensor node used for signal acquisition. A suitable experimental platform has been established to evaluate the advantages of the proposed method. All the different models were tested using a consistent set of hyperparameters and an output dataset generated from the principal component analysis technique. The best results achieved included models reaching 100% recall and a 92% F1 score.

One of the most critical items from the reliability and the State-of-Health (SOH) point of view of wireless sensor networks is represented by lithium batteries. Predicting the SOH of batteries in sensor-equipped smart grids is crucial for optimizing energy management, preventing failures, and extending battery lifespan. Accurate SOH estimation enhances grid reliability, reduces maintenance costs, and facilitates the efficient integration of renewable energy sources. In this article, a solution for SOH prediction and the estimation of the Remaining Useful Life (RUL) of lithium batteries is presented. The approach was implemented and tested using two training datasets: the first consists of raw data provided by the Prognostics Center of Excellence at NASA, comprising 168 records, while the second is based on the curve fitting of the measured data using a single exponential degradation model. Long Short-Term Memory networks (LSTMs) were trained using data from three different scenarios, where battery cycle consumption reached 30%, 50%, and 65% correspondingly. Various architectures and hyperparameters were explored to optimize the models’ performance. The key finding is that training one of the models with only 50 records (equivalent to 30% of battery usage) enables accurate SOH prediction, achieving a Mean Squared Error (MSE) of $1.68\times {10}^{-4}$ and Root Mean Squared Error (RMSE) of $1.30\times {10}^{-2}$. The best model trained with 110 records achieved an MSE of $2.51\times {10}^{-5}$ and an RMSE of $5.01\times {10}^{-3}$.