Abstract

Abstract
Network human operators' decision-making during grid outages requires significant attention and the ability to perceive real-time feedback from multiple information sources to minimize the number of control actions required to restore service, while maintaining the system and people safety. Data-driven event and alarm management have the potential to reduce human operator cognitive burden. However, the high complexity of events, the data semantics, and the large variety of equipment and technologies are key barriers for the application of Artificial Intelligence (AI) to raw SCADA data. In this context, this paper proposes a methodology to convert a large volume of alarm events into data mining terminology, creating the conditions for the application of modern AI techniques to alarm data. Moreover, this work also proposes two novel data-driven applications based on SCADA data: (i) identification of anomalous behaviors regarding the performance of the protection relays of primary substations, during circuit breaker tripping alarms in High Voltage (HV) and Medium Voltage (MV) lines; (ii) unsupervised learning to cluster similar events in HV line panels, classify new event logs based on the obtained clusters and membership grade with a control parameter that helps to identify rare events. Important aspects associated with data handling and pre-processing are also covered. The results for real data from a Distribution System Operator (DSO) showed: (i) that the proposed method can detect unexpected relay pickup events, e.g., one substation with nearly 41% of the circuit breaker alarms had an 'atypical' event in their context (revealed an overlooked problem on the electrification of a protection relay); (ii) capability to automatically detect and group issues into specific clusters, e.g., SF6 low-pressure alarms and blocks with abnormal profiles caused by event time-delay problems.

Abstract
This paper aims at contributing to personalize anesthetic drug administration during surgery. This study devel-ops an online robust model to predict the maintenance dose of atracurium necessary for the resulting effect, i.e. neuromuscular blockade, to attain a target profile. The model is based on the patient's neuromuscular blockade (NMB) response to the initial bolus only, overcoming the need for information on the patient's weight, age, height and Lean Body Mass usually associated to pharmacokinetic and pharmacodynamic models. To achieve this, a statistical analysis of the response of the patient to the initial bolus is carried out and a set of variables is established as predictors of the maintenance dose. The prediction is accomplished using Classification and Regression Trees, CART, which is a supervised learning method. Simulated data from a stochastic model for the NMB induced by atracurium is used as training set. All the 5000 doses predicted by the model lead to NMB level between 5% and 10%, which supports the proposed predictive model since it is clinically required that the steady state NMB level lies between this two values. The methodology is applied both to simulated and to clinical data sets and is found appropriate for online dose prediction.

Abstract

Abstract
Online incremental models for recommendation are nowadays pervasive in both the industry and the academia. However, there is not yet a standard evaluation methodology for the algorithms that maintain such models. Moreover, online evaluation methodologies available in the literature generally fall short on the statistical validation of results, since this validation is not trivially applicable to stream-based algorithms. We propose a k-fold validation framework for the pairwise comparison of recommendation algorithms that learn from user feedback streams, using prequential evaluation. Our proposal enables continuous statistical testing on adaptive-size sliding windows over the outcome of the prequential process, allowing practitioners and researchers to make decisions in real time based on solid statistical evidence. We present a set of experiments to gain insights on the sensitivity and robustness of two statistical tests-McNemar's and Wilcoxon signed rank-in a streaming data environment. Our results show that besides allowing a real-time, fine-grained online assessment, the online versions of the statistical tests are at least as robust as the batch versions, and definitely more robust than a simple prequential single-fold approach.