Presentations

Abstract

Standard SCADA systems, which typically record 10-minute statistical aggregates (means, maximums, minimums, and standard deviations), act as low-pass filters that smooth out high-frequency dynamics. Consequently, conventional diagnostic models often struggle to interpret the highly non-linear and chaotic dependencies present during state changes, such as start-ups or emergency stops. These periods represent moments of maximum mechanical and electrical stress, yet they are frequently filtered out to avoid false alarms, leaving critical early-warning signals unexploited. This work addresses this gap by proposing a generalized methodology centered on analyzing the temporal evolution of post-event data sequences by focusing on specific “slow” signals like temperatures. While the methodology is applicable to any operational transition, this study validates the approach using turbine start-ups as the primary anchor point. The objective is to reconstruct the "health signature" of the turbine during its most vulnerable moments by exploiting the richness of multivariate statistical aggregates rather than relying solely on average values. The research presents a comprehensive comparative study between two paradigms: more conventional Machine Learning baselines (such as DTW - Direct Time Warping) and a state-of-the-art Deep Learning approach based on the Transformer architecture. Under the premise that sequences of SCADA statistics can be interpreted as semantic structures—where a sequence of sensor readings forms a "sentence" describing the turbine's state—the study adapts techniques from Natural Language Processing (NLP). Unlike Recurrent Neural Networks (RNNs) or LSTMs, which process data sequentially and often forget long-term dependencies, the Transformer analyzes the entire transient sequence simultaneously. This allows the model to identify subtle, long-range correlations between thermal, mechanical, and electrical variables that traditional methods overlook. The results demonstrate that the Transformer-based approach may offer a superior capability for modeling the complex, non-stationary behavior of a healthy start-up. By accurately predicting the expected evolution of operating variables, the model uses reconstruction error to detect anomalous deviations before the system stabilizes. This capability is particularly effective in distinguishing between normal operational variability and genuine incipient faults. This study concludes that transferring the "semantic" analysis capabilities of Transformers to the domain of SCADA data represents a significant advancement for renewable energy reliability. By decoding the complex information embedded in transient phases, operators can unlock a new layer of diagnostic depth, shifting from reactive filtering to proactive, event-driven monitoring.