Multivariate time-series modelling of transformer health indicators using SCADA, DGA, alarm and maintenance records

Transformer health assessment requires the joint interpretation of electrical, thermal, chemical, mechanical, insulation and operational evidence. This paper presents a multivariate time-series modelling framework for transformer health indicators using supervisory control and data acquisition (SCADA), dissolved gas analysis (DGA), alarm, sensor and maintenance records. The model-ready dataset was constructed through timestamp alignment, engineering-unit validation, impossible-value filtering, missing-value handling, normalization, feature engineering and fixed-window segmentation. The baseline task predicts whether abnormal transformer-health evidence will occur within a six-step future horizon using a twenty-four-step observation window. The variables include load current, voltage, top-oil temperature, winding temperature, ambient temperature, dissolved gases, moisture, partial discharge, vibration, alarm history, cooling status, tap position and maintenance events. Exploratory analysis shows that transformer indicators are interdependent. Load is associated with thermal response, DGA indicators are linked with moisture and alarm history, and abnormal windows usually contain combined temporal evidence rather than isolated threshold exceedance. The cleaned sequence representation supports LSTM, iTransformer and Hybrid CNN-Attention prediction layers for early-warning maintenance decision support.