Fault early-warning and closed-loop thermal control for power transformers based on multisource sensor data fusion

Abstract

Power transformers are exposed to coupled electrical, thermal, chemical, and mechanical stresses; therefore, single-channel alarms cannot provide reliable early warning or energy-efficient thermal regulation. This paper proposes a multisource sensor fusion framework that jointly performs incipient-fault assessment and closed-loop top-oil temperature control. Dissolved gas analysis, top-oil and winding temperatures, ultra-high-frequency/acoustic partial-discharge pulses, vibration, load, ambient temperature, and cooling-power measurements are time-aligned, quality weighted, encoded by lightweight one-dimensional convolutional and bidirectional gated recurrent branches, and fused through gated cross-modal attention. The fused state drives both a calibrated risk head and a supervisory fan/pump controller. A reproducible semi-physical dataset containing 12,480 one-hour fusion windows was constructed, while the thermal-control loop was evaluated at a one-minute interval. On the chronological test period, the proposed controller achieved a temperature mean absolute error of 0.93 °C, a root-mean-square error of 1.21 °C, cooling energy of 37.6 kWh/day, an 8.7 min response time, and a maximum overshoot of 1.6 °C. Relative to threshold control, these values correspond to reductions of 67.3%, 22.8%, 72.4%, and 72.4% in tracking error, energy use, response time, and overshoot, respectively. The identified closed-loop pole magnitude was 0.622, and all 300 disturbance simulations remained bounded. The results demonstrate that reliability aware fusion can convert heterogeneous monitoring data into stable, timely, and energy-efficient transformer control actions.