In recent years, numerous machine learning methods have been developed to predict the remaining useful life (RUL) of wind turbine components. However, uncertainties in modelling the future progression of degradation often preclude accurate point forecasts of failure times. Quantifying this uncertainty is therefore crucial to ensuring reliable predictions as it empowers operators to make risk-informed maintenance decisions. This work proposes a probabilistic RUL forecasting framework that leverages a convolutional autoencoder (CAE) to extract health indicators (HIs) from supervisory control and data acquisition (SCADA) signals, accurately capturing component degradation over time. To facilitate HI extraction, a Convolutional Neural Network-based normal behaviour modelling framework is employed as a feature extractor, and residuals of component temperature signals, rather than the raw signals, are supplied to the CAE. These HIs are then fed into a Long Short-Term Memory-based conformalised quantile regression framework to probabilistically predict RUL, calibrating confidence intervals to reliably represent uncertainty. This proposed approach effectively models degradation while alleviating the impact of high noise in field data. Its application to two case studies demonstrates that, while achieving similar performance to existing approaches using the simulated Commercial Modular Aero-Propulsion System Simulation (CMAPSS) dataset, the proposed approach significantly outperforms when using a real SCADA dataset with gearbox failures, reducing point prediction errors by approximately 67 %. Furthermore, the generated prediction intervals are better calibrated and, on average, 42 % shorter, providing more informative and reliable uncertainty estimates.