Paste-filling is a crucial process in the mining industry. Traditional sensor devices often overlook model errors and struggle to measure the certainty of key quality variables. This study addresses these challenges by proposing a novel efficient data-driven forecasting framework to predict the concentration of deep cone thickeners (DDQRF). Conventional methods rely on normal regression models, minimizing residual mean square error to estimate underflow concentration, resulting in inaccuracies due to residual error accumulation in recursive strategies. Specifically, complex high-quality feature representation is essential for accurate prediction models, particularly for multiple horizon predictions necessary for hierarchical optimal control. Our framework introduces direct data-driven regression prediction, leveraging temporal machine learning models to extract features effectively. Unlike probabilistic Bayesian models, our approach offers efficient implementation and deployment, utilizing prediction intervals to quantify forecast uncertainty. In addition, the proposed model directly predicts multiple horizons, contrasting with traditional recursive single-point forecasting, offering enhanced training and memory efficiency. These characteristics are validated through industrial experiments on a deep cone thickener, comprehensively comparing performance with state-of-the-art counterparts.