Trend-Aware Dynamic Fusion of Regression and Classification for Supernatant Interface Prediction in Industrial Thickening Systems

Optimizing industrial thickening systems requires accurate supernatant interface height prediction. Existing methods struggle with complex dynamics and noise, limiting accuracy. This article proposes a novel hybrid deep learning framework that synergistically combines regression and classification methodologies to improve predictive performance. Specifically, a differential attention-enhanced long short-term memory network provides high-precision point prediction by capturing salient features and addressing uncertain time delays. Concurrently, a multiscale convolutional neural network with a convolutional block attention module extracts features and classifies trend direction. To integrate numerical and trend predictions, a trend-aware dynamic fusion mechanism is introduced, which explicitly ensures trend consistency and adaptively optimizes fusion weights using a deep deterministic policy gradient algorithm, yielding predictions accurate in value and reliable in direction. Real-world data from a large coal preparation plant demonstrates the superior performance of our prediction method compared to baseline methods. In addition, industrial experiments validate that the proposed model supports operator decision-making for the optimization and control of thickening systems.