Prediction of Cement Clinker f-CaO Based on Temporal Contrastive Learning and Adaptive Gated TCN-GRU

The calcination of cement clinker is the fundamental stage of the entire cement manufacturing process, where the concentration level of free calcium oxide f-CaO is universally recognized as the most critical parameter for assessing clinker quality and controlling energy efficiency. However, in current industrial practice, the measurement of f-CaO primarily depends on manual sampling and off-line laboratory analysis, which introduces a substantial feedback delay typically ranging from one to two hours. This feedback latency makes it impossible to satisfy the strict requirements for real-time closed-loop control, often resulting in quality fluctuations and increased fuel consumption. To overcome these limitations, this paper proposes a dual-layer temporal contrastive regression model, termed AGTC-TCN-GRU, which combines an adaptive gated fusion mechanism with Temporal Convolutional Networks TCN and Gated Recurrent Units GRU. To resolve the conflict between the vast amount of unannotated process data and the scarcity of quality labels, a matrix-based structuring technique is first introduced. This technique utilizes a matrix windowing strategy to transform raw process data into two-dimensional structured blocks, allowing the model to incorporate latent process information from unlabeled data to enrich the feature representation. The proposed architecture adopts a hierarchical modeling approach to distinguish between intra-block local features and inter-block global dependencies. Specifically, a TCN layer—incorporating causal convolutions, dilated convolutions, and residual blocks—is implemented to extract high-dimensional temporal features within each data block. The TCN consists of two layers with 48 and 24 channels respectively, and a kernel size of 3. A global average pooling operation is then applied along the temporal dimension to compress each data block into a fixed-dimensional local feature vector. Subsequently, a GRU layer is applied to the sequence of these feature vectors to model the long-term dynamic evolution and the substantial thermal inertia inherent in the continuous calcination process, thereby bridging the temporal gaps introduced by discrete data slicing. The GRU hidden size is set to 56. To enhance robustness under fluctuating operating conditions, an adaptive gated fusion module is designed. This module dynamically adjusts the fusion weights between local information and global trends based on the real-time input status, ensuring stable output during process disturbances. Furthermore, the framework integrates a temporal weighted contrastive learning mechanism that enforces a distance constraint in the feature space. By applying an exponential time-decay function to maintain feature consistency between temporally adjacent samples, the model suppresses high-frequency sensor noise and enhances representation stability. This soft weighting strategy avoids hard binary partitioning of samples and effectively preserves the continuity of operating condition transitions. Comprehensive experimental evaluations using a real-world industrial dataset from a cement production line confirm the performance of the AGTC-TCN-GRU model. The model achieves a determination coefficient R2 of 0.7970, a Root Mean Square Error RMSE of 0.1480, and a Mean Absolute Error MAE of 0.1153. These results demonstrate that the proposed method outperforms conventional models such as Support Vector Regression SVR and LightGBM, as well as several advanced deep learning architectures. This research provides a reliable technical foundation for real-time quality monitoring and energy-efficient control in cement clinker production.