MSTFormer: A novel long-term time series prediction model for electricity consumption in the cement calcination process

The cement calcination process is a critical and energy-intensive step in the cement production process, with electricity consumption accounting for the majority of overall energy consumption. Due to the combined influence of multiple factors such as process parameters, equipment status, and environmental disturbances, electricity consumption data exhibits complex characteristics. These include high time-delay, pronounced nonlinearity, and multivariate coupling, which pose significant challenges for long-term trend modeling and prediction. This paper addresses the challenge of long sequence time-series forecasting (LSTF) for electricity consumption in the cement calcination process by proposing a predictive model, MSTFormer. It integrates multi-channel time-series inputs with a spatio-temporal self-attention mechanism. The model first preprocesses the raw electricity consumption signals using variational modal decomposition. It then constructs a parallel multi-channel time-series input structure based on a weighted mechanism, effectively enhancing long-term dependency modeling capabilities. Subsequently, a novel spatiotemporal ProbSparse self-attention mechanism is designed, combining spatio-temporal embedding and causal convolution to adaptively extract global trends and local dynamic features, thereby enhancing the model’s ability to capture complex time-series patterns. Experimental results on actual cement production datasets demonstrate that MSTFormer outperforms mainstream long-sequence forecasting models across various prediction lengths. Compared with other decomposition methods, VMD-based decomposition reduces prediction errors by over 30%. For 15-minute prediction length, the model achieves an MAE of 0.063 and R² of 0.991. Even at 60-minute length, it improves prediction accuracy by over 24%. This study provides an effective approach for intelligent scheduling and energy optimization, offering theoretical and practical value for enhancing industrial energy efficiency.