A hybrid forecasting and fuzzy comprehensive evaluation approach for graded early warning of experimental task frequency

Abstract Accurate prediction and graded warning of experimental task frequencies are critical for resource optimization and operational safety in large-scale scientific facilities. To address this, we propose a novel graded early-warning framework that seamlessly integrates deep sequential forecasting with fuzzy comprehensive evaluation. A hybrid Transformer-LSTM neural network is developed to capture both long-range dependencies and local temporal patterns in multivariate task frequency sequences. To address the sensitivity of deep models to hyperparameters, we employ the bioinspired Black-winged Kite Algorithm (BKA) to jointly optimize the learning rate, hidden layer units, and regularization coefficients. BKA is selected for its Cauchy mutation and adaptive leadership mechanisms, which effectively avoid local optima and balance exploration-exploitation. Preliminary tests confirm that BKA outperforms classical methods such as PSO and GA. Subsequently, a fuzzy comprehensive evaluation model, incorporating entropy-based weights and membership functions derived from prediction residuals, maps the forecasting uncertainties onto a four-level early warning scale (Safe, Mild, Moderate, Severe). Digital simulation verification demonstrates that the proposed method achieves high prediction accuracy, with a root mean square error (RMSE) of 2.18 counts $$/\textrm{h}$$ and a coefficient of determination ( $$R^2$$ ) of 0.774 on the test set, outperforming baseline LSTM and PSO-LSTM models. Furthermore, the fuzzy warning system attains a precision of 0.91 and a recall of 0.93 for the Severe level, with an overall area under the ROC curve (AUC) of 0.94, confirming its reliable graded warning capability. The integration of data-driven prediction and fuzzy logic provides a powerful framework for active anomaly management in complex dynamic systems.