Accurate prediction of coal spontaneous combustion (CSC) temperatures is crucial for safe coal mine production. To further improve the accuracy of CSC temperature prediction and the interpretability of the model, this study proposes an interpretable Chebyshev chaotic mapping-Lévy flight-Enhanced pinhole imaging inverse learning-Adaptive weighted Osprey Optimization Algorithm optimized Bidirectional Long Short-Term Memory (CLEA-OOA-BiLSTM) framework for predicting CSC temperatures. First, we optimized the Osprey Optimization Algorithm (OOA) by incorporating the Chebyshev chaotic map, Lévy flights, enhanced pinhole imaging backpropagation, and an adaptive weighting strategy, thereby developing the CLEA-OOA algorithm. Through comparative experiments using eight benchmark test functions and four heuristic algorithms, we verified that CLEA-OOA achieves superior convergence accuracy and convergence speed. Subsequently, using data from the Dongtan Coal Mine as the subject of study, we employed CLEA-OOA to perform adaptive optimization of the BiLSTM hyperparameters. Using Spearman’s correlation analysis, C2H4/C2H6, CO, C2H4, CO/ΔO2, and O2 (%) were identified as key indicators for predicting the CSC temperatures. The results show that the CLEA-OOA-BiLSTM model achieves an R2 of 0.98, which is higher than that of all comparison models. The model’s MSE, RMSE, MAE, and MAPE are 11.26%, 3.36%, 2.73%, and 2.74%, respectively, demonstrating the model’s excellent error control capabilities. The results of the global and local SHapley Additive exPlanations (SHAP) interpretability analysis indicate that C2H4/C2H6 and CO are key contributing factors to the model’s decision-making, consistent with the oxidation mechanisms of CSC. The model was validated using coal mine data from multiple locations in Inner Mongolia, Shanxi, and Anhui, with R2 consistently reaching 0.985, demonstrating strong cross-regional generalization capabilities and practical engineering value. This framework provides a new method for CSC early warning and the intelligent development of mines.

Developing sustainability competence requires engineering graduates to weigh precision against cost, energy demand and regulatory limits, yet calculus instruction rewards exactness as the only correct answer, leaving students reluctant to settle for an adequate, resource-efficient solution. This study proposes the DRAP model (Dual-track Reasoning—Rigour-anchored Standard Operating Procedure (SOP)—Applied Practice—Post-practice Review), integrating the key competencies for sustainability framework, threshold concept theory and boundary management, and delivers it through a five-act chemical oxygen demand (COD) monitoring case. A quasi-experimental study was conducted with 114 first-year environmental engineering undergraduates (experimental n = 57; control n = 57; four intact classes delivered as two combined teaching units, one per condition). The primary outcome was the quality of the decision record produced in a worksheet format, covering boundary identification, decision compliance, cost–precision trade-off and risk anticipation and rated independently by three of the course’s teachers on anonymised submissions; class-level response entropy and individual-level integration choice described how reasoning was distributed. After controlling for pre-test cognitive state, the experimental group showed higher odds of selecting dual-track integration (OR = 3.57, 95% CI [1.38, 9.22], p = 0.017); SOP scores were higher (M = 79.83 vs. 66.17, p < 0.001, d = 0.97); and the difference between the two units persisted after controlling for baseline mathematical proficiency (partial η² = 0.24–0.30), most strongly on the cost–precision trade-off among the four dimensions. Embedding resource-efficiency and compliance constraints as decision premises within mathematical approximation offers a course-level route to the competences that SDG 4.7 and SDG 12 set out, without additional contact hours.