Winter wheat (Triticum aestivumL.) monoculture is conventionally cultivated followed by two to three months of summer fallow in the Loess Plateau. To develop a sustainable cropping system, we conducted a six-year field experiment to investigate the effect of leguminous green manure (LGM) instead of bare fallow on the yield and water use efficiency (WUE) of winter wheat and the soil water balance (SWB) in different precipitation years in a semi-arid region of northwest China. Results confirmed that planting LGM crop consumes soil water in the fallow season can bring varied effects to the subsequent wheat. The effect is positive or neutral when the annual precipitation is adequate, so that there is no significant reduction in the soil water supplied to wheat. If this is not the case, the effect is negative. On average, the LGM crop increased wheat yield and WUE by 13% and 28%, respectively, and had considerable potential for maintaining the SWB (0–200 cm) compared with fallow management. In conclusion, cultivation of the LGM crop is a better option than fallow to improve the productivity and WUE of the next crop and maintain the soil water balance in the normal and wet years in the Loess Plateau.

The carbon trading market is directed by policy and responsive to a multitude of factors, experiencing considerable price volatility that mirrors the supply–demand dynamics of greenhouse gas emissions rights and the economic expenses of carbon reduction measures. Improving prediction accuracy of carbon prices is crucial for investors and policymakers. It can help investors avoid risks and provide a basis for policymakers to formulate effective policies. The study employed a lightweight hybrid model to forecast carbon prices. Firstly, Marine Predators Algorithm Optimized Variational Mode Decomposition (MPA-VMD) was applied to decompose the original carbon price time series to obtain optimal parameters and subsequences. Secondly, the Minimal Gated Memory Network (MGM), a simplified network structure, was proposed to forecasting subsequences, reducing training time without compromising prediction accuracy. Thirdly, the MGM was integrated with quantile regression (QR) to predict the conditional quantiles of each subsequence. Fourthly, the probability density function was estimated based on the conditional quantiles of the carbon price using the Kernel Density Estimation (KDE) method. Finally, the final forecasting values of point prediction, interval prediction and comprehensive probability were obtained through the linear superposition of each subsequence, respectively. Experimental results showed that the performance of the proposed model was validated across five aspects: the superiority of the decomposition method, point prediction accuracy, suitable prediction interval, comprehensive probability prediction performance, and training time relative to six benchmark models.