In the cremation process, dynamic control of the combustion state has a significant impact on energy efficiency and pollutant emissions. To further reduce energy consumption while maintaining stable combustion and ensuring pollutant emissions meet regulatory standards, this paper proposes a combustion control algorithm that combines stage determination with rolling horizon optimization. First, based on sensor and infrared temperature measurement data, a real-time data preprocessing and stage determination module is established. By analyzing the variation rate of flue gas concentrations and the oxygen content, the combustion process is divided into three stages: ignition, self-combustion, and tail stage. Second, an oil valve duty cycle parameter is introduced during the self-combustion stage to achieve intermittent fuel supply, while continuous fuel supply is maintained during the ignition and tail stages. Then, using pre-trained XGBoost models for flue gas and energy consumption prediction, along with a pollutant classification model, a rolling horizon genetic algorithm optimizer is constructed. The optimizer minimizes the cumulative pollutant concentrations and energy consumption within a future time window to dynamically determine the optimal control parameter combination. The algorithm performs rolling optimization once per minute according to the updated stage determination results and outputs the current control strategy, thereby achieving near-global dynamic optimization control. Experimental results show that, under actual cremation conditions, the proposed method effectively reduces fuel consumption while maintaining emission compliance.
The rapid growth of live-streaming e-commerce has generated vast multi-source data, creating an urgent need for data-driven risk assessment from the retailer’s perspective. This study proposes a novel Bayesian network (BN) model enhanced by data mining techniques to evaluate sales risks in retailer-streamer collaborations. A systematic framework of 21 risk indicators is developed across product, team, platform, and external environment dimensions. To address data scarcity, quantitative survey data (137 valid responses, Cronbach’s α = 0.959) are integrated with expert knowledge using Dempster–Shafer evidence theory for BN parameterization. A case study demonstrates the model’s capabilities in risk prediction, diagnostic reasoning, and sensitivity analysis. Results show a total risk distribution of 24.5
Predicting network traffic in large-scale environments such as universities is essential for efficient resource allocation, early anomaly detection, and maintaining service quality. This study presents a hybrid forecasting model that combines seasonal autoregressive integrated moving average SARIMA with a bidirectional long short-term memory BiLSTM network. SARIMA captures the linear and seasonal structure of Wi-Fi access point AP traffic, while the BiLSTM learns the remaining nonlinear residual patterns. We evaluate the hybrid model against standalone SARIMA and BiLSTM baselines and show that it achieves superior forecasting accuracy. We also investigate transfer learning by adapting the pre-trained hybrid model to APs with limited historical data. Results demonstrate that the proposed framework delivers accurate, scalable, and data-efficient network traffic forecasting suitable for dynamic institutional settings.
Deep learning-based change detection is crucial for land use monitoring and disaster assessment. However, illumination differences, seasonal vegetation, and geometric deformations often introduce pseudo responses and boundary ambiguity in high-resolution imagery. To address these issues, we propose a Multi-scale Dynamic Interaction Network (MDINet) that integrates cross-temporal feature alignment, adaptive attention, and fine-grained boundary refinement in a unified framework. MDINet enhances temporal consistency and preserves structural details, enabling precise distinction between true and false changes. Experiments on LEVIR-CD and CDD-CD show state-of-the-art results, achieving F1-scores of 95.82
Factor mining in quantitative finance typically uses Information Coefficient (IC) to evaluate factors during optimization. However, IC evaluates entire stock rankings uniformly while practical trading focuses only on tail-ranked stocks, creating suboptimal factor selection. This paper proposes hybrid fitness functions incorporating tail return components alongside IC to align optimization with trading implementation. Using Genetic Programming on A-shares market data, we tested configurations combining IC with top-ranked returns (long-only) and separation percentages (long-short) at various weights. Results show that certain hybrid configurations substantially outperformed pure IC, with higher weights generally yielded better results. Improvements are further validated across expanded testing. The findings suggest that incorporating tail return information directly into fitness evaluation may improve factor discovery by aligning optimization with trading implementation.
