In the contemporary landscape of large language models (LLMs) development, it is crucial to address the challenges of deploying these models on hardware-constrained consumer electronic devices (CEDs), especially within complex dynamic task scheduling in multi-cloud environments (MCE). We propose a novel methodology leveraging a lightweight LLM to enhance task scheduling decisions in MCE.Our approach involves creating a task scheduling expert database informed by optimization objectives to fine-tune the lightweight LLM. This enables the model to generate a schedulable candidate set of tasks based on the current state of tasks and operational conditions within CEDs across MCE, optimizing scheduling decisions and enhancing overall efficiency. Simulations using both synthetic and real-world datasets demonstrate that our method outperforms three other algorithms in cost minimization, makespan reduction, and energy consumption. In summary, our methodology empowers CEDs to optimize the utilization of multi-cloud resources and harness the capabilities of lightweight LLMs to effectively minimize makespan, operational costs, and energy consumption during the task scheduling process, thereby facilitating efficient task scheduling.

Due to the fast, dynamic, and continuous arrival of data streams in the green Internet of Things (IoT) environment, the probability distribution of data streams changes over time. In real IoT scenarios such as unmanned aerial vehicle (UAV) detection and smart light switch control, data distribution changes have reduced the trained model’s accuracy for data streams problems classification, making it challenging to detect UAV intruders and predict whether energy-saving lamps in smart buildings are on or off. In this paper, an incremental ensemble classification method is proposed to improve prediction accuracy for green IoT. Specifically, a fuzzy rule-based classifier is combined with a dynamic weighting algorithm for improving classification accuracy. Moreover, the model is updated by incrementally learning the characteristics of data streams, which can effectively handle concept drift caused by data distribution changes in data streams. Experimental evaluations of UAV intrusion detection, smart buildings, and other datasets show that the proposed approach yields 2% higher area under the curve (AUC) and geometric mean (G-mean) than existing methods on UAV Detection and Occupancy datasets and 5% higher AUC and G-mean on five benchmarking datasets. For all datasets, the proposed approach yields 50% faster average training time than other methods.