Design and optimization of a lightweight machine learning algorithm for real-time anomaly detection and forecasting on resource-constrained edge sensor systems

Abstract

edge and IoT sensor deployments increasingly require on-site or near-site data quality control and forecasting, yet the computing resources available at remote monitoring stations are often limited. This paper addresses the design and optimization of a lightweight machine learning pipeline that combines a rolling z score anomaly filter with a gradient-boosted regression model for short-horizon forecasting, using continuous water-discharge monitoring in an arid region as a representative resource-constrained deployment case. The design is evaluated against persistence and ARIMA baselines using accuracy, RMSE, MAE, and inference time as joint design criteria. Results show that the proposed lightweight model attains 92.7% forecasting accuracy with an RMSE of 0.58 m3/s and an inference time of 0.46 s, outperforming both baselines in accuracy while remaining suitable for periodic retraining on modest edge-computing hardware, illustrating an effective algorithm-design strategy for resource-constrained monitoring systems.