Dynamic optimization of recurrent networks for wind speed prediction on edge devices

Accurate wind speed prediction (WSP) remains essential for optimizing energy management in small-scale domestic windmills. Server-dependent machine learning models, commonly deployed in wind farms, prove infeasible for domestic systems due to high costs and energy demands. While edge computing offers a viable alternative, current WSP methods prioritize hyperparameter optimization without constraining model size (MS), resulting in memory-intensive architectures incompatible with resource-limited devices. To address this gap, we propose a framework that co-optimizes the discrete hyperparameter spaces of Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU) and Temporal Convolutional Network (TCN) models under strict memory constraints. An adaptive Simulated Annealing algorithm with memory-based rejection (aSAR) navigates the discrete design space, employing nine objective functions that balance Mean Absolute Percentage Error (MAPE) against model compactness. Evaluations on wind datasets from Chile, Kazakhstan and Mongolia demonstrate that aSAR-optimized models reduce prediction errors by up to 54.17% and decrease MS by 98.75% relative to state-of-the-art techniques. The results highlight significant regional performance variations, underscoring the necessity of location-specific architecture selection. This work establishes a systematic approach for deploying memory-efficient WSP models on edge devices, advancing sustainable energy solutions for decentralized wind power systems.