Hardware-Adaptive Acceleration of Multi-Fidelity Surrogate Optimization for Real-Time Wind Farm Micro-Siting

Abstract

The optimization problem in wind farms micro-siting is computationally-demanding because of repetition of intricate wake interaction modelling in different wind and topography settings, which can be a major constraint to real-time and embedded implementation. Existing optimization models make use of high-fidelity modelling models which are run on general-purpose processors thus making them prohibitive in terms of latency and energy expenditure. In order to overcome these weaknesses, this paper will suggest a hardware-adaptive Multi-Fidelity Bayesian Optimization (MF-BO) framework to apply to real time wind farm micro-siting. The proposed method involves multi-fidelity surrogate modelling, i.e. low-, medium-, and high-fidelity wake models are incorporated into a co-kriging Gaussian process framework and a reconfigurable hardware acceleration platform is introduced which dynamically allocates computational workloads to heterogeneous processing resources. A strategy of selecting and acquisition fidelity that is hardware-aware is presented to trade-off accuracy, computational cost and latency constraints. The inference and acquisition evaluation kernels provide by the MF-BO are optimized using reconfigurable logic, allowing them to execute with efficient parallelism and mixed-precision to be applied appropriately to embedded devices. Large scale experimental analyses of realistic wind farm situations show the proposed framework can achieve significant reduction in latency and enhanced energy efficiency over CPU-only and single-fidelity benchmarks, with no loss or gain in optimization quality. The findings indicate that the annual energy production (AEP) gain varies coherently with accelerated convergence, which confirms the possibility of implementing the MF-BO-based micro-siting optimization in practise, where resources are limited and real-time optimization is required. This paper provides an indication of how software-hardware co-design can make future embedded optimization systems possible in renewable energy usages.