Quasi-Newton optimised Kolmogorov-Arnold Networks for wind farm power prediction.

Abstract

Having accurate and effective wind energy forecasting that can be easily incorporated into smart networks is important. Appropriate planning and energy generation predictions are necessary for these infrastructures. The production of wind energy is linked to instability and unpredictability. Wind energy forecasting has traditionally been performed using statistical models, but with the advent of artificial intelligence (AI), research indicates that AI is more accurate than the statical technique. In this study, the nominal power of six wind farms in China was predicted using Kolmogorov-Arnold Networks (KAN) and Multilayer Perceptron (MLP) models. KAN as an alternative to the conventional MLP, has the ability to handle problems with scalability, vanishing gradients, and interpretability associated with MLP. The KAN uses learnable B-Spline as activation functions prompting it to address the issues of the MLP. We employed the Radial Basis Function (RBF) with Gaussian kernels to approximate the 3-order B-spline basis. In most deep learning models stochastic gradient-based optimization algorithms such as Adaptive Moment Estimation (ADAM) and Stochastic Gradient Descent (SGD) optimizer are mostly employed, a quasi-Newton optimization technique Limited-memory Broyden-Fletcher-Goldfarb-Shanno algorithm LBFGS was employed in this work to approximate the Hessian matrix and estimate the parameter space's curvature. Also, in the preprocessing of the data, the Interquartile Range (IQR) technique is used to handle outliers and a clustering-based K-Nearest Neighbor (KNN) imputer to handle missing values. Based on different sites, the KAN-LBFGS shows superior performance based on the performance evaluation metrics with site 5 achieving MSE of 0.0039, RMSE of 0.0622, MAE of 0.0352, and DC of 0.9468. The study highlights the importance of the model's architecture, preprocessing and optimization techniques.