Comparison Study of Wind Flow Velocity Short-Term Forecasting Methods Based on Adaptive Models and Neural Networks
The article is devoted to the study and comparison analysis of wind flow velocity short-term forecasting methods based on adaptive models and neural networks. The research used data obtained from the Russian island. The adaptive Holt, Brown and Holt-Winters models are compared, and their best parameters are selected for the problem at hand. The paper gives the results of research on the influence of the number of hidden layers, the activation function, and the method of training a multilayer neural network on the accuracy in forecasting. The results showed a significant advantage of a rectified linear unit (ReLU) as an activation function relative to the sigmoid function and hyperbolic tangent, as well as the advantage of the Adam method over classic Gradient Descent. Adaptive methods showed MAPE 16%, and Neural networks showed 10%. Adaptive methods ensure quick tuning a task solution with a low risk of overfitting, so the accuracy between the training data and real-life accuracy will be the same. Artificial neural networks allow to obtain more accurate forecasts but require careful tuning and have the risk of overfitting as a significant accuracy decreasing in case of changing climatic conditions.