Robust Short-Term Wind Speed Forecasting Using Adaptive Shallow Neural Networks

Wind speed forecasting is necessary to integrate wind farms into power systems. In the past ten years, the forecasting models have become increasingly complex due to the development of artificial intelligence methods and computing power. Simultaneously, the robustness of models has decreased since complex models have a high risk of overfitting and decline in the accuracy if working conditions change significantly. This work aims to develop a machine learning model for short-term wind speed forecasting with acceptable accuracy but high robustness and the possibility of automatic online retraining. A shallow multilayer perceptron, trained only on retrospective data on wind speed, is proposed. The most significant results are combining simple neural network architecture with ReLU activation function, Adam training method developed for deep neural networks; and the automatic hyper-parameters selection using Grid search with open upper bounds. The model was trained on the data of the autumn period and tested on the winter data. A comparison was made with the simplest and most robust adaptive forecasting methods: Brown and Holt models. The significance of the obtained results is that shallow neural networks using ReLU, Adam, and Grid search are practically not inferior to adaptive models in terms of tuning speed and the risk of subsequent differences in accuracy between training data and data supplied during operation. At the same time, shallow neural networks make it possible to obtain more accurate forecasts, and due to their small size, they are trained quickly; and retraining can be performed automatically when new data arrives.