Design of Short-Term Wind Production Forecasting Model using Machine Learning Algorithms

Short-term (24h) wind production forecast is mainly used in energy trading in the day-ahead market (DAM) and intra-day market (IDM). The bid strategy of the market participants (wind producers) is based on wind production forecast which is one of the most important factors from an economic point of view. This paper presents the results of the proposed wind forecast model based on wind production data, weather data and supervised machine learning algorithms. The forecast model is built from scratch in Jupyter Lab with Python and Scikit-learn. In the development process of the forecast model, the most important algorithms (regression techniques), such as Linear regression, Ridge regression, Polynomial Ridge regression (order 4), Multilayer Perceptron regression, Decision Tree regression and Gradient Boosting regression are used. The data was collected, pre-processed, and used for the Machine Learning (ML) algorithms to prove the feasibility of the artificial intelligence applied to this field of work, with the final goal of improving the offers of the wind producers on the available energy markets. From the analysis of the wind production forecast results, it concluded that the most accurate model is the Polynomial Ridge Regression algorithm (order 4) with an error of 16.41%, measured with normalized mean absolute error (NMAE). The model accuracy could be improved by using deep learning algorithms such as Long-Short Term Memory (LSTM) or / add more features to the forecast model (forecast of the real wind speed).