Cross-location wind speed forecasting for wind energy applications using machine learning based models

TL;DR

This study develops machine learning models using Support Vector Machine and Random Forest to forecast wind speeds across different locations up to a year ahead, demonstrating high accuracy and consistency for wind energy applications.

Contribution

The paper introduces cross-location wind speed forecasting models that perform reliably over extended horizons, addressing the challenge of location-specific data limitations in wind energy forecasting.

Findings

80% predictions within 1.5 m/s RMSE up to a year ahead

Models perform consistently for different locations and times

Reliable forecasts up to 22 hours when trained on same location data

Abstract

The widespread utilisation of grid-integrated wind electricity necessitates accurate and reliable wind speed forecasting to ensure stable grid and quality power. Machine learning algorithm based wind speed forecasting models are getting increased attention in the literature owing to its superior ability to learn by effectively capturing the changing patterns from the data. Most of the reported wind forecasting models built on machine learning algorithms are location specific and tested against data adjacent to the training data. In this work, we develop the machine learning based wind speed forecasting models and analyse their performance when applied to data from different cross- locations up to a year ahead. Two distinct machine learning models based on Support Vector Machine (SVM) and Random Forest (RF) algorithms have been developed and tested separately for a relatively large geographical area. The results of analysis of 1-hour forecasts obtained at various cross-locations and points of time up to a year ahead show 80% of predictions within a Root Mean Square Error (RMSE) of 1.5 m/s, 95% within 2.5 m/s and 98% within an RMSEof 3.5 m/s. The 75% of 2-hour predictions are within RMSE of 1.5 m/s, 16-hour predictions within RMSE of 2.5 m / s and 48-hour predictions within RMSE of 3.5 m/s. When applied to thesame location of training data, the models generate reliable forecasts for periods up to 22 hours, with the added advantage that the models perform consistently throughout the year ahead horizon, independent of the lead time from the training data. The output of the analysis is highly promising to the wind energy industry in wind forecasting for locations where historical wind speed data are not available for model building and training.