Simultaneous multi-step wind speed prediction on multiple farms using multi-task deep learning

In this paper, we present the MUSONet model, which leverages information from different sources (in this case, wind farms) to perform a multi-step wind speed prediction. The main goal of this approach is improving the global prediction accuracy, specifically at longer prediction horizons. Thus, the proposed model is able to simultaneously predict the wind speed at three different prediction horizons ( <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline" overflow="scroll"> <mml:mn>6</mml:mn> </mml:math> h, <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline" overflow="scroll"> <mml:mn>12</mml:mn> </mml:math> h, and <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline" overflow="scroll"> <mml:mn>24</mml:mn> </mml:math> h), across three different wind farms located in Spain. We also evaluate the performance of the presented methodology by considering three different activation functions for hidden neurons in the neural network: Sigmoid, ReLU, and ELUs+2L. The results show that the proposed multi-source approach improves the performance of the single-source counterpart for the longer prediction horizons ( <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline" overflow="scroll"> <mml:mn>12</mml:mn> </mml:math> h and <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline" overflow="scroll"> <mml:mn>24</mml:mn> </mml:math> h). In addition, the proposed multi-source method reduces by over <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline" overflow="scroll"> <mml:mn>30</mml:mn> <mml:mi mathvariant="normal">%</mml:mi> </mml:math> the number of parameters compared to three single-source models (in this case, one model per wind farm), resulting in a simpler solution for the problem addressed and requiring much lower computational resources.