NeuraFirst's paper accepted at NeurIPS 2023 Workshop

We are thrilled to announce that our paper Symbolic regression for scientific discovery: an application to wind speed forecasting has been accepted for the NeurIPS 2023 North Africans in Machine Learning (NAML) workshop!

We are also excited to share with you that we have been invited to give a lightning talk about our paper, which will take place on December 11th.

In this blog post, we'll take you through the exciting journey of how we used symbolic regression to perform wind speed forecasting and share some key insights from our paper.

What is symbolic regression?

It is a set of techniques that strive to unveil an analytical formula from arbitrary data. The beauty of this approach lies in its inherent explainability and lightning-fast inference, which stands in stark contrast to the often opaque and resource-intensive nature of big neural network models.

In our paper, we harnessed symbolic regression to tackle wind speed forecasting, a critical task in weather prediction.

Dataset

We utilized data from the National Climatic Data Center (NCDC), focusing on five Danish cities and four weather features spanning from 2000 to 2010. The dataset had hourly time resolution and included temperature, pressure, wind speed, and wind direction as weather features.

The Equation Learner (EQL) Approach

EQL is a unique type of neural network that employs activation functions as primitive functions for the final analytical expression. It also implements a training protocol that encouraged sparsity while preventing overfitting, achieving a balance between accuracy and a compact final equation.

Results

Our final analytical equations revealed that the most influential weather feature in wind speed prediction is, unsurprisingly, wind speed itself. Furthermore, our work demonstrated how the wind speed of nearby cities can significantly impact the forecasting of a particular city, shedding light on the interconnected nature of weather patterns. Perhaps the most exciting finding was the superior inference time of our approach compared to conventional CNN-based networks. Additionally, an intriguing result emerged during the second phase of training, which appeared to learn an affine transformation.

In the figure below you can find the discovered analytical equation for each target city for 6h ahead wind speed predictions.

37th Annual NeurIPS

We are immensely proud of our research and grateful for the opportunity to present it at the 37th Annual NeurIPS (Advances of Neural Information Processing Systems), a conference widely considered the leading AI conference in the world. This year's NeurIPS promises to be an exceptional event, held from December 10th to 16th, 2023. You can find more details and the full program on the NeurIPS website.

Before we sign off, we'd like to extend our gratitude to the main contributing author of the paper:
Ismail Alaoui Abdellaoui.

For those eager to delve deeper into our work, you can read the abstract of our paper here.

Stay tuned for more updates as we prepare to present our findings at NeurIPS 2023!