The week 2 baseline model (linear) looks rather like a strawman proposal. There surely are better non-neural models to fit this data that are still simple enough to work as a baseline, e.g. a sigmoid function. I’m not very familiar with Python numerical libraries, but I bet there is one that can do sigmoid function fitting to the power vs. wind speed data. Wolfram Mathematica certainly can.

A sigmoid function can be a suitable alternative to the linear baseline model for fitting S-shaped data like power vs. wind speed. Python’s numerical libraries, such as SciPy, provide tools like curve_fit for fitting sigmoid functions to data. By using curve_fit, you can fit a sigmoid function to your power vs. wind speed data and extract the fitted parameters for analysis or further calculations.

I agree that wind speed vs power output seems to take on an S-shape. But I wouldn’t call this a strawman proposal or an over-simplified baseline that is ‘easy to beat’. In the lab, we create a multi-variate linear model in n-dimensional space, using multiple attributes - not just wind speed. The other attributes do not necessarily have the same ‘s-shape’ when plotted against power output, the way wind speed does. As such, a simple multi-variate linear model seems to be a fine starting point that is easy to set up and compute.

But then, one could argue that wind speed so dominates in feature importance, that perhaps a single-variable sigmoid model using wind speed would be a better baseline than a multi-variate linear model.