4
[Repost Q] Need help understanding this code on power consumption prediction using LSTM
(self.learnmachinelearning)
Original answer:
This is a time series prediction task. They are using a deep learning model with LSTM. There are too many points for fast NN training so they resample. From the results it looks like there is a bit of low pass filtering. This data is at least somewhat periodic and they haven't done a frequency analysis which they should have done imo. You don't need an NN for this type of prediction. I would not consider a deep learning strategy to be a good idea for a intro ML course project because they lack explainability.
For the second question, depends on the restrictions for your project. As an EE, I'd say state estimation (such as kalman filtering) or model fitting / regression (you create a model with some parameters the use an algorithm to optimize the parameters based on some data) but those would not really covered in intro ML courses (kalman filtering is usually under control but a lot of controls is ML). There are a lot of stats things that fit into ML as well.
If you have the stats background (you should as an eng student), regression is very useful. Linear regression is very simple https://onlinestatbook.com/2/regression/intro.html If that is too simple then look at curve fitting: https://ncss-wpengine.netdna-ssl.com/wp-content/themes/ncss/pdf/Procedures/NCSS/Introduction_to_Curve_Fitting.pdf
Imo, the some of the easiest ML algorithms to start with are classification. Linear SVM is an easy classifier to explain and you could apply it to some data set of your choosing https://towardsdatascience.com/https-medium-com-pupalerushikesh-svm-f4b42800e989
If you do want to do something with neural networks, you can showcase how they are nondeterministic. Just take a simple NN and run it with different initial weights and see how the results change. If you save all the results (for example, accuracy) and plot them as a histogram, then you'll see a normal or skewed distribution. You can also switch the training/val/test split to see what that does as well. (this will probably show you a big flaw with some of your classmates projects) You can explain how this impacts how we use NN (safety, etc) and why it happens. Here is something to get you started: https://machinelearningmastery.com/reproducible-results-neural-networks-keras/