AI stuff

[Joelius300]

I don't have an issue for it yet so this is just to collect some links without having to edit the readme:

• How (not) to use Machine Learning for time series forecasting: Avoiding the pitfalls
• Forecasting: Principles and Practice (by Rob J Hyndman and George Athanasopoulos)
• An Overview of ML.Net (by KTL Solutions)
• https://towardsdatascience.com/introducing-pytorch-forecasting-64de99b9ef46
• https://towardsdatascience.com/how-to-detect-and-deal-with-multicollinearity-9e02b18695f1
• https://medium.com/analytics-vidhya/tensorflow-2-tutorial-on-categorical-features-embedding-93dd81027ea9
• https://machinelearningmastery.com/how-to-develop-lstm-models-for-time-series-forecasting/
• https://machinelearningmastery.com/multivariate-time-series-forecasting-lstms-keras/

I'm tired and inexperienced but I see two possibilities: supervised learning with a linear regression model where you can learn lots of stuff including which variables are influencing which, or un supervised learning with LSTM which is more expensive, harder to understand but will most likely yield better results until you really understand what's going on and can brew up an efficient and smart linear regression model (nvm they're both supervised IIRC). So I think create a good LSTM model then fiddle around with the linear regression one, trying to improve it with the goal of beating the LSTM one.

Autocorrelation might be very interesting and an important topic when deciding how to model the system.

Ps. don't forget to try a model with as little features as possible so only those that you know are related like boiler, puffer, kessel and aussen (and seasonality and stuff).

Another question: currently there is a threshold of 45° C on the boiler which leads to puffer activity aka refilling the boiler as soon as it goes below that threshold. Question now is will the model pick that up by itself or is there feature engineering necessary? Someone with more knowledge and experience could probably answer this outright but maybe some trial and error will do nicely here.

[Joelius300]

A convo with an architect has shown me that it might be a good idea to collect weather forecast data alongside the current state of the heating unit to further enhance the AIs learning potential. One idea would be to store the predicted temperature for the place of residence in the next few hours. More granularity probably isn't necessary but maybe extending further than the next few hours would be interesting. It would definitely make sense to start storing this data as soon as possible because that'll be the data the AI will be trained on and the more data we have the better. However I am not quite sure how I would go about storing it because just five new columns with "predicition1h" "prediction2h" etc. seems terrible. I'd much rather use a second table and join them when training the dataset, this way I can also easily limit the amount of times weather data is fetched to 20 min for example instead of the 6 second rate that will be the norm for heating data going forward.

Also sun exposure could be very interesting especially considering that our house doesn't get as much sun as the average in the vicinity. But for that you'd need some sort of sensor on the roof or something like that.

[Joelius300]

Setup Grafana for an easy way to play around with the data, I think building a dashboard like that would be a lot easier than creating a custom graphing etc. solution for playing around with the data.
Of course the end-user still gets a cool custom application, otherwise where would be the fun in this project.

[Joelius300]

stuff someone said may be useful for frequency analysis

• periodogram (SciPy)
• windowing

[Joelius300]

dviz_hslu 99f77d7d5

[Joelius300]

In ML, always try simple first. Regression first, then more complex models. Try without feature engineering first, then add complexity. Etc.

[Joelius300]

https://www.featuretools.com/