Feature request: Add electricity price to dataset

[chris-y]

Hello, I've been playing around with this a little and I think it might be beneficial to the forecast model to add the electricity price to the dataset. (if I have this right, this is a "future regressor" in neuralprophet speak)

There should be some correlation between electricity price and load. This may negate the need to manually subtract the car charging load, as it will learn that these heavy loads only occur in times of lower electricity prices, which in turn will stop it forecasting heavy load where there has been past free electricity/power-up or daytime IOG slots.

[ragg987]

Yes please, it would be good if we can pass additional sensors to give context - in my case I've just set up a few energy sensors using PredAI, and am currently having a go at using these as a load input for Predbat.

One of these sensors relates to energy used for heating my house. The predictions would be much more useful if I could pass temperature forecast and actuals to Neural Prophet. Or even set and demand temperature for heating.

From the NP site:

Lagged regressors: Lagged observations (e.g temperature sensor) - linear or NN.
Future regressors: In advance known features (e.g. temperature forecast) - linear or NN.

I know it should be possible to do some energy modelling in Predheat, however that is giving an error - I've created a separate issue in github.

[ragg987]

Just thinking...

The ability to pass historical and forward looking data could well provide a (partial?) replacement for predheat as it currently stands (I know you are thinking about developing it further @springfall2008 e.g. to control heating). I am currently modelling my home heating energy with the following parameters, the figures in brackets are my estimate of gain over the next 24 hrs as an indication of how strongly the factors impact demand (negative is a loss).

• solar forecast > solar gain through windows (+25 kWh in the next 24hrs)
• solar shading applied (no shading applied in winter)
• wind speed forecast (-5 kWh)
• number of occupants (+15 kWh)
• fabric and ventilation losses (-57 kWh)
• external temperature forecast
• thermostat set temperature
• current house temperature
• ASHP power output vs input (COP)

In my example I estimate I will need a net input from the ASHP of 25kWh for the next 24hrs to keep the house warm, this translates to approx 6hrs of run-time on the ASHP.

I wonder if feeding all this data into NP would give a decent model over time? If so feeding it some forward-looking data like wind speed, external temperature, number of occupants etc might give a nice heating energy demand forecast?

The limitations of my current method is that it is quite static as it takes a daily snapshot of the forecast for the next 24h and converts that into ASHP on-time. Quite crude, but for our house (Passiv standard) it seems to be working, though I am having to constantly fine-tune my plan assumptions.

I cannot get predheat to work to confirm or improve my modelling - it throws an error - issue logged in preheat. Having said that predheat cannot currently factor in some of the above - e.g. solar gain has quite a large impact on energy required in my case.

[peter1rhodes]

This is kind of fundamental I think, and it may be that I misunderstand how predai works, but surely we need to give it quite a few sensors for it to be accurate? There's no way it can only look at the history of one sensor and predict its future well. I'd want to give it the outside temperate, binary sensors telling it whether we're at home, whether we're on holiday, the energy price of course, when certain devices have been switched on (e.g. a 3D printer, once on, will usually consume 150W for a few hours, the dishwasher is rarely run more than once in a day, etc).

I assumed when I first installed it that predai looks at all Home Assistant entities and makes its predictions from there, but I'm not sure that can be the case. So my fairly fundamental question is, how do you tell it what sensors to use to make its prediction(s)?

[peter1rhodes]

Looking at the closed issues, it appears that my question was asked in #4 , and answered by trefor:

It's pretty simplistic right now it just uses the specified sensor for the prediction. In future it maybe possible to include other data. The historical data of that sensor is used for the prediction.

So that answers that. It seems to me that this feature request is basically part of this more general request: to be able to provide other 'input' sensors for a given output prediction. I agree with this feature request! I think it could make the prediction quality, and thus predbat's performance, much better!