Time_series_ML

Analysis and prediction of time-series data of temperature and heating demand

Corresponding Presentation

https://docs.google.com/presentation/d/1eMNVJstHoVfn2XtqSaIOQUNjUOtHGVVfSz-riOgJmCM/edit?usp=sharing

Content

• EDA Notebook: Project_Notebook_EDA.IPYNB
• LSTM Notebook: Project_Notebook_LSTM.IPYNB + Project_Notebook_LSTM_Attention.IPYNB
• Prophet + SARIMAX Notebook: Project_Notebook_PROPHET&SARIMAX.ipynb

Data

1. when2heat_DE (column: DE_heat_demand_total):

   • Type: int
   • Description: Heat demand in Germany in MW for space and water heating

2. weather data

   • from 2004 to 2022
   • hourly data
   • two values per timestamp: TT_TU & RF_TU

Goal

Predict heating demand. -> with this information local electricity provider can better stabilise their grid. -> with more heatpumps used, heating demand will be more important for electricity grid than ever

Forecast Horizon: 48 hours

Input sequence length of 14days24hours and 30days24hours

Loss:

Mean Squared Error on MinMax scaled data range 0 - 1

Models

Univariate Models

Only use heating demand as input data:

• SARIMAX
• LSTM

Multivariate Models

• Prophet (+ Holidays)
• Prophet (+ Holidays + Temperature)

Model Comparison

Always compare 2day forecast-horizon and use mean of all predictions done in 2014. Use Min-Max Scaler(0,1).

Model Name	Hyperparameter	Additional features	MSE
SARIMAX	SARIMAX_v1	nothing added	0.000869
Prophet	Default-Values	holidays	0.007121
Prophet	prophet_v1	holidays	0.0029120
Prophet_with_temp	prophet_v2	holidays, temp	0.0025222
LSTM 8 units	lr=1e-2/Restart	N/A	0.001879228
LSTM + ATTENTION	lr=1e-2/Restart	N/A	0.001313091

Hyperparameter

To decide on the Hyperparameters mentioned in the below tables a gridsearch was executed. See the grid search in the AllExperiments.ipynb notebook.

SARIMAX_v1:

order	seasonal_order
(1, 0, 1)	(1, 1, 1, 24)

prophet_v1:

holidays_bool	weekly_seasonality	daily_seasonality_bool	changepoint_prior_scale	holidays_prior_scale	daily_fourier
True	50	False	0.3	0.3	5

prophet_v2:

holidays_bool	daily_seas	daily_seasonality_bool	changepoint_prior_scale	holidays_prior_scale	daily_fourier
True	10	False	0.3	0.1	5

LSTM: Architecture: LSTM_Heat_model.png

LSTM + ATTENTION Architecture: LSTM_Heat_model_attention.png