Deep Learning model to predict spikes in energy demand based on historical environmental and power consumption data. Hyperparameters are controlled via hydra and Logging is done with neptune.
Model currently implemented is a LSTM taking the last 7 days of environmental (Temperature, Windspeeds, Solargobalradiation) and power consumption data in 15-minute resolution. The model predicts the expected time of the daily maximum in electrical load in a 15 minute interval between 7AM and 8PM.
Energy data is retrieved from German Energy Providers, in this case the Thueringer Energienetze. Weather data is retrieved from the German Weather Forecasting Service Deutscher Wetterdienst.
Currently data is limited to 1y of data.
First, install dependencies (before you do that, install Miniconda.)
# clone spike
git clone https://github.com/thbuerg/spike
# install dependencies
cd spike
conda env create -f environment.yml
# install spike
pip install -e .
If you want to log to neptune, make sure your Neptune.ai API token is exported to the environment
Next, navigate to train.py and execute it.
# module folder
cd spike/spike/
# run preprocessing
python preprocessing.py +experiment.filepath=/path/to/your/data/
# run module
python train.py \
+experiment.filepath=/path/to/your/data/data_normed.csv \
+trainer.default_root_dir=/path/to/your/results/ \
+trainer.num_sanity_val_steps=1 +trainer.max_epochs=100 trainer.gpus=[0] experiment.batch_size=128 experiment.learning_rate=0.001@article{Time Series Forecasting for Energy Demand Spike Prediction (SPIKE),
title={Time Series Forecasting for Energy Demand Spike Prediction (SPIKE)},
author={Thore Buergel},
journal={github},
year={2021}
}