Using time series AI methods on IoT data captured from wind turbines.

We recommend to first setup conda environment with Python 3.7+ and then install dependencies using pip:

pip install -r requirements.txt

Download the data from the wind energy company Engie's open data page and put those csv files in data/ folder. Descriptions about the columns in this data is provided in data_description.tsv file. Once data is available we need to fill in missing values which will create data/imputed/ folder with same filenames but with all datetime instances filled.

cd code
python impute_missing_data.py

A streamlit app for visualizing wind turbine data is available in this repo. Line charts of two columns for four turbines are shown at once. User could also select the date range from which the charts need to be shown. A correlation matrix between the columns in the time series data is also shown in the app. At the end box plot of first column with second column as x-axis is also given. streamlit command can be used to start the app.

cd exploratory_data_analysis
streamlit run streamlit_app.py

For training and evaluating forecasting models, currently the split date is hardcoded to 2017-01-01. Data before this is used for trianing the models and after this date for evaluation. The wind turbine specific columns are chosen as the fields to be predicted and other weather data provided are used as covariates. We also create extra covariates from date-time.

We support using the covariates as past_covariates or future_covariates or both. One could also choose to use time or weather or both covariates. Currently using the repo one can train and evaluate following time-series forecasting models provided in the DARTS library:

• TCNModel
• TransformerModel
• NBEATSModel
• BlockRnnModel
• DLinearModel
• NLinearModel

An example of training an NBEATS model on R80711 turbine data with weather covariates

cd code
python run.py \
  --do_train \
  --expt_name darts \
  --turbine_name R80711 \
  --epochs 10 \
  --batch_size=128 \
  --model_type nbeats \
  --mode weather_covariates \
  --covariates past 

Trained model with be saved in models/darts_nbeats_weather_covariates_past_10.pt

To evaluate the model

cd code
python run.py \
  --do_eval \
  --turbine_name R80711 \
  --model_type nbeats \
  --model_path ../models/darts_nbeats_weather_covariates_past_10.pt \
  --forecast_horizon 12 \
  --results_path ../results/nbeats_eval_results.tsv

One csv file with MAE, MAPE and SMAPE scores for each column predictions will be saved in results_path.