The task is to predict day-ahead prouduction curve of a hydro power plant. The following figure showcases an example daily curve:

Notes:

• I always assume the prediction time to be at 00:00 of the previous day. At 31-Dec 00:00 I am making predictions for whole day 1-Jan 00:00 - 23:45. The time step t is 15 minutes long meaning there are 96 values to be predicted.
• Realized values used for prediction can only be from the previous day hence until 30-Dec 23:45. This applies to the completel daily curve prediction.
• Meteo features are used up until the point of prediction i.e. when predicting for 1-Jan 00:15, I use meteo features up until 1-Jan 00:15.

The project uses ploomber pipelining and has the following structure and tasks:

todo:

1. mypy check for static typing
2. testing task: data, methods...

• Variance of the observed target = huge, need to transform the target to do anything.
• (noisy target - hard for MSE loss functions etc etc...)
• most simple: averaging: hourly, daily -> some loss of information
• (IDEAS: filtering, GP transf.)

• daily mean transf. = significant loss of info, but doable for initial analysis
• full analysis in report: exploratory_target.ipynb

On daily, 1 step forward the residuals look good.

• NOTE: need to do 2 steps prediction, (curve prediction is day ahead)
• (every additional step, the error accumulates)

When I forward fill daily 2 step daily prediction into 15 minute target, the residuals look worse. NOTES:

• hold out is always from 11-Nov-2021, hence two months dataset
• Multiple ways of model checking for final model, weekly walk forward, daily walk forward.

Improving the daily with additional features:

• cyclicals: month, week_of_year encoded with sin, cos - usable in linear models
• (note: month 12, close to month 1)
• lagged daily target by two - (day ahead prediction)
• model: simple RF with few estimators fitted on residuals of true-daily - arima-daily
• hyperparams chosen by CV

ARIMA + RF on DALY RMSE=125 kw

ARIMA on DAILY RMSE=117 kw

RF + ARIMA was worse even on the daily so I did not bother forward filling and checking 15 min.

Full experiments in exploratory_darts_daily.ipynb

• 340 something days is obv. too small for any complex RNNS.
• Wanted to try if darts framework is any good
• Very simple RNN on daily learns approximately my baseline arima(2,1,1)
• Tried adding meteo covariates as well - no improvement

Hourly Full experiments in exploratory_darts_cnn.ipynb, exploratory_darts_hourly.ipynb

• generally I found it difficult to encode the day ahead nature of the problem using darts
• either there were leaks of realized Values that should not be known at the time of prediction
• or the network did not learn anything significant.
• Experimented with loss functions, early stoppers etc in hydro-timeseries.pytorch_utils
• Possible route to take but requires much more time and regularization.
• too few data points
• TCN, RNNs hard to do on a laptop
• IDEAS: multioutput RNN for regularization [15min, hourly, daily] as a target vector.

• Focus on transforming the target,e.g. remove trend and explain the residuals using meteo + other vars
• Use SMAPE as a model selection metric instead of others.

feature_manual.ipynb

The daily-arima(2,1,1) had the best error metrics hence I chose it as a transformer.

Transformed target

• Target is transformed with day-ahead prediction of daily arima
• some residual "trend" seem to remain

Full analysis in exloratory_meteo.ipynb

Multiplots - eyecheck for patterns between meteo and target

• positive correlations:

  • STRONG: rising temperature -> rising transpiration
  • WEAK: higher snow_mean -> higher soil_moisture
  • TARGET: soil moisture -> target

• negative correlations:

  • MEDIUM: drop in pressure -> high precip_mean (storm)

Significance is analyzed w.r.t. the transformed target with chosen transformer(i.e. arima-daily(2,1,1))

Top 10 raw meteo vars by abs correlation. soil_moisture, transpiration. Mostly stations 104, 81.

| soil_moisture_index_81  | 0.141967 |
| soil_moisture_index_104 | 0.140683 |
| soil_moisture_index_20  | 0.139955 |
| soil_moisture_index_134 | 0.135284 |
| soil_moisture_index_75  | 0.134490 |
| evapotranspiration_20   | 0.124984 |
| evapotranspiration_40   | 0.124436 |
| evapotranspiration_81   | 0.124313 |
| pressure_hpa_20         | 0.119161 |
| evapotranspiration_75   | 0.119053 |

(Tried means across meteo stations for simplicity of further processing - not a good idea. Loss of information.)

                    Results: Ordinary least squares
=======================================================================
Model:                OLS               Adj. R-squared:      0.039     
Dependent Variable:   0.0000            AIC:                 87034.5648
Date:                 2022-08-24 18:30  BIC:                 87134.7059
No. Observations:     31104             Log-Likelihood:      -43505.   
Df Model:             11                F-statistic:         116.7     
Df Residuals:         31092             Prob (F-statistic):  4.05e-263 
R-squared:            0.040             Scale:               0.96071   
-----------------------------------------------------------------------
                         Coef.  Std.Err.    t    P>|t|   [0.025  0.975]
-----------------------------------------------------------------------
const                    0.0000   0.0056  0.0000 1.0000 -0.0109  0.0109
temperature_mean         0.0674   0.0128  5.2602 0.0000  0.0423  0.0925
evapotranspiration_mean -0.1166   0.0137 -8.5238 0.0000 -0.1434 -0.0898
snow_mean                0.0230   0.0071  3.2265 0.0013  0.0090  0.0370
soil_moisture_mean       0.1617   0.0104 15.5221 0.0000  0.1413  0.1821
pressure_mean            0.1339   0.0060 22.2367 0.0000  0.1221  0.1457
precip_mean             -0.0086   0.0057 -1.5177 0.1291 -0.0198  0.0025
sin_month                0.3159   0.0353  8.9583 0.0000  0.2468  0.3850
cos_month               -0.2227   0.0392 -5.6795 0.0000 -0.2995 -0.1458
sin_week_of_year        -0.2408   0.0363 -6.6308 0.0000 -0.3120 -0.1696
cos_week_of_year         0.1502   0.0388  3.8676 0.0001  0.0741  0.2263
value_detr_daily_lag2   -0.0227   0.0057 -3.9963 0.0001 -0.0338 -0.0116
-----------------------------------------------------------------------
Omnibus:               4088.449       Durbin-Watson:          0.090    
Prob(Omnibus):         0.000          Jarque-Bera (JB):       17873.826
Skew:                  -0.586         Prob(JB):               0.000    
Kurtosis:              6.524          Condition No.:          22       
=======================================================================

Initially the raw features explained only ~4% of variance.

Both manual and auto generated features.

• Manual features include
  • ewms of all meteo, specific to station, mean, median across stations
  • lagged meteo vars,
  • lagged STL decomposition vars
  • lagged Value vars(lag 2 due to day ahead)
  • Datetime cyclicals sin, cos encoded
  • ...

• Tried: MutualInformation, ExtraTreesSelector, LGBMSelector, Lasso, in combination with RFE
• Most stable: Lasso(higher alpha)

• Walk forward checking takes time so initially I chose models using weekly walk forward of prediction month over the final 3-4 months.
• i.e. validation set is 30% of the original
• This creates 13 folds -> some statistic about model perf better than none.

with the best model I then perform day by day walk forward prediction on the hold out.

note: weights are unscaled.

Added auto generated features using tsfresh windowing.

• windows were performed on hourly, 15 minute steps etc.. of both realized values and meteo.
• made sure to window correctly
  • Value is windowed so that only D-1 and further back are known at the time of prediction for D+1
  • meteo windowed including the latest value.

• I used the same Lasso var selection
• auto generated features did not improve the arima-ridge model.

1. STL decomposition in feature-manual, add trend + seasonal + resid. Figure out the period.
2. Self defined objective function for LGBM, smape?
3. Multiple output NN for regularization - 15 min, 1 hourly, 1 daily.
4. Parquet instead of CSV
5. Long format instead of wide

1. Gaussian process to model the daily curve
2. Introduce temporality through prediction averaging:
   1. Daily - even true mean is not good enough
   2. Hourly?
   3. 30 minute avgs?
3. Parametric models for the daily curve
4. Filtering
5. Other transforms for the noisy target

1. ARIMA day ahead baseline - Done
2. ridge-arima improves daily arima with meteo for 15 minutes - Done
3. Gaussian process regressor - gpytorch
4. LGBM with better loss func. (optimize smape more directly?)
5. Multi output RNN - [15min, hourly, daily] - should regularize. (GRU)
6. Bayesian net
7. Tabnet
8. LSTMs - perhaps infeasible, only ~400 days to train for the 96 sequence lengths.

Requires Miniconda.

sudo apt install graphviz
conda env create -f environment.yml
conda activate hydro-power-prediction
pip install -e .
ploomber task load-data
ploomber task run-tests

conda env export -n hydro-power-prediction -f environment.yml --no-builds
conda env create -f environment.yml
conda env update --file environment.yml --prune

ploomber build

ploomber plot --backend pygraphviz

Raw: Trial 61 finished with value: 167.14923047681353 and parameters: {'boosting': 'gbdt', 'n_estimators': 910, 'max_depth': 13, 'min_data_in_leaf': 10, 'lambda_l1': 3.0, 'lambda_l2': 3.0}. Best is trial 61 with value: 167.14923047681353.

To set up the project inside slurm cluster, do one of the following:

srun -p ipu --gres=ipu:1 --cpus-per-task 128 --pty bash -i