Forecasting Hourly Electricity Prices

Various models and approaches are deployed to forecast hourly electricity prices provided by the EXIST Market Transparency Platform. First, lag-1 of the exogenous variable are used to make prediction with values at time t-1 to forecast the electricity price at time t (one hour later). For that, a baseline XGBM deployed to observe what pre-fine-tuning performance looks like. Afterwards, XGBM model tuned with various hyperparameters including but not limited to max_depth, n_estimators, subsamples and so on. Fine-tuned XGBM model revealed better performance in terms of root mean squared errors than what the baseline model achieved.

Finally, nn.LSTM model developed to use the capabilities of LSTM algorithm with long-term time-dependencies. Nn.LSTM Time-series model achieved better (less rmse) performance compared to fine-tuned XGBM model. Below you can find the model architecture. 48-hour look-back horizon is used to predict horizon of +1.

Data Description

Variable Name	Description
PTF	Market Clearing Price is the hourly energy price that is determined with respect to orders that are cleared according to total supply and demand
Volume	In the Day Ahead Market, it is the hourly total financial value of the matching bids. (Matching bid = Matching Offer)
bid_amount	Sum of hourly, block and flexible bid order volumes at 0 TL/MWh price step.
ask_amount	Sum of hourly, block and flexible sales order volumes at 4800 TL/MWh price step.
imbalance_delta	Imbalance Quantity

Model Architecture

XGBM Prediction Performance

nn.LSTM Prediction Performance

Model Performances

Model	RMSE
nn.LSTM	57.348
Fine-tuned XGBM	59.972
Baseline XGBM	61.42