Intraday Electricity Market Price Forecast Using Machine Learning
Overview
This Bachelor’s thesis develops short-horizon forecasts for Turkey’s intraday electricity market using hourly transparency data from the national exchange (EPIAS / EXIST), together with aligned day-ahead prices and system generation signals. The study situates the problem in Turkish market design (continuous intraday matching against day-ahead clearing) and follows a full workflow from exploratory analysis through anomaly-aware cleaning, dimension reduction, baselines, and deep sequence models.
Simple statistical benchmarks were fit mainly for one-step-ahead comparison; operationally, bids need at least a two-hour lead, so the main neural experiments target that horizon with a 168-hour lookback. After extreme prices were flagged with Prophet (fit separately before and after the 2018 FX shock) and imputed using the model’s expected level, five independent feature-selection routes were reconciled to a compact predictor set dominated by current and lagged day-ahead and intraday prices. Among LSTM, Bidirectional LSTM, GRU, CNN, and a CNN–LSTM hybrid, the hybrid achieved the lowest held-out mean absolute error; additional TensorBoard-driven tuning on that architecture reduced test MAE to roughly 21 TL/MWh on the thesis split.
Deliverables
- Market and data narrative: Literature and institutional background on Turkish electricity markets; descriptive analysis of intraday and day-ahead prices, FX effects, and generation mix over 2015–2021.
- Structured feature table: Hourly panel built from clearing prices, imbalances, planned and realized production, calendar fields, and multi-step lags (same-hour prior days and weeks).
- Anomaly workflow: Regime-split screening with Prophet, replacement of flagged hours using the library’s fitted forecast level, stabilizing volatile TL series post-2018.
- Feature selection: Intersection of rankings from exhaustive search, sequential forward/backward selection, ridge, and lasso, yielding nine consensus inputs for neural models.
- Baselines: Moving and weighted moving averages (multiple windows), exponential smoothing, autoregressive and ARMA-style specifications calibrated against one-step errors as described in the thesis.
- Deep models: TensorFlow/Keras implementations with MAE on a chronological 80/20 train–test partition; CNN–LSTM refined via structured hyperparameter trials.
- Code: Implementation and notebooks on GitHub.
- Written thesis: Full thesis (PDF).
Technology stack
- Python: End-to-end modeling, data assembly from transparency feeds, and evaluation.
- TensorFlow / Keras: LSTM, BiLSTM, GRU, CNN, and CNN–LSTM architectures; TensorBoard for training diagnostics.
- scikit-learn: Feature selection pipelines and classical baselines aligned with the thesis experiments.
- Prophet: Time-series anomaly detection and level adjustment for extreme intraday prints.
- NumPy / pandas workflow: Hourly alignment, lag construction, and train-test splitting without random shuffle (temporal order preserved).