Yılmaz et al. (2026) Prediction of hydroelectric power generation with machine learning and innovative combined deep learning techniques

Identification

• Journal: Stochastic Environmental Research and Risk Assessment
• Year: 2026
• Date: 2026-01-01
• Authors: Banu Yılmaz, Egemen Aras, Saeed Samadianfard
• DOI: 10.1007/s00477-025-03140-8

Research Groups

• Department of Civil Engineering, Karadeniz Technical University, Trabzon, Turkey
• Department of Civil Engineering, Bursa Technical University, Bursa, Turkey
• Department of Water Engineering, University of Tabriz, Tabriz, Iran

Short Summary

This study develops and evaluates machine learning and innovative combined deep learning techniques, including a novel Temporal Pattern Attention Feed-Forward Neural Network-Long Short-Term Memory (TPAFFNN-LSTM) model, for monthly hydroelectric power generation forecasting at the Altınkaya Dam Basin in Turkey, demonstrating its superior accuracy compared to other methods.

Objective

• To investigate the usability of atmospheric and hydrological parameters (precipitation, temperature, and flow) for monthly hydroelectric power generation forecasting at the Altınkaya Dam.
• To propose and evaluate an innovative hybrid deep learning model (TPAFFNN-LSTM) that combines a Temporal Pattern Attention (TPA) mechanism with Feed-Forward Neural Networks (FFNN) and Long Short-Term Memory (LSTM) networks, aiming for higher accuracy in forecasting compared to existing machine learning and deep learning methods.

Study Configuration

• Spatial Scale: Altınkaya Dam Basin, Kızılırmak River, Turkey. The dam has an installed capacity of 700 MW.
• Temporal Scale: Monthly data from 2013 to 2023 (11 years).

Methodology and Data

• Models used:
  • Deep Learning: Long Short-Term Memory (LSTM), Feed-Forward Neural Network (FFNN), Temporal Pattern Attention Feed-Forward Neural Network-Long Short-Term Memory (TPAFFNN-LSTM - proposed hybrid model).
  • Machine Learning: Random Forest (RF), Extreme Gradient Boosting (XGB).
  • Feature Selection: Least Absolute Shrinkage and Selection Operator (LASSO) regression.
  • Evaluation: Shapley Additive Explanations (SHAP) for sensitivity analysis, Regression Receiver Operating Characteristic (RROC) curve and Area Over the RROC Curve (AOC) for model performance.
  • Hyperparameter Optimization: Bayesian optimization for LSTM.
• Data sources:
  • Monthly average temperature and total precipitation: Turkish State Meteorological Service (MGM).
  • Monthly inflow data (in cubic hectometres, hm³): General Directorate of State Hydraulic Works (DSİ).
  • Real-time electricity generation data (in megawatt-hours, MWh): Transparency platform of Energy Exchange Istanbul (EPİAŞ).

Main Results

• The proposed TPAFFNN-LSTM method achieved the highest performance for monthly hydroelectric power generation forecasting, with a normalized Root Mean Square Error (nRMSE) of 0.16 and a Nash–Sutcliffe Efficiency (NSE) of 0.69.
• These results indicate that TPAFFNN-LSTM was 18% more successful in nRMSE and 19% more successful in NSE compared to other evaluated methods (FFNN, LSTM, RF, XGB).
• The NSE value of 0.69 for TPAFFNN-LSTM falls into the 'good' range (0.65 < NSE < 0.75), while FFNN and LSTM were in the 'adequate' range (0.50 < NSE < 0.65).
• RROC analysis showed that TPAFFNN-LSTM had the lowest Area Over the RROC Curve (AOC) value of 11.72, indicating superior accuracy and more balanced predictions compared to other models.
• LASSO regression effectively identified and selected the most influential input parameters, particularly lagged states of precipitation, temperature, and flow, reducing model complexity and preventing overfitting.
• SHAP analysis confirmed the significant influence of lagged predictors on the forecasting model's decisions, consistent with hydrological processes.
• TPAFFNN-LSTM demonstrated the highest success in estimating maximum monthly generation values and exhibited a standard deviation closest to the observed values, indicating better capture of data variability.

Contributions

• Introduction of an innovative hybrid deep learning model (TPAFFNN-LSTM) that synergistically combines a Temporal Pattern Attention (TPA) mechanism, Feed-Forward Neural Networks (FFNN), and Long Short-Term Memory (LSTM) networks to enhance the accuracy and interpretability of hydroelectric power generation forecasting.
• Demonstration of the TPAFFNN-LSTM model's superior predictive performance (18-19% more successful in nRMSE and NSE) compared to classical machine learning (Random Forest, XGBoost) and other deep learning (FFNN, LSTM) methods.
• Application of LASSO regression for effective feature selection and complexity reduction, and comprehensive evaluation using SHAP and RROC analyses, providing detailed insights into model interpretability and performance in the context of hydropower forecasting.
• Provision of a robust and generalizable framework for monthly energy production forecasting that can be adapted to other hydropower systems and basins, offering practical value for dam operators, energy managers, and policymakers in managing energy resources under changing climate conditions.
• This study is the first to apply LSTM and the innovative TPAFFNN-LSTM methods with the specific set of input parameters for the Altınkaya Dam Basin, addressing a gap in the literature for this critical region.

Funding

Not explicitly stated in the paper.

Citation

@article{Yılmaz2026Prediction,
  author = {Yılmaz, Banu and Aras, Egemen and Samadianfard, Saeed},
  title = {Prediction of hydroelectric power generation with machine learning and innovative combined deep learning techniques},
  journal = {Stochastic Environmental Research and Risk Assessment},
  year = {2026},
  doi = {10.1007/s00477-025-03140-8},
  url = {https://doi.org/10.1007/s00477-025-03140-8}
}