This project is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0) license.
🔗 Live Dashboard: Access the Streamlit Application
This project focuses on the performance optimization of a Combined Cycle Power Plant (CCPP). By developing a digital twin using historical sensor data, the system predicts net hourly electrical power output (PE) and identifies the optimal operating setpoints to maximize efficiency under fluctuating ambient conditions.
As a Mechanical Engineer, understanding the underlying thermodynamics is crucial to validate the machine learning model's behavior. Ambient parameters deeply affect the gas turbine cycle (Brayton) and the steam turbine cycle (Rankine):
- Ambient Temperature (AT): Higher ambient temperatures reduce air density. Since the gas turbine is a constant-volume machine, a lower air mass flow rate enters the compressor, decreasing the overall power output and efficiency.
- Relative Humidity (RH) & Moisture Effects: In thermal systems, high humidity levels alter the specific heat capacity of the working fluid. When air contains a high concentration of water vapor, additional energy is required just to heat up this non-reactive water mass up to the combustion temperature. In gas turbines or combustion furnaces, this creates an energy penalty (parasitic heat load), reducing the net power output.
- Exhaust Vacuum (V): A lower vacuum pressure in the condenser increases the overall enthalpy drop across the steam turbine, directly boosting the net generated Megawatts (MW).
- Prescriptive Optimization: Implementation of a mathematical optimization module using
scipy.optimize(SLSQP) to calculate the ideal exhaust vacuum (V) setpoint, maximizing power output based on real-time ambient inputs. - Mathematical Smoothing: Utilizing a Polynomial Regression (Degree 2 Pipeline) to guarantee a continuous, differentiable surface. This mathematical pivot prevents gradient-based optimizers from getting trapped in local minima or discrete "steps" common in tree-based algorithms like XGBoost.
- Feature Engineering: Advanced thermodynamic indices integration (Air Density Index, AT-V Interaction, Specific Humidity) to enhance the model's physical intuition and convergence.
- Interactive Dashboard: A Streamlit-based UI that allows operators to perform real-time "What-if" analysis and optimization tasks.
- Language: Python
- ML & Regressions: Scikit-learn (PolynomialFeatures, LinearRegression, XGBRegressor,RandomForestRegressor, GradientBoostingRegressor)
- Optimization Engine: SciPy (SLSQP algorithm)
- UI/Dashboard: Streamlit
- Data Engineering: Pandas, NumPy
- Visualization: Matplotlib, Seaborn
The Polynomial Pipeline achieves an outstanding balance between predictive accuracy and prescriptive utility:
-
$R^2$ Score: 0.9419 (explaining over 94% of the operational variance). - RMSE: 4.1060 MW (~1% relative error across the typical production range).
The digital twin correctly identifies the underlying thermodynamic principle: minimizing exhaust vacuum (
/data: Contains the CCPP dataset./notebooks: Exploratory Data Analysis (EDA), model benchmarking (XGBoost vs. Random Forest vs. Polynomial), and sensitivity analysis./img: Images, graphics and correlation matrixes obtained from the EDA process./deployment: The app script used for the streamlit demo and the pre-trained and serialized Polynomial Pipeline ready for production.requirements.txt: List of dependencies required to run the environment.
The interactive console acts as a web-based Digital Twin for control-room decision-making. Operators can adjust weather conditions via sliders, and the optimization engine automatically calculates the target operation threshold.
- Prescriptive Success: Transitioning from a tree-based model (XGBoost) to a smooth, continuous model (Polynomial) resolved numerical optimization deadlocks, unlocking stable gradient calculations.
- Industrial Synergy: Proves that combining data science frameworks with mechanical engineering domain knowledge produces robust, physically stable models.
- Future Iterations: Future updates will target adding a financial constraint module (calculating cooling tower pumping costs vs. MW gains) to find the absolute economic optimum, rather than just the thermodynamic maximum.
The data used in this project is sourced from the Combined Cycle Power Plant dataset hosted by the UCI Machine Learning Repository:
P. Tufekci and H. Kaya. "Combined Cycle Power Plant," UCI Machine Learning Repository, 2014. [Online]. Available: https://doi.org/10.24432/C5002N.
Julen Neila Garcia | [www.linkedin.com/in/julen-neila-garcia-a42304268]
Passionate about bridging the gap between Mechanical Engineering and Industrial Data Science.