Inspiration
As the global push for clean energy accelerates, we realized that optimizing our existing energy infrastructure is just as critical as building new ones. Combined Cycle Power Plants (CCPP) are incredibly efficient because they use both gas and steam turbines, but their output is highly volatile and sensitive to subtle atmospheric changes—specifically temperature, pressure, humidity, and exhaust vacuum. When grid operators can't accurately predict a plant's energy output, they are forced to over-generate power (wasting massive amounts of fossil fuels) or under-generate (causing grid instability and blackouts). We were inspired to bridge the gap between heavy industrial infrastructure and modern artificial intelligence. We wanted to build a tool that removes the guesswork from power generation, empowering operators to optimize dispatch, cut down carbon emissions, and save millions in operational costs.
What it does
PowerOpt AI is a real-time, AI-native web dashboard that acts as the "brain" for a Combined Cycle Power Plant.
Plant operators input real-time environmental metrics (Ambient Temperature $T$, Ambient Pressure $AP$, Relative Humidity $RH$, and Exhaust Vacuum $V$), and the underlying Random Forest Regression model instantly forecasts the exact Net Hourly Electrical Energy Output ($PE$) in Megawatts (MW).
Beyond just raw numbers, the dashboard features a rich "Glassmorphism" UI that provides:
Live Output Tracking: Dynamic Chart.js visualizations that track predicted outputs over time. Deep Analytics: Exploratory Data Analysis (EDA) breakdowns showing how specific variables negatively or positively correlate with power output. Model Evaluation: Full transparency into the ML lifecycle, allowing operators to see exactly why the Random Forest model (achieving a $96.16%$ $R^2$ score) outperforms traditional linear and polynomial algorithms.
How we built it
We approached this as a true end-to-end full-stack AI problem:
The Brain (Machine Learning): We started with a rich dataset of 9,568 data points collected over 6 years. Using scikit-learn, pandas, and numpy, we tested five distinct regression algorithms. Random Forest ultimately won out because of its natural ability to handle the non-linear relationship between variables like Exhaust Vacuum and Energy Output without overfitting. We serialized this trained model using joblib. The Nervous System (Backend API): We built a lightweight, asynchronous Flask server in Python. This REST API ingests the JSON payload from the frontend, runs it through the rf_model.joblib, and returns the precise MW prediction in milliseconds. The Face (Frontend UI/UX): We consciously avoided heavy frameworks like React or Angular, opting instead for pure HTML5, Vanilla JavaScript, and raw CSS3. We implemented a stunning, GPU-accelerated Glassmorphism design system to give the dashboard a premium, futuristic feel. We used Chart.js for the live data visualization and embedded micro-animations to make the interface feel alive and responsive.
Challenges we ran into
Model Selection & Overfitting: In our initial tests, Polynomial Regression showed promise, but it began to overfit when handling extreme outliers in temperature and pressure. We had to pivot to an ensemble method (Random Forest) and carefully tune our hyperparameters (n_estimators=100) to balance variance and bias. Frontend-to-Backend Latency: Because the dashboard is meant for real-time grid adjustments, we needed the prediction to return instantly. We initially struggled with slow response times when loading the model into memory. We solved this by pre-loading the .joblib file into the Flask app's global state at server startup, rather than loading it on every /predict request. UI Architecture: Designing a complex dashboard using only Vanilla CSS and JS without a framework was a massive challenge. Managing the DOM states for the single-page application (SPA) navigation and ensuring the charts didn't re-render redundantly required writing very clean, modular JavaScript.
Accomplishments that we're proud of
Incredible Accuracy: We are incredibly proud of achieving an $R^2$ score of $96.16%$, which translates to predictions that are accurate enough to be safely deployed in real-world energy markets. Zero-Dependency Frontend: We built a highly dynamic, reactive, and visually stunning SPA dashboard entirely from scratch using Vanilla HTML/CSS/JS, proving that you don't always need heavy frameworks to create a premium user experience. Translating Data to Design: We successfully transformed a dry, mathematical dataset of atmospheric variables into a beautiful, engaging tool that a non-technical plant operator could immediately understand and use.
What we learned
The Power of Ensemble Learning: We learned firsthand why ensemble methods like Random Forests are the industry standard for tabular data. Watching it effortlessly map the non-linear relationships that stumped the Multiple Linear Regression model was a massive "aha!" moment. State Management without Frameworks: We significantly leveled up our Vanilla JS skills, specifically regarding asynchronous API calls, DOM manipulation, and handling application state across multiple views. Bridging the Gap: Most importantly, we learned how to bridge the gap between data science and product design. A highly accurate model is useless if the end-user can't understand it. Designing the UI was just as important as tuning the algorithm.
What's next for PowerOpt AI: NextGen Power Predictor
IoT Sensor Integration: We plan to move away from manual data entry by building direct API integrations with physical IoT sensors inside power plants, allowing the dashboard to update autonomously every second. Cost-Benefit Forecasting Module: We want to integrate real-time energy market pricing. By combining our MW predictions with live energy prices, we can tell operators exactly how much revenue they are generating or losing per minute. Expanded Architectures: We plan to adapt the backend to support other types of power generation facilities, such as Solar and Wind farms, by creating a modular ML pipeline that detects the facility type and loads the appropriate predictive model.
Log in or sign up for Devpost to join the conversation.