AI-driven prediction of wind energy production using machine learning, deep neural networks, and quantum AI — tested across two geographically distinct datasets.
This project aims to predict wind power generation using historical data and environmental variables such as wind speed, temperature, direction, and atmospheric conditions.
Two real-world datasets were used — from Adrar, Algeria (6 035 records, 15-min frequency) and Yalova, Turkey (50 530 records, 10-min frequency) — enabling cross-geographical generalisation testing.
The system leverages a broad spectrum of machine learning models to generate accurate and reliable forecasts, from classical statistical regression to quantum neural networks.
Wind energy production is highly variable and directly dependent on environmental conditions, making consistent supply difficult.
Accurate forecasting is essential for maintaining grid stability and preventing blackouts caused by sudden supply-demand imbalances.
Reliable predictions enable better energy management, storage allocation, and economic planning for renewable energy operators.
Classical statistical methods are often insufficient for capturing the non-linear dynamics of wind power — demanding intelligent predictive systems.
A rigorous six-stage process from raw data to actionable forecasts.
Wind speed, direction, and power records from Adrar (Algeria) and Yalova (Turkey) wind farms, sampled at 10–15 min intervals.
Missing value imputation via Random Forest & forward-fill; anomaly detection using Isolation Forest, One-Class SVM, LOF, and Elliptic Envelope.
Derived variables: wind deviation, usability factor, effective theoretical power, day/night flag, season, and meteorological enrichment via Meteostat.
15 models trained: statistical (Linear/Ridge/Lasso), boosting (XGBoost/LightGBM/GB), recurrent (RNN/LSTM/GRU), hybrid (CNN+LSTM/GRU), and Quantum AI.
Models generate short-term wind power predictions on held-out test sets for both geographic contexts.
Multi-metric assessment: RMSE, MAE, R², MAPE, SMAPE — across global and per-power-range strata for both datasets.
From classical statistics to cutting-edge quantum neural networks.
Baseline model establishing a linear relationship between variables and power output.
L2-regularised regression that stabilises coefficients for correlated inputs.
L1-regularised regression performing automatic feature selection by zeroing coefficients.
High-performance gradient boosted trees, excellent at handling noisy and non-linear data.
Leaf-wise tree growth for faster training and lower memory usage with strong accuracy.
Ensemble of weak learners sequentially correcting residuals to minimise prediction error.
Recurrent architecture capturing sequential dependencies in time-series energy data.
Long Short-Term Memory with forget/input/output gates for long-range temporal patterns.
Gated Recurrent Unit — a faster, streamlined alternative to LSTM with two gates.
Convolutional feature extraction combined with LSTM temporal modelling for complex time series.
Lightweight hybrid combining spatial pattern detection with efficient recurrent processing.
Parameterised quantum circuits + classical deep learning layers via TensorFlow Quantum. Achieves near-classical performance with fewer neurons.
Gradient Boosting and XGBoost achieved R² = 0.99 with RMSE ≈ 0.07 on the Algerian dataset — the top performers across both test sites.
Simulate a wind power prediction based on typical input conditions.
⚠ This is a UI simulation for portfolio demonstration purposes.
The project is fully reproducible with detailed Jupyter notebooks covering every stage of the pipeline.
This project demonstrates the ability to design and implement complete AI pipelines for real-world energy forecasting, from raw time-series data to production-ready predictive models. The work combines deep data science expertise — spanning classical statistics, gradient boosting, recurrent networks, and quantum AI — with a focus on practical impact in the renewable energy sector.
Future directions include richer multi-site datasets, transformer-based architectures, and more mature quantum hardware environments to further unlock the potential of quantum machine learning.