Background
Forecasting renewable energy dynamics is crucial for designing smarter and more sustainable grids.
Most studies separately analyse energy consumption or renewable production, but rarely combine them into a unified measure of Green Energy Demand (GED), the proportion of electricity consumption that can be met by renewables.
This project introduces a new daily-resolution GED dataset for the UK (2000–2023), overcoming the lack of fine-grained data in official statistics (which are usually monthly or yearly). The dataset integrates:
- Energy consumption and generation (renewable + non-renewable)
- Weather features (temperature, wind, solar, daylight)
- Electricity prices
Research Goals
- Construct a daily GED dataset from heterogeneous sources.
- Benchmark classical forecasting methods (ARIMA, SARIMAX + GARCH).
- Develop deep learning models capable of handling long-term seasonality and short-term volatility.
- Evaluate performance with and without solar energy inputs, which introduce strong fluctuations.
Methodology
- ARIMA (5,0,2): Univariate baseline, capturing medium-term autocorrelations.
- SARIMAX + GARCH: Incorporated exogenous weather and market variables, then modelled residual volatility.
- Baseline LSTM: One-branch recurrent neural network with 30-day input windows.
- Fourier-Based Dual-Branch LSTM with Attention:
- Trend branch: Seasonal patterns represented by Fourier-transformed GED.
- Residual branch: Weather and price features (with solar inputs in some configurations).
- Attention fusion: Learnable weighting to adaptively balance trend and volatility signals.
Results
- ARIMA: Smooth forecasts but failed to capture volatility (R² = –1.17).
- SARIMAX + GARCH: Performed well without solar (R² ≈ 0.49) but collapsed when solar inputs were included.
- Baseline LSTM: Reasonable at seasonal trends but underfit volatility (R² ≈ 0.06).
- Fourier-Based Dual-Branch LSTM with Attention:
- Best-performing model, especially with solar inputs (R² ≈ 0.63).
- Trend branch captured long-term seasonality.
- Residual branch handled short-term, weather-driven fluctuations.
- Attention allowed the model to dynamically weigh both signals.
Contributions
- Introduced a new high-resolution GED dataset for the UK.
- Proposed a volatility-aware hybrid deep learning model combining Fourier decomposition, dual-branch LSTM, and attention fusion.
- Demonstrated that treating solar features through a dedicated residual pathway avoids overfitting and improves robustness.
Impact
This framework shows that hybrid, volatility-sensitive models can outperform traditional methods in forecasting renewable energy demand. Such models are highly relevant for:
- Sustainable grid planning
- Demand-side management
- Integration of renewables under uncertain and volatile conditions.