Hybrid 15-Minute

Overview

The hybrid 15-minute approach combines hourly features (expanded to 15-minute resolution) with native 15-minute data where available. This maximizes the feature set by leveraging all data regardless of native resolution, offering the best of both the expanded and pure15 approaches.

How It Works

1. Hourly features: Standard hourly indicators (interconnections, some weather) expanded to 15-minute resolution by repeating each value four times
2. Native 15-min features: Quarter-hourly prices, demand, and generation from the ree_15min table
3. Merge: Both feature sets combined into a single training matrix
4. Training: Direct multi-horizon models on the combined dataset
5. Output: 96 predictions per day with intra-hour variation

Feature Merging Strategy

Hourly data (expanded):          Native 15-min data:
├─ interconnections (×4)         ├─ 15-min prices
├─ weather hourly (×4)           ├─ 15-min demand
├─ commodities (×4)              ├─ 15-min generation
└─ hourly lags (×4)              └─ 15-min lags/rolling

         ↓ combine_first ↓

Merged training matrix:
├─ All hourly features (expanded)
├─ All native 15-min features
├─ 15-min cyclical encoding (quarter_sin/cos)
└─ Standard temporal features

The combine_first strategy ensures that native 15-minute data takes precedence where available, falling back to expanded hourly data for indicators that lack sub-hourly resolution.

Advantages

• Maximum feature coverage — uses all available data from both resolutions
• Intra-hour variation — native 15-minute prices capture real sub-hourly dynamics
• Best of both worlds — hourly indicators (interconnections, commodities) that aren’t available at 15-min are still included via expansion
• Larger effective training set — combines samples from both data sources

Limitations

• Mixed resolution artifacts — expanded hourly features are flat within each hour, which may confuse models when combined with varying 15-min features
• Implementation complexity — requires careful alignment of two data sources at different native resolutions
• Potential redundancy — some features appear at both resolutions, requiring deduplication
• Data history requirement — still needs 6+ months of 15-minute data for the native component

When to Use

The hybrid15 approach is appropriate when:

• Both hourly and 15-minute data are available with sufficient history
• Maximum feature coverage is a priority
• The use case requires intra-hour variation (not just flat expanded profiles)
• Computational resources allow for the larger combined feature set

Comparison Summary

Aspect	Expanded	Pure15	Hybrid15
Hourly features	Yes (native)	No	Yes (expanded)
15-min features	No	Yes (native)	Yes (native)
Intra-hour variation	None	Full	Full
Feature count	Highest (hourly)	Lowest	Highest (combined)
Data requirement	Minimal	6+ months 15-min	6+ months 15-min
Resolution mixing	None	None	Yes
Complexity	Low	Medium	High

Future Direction

As the OMIE MTU15 transition matures and more indicators become available at native 15-minute resolution, the hybrid approach will gradually converge with the pure15 approach. The hybrid strategy serves as a bridge during the transition period when data availability is mixed.