Pure 15-Minute

Overview

The pure 15-minute approach trains models exclusively on native 15-minute resolution data. Unlike the expanded approach (which replicates hourly values), pure15 learns genuine sub-hourly price dynamics, producing forecasts that capture intra-hour variation from ramp events, solar intermittency, and demand fluctuations.

How It Works

Data source: The ree_15min table containing quarter-hourly prices, demand, and generation
Feature engineering: Lags and rolling statistics computed at 15-minute scale
Training: Direct multi-horizon models with 15-minute-native samples
Output: 96 predictions per day with genuine intra-hour variation

15-Minute Feature Scaling

All temporal features are rescaled to 15-minute steps:

Feature	Hourly	Pure15
1-hour lag	`shift(1)`	`shift(4)`
24-hour lag	`shift(24)`	`shift(96)`
7-day lag	`shift(168)`	`shift(672)`
24h rolling mean	`window=24`	`window=96`
7d rolling mean	`window=168`	`window=672`

Quarter-of-Day Encoding

Instead of the standard 24-period hour cycle, pure15 uses a 96-period quarter-of-day cycle:

quarter_of_day = hour × 4 + minute ÷ 15     (range: 0–95)
target_quarter_sin = sin(2π × quarter / 96)
target_quarter_cos = cos(2π × quarter / 96)

This preserves cyclical continuity at quarter-hourly resolution — quarter 95 (23:45) is adjacent to quarter 0 (00:00).

Horizon Groups (15-Minute)

Pure15 uses day-level horizon groups, each covering 96 quarter-hourly periods:

Group	Quarters Ahead	Target Period
D1	1–96	D+1 full day
D2	97–192	D+2 full day
D3	193–288	D+3 full day
D4	289–384	D+4 full day
D5	385–480	D+5 full day
D6	481–576	D+6 full day
D7	577–672	D+7 full day

Origin Step Optimization

To manage training data volume (4× more samples than hourly), origins are sampled every 4th 15-minute step (i.e., at hourly intervals):

All 15-min steps:  00:00, 00:15, 00:30, 00:45, 01:00, ...
Training origins:  00:00,                         01:00, ...  (origin_step=4)

This reduces training set size by 75% while maintaining daily coverage.

Advantages

Genuine intra-hour variation — captures real 15-minute price dynamics
Sub-hourly pattern learning — can model solar ramp events, demand spikes within hours
Temporal consistency — features and targets at the same resolution, no mixing

Limitations

Reduced feature set — only indicators available at 15-minute resolution can be used (interconnection data, some weather variables are hourly-only)
Data history requirement — needs 6+ months of 15-minute data for reliable training
Training cost — 4× more samples per origin (mitigated by origin step)
Newer, less proven — shorter track record than hourly models

When to Use

The pure15 approach is appropriate when:

Sufficient 15-minute historical data exists (6+ months since OMIE MTU15 transition)
Intra-hour price variation is important for the use case (e.g., battery optimization)
Feature set limitations are acceptable
Maximum temporal consistency is desired