Skip to content

Hourly Approach

Overview

The hourly approach is the default forecasting mode, producing 24 price predictions per day at 60-minute resolution. It aligns with the traditional OMIE hourly market clearing and serves as the foundation upon which the 15-minute approaches are built.

How It Works

  1. Data resolution: All input data (prices, demand, generation, weather) at hourly granularity
  2. Feature engineering: Hourly lags, 24h/168h rolling statistics, cyclical hour/day/month encoding
  3. Direct prediction: Separate model per horizon group, each predicting hourly target prices
  4. Output: 24 hourly prices for D+1 (day-ahead) or 144 hourly prices for D+2–D+7 (strategic)

Horizon Groups

The hourly approach uses two sets of horizon groups depending on the run mode:

Day-ahead (D+1):

GroupTarget HoursDescription
DA1D+1 00:00–11:00Morning block (12 hours)
DA2D+1 12:00–23:00Afternoon/evening block (12 hours)

Strategic (D+2–D+7):

GroupTarget PeriodDescription
S1D+2 full day24 hours
S2D+3 full day24 hours
S3D+4 full day24 hours
S4D+5 full day24 hours
S5D+6–D+748 hours combined

Feature Set

The hourly approach uses the full feature inventory:

  • Temporal: hour_sin/cos, dow_sin/cos, month_sin/cos, is_weekend, is_holiday
  • Price dynamics: Lags at 1h, 2h, 3h, 24h, 48h, 168h; rolling means and standard deviations
  • Demand/generation: Current demand, renewable/thermal shares, residual demand
  • Weather: Temperature, wind speed, solar radiation, degree days, interaction terms
  • Commodities: Gas price, carbon price, marginal cost proxy, spark spread
  • D+1 prices: (Strategic mode only) Published day-ahead price statistics

Temporal Encoding

Hours, days, and months are encoded using sin/cos pairs to preserve cyclical continuity:

hour_sin = sin(2π × hour / 24)
hour_cos = cos(2π × hour / 24)

This ensures that hour 23 and hour 0 are treated as adjacent (which they are) rather than distant (which ordinal encoding would imply).

When to Use

The hourly approach is appropriate when:

  • The target market clears at hourly resolution (traditional OMIE)
  • Maximum feature coverage is needed (all indicators available hourly)
  • Simplicity and interpretability are priorities
  • Historical comparison with pre-MTU15 forecasts is required

Relationship to 15-Minute Approaches

The hourly approach provides the baseline that the three 15-minute approaches extend:

  • Expanded: Takes hourly predictions and replicates each to 4 quarter-hourly slots
  • Pure15: Rebuilds the entire pipeline at 15-minute resolution
  • Hybrid15: Combines hourly-expanded features with native 15-minute data

See the individual approach pages for details on each strategy.