Spaces:
Sleeping
Sleeping
File size: 12,466 Bytes
f4be780 d080539 f4be780 d080539 27ce714 069627f 27ce714 069627f a2c07c4 069627f f4be780 d080539 f4be780 5ff6f25 f4be780 de602fd 2e13800 de602fd 2e13800 0bb8c81 2e13800 de602fd f4be780 27ce714 069627f 27ce714 069627f f4be780 04dcb8d f4be780 069627f a2c07c4 f4be780 069627f a2c07c4 f4be780 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 |
#!/usr/bin/env python3
"""
Dynamic Forecast Module v1.8.0 - Context Window (47 Days / 1.5 Months)
Time-aware data extraction for forecasting with run-date awareness.
Purpose: Prevent data leakage by extracting data AS IT WAS KNOWN at run time.
Key Concepts:
- run_date: When the forecast is made (e.g., "2025-09-30 23:00")
- forecast_horizon: Always 14 days (D+1 to D+14, fixed at 336 hours)
- context_window: Historical data before run_date (1,125 hours = 47 days / 1.5 months, fits A100-80GB)
- future_covariates: ALL 2,514 features (leveraging Chronos-2 past-only masking)
* 603 full-horizon (known future)
* 12 partial D+1 (masked D+2-D+14)
* 1,899 historical (masked as past-only covariates)
Chronos-2 Past-Only Covariate Masking:
- Historical features have NaN future values → Chronos-2 sets mask=0
- Model learns cross-feature correlations from historical context
- Attention mechanism uses dimensional structure even when values masked
- Enables learning of CNEC/volatility patterns without future knowledge
"""
from typing import Dict, Tuple, Optional
import pandas as pd
import polars as pl
import numpy as np
from datetime import datetime, timedelta
from src.forecasting.feature_availability import FeatureAvailability
class DynamicForecast:
"""
Handles time-aware data extraction for forecasting.
Ensures no data leakage by only using data available at run_date.
"""
def __init__(
self,
dataset: pl.DataFrame,
context_hours: int = 1125, # 1,125 hours = 46.9 days (1.5 months, fits A100-80GB)
forecast_hours: int = 336 # Fixed at 14 days
):
"""
Initialize dynamic forecast handler.
Args:
dataset: Polars DataFrame with all features
context_hours: Hours of historical context (default 1440 = 60 days)
forecast_hours: Forecast horizon in hours (default 336 = 14 days)
"""
self.dataset = dataset
self.context_hours = context_hours
self.forecast_hours = forecast_hours
# Categorize features on initialization
self.categories = FeatureAvailability.categorize_features(dataset.columns)
# Validate categorization
is_valid, warnings = FeatureAvailability.validate_categorization(
self.categories, verbose=False
)
if not is_valid:
print("[!] WARNING: Feature categorization issues detected")
for w in warnings:
print(f" - {w}")
def prepare_forecast_data(
self,
run_date: datetime,
border: str
) -> Tuple[pd.DataFrame, pd.DataFrame]:
"""
Prepare context and future data for a single border forecast.
Args:
run_date: When the forecast is made (all data before this is historical)
border: Border to forecast (e.g., "AT_CZ")
Returns:
Tuple of (context_data, future_data):
- context_data: Historical features + target (pandas DataFrame)
- future_data: Future covariates only (pandas DataFrame)
"""
# Step 1: Extract historical context
context_data = self._extract_context(run_date, border)
# Step 2: Extract future covariates
future_data = self._extract_future_covariates(run_date, border)
# Step 3: Apply availability masking
future_data = self._apply_masking(future_data, run_date)
# Step 4: Align dtypes between context and future
# Chronos-2 requires matching dtypes for columns that appear in both DataFrames
# After masking, int columns may become float due to NaN values
# Solution: Convert ALL numeric columns to float64 in both DataFrames
import pandas as pd
common_cols = set(context_data.columns) & set(future_data.columns)
for col in common_cols:
if col in ['timestamp', 'border']:
continue # Skip non-numeric columns
# Convert both context and future to float64 for consistency
# This ensures Chronos-2's validation passes (requires matching dtypes)
# Use pd.to_numeric() which handles NaN gracefully (unlike .astype())
context_data[col] = pd.to_numeric(context_data[col], errors='coerce').astype('float64')
future_data[col] = pd.to_numeric(future_data[col], errors='coerce').astype('float64')
return context_data, future_data
def _extract_context(
self,
run_date: datetime,
border: str
) -> pd.DataFrame:
"""
Extract historical context data.
Context includes:
- All features (full+partial+historical) up to run_date
- Target values up to run_date
Args:
run_date: Cutoff timestamp
border: Border identifier
Returns:
Pandas DataFrame with columns: timestamp, border, target, all_features
"""
# Calculate context window
context_start = run_date - timedelta(hours=self.context_hours)
# Filter data
context_df = self.dataset.filter(
(pl.col('timestamp') >= context_start) &
(pl.col('timestamp') < run_date)
)
# Select target column for this border
target_col = f'target_border_{border}'
# All features (we'll use all for context, Chronos-2 handles it)
all_features = (
self.categories['full_horizon_d14'] +
self.categories['partial_d1'] +
self.categories['historical']
)
# Build context DataFrame
context_cols = ['timestamp', target_col] + all_features
context_data = context_df.select(context_cols).to_pandas()
# Add border identifier and rename target
context_data['border'] = border
context_data = context_data.rename(columns={target_col: 'target'})
# Reorder: timestamp, border, target, features
context_data = context_data[['timestamp', 'border', 'target'] + all_features]
return context_data
def _extract_future_covariates(
self,
run_date: datetime,
border: str
) -> pd.DataFrame:
"""
Extract future covariate data for D+1 to D+14.
Future covariates include ALL 2,514 features using Chronos-2's past-only masking:
- Full-horizon D+14: 603 features (known future values)
- Partial D+1: 12 features (load forecasts, masked D+2-D+14)
- Historical: 1,899 features (MASKED as past-only covariates)
Past-only covariates leverage Chronos-2's mask-based attention:
- Future values are NaN (unknown)
- Chronos-2 sets mask=0 for these dimensions
- Model learns cross-feature correlations from historical context
- Attention mechanism uses structure even when future values masked
Args:
run_date: Forecast run timestamp
border: Border identifier
Returns:
Pandas DataFrame with columns: timestamp, border, future_features
"""
# Calculate future window
# IMPORTANT: Chronos-2 predict_df() expects future_df to start at the LAST context timestamp,
# not the first forecast timestamp. See dataset.py:549 assertion.
forecast_start = run_date # Start at last context timestamp
forecast_end = forecast_start + timedelta(hours=self.forecast_hours - 1)
# Filter data
future_df = self.dataset.filter(
(pl.col('timestamp') >= forecast_start) &
(pl.col('timestamp') <= forecast_end)
)
# Include ALL features (3,043 total) to leverage past-only covariate masking
# Historical features will be NaN in future → Chronos-2 masks them automatically
future_features = (
self.categories['full_horizon_d14'] + # 603 known-future
self.categories['partial_d1'] + # 12 partial
self.categories['historical'] # ~2,428 past-only (MASKED!)
)
# Build future DataFrame
future_cols = ['timestamp'] + future_features
future_data = future_df.select(future_cols).to_pandas()
# Add border identifier
future_data['border'] = border
# Reorder: timestamp, border, features
future_data = future_data[['timestamp', 'border'] + future_features]
return future_data
def _apply_masking(
self,
future_data: pd.DataFrame,
run_date: datetime
) -> pd.DataFrame:
"""
Apply availability masking for partial features.
Masking:
- Load forecasts (12 features): Available D+1 only, masked D+2-D+14
- LTA (40 features): Forward-fill from last known value
Args:
future_data: DataFrame with future covariates
run_date: Forecast run timestamp
Returns:
DataFrame with masking applied
"""
# Calculate D+1 cutoff (24 hours after run_date)
d1_cutoff = run_date + timedelta(hours=24)
# Mask load forecasts for D+2 onwards
for col in self.categories['partial_d1']:
# Set to NaN (or 0) for hours beyond D+1
mask = future_data['timestamp'] > d1_cutoff
future_data.loc[mask, col] = np.nan # Chronos-2 handles NaN
# Forward-fill LTA values
# Note: LTA values in dataset should already be forward-filled during
# feature engineering, but we ensure consistency here
lta_cols = [c for c in self.categories['full_horizon_d14']
if c.startswith('lta_')]
# LTA is constant across forecast horizon (use first value)
if len(lta_cols) > 0 and len(future_data) > 0:
first_values = future_data[lta_cols].iloc[0]
for col in lta_cols:
future_data[col] = first_values[col]
return future_data
def validate_no_leakage(
self,
context_data: pd.DataFrame,
future_data: pd.DataFrame,
run_date: datetime
) -> Tuple[bool, list]:
"""
Validate that no data leakage exists.
Checks:
1. All context timestamps < run_date
2. All future timestamps >= run_date + 1 hour
3. No overlap between context and future
4. Future data only contains future covariates
Args:
context_data: Historical context
future_data: Future covariates
run_date: Forecast run timestamp
Returns:
Tuple of (is_valid, errors)
"""
errors = []
# Check 1: Context timestamps
if context_data['timestamp'].max() >= run_date:
errors.append(
f"Context data leaks into future: max timestamp "
f"{context_data['timestamp'].max()} >= run_date {run_date}"
)
# Check 2: Future timestamps
forecast_start = run_date + timedelta(hours=1)
if future_data['timestamp'].min() < forecast_start:
errors.append(
f"Future data includes historical: min timestamp "
f"{future_data['timestamp'].min()} < forecast_start {forecast_start}"
)
# Check 3: No overlap
if (context_data['timestamp'].max() >= future_data['timestamp'].min()):
errors.append("Overlap detected between context and future data")
# Check 4: Future columns
future_features = set(
self.categories['full_horizon_d14'] +
self.categories['partial_d1']
)
future_cols = set(future_data.columns) - {'timestamp', 'border'}
if not future_cols.issubset(future_features):
extra_cols = future_cols - future_features
errors.append(
f"Future data contains non-future features: {extra_cols}"
)
is_valid = len(errors) == 0
return is_valid, errors
def get_feature_summary(self) -> Dict[str, int]:
"""
Get summary of feature categorization.
Returns:
Dictionary with feature counts by category
"""
return {
'full_horizon_d14': len(self.categories['full_horizon_d14']),
'partial_d1': len(self.categories['partial_d1']),
'historical': len(self.categories['historical']),
'total': sum(len(v) for v in self.categories.values())
}
|