fbmc-chronos2 / evaluate_forecasts.py
Evgueni Poloukarov
fix: add Windows multiprocessing protection and validation methodology
e5f4fec
raw
history blame
9.67 kB
#!/usr/bin/env python3
"""
Holdout Evaluation of Chronos 2 Zero-Shot Forecasts
Forecasts Sept 1-14, 2025 using context up to Aug 31, 2025
Compares against actual values to calculate MAE, RMSE, MAPE
"""
import pandas as pd
import numpy as np
import polars as pl
from datetime import datetime, timedelta
from chronos import Chronos2Pipeline
import torch
import time
import os
def main():
print("="*60)
print("CHRONOS 2 ZERO-SHOT EVALUATION")
print("="*60)
total_start = time.time()
# Step 1: Load dataset
print("\n[1/6] Loading dataset from local cache...")
start_time = time.time()
from datasets import load_dataset
# Load HF token from environment
hf_token = os.getenv("HF_TOKEN")
if not hf_token:
raise ValueError("HF_TOKEN not found in environment. Please set HF_TOKEN.")
dataset = load_dataset(
"evgueni-p/fbmc-features-24month",
split="train",
token=hf_token
)
df = pl.from_pandas(dataset.to_pandas())
# Ensure timestamp is datetime
if df['timestamp'].dtype == pl.String:
df = df.with_columns(pl.col('timestamp').str.to_datetime())
elif df['timestamp'].dtype != pl.Datetime:
df = df.with_columns(pl.col('timestamp').cast(pl.Datetime))
print(f"[OK] Loaded {len(df)} rows, {len(df.columns)} columns")
print(f" Date range: {df['timestamp'].min()} to {df['timestamp'].max()}")
print(f" Load time: {time.time() - start_time:.1f}s")
# Step 2: Identify target borders
print("\n[2/6] Identifying target borders...")
target_cols = [col for col in df.columns if col.startswith('target_border_')]
borders = [col.replace('target_border_', '') for col in target_cols]
print(f"[OK] Found {len(borders)} borders")
# Step 3: Define evaluation period
print("\n[3/6] Setting up holdout evaluation...")
# Holdout: Forecast Sept 1-14, 2025 using context up to Aug 31, 2025
holdout_end = datetime(2025, 8, 31, 23, 0, 0) # Last context timestamp
forecast_start = datetime(2025, 9, 1, 0, 0, 0) # First forecast timestamp
forecast_end = datetime(2025, 9, 14, 23, 0, 0) # Last forecast timestamp
context_hours = 512
prediction_hours = 336 # 14 days
print(f" Holdout evaluation period:")
print(f" Context: up to {holdout_end}")
print(f" Forecast: {forecast_start} to {forecast_end} (14 days)")
print(f" Context window: {context_hours} hours")
# Step 4: Extract actual values for evaluation
print("\n[4/6] Extracting actual values for evaluation period...")
actual_df = df.filter(
(pl.col('timestamp') >= forecast_start) &
(pl.col('timestamp') <= forecast_end)
)
print(f"[OK] Extracted {len(actual_df)} hours of actual values")
# Step 5: Load model
print("\n[5/6] Loading Chronos 2 model...")
model_start = time.time()
# Note: Running locally, will use CPU if CUDA not available
device = 'cuda' if torch.cuda.is_available() else 'cpu'
print(f" Using device: {device}")
pipeline = Chronos2Pipeline.from_pretrained(
'amazon/chronos-2',
device_map=device,
dtype=torch.float32 if device == 'cuda' else torch.float32
)
model_time = time.time() - model_start
print(f"[OK] Model loaded in {model_time:.1f}s")
# Step 6: Run inference for all borders and calculate metrics
print(f"\n[6/6] Running holdout evaluation for {len(borders)} borders...")
print(f" Progress:")
results = []
inference_times = []
for i, border in enumerate(borders, 1):
border_start = time.time()
# Get context data (up to Aug 31, 2025)
context_start = holdout_end - timedelta(hours=context_hours - 1)
context_df = df.filter(
(pl.col('timestamp') >= context_start) &
(pl.col('timestamp') <= holdout_end)
)
# Prepare context DataFrame
target_col = f'target_border_{border}'
context_data = context_df.select([
'timestamp',
pl.lit(border).alias('border'),
pl.col(target_col).alias('target')
]).to_pandas()
# Prepare future data
future_timestamps = pd.date_range(
start=forecast_start,
periods=prediction_hours,
freq='h'
)
future_data = pd.DataFrame({
'timestamp': future_timestamps,
'border': [border] * prediction_hours,
'target': [np.nan] * prediction_hours
})
# Combine and predict
combined_df = pd.concat([context_data, future_data], ignore_index=True)
try:
forecasts = pipeline.predict_df(
df=combined_df,
prediction_length=prediction_hours,
id_column='border',
timestamp_column='timestamp',
target='target'
)
# Get actual values for this border
actual_values = actual_df.select([
'timestamp',
pl.col(target_col).alias('actual')
]).to_pandas()
# Merge forecasts with actuals
merged = forecasts.merge(actual_values, on='timestamp', how='left')
# Calculate metrics using median (0.5 quantile) as point forecast
if '0.5' in merged.columns and 'actual' in merged.columns:
# Remove any rows with missing values
valid_data = merged[['0.5', 'actual']].dropna()
if len(valid_data) > 0:
mae = np.mean(np.abs(valid_data['0.5'] - valid_data['actual']))
rmse = np.sqrt(np.mean((valid_data['0.5'] - valid_data['actual'])**2))
mape = np.mean(np.abs((valid_data['0.5'] - valid_data['actual']) / (valid_data['actual'] + 1e-10))) * 100
results.append({
'border': border,
'mae': mae,
'rmse': rmse,
'mape': mape,
'n_points': len(valid_data),
'inference_time': time.time() - border_start
})
inference_times.append(time.time() - border_start)
status = "[OK]" if mae <= 150 else "[!]" # Target: <150 MW
print(f" [{i:2d}/{len(borders)}] {border:15s} - MAE: {mae:6.1f} MW {status}")
else:
print(f" [{i:2d}/{len(borders)}] {border:15s} - SKIPPED (no valid data)")
else:
print(f" [{i:2d}/{len(borders)}] {border:15s} - FAILED (missing columns)")
except Exception as e:
print(f" [{i:2d}/{len(borders)}] {border:15s} - ERROR: {e}")
inference_time = time.time() - model_start - model_time
# Step 7: Calculate and display summary statistics
print("\n" + "="*60)
print("EVALUATION RESULTS SUMMARY")
print("="*60)
if results:
results_df = pd.DataFrame(results)
print(f"\nBorders evaluated: {len(results)}/{len(borders)}")
print(f"Total inference time: {inference_time:.1f}s ({inference_time / 60:.2f} min)")
print(f"Average per border: {np.mean(inference_times):.2f}s")
print(f"\n*** OVERALL METRICS ***")
print(f"Mean MAE: {results_df['mae'].mean():.2f} MW (Target: ≤134 MW)")
print(f"Mean RMSE: {results_df['rmse'].mean():.2f} MW")
print(f"Mean MAPE: {results_df['mape'].mean():.2f}%")
print(f"\n*** DISTRIBUTION ***")
print(f"MAE: Min={results_df['mae'].min():.2f}, Median={results_df['mae'].median():.2f}, Max={results_df['mae'].max():.2f}")
print(f"RMSE: Min={results_df['rmse'].min():.2f}, Median={results_df['rmse'].median():.2f}, Max={results_df['rmse'].max():.2f}")
print(f"MAPE: Min={results_df['mape'].min():.2f}%, Median={results_df['mape'].median():.2f}%, Max={results_df['mape'].max():.2f}%")
# Target achievement
below_target = (results_df['mae'] <= 150).sum()
print(f"\n*** TARGET ACHIEVEMENT ***")
print(f"Borders with MAE ≤150 MW: {below_target}/{len(results)} ({below_target/len(results)*100:.1f}%)")
# Best and worst performers
print(f"\n*** TOP 5 BEST PERFORMERS (Lowest MAE) ***")
best = results_df.nsmallest(5, 'mae')[['border', 'mae', 'rmse', 'mape']]
for idx, row in best.iterrows():
print(f" {row['border']:15s}: MAE={row['mae']:6.1f} MW, RMSE={row['rmse']:6.1f} MW, MAPE={row['mape']:5.1f}%")
print(f"\n*** TOP 5 WORST PERFORMERS (Highest MAE) ***")
worst = results_df.nlargest(5, 'mae')[['border', 'mae', 'rmse', 'mape']]
for idx, row in worst.iterrows():
print(f" {row['border']:15s}: MAE={row['mae']:6.1f} MW, RMSE={row['rmse']:6.1f} MW, MAPE={row['mape']:5.1f}%")
# Save results
output_file = 'results/evaluation_results.csv'
results_df.to_csv(output_file, index=False)
print(f"\n[OK] Detailed results saved to: {output_file}")
print("="*60)
if results_df['mae'].mean() <= 134:
print("[OK] TARGET ACHIEVED! Mean MAE ≤134 MW")
else:
print(f"[!] Target not met. Mean MAE: {results_df['mae'].mean():.2f} MW (Target: ≤134 MW)")
print(" Consider fine-tuning for Phase 2")
print("="*60)
else:
print("[!] No results to evaluate")
# Total time
total_time = time.time() - total_start
print(f"\nTotal evaluation time: {total_time:.1f}s ({total_time / 60:.1f} min)")
if __name__ == '__main__':
main()