HELIOEXPECT ML-CORRECTED SOLAR FORECAST

Solar forecasting errors don’t just hurt accuracy metrics — they cost real money.

Deviation penalties, imbalance charges, suboptimal schedules, and missed trading opportunities all stem from one root cause: forecasts that fail to reflect how your plant actually behaves.

At HelioExpect, we combine the reliability of physics with the adaptability of machine learning to deliver forecasts that improve with every day of operation.

The Real Problem With Solar Forecasting

Even the most advanced physics-based models struggle in the real world.

Solar forecast errors lead to:

• Deviation penalties from grid operators
• Imbalance charges in power markets
• Suboptimal day-ahead scheduling
• Missed trading and arbitrage opportunities

Physics-based models provide a strong baseline, but they cannot capture:

• Site-specific equipment behavior
• Local microclimate effects
• Systematic prediction biases
• How your plant responds to weather and operations

This is where most forecasting solutions stop — and where accuracy plateaus.

Our Core Philosophy

Physics gives you a foundation.
Machine learning gives you adaptation.

HelioExpect starts with a physics-based forecast and then applies machine learning trained on your plant’s actual performance to correct systematic errors.

The result is not a black box — but a physics-informed, data-driven correction layer that continuously improves accuracy.

The HelioExpect Forecasting Workflow

PHYSICS-BASED BASELINE
↓
HISTORICAL PLANT DATA
↓
INTELLIGENT FEATURE ENGINEERING
↓
OPTIMIZED MODEL TRAINING
↓
ML-CORRECTED FORECAST

Each step is designed to respect physical reality while learning from real-world behavior.

Step 1: Physics-Based Baseline

We begin with established solar physics:

• Solar position and irradiance modeling
• Atmospheric transmission calculations
• Panel temperature and efficiency curves
• System loss factors

This provides a scientifically grounded starting point that behaves reasonably under all conditions.

Step 2: Historical Plant Data

Next, we collect your plant’s actual performance:

• 90 days of SCADA generation data (5-min or 15-min intervals)
• Corresponding weather parameters:
  • GHI, DNI
  • Temperature, humidity
  • Wind speed
• Fully time-aligned and quality-checked

This is the data that captures how your plant really behaves — not how it’s supposed to behave on paper.

Step 3: Intelligent Feature Engineering

We transform raw data into features that encode domain knowledge.

Temporal Patterns

• Hour-of-day and day-of-week effects
• Cyclic encodings for smooth daily transitions
• Block-wise baselines using 21-day rolling statistics

Weather–Power Relationships

• Direct irradiance response
• Temperature efficiency effects
• Humidity-driven light scattering
• Wind effects on panel cooling and tracker behavior

Performance History

• Previous-day same-time-block generation
• 7-day rolling statistics per block
• Recent performance trends

These features teach the model how solar plants actually behave — not just how equations predict they should.

Step 4: Intelligent Model Training

This is where most systems fall short — and where HelioExpect differentiates.

Automated Parameter Optimization

• Learns from each training iteration
• Focuses on promising configurations
• Discovers optimal parameter combinations
• Typically delivers 5–15% higher accuracy than manual tuning

Leakage-Safe Validation

• Date-grouped cross-validation
• Prevents future data leakage
• Ensures realistic, deployable performance

Physics-Informed Constraints

• More irradiance → more power (never the reverse)
• Temperature effects follow known physics
• Predictions bounded by installed capacity

Daylight-Weighted Training

• Higher weight on daylight hours
• Nighttime values automatically zeroed
• Focus on periods that affect scheduling and penalties

Outlier-Robust Learning

• Specialized loss functions
• Equipment outages don’t corrupt training
• Sensor anomalies don’t distort predictions

Step 5: Bias Correction

Even strong ML models can drift.

We apply block-wise bias correction that:

• Analyzes recent forecast errors by time-of-day
• Corrects systematic over/under-prediction
• Adapts to current plant performance

This keeps forecasts aligned with operational reality.

Step 6: The ML-Corrected Forecast

The final output combines:

• Physics-based baseline
• Machine-learned corrections
• Recent bias adjustments
• Physical constraints and daylight masking

This ensures forecasts are accurate, stable, and physically plausible.

Why Physics + ML Works Better

Accuracy Improvements We See in Practice

Typical improvements after adding ML correction:

Actual results depend on data quality, baseline accuracy, and site characteristics.

Operational Benefits

Lower Deviation Costs

More accurate schedules mean fewer penalties from grid operators.

Better Market Performance

Confident day-ahead bids based on reliable forecasts.

Improved Planning

Trustworthy predictions for maintenance and operations.

Performance Insights

Understand exactly how your plant responds to weather and conditions.

Data Requirements

Minimum

• 30–60 days of SCADA data (15-min intervals)
• Plant specifications

Recommended

• 90+ days of historical data
• 5-minute resolution for higher accuracy

We handle:

• Data validation
• Timestamp alignment
• Missing data interpolation

Continuous Improvement by Design

HelioExpect models don’t stay static:

• Retrained daily with fresh data
• Cached models remain valid for 18 hours
• Bias correction updates continuously
• Seasonal patterns improve as data accumulates

Accuracy compounds over time.