Increasing the Accuracy of Hourly Multi-Output Solar Power Forecast with Physics-Informed Machine Learning

Area of Science:

• Renewable Energy Systems
• Machine Learning Applications
• Solar Power Forecasting

Background:

• Machine learning (ML) methods show promise for photovoltaic (PV) power forecasting, typically up to one day ahead.
• Accurate long-term PV power forecasting is crucial for grid stability and energy management.

Purpose of the Study:

• To introduce a generic physical model into ML predictors for enhanced PV power forecasting from one to three days ahead.
• To evaluate the effectiveness of physics-informed features in improving ML-based PV power prediction.
• To present a generalized methodology for evaluating physics-informed approaches in PV forecasting.

Main Methods:

• Incorporated a generic physical model of a PV system into ML predictors.
• Recombined basic power, wind speed, and air temperature measurements into physics-informed features.
• Evaluated five ML methods (Random Forest, SVM, CNN, LSTM, CNN-LSTM) using a case study in Denmark.
• Assessed the impact of feature engineering and forecasting horizon on prediction accuracy.

Main Results:

• Physics-informed features consistently improved the performance of all evaluated ML methods for PV power forecasting.
• The proposed approach enhanced forecasting accuracy across different time horizons (1-3 days).
• A threshold was identified for the optimal number of previous samples, exhibiting a convex relationship.

Conclusions:

• Integrating physics-informed features significantly enhances ML-based PV power forecasting accuracy.
• The methodology is effective across various ML algorithms and forecasting horizons.
• Feature engineering based on physical principles simplifies model training and improves generalization.