Adaptive Online Learning for Short-Term PV Power Prediction

Co-Supervised by: Fereshteh Jafari

If you are interested in this topic or have further questions, do not hesitate to contact fereshteh.jafari@unibe.ch.

Background / Context

Traditional photovoltaic (PV) power prediction models are trained offline using historical data and remain static during deployment. However, solar energy systems operate in dynamic environments where weather patterns, seasonal changes, equipment aging, and new PV installations continuously affect performance. Static models cannot adapt to these evolving conditions, leading to degraded prediction accuracy over time. Online learning approaches offer a solution by enabling models to continuously update their parameters as new data becomes available. This capability is particularly valuable for short-term forecasting (10 minutes to 2 hours ahead), where rapid weather changes can significantly impact prediction accuracy and energy management decisions.

Research Question(s) / Goals

The primary goal is to develop and evaluate an adaptive online learning system for short- term PV power prediction that can:

Continuously adapt to changing environmental conditions and system characteristics
Maintain or improve prediction accuracy compared to static models over extended periods
Operate efficiently in real-time with minimal computational overhead
Handle concept drift and seasonal variations in solar generation patterns

Approach / Methods

The student will use an offline trained neural network for short-term PV power prediction.

In a first approach, the offline training will be replaced by an online training scheme, on the same data archive, starting from initial weight and bias settings. The objective is to measure how fast the online training method converges with the offline training results.
Then, the data archive will be modified to study the capacity of the online training to adapt to changing environmental conditions and system characteristics.
Finally, both the online training and the offline trained networks will be run in real time on new data, and the prediction performance will be compared.
Different online training methods (Markov chains, dynamic programming, stochastic gradient descent, etc) will be tested.
Computational load and memory requirements shall be analyzed.

Expected Contributions / Outcomes

A functioning online learning system for PV power prediction
Comprehensive evaluation demonstrating adaptation benefits over static models
Analysis of trade-offs between adaptation speed and prediction stability
Recommendations for optimal hyperparameter settings and update strategies
Documentation of system performance under various weather conditions and seasonal changes

Required Skills / Prerequisites

Python programming and familiarity with machine learning libraries (scikit-learn, numpy, pandas)
Basic understanding of regression algorithms and time series forecasting
Knowledge of data preprocessing and feature engineering
Experience with data visualization tools (matplotlib, seaborn)
Understanding of model evaluation metrics and statistical analysis

Possible Extensions

Integration with real-time weather data APIs
Implementation of ensemble methods combining multiple online learners
Extension to multi-step ahead forecasting

Further Reading / Starting Literature

Bottou, L. (2010). “Large-scale machine learning with stochastic gradient descent.” Proceedings of COMPSTAT’2010.
Antonanzas, J., et al. (2016). “Review of photovoltaic power forecasting.” Solar Energy, 136, 78-111.
Losing, V., Hammer, B., & Wersing, H. (2018). “Incremental on-line learning: A review and comparison of state of the art algorithms.” Neurocomputing, 275, 1261-1274.