Data-Driven Peak Voltage Forecasting in PV-Integrated Grids: Enabling Smarter Grid Interventions
As solar adoption grows, local grids face increasing voltage challenges. This thesis shows how machine learning can predict peak voltages in grids with high PV penetration, enabling smarter interventions and reducing reliance on static curtailment.
Topic
This thesis addresses the challenge of maintaining voltage stability in local distribution grids with increasing photovoltaic (PV) integration. Traditional mitigation strategies, such as static curtailment or fixed inverter settings, often lack precision and result in unnecessary energy losses. In some cases, the approval of new PV systems is even withheld to ensure voltage stability. This study investigates whether machine learning can improve voltage forecasting using real operational and weather data. It also examines how PV-related structural features, such as installed capacity and system count, can enhance model accuracy. These insights may provide grid operators with more detailed information for targeted interventions, supporting more efficient grid management.
Relevance
The rapid growth of decentralized PV systems is reshaping the Swiss energy landscape and creating new challenges for local grid operation. In residential areas, voltage peaks are becoming more frequent, while existing mitigation strategies remain static and untargeted. This thesis responds to an urgent operational need by exploring predictive, data-driven alternatives that align with ongoing regulatory reforms and support a more flexible and efficient integration of solar energy into local grids.
Results
The analysis shows that machine learning can accurately forecast peak voltage in PV-integrated distribution grids. Among the models tested, XGBoost delivered the highest accuracy and robustness, outperforming Random Forest and Artificial Neural Networks. Including PV-related features, such as installed capacity and system count, improved model performance. The results confirm the technical feasibility of data-driven forecasting and highlight the importance of structural PV characteristics for accurate voltage prediction.
Implications for Practitioners
- Machine learning models such as XGBoost should be considered for voltage forecasting in low-voltage grids with high PV integration
- Predictive forecasting can enable distribution grid operators to move beyond fixed inverter settings and blanket curtailment by allowing targeted, location-specific interventions
- Existing infrastructure, such as ripple control systems, can be enhanced through dynamic, condition-based activation that responds to predicted voltage events
- Systematic and targeted measurement practices should be developed to close current data gaps and to enable reliable, data-driven grid management
Methods
Three machine learning models were examined to predict peak voltage in PV-integrated low-voltage grids: XGBoost, Random Forest and Artificial Neural Networks. Real operational data from 15 Swiss transformer stations were combined with surplus feed-in measurements and historical weather data. Features included power flows, voltage levels and PV-specific structural data such as installed capacity and system count. Model performance was evaluated using R², MAE and RMSE. To ensure generalizability, the models were evaluated on data from transformer stations that were not part of the training set. SHAP values were used to interpret model behavior and assess the relevance of individual PV-related features.