Coordination and planning of Smart grids with material-specific photovoltaic systems using machine learning technologies

Abstract

The global energy transition necessitates the large-scale integration of solar photovoltaics (PV), creating challenges for the planning and operation of power distribution networks. This complexity is amplified by the arrival of emerging PV technologies, such as perovskite and organic solar cells, which exhibit performance characteristics fundamentally different from conventional silicon and are poorly represented by existing models. Focusing on Distribution System Operators and PV researchers, this thesis bridges the gap between PV material science and grid engineering. It delivers a multi-scale framework that quantifies the network-level impact of emerging PV, offering a pathway to decarbonisation that optimises asset deployment and operational cost.

A novel generalised PV model is proposed, introducing a material-specific performance coefficient that translates PV material behaviours into grid impacts. This highlights how material choice directly dictates the Self-Sufficiency Ratio and total network power losses, depending on location characteristics. System-level simulations reveal that relying on conventional, material-agnostic models leads to an underestimation of energy yield in high-latitude regions. Neglecting material-specific responses to different irradiance levels results in suboptimal sizing and siting decisions, ultimately compromising grid efficiency and increasing reliance on external power imports. By evaluating distinct PV materials with the proposed PV model against low-light efficiency and temperature coefficients, the framework identifies the optimal material match for local climates.

To scale this analysis, the Material-Aware Multi-Attention Spatio-Temporal Graph Attention Network (MAMSTGAT) is introduced. This proposed framework, concurrently learns three coupled objectives bus load classification, Loss
Sensitivity Factor (LSF) prediction, and optimal PV sizing. This enables a holistic planning solution that outperforms single-task baselines. By inherently modelling the grid's topology, MAMSTGAT reveals how evaluating distinct PV
technologies against potential PV generation based on irradiance and temperature response, allows for optimal deployment based on location and climate characteristics.

Optimal deployment is insufficient without effective management. Therefore, this work integrates long-term strategic planning to establish optimal capacity and location with real-time operational control, ensuring that the deployed assets can dynamically adapt to stochastic weather fluctuations and market incentives. To achieve this, a Model Predictive Control framework is developed. Crucially, the framework optimises the energy generation based on the material information by forecasting multi-horizon probabilistic weather predictions generated via a Temporal Fusion Transformer into the material-specific performance models, reducing forecasting errors by up to 44% compared to standard models. This coupling ensures that the unique material capabilities of different PVs are accurately identified in the dispatch schedule, achieving cost minimisation by the increased forecast precision. The framework successfully co-optimises the dispatch of material-specific PV, BESS (accounting for long-term degradation costs), and an incentive-based demand response program to minimise daily operational expenditure.

By connecting advances in PV material science with data-driven grid management, this thesis provides a comprehensive and validated framework for the planning, forecasting, and coordination required improve solar technologies integration in future smart grids.