Advanced Modeling Approaches for Photovoltaic Energy Forecasting: Multi-task Modeling and Event-Driven Spiking Transformers
Alhammadi, Ayesha Abdulla Ibrahim
Alhammadi, Ayesha Abdulla Ibrahim
Citations
Altmetric:
Supervisor
Department
Machine Learning
Embargo End Date
Type
Thesis
Date
2026
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Accurate forecasting of photovoltaic (PV) energy generation is essential for ensuring reliable and efficient operation of modern smart grid systems. However, PV power output is highly variable and influenced by complex nonlinear interactions among meteorological and environmental factors, making forecasting challenging. Current methodologies are typically centered on a single problem dimension, such as temporal models, thereby neglecting spatial relations, system interactions, or efficient computation. This thesis addresses these challenges through the development and evaluation of two complementary models for forecasting PV energy, which explicitly address the key limitations of existing forecasting methodologies. To start with, a multi-task forecasting paradigm is introduced, which allows capturing of both electricity production and consumption under a single architecture and generating independent predictions for each task, thus gaining a better insight into supply-demand dynamics in a smart grid environment. Secondly, a novel event-driven spiking transformer architecture is proposed, which is capable of effectively capturing long-term temporal dependencies without performing redundant calculations due to sparse neural activations. The general hypothesis is that the use of sophisticated approaches, such as multi-task system modeling and event-driven temporal processing, leads to consistent reductions in forecasting errors across all tasks compared to traditional techniques. Experimental outcomes demonstrate that the proposed multi-task framework reduces the forecasting error compared to baseline models, achieving improvement in MAE and RMSE values of 0.1378 and 0.1386 for generation forecasting, and 0.1393 and 0.1354 for consumption forecasting, respectively. For the spiking transformer, the proposed model achieves MAE/RMSE values of 0.0326/0.0676 on DKASC-ASA-1B, 0.0438/0.0865 on DKASC-ASA-1A, and 0.0227/0.0433 on DKASC-ASA-2, demonstrating consistently low MAE and RMSE across all evaluated datasets while maintaining efficient temporal modeling. Overall, this thesis presents a detailed examination of advanced techniques used in the prediction of PV energy systems, along with the introduction of novel methods that will enhance the robustness of energy predictions.
Citation
Alhammadi, Ayesha Abdulla Ibrahim, "Advanced Modeling Approaches for Photovoltaic Energy Forecasting: Multi-task Modeling and Event-Driven Spiking Transformers," M.S. Thesis, Machine Learning, MBZUAI, 2026.
Source
Conference
Keywords
Spiking neural networks, Renewable energy forecasting, Electricity consumption prediction, Transformer models, Time-series forecasting, Smart grid