Blog 18: Paper presentation at ICRERA 2025, Vienna
Abdulaziz, 28 November 2025
In late October 2025, I travelled to Vienna, Austria, to attend the 14th International Conference on Renewable Energy Research and Applications (ICRERA 2025). I presented my paper titled “Machine Learning for Solar Power Prediction: Leveraging Large Scale Real Solar Generation Data and Weather Inputs in Saudi Arabia”.
This work is part of my PhD, where I study how artificial intelligence can support energy planning and help balance supply and demand in large scale systems. My research focuses on solar forecasting, predicting how much solar energy will be generated in the next few hours which is important for improving grid stability in renewable systems.
​
The study used two years of hourly data from a large solar power plant in Saudi Arabia, combined with NASA weather data. Saudi Arabia’s consistent and strong solar radiation makes it an ideal region to study forecasting reliability. The dataset included features such as surface solar radiation (the total sunlight reaching the ground), clearness index (how clear the sky is), air temperature, humidity, and wind speed.
​
I developed sixteen different combinations of input features using both historical power data and weather variables. The models tested included one classical machine learning model Support Vector Regression (SVR) and three deep learning models: LSTM, BLSTM, and CNN LSTM. The aim was to identify which inputs and model type produced the most accurate forecasts.
Attendance of the event and receipt of the presentation certificate.
The SVR model gave the lowest average error when using two lagged power values (one hour and two hours before) together with surface solar radiation. It achieved a Mean Absolute Error of about 7.8 MW and an R^2 of 0.96, showing very high accuracy (R^2 is a statistical metric used for determining how well the model can predict outcomes). The best deep learning model, CNN LSTM, reached a slightly higher R^2 of around 0.97 using only the clearness index. This showed that deep learning could automatically learn time patterns without explicit lagged inputs. Overall, sunlight-related features were the most influential, while variables such as temperature and humidity had a smaller effect.
​
During my presentation, I received valuable feedback suggesting the inclusion of a simple baseline method, such as a statistical model with rolling average, and evaluating how model accuracy changes at different with ML models. These insights will help refine my work. An highlight of the conference was a tutorial on AI applications for energy forecasting and optimisation delivered by university professors. Their work closely relates to my PhD topic, and I had the opportunity to discuss how similar techniques could be applied to decision support in energy planning.
​
Attending ICRERA 2025 was a highly rewarding experience. It allowed me to share my research, learn from experts in the field, and build new professional connections. The feedback and discussions I had in Vienna will play an important role in guiding the next steps of my PhD.
abdulaziz dot alhayd at kcl dot ac dot uk
​