Satellite-Based Solar Radiation & Grid Congestion Analysis Project
AI & Data
Semester programme:AI, Machine Learning & Data
Research group:Sustainable Data & AI Application
Project group members:Nikola Momchilov
Priyanka Darbari
Hanh Truong
Ahn Huynh
Mohammed Aslan
Project description
The project explores whether you can look at a city from space and figure out where the electricity grid is struggling. As more people install solar panels, some parts of the grid get overloaded because the infrastructure was never designed for the flow of electricity.
Using satellite images, public data, and machine learning, the project predicts which areas in Amsterdam face the highest risk of grid congestion, and then uses solar radiation data to explain why those areas are under stress.
Context
The project is a data science and AI application in the energy domain. It combines satellite imagery, geospatial data processing, and machine learning to solve a real-world classification problem: predicting grid congestion risk from publicly available spatial data.
The work involves processing multispectral satellite images, extracting features through spectral indices, integrating multiple open datasets (CBS statistics, Liander energy data, EUMETSAT radiation data), and training a supervised learning model to classify areas by risk level. It falls within the HBO-ICT field of applied AI and data engineering, using Python, Jupyter notebooks, and web technologies to turn raw satellite and tabular data into an interactive analytical tool.
Results
Product Outcomes
The project delivers a complete AI pipeline that predicts electricity grid congestion risk from satellite imagery and public spatial data. A machine learning model analyses Sentinel-2 satellite images of Amsterdam, extracts spatial features like vegetation density and urban surface cover, and combines them with demographic statistics to classify each 1x1 km area as low, medium, or high congestion risk. A second pipeline estimates solar radiation from Meteosat weather satellite data and overlays it on the same map. The results are presented through an interactive web application where users can explore congestion risk, solar radiation patterns, and the underlying spatial factors for any area in the study region.
Key Insights
The central finding is that solar radiation is identical across all of Amsterdam. Every neighbourhood receives the same amount of sunlight. Yet congestion risk varies dramatically between areas. This proves that grid overload is not a natural resource problem but a human infrastructure problem: it depends on how many solar panels people install and how much electricity they consume, not on how much sun their neighbourhood gets.
The combined analysis classifies congestion zones into two types: areas where the grid is overloaded by electricity demand, and areas where the grid is overloaded by solar panels feeding power back into cables that were not designed for it. Each type requires a fundamentally different investment strategy, and the tool identifies which solution fits which area.
Validation
The model achieves an F1 score of 0.578 on cross-validation and 0.55 on temporally held-out satellite data, confirming it learns real spatial patterns rather than noise. Predictions show broad spatial agreement with the official Dutch grid capacity map published by Netbeheer Nederland. Radiation values are consistent with national reference measurements. All analysis uses only publicly available data and open-source tools, making the methodology fully reproducible.
TRL Positioning
The project is positioned at TRL 4: technology validated in a laboratory environment using real data. Two working pipelines process actual satellite imagery and climate data for Amsterdam, producing validated spatial predictions. An interactive prototype demonstrates how the outputs can support energy planning decisions. The path to TRL 5 involves evaluation by grid operators and testing the methodology on additional cities to confirm it generalises beyond Amsterdam.