Wind resource · Weather & climate data validation · Research software
I build and validate models of weather and climate data, and the software that makes them usable by other people. My PhD at Imperial College London corrected ERA5 reanalysis wind against observed generation across 12 European countries, then measured what that correction changes inside European energy system models.
I released the method as PyVWF, an open-source Python library with continuous integration, typing, documentation and a Zenodo DOI. I am looking for data science, modelling or research software roles in climate, energy, weather and climate risk.
Corrected ERA5 reanalysis wind against observed generation and turned it into capacity factors, with wake losses and spatial resolution treated as modelling choices that move results.
Corrections from 1,729 control points in 12 countries, gridded over 23,989 ERA5 cells; capacity-factor MAE cut by 16 to 79% against uncorrected ERA5.
Owning the question of how far a dataset can be trusted, and for which decision. Bias, error decomposition and validation designs that do not leak.
Spatial and leave-one-country-out cross-validation; regions the observations could not support were masked, not published.
Pre-registered tests, failed predictions published as failures, and the simpler model kept when the complex one does not earn its place.
Ran the Oasis LMF PiWind catastrophe model end to end with an exposure sensitivity study on AAL and EP curves.
Gridded climate fields, interpolation and projections, and supervised classification of satellite imagery.
Credited App Developer on I-Hex, the Island Health Explorer: land-cover and benthic classification over Maldivian atolls in Google Earth Engine.
Turning research code into packages other people can install, test, cite and extend, and working inside other people's codebases to their conventions.
PyVWF: CI, typing, Sphinx docs built with warnings as errors, Docker, semantic versioning and a Zenodo DOI. A public fork of PyPSA-Eur.
A Python rewrite of the Virtual Wind Farm model behind the wind simulations on Renewables.ninja, implementing the granular bias-correction method from my 2024 Energy paper. It trains correction factors against observed generation at turbine, regional or national scale and applies them to ERA5. I wrote every commit in the v0.5.1 release, which carries a Zenodo DOI.
Python · Xarray · Dask · SciPy · scikit-learn · pytest · GitHub Actions · ruff · mypy · Sphinx · Docker
Extended the open European energy system model with a density-tiered wake-loss correction, the gridded PyVWF corrections, and a spatial splitting algorithm that decouples offshore resource resolution from network topology. Factorial studies (bias × wake × resolution) across a regional North Sea system and a 25-country network. Every manuscript figure regenerates from one script, and the headline claims are held under automated test.
Python · PyPSA · Atlite · Snakemake · Gurobi · GeoPandas
A constrained model meant to transfer corrections to regions with no observations. Gates and thirteen predictions were fixed before each run. As registered it beat uncorrected ERA5 in 5 of 5 held-out regions, with thin margins in two. Seven predictions failed and are published as failures. Two physics extensions were built, tested and rejected.
PyTorch · pre-registration · leave-one-region-out validation
Reproduced the Oasis LMF PiWind example model end to end, then ran a single-site exposure sensitivity study comparing AAL and EP curves, ground-up against insured. It is a reproduction and a sensitivity study, not a catastrophe model; the event set and vulnerability curves are PiWind's example data.
Python · oasislmf · AAL · OEP / AEP
A Google Earth Engine suite giving policymakers and island communities free access to satellite and climate data for the Maldives. I worked on the supervised land-cover and benthic-feature classification, and digitised and assembled its 861-point labelled training set across nine classes. The classified imagery uses separate Landsat and Sentinel-2 classifiers, with per-atoll median compositing to handle cloud.
Google Earth Engine · JavaScript · Landsat · Sentinel-2 · pandas
The codebase behind the MOR-GANs paper, extended past publication: a PyTorch port, and benchmarks against Gaussian process, mixture density network and conditional diffusion baselines on twelve datasets over five seeds. The README reports where the newer baselines win.
TensorFlow · Keras · PyTorch · WGAN-GP
An evaluation harness testing whether populations of LLM personas reproduce the distribution of opinion in two nationally representative UK surveys. Five prompt-level persona methods, two Claude models, distribution metrics (TVD, JSD) with bootstrap confidence intervals, and diagnostics for a model being right for the wrong reasons.
Python · Anthropic API · Docker · bootstrap CIs
I tested whether machine learning could beat my geometric correction. Random cross-validation gave R² of 0.35 to 0.41. Neighbouring sites share weather, so the folds were not independent. Under spatial cross-validation the best model scored 0.295, and on an unseen country the tree models fell below zero. I reported the lower figures and kept interpolation as the published method.
For the physics-informed work, the gates and thirteen predictions were fixed before each run. Seven predictions failed; they are published as failures. A configuration chosen after seeing results was re-tested on four new regions, lost to the simpler one in three, and the simpler one became the headline.
PyVWF started as research code only I could read. When MSc and PhD students began building on it, I made it a package. When a student could not load their own data, I wrote a documented input framework with an example notebook, then used the same pipeline to integrate a lot more data myself.
Imperial College London · EPSRC and Shell Industrial CASE award
Shell
Imperial College London
Link Humans
Imperial College London · Earth Science & Engineering
Thesis: Data Science-Enhanced Wind Power Modelling: From Reanalysis Correction to Energy System Representation. Findings
Imperial College London
Top marks in Modern Programming Methods (92.4) and Inversion and Optimisation (82.2). Dissertation on Wasserstein GANs for multi-output regression, which led to a co-authored paper.
University of Surrey
Benmoufok, E.F., Warder, S.C., Zhu, E., Bhaskaran, B., Staffell, I., Piggott, M.D.
Energy, 313, 133759
View publicationWang, Y., Warder, S.C., Benmoufok, E.F., Wynn, A., Buxton, O.R.H., Staffell, I., Piggott, M.D.
Energy Conversion and Management, 352, 121066
View publicationPhillips, T.R.F., Heaney, C.E., Benmoufok, E., Li, Q., Hua, L., Porter, A.E., Chung, K.F., Pain, C.C.
Applied Sciences, 12(18), 9209
View publication