Aurelian-Andrei Panait

Intraday vs Imbalance Arbitrage on the Romanian Power Market

Trained a Gradient Boosting classifier to predict the sign of grid imbalance on the Romanian power market 1h45min before delivery, using forecast errors, lagged grid state, and time encodings under strict no-lookahead constraints (FCST_LAG=4, REAL_LAG=7 quarters). On 87,934 fifteen-minute intervals between July 2022 and January 2025, the model reaches 74.3% direction accuracy on validation and 67.8% trade-level win rate on a never-seen July–December 2024 backtest. Converted into a confidence-weighted intraday strategy capped at 10 MWh per interval, it generates 46.66 M RON of cumulative PnL across 17,308 trades (profit factor 2.99, max drawdown −695.8 kRON, capturing 31.2% of the oracle benchmark). Recruitment case study for ENGIE (March 2026).

Post-Activation Control of an Industrial Demand Response Event

Reproduced the full contrôle du réalisé chain for a 50 MW curtailment activation (24/01/2024) on a 200 MW industrial client, using RTE's Médiane 10J baseline method over 10 reference days. At 10-minute resolution across 12 timesteps, the client delivered 96.33 MWh of curtailment out of 100 MWh expected (96.3% realisation rate, well above the 80% compliance threshold), earning 33,715.88 € in remuneration. Per-timestep analysis revealed a ramping signature — first-hour under-delivery offset by second-hour over-delivery — consistent with physical inertia of industrial loads. Recruitment case study for a French renewable energy aggregator (April 2026).

Post-Delivery Control of a Three-Park Renewable Portfolio

Analysed 15-minute operational data (one week, 864 intervals) for a three-park renewable portfolio (15 MW solar + 13+85 MW wind) to audit post-delivery market compliance on the French day-ahead framework (SPOT/PREP/PREN, TAO adjustment orders). Applied an imbalance-and-settlement model in Python/Pandas and identified two distinct PARC_3 failure modes: complete absence of response to negative SPOT (139.8 MWh vs 1.5 MWh on PARC_2) and premature restart du ring TAO orders (237 MWh vs 1.44 MWh). Weekly imbalance cost −115,914 € concentrated on 2 days. Forecast-quality comparison (54.0% vs 53.3% relative error) rules out the forecasting model as root cause. Recruitment case study for a French renewable energy aggregator (April 2026).

CV Matcher — AI-Powered Application Workflow

A personal full-stack web app that matches new job descriptions against a curated library of 31 tailored CVs. Two-stage retrieval: cosine similarity over term frequencies (browser-side pre-filter) narrows the library from 31 to 8, then Claude Sonnet 4 (Python serverless on Vercel) re-ranks to a top 3 with structured, actionable output — transferable bullets, missing skills, wrong-industry markers, and surgical edits. Built with the help of Claude (2026).

Artificial Intelligence–Based Analysis of Solar Shading

Neural network models trained to predict building energy consumption under various solar shading configurations. The shading-aware MLP achieves R²=0.9999 and a 72% RMSE reduction over the baseline. Supervised by Ayoub Hannad, PhD — MINES Paris · PSL (2025–2026).

V2G Fleet Dispatch Optimization for Primary Frequency Containment Reserve (FCR)

Modelled and compared Uniform vs. Smart V2G dispatch strategies for a 150-EV fleet participating in Primary Frequency Containment Reserves (FCR). Vectorized simulation in NumPy/Pandas over real grid frequency data (France 2019–2021, 10s resolution). Smart strategy: efficiency 95% vs. 86% (Uniform), OBC operating time reduced to ~6.7% (bang-bang control). Battery degradation via Thévenin equivalent circuit and throughput-based cost model (€150/kWh, 3000 cycles). Net revenue ~€4/EV/month after depreciation. Supervised by Pierre Dumont, PhD — MINES Paris · PSL (2026).