This repository implements a deterministic (perfect-foresight) day-ahead co-optimization of a Battery Energy Storage System (BESS) between:

• Day-Ahead (DA) energy arbitrage,
• FCR (primary frequency reserve, symmetric),
• aFRR (secondary automatic reserve, UP / DOWN products).

The goal is to compute, day by day, an optimal schedule (SoC, charge/discharge, reserve bids) on a 15-minute grid while respecting battery and reserve feasibility constraints.

This work is largely inspired by a reference thesis used throughout the project : "Optimizing Residential Battery Energy Storage Systems Across Frequency Regulation Markets and Energy Arbitrage" by Elias Schuhmacher and Eric Rosen wrote in 2025.

An example of a day-ahead optimized schedule (January 15, 2021):

Details of three strategies PnL (January 15, 2021):

Comparison of three strategies PnL over January:

A BESS can earn:

• Energy arbitrage: charge low, discharge high.
• Reserve capacity payments: get paid for being available to provide frequency control (even if not activated).

Because energy and reserve share the same power (MW) and energy (MWh) limits, the trade-off is non-trivial: allocating headroom to reserves may reduce arbitrage, and aggressive arbitrage may violate reserve deliverability.

Two public datasets are used:

• French day-ahead prices (from Ember) originally hourly, in €/MWh.
• French balancing capacity remuneration (from RTE Services) 15-min, in €/MW/15min.

Please find the links to the data below:

• https://ember-energy.org/data/european-wholesale-electricity-price-data
• https://www.services-rte.com/fr/telechargez-les-donnees-publiees-par-rte.html?category=market&type=balancing_capacity&subType=procured_reserves

We operate everything on a UTC 15-minute grid (96 steps/day). DA hourly prices are expanded to 15-min as piecewise-constant; reserve products are kept for FCR, aFRR_UP, aFRR_DOWN; small gaps are forward-filled. Processed datasets are stored as compact Parquet files for fast reproducibility in the repo (data directory).

We follow a standard battery model (SoC dynamics + power/energy constraints), grounded in the reference thesis and adapted to an industrial/utility scale.

Typical parameters used in experiments:

• Power: 10 MW
• Energy: 20 MWh
• SoC bounds: 10% – 90%
• Efficiencies: 90% per conversion step
• Degradation proxy: linear throughput penalty, ~15 €/MWh

We consider three products settled on the same 15-min grid:

• Energy (DA): scheduled charging/discharging at price $\pi_t$ (€/MWh)
• FCR capacity: symmetric reserve, assumed net SoC impact ~ 0 on average (but consumes power headroom)
• aFRR capacity: UP/DOWN products, modeled with a simplified activation mechanism

We solve one optimization per day with:

• $T = 96$ time steps, $\Delta t = 0.25$ hours
• initial SoC $S_0$ coming from the previous day (optional multi-day linking)

• $P_t^{ch} \ge 0$, $P_t^{dis} \ge 0$ : charge/discharge power (MW)
• $R_t^{fcr} \ge 0$ : FCR capacity (MW)
• $R_t^{up} \ge 0$, $R_t^{down} \ge 0$ : aFRR UP/DOWN capacity (MW)
• $A_t^{up} \ge 0$, $A_t^{down} \ge 0$ : aFRR activation power (MW)
• $S_t$ : state-of-charge (MWh)

$\text{Revenues}=\sum_{t=1}^{T}\Big[\pi_t,(P_t^{dis}-P_t^{ch}),\Delta t +\rho_t^{fcr}R_t^{fcr} +\rho_t^{up}R_t^{up} +\rho_t^{down}R_t^{down} \Big]$

$\text{Costs}=\sum_{t=1}^{T} c_{th},(P_t^{ch}+P_t^{dis}),\Delta t$

$\max \ \text{Profit}=\max\big(\text{Revenues}-\text{Costs}\big)$

$S_{t+1}=S_t+\eta_{ch}(P_t^{ch}+A_t^{down})\Delta t -\frac{1}{\eta_{dis}}(P_t^{dis}+A_t^{up})\Delta t$

$S_{\min}\le S_t \le S_{\max}\qquad\forall t$

Directional aFRR consumes headroom in the corresponding direction: $P_t^{dis}+R_t^{up}\le P_{\max}^{dis},\qquad P_t^{ch}+R_t^{down}\le P_{\max}^{ch}$

For FCR, we enforce symmetric deliverability (headroom in both directions): $P_t^{dis}+R_t^{up}+\gamma R_t^{fcr}\le P_{\max}^{dis},\qquad P_t^{ch}+R_t^{down}+\gamma R_t^{fcr}\le P_{\max}^{ch}$

Activation cannot exceed capacity: $0\le A_t^{up}\le R_t^{up},\qquad 0\le A_t^{down}\le R_t^{down}$

We enforce a uniform activation in time with 25% activation ratios ($\alpha_{up}=0.25$, $\alpha_{down}=0.25$) : $\sum_{t=1}^{T}A_t^{up}=\alpha_{up}\sum_{t=1}^{T}R_t^{up},\qquad \sum_{t=1}^{T}A_t^{down}=\alpha_{down}\sum_{t=1}^{T}R_t^{down}$

Upward reserve requires enough energy above $S_{\min}$, downward reserve requires enough empty space below $S_{\max}$: $S_t \ge S_{\min}+(\gamma R_t^{fcr}+R_t^{up}),\tau,\qquad S_t \le S_{\max}-(\gamma R_t^{fcr}+R_t^{down}),\tau$

with $\tau$ a deliverability horizon (here 15 minutes).

$S^{(d+1)}_0=S^{(d)}_T$

• Python + CVXPY formulation
• Solved as a linear program, using ECOS solver
• Modular, object-oriented structure:
  • Battery stores technical parameters
  • DaySolver solves one day
  • DaySolution stores schedule + metrics + status
  • Orchestrator runs day-by-day and aggregates results

This repo is uv-friendly !

# 1) Create venv
uv venv .venv
.venv\Scripts\activate

# 2) Install dependencies
uv sync

# 3) Run
uv run python main.py