MarkovESValuation

Analytical stochastic dynamic programming (SDP) algorithm with Markov process price model for energy storage price arbitrage, implemented in MATLAB with python for data pre-processing and Markov process model training.

Features

Training Markov process model for real-time price (Python) and testing real-time arbitrage with historical price data (MATLAB).
Standard Train, Test, Markov process model for different cases in NYISO.

Requirements

Python 3.6+
MATLAB
Julia 1.6.2 and Gurobi for independent perfect information arbitrage benchmark

Usage

Training and testing data in NYISO's four zones (NYC, LONGIL, NORTH, WEST) already include in this repo, also trained Markov process models for different cases.

Markov process training

Use pre_processing_main.py:

Set training data set duration in line 12,13;
Set location in line 13,14;
Set model training for real-time model or DAP-RTP model in line 15;
Other settings are in line 37 (optional_kwargs), set summer duration, time step, price nodes gap and upper/lower bounds.

Real-time arbitrage with real-time Markov process model

For real-time Markov model, Uuse main_RT.m; for DAP-RTP bias Markov process model use main_DA_bias.m:

Set location, price node number and gap and cases (base, independent, with seasonal/weekly pattern), and Markov model training dataset in line 1-23.

Power-to-energy ratio (P/E), efficiency, presumed marginal discharge cost can be set in line 67-74 (main_RT.m)/line 69-76 (main_DA_bias.m).

Benchmark simulation

For day-ahead benchmark (BEN-DA), use main_BEN_DA.m;

For perfect information benchmark, use test_arb.jl. Similar result can be obtained by change lambda(t) to lambda_DA(t) in main_BEN_DA.m, Line 49.