Skip to content

Code for "Efficient Trading of Aggregate Bidirectional EV Charging Flexibility with Reinforcement Learning". It contains incentive-compatible contracts, flexibility aggregation and disaggregation algorithms, and a RL trading agent for profit maximization in real-world scenarios.

License

Notifications You must be signed in to change notification settings

sustainable-computing/VPP-VariableTermContracts

 
 

Repository files navigation

EvCharge ⚡ 🚗 ⚡

This is the accompanying repo for the paper "Efficient Trading of Aggregate Bidirectional EV Charging Flexibility with Reinforcement Learning", due to appear on ACM's e-Energy 2024 proceedings. You can find it in this repo as AggregateFlex_eEnergy24.pdf.

🌩️ Abstract

We study a virtual power plant (VPP) that trades the bidirectional charging flexibility of privately owned plug-in electric vehicles (EVs) in a real-time electricity market to maximize its profit. To incentivize EVs to allow bidirectional charging, we design incentive-compatible, variable-term contracts between the VPP and EVs. Through deliberate aggregation of the energy storage capacity of individual EVs, we learn a reinforcement learning (RL) policy to efficiently trade the flexibility, independent of the number of accepted contracts and connected EVs. The proposed aggregation method ensures the satisfaction of individual EV charging requirements by constraining the optimal action returned by the RL policy within certain bounds. We then develop a disaggregation scheme to allocate power to bidirectional chargers in a proportionally fair manner, given the total amount of energy traded in the market. Evaluation on a real-world dataset demonstrates robust performance of the proposed method despite uncertainties in electricity prices and shifts in the distribution of EV mobility.

🔭 Overview

The code is divided into these parts

  • 📜 ContractDesign/: The code and notebooks used to generate and analyze the V2G contracts.
  • 🔌 ElectricityMarkets/: The analysis for the electricity price dataset at the proper time resolution.
  • 🏋️ EvGym/: The environment and the RL agents, more details in a later section.
  • 🧪 ExpLogs/: Records of the experiments used in the paper.
  • 🔬 ResultsAnalysis/: Notebooks used for analysis and visualization.
  • 📚 data/: Datasets used simulations.
  • ♟️ scripts/: The Bash scripts used for experiments, multiple simulation runs.
  • time/: Results for time profiling of different agents.
  • 🍲 PreprocElaad.ipynb: Notebook for preprocessing Elaad, charging sessions dataset
  • RunChargeWorld.py: Script for running simulations without RL.
  • 🌟 RunSACChargeWorld.py: Script for running simulation with RL

Additionally a requirements.txt file is provided. Using a virtualenv is recommended.

⚙️ Parameters

Parameter Value Description
--agent SAC-sagg Agent to use for real-time scheduling
--save-name sac_a Name used for logs, results, etc.
--pred-noise 0.00 Nosie for price predictions in training
--seed 42 Seed for random number generators
--years 200 Number of episodes to train
--batch-size 512 Batch size to sample from the replay buffer
--alpha 0.02 Temperature parameter in SAC
--policy-frequency 4 How often to update the policy (timesteps)
--target-network-frequency 2 How often to update the second Q NN (timesteps)
--disagg PF (Proportional fairness) Disaggregation algorithm
--buffer-size 1e6 Number of experiences to save in replay buffer
--save-agent True Save the weights of the trained agent
--general True Run training mode (False is for deployment)

🧠 Architecture

🤺 Actor

The architecture for the actor, the policy network.

Layer In Out
Linear (ReLU) 59 256
Linear (ReLU) 256 256
Head 1, Mean: Linear (Sigmoid) 256 1
Head 2, Logstd: Linear (Tanh) 256 1

🕵️ Critic

The architecture for the two critics, soft Q networks.

Layer In Out
Linear (ReLU) 60 256
Linear (ReLU) 256 256
Linear 256 1

⛷️ Running the Scripts

The command to run the experiment with the parameters is shown below.

python3 RunSACChargeWorld.py --agent SAC-sagg --save-name sac_a 
                             --pred-noise 0.00 --seed 42 --years 200 
                             --batch-size 512 --alpha 0.02 
                             --policy-frequency 4  
                             --target-network-frequency 2 
                             --disagg PF --save-agent True 
                             --general True

🌇 Notes for RunSACChargeWorld.py

This is of how we train our implementation of Aggregate SAC. First, we import some general modules. Then we import the user-defined modules, mainly the environment (ChargeWorldEnv), the actor (agentSAC_sagg), and the critic (SoftQNetwork).

In the body of the program, we initialize ChargeWorldEnv with the dataset that contains the charging sessions (df_sessions), the dataset that contains the real-time prices (df_prices), the contract parameters (contract_info), and a random number generator (rng).

For the agent, we initialize the actor (agentSAC_agg) with the price dataset (df_price), arguments read from the command line (args), and the device (device). This device is needed for certain PyTorch functionalities. The critic is composed of two Q networks (SoftQNetwork). Additionally, soft actor-critic uses a replay buffer (rb).

We can train the agent for many episodes, each with a predetermined number of timesteps. Similar to Farama's Gym, the environment is initialized with a world.reset(). During training, the agent receives an observation from the environment, and it outputs an action with agent.get_action(). The environment receives the action and moves forward one timestep with world.step(). The loop keeps going on until all the timesteps are completed for all the episodes. At each iteration, the training of the agent is performed.

The charging sessions dataset, real-time prices dataset and state are implemented in Pandas DataFrames. Additionally, the environment also receives a Pandas DataFrame for the action. Conversely, the agent works mainly with PyTorch Tensors. To convert the state DataFrame into a PyTorch Tensor, we employ agent.df_to_state(). Similarly, to convert the agent's action into the required Pandas format that the environment prefers, we use agent.action_to_env().

About

Code for "Efficient Trading of Aggregate Bidirectional EV Charging Flexibility with Reinforcement Learning". It contains incentive-compatible contracts, flexibility aggregation and disaggregation algorithms, and a RL trading agent for profit maximization in real-world scenarios.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Jupyter Notebook 89.5%
  • Shell 5.5%
  • Python 5.0%