DTU Studieprojekt - Economic Dispatch of a Multi-Energy System through Deep Reinforcement Learning

Economic Dispatch of a Multi-Energy System through Deep Reinforcement Learning

Udbyder
Vejleder
Sted
København og omegn
Context
In order to achieve the 100% renewable energy goal in Denmark, where all energy sectors will be supplied with renewable energy in 2050, a multi-energy system (MES) is necessary to make all sectors strongly coupled. An economic dispatch problem of the MES determines the power and heat output of cost-optimized generation units. Historically, the economic dispatch of the MES considers the network constraints. However, the physical modelling of heating and power networks is traditionally nonconvex. This can make the optimization problem a challenge due to the limited accuracy of network parameters and highly computational burden. On the other hand, uncertainties such as wind power generation and heat loads need to be considered during the operation. Such uncertainties are traditionally modelled by stochastic or robust optimization. Such optimization methods require the probability distribution functions of the uncertainties and also have a highly computational burden.

Recently, Deep Reinforcement Learning (DRL) has been applied in the area of power system dispatch. The DRL comprises multiple agents that can optimally dispatch the generating units by learning reactions from the environment. In this way, complex physical modelling of networks and the generation of uncertainties can be avoided during optimization.

Therefore, DRL applied to an MES is necessary to achieve an optimal dispatch of heat and power in an efficient and accurate way. Such a method can contribute to the coupling of heat and power system with lowest operating costs and ensure a secure operation of the MES.

• 
  Objectives
• Methodology
  As a first step, the current scientific literature on DRL applied to energy systems will be reviewed in order to determine the operational strategy. Next, the state space, action space and reward function will be formulated for an agent that reacts to the environment. Then DRL method will be developed to optimize the MES operation. The economic dispatch of the MES system should consider both the cost of generation and load, as well as maintain the energy system balance. System performance will be evaluated based on operating costs. The DRL method will be tested and the results will be compared to the current operational data.
• 
  Expected results

1. Construction of a Markov Decision Process to formulate the economic dispatch problem for a MES.

2. Development of a DRL process for the MES to determine the operational strategies of the MES with lowest operation costs and secure operation.

3. Simulation of MES operation performance.

4. Compared to the current operation data of the MES in Denmark.

The expected outcome includes a review of relevant literature, an overview of the theory of the used methods, and documentation of the implementation and results.

Prerequisites
Deep learning, Reinforcement learning, Python programming, and optimization

Students can further apply the methods in other areas in the future, such as banking.

I samarbejde med
Villum Fonden

Forudsætninger
Deep learning, Reinforcement learning, Python programming, and optimization

Emneord

• Elektroteknologi
• Energi
• Matematik
• Antenner
• Elektromagnetisme
• Elektronik
• Lyd
• Mikrobølgeteknologi
• Robotteknik og automation
• Bioenergi
• Brændselsceller
• Elforsyning
• Energieffektivisering
• Energilagring
• Energiproduktion
• Energisystemer
• Kraftværker
• Solenergi
• Vindenergi
• Kortlægning og opmåling
• Billedanalyse
• Computerberegning
• Dataanalyse
• Hardware og komponenter
• Software og programmering
• Telekommunikation
• Geometri
• Matematisk analyse
• Matematisk logik
• Matematisk modellering
• Statistik
• Høreapparater
• Medicinske apparater og systemer
• Klimatilpasning
• Rumteknologi og instrumenter
• Satellitter

• deeplearning
• economicdispatch
• energysytem
• powersystems

Navn
Jiawei Wang

Stilling
Postdoc

Mail
jiawang@elektro.dtu.dk

Vejleder-info
Kandidatuddannelsen i Elektroteknologi
Vejleder
Jiawei Wang

Medvejledere
Pierre Pinson, Xiufeng Liu

Type
Kandidatspeciale, Specialkursus

Kandidatuddannelsen i Matematisk Modellering og Computing
Vejleder
Jiawei Wang

Medvejledere
Pierre Pinson, Xiufeng Liu

Type
Kandidatspeciale, Specialkursus

Kandidatuddannelsen i Bæredygtig Energi
Vejleder
Jiawei Wang

Medvejledere
Pierre Pinson, Xiufeng Liu

Type
Kandidatspeciale, Specialkursus