Back to search

Cornell University

Improving the Computational Efficiency of Approximate Dynamic Programming using Neural Networks, with Application to a Multi-Reservoir Hydropower and Wind Power System

Abstract

dc:description.abstract

This dissertation develops a more computationally efficient Approximate Dynamic Programming (ADP) method that can be applied to optimal control problems that are stochastic, nonconvex, and either finite or rolling horizon with many stages. The primary application is to a realistic model of Bonneville Power Administration, which is a large wind power and hydropower producer in the Pacific Northwest. The objective is to determine the optimal amount of power to buy and sell on the day and hour ahead markets conditioned on the wind power forecast, as well as how much water to release from each reservoir. The second Chapter in this dissertation develops the Fitting via Unimodal Approximation Optimization (FUA) method for more accurately approximating the value function in ADP with a Feedforward Neural Network with one hidden layer (FFNN1). A major part of FUA is the new Unimodal Approximation Optimization (UAO) algorithm that is used to perform FFNN1 hyperparameter optimization. UAO can be applied to optimization problems with a discrete domain and a noisy unimodal objective function, and it is proven that UAO converges almost surely to the correct solution. Results on two control problems with 4, 12 and 15 state space dimensions show that approximating the value function in ADP with an FFNN1 using FUA yields a more accurate control solution in less time as compared to using other methods of fitting an FFNN1. Chapter 3 presents the Long Term Generation method that generates long-term synthetic wind power scenarios conditioned on historical sequential short-term wind power forecasts. Power systems with wind power integration can be simulated on this data in order to more accurately evaluate the performance of their control policies. Additionally, the Joint Distribution Comparison test is developed to evaluate the quality of these synthetic scenarios. Finally, in Chapter 4 a stochastic rolling horizon model of Bonneville Power Administration is developed and the resulting control problem is solved using the ADP algorithm developed in Chapter 2 and is evaluated using data generated in Chapter 3. The stochastic control formulation has a nonlinear objective function, 24 decision variables, and 16 state space dimensions.

Degree

thesis:*
Name thesis:degree_name
Ph. D., Applied Mathematics
Level thesis:degree_level
Doctor of Philosophy
Discipline thesis:degree_discipline
Applied Mathematics
Grantor
Cornell University
Year dc:date.issued
2017

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Perline, Kyle Robert
Chair dc:contributor.chair
  • Shoemaker, Christine Ann
Committee members dc:contributor.committeemember
  • Topaloglu, Huseyin
  • Vladimirsky, Alexander B.

Subjects

dc:subject × 4

Rights

dc:rights
Statement dc:rights
  • Attribution 4.0 International
Language dc:language.iso
en_US

Identifiers

dc:identifier.*
Dc Identifier Bibid
10474156
Dc Identifier Other
ProQuest Submission ID: 10637
ProQuest Publication ID: 10682308

Chain of custody

source
Harvested from
Cornell University
Base URL
ecommons.cornell.edu/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Perline, Kyle Robert. Improving the Computational Efficiency of Approximate Dynamic Programming using Neural Networks, with Application to a Multi-Reservoir Hydropower and Wind Power System. Doctor of Philosophy thesis, Cornell University, 2017. https://hdl.handle.net/1813/59053