Research On Dynamic Lot Sizing Decision Based On Uncertain Time-varying Demand

Dynamic lot sizing problem (LSP) is an important topic in the research of production and inventory operation and underlines the basis for optimaization of firms as well as the whole supply chain they belong to. Based on time-varying demand, dynamic lot sizing decision faces two basic problems:one is to decide when to produce/order (or provide service), and the other how many items to produce/order (or how well the service can be). As for time-varying demand, it is categorized into two types of demand:one is deterministic time-varying demand and the other uncertain time-varying demand. Many of recent studies focused on LSP with the first type while a few of them with the second one. Following with the definition and analysis of the term "uncertain time-varying demand", the focus research on uncapacitated single item lot sizing problem (USILSP) based on stochastic time-varying demand and on unknown-distribution one is determined, and that of solution could be a new start as well as the base stone for studying LSP with more complex cases such as uncapacitated LSP, multi-item or multi-stage LSP, and LSP in supply chain.Furthur more, multi-item or multi-stage LSP on rolling horizon and on high degree of nonlinear and nonstationary demand are also examined. To sharpen the knowledge of the research topics, with the review of literature about time-varying demand, stochastic demand, and forecast methods for dealing with unknown distribution demand, multi-stage LSP, and supply chain LSP, the LSP under uncertain time-varying demand are divided into three subproblems according to the demand property of stationarity and non-stationarity as well as of known probability and unknown probability. The first one is stochastic LSP based on stationary normal distribution, and the second one stochastic LSP on nonstationary erlang distribution, and then the third one uncertain LSP on unkown probability distribution and high degree of non-stationarity and nonlinearity demand.To solve the first subproblem metioned above, combining with consideration of truncated horizon effect and with the idea of USILSP algorithms especially including PPB (Part Period Balancing) and SM (Silver-Meal) algorithm, we proposed a newly designed algorithm called T-period Skipping Algorithm (TSA) through the implementation of "look back and then look forward" procedure. With series of assumptions and related rational proof, the framework of TSA has been initially established. The corresponding numerical tests are designed and processed by comparison of TSA and some of other mainstream algorithms, and the results shows that TSA has relatively better cost performance and is more stable. As to the second subproblem, optimal model based on erlang distribution is proposed and the specified bisection search algorithm designed for the solution to the determination of optimal cummulated quantity, thus facilitating the corresponding optimal LSP decision.For the third subproblem, through use of demand forecast methods, LSP under unknown demand can be transformed into the one under deterministic demand. It can be divided into two parts:one is to find more optimal demand forecast methods, and the other to study the impacts on lot sizing decision by forecast error, therefore showing better forecast method facilitate better lot sizing decision. With application of hybrid forecast methods containing the mix method of SARIMA (Seasonal Auto Regression Integrated Moving Average) and SVM (Support Vector Machine) and the other mix of ARIMA and EEMD (Ensemble Empirical Mode Decompositon), we find those two mix methods perform better based on the demand respectively with significant seasonal and non-linear property and with high degree of fluctuation and nonlinearity comparing with non-mix forecast methods. Moreover, after using model construction and relevant analysis, different LSP algorithms are selected and the stong relationship between those cost performances and forecast errors are tested by numerical experiments. Hence, for the case of sigificant forecast error, the two mixed forecast methods metioned above can effectively improve forecast performance and cut down lot sizing relevant cost. Those lot sizing algorithms can also function well.Finally, this thesis dicusses the main contribution of our research work and provides some future possible extensions.