This thesis focuses on the two fundamental issues in manufacturing. One is to forecast the demand quantity of products in the coming time periods. After the demand forecasting, the next key important step is to schedule the corresponding production planning.

In the first part, we consider a long-term demand forecasting problem. Due to the fact that the sample data may be very limited and of high variation, traditional prediction models can not perform well here. Inspired by the ensemble idea in Random Forest, we apply this ensemble idea together with the random subset sampling on the training data and propose a new forecasting model named Data-ensemble Random Forest (DERF). We prove that this new model retains the convergence property of Random Forest, i.e., as the number of t...[ Read more ]

This thesis focuses on the two fundamental issues in manufacturing. One is to forecast the demand quantity of products in the coming time periods. After the demand forecasting, the next key important step is to schedule the corresponding production planning.

In the first part, we consider a long-term demand forecasting problem. Due to the fact that the sample data may be very limited and of high variation, traditional prediction models can not perform well here. Inspired by the ensemble idea in Random Forest, we apply this ensemble idea together with the random subset sampling on the training data and propose a new forecasting model named Data-ensemble Random Forest (DERF). We prove that this new model retains the convergence property of Random Forest, i.e., as the number of training rounds increases, DERF will alway converge. Although DERF is based on Random Forest, its construction idea could also be extended to any other model. By numerical experiment, we show that DERF achieves a significant improvement in performance compared with the traditional prediction models, such as ARIMA, Random Forest, Gradient Boosting Decision Tree, and Long Term Short Memory.

In the second part, in order to solve a large scale production scheduling integer programming, whose objective is to minimize the total holding cost, within a reasonable time, we decompose this problem into four steps. We first regard all different plants as a whole one, and obtain an upper bound of the amount of each product that should be produced. Then, we assign the above production tasks for each plant based on its own constraints. Afterwards, other decision variables such as the supply amount of each product, and the transfer amount of each product between plants will be updated accordingly. Numerically, the gap between the optimal solution of the original problem and the one obtained by our decomposition algorithm is quite small.