The objective of this thesis (related to the AWARD project in the MEDIANETS lab) is to optimize item placement in a warehouse of a large logistics provider. Nowadays due to the expanding market and increasing need for fast and accurate logistics, warehouses and their efficient operation becomes more and more important. It is therefore important to place products in an optimal way in the warehouse for incoming orders to be processed as fast as possible. Prediction of item order quantities is relevant, because it determines the amount of storage space that needs to be allocated for a given item. The order quantities expected in the future can be determined through modeling order quantity timeseries and performing a prediction using these models.
The first step to make accurate predictions was to filter out the non dominant items from the order dataset. Non dominant items have incomlete time series due to their low order amount. As a next step I applied different models to the extracted product timeseries and compared them. For this problem I used the Python programming language and several libraries, such as pandas, sklearn, numpy and pyramid. The implementation of prediction models exist in these libraries. I tried ARIMA, neural network regression, k-nearest neighbor regression and decision tree regression to find the best method for predicting future order counts for dominant items in the dataset.
First an ARIMA model was tried, but it was not able to predict based on the given historical data. After that three regressors were trained on the dominant products’ timeseries and compared to each other. Based on the results it can be stated that for these timeseries the KNN regressor has the best performance. The processing time of the KNN algorithm was low as well as the memory requirements associated with the training and prediction phase. This method is suitable is resource constraints are strict or the prediction needs to be done real time. The prediction models trained were used in the AWARD project to help in the item placement optimalization of a real warehouse.